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 */
114 struct scan_frame *this_prev_frame; /* this previous frame */
115 struct scan_frame *prev_frame; /* previous frame */
116 struct scan_frame *next_frame; /* next frame */
119 /* Certain characters are output as a sequence with the first being a
121 #define isBACKSLASHED_PUNCT(c) strchr("-[]\\^", c)
124 struct RExC_state_t {
125 U32 flags; /* RXf_* are we folding, multilining? */
126 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
127 char *precomp; /* uncompiled string. */
128 char *precomp_end; /* pointer to end of uncompiled string. */
129 REGEXP *rx_sv; /* The SV that is the regexp. */
130 regexp *rx; /* perl core regexp structure */
131 regexp_internal *rxi; /* internal data for regexp object
133 char *start; /* Start of input for compile */
134 char *end; /* End of input for compile */
135 char *parse; /* Input-scan pointer. */
136 char *adjusted_start; /* 'start', adjusted. See code use */
137 STRLEN precomp_adj; /* an offset beyond precomp. See code use */
138 SSize_t whilem_seen; /* number of WHILEM in this expr */
139 regnode *emit_start; /* Start of emitted-code area */
140 regnode *emit_bound; /* First regnode outside of the
142 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
143 implies compiling, so don't emit */
144 regnode_ssc emit_dummy; /* placeholder for emit to point to;
145 large enough for the largest
146 non-EXACTish node, so can use it as
148 I32 naughty; /* How bad is this pattern? */
149 I32 sawback; /* Did we see \1, ...? */
151 SSize_t size; /* Code size. */
152 I32 npar; /* Capture buffer count, (OPEN) plus
153 one. ("par" 0 is the whole
155 I32 nestroot; /* root parens we are in - used by
159 regnode **open_parens; /* pointers to open parens */
160 regnode **close_parens; /* pointers to close parens */
161 regnode *end_op; /* END node in program */
162 I32 utf8; /* whether the pattern is utf8 or not */
163 I32 orig_utf8; /* whether the pattern was originally in utf8 */
164 /* XXX use this for future optimisation of case
165 * where pattern must be upgraded to utf8. */
166 I32 uni_semantics; /* If a d charset modifier should use unicode
167 rules, even if the pattern is not in
169 HV *paren_names; /* Paren names */
171 regnode **recurse; /* Recurse regops */
172 I32 recurse_count; /* Number of recurse regops we have generated */
173 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
175 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
178 I32 override_recoding;
180 I32 recode_x_to_native;
182 I32 in_multi_char_class;
183 struct reg_code_blocks *code_blocks;/* positions of literal (?{})
185 int code_index; /* next code_blocks[] slot */
186 SSize_t maxlen; /* mininum possible number of chars in string to match */
187 scan_frame *frame_head;
188 scan_frame *frame_last;
191 #ifdef ADD_TO_REGEXEC
192 char *starttry; /* -Dr: where regtry was called. */
193 #define RExC_starttry (pRExC_state->starttry)
195 SV *runtime_code_qr; /* qr with the runtime code blocks */
197 const char *lastparse;
199 AV *paren_name_list; /* idx -> name */
200 U32 study_chunk_recursed_count;
203 #define RExC_lastparse (pRExC_state->lastparse)
204 #define RExC_lastnum (pRExC_state->lastnum)
205 #define RExC_paren_name_list (pRExC_state->paren_name_list)
206 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
207 #define RExC_mysv (pRExC_state->mysv1)
208 #define RExC_mysv1 (pRExC_state->mysv1)
209 #define RExC_mysv2 (pRExC_state->mysv2)
212 bool seen_unfolded_sharp_s;
218 #define RExC_flags (pRExC_state->flags)
219 #define RExC_pm_flags (pRExC_state->pm_flags)
220 #define RExC_precomp (pRExC_state->precomp)
221 #define RExC_precomp_adj (pRExC_state->precomp_adj)
222 #define RExC_adjusted_start (pRExC_state->adjusted_start)
223 #define RExC_precomp_end (pRExC_state->precomp_end)
224 #define RExC_rx_sv (pRExC_state->rx_sv)
225 #define RExC_rx (pRExC_state->rx)
226 #define RExC_rxi (pRExC_state->rxi)
227 #define RExC_start (pRExC_state->start)
228 #define RExC_end (pRExC_state->end)
229 #define RExC_parse (pRExC_state->parse)
230 #define RExC_whilem_seen (pRExC_state->whilem_seen)
232 /* Set during the sizing pass when there is a LATIN SMALL LETTER SHARP S in any
233 * EXACTF node, hence was parsed under /di rules. If later in the parse,
234 * something forces the pattern into using /ui rules, the sharp s should be
235 * folded into the sequence 'ss', which takes up more space than previously
236 * calculated. This means that the sizing pass needs to be restarted. (The
237 * node also becomes an EXACTFU_SS.) For all other characters, an EXACTF node
238 * that gets converted to /ui (and EXACTFU) occupies the same amount of space,
239 * so there is no need to resize [perl #125990]. */
240 #define RExC_seen_unfolded_sharp_s (pRExC_state->seen_unfolded_sharp_s)
242 #ifdef RE_TRACK_PATTERN_OFFSETS
243 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
246 #define RExC_emit (pRExC_state->emit)
247 #define RExC_emit_dummy (pRExC_state->emit_dummy)
248 #define RExC_emit_start (pRExC_state->emit_start)
249 #define RExC_emit_bound (pRExC_state->emit_bound)
250 #define RExC_sawback (pRExC_state->sawback)
251 #define RExC_seen (pRExC_state->seen)
252 #define RExC_size (pRExC_state->size)
253 #define RExC_maxlen (pRExC_state->maxlen)
254 #define RExC_npar (pRExC_state->npar)
255 #define RExC_nestroot (pRExC_state->nestroot)
256 #define RExC_extralen (pRExC_state->extralen)
257 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
258 #define RExC_utf8 (pRExC_state->utf8)
259 #define RExC_uni_semantics (pRExC_state->uni_semantics)
260 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
261 #define RExC_open_parens (pRExC_state->open_parens)
262 #define RExC_close_parens (pRExC_state->close_parens)
263 #define RExC_end_op (pRExC_state->end_op)
264 #define RExC_paren_names (pRExC_state->paren_names)
265 #define RExC_recurse (pRExC_state->recurse)
266 #define RExC_recurse_count (pRExC_state->recurse_count)
267 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
268 #define RExC_study_chunk_recursed_bytes \
269 (pRExC_state->study_chunk_recursed_bytes)
270 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
271 #define RExC_contains_locale (pRExC_state->contains_locale)
273 # define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
275 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
276 #define RExC_frame_head (pRExC_state->frame_head)
277 #define RExC_frame_last (pRExC_state->frame_last)
278 #define RExC_frame_count (pRExC_state->frame_count)
279 #define RExC_strict (pRExC_state->strict)
280 #define RExC_study_started (pRExC_state->study_started)
281 #define RExC_warn_text (pRExC_state->warn_text)
282 #define RExC_in_script_run (pRExC_state->in_script_run)
284 /* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
285 * a flag to disable back-off on the fixed/floating substrings - if it's
286 * a high complexity pattern we assume the benefit of avoiding a full match
287 * is worth the cost of checking for the substrings even if they rarely help.
289 #define RExC_naughty (pRExC_state->naughty)
290 #define TOO_NAUGHTY (10)
291 #define MARK_NAUGHTY(add) \
292 if (RExC_naughty < TOO_NAUGHTY) \
293 RExC_naughty += (add)
294 #define MARK_NAUGHTY_EXP(exp, add) \
295 if (RExC_naughty < TOO_NAUGHTY) \
296 RExC_naughty += RExC_naughty / (exp) + (add)
298 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
299 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
300 ((*s) == '{' && regcurly(s)))
303 * Flags to be passed up and down.
305 #define WORST 0 /* Worst case. */
306 #define HASWIDTH 0x01 /* Known to match non-null strings. */
308 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
309 * character. (There needs to be a case: in the switch statement in regexec.c
310 * for any node marked SIMPLE.) Note that this is not the same thing as
313 #define SPSTART 0x04 /* Starts with * or + */
314 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
315 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
316 #define RESTART_PASS1 0x20 /* Need to restart sizing pass */
317 #define NEED_UTF8 0x40 /* In conjunction with RESTART_PASS1, need to
318 calcuate sizes as UTF-8 */
320 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
322 /* whether trie related optimizations are enabled */
323 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
324 #define TRIE_STUDY_OPT
325 #define FULL_TRIE_STUDY
331 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
332 #define PBITVAL(paren) (1 << ((paren) & 7))
333 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
334 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
335 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
337 #define REQUIRE_UTF8(flagp) STMT_START { \
340 *flagp = RESTART_PASS1|NEED_UTF8; \
345 /* Change from /d into /u rules, and restart the parse if we've already seen
346 * something whose size would increase as a result, by setting *flagp and
347 * returning 'restart_retval'. RExC_uni_semantics is a flag that indicates
348 * we've changed to /u during the parse. */
349 #define REQUIRE_UNI_RULES(flagp, restart_retval) \
351 if (DEPENDS_SEMANTICS) { \
353 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
354 RExC_uni_semantics = 1; \
355 if (RExC_seen_unfolded_sharp_s) { \
356 *flagp |= RESTART_PASS1; \
357 return restart_retval; \
362 /* Executes a return statement with the value 'X', if 'flags' contains any of
363 * 'RESTART_PASS1', 'NEED_UTF8', or 'extra'. If so, *flagp is set to those
365 #define RETURN_X_ON_RESTART_OR_FLAGS(X, flags, flagp, extra) \
367 if ((flags) & (RESTART_PASS1|NEED_UTF8|(extra))) { \
368 *(flagp) = (flags) & (RESTART_PASS1|NEED_UTF8|(extra)); \
373 #define RETURN_NULL_ON_RESTART_OR_FLAGS(flags,flagp,extra) \
374 RETURN_X_ON_RESTART_OR_FLAGS(NULL,flags,flagp,extra)
376 #define RETURN_X_ON_RESTART(X, flags,flagp) \
377 RETURN_X_ON_RESTART_OR_FLAGS( X, flags, flagp, 0)
380 #define RETURN_NULL_ON_RESTART_FLAGP_OR_FLAGS(flagp,extra) \
381 if (*(flagp) & (RESTART_PASS1|(extra))) return NULL
383 #define MUST_RESTART(flags) ((flags) & (RESTART_PASS1))
385 #define RETURN_NULL_ON_RESTART(flags,flagp) \
386 RETURN_X_ON_RESTART(NULL, flags,flagp)
387 #define RETURN_NULL_ON_RESTART_FLAGP(flagp) \
388 RETURN_NULL_ON_RESTART_FLAGP_OR_FLAGS(flagp,0)
390 /* This converts the named class defined in regcomp.h to its equivalent class
391 * number defined in handy.h. */
392 #define namedclass_to_classnum(class) ((int) ((class) / 2))
393 #define classnum_to_namedclass(classnum) ((classnum) * 2)
395 #define _invlist_union_complement_2nd(a, b, output) \
396 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
397 #define _invlist_intersection_complement_2nd(a, b, output) \
398 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
400 /* About scan_data_t.
402 During optimisation we recurse through the regexp program performing
403 various inplace (keyhole style) optimisations. In addition study_chunk
404 and scan_commit populate this data structure with information about
405 what strings MUST appear in the pattern. We look for the longest
406 string that must appear at a fixed location, and we look for the
407 longest string that may appear at a floating location. So for instance
412 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
413 strings (because they follow a .* construct). study_chunk will identify
414 both FOO and BAR as being the longest fixed and floating strings respectively.
416 The strings can be composites, for instance
420 will result in a composite fixed substring 'foo'.
422 For each string some basic information is maintained:
425 This is the position the string must appear at, or not before.
426 It also implicitly (when combined with minlenp) tells us how many
427 characters must match before the string we are searching for.
428 Likewise when combined with minlenp and the length of the string it
429 tells us how many characters must appear after the string we have
433 Only used for floating strings. This is the rightmost point that
434 the string can appear at. If set to SSize_t_MAX it indicates that the
435 string can occur infinitely far to the right.
436 For fixed strings, it is equal to min_offset.
439 A pointer to the minimum number of characters of the pattern that the
440 string was found inside. This is important as in the case of positive
441 lookahead or positive lookbehind we can have multiple patterns
446 The minimum length of the pattern overall is 3, the minimum length
447 of the lookahead part is 3, but the minimum length of the part that
448 will actually match is 1. So 'FOO's minimum length is 3, but the
449 minimum length for the F is 1. This is important as the minimum length
450 is used to determine offsets in front of and behind the string being
451 looked for. Since strings can be composites this is the length of the
452 pattern at the time it was committed with a scan_commit. Note that
453 the length is calculated by study_chunk, so that the minimum lengths
454 are not known until the full pattern has been compiled, thus the
455 pointer to the value.
459 In the case of lookbehind the string being searched for can be
460 offset past the start point of the final matching string.
461 If this value was just blithely removed from the min_offset it would
462 invalidate some of the calculations for how many chars must match
463 before or after (as they are derived from min_offset and minlen and
464 the length of the string being searched for).
465 When the final pattern is compiled and the data is moved from the
466 scan_data_t structure into the regexp structure the information
467 about lookbehind is factored in, with the information that would
468 have been lost precalculated in the end_shift field for the
471 The fields pos_min and pos_delta are used to store the minimum offset
472 and the delta to the maximum offset at the current point in the pattern.
476 struct scan_data_substrs {
477 SV *str; /* longest substring found in pattern */
478 SSize_t min_offset; /* earliest point in string it can appear */
479 SSize_t max_offset; /* latest point in string it can appear */
480 SSize_t *minlenp; /* pointer to the minlen relevant to the string */
481 SSize_t lookbehind; /* is the pos of the string modified by LB */
482 I32 flags; /* per substring SF_* and SCF_* flags */
485 typedef struct scan_data_t {
486 /*I32 len_min; unused */
487 /*I32 len_delta; unused */
491 SSize_t last_end; /* min value, <0 unless valid. */
492 SSize_t last_start_min;
493 SSize_t last_start_max;
494 U8 cur_is_floating; /* whether the last_* values should be set as
495 * the next fixed (0) or floating (1)
498 /* [0] is longest fixed substring so far, [1] is longest float so far */
499 struct scan_data_substrs substrs[2];
501 I32 flags; /* common SF_* and SCF_* flags */
503 SSize_t *last_closep;
504 regnode_ssc *start_class;
508 * Forward declarations for pregcomp()'s friends.
511 static const scan_data_t zero_scan_data = {
512 0, 0, NULL, 0, 0, 0, 0,
514 { NULL, 0, 0, 0, 0, 0 },
515 { NULL, 0, 0, 0, 0, 0 },
522 #define SF_BEFORE_SEOL 0x0001
523 #define SF_BEFORE_MEOL 0x0002
524 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
526 #define SF_IS_INF 0x0040
527 #define SF_HAS_PAR 0x0080
528 #define SF_IN_PAR 0x0100
529 #define SF_HAS_EVAL 0x0200
532 /* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
533 * longest substring in the pattern. When it is not set the optimiser keeps
534 * track of position, but does not keep track of the actual strings seen,
536 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
539 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
540 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
541 * turned off because of the alternation (BRANCH). */
542 #define SCF_DO_SUBSTR 0x0400
544 #define SCF_DO_STCLASS_AND 0x0800
545 #define SCF_DO_STCLASS_OR 0x1000
546 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
547 #define SCF_WHILEM_VISITED_POS 0x2000
549 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
550 #define SCF_SEEN_ACCEPT 0x8000
551 #define SCF_TRIE_DOING_RESTUDY 0x10000
552 #define SCF_IN_DEFINE 0x20000
557 #define UTF cBOOL(RExC_utf8)
559 /* The enums for all these are ordered so things work out correctly */
560 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
561 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
562 == REGEX_DEPENDS_CHARSET)
563 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
564 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
565 >= REGEX_UNICODE_CHARSET)
566 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
567 == REGEX_ASCII_RESTRICTED_CHARSET)
568 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
569 >= REGEX_ASCII_RESTRICTED_CHARSET)
570 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
571 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
573 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
575 /* For programs that want to be strictly Unicode compatible by dying if any
576 * attempt is made to match a non-Unicode code point against a Unicode
578 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
580 #define OOB_NAMEDCLASS -1
582 /* There is no code point that is out-of-bounds, so this is problematic. But
583 * its only current use is to initialize a variable that is always set before
585 #define OOB_UNICODE 0xDEADBEEF
587 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
590 /* length of regex to show in messages that don't mark a position within */
591 #define RegexLengthToShowInErrorMessages 127
594 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
595 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
596 * op/pragma/warn/regcomp.
598 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
599 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
601 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
602 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
604 /* The code in this file in places uses one level of recursion with parsing
605 * rebased to an alternate string constructed by us in memory. This can take
606 * the form of something that is completely different from the input, or
607 * something that uses the input as part of the alternate. In the first case,
608 * there should be no possibility of an error, as we are in complete control of
609 * the alternate string. But in the second case we don't control the input
610 * portion, so there may be errors in that. Here's an example:
612 * is handled specially because \x{df} folds to a sequence of more than one
613 * character, 'ss'. What is done is to create and parse an alternate string,
614 * which looks like this:
615 * /(?:\x{DF}|[abc\x{DF}def])/ui
616 * where it uses the input unchanged in the middle of something it constructs,
617 * which is a branch for the DF outside the character class, and clustering
618 * parens around the whole thing. (It knows enough to skip the DF inside the
619 * class while in this substitute parse.) 'abc' and 'def' may have errors that
620 * need to be reported. The general situation looks like this:
623 * Input: ----------------------------------------------------
624 * Constructed: ---------------------------------------------------
627 * The input string sI..eI is the input pattern. The string sC..EC is the
628 * constructed substitute parse string. The portions sC..tC and eC..EC are
629 * constructed by us. The portion tC..eC is an exact duplicate of the input
630 * pattern tI..eI. In the diagram, these are vertically aligned. Suppose that
631 * while parsing, we find an error at xC. We want to display a message showing
632 * the real input string. Thus we need to find the point xI in it which
633 * corresponds to xC. xC >= tC, since the portion of the string sC..tC has
634 * been constructed by us, and so shouldn't have errors. We get:
636 * xI = sI + (tI - sI) + (xC - tC)
638 * and, the offset into sI is:
640 * (xI - sI) = (tI - sI) + (xC - tC)
642 * When the substitute is constructed, we save (tI -sI) as RExC_precomp_adj,
643 * and we save tC as RExC_adjusted_start.
645 * During normal processing of the input pattern, everything points to that,
646 * with RExC_precomp_adj set to 0, and RExC_adjusted_start set to sI.
649 #define tI_sI RExC_precomp_adj
650 #define tC RExC_adjusted_start
651 #define sC RExC_precomp
652 #define xI_offset(xC) ((IV) (tI_sI + (xC - tC)))
653 #define xI(xC) (sC + xI_offset(xC))
654 #define eC RExC_precomp_end
656 #define REPORT_LOCATION_ARGS(xC) \
658 (xI(xC) > eC) /* Don't run off end */ \
659 ? eC - sC /* Length before the <--HERE */ \
660 : ((xI_offset(xC) >= 0) \
662 : (Perl_croak(aTHX_ "panic: %s: %d: negative offset: %" \
663 IVdf " trying to output message for " \
665 __FILE__, __LINE__, xI_offset(xC), \
666 ((int) (eC - sC)), sC), 0)), \
667 sC), /* The input pattern printed up to the <--HERE */ \
669 (xI(xC) > eC) ? 0 : eC - xI(xC), /* Length after <--HERE */ \
670 (xI(xC) > eC) ? eC : xI(xC)) /* pattern after <--HERE */
672 /* Used to point after bad bytes for an error message, but avoid skipping
673 * past a nul byte. */
674 #define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
677 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
678 * arg. Show regex, up to a maximum length. If it's too long, chop and add
681 #define _FAIL(code) STMT_START { \
682 const char *ellipses = ""; \
683 IV len = RExC_precomp_end - RExC_precomp; \
686 SAVEFREESV(RExC_rx_sv); \
687 if (len > RegexLengthToShowInErrorMessages) { \
688 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
689 len = RegexLengthToShowInErrorMessages - 10; \
695 #define FAIL(msg) _FAIL( \
696 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
697 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
699 #define FAIL2(msg,arg) _FAIL( \
700 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
701 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
704 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
706 #define Simple_vFAIL(m) STMT_START { \
707 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
708 m, REPORT_LOCATION_ARGS(RExC_parse)); \
712 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
714 #define vFAIL(m) STMT_START { \
716 SAVEFREESV(RExC_rx_sv); \
721 * Like Simple_vFAIL(), but accepts two arguments.
723 #define Simple_vFAIL2(m,a1) STMT_START { \
724 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
725 REPORT_LOCATION_ARGS(RExC_parse)); \
729 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
731 #define vFAIL2(m,a1) STMT_START { \
733 SAVEFREESV(RExC_rx_sv); \
734 Simple_vFAIL2(m, a1); \
739 * Like Simple_vFAIL(), but accepts three arguments.
741 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
742 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
743 REPORT_LOCATION_ARGS(RExC_parse)); \
747 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
749 #define vFAIL3(m,a1,a2) STMT_START { \
751 SAVEFREESV(RExC_rx_sv); \
752 Simple_vFAIL3(m, a1, a2); \
756 * Like Simple_vFAIL(), but accepts four arguments.
758 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
759 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
760 REPORT_LOCATION_ARGS(RExC_parse)); \
763 #define vFAIL4(m,a1,a2,a3) STMT_START { \
765 SAVEFREESV(RExC_rx_sv); \
766 Simple_vFAIL4(m, a1, a2, a3); \
769 /* A specialized version of vFAIL2 that works with UTF8f */
770 #define vFAIL2utf8f(m, a1) STMT_START { \
772 SAVEFREESV(RExC_rx_sv); \
773 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
774 REPORT_LOCATION_ARGS(RExC_parse)); \
777 #define vFAIL3utf8f(m, a1, a2) STMT_START { \
779 SAVEFREESV(RExC_rx_sv); \
780 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
781 REPORT_LOCATION_ARGS(RExC_parse)); \
784 /* These have asserts in them because of [perl #122671] Many warnings in
785 * regcomp.c can occur twice. If they get output in pass1 and later in that
786 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
787 * would get output again. So they should be output in pass2, and these
788 * asserts make sure new warnings follow that paradigm. */
790 /* m is not necessarily a "literal string", in this macro */
791 #define reg_warn_non_literal_string(loc, m) STMT_START { \
792 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
793 "%s" REPORT_LOCATION, \
794 m, REPORT_LOCATION_ARGS(loc)); \
797 #define ckWARNreg(loc,m) STMT_START { \
798 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
800 REPORT_LOCATION_ARGS(loc)); \
803 #define vWARN(loc, m) STMT_START { \
804 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
806 REPORT_LOCATION_ARGS(loc)); \
809 #define vWARN_dep(loc, m) STMT_START { \
810 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
812 REPORT_LOCATION_ARGS(loc)); \
815 #define ckWARNdep(loc,m) STMT_START { \
816 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
818 REPORT_LOCATION_ARGS(loc)); \
821 #define ckWARNregdep(loc,m) STMT_START { \
822 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
825 REPORT_LOCATION_ARGS(loc)); \
828 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
829 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
831 a1, REPORT_LOCATION_ARGS(loc)); \
834 #define ckWARN2reg(loc, m, a1) STMT_START { \
835 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
837 a1, REPORT_LOCATION_ARGS(loc)); \
840 #define vWARN3(loc, m, a1, a2) STMT_START { \
841 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
843 a1, a2, REPORT_LOCATION_ARGS(loc)); \
846 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
847 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
850 REPORT_LOCATION_ARGS(loc)); \
853 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
854 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
857 REPORT_LOCATION_ARGS(loc)); \
860 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
861 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
864 REPORT_LOCATION_ARGS(loc)); \
867 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
868 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
871 REPORT_LOCATION_ARGS(loc)); \
874 /* Macros for recording node offsets. 20001227 mjd@plover.com
875 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
876 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
877 * Element 0 holds the number n.
878 * Position is 1 indexed.
880 #ifndef RE_TRACK_PATTERN_OFFSETS
881 #define Set_Node_Offset_To_R(node,byte)
882 #define Set_Node_Offset(node,byte)
883 #define Set_Cur_Node_Offset
884 #define Set_Node_Length_To_R(node,len)
885 #define Set_Node_Length(node,len)
886 #define Set_Node_Cur_Length(node,start)
887 #define Node_Offset(n)
888 #define Node_Length(n)
889 #define Set_Node_Offset_Length(node,offset,len)
890 #define ProgLen(ri) ri->u.proglen
891 #define SetProgLen(ri,x) ri->u.proglen = x
893 #define ProgLen(ri) ri->u.offsets[0]
894 #define SetProgLen(ri,x) ri->u.offsets[0] = x
895 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
897 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
898 __LINE__, (int)(node), (int)(byte))); \
900 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
903 RExC_offsets[2*(node)-1] = (byte); \
908 #define Set_Node_Offset(node,byte) \
909 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
910 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
912 #define Set_Node_Length_To_R(node,len) STMT_START { \
914 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
915 __LINE__, (int)(node), (int)(len))); \
917 Perl_croak(aTHX_ "value of node is %d in Length macro", \
920 RExC_offsets[2*(node)] = (len); \
925 #define Set_Node_Length(node,len) \
926 Set_Node_Length_To_R((node)-RExC_emit_start, len)
927 #define Set_Node_Cur_Length(node, start) \
928 Set_Node_Length(node, RExC_parse - start)
930 /* Get offsets and lengths */
931 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
932 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
934 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
935 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
936 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
940 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
941 #define EXPERIMENTAL_INPLACESCAN
942 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
946 Perl_re_printf(pTHX_ const char *fmt, ...)
950 PerlIO *f= Perl_debug_log;
951 PERL_ARGS_ASSERT_RE_PRINTF;
953 result = PerlIO_vprintf(f, fmt, ap);
959 Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
963 PerlIO *f= Perl_debug_log;
964 PERL_ARGS_ASSERT_RE_INDENTF;
966 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
967 result = PerlIO_vprintf(f, fmt, ap);
971 #endif /* DEBUGGING */
973 #define DEBUG_RExC_seen() \
974 DEBUG_OPTIMISE_MORE_r({ \
975 Perl_re_printf( aTHX_ "RExC_seen: "); \
977 if (RExC_seen & REG_ZERO_LEN_SEEN) \
978 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
980 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
981 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
983 if (RExC_seen & REG_GPOS_SEEN) \
984 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
986 if (RExC_seen & REG_RECURSE_SEEN) \
987 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
989 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
990 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
992 if (RExC_seen & REG_VERBARG_SEEN) \
993 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
995 if (RExC_seen & REG_CUTGROUP_SEEN) \
996 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
998 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
999 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
1001 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
1002 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
1004 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
1005 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
1007 Perl_re_printf( aTHX_ "\n"); \
1010 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
1011 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
1016 S_debug_show_study_flags(pTHX_ U32 flags, const char *open_str,
1017 const char *close_str)
1022 Perl_re_printf( aTHX_ "%s", open_str);
1023 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_SEOL);
1024 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_MEOL);
1025 DEBUG_SHOW_STUDY_FLAG(flags, SF_IS_INF);
1026 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_PAR);
1027 DEBUG_SHOW_STUDY_FLAG(flags, SF_IN_PAR);
1028 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_EVAL);
1029 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_SUBSTR);
1030 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_AND);
1031 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_OR);
1032 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS);
1033 DEBUG_SHOW_STUDY_FLAG(flags, SCF_WHILEM_VISITED_POS);
1034 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_RESTUDY);
1035 DEBUG_SHOW_STUDY_FLAG(flags, SCF_SEEN_ACCEPT);
1036 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_DOING_RESTUDY);
1037 DEBUG_SHOW_STUDY_FLAG(flags, SCF_IN_DEFINE);
1038 Perl_re_printf( aTHX_ "%s", close_str);
1043 S_debug_studydata(pTHX_ const char *where, scan_data_t *data,
1044 U32 depth, int is_inf)
1046 GET_RE_DEBUG_FLAGS_DECL;
1048 DEBUG_OPTIMISE_MORE_r({
1051 Perl_re_indentf(aTHX_ "%s: Pos:%" IVdf "/%" IVdf " Flags: 0x%" UVXf,
1055 (IV)data->pos_delta,
1059 S_debug_show_study_flags(aTHX_ data->flags," [","]");
1061 Perl_re_printf( aTHX_
1062 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s",
1064 (IV)(data->last_closep ? *((data)->last_closep) : -1),
1065 is_inf ? "INF " : ""
1068 if (data->last_found) {
1070 Perl_re_printf(aTHX_
1071 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf,
1072 SvPVX_const(data->last_found),
1074 (IV)data->last_start_min,
1075 (IV)data->last_start_max
1078 for (i = 0; i < 2; i++) {
1079 Perl_re_printf(aTHX_
1080 " %s%s: '%s' @ %" IVdf "/%" IVdf,
1081 data->cur_is_floating == i ? "*" : "",
1082 i ? "Float" : "Fixed",
1083 SvPVX_const(data->substrs[i].str),
1084 (IV)data->substrs[i].min_offset,
1085 (IV)data->substrs[i].max_offset
1087 S_debug_show_study_flags(aTHX_ data->substrs[i].flags," [","]");
1091 Perl_re_printf( aTHX_ "\n");
1097 S_debug_peep(pTHX_ const char *str, const RExC_state_t *pRExC_state,
1098 regnode *scan, U32 depth, U32 flags)
1100 GET_RE_DEBUG_FLAGS_DECL;
1107 Next = regnext(scan);
1108 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
1109 Perl_re_indentf( aTHX_ "%s>%3d: %s (%d)",
1112 REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),
1113 Next ? (REG_NODE_NUM(Next)) : 0 );
1114 S_debug_show_study_flags(aTHX_ flags," [ ","]");
1115 Perl_re_printf( aTHX_ "\n");
1120 # define DEBUG_STUDYDATA(where, data, depth, is_inf) \
1121 S_debug_studydata(aTHX_ where, data, depth, is_inf)
1123 # define DEBUG_PEEP(str, scan, depth, flags) \
1124 S_debug_peep(aTHX_ str, pRExC_state, scan, depth, flags)
1127 # define DEBUG_STUDYDATA(where, data, depth, is_inf) NOOP
1128 # define DEBUG_PEEP(str, scan, depth, flags) NOOP
1132 /* =========================================================
1133 * BEGIN edit_distance stuff.
1135 * This calculates how many single character changes of any type are needed to
1136 * transform a string into another one. It is taken from version 3.1 of
1138 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1141 /* Our unsorted dictionary linked list. */
1142 /* Note we use UVs, not chars. */
1147 struct dictionary* next;
1149 typedef struct dictionary item;
1152 PERL_STATIC_INLINE item*
1153 push(UV key,item* curr)
1156 Newx(head, 1, item);
1164 PERL_STATIC_INLINE item*
1165 find(item* head, UV key)
1167 item* iterator = head;
1169 if (iterator->key == key){
1172 iterator = iterator->next;
1178 PERL_STATIC_INLINE item*
1179 uniquePush(item* head,UV key)
1181 item* iterator = head;
1184 if (iterator->key == key) {
1187 iterator = iterator->next;
1190 return push(key,head);
1193 PERL_STATIC_INLINE void
1194 dict_free(item* head)
1196 item* iterator = head;
1199 item* temp = iterator;
1200 iterator = iterator->next;
1207 /* End of Dictionary Stuff */
1209 /* All calculations/work are done here */
1211 S_edit_distance(const UV* src,
1213 const STRLEN x, /* length of src[] */
1214 const STRLEN y, /* length of tgt[] */
1215 const SSize_t maxDistance
1219 UV swapCount,swapScore,targetCharCount,i,j;
1221 UV score_ceil = x + y;
1223 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1225 /* intialize matrix start values */
1226 Newx(scores, ( (x + 2) * (y + 2)), UV);
1227 scores[0] = score_ceil;
1228 scores[1 * (y + 2) + 0] = score_ceil;
1229 scores[0 * (y + 2) + 1] = score_ceil;
1230 scores[1 * (y + 2) + 1] = 0;
1231 head = uniquePush(uniquePush(head,src[0]),tgt[0]);
1236 for (i=1;i<=x;i++) {
1238 head = uniquePush(head,src[i]);
1239 scores[(i+1) * (y + 2) + 1] = i;
1240 scores[(i+1) * (y + 2) + 0] = score_ceil;
1243 for (j=1;j<=y;j++) {
1246 head = uniquePush(head,tgt[j]);
1247 scores[1 * (y + 2) + (j + 1)] = j;
1248 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1251 targetCharCount = find(head,tgt[j-1])->value;
1252 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1254 if (src[i-1] != tgt[j-1]){
1255 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));
1259 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1263 find(head,src[i-1])->value = i;
1267 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1270 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1274 /* END of edit_distance() stuff
1275 * ========================================================= */
1277 /* is c a control character for which we have a mnemonic? */
1278 #define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
1281 S_cntrl_to_mnemonic(const U8 c)
1283 /* Returns the mnemonic string that represents character 'c', if one
1284 * exists; NULL otherwise. The only ones that exist for the purposes of
1285 * this routine are a few control characters */
1288 case '\a': return "\\a";
1289 case '\b': return "\\b";
1290 case ESC_NATIVE: return "\\e";
1291 case '\f': return "\\f";
1292 case '\n': return "\\n";
1293 case '\r': return "\\r";
1294 case '\t': return "\\t";
1300 /* Mark that we cannot extend a found fixed substring at this point.
1301 Update the longest found anchored substring or the longest found
1302 floating substrings if needed. */
1305 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1306 SSize_t *minlenp, int is_inf)
1308 const STRLEN l = CHR_SVLEN(data->last_found);
1309 SV * const longest_sv = data->substrs[data->cur_is_floating].str;
1310 const STRLEN old_l = CHR_SVLEN(longest_sv);
1311 GET_RE_DEBUG_FLAGS_DECL;
1313 PERL_ARGS_ASSERT_SCAN_COMMIT;
1315 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
1316 const U8 i = data->cur_is_floating;
1317 SvSetMagicSV(longest_sv, data->last_found);
1318 data->substrs[i].min_offset = l ? data->last_start_min : data->pos_min;
1321 data->substrs[0].max_offset = data->substrs[0].min_offset;
1323 data->substrs[1].max_offset = (l
1324 ? data->last_start_max
1325 : (data->pos_delta > SSize_t_MAX - data->pos_min
1327 : data->pos_min + data->pos_delta));
1329 || (STRLEN)data->substrs[1].max_offset > (STRLEN)SSize_t_MAX)
1330 data->substrs[1].max_offset = SSize_t_MAX;
1333 if (data->flags & SF_BEFORE_EOL)
1334 data->substrs[i].flags |= (data->flags & SF_BEFORE_EOL);
1336 data->substrs[i].flags &= ~SF_BEFORE_EOL;
1337 data->substrs[i].minlenp = minlenp;
1338 data->substrs[i].lookbehind = 0;
1341 SvCUR_set(data->last_found, 0);
1343 SV * const sv = data->last_found;
1344 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1345 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1350 data->last_end = -1;
1351 data->flags &= ~SF_BEFORE_EOL;
1352 DEBUG_STUDYDATA("commit", data, 0, is_inf);
1355 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1356 * list that describes which code points it matches */
1359 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1361 /* Set the SSC 'ssc' to match an empty string or any code point */
1363 PERL_ARGS_ASSERT_SSC_ANYTHING;
1365 assert(is_ANYOF_SYNTHETIC(ssc));
1367 /* mortalize so won't leak */
1368 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
1369 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1373 S_ssc_is_anything(const regnode_ssc *ssc)
1375 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1376 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1377 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1378 * in any way, so there's no point in using it */
1383 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1385 assert(is_ANYOF_SYNTHETIC(ssc));
1387 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1391 /* See if the list consists solely of the range 0 - Infinity */
1392 invlist_iterinit(ssc->invlist);
1393 ret = invlist_iternext(ssc->invlist, &start, &end)
1397 invlist_iterfinish(ssc->invlist);
1403 /* If e.g., both \w and \W are set, matches everything */
1404 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1406 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1407 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1417 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1419 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1420 * string, any code point, or any posix class under locale */
1422 PERL_ARGS_ASSERT_SSC_INIT;
1424 Zero(ssc, 1, regnode_ssc);
1425 set_ANYOF_SYNTHETIC(ssc);
1426 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1429 /* If any portion of the regex is to operate under locale rules that aren't
1430 * fully known at compile time, initialization includes it. The reason
1431 * this isn't done for all regexes is that the optimizer was written under
1432 * the assumption that locale was all-or-nothing. Given the complexity and
1433 * lack of documentation in the optimizer, and that there are inadequate
1434 * test cases for locale, many parts of it may not work properly, it is
1435 * safest to avoid locale unless necessary. */
1436 if (RExC_contains_locale) {
1437 ANYOF_POSIXL_SETALL(ssc);
1440 ANYOF_POSIXL_ZERO(ssc);
1445 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1446 const regnode_ssc *ssc)
1448 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1449 * to the list of code points matched, and locale posix classes; hence does
1450 * not check its flags) */
1455 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1457 assert(is_ANYOF_SYNTHETIC(ssc));
1459 invlist_iterinit(ssc->invlist);
1460 ret = invlist_iternext(ssc->invlist, &start, &end)
1464 invlist_iterfinish(ssc->invlist);
1470 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1478 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1479 const regnode_charclass* const node)
1481 /* Returns a mortal inversion list defining which code points are matched
1482 * by 'node', which is of type ANYOF. Handles complementing the result if
1483 * appropriate. If some code points aren't knowable at this time, the
1484 * returned list must, and will, contain every code point that is a
1488 SV* only_utf8_locale_invlist = NULL;
1490 const U32 n = ARG(node);
1491 bool new_node_has_latin1 = FALSE;
1493 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1495 /* Look at the data structure created by S_set_ANYOF_arg() */
1496 if (n != ANYOF_ONLY_HAS_BITMAP) {
1497 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1498 AV * const av = MUTABLE_AV(SvRV(rv));
1499 SV **const ary = AvARRAY(av);
1500 assert(RExC_rxi->data->what[n] == 's');
1502 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1503 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1505 else if (ary[0] && ary[0] != &PL_sv_undef) {
1507 /* Here, no compile-time swash, and there are things that won't be
1508 * known until runtime -- we have to assume it could be anything */
1509 invlist = sv_2mortal(_new_invlist(1));
1510 return _add_range_to_invlist(invlist, 0, UV_MAX);
1512 else if (ary[3] && ary[3] != &PL_sv_undef) {
1514 /* Here no compile-time swash, and no run-time only data. Use the
1515 * node's inversion list */
1516 invlist = sv_2mortal(invlist_clone(ary[3]));
1519 /* Get the code points valid only under UTF-8 locales */
1520 if ((ANYOF_FLAGS(node) & ANYOFL_FOLD)
1521 && ary[2] && ary[2] != &PL_sv_undef)
1523 only_utf8_locale_invlist = ary[2];
1528 invlist = sv_2mortal(_new_invlist(0));
1531 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1532 * code points, and an inversion list for the others, but if there are code
1533 * points that should match only conditionally on the target string being
1534 * UTF-8, those are placed in the inversion list, and not the bitmap.
1535 * Since there are circumstances under which they could match, they are
1536 * included in the SSC. But if the ANYOF node is to be inverted, we have
1537 * to exclude them here, so that when we invert below, the end result
1538 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1539 * have to do this here before we add the unconditionally matched code
1541 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1542 _invlist_intersection_complement_2nd(invlist,
1547 /* Add in the points from the bit map */
1548 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1549 if (ANYOF_BITMAP_TEST(node, i)) {
1550 unsigned int start = i++;
1552 for (; i < NUM_ANYOF_CODE_POINTS && ANYOF_BITMAP_TEST(node, i); ++i) {
1555 invlist = _add_range_to_invlist(invlist, start, i-1);
1556 new_node_has_latin1 = TRUE;
1560 /* If this can match all upper Latin1 code points, have to add them
1561 * as well. But don't add them if inverting, as when that gets done below,
1562 * it would exclude all these characters, including the ones it shouldn't
1563 * that were added just above */
1564 if (! (ANYOF_FLAGS(node) & ANYOF_INVERT) && OP(node) == ANYOFD
1565 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1567 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1570 /* Similarly for these */
1571 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1572 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1575 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1576 _invlist_invert(invlist);
1578 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOFL_FOLD) {
1580 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1581 * locale. We can skip this if there are no 0-255 at all. */
1582 _invlist_union(invlist, PL_Latin1, &invlist);
1585 /* Similarly add the UTF-8 locale possible matches. These have to be
1586 * deferred until after the non-UTF-8 locale ones are taken care of just
1587 * above, or it leads to wrong results under ANYOF_INVERT */
1588 if (only_utf8_locale_invlist) {
1589 _invlist_union_maybe_complement_2nd(invlist,
1590 only_utf8_locale_invlist,
1591 ANYOF_FLAGS(node) & ANYOF_INVERT,
1598 /* These two functions currently do the exact same thing */
1599 #define ssc_init_zero ssc_init
1601 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1602 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1604 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1605 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1606 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1609 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1610 const regnode_charclass *and_with)
1612 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1613 * another SSC or a regular ANYOF class. Can create false positives. */
1618 PERL_ARGS_ASSERT_SSC_AND;
1620 assert(is_ANYOF_SYNTHETIC(ssc));
1622 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1623 * the code point inversion list and just the relevant flags */
1624 if (is_ANYOF_SYNTHETIC(and_with)) {
1625 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1626 anded_flags = ANYOF_FLAGS(and_with);
1628 /* XXX This is a kludge around what appears to be deficiencies in the
1629 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1630 * there are paths through the optimizer where it doesn't get weeded
1631 * out when it should. And if we don't make some extra provision for
1632 * it like the code just below, it doesn't get added when it should.
1633 * This solution is to add it only when AND'ing, which is here, and
1634 * only when what is being AND'ed is the pristine, original node
1635 * matching anything. Thus it is like adding it to ssc_anything() but
1636 * only when the result is to be AND'ed. Probably the same solution
1637 * could be adopted for the same problem we have with /l matching,
1638 * which is solved differently in S_ssc_init(), and that would lead to
1639 * fewer false positives than that solution has. But if this solution
1640 * creates bugs, the consequences are only that a warning isn't raised
1641 * that should be; while the consequences for having /l bugs is
1642 * incorrect matches */
1643 if (ssc_is_anything((regnode_ssc *)and_with)) {
1644 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1648 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1649 if (OP(and_with) == ANYOFD) {
1650 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1653 anded_flags = ANYOF_FLAGS(and_with)
1654 &( ANYOF_COMMON_FLAGS
1655 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1656 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1657 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(and_with))) {
1659 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1664 ANYOF_FLAGS(ssc) &= anded_flags;
1666 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1667 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1668 * 'and_with' may be inverted. When not inverted, we have the situation of
1670 * (C1 | P1) & (C2 | P2)
1671 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1672 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1673 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1674 * <= ((C1 & C2) | P1 | P2)
1675 * Alternatively, the last few steps could be:
1676 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1677 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1678 * <= (C1 | C2 | (P1 & P2))
1679 * We favor the second approach if either P1 or P2 is non-empty. This is
1680 * because these components are a barrier to doing optimizations, as what
1681 * they match cannot be known until the moment of matching as they are
1682 * dependent on the current locale, 'AND"ing them likely will reduce or
1684 * But we can do better if we know that C1,P1 are in their initial state (a
1685 * frequent occurrence), each matching everything:
1686 * (<everything>) & (C2 | P2) = C2 | P2
1687 * Similarly, if C2,P2 are in their initial state (again a frequent
1688 * occurrence), the result is a no-op
1689 * (C1 | P1) & (<everything>) = C1 | P1
1692 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1693 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1694 * <= (C1 & ~C2) | (P1 & ~P2)
1697 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1698 && ! is_ANYOF_SYNTHETIC(and_with))
1702 ssc_intersection(ssc,
1704 FALSE /* Has already been inverted */
1707 /* If either P1 or P2 is empty, the intersection will be also; can skip
1709 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
1710 ANYOF_POSIXL_ZERO(ssc);
1712 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1714 /* Note that the Posix class component P from 'and_with' actually
1716 * P = Pa | Pb | ... | Pn
1717 * where each component is one posix class, such as in [\w\s].
1719 * ~P = ~(Pa | Pb | ... | Pn)
1720 * = ~Pa & ~Pb & ... & ~Pn
1721 * <= ~Pa | ~Pb | ... | ~Pn
1722 * The last is something we can easily calculate, but unfortunately
1723 * is likely to have many false positives. We could do better
1724 * in some (but certainly not all) instances if two classes in
1725 * P have known relationships. For example
1726 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1728 * :lower: & :print: = :lower:
1729 * And similarly for classes that must be disjoint. For example,
1730 * since \s and \w can have no elements in common based on rules in
1731 * the POSIX standard,
1732 * \w & ^\S = nothing
1733 * Unfortunately, some vendor locales do not meet the Posix
1734 * standard, in particular almost everything by Microsoft.
1735 * The loop below just changes e.g., \w into \W and vice versa */
1737 regnode_charclass_posixl temp;
1738 int add = 1; /* To calculate the index of the complement */
1740 Zero(&temp, 1, regnode_charclass_posixl);
1741 ANYOF_POSIXL_ZERO(&temp);
1742 for (i = 0; i < ANYOF_MAX; i++) {
1744 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1745 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1747 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1748 ANYOF_POSIXL_SET(&temp, i + add);
1750 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1752 ANYOF_POSIXL_AND(&temp, ssc);
1754 } /* else ssc already has no posixes */
1755 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1756 in its initial state */
1757 else if (! is_ANYOF_SYNTHETIC(and_with)
1758 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1760 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1761 * copy it over 'ssc' */
1762 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1763 if (is_ANYOF_SYNTHETIC(and_with)) {
1764 StructCopy(and_with, ssc, regnode_ssc);
1767 ssc->invlist = anded_cp_list;
1768 ANYOF_POSIXL_ZERO(ssc);
1769 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1770 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1774 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1775 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
1777 /* One or the other of P1, P2 is non-empty. */
1778 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1779 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1781 ssc_union(ssc, anded_cp_list, FALSE);
1783 else { /* P1 = P2 = empty */
1784 ssc_intersection(ssc, anded_cp_list, FALSE);
1790 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1791 const regnode_charclass *or_with)
1793 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1794 * another SSC or a regular ANYOF class. Can create false positives if
1795 * 'or_with' is to be inverted. */
1800 PERL_ARGS_ASSERT_SSC_OR;
1802 assert(is_ANYOF_SYNTHETIC(ssc));
1804 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1805 * the code point inversion list and just the relevant flags */
1806 if (is_ANYOF_SYNTHETIC(or_with)) {
1807 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1808 ored_flags = ANYOF_FLAGS(or_with);
1811 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1812 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1813 if (OP(or_with) != ANYOFD) {
1815 |= ANYOF_FLAGS(or_with)
1816 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1817 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1818 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(or_with))) {
1820 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1825 ANYOF_FLAGS(ssc) |= ored_flags;
1827 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1828 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1829 * 'or_with' may be inverted. When not inverted, we have the simple
1830 * situation of computing:
1831 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1832 * If P1|P2 yields a situation with both a class and its complement are
1833 * set, like having both \w and \W, this matches all code points, and we
1834 * can delete these from the P component of the ssc going forward. XXX We
1835 * might be able to delete all the P components, but I (khw) am not certain
1836 * about this, and it is better to be safe.
1839 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1840 * <= (C1 | P1) | ~C2
1841 * <= (C1 | ~C2) | P1
1842 * (which results in actually simpler code than the non-inverted case)
1845 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1846 && ! is_ANYOF_SYNTHETIC(or_with))
1848 /* We ignore P2, leaving P1 going forward */
1849 } /* else Not inverted */
1850 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
1851 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1852 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1854 for (i = 0; i < ANYOF_MAX; i += 2) {
1855 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1857 ssc_match_all_cp(ssc);
1858 ANYOF_POSIXL_CLEAR(ssc, i);
1859 ANYOF_POSIXL_CLEAR(ssc, i+1);
1867 FALSE /* Already has been inverted */
1871 PERL_STATIC_INLINE void
1872 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1874 PERL_ARGS_ASSERT_SSC_UNION;
1876 assert(is_ANYOF_SYNTHETIC(ssc));
1878 _invlist_union_maybe_complement_2nd(ssc->invlist,
1884 PERL_STATIC_INLINE void
1885 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1887 const bool invert2nd)
1889 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1891 assert(is_ANYOF_SYNTHETIC(ssc));
1893 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1899 PERL_STATIC_INLINE void
1900 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1902 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1904 assert(is_ANYOF_SYNTHETIC(ssc));
1906 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1909 PERL_STATIC_INLINE void
1910 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1912 /* AND just the single code point 'cp' into the SSC 'ssc' */
1914 SV* cp_list = _new_invlist(2);
1916 PERL_ARGS_ASSERT_SSC_CP_AND;
1918 assert(is_ANYOF_SYNTHETIC(ssc));
1920 cp_list = add_cp_to_invlist(cp_list, cp);
1921 ssc_intersection(ssc, cp_list,
1922 FALSE /* Not inverted */
1924 SvREFCNT_dec_NN(cp_list);
1927 PERL_STATIC_INLINE void
1928 S_ssc_clear_locale(regnode_ssc *ssc)
1930 /* Set the SSC 'ssc' to not match any locale things */
1931 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1933 assert(is_ANYOF_SYNTHETIC(ssc));
1935 ANYOF_POSIXL_ZERO(ssc);
1936 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1939 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1942 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1944 /* The synthetic start class is used to hopefully quickly winnow down
1945 * places where a pattern could start a match in the target string. If it
1946 * doesn't really narrow things down that much, there isn't much point to
1947 * having the overhead of using it. This function uses some very crude
1948 * heuristics to decide if to use the ssc or not.
1950 * It returns TRUE if 'ssc' rules out more than half what it considers to
1951 * be the "likely" possible matches, but of course it doesn't know what the
1952 * actual things being matched are going to be; these are only guesses
1954 * For /l matches, it assumes that the only likely matches are going to be
1955 * in the 0-255 range, uniformly distributed, so half of that is 127
1956 * For /a and /d matches, it assumes that the likely matches will be just
1957 * the ASCII range, so half of that is 63
1958 * For /u and there isn't anything matching above the Latin1 range, it
1959 * assumes that that is the only range likely to be matched, and uses
1960 * half that as the cut-off: 127. If anything matches above Latin1,
1961 * it assumes that all of Unicode could match (uniformly), except for
1962 * non-Unicode code points and things in the General Category "Other"
1963 * (unassigned, private use, surrogates, controls and formats). This
1964 * is a much large number. */
1966 U32 count = 0; /* Running total of number of code points matched by
1968 UV start, end; /* Start and end points of current range in inversion
1970 const U32 max_code_points = (LOC)
1972 : (( ! UNI_SEMANTICS
1973 || invlist_highest(ssc->invlist) < 256)
1976 const U32 max_match = max_code_points / 2;
1978 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1980 invlist_iterinit(ssc->invlist);
1981 while (invlist_iternext(ssc->invlist, &start, &end)) {
1982 if (start >= max_code_points) {
1985 end = MIN(end, max_code_points - 1);
1986 count += end - start + 1;
1987 if (count >= max_match) {
1988 invlist_iterfinish(ssc->invlist);
1998 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
2000 /* The inversion list in the SSC is marked mortal; now we need a more
2001 * permanent copy, which is stored the same way that is done in a regular
2002 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
2005 SV* invlist = invlist_clone(ssc->invlist);
2007 PERL_ARGS_ASSERT_SSC_FINALIZE;
2009 assert(is_ANYOF_SYNTHETIC(ssc));
2011 /* The code in this file assumes that all but these flags aren't relevant
2012 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
2013 * by the time we reach here */
2014 assert(! (ANYOF_FLAGS(ssc)
2015 & ~( ANYOF_COMMON_FLAGS
2016 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
2017 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
2019 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
2021 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
2022 NULL, NULL, NULL, FALSE);
2024 /* Make sure is clone-safe */
2025 ssc->invlist = NULL;
2027 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
2028 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
2031 if (RExC_contains_locale) {
2035 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
2038 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
2039 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
2040 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
2041 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
2042 ? (TRIE_LIST_CUR( idx ) - 1) \
2048 dump_trie(trie,widecharmap,revcharmap)
2049 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
2050 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
2052 These routines dump out a trie in a somewhat readable format.
2053 The _interim_ variants are used for debugging the interim
2054 tables that are used to generate the final compressed
2055 representation which is what dump_trie expects.
2057 Part of the reason for their existence is to provide a form
2058 of documentation as to how the different representations function.
2063 Dumps the final compressed table form of the trie to Perl_debug_log.
2064 Used for debugging make_trie().
2068 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
2069 AV *revcharmap, U32 depth)
2072 SV *sv=sv_newmortal();
2073 int colwidth= widecharmap ? 6 : 4;
2075 GET_RE_DEBUG_FLAGS_DECL;
2077 PERL_ARGS_ASSERT_DUMP_TRIE;
2079 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
2080 depth+1, "Match","Base","Ofs" );
2082 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2083 SV ** const tmp = av_fetch( revcharmap, state, 0);
2085 Perl_re_printf( aTHX_ "%*s",
2087 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2088 PL_colors[0], PL_colors[1],
2089 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2090 PERL_PV_ESCAPE_FIRSTCHAR
2095 Perl_re_printf( aTHX_ "\n");
2096 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
2098 for( state = 0 ; state < trie->uniquecharcount ; state++ )
2099 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2100 Perl_re_printf( aTHX_ "\n");
2102 for( state = 1 ; state < trie->statecount ; state++ ) {
2103 const U32 base = trie->states[ state ].trans.base;
2105 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
2107 if ( trie->states[ state ].wordnum ) {
2108 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
2110 Perl_re_printf( aTHX_ "%6s", "" );
2113 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
2118 while( ( base + ofs < trie->uniquecharcount ) ||
2119 ( base + ofs - trie->uniquecharcount < trie->lasttrans
2120 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2124 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
2126 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2127 if ( ( base + ofs >= trie->uniquecharcount )
2128 && ( base + ofs - trie->uniquecharcount
2130 && trie->trans[ base + ofs
2131 - trie->uniquecharcount ].check == state )
2133 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
2134 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2137 Perl_re_printf( aTHX_ "%*s",colwidth," ." );
2141 Perl_re_printf( aTHX_ "]");
2144 Perl_re_printf( aTHX_ "\n" );
2146 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
2148 for (word=1; word <= trie->wordcount; word++) {
2149 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2150 (int)word, (int)(trie->wordinfo[word].prev),
2151 (int)(trie->wordinfo[word].len));
2153 Perl_re_printf( aTHX_ "\n" );
2156 Dumps a fully constructed but uncompressed trie in list form.
2157 List tries normally only are used for construction when the number of
2158 possible chars (trie->uniquecharcount) is very high.
2159 Used for debugging make_trie().
2162 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2163 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2167 SV *sv=sv_newmortal();
2168 int colwidth= widecharmap ? 6 : 4;
2169 GET_RE_DEBUG_FLAGS_DECL;
2171 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2173 /* print out the table precompression. */
2174 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
2176 Perl_re_indentf( aTHX_ "%s",
2177 depth+1, "------:-----+-----------------\n" );
2179 for( state=1 ; state < next_alloc ; state ++ ) {
2182 Perl_re_indentf( aTHX_ " %4" UVXf " :",
2183 depth+1, (UV)state );
2184 if ( ! trie->states[ state ].wordnum ) {
2185 Perl_re_printf( aTHX_ "%5s| ","");
2187 Perl_re_printf( aTHX_ "W%4x| ",
2188 trie->states[ state ].wordnum
2191 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2192 SV ** const tmp = av_fetch( revcharmap,
2193 TRIE_LIST_ITEM(state,charid).forid, 0);
2195 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
2197 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2199 PL_colors[0], PL_colors[1],
2200 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2201 | PERL_PV_ESCAPE_FIRSTCHAR
2203 TRIE_LIST_ITEM(state,charid).forid,
2204 (UV)TRIE_LIST_ITEM(state,charid).newstate
2207 Perl_re_printf( aTHX_ "\n%*s| ",
2208 (int)((depth * 2) + 14), "");
2211 Perl_re_printf( aTHX_ "\n");
2216 Dumps a fully constructed but uncompressed trie in table form.
2217 This is the normal DFA style state transition table, with a few
2218 twists to facilitate compression later.
2219 Used for debugging make_trie().
2222 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2223 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2228 SV *sv=sv_newmortal();
2229 int colwidth= widecharmap ? 6 : 4;
2230 GET_RE_DEBUG_FLAGS_DECL;
2232 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2235 print out the table precompression so that we can do a visual check
2236 that they are identical.
2239 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
2241 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2242 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2244 Perl_re_printf( aTHX_ "%*s",
2246 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2247 PL_colors[0], PL_colors[1],
2248 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2249 PERL_PV_ESCAPE_FIRSTCHAR
2255 Perl_re_printf( aTHX_ "\n");
2256 Perl_re_indentf( aTHX_ "State+-", depth+1 );
2258 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2259 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
2262 Perl_re_printf( aTHX_ "\n" );
2264 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2266 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
2268 (UV)TRIE_NODENUM( state ) );
2270 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2271 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2273 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
2275 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
2277 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2278 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
2279 (UV)trie->trans[ state ].check );
2281 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
2282 (UV)trie->trans[ state ].check,
2283 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2291 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2292 startbranch: the first branch in the whole branch sequence
2293 first : start branch of sequence of branch-exact nodes.
2294 May be the same as startbranch
2295 last : Thing following the last branch.
2296 May be the same as tail.
2297 tail : item following the branch sequence
2298 count : words in the sequence
2299 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2300 depth : indent depth
2302 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2304 A trie is an N'ary tree where the branches are determined by digital
2305 decomposition of the key. IE, at the root node you look up the 1st character and
2306 follow that branch repeat until you find the end of the branches. Nodes can be
2307 marked as "accepting" meaning they represent a complete word. Eg:
2311 would convert into the following structure. Numbers represent states, letters
2312 following numbers represent valid transitions on the letter from that state, if
2313 the number is in square brackets it represents an accepting state, otherwise it
2314 will be in parenthesis.
2316 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2320 (1) +-i->(6)-+-s->[7]
2322 +-s->(3)-+-h->(4)-+-e->[5]
2324 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2326 This shows that when matching against the string 'hers' we will begin at state 1
2327 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2328 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2329 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2330 single traverse. We store a mapping from accepting to state to which word was
2331 matched, and then when we have multiple possibilities we try to complete the
2332 rest of the regex in the order in which they occurred in the alternation.
2334 The only prior NFA like behaviour that would be changed by the TRIE support is
2335 the silent ignoring of duplicate alternations which are of the form:
2337 / (DUPE|DUPE) X? (?{ ... }) Y /x
2339 Thus EVAL blocks following a trie may be called a different number of times with
2340 and without the optimisation. With the optimisations dupes will be silently
2341 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2342 the following demonstrates:
2344 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2346 which prints out 'word' three times, but
2348 'words'=~/(word|word|word)(?{ print $1 })S/
2350 which doesnt print it out at all. This is due to other optimisations kicking in.
2352 Example of what happens on a structural level:
2354 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2356 1: CURLYM[1] {1,32767}(18)
2367 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2368 and should turn into:
2370 1: CURLYM[1] {1,32767}(18)
2372 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2380 Cases where tail != last would be like /(?foo|bar)baz/:
2390 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2391 and would end up looking like:
2394 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2401 d = uvchr_to_utf8_flags(d, uv, 0);
2403 is the recommended Unicode-aware way of saying
2408 #define TRIE_STORE_REVCHAR(val) \
2411 SV *zlopp = newSV(UTF8_MAXBYTES); \
2412 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2413 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2414 SvCUR_set(zlopp, kapow - flrbbbbb); \
2417 av_push(revcharmap, zlopp); \
2419 char ooooff = (char)val; \
2420 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2424 /* This gets the next character from the input, folding it if not already
2426 #define TRIE_READ_CHAR STMT_START { \
2429 /* if it is UTF then it is either already folded, or does not need \
2431 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2433 else if (folder == PL_fold_latin1) { \
2434 /* This folder implies Unicode rules, which in the range expressible \
2435 * by not UTF is the lower case, with the two exceptions, one of \
2436 * which should have been taken care of before calling this */ \
2437 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2438 uvc = toLOWER_L1(*uc); \
2439 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2442 /* raw data, will be folded later if needed */ \
2450 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2451 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2452 U32 ging = TRIE_LIST_LEN( state ) * 2; \
2453 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2454 TRIE_LIST_LEN( state ) = ging; \
2456 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2457 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2458 TRIE_LIST_CUR( state )++; \
2461 #define TRIE_LIST_NEW(state) STMT_START { \
2462 Newx( trie->states[ state ].trans.list, \
2463 4, reg_trie_trans_le ); \
2464 TRIE_LIST_CUR( state ) = 1; \
2465 TRIE_LIST_LEN( state ) = 4; \
2468 #define TRIE_HANDLE_WORD(state) STMT_START { \
2469 U16 dupe= trie->states[ state ].wordnum; \
2470 regnode * const noper_next = regnext( noper ); \
2473 /* store the word for dumping */ \
2475 if (OP(noper) != NOTHING) \
2476 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2478 tmp = newSVpvn_utf8( "", 0, UTF ); \
2479 av_push( trie_words, tmp ); \
2483 trie->wordinfo[curword].prev = 0; \
2484 trie->wordinfo[curword].len = wordlen; \
2485 trie->wordinfo[curword].accept = state; \
2487 if ( noper_next < tail ) { \
2489 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2491 trie->jump[curword] = (U16)(noper_next - convert); \
2493 jumper = noper_next; \
2495 nextbranch= regnext(cur); \
2499 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2500 /* chain, so that when the bits of chain are later */\
2501 /* linked together, the dups appear in the chain */\
2502 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2503 trie->wordinfo[dupe].prev = curword; \
2505 /* we haven't inserted this word yet. */ \
2506 trie->states[ state ].wordnum = curword; \
2511 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2512 ( ( base + charid >= ucharcount \
2513 && base + charid < ubound \
2514 && state == trie->trans[ base - ucharcount + charid ].check \
2515 && trie->trans[ base - ucharcount + charid ].next ) \
2516 ? trie->trans[ base - ucharcount + charid ].next \
2517 : ( state==1 ? special : 0 ) \
2520 #define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2522 TRIE_BITMAP_SET(trie, uvc); \
2523 /* store the folded codepoint */ \
2525 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2528 /* store first byte of utf8 representation of */ \
2529 /* variant codepoints */ \
2530 if (! UVCHR_IS_INVARIANT(uvc)) { \
2531 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2536 #define MADE_JUMP_TRIE 2
2537 #define MADE_EXACT_TRIE 4
2540 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2541 regnode *first, regnode *last, regnode *tail,
2542 U32 word_count, U32 flags, U32 depth)
2544 /* first pass, loop through and scan words */
2545 reg_trie_data *trie;
2546 HV *widecharmap = NULL;
2547 AV *revcharmap = newAV();
2553 regnode *jumper = NULL;
2554 regnode *nextbranch = NULL;
2555 regnode *convert = NULL;
2556 U32 *prev_states; /* temp array mapping each state to previous one */
2557 /* we just use folder as a flag in utf8 */
2558 const U8 * folder = NULL;
2560 /* in the below add_data call we are storing either 'tu' or 'tuaa'
2561 * which stands for one trie structure, one hash, optionally followed
2564 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuaa"));
2565 AV *trie_words = NULL;
2566 /* along with revcharmap, this only used during construction but both are
2567 * useful during debugging so we store them in the struct when debugging.
2570 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2571 STRLEN trie_charcount=0;
2573 SV *re_trie_maxbuff;
2574 GET_RE_DEBUG_FLAGS_DECL;
2576 PERL_ARGS_ASSERT_MAKE_TRIE;
2578 PERL_UNUSED_ARG(depth);
2582 case EXACT: case EXACTL: break;
2586 case EXACTFLU8: folder = PL_fold_latin1; break;
2587 case EXACTF: folder = PL_fold; break;
2588 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2591 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2593 trie->startstate = 1;
2594 trie->wordcount = word_count;
2595 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2596 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2597 if (flags == EXACT || flags == EXACTL)
2598 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2599 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2600 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2603 trie_words = newAV();
2606 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2607 assert(re_trie_maxbuff);
2608 if (!SvIOK(re_trie_maxbuff)) {
2609 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2611 DEBUG_TRIE_COMPILE_r({
2612 Perl_re_indentf( aTHX_
2613 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2615 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2616 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2619 /* Find the node we are going to overwrite */
2620 if ( first == startbranch && OP( last ) != BRANCH ) {
2621 /* whole branch chain */
2624 /* branch sub-chain */
2625 convert = NEXTOPER( first );
2628 /* -- First loop and Setup --
2630 We first traverse the branches and scan each word to determine if it
2631 contains widechars, and how many unique chars there are, this is
2632 important as we have to build a table with at least as many columns as we
2635 We use an array of integers to represent the character codes 0..255
2636 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2637 the native representation of the character value as the key and IV's for
2640 *TODO* If we keep track of how many times each character is used we can
2641 remap the columns so that the table compression later on is more
2642 efficient in terms of memory by ensuring the most common value is in the
2643 middle and the least common are on the outside. IMO this would be better
2644 than a most to least common mapping as theres a decent chance the most
2645 common letter will share a node with the least common, meaning the node
2646 will not be compressible. With a middle is most common approach the worst
2647 case is when we have the least common nodes twice.
2651 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2652 regnode *noper = NEXTOPER( cur );
2656 U32 wordlen = 0; /* required init */
2657 STRLEN minchars = 0;
2658 STRLEN maxchars = 0;
2659 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2662 if (OP(noper) == NOTHING) {
2663 /* skip past a NOTHING at the start of an alternation
2664 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2666 regnode *noper_next= regnext(noper);
2667 if (noper_next < tail)
2671 if ( noper < tail &&
2673 OP(noper) == flags ||
2676 OP(noper) == EXACTFU_SS
2680 uc= (U8*)STRING(noper);
2681 e= uc + STR_LEN(noper);
2688 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2689 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2690 regardless of encoding */
2691 if (OP( noper ) == EXACTFU_SS) {
2692 /* false positives are ok, so just set this */
2693 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2697 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2699 TRIE_CHARCOUNT(trie)++;
2702 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2703 * is in effect. Under /i, this character can match itself, or
2704 * anything that folds to it. If not under /i, it can match just
2705 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2706 * all fold to k, and all are single characters. But some folds
2707 * expand to more than one character, so for example LATIN SMALL
2708 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2709 * the string beginning at 'uc' is 'ffi', it could be matched by
2710 * three characters, or just by the one ligature character. (It
2711 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2712 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2713 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2714 * match.) The trie needs to know the minimum and maximum number
2715 * of characters that could match so that it can use size alone to
2716 * quickly reject many match attempts. The max is simple: it is
2717 * the number of folded characters in this branch (since a fold is
2718 * never shorter than what folds to it. */
2722 /* And the min is equal to the max if not under /i (indicated by
2723 * 'folder' being NULL), or there are no multi-character folds. If
2724 * there is a multi-character fold, the min is incremented just
2725 * once, for the character that folds to the sequence. Each
2726 * character in the sequence needs to be added to the list below of
2727 * characters in the trie, but we count only the first towards the
2728 * min number of characters needed. This is done through the
2729 * variable 'foldlen', which is returned by the macros that look
2730 * for these sequences as the number of bytes the sequence
2731 * occupies. Each time through the loop, we decrement 'foldlen' by
2732 * how many bytes the current char occupies. Only when it reaches
2733 * 0 do we increment 'minchars' or look for another multi-character
2735 if (folder == NULL) {
2738 else if (foldlen > 0) {
2739 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2744 /* See if *uc is the beginning of a multi-character fold. If
2745 * so, we decrement the length remaining to look at, to account
2746 * for the current character this iteration. (We can use 'uc'
2747 * instead of the fold returned by TRIE_READ_CHAR because for
2748 * non-UTF, the latin1_safe macro is smart enough to account
2749 * for all the unfolded characters, and because for UTF, the
2750 * string will already have been folded earlier in the
2751 * compilation process */
2753 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2754 foldlen -= UTF8SKIP(uc);
2757 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2762 /* The current character (and any potential folds) should be added
2763 * to the possible matching characters for this position in this
2767 U8 folded= folder[ (U8) uvc ];
2768 if ( !trie->charmap[ folded ] ) {
2769 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2770 TRIE_STORE_REVCHAR( folded );
2773 if ( !trie->charmap[ uvc ] ) {
2774 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2775 TRIE_STORE_REVCHAR( uvc );
2778 /* store the codepoint in the bitmap, and its folded
2780 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
2781 set_bit = 0; /* We've done our bit :-) */
2785 /* XXX We could come up with the list of code points that fold
2786 * to this using PL_utf8_foldclosures, except not for
2787 * multi-char folds, as there may be multiple combinations
2788 * there that could work, which needs to wait until runtime to
2789 * resolve (The comment about LIGATURE FFI above is such an
2794 widecharmap = newHV();
2796 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2799 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
2801 if ( !SvTRUE( *svpp ) ) {
2802 sv_setiv( *svpp, ++trie->uniquecharcount );
2803 TRIE_STORE_REVCHAR(uvc);
2806 } /* end loop through characters in this branch of the trie */
2808 /* We take the min and max for this branch and combine to find the min
2809 * and max for all branches processed so far */
2810 if( cur == first ) {
2811 trie->minlen = minchars;
2812 trie->maxlen = maxchars;
2813 } else if (minchars < trie->minlen) {
2814 trie->minlen = minchars;
2815 } else if (maxchars > trie->maxlen) {
2816 trie->maxlen = maxchars;
2818 } /* end first pass */
2819 DEBUG_TRIE_COMPILE_r(
2820 Perl_re_indentf( aTHX_
2821 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2823 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2824 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2825 (int)trie->minlen, (int)trie->maxlen )
2829 We now know what we are dealing with in terms of unique chars and
2830 string sizes so we can calculate how much memory a naive
2831 representation using a flat table will take. If it's over a reasonable
2832 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2833 conservative but potentially much slower representation using an array
2836 At the end we convert both representations into the same compressed
2837 form that will be used in regexec.c for matching with. The latter
2838 is a form that cannot be used to construct with but has memory
2839 properties similar to the list form and access properties similar
2840 to the table form making it both suitable for fast searches and
2841 small enough that its feasable to store for the duration of a program.
2843 See the comment in the code where the compressed table is produced
2844 inplace from the flat tabe representation for an explanation of how
2845 the compression works.
2850 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2853 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2854 > SvIV(re_trie_maxbuff) )
2857 Second Pass -- Array Of Lists Representation
2859 Each state will be represented by a list of charid:state records
2860 (reg_trie_trans_le) the first such element holds the CUR and LEN
2861 points of the allocated array. (See defines above).
2863 We build the initial structure using the lists, and then convert
2864 it into the compressed table form which allows faster lookups
2865 (but cant be modified once converted).
2868 STRLEN transcount = 1;
2870 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
2873 trie->states = (reg_trie_state *)
2874 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2875 sizeof(reg_trie_state) );
2879 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2881 regnode *noper = NEXTOPER( cur );
2882 U32 state = 1; /* required init */
2883 U16 charid = 0; /* sanity init */
2884 U32 wordlen = 0; /* required init */
2886 if (OP(noper) == NOTHING) {
2887 regnode *noper_next= regnext(noper);
2888 if (noper_next < tail)
2892 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
2893 const U8 *uc= (U8*)STRING(noper);
2894 const U8 *e= uc + STR_LEN(noper);
2896 for ( ; uc < e ; uc += len ) {
2901 charid = trie->charmap[ uvc ];
2903 SV** const svpp = hv_fetch( widecharmap,
2910 charid=(U16)SvIV( *svpp );
2913 /* charid is now 0 if we dont know the char read, or
2914 * nonzero if we do */
2921 if ( !trie->states[ state ].trans.list ) {
2922 TRIE_LIST_NEW( state );
2925 check <= TRIE_LIST_USED( state );
2928 if ( TRIE_LIST_ITEM( state, check ).forid
2931 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2936 newstate = next_alloc++;
2937 prev_states[newstate] = state;
2938 TRIE_LIST_PUSH( state, charid, newstate );
2943 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
2947 TRIE_HANDLE_WORD(state);
2949 } /* end second pass */
2951 /* next alloc is the NEXT state to be allocated */
2952 trie->statecount = next_alloc;
2953 trie->states = (reg_trie_state *)
2954 PerlMemShared_realloc( trie->states,
2956 * sizeof(reg_trie_state) );
2958 /* and now dump it out before we compress it */
2959 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2960 revcharmap, next_alloc,
2964 trie->trans = (reg_trie_trans *)
2965 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2972 for( state=1 ; state < next_alloc ; state ++ ) {
2976 DEBUG_TRIE_COMPILE_MORE_r(
2977 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
2981 if (trie->states[state].trans.list) {
2982 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2986 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2987 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2988 if ( forid < minid ) {
2990 } else if ( forid > maxid ) {
2994 if ( transcount < tp + maxid - minid + 1) {
2996 trie->trans = (reg_trie_trans *)
2997 PerlMemShared_realloc( trie->trans,
2999 * sizeof(reg_trie_trans) );
3000 Zero( trie->trans + (transcount / 2),
3004 base = trie->uniquecharcount + tp - minid;
3005 if ( maxid == minid ) {
3007 for ( ; zp < tp ; zp++ ) {
3008 if ( ! trie->trans[ zp ].next ) {
3009 base = trie->uniquecharcount + zp - minid;
3010 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
3012 trie->trans[ zp ].check = state;
3018 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
3020 trie->trans[ tp ].check = state;
3025 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3026 const U32 tid = base
3027 - trie->uniquecharcount
3028 + TRIE_LIST_ITEM( state, idx ).forid;
3029 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
3031 trie->trans[ tid ].check = state;
3033 tp += ( maxid - minid + 1 );
3035 Safefree(trie->states[ state ].trans.list);
3038 DEBUG_TRIE_COMPILE_MORE_r(
3039 Perl_re_printf( aTHX_ " base: %d\n",base);
3042 trie->states[ state ].trans.base=base;
3044 trie->lasttrans = tp + 1;
3048 Second Pass -- Flat Table Representation.
3050 we dont use the 0 slot of either trans[] or states[] so we add 1 to
3051 each. We know that we will need Charcount+1 trans at most to store
3052 the data (one row per char at worst case) So we preallocate both
3053 structures assuming worst case.
3055 We then construct the trie using only the .next slots of the entry
3058 We use the .check field of the first entry of the node temporarily
3059 to make compression both faster and easier by keeping track of how
3060 many non zero fields are in the node.
3062 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
3065 There are two terms at use here: state as a TRIE_NODEIDX() which is
3066 a number representing the first entry of the node, and state as a
3067 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
3068 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
3069 if there are 2 entrys per node. eg:
3077 The table is internally in the right hand, idx form. However as we
3078 also have to deal with the states array which is indexed by nodenum
3079 we have to use TRIE_NODENUM() to convert.
3082 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
3085 trie->trans = (reg_trie_trans *)
3086 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
3087 * trie->uniquecharcount + 1,
3088 sizeof(reg_trie_trans) );
3089 trie->states = (reg_trie_state *)
3090 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3091 sizeof(reg_trie_state) );
3092 next_alloc = trie->uniquecharcount + 1;
3095 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3097 regnode *noper = NEXTOPER( cur );
3099 U32 state = 1; /* required init */
3101 U16 charid = 0; /* sanity init */
3102 U32 accept_state = 0; /* sanity init */
3104 U32 wordlen = 0; /* required init */
3106 if (OP(noper) == NOTHING) {
3107 regnode *noper_next= regnext(noper);
3108 if (noper_next < tail)
3112 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
3113 const U8 *uc= (U8*)STRING(noper);
3114 const U8 *e= uc + STR_LEN(noper);
3116 for ( ; uc < e ; uc += len ) {
3121 charid = trie->charmap[ uvc ];
3123 SV* const * const svpp = hv_fetch( widecharmap,
3127 charid = svpp ? (U16)SvIV(*svpp) : 0;
3131 if ( !trie->trans[ state + charid ].next ) {
3132 trie->trans[ state + charid ].next = next_alloc;
3133 trie->trans[ state ].check++;
3134 prev_states[TRIE_NODENUM(next_alloc)]
3135 = TRIE_NODENUM(state);
3136 next_alloc += trie->uniquecharcount;
3138 state = trie->trans[ state + charid ].next;
3140 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3142 /* charid is now 0 if we dont know the char read, or
3143 * nonzero if we do */
3146 accept_state = TRIE_NODENUM( state );
3147 TRIE_HANDLE_WORD(accept_state);
3149 } /* end second pass */
3151 /* and now dump it out before we compress it */
3152 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3154 next_alloc, depth+1));
3158 * Inplace compress the table.*
3160 For sparse data sets the table constructed by the trie algorithm will
3161 be mostly 0/FAIL transitions or to put it another way mostly empty.
3162 (Note that leaf nodes will not contain any transitions.)
3164 This algorithm compresses the tables by eliminating most such
3165 transitions, at the cost of a modest bit of extra work during lookup:
3167 - Each states[] entry contains a .base field which indicates the
3168 index in the state[] array wheres its transition data is stored.
3170 - If .base is 0 there are no valid transitions from that node.
3172 - If .base is nonzero then charid is added to it to find an entry in
3175 -If trans[states[state].base+charid].check!=state then the
3176 transition is taken to be a 0/Fail transition. Thus if there are fail
3177 transitions at the front of the node then the .base offset will point
3178 somewhere inside the previous nodes data (or maybe even into a node
3179 even earlier), but the .check field determines if the transition is
3183 The following process inplace converts the table to the compressed
3184 table: We first do not compress the root node 1,and mark all its
3185 .check pointers as 1 and set its .base pointer as 1 as well. This
3186 allows us to do a DFA construction from the compressed table later,
3187 and ensures that any .base pointers we calculate later are greater
3190 - We set 'pos' to indicate the first entry of the second node.
3192 - We then iterate over the columns of the node, finding the first and
3193 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3194 and set the .check pointers accordingly, and advance pos
3195 appropriately and repreat for the next node. Note that when we copy
3196 the next pointers we have to convert them from the original
3197 NODEIDX form to NODENUM form as the former is not valid post
3200 - If a node has no transitions used we mark its base as 0 and do not
3201 advance the pos pointer.
3203 - If a node only has one transition we use a second pointer into the
3204 structure to fill in allocated fail transitions from other states.
3205 This pointer is independent of the main pointer and scans forward
3206 looking for null transitions that are allocated to a state. When it
3207 finds one it writes the single transition into the "hole". If the
3208 pointer doesnt find one the single transition is appended as normal.
3210 - Once compressed we can Renew/realloc the structures to release the
3213 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3214 specifically Fig 3.47 and the associated pseudocode.
3218 const U32 laststate = TRIE_NODENUM( next_alloc );
3221 trie->statecount = laststate;
3223 for ( state = 1 ; state < laststate ; state++ ) {
3225 const U32 stateidx = TRIE_NODEIDX( state );
3226 const U32 o_used = trie->trans[ stateidx ].check;
3227 U32 used = trie->trans[ stateidx ].check;
3228 trie->trans[ stateidx ].check = 0;
3231 used && charid < trie->uniquecharcount;
3234 if ( flag || trie->trans[ stateidx + charid ].next ) {
3235 if ( trie->trans[ stateidx + charid ].next ) {
3237 for ( ; zp < pos ; zp++ ) {
3238 if ( ! trie->trans[ zp ].next ) {
3242 trie->states[ state ].trans.base
3244 + trie->uniquecharcount
3246 trie->trans[ zp ].next
3247 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3249 trie->trans[ zp ].check = state;
3250 if ( ++zp > pos ) pos = zp;
3257 trie->states[ state ].trans.base
3258 = pos + trie->uniquecharcount - charid ;
3260 trie->trans[ pos ].next
3261 = SAFE_TRIE_NODENUM(
3262 trie->trans[ stateidx + charid ].next );
3263 trie->trans[ pos ].check = state;
3268 trie->lasttrans = pos + 1;
3269 trie->states = (reg_trie_state *)
3270 PerlMemShared_realloc( trie->states, laststate
3271 * sizeof(reg_trie_state) );
3272 DEBUG_TRIE_COMPILE_MORE_r(
3273 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
3275 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3279 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3282 } /* end table compress */
3284 DEBUG_TRIE_COMPILE_MORE_r(
3285 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
3287 (UV)trie->statecount,
3288 (UV)trie->lasttrans)
3290 /* resize the trans array to remove unused space */
3291 trie->trans = (reg_trie_trans *)
3292 PerlMemShared_realloc( trie->trans, trie->lasttrans
3293 * sizeof(reg_trie_trans) );
3295 { /* Modify the program and insert the new TRIE node */
3296 U8 nodetype =(U8)(flags & 0xFF);
3300 regnode *optimize = NULL;
3301 #ifdef RE_TRACK_PATTERN_OFFSETS
3304 U32 mjd_nodelen = 0;
3305 #endif /* RE_TRACK_PATTERN_OFFSETS */
3306 #endif /* DEBUGGING */
3308 This means we convert either the first branch or the first Exact,
3309 depending on whether the thing following (in 'last') is a branch
3310 or not and whther first is the startbranch (ie is it a sub part of
3311 the alternation or is it the whole thing.)
3312 Assuming its a sub part we convert the EXACT otherwise we convert
3313 the whole branch sequence, including the first.
3315 /* Find the node we are going to overwrite */
3316 if ( first != startbranch || OP( last ) == BRANCH ) {
3317 /* branch sub-chain */
3318 NEXT_OFF( first ) = (U16)(last - first);
3319 #ifdef RE_TRACK_PATTERN_OFFSETS
3321 mjd_offset= Node_Offset((convert));
3322 mjd_nodelen= Node_Length((convert));
3325 /* whole branch chain */
3327 #ifdef RE_TRACK_PATTERN_OFFSETS
3330 const regnode *nop = NEXTOPER( convert );
3331 mjd_offset= Node_Offset((nop));
3332 mjd_nodelen= Node_Length((nop));
3336 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
3338 (UV)mjd_offset, (UV)mjd_nodelen)
3341 /* But first we check to see if there is a common prefix we can
3342 split out as an EXACT and put in front of the TRIE node. */
3343 trie->startstate= 1;
3344 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3345 /* we want to find the first state that has more than
3346 * one transition, if that state is not the first state
3347 * then we have a common prefix which we can remove.
3350 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3352 I32 first_ofs = -1; /* keeps track of the ofs of the first
3353 transition, -1 means none */
3355 const U32 base = trie->states[ state ].trans.base;
3357 /* does this state terminate an alternation? */
3358 if ( trie->states[state].wordnum )
3361 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3362 if ( ( base + ofs >= trie->uniquecharcount ) &&
3363 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3364 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3366 if ( ++count > 1 ) {
3367 /* we have more than one transition */
3370 /* if this is the first state there is no common prefix
3371 * to extract, so we can exit */
3372 if ( state == 1 ) break;
3373 tmp = av_fetch( revcharmap, ofs, 0);
3374 ch = (U8*)SvPV_nolen_const( *tmp );
3376 /* if we are on count 2 then we need to initialize the
3377 * bitmap, and store the previous char if there was one
3380 /* clear the bitmap */
3381 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3383 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
3386 if (first_ofs >= 0) {
3387 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
3388 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3390 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
3392 Perl_re_printf( aTHX_ "%s", (char*)ch)
3396 /* store the current firstchar in the bitmap */
3397 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
3398 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
3404 /* This state has only one transition, its transition is part
3405 * of a common prefix - we need to concatenate the char it
3406 * represents to what we have so far. */
3407 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
3409 char *ch = SvPV( *tmp, len );
3411 SV *sv=sv_newmortal();
3412 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
3414 (UV)state, (UV)first_ofs,
3415 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3416 PL_colors[0], PL_colors[1],
3417 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3418 PERL_PV_ESCAPE_FIRSTCHAR
3423 OP( convert ) = nodetype;
3424 str=STRING(convert);
3427 STR_LEN(convert) += len;
3433 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
3438 trie->prefixlen = (state-1);
3440 regnode *n = convert+NODE_SZ_STR(convert);
3441 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3442 trie->startstate = state;
3443 trie->minlen -= (state - 1);
3444 trie->maxlen -= (state - 1);
3446 /* At least the UNICOS C compiler choked on this
3447 * being argument to DEBUG_r(), so let's just have
3450 #ifdef PERL_EXT_RE_BUILD
3456 regnode *fix = convert;
3457 U32 word = trie->wordcount;
3459 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3460 while( ++fix < n ) {
3461 Set_Node_Offset_Length(fix, 0, 0);
3464 SV ** const tmp = av_fetch( trie_words, word, 0 );
3466 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3467 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3469 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3477 NEXT_OFF(convert) = (U16)(tail - convert);
3478 DEBUG_r(optimize= n);
3484 if ( trie->maxlen ) {
3485 NEXT_OFF( convert ) = (U16)(tail - convert);
3486 ARG_SET( convert, data_slot );
3487 /* Store the offset to the first unabsorbed branch in
3488 jump[0], which is otherwise unused by the jump logic.
3489 We use this when dumping a trie and during optimisation. */
3491 trie->jump[0] = (U16)(nextbranch - convert);
3493 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3494 * and there is a bitmap
3495 * and the first "jump target" node we found leaves enough room
3496 * then convert the TRIE node into a TRIEC node, with the bitmap
3497 * embedded inline in the opcode - this is hypothetically faster.
3499 if ( !trie->states[trie->startstate].wordnum
3501 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3503 OP( convert ) = TRIEC;
3504 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3505 PerlMemShared_free(trie->bitmap);
3508 OP( convert ) = TRIE;
3510 /* store the type in the flags */
3511 convert->flags = nodetype;
3515 + regarglen[ OP( convert ) ];
3517 /* XXX We really should free up the resource in trie now,
3518 as we won't use them - (which resources?) dmq */
3520 /* needed for dumping*/
3521 DEBUG_r(if (optimize) {
3522 regnode *opt = convert;
3524 while ( ++opt < optimize) {
3525 Set_Node_Offset_Length(opt,0,0);
3528 Try to clean up some of the debris left after the
3531 while( optimize < jumper ) {
3532 mjd_nodelen += Node_Length((optimize));
3533 OP( optimize ) = OPTIMIZED;
3534 Set_Node_Offset_Length(optimize,0,0);
3537 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
3539 } /* end node insert */
3541 /* Finish populating the prev field of the wordinfo array. Walk back
3542 * from each accept state until we find another accept state, and if
3543 * so, point the first word's .prev field at the second word. If the
3544 * second already has a .prev field set, stop now. This will be the
3545 * case either if we've already processed that word's accept state,
3546 * or that state had multiple words, and the overspill words were
3547 * already linked up earlier.
3554 for (word=1; word <= trie->wordcount; word++) {
3556 if (trie->wordinfo[word].prev)
3558 state = trie->wordinfo[word].accept;
3560 state = prev_states[state];
3563 prev = trie->states[state].wordnum;
3567 trie->wordinfo[word].prev = prev;
3569 Safefree(prev_states);
3573 /* and now dump out the compressed format */
3574 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3576 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3578 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3579 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3581 SvREFCNT_dec_NN(revcharmap);
3585 : trie->startstate>1
3591 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3593 /* The Trie is constructed and compressed now so we can build a fail array if
3596 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3598 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3602 We find the fail state for each state in the trie, this state is the longest
3603 proper suffix of the current state's 'word' that is also a proper prefix of
3604 another word in our trie. State 1 represents the word '' and is thus the
3605 default fail state. This allows the DFA not to have to restart after its
3606 tried and failed a word at a given point, it simply continues as though it
3607 had been matching the other word in the first place.
3609 'abcdgu'=~/abcdefg|cdgu/
3610 When we get to 'd' we are still matching the first word, we would encounter
3611 'g' which would fail, which would bring us to the state representing 'd' in
3612 the second word where we would try 'g' and succeed, proceeding to match
3615 /* add a fail transition */
3616 const U32 trie_offset = ARG(source);
3617 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3619 const U32 ucharcount = trie->uniquecharcount;
3620 const U32 numstates = trie->statecount;
3621 const U32 ubound = trie->lasttrans + ucharcount;
3625 U32 base = trie->states[ 1 ].trans.base;
3628 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3630 GET_RE_DEBUG_FLAGS_DECL;
3632 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3633 PERL_UNUSED_CONTEXT;
3635 PERL_UNUSED_ARG(depth);
3638 if ( OP(source) == TRIE ) {
3639 struct regnode_1 *op = (struct regnode_1 *)
3640 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3641 StructCopy(source,op,struct regnode_1);
3642 stclass = (regnode *)op;
3644 struct regnode_charclass *op = (struct regnode_charclass *)
3645 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3646 StructCopy(source,op,struct regnode_charclass);
3647 stclass = (regnode *)op;
3649 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3651 ARG_SET( stclass, data_slot );
3652 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3653 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3654 aho->trie=trie_offset;
3655 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3656 Copy( trie->states, aho->states, numstates, reg_trie_state );
3657 Newx( q, numstates, U32);
3658 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3661 /* initialize fail[0..1] to be 1 so that we always have
3662 a valid final fail state */
3663 fail[ 0 ] = fail[ 1 ] = 1;
3665 for ( charid = 0; charid < ucharcount ; charid++ ) {
3666 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3668 q[ q_write ] = newstate;
3669 /* set to point at the root */
3670 fail[ q[ q_write++ ] ]=1;
3673 while ( q_read < q_write) {
3674 const U32 cur = q[ q_read++ % numstates ];
3675 base = trie->states[ cur ].trans.base;
3677 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3678 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3680 U32 fail_state = cur;
3683 fail_state = fail[ fail_state ];
3684 fail_base = aho->states[ fail_state ].trans.base;
3685 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3687 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3688 fail[ ch_state ] = fail_state;
3689 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3691 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3693 q[ q_write++ % numstates] = ch_state;
3697 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3698 when we fail in state 1, this allows us to use the
3699 charclass scan to find a valid start char. This is based on the principle
3700 that theres a good chance the string being searched contains lots of stuff
3701 that cant be a start char.
3703 fail[ 0 ] = fail[ 1 ] = 0;
3704 DEBUG_TRIE_COMPILE_r({
3705 Perl_re_indentf( aTHX_ "Stclass Failtable (%" UVuf " states): 0",
3706 depth, (UV)numstates
3708 for( q_read=1; q_read<numstates; q_read++ ) {
3709 Perl_re_printf( aTHX_ ", %" UVuf, (UV)fail[q_read]);
3711 Perl_re_printf( aTHX_ "\n");
3714 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3719 /* The below joins as many adjacent EXACTish nodes as possible into a single
3720 * one. The regop may be changed if the node(s) contain certain sequences that
3721 * require special handling. The joining is only done if:
3722 * 1) there is room in the current conglomerated node to entirely contain the
3724 * 2) they are the exact same node type
3726 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3727 * these get optimized out
3729 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3730 * as possible, even if that means splitting an existing node so that its first
3731 * part is moved to the preceeding node. This would maximise the efficiency of
3732 * memEQ during matching.
3734 * If a node is to match under /i (folded), the number of characters it matches
3735 * can be different than its character length if it contains a multi-character
3736 * fold. *min_subtract is set to the total delta number of characters of the
3739 * And *unfolded_multi_char is set to indicate whether or not the node contains
3740 * an unfolded multi-char fold. This happens when it won't be known until
3741 * runtime whether the fold is valid or not; namely
3742 * 1) for EXACTF nodes that contain LATIN SMALL LETTER SHARP S, as only if the
3743 * target string being matched against turns out to be UTF-8 is that fold
3745 * 2) for EXACTFL nodes whose folding rules depend on the locale in force at
3747 * (Multi-char folds whose components are all above the Latin1 range are not
3748 * run-time locale dependent, and have already been folded by the time this
3749 * function is called.)
3751 * This is as good a place as any to discuss the design of handling these
3752 * multi-character fold sequences. It's been wrong in Perl for a very long
3753 * time. There are three code points in Unicode whose multi-character folds
3754 * were long ago discovered to mess things up. The previous designs for
3755 * dealing with these involved assigning a special node for them. This
3756 * approach doesn't always work, as evidenced by this example:
3757 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3758 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3759 * would match just the \xDF, it won't be able to handle the case where a
3760 * successful match would have to cross the node's boundary. The new approach
3761 * that hopefully generally solves the problem generates an EXACTFU_SS node
3762 * that is "sss" in this case.
3764 * It turns out that there are problems with all multi-character folds, and not
3765 * just these three. Now the code is general, for all such cases. The
3766 * approach taken is:
3767 * 1) This routine examines each EXACTFish node that could contain multi-
3768 * character folded sequences. Since a single character can fold into
3769 * such a sequence, the minimum match length for this node is less than
3770 * the number of characters in the node. This routine returns in
3771 * *min_subtract how many characters to subtract from the the actual
3772 * length of the string to get a real minimum match length; it is 0 if
3773 * there are no multi-char foldeds. This delta is used by the caller to
3774 * adjust the min length of the match, and the delta between min and max,
3775 * so that the optimizer doesn't reject these possibilities based on size
3777 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3778 * is used for an EXACTFU node that contains at least one "ss" sequence in
3779 * it. For non-UTF-8 patterns and strings, this is the only case where
3780 * there is a possible fold length change. That means that a regular
3781 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3782 * with length changes, and so can be processed faster. regexec.c takes
3783 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3784 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3785 * known until runtime). This saves effort in regex matching. However,
3786 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3787 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3788 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3789 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3790 * possibilities for the non-UTF8 patterns are quite simple, except for
3791 * the sharp s. All the ones that don't involve a UTF-8 target string are
3792 * members of a fold-pair, and arrays are set up for all of them so that
3793 * the other member of the pair can be found quickly. Code elsewhere in
3794 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3795 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3796 * described in the next item.
3797 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3798 * validity of the fold won't be known until runtime, and so must remain
3799 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFAA
3800 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3801 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3802 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3803 * The reason this is a problem is that the optimizer part of regexec.c
3804 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3805 * that a character in the pattern corresponds to at most a single
3806 * character in the target string. (And I do mean character, and not byte
3807 * here, unlike other parts of the documentation that have never been
3808 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3809 * EXACTFL nodes can match the two character string 'ss'; in EXACTFAA
3810 * nodes it can match "\x{17F}\x{17F}". These, along with other ones in
3811 * EXACTFL nodes, violate the assumption, and they are the only instances
3812 * where it is violated. I'm reluctant to try to change the assumption,
3813 * as the code involved is impenetrable to me (khw), so instead the code
3814 * here punts. This routine examines EXACTFL nodes, and (when the pattern
3815 * isn't UTF-8) EXACTF and EXACTFAA for such unfolded folds, and returns a
3816 * boolean indicating whether or not the node contains such a fold. When
3817 * it is true, the caller sets a flag that later causes the optimizer in
3818 * this file to not set values for the floating and fixed string lengths,
3819 * and thus avoids the optimizer code in regexec.c that makes the invalid
3820 * assumption. Thus, there is no optimization based on string lengths for
3821 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3822 * EXACTF and EXACTFAA nodes that contain the sharp s. (The reason the
3823 * assumption is wrong only in these cases is that all other non-UTF-8
3824 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3825 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3826 * EXACTF nodes because we don't know at compile time if it actually
3827 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3828 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3829 * always matches; and EXACTFAA where it never does. In an EXACTFAA node
3830 * in a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3831 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3832 * string would require the pattern to be forced into UTF-8, the overhead
3833 * of which we want to avoid. Similarly the unfolded multi-char folds in
3834 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3837 * Similarly, the code that generates tries doesn't currently handle
3838 * not-already-folded multi-char folds, and it looks like a pain to change
3839 * that. Therefore, trie generation of EXACTFAA nodes with the sharp s
3840 * doesn't work. Instead, such an EXACTFAA is turned into a new regnode,
3841 * EXACTFAA_NO_TRIE, which the trie code knows not to handle. Most people
3842 * using /iaa matching will be doing so almost entirely with ASCII
3843 * strings, so this should rarely be encountered in practice */
3845 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3846 if (PL_regkind[OP(scan)] == EXACT) \
3847 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3850 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3851 UV *min_subtract, bool *unfolded_multi_char,
3852 U32 flags,regnode *val, U32 depth)
3854 /* Merge several consecutive EXACTish nodes into one. */
3855 regnode *n = regnext(scan);
3857 regnode *next = scan + NODE_SZ_STR(scan);
3861 regnode *stop = scan;
3862 GET_RE_DEBUG_FLAGS_DECL;
3864 PERL_UNUSED_ARG(depth);
3867 PERL_ARGS_ASSERT_JOIN_EXACT;
3868 #ifndef EXPERIMENTAL_INPLACESCAN
3869 PERL_UNUSED_ARG(flags);
3870 PERL_UNUSED_ARG(val);
3872 DEBUG_PEEP("join", scan, depth, 0);
3874 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3875 * EXACT ones that are mergeable to the current one. */
3877 && (PL_regkind[OP(n)] == NOTHING
3878 || (stringok && OP(n) == OP(scan)))
3880 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3883 if (OP(n) == TAIL || n > next)
3885 if (PL_regkind[OP(n)] == NOTHING) {
3886 DEBUG_PEEP("skip:", n, depth, 0);
3887 NEXT_OFF(scan) += NEXT_OFF(n);
3888 next = n + NODE_STEP_REGNODE;
3895 else if (stringok) {
3896 const unsigned int oldl = STR_LEN(scan);
3897 regnode * const nnext = regnext(n);
3899 /* XXX I (khw) kind of doubt that this works on platforms (should
3900 * Perl ever run on one) where U8_MAX is above 255 because of lots
3901 * of other assumptions */
3902 /* Don't join if the sum can't fit into a single node */
3903 if (oldl + STR_LEN(n) > U8_MAX)
3906 DEBUG_PEEP("merg", n, depth, 0);
3909 NEXT_OFF(scan) += NEXT_OFF(n);
3910 STR_LEN(scan) += STR_LEN(n);
3911 next = n + NODE_SZ_STR(n);
3912 /* Now we can overwrite *n : */
3913 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3921 #ifdef EXPERIMENTAL_INPLACESCAN
3922 if (flags && !NEXT_OFF(n)) {
3923 DEBUG_PEEP("atch", val, depth, 0);
3924 if (reg_off_by_arg[OP(n)]) {
3925 ARG_SET(n, val - n);
3928 NEXT_OFF(n) = val - n;
3936 *unfolded_multi_char = FALSE;
3938 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3939 * can now analyze for sequences of problematic code points. (Prior to
3940 * this final joining, sequences could have been split over boundaries, and
3941 * hence missed). The sequences only happen in folding, hence for any
3942 * non-EXACT EXACTish node */
3943 if (OP(scan) != EXACT && OP(scan) != EXACTL) {
3944 U8* s0 = (U8*) STRING(scan);
3946 U8* s_end = s0 + STR_LEN(scan);
3948 int total_count_delta = 0; /* Total delta number of characters that
3949 multi-char folds expand to */
3951 /* One pass is made over the node's string looking for all the
3952 * possibilities. To avoid some tests in the loop, there are two main
3953 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3958 if (OP(scan) == EXACTFL) {
3961 /* An EXACTFL node would already have been changed to another
3962 * node type unless there is at least one character in it that
3963 * is problematic; likely a character whose fold definition
3964 * won't be known until runtime, and so has yet to be folded.
3965 * For all but the UTF-8 locale, folds are 1-1 in length, but
3966 * to handle the UTF-8 case, we need to create a temporary
3967 * folded copy using UTF-8 locale rules in order to analyze it.
3968 * This is because our macros that look to see if a sequence is
3969 * a multi-char fold assume everything is folded (otherwise the
3970 * tests in those macros would be too complicated and slow).
3971 * Note that here, the non-problematic folds will have already
3972 * been done, so we can just copy such characters. We actually
3973 * don't completely fold the EXACTFL string. We skip the
3974 * unfolded multi-char folds, as that would just create work
3975 * below to figure out the size they already are */
3977 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3980 STRLEN s_len = UTF8SKIP(s);
3981 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3982 Copy(s, d, s_len, U8);
3985 else if (is_FOLDS_TO_MULTI_utf8(s)) {
3986 *unfolded_multi_char = TRUE;
3987 Copy(s, d, s_len, U8);
3990 else if (isASCII(*s)) {
3991 *(d++) = toFOLD(*s);
3995 _toFOLD_utf8_flags(s, s_end, d, &len, FOLD_FLAGS_FULL);
4001 /* Point the remainder of the routine to look at our temporary
4005 } /* End of creating folded copy of EXACTFL string */
4007 /* Examine the string for a multi-character fold sequence. UTF-8
4008 * patterns have all characters pre-folded by the time this code is
4010 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
4011 length sequence we are looking for is 2 */
4013 int count = 0; /* How many characters in a multi-char fold */
4014 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
4015 if (! len) { /* Not a multi-char fold: get next char */
4020 /* Nodes with 'ss' require special handling, except for
4021 * EXACTFAA-ish for which there is no multi-char fold to this */
4022 if (len == 2 && *s == 's' && *(s+1) == 's'
4023 && OP(scan) != EXACTFAA
4024 && OP(scan) != EXACTFAA_NO_TRIE)
4027 if (OP(scan) != EXACTFL) {
4028 OP(scan) = EXACTFU_SS;
4032 else { /* Here is a generic multi-char fold. */
4033 U8* multi_end = s + len;
4035 /* Count how many characters are in it. In the case of
4036 * /aa, no folds which contain ASCII code points are
4037 * allowed, so check for those, and skip if found. */
4038 if (OP(scan) != EXACTFAA && OP(scan) != EXACTFAA_NO_TRIE) {
4039 count = utf8_length(s, multi_end);
4043 while (s < multi_end) {
4046 goto next_iteration;
4056 /* The delta is how long the sequence is minus 1 (1 is how long
4057 * the character that folds to the sequence is) */
4058 total_count_delta += count - 1;
4062 /* We created a temporary folded copy of the string in EXACTFL
4063 * nodes. Therefore we need to be sure it doesn't go below zero,
4064 * as the real string could be shorter */
4065 if (OP(scan) == EXACTFL) {
4066 int total_chars = utf8_length((U8*) STRING(scan),
4067 (U8*) STRING(scan) + STR_LEN(scan));
4068 if (total_count_delta > total_chars) {
4069 total_count_delta = total_chars;
4073 *min_subtract += total_count_delta;
4076 else if (OP(scan) == EXACTFAA) {
4078 /* Non-UTF-8 pattern, EXACTFAA node. There can't be a multi-char
4079 * fold to the ASCII range (and there are no existing ones in the
4080 * upper latin1 range). But, as outlined in the comments preceding
4081 * this function, we need to flag any occurrences of the sharp s.
4082 * This character forbids trie formation (because of added
4084 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
4085 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
4086 || UNICODE_DOT_DOT_VERSION > 0)
4088 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
4089 OP(scan) = EXACTFAA_NO_TRIE;
4090 *unfolded_multi_char = TRUE;
4098 /* Non-UTF-8 pattern, not EXACTFAA node. Look for the multi-char
4099 * folds that are all Latin1. As explained in the comments
4100 * preceding this function, we look also for the sharp s in EXACTF
4101 * and EXACTFL nodes; it can be in the final position. Otherwise
4102 * we can stop looking 1 byte earlier because have to find at least
4103 * two characters for a multi-fold */
4104 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
4109 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
4110 if (! len) { /* Not a multi-char fold. */
4111 if (*s == LATIN_SMALL_LETTER_SHARP_S
4112 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
4114 *unfolded_multi_char = TRUE;
4121 && isALPHA_FOLD_EQ(*s, 's')
4122 && isALPHA_FOLD_EQ(*(s+1), 's'))
4125 /* EXACTF nodes need to know that the minimum length
4126 * changed so that a sharp s in the string can match this
4127 * ss in the pattern, but they remain EXACTF nodes, as they
4128 * won't match this unless the target string is is UTF-8,
4129 * which we don't know until runtime. EXACTFL nodes can't
4130 * transform into EXACTFU nodes */
4131 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
4132 OP(scan) = EXACTFU_SS;
4136 *min_subtract += len - 1;
4144 /* Allow dumping but overwriting the collection of skipped
4145 * ops and/or strings with fake optimized ops */
4146 n = scan + NODE_SZ_STR(scan);
4154 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl", scan, depth, 0);});
4158 /* REx optimizer. Converts nodes into quicker variants "in place".
4159 Finds fixed substrings. */
4161 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
4162 to the position after last scanned or to NULL. */
4164 #define INIT_AND_WITHP \
4165 assert(!and_withp); \
4166 Newx(and_withp,1, regnode_ssc); \
4167 SAVEFREEPV(and_withp)
4171 S_unwind_scan_frames(pTHX_ const void *p)
4173 scan_frame *f= (scan_frame *)p;
4175 scan_frame *n= f->next_frame;
4181 /* the return from this sub is the minimum length that could possibly match */
4183 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
4184 SSize_t *minlenp, SSize_t *deltap,
4189 regnode_ssc *and_withp,
4190 U32 flags, U32 depth)
4191 /* scanp: Start here (read-write). */
4192 /* deltap: Write maxlen-minlen here. */
4193 /* last: Stop before this one. */
4194 /* data: string data about the pattern */
4195 /* stopparen: treat close N as END */
4196 /* recursed: which subroutines have we recursed into */
4197 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
4199 /* There must be at least this number of characters to match */
4202 regnode *scan = *scanp, *next;
4204 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
4205 int is_inf_internal = 0; /* The studied chunk is infinite */
4206 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
4207 scan_data_t data_fake;
4208 SV *re_trie_maxbuff = NULL;
4209 regnode *first_non_open = scan;
4210 SSize_t stopmin = SSize_t_MAX;
4211 scan_frame *frame = NULL;
4212 GET_RE_DEBUG_FLAGS_DECL;
4214 PERL_ARGS_ASSERT_STUDY_CHUNK;
4215 RExC_study_started= 1;
4217 Zero(&data_fake, 1, scan_data_t);
4220 while (first_non_open && OP(first_non_open) == OPEN)
4221 first_non_open=regnext(first_non_open);
4227 RExC_study_chunk_recursed_count++;
4229 DEBUG_OPTIMISE_MORE_r(
4231 Perl_re_indentf( aTHX_ "study_chunk stopparen=%ld recursed_count=%lu depth=%lu recursed_depth=%lu scan=%p last=%p",
4232 depth, (long)stopparen,
4233 (unsigned long)RExC_study_chunk_recursed_count,
4234 (unsigned long)depth, (unsigned long)recursed_depth,
4237 if (recursed_depth) {
4240 for ( j = 0 ; j < recursed_depth ; j++ ) {
4241 for ( i = 0 ; i < (U32)RExC_npar ; i++ ) {
4243 PAREN_TEST(RExC_study_chunk_recursed +
4244 ( j * RExC_study_chunk_recursed_bytes), i )
4247 !PAREN_TEST(RExC_study_chunk_recursed +
4248 (( j - 1 ) * RExC_study_chunk_recursed_bytes), i)
4251 Perl_re_printf( aTHX_ " %d",(int)i);
4255 if ( j + 1 < recursed_depth ) {
4256 Perl_re_printf( aTHX_ ",");
4260 Perl_re_printf( aTHX_ "\n");
4263 while ( scan && OP(scan) != END && scan < last ){
4264 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
4265 node length to get a real minimum (because
4266 the folded version may be shorter) */
4267 bool unfolded_multi_char = FALSE;
4268 /* Peephole optimizer: */
4269 DEBUG_STUDYDATA("Peep", data, depth, is_inf);
4270 DEBUG_PEEP("Peep", scan, depth, flags);
4273 /* The reason we do this here is that we need to deal with things like
4274 * /(?:f)(?:o)(?:o)/ which cant be dealt with by the normal EXACT
4275 * parsing code, as each (?:..) is handled by a different invocation of
4278 JOIN_EXACT(scan,&min_subtract, &unfolded_multi_char, 0);
4280 /* Follow the next-chain of the current node and optimize
4281 away all the NOTHINGs from it. */
4282 if (OP(scan) != CURLYX) {
4283 const int max = (reg_off_by_arg[OP(scan)]
4285 /* I32 may be smaller than U16 on CRAYs! */
4286 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
4287 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
4291 /* Skip NOTHING and LONGJMP. */
4292 while ((n = regnext(n))
4293 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
4294 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
4295 && off + noff < max)
4297 if (reg_off_by_arg[OP(scan)])
4300 NEXT_OFF(scan) = off;
4303 /* The principal pseudo-switch. Cannot be a switch, since we
4304 look into several different things. */
4305 if ( OP(scan) == DEFINEP ) {
4307 SSize_t deltanext = 0;
4308 SSize_t fake_last_close = 0;
4309 I32 f = SCF_IN_DEFINE;
4311 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4312 scan = regnext(scan);
4313 assert( OP(scan) == IFTHEN );
4314 DEBUG_PEEP("expect IFTHEN", scan, depth, flags);
4316 data_fake.last_closep= &fake_last_close;
4318 next = regnext(scan);
4319 scan = NEXTOPER(NEXTOPER(scan));
4320 DEBUG_PEEP("scan", scan, depth, flags);
4321 DEBUG_PEEP("next", next, depth, flags);
4323 /* we suppose the run is continuous, last=next...
4324 * NOTE we dont use the return here! */
4325 /* DEFINEP study_chunk() recursion */
4326 (void)study_chunk(pRExC_state, &scan, &minlen,
4327 &deltanext, next, &data_fake, stopparen,
4328 recursed_depth, NULL, f, depth+1);
4333 OP(scan) == BRANCH ||
4334 OP(scan) == BRANCHJ ||
4337 next = regnext(scan);
4340 /* The op(next)==code check below is to see if we
4341 * have "BRANCH-BRANCH", "BRANCHJ-BRANCHJ", "IFTHEN-IFTHEN"
4342 * IFTHEN is special as it might not appear in pairs.
4343 * Not sure whether BRANCH-BRANCHJ is possible, regardless
4344 * we dont handle it cleanly. */
4345 if (OP(next) == code || code == IFTHEN) {
4346 /* NOTE - There is similar code to this block below for
4347 * handling TRIE nodes on a re-study. If you change stuff here
4348 * check there too. */
4349 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
4351 regnode * const startbranch=scan;
4353 if (flags & SCF_DO_SUBSTR) {
4354 /* Cannot merge strings after this. */
4355 scan_commit(pRExC_state, data, minlenp, is_inf);
4358 if (flags & SCF_DO_STCLASS)
4359 ssc_init_zero(pRExC_state, &accum);
4361 while (OP(scan) == code) {
4362 SSize_t deltanext, minnext, fake;
4364 regnode_ssc this_class;
4366 DEBUG_PEEP("Branch", scan, depth, flags);
4369 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4371 data_fake.whilem_c = data->whilem_c;
4372 data_fake.last_closep = data->last_closep;
4375 data_fake.last_closep = &fake;
4377 data_fake.pos_delta = delta;
4378 next = regnext(scan);
4380 scan = NEXTOPER(scan); /* everything */
4381 if (code != BRANCH) /* everything but BRANCH */
4382 scan = NEXTOPER(scan);
4384 if (flags & SCF_DO_STCLASS) {
4385 ssc_init(pRExC_state, &this_class);
4386 data_fake.start_class = &this_class;
4387 f = SCF_DO_STCLASS_AND;
4389 if (flags & SCF_WHILEM_VISITED_POS)
4390 f |= SCF_WHILEM_VISITED_POS;
4392 /* we suppose the run is continuous, last=next...*/
4393 /* recurse study_chunk() for each BRANCH in an alternation */
4394 minnext = study_chunk(pRExC_state, &scan, minlenp,
4395 &deltanext, next, &data_fake, stopparen,
4396 recursed_depth, NULL, f,depth+1);
4400 if (deltanext == SSize_t_MAX) {
4401 is_inf = is_inf_internal = 1;
4403 } else if (max1 < minnext + deltanext)
4404 max1 = minnext + deltanext;
4406 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4408 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4409 if ( stopmin > minnext)
4410 stopmin = min + min1;
4411 flags &= ~SCF_DO_SUBSTR;
4413 data->flags |= SCF_SEEN_ACCEPT;
4416 if (data_fake.flags & SF_HAS_EVAL)
4417 data->flags |= SF_HAS_EVAL;
4418 data->whilem_c = data_fake.whilem_c;
4420 if (flags & SCF_DO_STCLASS)
4421 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
4423 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
4425 if (flags & SCF_DO_SUBSTR) {
4426 data->pos_min += min1;
4427 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
4428 data->pos_delta = SSize_t_MAX;
4430 data->pos_delta += max1 - min1;
4431 if (max1 != min1 || is_inf)
4432 data->cur_is_floating = 1;
4435 if (delta == SSize_t_MAX
4436 || SSize_t_MAX - delta - (max1 - min1) < 0)
4437 delta = SSize_t_MAX;
4439 delta += max1 - min1;
4440 if (flags & SCF_DO_STCLASS_OR) {
4441 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
4443 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4444 flags &= ~SCF_DO_STCLASS;
4447 else if (flags & SCF_DO_STCLASS_AND) {
4449 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
4450 flags &= ~SCF_DO_STCLASS;
4453 /* Switch to OR mode: cache the old value of
4454 * data->start_class */
4456 StructCopy(data->start_class, and_withp, regnode_ssc);
4457 flags &= ~SCF_DO_STCLASS_AND;
4458 StructCopy(&accum, data->start_class, regnode_ssc);
4459 flags |= SCF_DO_STCLASS_OR;
4463 if (PERL_ENABLE_TRIE_OPTIMISATION &&
4464 OP( startbranch ) == BRANCH )
4468 Assuming this was/is a branch we are dealing with: 'scan'
4469 now points at the item that follows the branch sequence,
4470 whatever it is. We now start at the beginning of the
4471 sequence and look for subsequences of
4477 which would be constructed from a pattern like
4480 If we can find such a subsequence we need to turn the first
4481 element into a trie and then add the subsequent branch exact
4482 strings to the trie.
4486 1. patterns where the whole set of branches can be
4489 2. patterns where only a subset can be converted.
4491 In case 1 we can replace the whole set with a single regop
4492 for the trie. In case 2 we need to keep the start and end
4495 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
4496 becomes BRANCH TRIE; BRANCH X;
4498 There is an additional case, that being where there is a
4499 common prefix, which gets split out into an EXACT like node
4500 preceding the TRIE node.
4502 If x(1..n)==tail then we can do a simple trie, if not we make
4503 a "jump" trie, such that when we match the appropriate word
4504 we "jump" to the appropriate tail node. Essentially we turn
4505 a nested if into a case structure of sorts.
4510 if (!re_trie_maxbuff) {
4511 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
4512 if (!SvIOK(re_trie_maxbuff))
4513 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
4515 if ( SvIV(re_trie_maxbuff)>=0 ) {
4517 regnode *first = (regnode *)NULL;
4518 regnode *last = (regnode *)NULL;
4519 regnode *tail = scan;
4523 /* var tail is used because there may be a TAIL
4524 regop in the way. Ie, the exacts will point to the
4525 thing following the TAIL, but the last branch will
4526 point at the TAIL. So we advance tail. If we
4527 have nested (?:) we may have to move through several
4531 while ( OP( tail ) == TAIL ) {
4532 /* this is the TAIL generated by (?:) */
4533 tail = regnext( tail );
4537 DEBUG_TRIE_COMPILE_r({
4538 regprop(RExC_rx, RExC_mysv, tail, NULL, pRExC_state);
4539 Perl_re_indentf( aTHX_ "%s %" UVuf ":%s\n",
4541 "Looking for TRIE'able sequences. Tail node is ",
4542 (UV)(tail - RExC_emit_start),
4543 SvPV_nolen_const( RExC_mysv )
4549 Step through the branches
4550 cur represents each branch,
4551 noper is the first thing to be matched as part
4553 noper_next is the regnext() of that node.
4555 We normally handle a case like this
4556 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
4557 support building with NOJUMPTRIE, which restricts
4558 the trie logic to structures like /FOO|BAR/.
4560 If noper is a trieable nodetype then the branch is
4561 a possible optimization target. If we are building
4562 under NOJUMPTRIE then we require that noper_next is
4563 the same as scan (our current position in the regex
4566 Once we have two or more consecutive such branches
4567 we can create a trie of the EXACT's contents and
4568 stitch it in place into the program.
4570 If the sequence represents all of the branches in
4571 the alternation we replace the entire thing with a
4574 Otherwise when it is a subsequence we need to
4575 stitch it in place and replace only the relevant
4576 branches. This means the first branch has to remain
4577 as it is used by the alternation logic, and its
4578 next pointer, and needs to be repointed at the item
4579 on the branch chain following the last branch we
4580 have optimized away.
4582 This could be either a BRANCH, in which case the
4583 subsequence is internal, or it could be the item
4584 following the branch sequence in which case the
4585 subsequence is at the end (which does not
4586 necessarily mean the first node is the start of the
4589 TRIE_TYPE(X) is a define which maps the optype to a
4593 ----------------+-----------
4597 EXACTFU_SS | EXACTFU
4600 EXACTFLU8 | EXACTFLU8
4604 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) \
4606 : ( EXACT == (X) ) \
4608 : ( EXACTFU == (X) || EXACTFU_SS == (X) ) \
4610 : ( EXACTFAA == (X) ) \
4612 : ( EXACTL == (X) ) \
4614 : ( EXACTFLU8 == (X) ) \
4618 /* dont use tail as the end marker for this traverse */
4619 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
4620 regnode * const noper = NEXTOPER( cur );
4621 U8 noper_type = OP( noper );
4622 U8 noper_trietype = TRIE_TYPE( noper_type );
4623 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
4624 regnode * const noper_next = regnext( noper );
4625 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4626 U8 noper_next_trietype = (noper_next && noper_next < tail) ? TRIE_TYPE( noper_next_type ) :0;
4629 DEBUG_TRIE_COMPILE_r({
4630 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4631 Perl_re_indentf( aTHX_ "- %d:%s (%d)",
4633 REG_NODE_NUM(cur), SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur) );
4635 regprop(RExC_rx, RExC_mysv, noper, NULL, pRExC_state);
4636 Perl_re_printf( aTHX_ " -> %d:%s",
4637 REG_NODE_NUM(noper), SvPV_nolen_const(RExC_mysv));
4640 regprop(RExC_rx, RExC_mysv, noper_next, NULL, pRExC_state);
4641 Perl_re_printf( aTHX_ "\t=> %d:%s\t",
4642 REG_NODE_NUM(noper_next), SvPV_nolen_const(RExC_mysv));
4644 Perl_re_printf( aTHX_ "(First==%d,Last==%d,Cur==%d,tt==%s,ntt==%s,nntt==%s)\n",
4645 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
4646 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
4650 /* Is noper a trieable nodetype that can be merged
4651 * with the current trie (if there is one)? */
4655 ( noper_trietype == NOTHING )
4656 || ( trietype == NOTHING )
4657 || ( trietype == noper_trietype )
4660 && noper_next >= tail
4664 /* Handle mergable triable node Either we are
4665 * the first node in a new trieable sequence,
4666 * in which case we do some bookkeeping,
4667 * otherwise we update the end pointer. */
4670 if ( noper_trietype == NOTHING ) {
4671 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
4672 regnode * const noper_next = regnext( noper );
4673 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4674 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
4677 if ( noper_next_trietype ) {
4678 trietype = noper_next_trietype;
4679 } else if (noper_next_type) {
4680 /* a NOTHING regop is 1 regop wide.
4681 * We need at least two for a trie
4682 * so we can't merge this in */
4686 trietype = noper_trietype;
4689 if ( trietype == NOTHING )
4690 trietype = noper_trietype;
4695 } /* end handle mergable triable node */
4697 /* handle unmergable node -
4698 * noper may either be a triable node which can
4699 * not be tried together with the current trie,
4700 * or a non triable node */
4702 /* If last is set and trietype is not
4703 * NOTHING then we have found at least two
4704 * triable branch sequences in a row of a
4705 * similar trietype so we can turn them
4706 * into a trie. If/when we allow NOTHING to
4707 * start a trie sequence this condition
4708 * will be required, and it isn't expensive
4709 * so we leave it in for now. */
4710 if ( trietype && trietype != NOTHING )
4711 make_trie( pRExC_state,
4712 startbranch, first, cur, tail,
4713 count, trietype, depth+1 );
4714 last = NULL; /* note: we clear/update
4715 first, trietype etc below,
4716 so we dont do it here */
4720 && noper_next >= tail
4723 /* noper is triable, so we can start a new
4727 trietype = noper_trietype;
4729 /* if we already saw a first but the
4730 * current node is not triable then we have
4731 * to reset the first information. */
4736 } /* end handle unmergable node */
4737 } /* loop over branches */
4738 DEBUG_TRIE_COMPILE_r({
4739 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4740 Perl_re_indentf( aTHX_ "- %s (%d) <SCAN FINISHED> ",
4741 depth+1, SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4742 Perl_re_printf( aTHX_ "(First==%d, Last==%d, Cur==%d, tt==%s)\n",
4743 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
4744 PL_reg_name[trietype]
4748 if ( last && trietype ) {
4749 if ( trietype != NOTHING ) {
4750 /* the last branch of the sequence was part of
4751 * a trie, so we have to construct it here
4752 * outside of the loop */
4753 made= make_trie( pRExC_state, startbranch,
4754 first, scan, tail, count,
4755 trietype, depth+1 );
4756 #ifdef TRIE_STUDY_OPT
4757 if ( ((made == MADE_EXACT_TRIE &&
4758 startbranch == first)
4759 || ( first_non_open == first )) &&
4761 flags |= SCF_TRIE_RESTUDY;
4762 if ( startbranch == first
4765 RExC_seen &=~REG_TOP_LEVEL_BRANCHES_SEEN;
4770 /* at this point we know whatever we have is a
4771 * NOTHING sequence/branch AND if 'startbranch'
4772 * is 'first' then we can turn the whole thing
4775 if ( startbranch == first ) {
4777 /* the entire thing is a NOTHING sequence,
4778 * something like this: (?:|) So we can
4779 * turn it into a plain NOTHING op. */
4780 DEBUG_TRIE_COMPILE_r({
4781 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4782 Perl_re_indentf( aTHX_ "- %s (%d) <NOTHING BRANCH SEQUENCE>\n",
4784 SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4787 OP(startbranch)= NOTHING;
4788 NEXT_OFF(startbranch)= tail - startbranch;
4789 for ( opt= startbranch + 1; opt < tail ; opt++ )
4793 } /* end if ( last) */
4794 } /* TRIE_MAXBUF is non zero */
4799 else if ( code == BRANCHJ ) { /* single branch is optimized. */
4800 scan = NEXTOPER(NEXTOPER(scan));
4801 } else /* single branch is optimized. */
4802 scan = NEXTOPER(scan);
4804 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB) {
4806 regnode *start = NULL;
4807 regnode *end = NULL;
4808 U32 my_recursed_depth= recursed_depth;
4810 if (OP(scan) != SUSPEND) { /* GOSUB */
4811 /* Do setup, note this code has side effects beyond
4812 * the rest of this block. Specifically setting
4813 * RExC_recurse[] must happen at least once during
4816 RExC_recurse[ARG2L(scan)] = scan;
4817 start = RExC_open_parens[paren];
4818 end = RExC_close_parens[paren];
4820 /* NOTE we MUST always execute the above code, even
4821 * if we do nothing with a GOSUB */
4823 ( flags & SCF_IN_DEFINE )
4826 (is_inf_internal || is_inf || (data && data->flags & SF_IS_INF))
4828 ( (flags & (SCF_DO_STCLASS | SCF_DO_SUBSTR)) == 0 )
4831 /* no need to do anything here if we are in a define. */
4832 /* or we are after some kind of infinite construct
4833 * so we can skip recursing into this item.
4834 * Since it is infinite we will not change the maxlen
4835 * or delta, and if we miss something that might raise
4836 * the minlen it will merely pessimise a little.
4838 * Iow /(?(DEFINE)(?<foo>foo|food))a+(?&foo)/
4839 * might result in a minlen of 1 and not of 4,
4840 * but this doesn't make us mismatch, just try a bit
4841 * harder than we should.
4843 scan= regnext(scan);
4850 !PAREN_TEST(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes), paren)
4852 /* it is quite possible that there are more efficient ways
4853 * to do this. We maintain a bitmap per level of recursion
4854 * of which patterns we have entered so we can detect if a
4855 * pattern creates a possible infinite loop. When we
4856 * recurse down a level we copy the previous levels bitmap
4857 * down. When we are at recursion level 0 we zero the top
4858 * level bitmap. It would be nice to implement a different
4859 * more efficient way of doing this. In particular the top
4860 * level bitmap may be unnecessary.
4862 if (!recursed_depth) {
4863 Zero(RExC_study_chunk_recursed, RExC_study_chunk_recursed_bytes, U8);
4865 Copy(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes),
4866 RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes),
4867 RExC_study_chunk_recursed_bytes, U8);
4869 /* we havent recursed into this paren yet, so recurse into it */
4870 DEBUG_STUDYDATA("gosub-set", data, depth, is_inf);
4871 PAREN_SET(RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes), paren);
4872 my_recursed_depth= recursed_depth + 1;
4874 DEBUG_STUDYDATA("gosub-inf", data, depth, is_inf);
4875 /* some form of infinite recursion, assume infinite length
4877 if (flags & SCF_DO_SUBSTR) {
4878 scan_commit(pRExC_state, data, minlenp, is_inf);
4879 data->cur_is_floating = 1;
4881 is_inf = is_inf_internal = 1;
4882 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4883 ssc_anything(data->start_class);
4884 flags &= ~SCF_DO_STCLASS;
4886 start= NULL; /* reset start so we dont recurse later on. */
4891 end = regnext(scan);
4894 scan_frame *newframe;
4896 if (!RExC_frame_last) {
4897 Newxz(newframe, 1, scan_frame);
4898 SAVEDESTRUCTOR_X(S_unwind_scan_frames, newframe);
4899 RExC_frame_head= newframe;
4901 } else if (!RExC_frame_last->next_frame) {
4902 Newxz(newframe,1,scan_frame);
4903 RExC_frame_last->next_frame= newframe;
4904 newframe->prev_frame= RExC_frame_last;
4907 newframe= RExC_frame_last->next_frame;
4909 RExC_frame_last= newframe;
4911 newframe->next_regnode = regnext(scan);
4912 newframe->last_regnode = last;
4913 newframe->stopparen = stopparen;
4914 newframe->prev_recursed_depth = recursed_depth;
4915 newframe->this_prev_frame= frame;
4917 DEBUG_STUDYDATA("frame-new", data, depth, is_inf);
4918 DEBUG_PEEP("fnew", scan, depth, flags);
4925 recursed_depth= my_recursed_depth;
4930 else if (OP(scan) == EXACT || OP(scan) == EXACTL) {
4931 SSize_t l = STR_LEN(scan);
4935 const U8 * const s = (U8*)STRING(scan);
4936 uc = utf8_to_uvchr_buf(s, s + l, NULL);
4937 l = utf8_length(s, s + l);
4939 uc = *((U8*)STRING(scan));
4942 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
4943 /* The code below prefers earlier match for fixed
4944 offset, later match for variable offset. */
4945 if (data->last_end == -1) { /* Update the start info. */
4946 data->last_start_min = data->pos_min;
4947 data->last_start_max = is_inf
4948 ? SSize_t_MAX : data->pos_min + data->pos_delta;
4950 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
4952 SvUTF8_on(data->last_found);
4954 SV * const sv = data->last_found;
4955 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4956 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4957 if (mg && mg->mg_len >= 0)
4958 mg->mg_len += utf8_length((U8*)STRING(scan),
4959 (U8*)STRING(scan)+STR_LEN(scan));
4961 data->last_end = data->pos_min + l;
4962 data->pos_min += l; /* As in the first entry. */
4963 data->flags &= ~SF_BEFORE_EOL;
4966 /* ANDing the code point leaves at most it, and not in locale, and
4967 * can't match null string */
4968 if (flags & SCF_DO_STCLASS_AND) {
4969 ssc_cp_and(data->start_class, uc);
4970 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4971 ssc_clear_locale(data->start_class);
4973 else if (flags & SCF_DO_STCLASS_OR) {
4974 ssc_add_cp(data->start_class, uc);
4975 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4977 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4978 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4980 flags &= ~SCF_DO_STCLASS;
4982 else if (PL_regkind[OP(scan)] == EXACT) {
4983 /* But OP != EXACT!, so is EXACTFish */
4984 SSize_t l = STR_LEN(scan);
4985 const U8 * s = (U8*)STRING(scan);
4987 /* Search for fixed substrings supports EXACT only. */
4988 if (flags & SCF_DO_SUBSTR) {
4990 scan_commit(pRExC_state, data, minlenp, is_inf);
4993 l = utf8_length(s, s + l);
4995 if (unfolded_multi_char) {
4996 RExC_seen |= REG_UNFOLDED_MULTI_SEEN;
4998 min += l - min_subtract;
5000 delta += min_subtract;
5001 if (flags & SCF_DO_SUBSTR) {
5002 data->pos_min += l - min_subtract;
5003 if (data->pos_min < 0) {
5006 data->pos_delta += min_subtract;
5008 data->cur_is_floating = 1; /* float */
5012 if (flags & SCF_DO_STCLASS) {
5013 SV* EXACTF_invlist = _make_exactf_invlist(pRExC_state, scan);
5015 assert(EXACTF_invlist);
5016 if (flags & SCF_DO_STCLASS_AND) {
5017 if (OP(scan) != EXACTFL)
5018 ssc_clear_locale(data->start_class);
5019 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5020 ANYOF_POSIXL_ZERO(data->start_class);
5021 ssc_intersection(data->start_class, EXACTF_invlist, FALSE);
5023 else { /* SCF_DO_STCLASS_OR */
5024 ssc_union(data->start_class, EXACTF_invlist, FALSE);
5025 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5027 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5028 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5030 flags &= ~SCF_DO_STCLASS;
5031 SvREFCNT_dec(EXACTF_invlist);
5034 else if (REGNODE_VARIES(OP(scan))) {
5035 SSize_t mincount, maxcount, minnext, deltanext, pos_before = 0;
5036 I32 fl = 0, f = flags;
5037 regnode * const oscan = scan;
5038 regnode_ssc this_class;
5039 regnode_ssc *oclass = NULL;
5040 I32 next_is_eval = 0;
5042 switch (PL_regkind[OP(scan)]) {
5043 case WHILEM: /* End of (?:...)* . */
5044 scan = NEXTOPER(scan);
5047 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
5048 next = NEXTOPER(scan);
5049 if (OP(next) == EXACT
5050 || OP(next) == EXACTL
5051 || (flags & SCF_DO_STCLASS))
5054 maxcount = REG_INFTY;
5055 next = regnext(scan);
5056 scan = NEXTOPER(scan);
5060 if (flags & SCF_DO_SUBSTR)
5065 if (flags & SCF_DO_STCLASS) {
5067 maxcount = REG_INFTY;
5068 next = regnext(scan);
5069 scan = NEXTOPER(scan);
5072 if (flags & SCF_DO_SUBSTR) {
5073 scan_commit(pRExC_state, data, minlenp, is_inf);
5074 /* Cannot extend fixed substrings */
5075 data->cur_is_floating = 1; /* float */
5077 is_inf = is_inf_internal = 1;
5078 scan = regnext(scan);
5079 goto optimize_curly_tail;
5081 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
5082 && (scan->flags == stopparen))
5087 mincount = ARG1(scan);
5088 maxcount = ARG2(scan);
5090 next = regnext(scan);
5091 if (OP(scan) == CURLYX) {
5092 I32 lp = (data ? *(data->last_closep) : 0);
5093 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
5095 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
5096 next_is_eval = (OP(scan) == EVAL);
5098 if (flags & SCF_DO_SUBSTR) {
5100 scan_commit(pRExC_state, data, minlenp, is_inf);
5101 /* Cannot extend fixed substrings */
5102 pos_before = data->pos_min;
5106 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
5108 data->flags |= SF_IS_INF;
5110 if (flags & SCF_DO_STCLASS) {
5111 ssc_init(pRExC_state, &this_class);
5112 oclass = data->start_class;
5113 data->start_class = &this_class;
5114 f |= SCF_DO_STCLASS_AND;
5115 f &= ~SCF_DO_STCLASS_OR;
5117 /* Exclude from super-linear cache processing any {n,m}
5118 regops for which the combination of input pos and regex
5119 pos is not enough information to determine if a match
5122 For example, in the regex /foo(bar\s*){4,8}baz/ with the
5123 regex pos at the \s*, the prospects for a match depend not
5124 only on the input position but also on how many (bar\s*)
5125 repeats into the {4,8} we are. */
5126 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
5127 f &= ~SCF_WHILEM_VISITED_POS;
5129 /* This will finish on WHILEM, setting scan, or on NULL: */
5130 /* recurse study_chunk() on loop bodies */
5131 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
5132 last, data, stopparen, recursed_depth, NULL,
5134 ? (f & ~SCF_DO_SUBSTR)
5138 if (flags & SCF_DO_STCLASS)
5139 data->start_class = oclass;
5140 if (mincount == 0 || minnext == 0) {
5141 if (flags & SCF_DO_STCLASS_OR) {
5142 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5144 else if (flags & SCF_DO_STCLASS_AND) {
5145 /* Switch to OR mode: cache the old value of
5146 * data->start_class */
5148 StructCopy(data->start_class, and_withp, regnode_ssc);
5149 flags &= ~SCF_DO_STCLASS_AND;
5150 StructCopy(&this_class, data->start_class, regnode_ssc);
5151 flags |= SCF_DO_STCLASS_OR;
5152 ANYOF_FLAGS(data->start_class)
5153 |= SSC_MATCHES_EMPTY_STRING;
5155 } else { /* Non-zero len */
5156 if (flags & SCF_DO_STCLASS_OR) {
5157 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5158 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5160 else if (flags & SCF_DO_STCLASS_AND)
5161 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5162 flags &= ~SCF_DO_STCLASS;
5164 if (!scan) /* It was not CURLYX, but CURLY. */
5166 if (((flags & (SCF_TRIE_DOING_RESTUDY|SCF_DO_SUBSTR))==SCF_DO_SUBSTR)
5167 /* ? quantifier ok, except for (?{ ... }) */
5168 && (next_is_eval || !(mincount == 0 && maxcount == 1))
5169 && (minnext == 0) && (deltanext == 0)
5170 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
5171 && maxcount <= REG_INFTY/3) /* Complement check for big
5174 /* Fatal warnings may leak the regexp without this: */
5175 SAVEFREESV(RExC_rx_sv);
5176 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP),
5177 "Quantifier unexpected on zero-length expression "
5178 "in regex m/%" UTF8f "/",
5179 UTF8fARG(UTF, RExC_precomp_end - RExC_precomp,
5181 (void)ReREFCNT_inc(RExC_rx_sv);
5184 min += minnext * mincount;
5185 is_inf_internal |= deltanext == SSize_t_MAX
5186 || (maxcount == REG_INFTY && minnext + deltanext > 0);
5187 is_inf |= is_inf_internal;
5189 delta = SSize_t_MAX;
5191 delta += (minnext + deltanext) * maxcount
5192 - minnext * mincount;
5194 /* Try powerful optimization CURLYX => CURLYN. */
5195 if ( OP(oscan) == CURLYX && data
5196 && data->flags & SF_IN_PAR
5197 && !(data->flags & SF_HAS_EVAL)
5198 && !deltanext && minnext == 1 ) {
5199 /* Try to optimize to CURLYN. */
5200 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
5201 regnode * const nxt1 = nxt;
5208 if (!REGNODE_SIMPLE(OP(nxt))
5209 && !(PL_regkind[OP(nxt)] == EXACT
5210 && STR_LEN(nxt) == 1))
5216 if (OP(nxt) != CLOSE)
5218 if (RExC_open_parens) {
5219 RExC_open_parens[ARG(nxt1)]=oscan; /*open->CURLYM*/
5220 RExC_close_parens[ARG(nxt1)]=nxt+2; /*close->while*/
5222 /* Now we know that nxt2 is the only contents: */
5223 oscan->flags = (U8)ARG(nxt);
5225 OP(nxt1) = NOTHING; /* was OPEN. */
5228 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5229 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
5230 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
5231 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5232 OP(nxt + 1) = OPTIMIZED; /* was count. */
5233 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
5238 /* Try optimization CURLYX => CURLYM. */
5239 if ( OP(oscan) == CURLYX && data
5240 && !(data->flags & SF_HAS_PAR)
5241 && !(data->flags & SF_HAS_EVAL)
5242 && !deltanext /* atom is fixed width */
5243 && minnext != 0 /* CURLYM can't handle zero width */
5245 /* Nor characters whose fold at run-time may be
5246 * multi-character */
5247 && ! (RExC_seen & REG_UNFOLDED_MULTI_SEEN)
5249 /* XXXX How to optimize if data == 0? */
5250 /* Optimize to a simpler form. */
5251 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
5255 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
5256 && (OP(nxt2) != WHILEM))
5258 OP(nxt2) = SUCCEED; /* Whas WHILEM */
5259 /* Need to optimize away parenths. */
5260 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
5261 /* Set the parenth number. */
5262 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
5264 oscan->flags = (U8)ARG(nxt);
5265 if (RExC_open_parens) {
5266 RExC_open_parens[ARG(nxt1)]=oscan; /*open->CURLYM*/
5267 RExC_close_parens[ARG(nxt1)]=nxt2+1; /*close->NOTHING*/
5269 OP(nxt1) = OPTIMIZED; /* was OPEN. */
5270 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5273 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5274 OP(nxt + 1) = OPTIMIZED; /* was count. */
5275 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
5276 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
5279 while ( nxt1 && (OP(nxt1) != WHILEM)) {
5280 regnode *nnxt = regnext(nxt1);
5282 if (reg_off_by_arg[OP(nxt1)])
5283 ARG_SET(nxt1, nxt2 - nxt1);
5284 else if (nxt2 - nxt1 < U16_MAX)
5285 NEXT_OFF(nxt1) = nxt2 - nxt1;
5287 OP(nxt) = NOTHING; /* Cannot beautify */
5292 /* Optimize again: */
5293 /* recurse study_chunk() on optimised CURLYX => CURLYM */
5294 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
5295 NULL, stopparen, recursed_depth, NULL, 0,depth+1);
5300 else if ((OP(oscan) == CURLYX)
5301 && (flags & SCF_WHILEM_VISITED_POS)
5302 /* See the comment on a similar expression above.
5303 However, this time it's not a subexpression
5304 we care about, but the expression itself. */
5305 && (maxcount == REG_INFTY)
5307 /* This stays as CURLYX, we can put the count/of pair. */
5308 /* Find WHILEM (as in regexec.c) */
5309 regnode *nxt = oscan + NEXT_OFF(oscan);
5311 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
5313 nxt = PREVOPER(nxt);
5314 if (nxt->flags & 0xf) {
5315 /* we've already set whilem count on this node */
5316 } else if (++data->whilem_c < 16) {
5317 assert(data->whilem_c <= RExC_whilem_seen);
5318 nxt->flags = (U8)(data->whilem_c
5319 | (RExC_whilem_seen << 4)); /* On WHILEM */
5322 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
5324 if (flags & SCF_DO_SUBSTR) {
5325 SV *last_str = NULL;
5326 STRLEN last_chrs = 0;
5327 int counted = mincount != 0;
5329 if (data->last_end > 0 && mincount != 0) { /* Ends with a
5331 SSize_t b = pos_before >= data->last_start_min
5332 ? pos_before : data->last_start_min;
5334 const char * const s = SvPV_const(data->last_found, l);
5335 SSize_t old = b - data->last_start_min;
5338 old = utf8_hop((U8*)s, old) - (U8*)s;
5340 /* Get the added string: */
5341 last_str = newSVpvn_utf8(s + old, l, UTF);
5342 last_chrs = UTF ? utf8_length((U8*)(s + old),
5343 (U8*)(s + old + l)) : l;
5344 if (deltanext == 0 && pos_before == b) {
5345 /* What was added is a constant string */
5348 SvGROW(last_str, (mincount * l) + 1);
5349 repeatcpy(SvPVX(last_str) + l,
5350 SvPVX_const(last_str), l,
5352 SvCUR_set(last_str, SvCUR(last_str) * mincount);
5353 /* Add additional parts. */
5354 SvCUR_set(data->last_found,
5355 SvCUR(data->last_found) - l);
5356 sv_catsv(data->last_found, last_str);
5358 SV * sv = data->last_found;
5360 SvUTF8(sv) && SvMAGICAL(sv) ?
5361 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5362 if (mg && mg->mg_len >= 0)
5363 mg->mg_len += last_chrs * (mincount-1);
5365 last_chrs *= mincount;
5366 data->last_end += l * (mincount - 1);
5369 /* start offset must point into the last copy */
5370 data->last_start_min += minnext * (mincount - 1);
5371 data->last_start_max =
5374 : data->last_start_max +
5375 (maxcount - 1) * (minnext + data->pos_delta);
5378 /* It is counted once already... */
5379 data->pos_min += minnext * (mincount - counted);
5381 Perl_re_printf( aTHX_ "counted=%" UVuf " deltanext=%" UVuf
5382 " SSize_t_MAX=%" UVuf " minnext=%" UVuf
5383 " maxcount=%" UVuf " mincount=%" UVuf "\n",
5384 (UV)counted, (UV)deltanext, (UV)SSize_t_MAX, (UV)minnext, (UV)maxcount,
5386 if (deltanext != SSize_t_MAX)
5387 Perl_re_printf( aTHX_ "LHS=%" UVuf " RHS=%" UVuf "\n",
5388 (UV)(-counted * deltanext + (minnext + deltanext) * maxcount
5389 - minnext * mincount), (UV)(SSize_t_MAX - data->pos_delta));
5391 if (deltanext == SSize_t_MAX
5392 || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= SSize_t_MAX - data->pos_delta)
5393 data->pos_delta = SSize_t_MAX;
5395 data->pos_delta += - counted * deltanext +
5396 (minnext + deltanext) * maxcount - minnext * mincount;
5397 if (mincount != maxcount) {
5398 /* Cannot extend fixed substrings found inside
5400 scan_commit(pRExC_state, data, minlenp, is_inf);
5401 if (mincount && last_str) {
5402 SV * const sv = data->last_found;
5403 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
5404 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5408 sv_setsv(sv, last_str);
5409 data->last_end = data->pos_min;
5410 data->last_start_min = data->pos_min - last_chrs;
5411 data->last_start_max = is_inf
5413 : data->pos_min + data->pos_delta - last_chrs;
5415 data->cur_is_floating = 1; /* float */
5417 SvREFCNT_dec(last_str);
5419 if (data && (fl & SF_HAS_EVAL))
5420 data->flags |= SF_HAS_EVAL;
5421 optimize_curly_tail:
5422 if (OP(oscan) != CURLYX) {
5423 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
5425 NEXT_OFF(oscan) += NEXT_OFF(next);
5431 Perl_croak(aTHX_ "panic: unexpected varying REx opcode %d",
5436 if (flags & SCF_DO_SUBSTR) {
5437 /* Cannot expect anything... */
5438 scan_commit(pRExC_state, data, minlenp, is_inf);
5439 data->cur_is_floating = 1; /* float */
5441 is_inf = is_inf_internal = 1;
5442 if (flags & SCF_DO_STCLASS_OR) {
5443 if (OP(scan) == CLUMP) {
5444 /* Actually is any start char, but very few code points
5445 * aren't start characters */
5446 ssc_match_all_cp(data->start_class);
5449 ssc_anything(data->start_class);
5452 flags &= ~SCF_DO_STCLASS;
5456 else if (OP(scan) == LNBREAK) {
5457 if (flags & SCF_DO_STCLASS) {
5458 if (flags & SCF_DO_STCLASS_AND) {
5459 ssc_intersection(data->start_class,
5460 PL_XPosix_ptrs[_CC_VERTSPACE], FALSE);
5461 ssc_clear_locale(data->start_class);
5462 ANYOF_FLAGS(data->start_class)
5463 &= ~SSC_MATCHES_EMPTY_STRING;
5465 else if (flags & SCF_DO_STCLASS_OR) {
5466 ssc_union(data->start_class,
5467 PL_XPosix_ptrs[_CC_VERTSPACE],
5469 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5471 /* See commit msg for
5472 * 749e076fceedeb708a624933726e7989f2302f6a */
5473 ANYOF_FLAGS(data->start_class)
5474 &= ~SSC_MATCHES_EMPTY_STRING;
5476 flags &= ~SCF_DO_STCLASS;
5479 if (delta != SSize_t_MAX)
5480 delta++; /* Because of the 2 char string cr-lf */
5481 if (flags & SCF_DO_SUBSTR) {
5482 /* Cannot expect anything... */
5483 scan_commit(pRExC_state, data, minlenp, is_inf);
5485 if (data->pos_delta != SSize_t_MAX) {
5486 data->pos_delta += 1;
5488 data->cur_is_floating = 1; /* float */
5491 else if (REGNODE_SIMPLE(OP(scan))) {
5493 if (flags & SCF_DO_SUBSTR) {
5494 scan_commit(pRExC_state, data, minlenp, is_inf);
5498 if (flags & SCF_DO_STCLASS) {
5500 SV* my_invlist = NULL;
5503 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5504 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5506 /* Some of the logic below assumes that switching
5507 locale on will only add false positives. */
5512 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d",
5516 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5517 ssc_match_all_cp(data->start_class);
5522 SV* REG_ANY_invlist = _new_invlist(2);
5523 REG_ANY_invlist = add_cp_to_invlist(REG_ANY_invlist,
5525 if (flags & SCF_DO_STCLASS_OR) {
5526 ssc_union(data->start_class,
5528 TRUE /* TRUE => invert, hence all but \n
5532 else if (flags & SCF_DO_STCLASS_AND) {
5533 ssc_intersection(data->start_class,
5535 TRUE /* TRUE => invert */
5537 ssc_clear_locale(data->start_class);
5539 SvREFCNT_dec_NN(REG_ANY_invlist);
5546 if (flags & SCF_DO_STCLASS_AND)
5547 ssc_and(pRExC_state, data->start_class,
5548 (regnode_charclass *) scan);
5550 ssc_or(pRExC_state, data->start_class,
5551 (regnode_charclass *) scan);
5556 SV* cp_list = get_ANYOFM_contents(scan);
5558 if (flags & SCF_DO_STCLASS_OR) {
5559 ssc_union(data->start_class,
5561 FALSE /* don't invert */
5564 else if (flags & SCF_DO_STCLASS_AND) {
5565 ssc_intersection(data->start_class,
5567 FALSE /* don't invert */
5571 SvREFCNT_dec_NN(cp_list);
5580 namedclass = classnum_to_namedclass(FLAGS(scan)) + invert;
5581 if (flags & SCF_DO_STCLASS_AND) {
5582 bool was_there = cBOOL(
5583 ANYOF_POSIXL_TEST(data->start_class,
5585 ANYOF_POSIXL_ZERO(data->start_class);
5586 if (was_there) { /* Do an AND */
5587 ANYOF_POSIXL_SET(data->start_class, namedclass);
5589 /* No individual code points can now match */
5590 data->start_class->invlist
5591 = sv_2mortal(_new_invlist(0));
5594 int complement = namedclass + ((invert) ? -1 : 1);
5596 assert(flags & SCF_DO_STCLASS_OR);
5598 /* If the complement of this class was already there,
5599 * the result is that they match all code points,
5600 * (\d + \D == everything). Remove the classes from
5601 * future consideration. Locale is not relevant in
5603 if (ANYOF_POSIXL_TEST(data->start_class, complement)) {
5604 ssc_match_all_cp(data->start_class);
5605 ANYOF_POSIXL_CLEAR(data->start_class, namedclass);
5606 ANYOF_POSIXL_CLEAR(data->start_class, complement);
5608 else { /* The usual case; just add this class to the
5610 ANYOF_POSIXL_SET(data->start_class, namedclass);
5619 my_invlist = invlist_clone(PL_Posix_ptrs[_CC_ASCII]);
5621 /* This can be handled as a Posix class */
5622 goto join_posix_and_ascii;
5624 case NPOSIXA: /* For these, we always know the exact set of
5629 assert(FLAGS(scan) != _CC_ASCII);
5630 my_invlist = invlist_clone(PL_Posix_ptrs[FLAGS(scan)]);
5631 goto join_posix_and_ascii;
5639 my_invlist = invlist_clone(PL_XPosix_ptrs[FLAGS(scan)]);
5641 /* NPOSIXD matches all upper Latin1 code points unless the
5642 * target string being matched is UTF-8, which is
5643 * unknowable until match time. Since we are going to
5644 * invert, we want to get rid of all of them so that the
5645 * inversion will match all */
5646 if (OP(scan) == NPOSIXD) {
5647 _invlist_subtract(my_invlist, PL_UpperLatin1,
5651 join_posix_and_ascii:
5653 if (flags & SCF_DO_STCLASS_AND) {
5654 ssc_intersection(data->start_class, my_invlist, invert);
5655 ssc_clear_locale(data->start_class);
5658 assert(flags & SCF_DO_STCLASS_OR);
5659 ssc_union(data->start_class, my_invlist, invert);
5661 SvREFCNT_dec(my_invlist);
5663 if (flags & SCF_DO_STCLASS_OR)
5664 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5665 flags &= ~SCF_DO_STCLASS;
5668 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
5669 data->flags |= (OP(scan) == MEOL
5672 scan_commit(pRExC_state, data, minlenp, is_inf);
5675 else if ( PL_regkind[OP(scan)] == BRANCHJ
5676 /* Lookbehind, or need to calculate parens/evals/stclass: */
5677 && (scan->flags || data || (flags & SCF_DO_STCLASS))
5678 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM))
5680 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5681 || OP(scan) == UNLESSM )
5683 /* Negative Lookahead/lookbehind
5684 In this case we can't do fixed string optimisation.
5687 SSize_t deltanext, minnext, fake = 0;
5692 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5694 data_fake.whilem_c = data->whilem_c;
5695 data_fake.last_closep = data->last_closep;
5698 data_fake.last_closep = &fake;
5699 data_fake.pos_delta = delta;
5700 if ( flags & SCF_DO_STCLASS && !scan->flags
5701 && OP(scan) == IFMATCH ) { /* Lookahead */
5702 ssc_init(pRExC_state, &intrnl);
5703 data_fake.start_class = &intrnl;
5704 f |= SCF_DO_STCLASS_AND;
5706 if (flags & SCF_WHILEM_VISITED_POS)
5707 f |= SCF_WHILEM_VISITED_POS;
5708 next = regnext(scan);
5709 nscan = NEXTOPER(NEXTOPER(scan));
5711 /* recurse study_chunk() for lookahead body */
5712 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
5713 last, &data_fake, stopparen,
5714 recursed_depth, NULL, f, depth+1);
5717 FAIL("Variable length lookbehind not implemented");
5719 else if (minnext > (I32)U8_MAX) {
5720 FAIL2("Lookbehind longer than %" UVuf " not implemented",
5723 scan->flags = (U8)minnext;
5726 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5728 if (data_fake.flags & SF_HAS_EVAL)
5729 data->flags |= SF_HAS_EVAL;
5730 data->whilem_c = data_fake.whilem_c;
5732 if (f & SCF_DO_STCLASS_AND) {
5733 if (flags & SCF_DO_STCLASS_OR) {
5734 /* OR before, AND after: ideally we would recurse with
5735 * data_fake to get the AND applied by study of the
5736 * remainder of the pattern, and then derecurse;
5737 * *** HACK *** for now just treat as "no information".
5738 * See [perl #56690].
5740 ssc_init(pRExC_state, data->start_class);
5742 /* AND before and after: combine and continue. These
5743 * assertions are zero-length, so can match an EMPTY
5745 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5746 ANYOF_FLAGS(data->start_class)
5747 |= SSC_MATCHES_EMPTY_STRING;
5751 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5753 /* Positive Lookahead/lookbehind
5754 In this case we can do fixed string optimisation,
5755 but we must be careful about it. Note in the case of
5756 lookbehind the positions will be offset by the minimum
5757 length of the pattern, something we won't know about
5758 until after the recurse.
5760 SSize_t deltanext, fake = 0;
5764 /* We use SAVEFREEPV so that when the full compile
5765 is finished perl will clean up the allocated
5766 minlens when it's all done. This way we don't
5767 have to worry about freeing them when we know
5768 they wont be used, which would be a pain.
5771 Newx( minnextp, 1, SSize_t );
5772 SAVEFREEPV(minnextp);
5775 StructCopy(data, &data_fake, scan_data_t);
5776 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
5779 scan_commit(pRExC_state, &data_fake, minlenp, is_inf);
5780 data_fake.last_found=newSVsv(data->last_found);
5784 data_fake.last_closep = &fake;
5785 data_fake.flags = 0;
5786 data_fake.substrs[0].flags = 0;
5787 data_fake.substrs[1].flags = 0;
5788 data_fake.pos_delta = delta;
5790 data_fake.flags |= SF_IS_INF;
5791 if ( flags & SCF_DO_STCLASS && !scan->flags
5792 && OP(scan) == IFMATCH ) { /* Lookahead */
5793 ssc_init(pRExC_state, &intrnl);
5794 data_fake.start_class = &intrnl;
5795 f |= SCF_DO_STCLASS_AND;
5797 if (flags & SCF_WHILEM_VISITED_POS)
5798 f |= SCF_WHILEM_VISITED_POS;
5799 next = regnext(scan);
5800 nscan = NEXTOPER(NEXTOPER(scan));
5802 /* positive lookahead study_chunk() recursion */
5803 *minnextp = study_chunk(pRExC_state, &nscan, minnextp,
5804 &deltanext, last, &data_fake,
5805 stopparen, recursed_depth, NULL,
5809 FAIL("Variable length lookbehind not implemented");
5811 else if (*minnextp > (I32)U8_MAX) {
5812 FAIL2("Lookbehind longer than %" UVuf " not implemented",
5815 scan->flags = (U8)*minnextp;
5820 if (f & SCF_DO_STCLASS_AND) {
5821 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5822 ANYOF_FLAGS(data->start_class) |= SSC_MATCHES_EMPTY_STRING;
5825 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5827 if (data_fake.flags & SF_HAS_EVAL)
5828 data->flags |= SF_HAS_EVAL;
5829 data->whilem_c = data_fake.whilem_c;
5830 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
5832 if (RExC_rx->minlen<*minnextp)
5833 RExC_rx->minlen=*minnextp;
5834 scan_commit(pRExC_state, &data_fake, minnextp, is_inf);
5835 SvREFCNT_dec_NN(data_fake.last_found);
5837 for (i = 0; i < 2; i++) {
5838 if (data_fake.substrs[i].minlenp != minlenp) {
5839 data->substrs[i].min_offset =
5840 data_fake.substrs[i].min_offset;
5841 data->substrs[i].max_offset =
5842 data_fake.substrs[i].max_offset;
5843 data->substrs[i].minlenp =
5844 data_fake.substrs[i].minlenp;
5845 data->substrs[i].lookbehind += scan->flags;
5854 else if (OP(scan) == OPEN) {
5855 if (stopparen != (I32)ARG(scan))
5858 else if (OP(scan) == CLOSE) {
5859 if (stopparen == (I32)ARG(scan)) {
5862 if ((I32)ARG(scan) == is_par) {
5863 next = regnext(scan);
5865 if ( next && (OP(next) != WHILEM) && next < last)
5866 is_par = 0; /* Disable optimization */
5869 *(data->last_closep) = ARG(scan);
5871 else if (OP(scan) == EVAL) {
5873 data->flags |= SF_HAS_EVAL;
5875 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
5876 if (flags & SCF_DO_SUBSTR) {
5877 scan_commit(pRExC_state, data, minlenp, is_inf);
5878 flags &= ~SCF_DO_SUBSTR;
5880 if (data && OP(scan)==ACCEPT) {
5881 data->flags |= SCF_SEEN_ACCEPT;
5886 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
5888 if (flags & SCF_DO_SUBSTR) {
5889 scan_commit(pRExC_state, data, minlenp, is_inf);
5890 data->cur_is_floating = 1; /* float */
5892 is_inf = is_inf_internal = 1;
5893 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5894 ssc_anything(data->start_class);
5895 flags &= ~SCF_DO_STCLASS;
5897 else if (OP(scan) == GPOS) {
5898 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
5899 !(delta || is_inf || (data && data->pos_delta)))
5901 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
5902 RExC_rx->intflags |= PREGf_ANCH_GPOS;
5903 if (RExC_rx->gofs < (STRLEN)min)
5904 RExC_rx->gofs = min;
5906 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
5910 #ifdef TRIE_STUDY_OPT
5911 #ifdef FULL_TRIE_STUDY
5912 else if (PL_regkind[OP(scan)] == TRIE) {
5913 /* NOTE - There is similar code to this block above for handling
5914 BRANCH nodes on the initial study. If you change stuff here
5916 regnode *trie_node= scan;
5917 regnode *tail= regnext(scan);
5918 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5919 SSize_t max1 = 0, min1 = SSize_t_MAX;
5922 if (flags & SCF_DO_SUBSTR) { /* XXXX Add !SUSPEND? */
5923 /* Cannot merge strings after this. */
5924 scan_commit(pRExC_state, data, minlenp, is_inf);
5926 if (flags & SCF_DO_STCLASS)
5927 ssc_init_zero(pRExC_state, &accum);
5933 const regnode *nextbranch= NULL;
5936 for ( word=1 ; word <= trie->wordcount ; word++)
5938 SSize_t deltanext=0, minnext=0, f = 0, fake;
5939 regnode_ssc this_class;
5941 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5943 data_fake.whilem_c = data->whilem_c;
5944 data_fake.last_closep = data->last_closep;
5947 data_fake.last_closep = &fake;
5948 data_fake.pos_delta = delta;
5949 if (flags & SCF_DO_STCLASS) {
5950 ssc_init(pRExC_state, &this_class);
5951 data_fake.start_class = &this_class;
5952 f = SCF_DO_STCLASS_AND;
5954 if (flags & SCF_WHILEM_VISITED_POS)
5955 f |= SCF_WHILEM_VISITED_POS;
5957 if (trie->jump[word]) {
5959 nextbranch = trie_node + trie->jump[0];
5960 scan= trie_node + trie->jump[word];
5961 /* We go from the jump point to the branch that follows
5962 it. Note this means we need the vestigal unused
5963 branches even though they arent otherwise used. */
5964 /* optimise study_chunk() for TRIE */
5965 minnext = study_chunk(pRExC_state, &scan, minlenp,
5966 &deltanext, (regnode *)nextbranch, &data_fake,
5967 stopparen, recursed_depth, NULL, f,depth+1);
5969 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
5970 nextbranch= regnext((regnode*)nextbranch);
5972 if (min1 > (SSize_t)(minnext + trie->minlen))
5973 min1 = minnext + trie->minlen;
5974 if (deltanext == SSize_t_MAX) {
5975 is_inf = is_inf_internal = 1;
5977 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
5978 max1 = minnext + deltanext + trie->maxlen;
5980 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5982 if (data_fake.flags & SCF_SEEN_ACCEPT) {
5983 if ( stopmin > min + min1)
5984 stopmin = min + min1;
5985 flags &= ~SCF_DO_SUBSTR;
5987 data->flags |= SCF_SEEN_ACCEPT;
5990 if (data_fake.flags & SF_HAS_EVAL)
5991 data->flags |= SF_HAS_EVAL;
5992 data->whilem_c = data_fake.whilem_c;
5994 if (flags & SCF_DO_STCLASS)
5995 ssc_or(pRExC_state, &accum, (regnode_charclass *) &this_class);
5998 if (flags & SCF_DO_SUBSTR) {
5999 data->pos_min += min1;
6000 data->pos_delta += max1 - min1;
6001 if (max1 != min1 || is_inf)
6002 data->cur_is_floating = 1; /* float */
6005 if (delta != SSize_t_MAX) {
6006 if (SSize_t_MAX - (max1 - min1) >= delta)
6007 delta += max1 - min1;
6009 delta = SSize_t_MAX;
6011 if (flags & SCF_DO_STCLASS_OR) {
6012 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &accum);
6014 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6015 flags &= ~SCF_DO_STCLASS;
6018 else if (flags & SCF_DO_STCLASS_AND) {
6020 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
6021 flags &= ~SCF_DO_STCLASS;
6024 /* Switch to OR mode: cache the old value of
6025 * data->start_class */
6027 StructCopy(data->start_class, and_withp, regnode_ssc);
6028 flags &= ~SCF_DO_STCLASS_AND;
6029 StructCopy(&accum, data->start_class, regnode_ssc);
6030 flags |= SCF_DO_STCLASS_OR;
6037 else if (PL_regkind[OP(scan)] == TRIE) {
6038 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
6041 min += trie->minlen;
6042 delta += (trie->maxlen - trie->minlen);
6043 flags &= ~SCF_DO_STCLASS; /* xxx */
6044 if (flags & SCF_DO_SUBSTR) {
6045 /* Cannot expect anything... */
6046 scan_commit(pRExC_state, data, minlenp, is_inf);
6047 data->pos_min += trie->minlen;
6048 data->pos_delta += (trie->maxlen - trie->minlen);
6049 if (trie->maxlen != trie->minlen)
6050 data->cur_is_floating = 1; /* float */
6052 if (trie->jump) /* no more substrings -- for now /grr*/
6053 flags &= ~SCF_DO_SUBSTR;
6055 #endif /* old or new */
6056 #endif /* TRIE_STUDY_OPT */
6058 /* Else: zero-length, ignore. */
6059 scan = regnext(scan);
6064 /* we need to unwind recursion. */
6067 DEBUG_STUDYDATA("frame-end", data, depth, is_inf);
6068 DEBUG_PEEP("fend", scan, depth, flags);
6070 /* restore previous context */
6071 last = frame->last_regnode;
6072 scan = frame->next_regnode;
6073 stopparen = frame->stopparen;
6074 recursed_depth = frame->prev_recursed_depth;
6076 RExC_frame_last = frame->prev_frame;
6077 frame = frame->this_prev_frame;
6078 goto fake_study_recurse;
6082 DEBUG_STUDYDATA("pre-fin", data, depth, is_inf);
6085 *deltap = is_inf_internal ? SSize_t_MAX : delta;
6087 if (flags & SCF_DO_SUBSTR && is_inf)
6088 data->pos_delta = SSize_t_MAX - data->pos_min;
6089 if (is_par > (I32)U8_MAX)
6091 if (is_par && pars==1 && data) {
6092 data->flags |= SF_IN_PAR;
6093 data->flags &= ~SF_HAS_PAR;
6095 else if (pars && data) {
6096 data->flags |= SF_HAS_PAR;
6097 data->flags &= ~SF_IN_PAR;
6099 if (flags & SCF_DO_STCLASS_OR)
6100 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6101 if (flags & SCF_TRIE_RESTUDY)
6102 data->flags |= SCF_TRIE_RESTUDY;
6104 DEBUG_STUDYDATA("post-fin", data, depth, is_inf);
6107 SSize_t final_minlen= min < stopmin ? min : stopmin;
6109 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN)) {
6110 if (final_minlen > SSize_t_MAX - delta)
6111 RExC_maxlen = SSize_t_MAX;
6112 else if (RExC_maxlen < final_minlen + delta)
6113 RExC_maxlen = final_minlen + delta;
6115 return final_minlen;
6117 NOT_REACHED; /* NOTREACHED */
6121 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
6123 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
6125 PERL_ARGS_ASSERT_ADD_DATA;
6127 Renewc(RExC_rxi->data,
6128 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
6129 char, struct reg_data);
6131 Renew(RExC_rxi->data->what, count + n, U8);
6133 Newx(RExC_rxi->data->what, n, U8);
6134 RExC_rxi->data->count = count + n;
6135 Copy(s, RExC_rxi->data->what + count, n, U8);
6139 /*XXX: todo make this not included in a non debugging perl, but appears to be
6140 * used anyway there, in 'use re' */
6141 #ifndef PERL_IN_XSUB_RE
6143 Perl_reginitcolors(pTHX)
6145 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
6147 char *t = savepv(s);
6151 t = strchr(t, '\t');
6157 PL_colors[i] = t = (char *)"";
6162 PL_colors[i++] = (char *)"";
6169 #ifdef TRIE_STUDY_OPT
6170 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
6173 (data.flags & SCF_TRIE_RESTUDY) \
6181 #define CHECK_RESTUDY_GOTO_butfirst
6185 * pregcomp - compile a regular expression into internal code
6187 * Decides which engine's compiler to call based on the hint currently in
6191 #ifndef PERL_IN_XSUB_RE
6193 /* return the currently in-scope regex engine (or the default if none) */
6195 regexp_engine const *
6196 Perl_current_re_engine(pTHX)
6198 if (IN_PERL_COMPILETIME) {
6199 HV * const table = GvHV(PL_hintgv);
6202 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
6203 return &PL_core_reg_engine;
6204 ptr = hv_fetchs(table, "regcomp", FALSE);
6205 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
6206 return &PL_core_reg_engine;
6207 return INT2PTR(regexp_engine*,SvIV(*ptr));
6211 if (!PL_curcop->cop_hints_hash)
6212 return &PL_core_reg_engine;
6213 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
6214 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
6215 return &PL_core_reg_engine;
6216 return INT2PTR(regexp_engine*,SvIV(ptr));
6222 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
6224 regexp_engine const *eng = current_re_engine();
6225 GET_RE_DEBUG_FLAGS_DECL;
6227 PERL_ARGS_ASSERT_PREGCOMP;
6229 /* Dispatch a request to compile a regexp to correct regexp engine. */
6231 Perl_re_printf( aTHX_ "Using engine %" UVxf "\n",
6234 return CALLREGCOMP_ENG(eng, pattern, flags);
6238 /* public(ish) entry point for the perl core's own regex compiling code.
6239 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
6240 * pattern rather than a list of OPs, and uses the internal engine rather
6241 * than the current one */
6244 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
6246 SV *pat = pattern; /* defeat constness! */
6247 PERL_ARGS_ASSERT_RE_COMPILE;
6248 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
6249 #ifdef PERL_IN_XSUB_RE
6252 &PL_core_reg_engine,
6254 NULL, NULL, rx_flags, 0);
6259 S_free_codeblocks(pTHX_ struct reg_code_blocks *cbs)
6263 if (--cbs->refcnt > 0)
6265 for (n = 0; n < cbs->count; n++) {
6266 REGEXP *rx = cbs->cb[n].src_regex;
6267 cbs->cb[n].src_regex = NULL;
6275 static struct reg_code_blocks *
6276 S_alloc_code_blocks(pTHX_ int ncode)
6278 struct reg_code_blocks *cbs;
6279 Newx(cbs, 1, struct reg_code_blocks);
6282 SAVEDESTRUCTOR_X(S_free_codeblocks, cbs);
6284 Newx(cbs->cb, ncode, struct reg_code_block);
6291 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
6292 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
6293 * point to the realloced string and length.
6295 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
6299 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
6300 char **pat_p, STRLEN *plen_p, int num_code_blocks)
6302 U8 *const src = (U8*)*pat_p;
6307 GET_RE_DEBUG_FLAGS_DECL;
6309 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6310 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
6312 Newx(dst, *plen_p * 2 + 1, U8);
6315 while (s < *plen_p) {
6316 append_utf8_from_native_byte(src[s], &d);
6318 if (n < num_code_blocks) {
6319 assert(pRExC_state->code_blocks);
6320 if (!do_end && pRExC_state->code_blocks->cb[n].start == s) {
6321 pRExC_state->code_blocks->cb[n].start = d - dst - 1;
6322 assert(*(d - 1) == '(');
6325 else if (do_end && pRExC_state->code_blocks->cb[n].end == s) {
6326 pRExC_state->code_blocks->cb[n].end = d - dst - 1;
6327 assert(*(d - 1) == ')');
6336 *pat_p = (char*) dst;
6338 RExC_orig_utf8 = RExC_utf8 = 1;
6343 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
6344 * while recording any code block indices, and handling overloading,
6345 * nested qr// objects etc. If pat is null, it will allocate a new
6346 * string, or just return the first arg, if there's only one.
6348 * Returns the malloced/updated pat.
6349 * patternp and pat_count is the array of SVs to be concatted;
6350 * oplist is the optional list of ops that generated the SVs;
6351 * recompile_p is a pointer to a boolean that will be set if
6352 * the regex will need to be recompiled.
6353 * delim, if non-null is an SV that will be inserted between each element
6357 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
6358 SV *pat, SV ** const patternp, int pat_count,
6359 OP *oplist, bool *recompile_p, SV *delim)
6363 bool use_delim = FALSE;
6364 bool alloced = FALSE;
6366 /* if we know we have at least two args, create an empty string,
6367 * then concatenate args to that. For no args, return an empty string */
6368 if (!pat && pat_count != 1) {
6374 for (svp = patternp; svp < patternp + pat_count; svp++) {
6377 STRLEN orig_patlen = 0;
6379 SV *msv = use_delim ? delim : *svp;
6380 if (!msv) msv = &PL_sv_undef;
6382 /* if we've got a delimiter, we go round the loop twice for each
6383 * svp slot (except the last), using the delimiter the second
6392 if (SvTYPE(msv) == SVt_PVAV) {
6393 /* we've encountered an interpolated array within
6394 * the pattern, e.g. /...@a..../. Expand the list of elements,
6395 * then recursively append elements.
6396 * The code in this block is based on S_pushav() */
6398 AV *const av = (AV*)msv;
6399 const SSize_t maxarg = AvFILL(av) + 1;
6403 assert(oplist->op_type == OP_PADAV
6404 || oplist->op_type == OP_RV2AV);
6405 oplist = OpSIBLING(oplist);
6408 if (SvRMAGICAL(av)) {
6411 Newx(array, maxarg, SV*);
6413 for (i=0; i < maxarg; i++) {
6414 SV ** const svp = av_fetch(av, i, FALSE);
6415 array[i] = svp ? *svp : &PL_sv_undef;
6419 array = AvARRAY(av);
6421 pat = S_concat_pat(aTHX_ pRExC_state, pat,
6422 array, maxarg, NULL, recompile_p,
6424 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
6430 /* we make the assumption here that each op in the list of
6431 * op_siblings maps to one SV pushed onto the stack,
6432 * except for code blocks, with have both an OP_NULL and
6434 * This allows us to match up the list of SVs against the
6435 * list of OPs to find the next code block.
6437 * Note that PUSHMARK PADSV PADSV ..
6439 * PADRANGE PADSV PADSV ..
6440 * so the alignment still works. */
6443 if (oplist->op_type == OP_NULL
6444 && (oplist->op_flags & OPf_SPECIAL))
6446 assert(n < pRExC_state->code_blocks->count);
6447 pRExC_state->code_blocks->cb[n].start = pat ? SvCUR(pat) : 0;
6448 pRExC_state->code_blocks->cb[n].block = oplist;
6449 pRExC_state->code_blocks->cb[n].src_regex = NULL;
6452 oplist = OpSIBLING(oplist); /* skip CONST */
6455 oplist = OpSIBLING(oplist);;
6458 /* apply magic and QR overloading to arg */
6461 if (SvROK(msv) && SvAMAGIC(msv)) {
6462 SV *sv = AMG_CALLunary(msv, regexp_amg);
6466 if (SvTYPE(sv) != SVt_REGEXP)
6467 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
6472 /* try concatenation overload ... */
6473 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
6474 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
6477 /* overloading involved: all bets are off over literal
6478 * code. Pretend we haven't seen it */
6480 pRExC_state->code_blocks->count -= n;
6484 /* ... or failing that, try "" overload */
6485 while (SvAMAGIC(msv)
6486 && (sv = AMG_CALLunary(msv, string_amg))
6490 && SvRV(msv) == SvRV(sv))
6495 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
6499 /* this is a partially unrolled
6500 * sv_catsv_nomg(pat, msv);
6501 * that allows us to adjust code block indices if
6504 char *dst = SvPV_force_nomg(pat, dlen);
6506 if (SvUTF8(msv) && !SvUTF8(pat)) {
6507 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
6508 sv_setpvn(pat, dst, dlen);
6511 sv_catsv_nomg(pat, msv);
6515 /* We have only one SV to process, but we need to verify
6516 * it is properly null terminated or we will fail asserts
6517 * later. In theory we probably shouldn't get such SV's,
6518 * but if we do we should handle it gracefully. */
6519 if ( SvTYPE(msv) != SVt_PV || (SvLEN(msv) > SvCUR(msv) && *(SvEND(msv)) == 0) || SvIsCOW_shared_hash(msv) ) {
6520 /* not a string, or a string with a trailing null */
6523 /* a string with no trailing null, we need to copy it
6524 * so it has a trailing null */
6525 pat = sv_2mortal(newSVsv(msv));
6530 pRExC_state->code_blocks->cb[n-1].end = SvCUR(pat)-1;
6533 /* extract any code blocks within any embedded qr//'s */
6534 if (rx && SvTYPE(rx) == SVt_REGEXP
6535 && RX_ENGINE((REGEXP*)rx)->op_comp)
6538 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
6539 if (ri->code_blocks && ri->code_blocks->count) {
6541 /* the presence of an embedded qr// with code means
6542 * we should always recompile: the text of the
6543 * qr// may not have changed, but it may be a
6544 * different closure than last time */
6546 if (pRExC_state->code_blocks) {
6547 int new_count = pRExC_state->code_blocks->count
6548 + ri->code_blocks->count;
6549 Renew(pRExC_state->code_blocks->cb,
6550 new_count, struct reg_code_block);
6551 pRExC_state->code_blocks->count = new_count;
6554 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_
6555 ri->code_blocks->count);
6557 for (i=0; i < ri->code_blocks->count; i++) {
6558 struct reg_code_block *src, *dst;
6559 STRLEN offset = orig_patlen
6560 + ReANY((REGEXP *)rx)->pre_prefix;
6561 assert(n < pRExC_state->code_blocks->count);
6562 src = &ri->code_blocks->cb[i];
6563 dst = &pRExC_state->code_blocks->cb[n];
6564 dst->start = src->start + offset;
6565 dst->end = src->end + offset;
6566 dst->block = src->block;
6567 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
6576 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
6585 /* see if there are any run-time code blocks in the pattern.
6586 * False positives are allowed */
6589 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6590 char *pat, STRLEN plen)
6595 PERL_UNUSED_CONTEXT;
6597 for (s = 0; s < plen; s++) {
6598 if ( pRExC_state->code_blocks
6599 && n < pRExC_state->code_blocks->count
6600 && s == pRExC_state->code_blocks->cb[n].start)
6602 s = pRExC_state->code_blocks->cb[n].end;
6606 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
6608 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
6610 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
6617 /* Handle run-time code blocks. We will already have compiled any direct
6618 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
6619 * copy of it, but with any literal code blocks blanked out and
6620 * appropriate chars escaped; then feed it into
6622 * eval "qr'modified_pattern'"
6626 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
6630 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
6632 * After eval_sv()-ing that, grab any new code blocks from the returned qr
6633 * and merge them with any code blocks of the original regexp.
6635 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
6636 * instead, just save the qr and return FALSE; this tells our caller that
6637 * the original pattern needs upgrading to utf8.
6641 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6642 char *pat, STRLEN plen)
6646 GET_RE_DEBUG_FLAGS_DECL;
6648 if (pRExC_state->runtime_code_qr) {
6649 /* this is the second time we've been called; this should
6650 * only happen if the main pattern got upgraded to utf8
6651 * during compilation; re-use the qr we compiled first time
6652 * round (which should be utf8 too)
6654 qr = pRExC_state->runtime_code_qr;
6655 pRExC_state->runtime_code_qr = NULL;
6656 assert(RExC_utf8 && SvUTF8(qr));
6662 int newlen = plen + 7; /* allow for "qr''xx\0" extra chars */
6666 /* determine how many extra chars we need for ' and \ escaping */
6667 for (s = 0; s < plen; s++) {
6668 if (pat[s] == '\'' || pat[s] == '\\')
6672 Newx(newpat, newlen, char);
6674 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
6676 for (s = 0; s < plen; s++) {
6677 if ( pRExC_state->code_blocks
6678 && n < pRExC_state->code_blocks->count
6679 && s == pRExC_state->code_blocks->cb[n].start)
6681 /* blank out literal code block */
6682 assert(pat[s] == '(');
6683 while (s <= pRExC_state->code_blocks->cb[n].end) {
6691 if (pat[s] == '\'' || pat[s] == '\\')
6696 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED) {
6698 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED_MORE) {
6704 Perl_re_printf( aTHX_
6705 "%sre-parsing pattern for runtime code:%s %s\n",
6706 PL_colors[4],PL_colors[5],newpat);
6709 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
6715 PUSHSTACKi(PERLSI_REQUIRE);
6716 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
6717 * parsing qr''; normally only q'' does this. It also alters
6719 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
6720 SvREFCNT_dec_NN(sv);
6725 SV * const errsv = ERRSV;
6726 if (SvTRUE_NN(errsv))
6727 /* use croak_sv ? */
6728 Perl_croak_nocontext("%" SVf, SVfARG(errsv));
6730 assert(SvROK(qr_ref));
6732 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
6733 /* the leaving below frees the tmp qr_ref.
6734 * Give qr a life of its own */
6742 if (!RExC_utf8 && SvUTF8(qr)) {
6743 /* first time through; the pattern got upgraded; save the
6744 * qr for the next time through */
6745 assert(!pRExC_state->runtime_code_qr);
6746 pRExC_state->runtime_code_qr = qr;
6751 /* extract any code blocks within the returned qr// */
6754 /* merge the main (r1) and run-time (r2) code blocks into one */
6756 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
6757 struct reg_code_block *new_block, *dst;
6758 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
6762 if (!r2->code_blocks || !r2->code_blocks->count) /* we guessed wrong */
6764 SvREFCNT_dec_NN(qr);
6768 if (!r1->code_blocks)
6769 r1->code_blocks = S_alloc_code_blocks(aTHX_ 0);
6771 r1c = r1->code_blocks->count;
6772 r2c = r2->code_blocks->count;
6774 Newx(new_block, r1c + r2c, struct reg_code_block);
6778 while (i1 < r1c || i2 < r2c) {
6779 struct reg_code_block *src;
6783 src = &r2->code_blocks->cb[i2++];
6787 src = &r1->code_blocks->cb[i1++];
6788 else if ( r1->code_blocks->cb[i1].start
6789 < r2->code_blocks->cb[i2].start)
6791 src = &r1->code_blocks->cb[i1++];
6792 assert(src->end < r2->code_blocks->cb[i2].start);
6795 assert( r1->code_blocks->cb[i1].start
6796 > r2->code_blocks->cb[i2].start);
6797 src = &r2->code_blocks->cb[i2++];
6799 assert(src->end < r1->code_blocks->cb[i1].start);
6802 assert(pat[src->start] == '(');
6803 assert(pat[src->end] == ')');
6804 dst->start = src->start;
6805 dst->end = src->end;
6806 dst->block = src->block;
6807 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
6811 r1->code_blocks->count += r2c;
6812 Safefree(r1->code_blocks->cb);
6813 r1->code_blocks->cb = new_block;
6816 SvREFCNT_dec_NN(qr);
6822 S_setup_longest(pTHX_ RExC_state_t *pRExC_state,
6823 struct reg_substr_datum *rsd,
6824 struct scan_data_substrs *sub,
6825 STRLEN longest_length)
6827 /* This is the common code for setting up the floating and fixed length
6828 * string data extracted from Perl_re_op_compile() below. Returns a boolean
6829 * as to whether succeeded or not */
6833 bool eol = cBOOL(sub->flags & SF_BEFORE_EOL);
6834 bool meol = cBOOL(sub->flags & SF_BEFORE_MEOL);
6836 if (! (longest_length
6837 || (eol /* Can't have SEOL and MULTI */
6838 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
6840 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
6841 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
6846 /* copy the information about the longest from the reg_scan_data
6847 over to the program. */
6848 if (SvUTF8(sub->str)) {
6850 rsd->utf8_substr = sub->str;
6852 rsd->substr = sub->str;
6853 rsd->utf8_substr = NULL;
6855 /* end_shift is how many chars that must be matched that
6856 follow this item. We calculate it ahead of time as once the
6857 lookbehind offset is added in we lose the ability to correctly
6859 ml = sub->minlenp ? *(sub->minlenp) : (SSize_t)longest_length;
6860 rsd->end_shift = ml - sub->min_offset
6862 /* XXX SvTAIL is always false here - did you mean FBMcf_TAIL
6864 + (SvTAIL(sub->str) != 0)
6868 t = (eol/* Can't have SEOL and MULTI */
6869 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
6870 fbm_compile(sub->str, t ? FBMcf_TAIL : 0);
6876 * Perl_re_op_compile - the perl internal RE engine's function to compile a
6877 * regular expression into internal code.
6878 * The pattern may be passed either as:
6879 * a list of SVs (patternp plus pat_count)
6880 * a list of OPs (expr)
6881 * If both are passed, the SV list is used, but the OP list indicates
6882 * which SVs are actually pre-compiled code blocks
6884 * The SVs in the list have magic and qr overloading applied to them (and
6885 * the list may be modified in-place with replacement SVs in the latter
6888 * If the pattern hasn't changed from old_re, then old_re will be
6891 * eng is the current engine. If that engine has an op_comp method, then
6892 * handle directly (i.e. we assume that op_comp was us); otherwise, just
6893 * do the initial concatenation of arguments and pass on to the external
6896 * If is_bare_re is not null, set it to a boolean indicating whether the
6897 * arg list reduced (after overloading) to a single bare regex which has
6898 * been returned (i.e. /$qr/).
6900 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
6902 * pm_flags contains the PMf_* flags, typically based on those from the
6903 * pm_flags field of the related PMOP. Currently we're only interested in
6904 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
6906 * We can't allocate space until we know how big the compiled form will be,
6907 * but we can't compile it (and thus know how big it is) until we've got a
6908 * place to put the code. So we cheat: we compile it twice, once with code
6909 * generation turned off and size counting turned on, and once "for real".
6910 * This also means that we don't allocate space until we are sure that the
6911 * thing really will compile successfully, and we never have to move the
6912 * code and thus invalidate pointers into it. (Note that it has to be in
6913 * one piece because free() must be able to free it all.) [NB: not true in perl]
6915 * Beware that the optimization-preparation code in here knows about some
6916 * of the structure of the compiled regexp. [I'll say.]
6920 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
6921 OP *expr, const regexp_engine* eng, REGEXP *old_re,
6922 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
6926 regexp_internal *ri;
6934 SV** new_patternp = patternp;
6936 /* these are all flags - maybe they should be turned
6937 * into a single int with different bit masks */
6938 I32 sawlookahead = 0;
6943 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
6945 bool runtime_code = 0;
6947 RExC_state_t RExC_state;
6948 RExC_state_t * const pRExC_state = &RExC_state;
6949 #ifdef TRIE_STUDY_OPT
6951 RExC_state_t copyRExC_state;
6953 GET_RE_DEBUG_FLAGS_DECL;
6955 PERL_ARGS_ASSERT_RE_OP_COMPILE;
6957 DEBUG_r(if (!PL_colorset) reginitcolors());
6959 /* Initialize these here instead of as-needed, as is quick and avoids
6960 * having to test them each time otherwise */
6961 if (! PL_InBitmap) {
6963 char * dump_len_string;
6966 /* This is calculated here, because the Perl program that generates the
6967 * static global ones doesn't currently have access to
6968 * NUM_ANYOF_CODE_POINTS */
6969 PL_InBitmap = _new_invlist(2);
6970 PL_InBitmap = _add_range_to_invlist(PL_InBitmap, 0,
6971 NUM_ANYOF_CODE_POINTS - 1);
6973 dump_len_string = PerlEnv_getenv("PERL_DUMP_RE_MAX_LEN");
6974 if ( ! dump_len_string
6975 || ! grok_atoUV(dump_len_string, (UV *)&PL_dump_re_max_len, NULL))
6977 PL_dump_re_max_len = 60; /* A reasonable default */
6982 pRExC_state->warn_text = NULL;
6983 pRExC_state->code_blocks = NULL;
6986 *is_bare_re = FALSE;
6988 if (expr && (expr->op_type == OP_LIST ||
6989 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
6990 /* allocate code_blocks if needed */
6994 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o))
6995 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
6996 ncode++; /* count of DO blocks */
6999 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_ ncode);
7003 /* compile-time pattern with just OP_CONSTs and DO blocks */
7008 /* find how many CONSTs there are */
7011 if (expr->op_type == OP_CONST)
7014 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
7015 if (o->op_type == OP_CONST)
7019 /* fake up an SV array */
7021 assert(!new_patternp);
7022 Newx(new_patternp, n, SV*);
7023 SAVEFREEPV(new_patternp);
7027 if (expr->op_type == OP_CONST)
7028 new_patternp[n] = cSVOPx_sv(expr);
7030 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
7031 if (o->op_type == OP_CONST)
7032 new_patternp[n++] = cSVOPo_sv;
7037 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7038 "Assembling pattern from %d elements%s\n", pat_count,
7039 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
7041 /* set expr to the first arg op */
7043 if (pRExC_state->code_blocks && pRExC_state->code_blocks->count
7044 && expr->op_type != OP_CONST)
7046 expr = cLISTOPx(expr)->op_first;
7047 assert( expr->op_type == OP_PUSHMARK
7048 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
7049 || expr->op_type == OP_PADRANGE);
7050 expr = OpSIBLING(expr);
7053 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
7054 expr, &recompile, NULL);
7056 /* handle bare (possibly after overloading) regex: foo =~ $re */
7061 if (SvTYPE(re) == SVt_REGEXP) {
7065 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7066 "Precompiled pattern%s\n",
7067 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
7073 exp = SvPV_nomg(pat, plen);
7075 if (!eng->op_comp) {
7076 if ((SvUTF8(pat) && IN_BYTES)
7077 || SvGMAGICAL(pat) || SvAMAGIC(pat))
7079 /* make a temporary copy; either to convert to bytes,
7080 * or to avoid repeating get-magic / overloaded stringify */
7081 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
7082 (IN_BYTES ? 0 : SvUTF8(pat)));
7084 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
7087 /* ignore the utf8ness if the pattern is 0 length */
7088 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
7090 RExC_uni_semantics = 0;
7091 RExC_seen_unfolded_sharp_s = 0;
7092 RExC_contains_locale = 0;
7093 RExC_strict = cBOOL(pm_flags & RXf_PMf_STRICT);
7094 RExC_in_script_run = 0;
7095 RExC_study_started = 0;
7096 pRExC_state->runtime_code_qr = NULL;
7097 RExC_frame_head= NULL;
7098 RExC_frame_last= NULL;
7099 RExC_frame_count= 0;
7102 RExC_mysv1= sv_newmortal();
7103 RExC_mysv2= sv_newmortal();
7106 SV *dsv= sv_newmortal();
7107 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, PL_dump_re_max_len);
7108 Perl_re_printf( aTHX_ "%sCompiling REx%s %s\n",
7109 PL_colors[4],PL_colors[5],s);
7113 /* we jump here if we have to recompile, e.g., from upgrading the pattern
7116 if ((pm_flags & PMf_USE_RE_EVAL)
7117 /* this second condition covers the non-regex literal case,
7118 * i.e. $foo =~ '(?{})'. */
7119 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
7121 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
7123 /* return old regex if pattern hasn't changed */
7124 /* XXX: note in the below we have to check the flags as well as the
7127 * Things get a touch tricky as we have to compare the utf8 flag
7128 * independently from the compile flags. */
7132 && !!RX_UTF8(old_re) == !!RExC_utf8
7133 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
7134 && RX_PRECOMP(old_re)
7135 && RX_PRELEN(old_re) == plen
7136 && memEQ(RX_PRECOMP(old_re), exp, plen)
7137 && !runtime_code /* with runtime code, always recompile */ )
7142 rx_flags = orig_rx_flags;
7144 if ( initial_charset == REGEX_DEPENDS_CHARSET
7145 && (RExC_utf8 ||RExC_uni_semantics))
7148 /* Set to use unicode semantics if the pattern is in utf8 and has the
7149 * 'depends' charset specified, as it means unicode when utf8 */
7150 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
7154 RExC_precomp_adj = 0;
7155 RExC_flags = rx_flags;
7156 RExC_pm_flags = pm_flags;
7159 assert(TAINTING_get || !TAINT_get);
7161 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
7163 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
7164 /* whoops, we have a non-utf8 pattern, whilst run-time code
7165 * got compiled as utf8. Try again with a utf8 pattern */
7166 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7167 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7168 goto redo_first_pass;
7171 assert(!pRExC_state->runtime_code_qr);
7177 RExC_in_lookbehind = 0;
7178 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
7181 RExC_recode_x_to_native = 0;
7183 RExC_in_multi_char_class = 0;
7185 /* First pass: determine size, legality. */
7187 RExC_start = RExC_adjusted_start = exp;
7188 RExC_end = exp + plen;
7189 RExC_precomp_end = RExC_end;
7194 RExC_emit = (regnode *) &RExC_emit_dummy;
7195 RExC_whilem_seen = 0;
7196 RExC_open_parens = NULL;
7197 RExC_close_parens = NULL;
7199 RExC_paren_names = NULL;
7201 RExC_paren_name_list = NULL;
7203 RExC_recurse = NULL;
7204 RExC_study_chunk_recursed = NULL;
7205 RExC_study_chunk_recursed_bytes= 0;
7206 RExC_recurse_count = 0;
7207 pRExC_state->code_index = 0;
7209 /* This NUL is guaranteed because the pattern comes from an SV*, and the sv
7210 * code makes sure the final byte is an uncounted NUL. But should this
7211 * ever not be the case, lots of things could read beyond the end of the
7212 * buffer: loops like
7213 * while(isFOO(*RExC_parse)) RExC_parse++;
7214 * strchr(RExC_parse, "foo");
7215 * etc. So it is worth noting. */
7216 assert(*RExC_end == '\0');
7219 Perl_re_printf( aTHX_ "Starting first pass (sizing)\n");
7221 RExC_lastparse=NULL;
7224 if (reg(pRExC_state, 0, &flags,1) == NULL) {
7225 /* It's possible to write a regexp in ascii that represents Unicode
7226 codepoints outside of the byte range, such as via \x{100}. If we
7227 detect such a sequence we have to convert the entire pattern to utf8
7228 and then recompile, as our sizing calculation will have been based
7229 on 1 byte == 1 character, but we will need to use utf8 to encode
7230 at least some part of the pattern, and therefore must convert the whole
7233 if (MUST_RESTART(flags)) {
7234 if (flags & NEED_UTF8) {
7235 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7236 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7237 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "Need to redo pass 1 after upgrade\n"));
7240 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "Need to redo pass 1\n"));
7243 goto redo_first_pass;
7245 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#" UVxf, (UV) flags);
7249 Perl_re_printf( aTHX_
7250 "Required size %" IVdf " nodes\n"
7251 "Starting second pass (creation)\n",
7254 RExC_lastparse=NULL;
7257 /* The first pass could have found things that force Unicode semantics */
7258 if ((RExC_utf8 || RExC_uni_semantics)
7259 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
7261 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
7264 /* Small enough for pointer-storage convention?
7265 If extralen==0, this means that we will not need long jumps. */
7266 if (RExC_size >= 0x10000L && RExC_extralen)
7267 RExC_size += RExC_extralen;
7270 if (RExC_whilem_seen > 15)
7271 RExC_whilem_seen = 15;
7273 /* Allocate space and zero-initialize. Note, the two step process
7274 of zeroing when in debug mode, thus anything assigned has to
7275 happen after that */
7276 rx = (REGEXP*) newSV_type(SVt_REGEXP);
7278 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
7279 char, regexp_internal);
7280 if ( r == NULL || ri == NULL )
7281 FAIL("Regexp out of space");
7283 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
7284 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
7287 /* bulk initialize base fields with 0. */
7288 Zero(ri, sizeof(regexp_internal), char);
7291 /* non-zero initialization begins here */
7294 r->extflags = rx_flags;
7295 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
7297 if (pm_flags & PMf_IS_QR) {
7298 ri->code_blocks = pRExC_state->code_blocks;
7299 if (ri->code_blocks)
7300 ri->code_blocks->refcnt++;
7304 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
7305 bool has_charset = (get_regex_charset(r->extflags)
7306 != REGEX_DEPENDS_CHARSET);
7308 /* The caret is output if there are any defaults: if not all the STD
7309 * flags are set, or if no character set specifier is needed */
7311 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
7313 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
7314 == REG_RUN_ON_COMMENT_SEEN);
7315 U8 reganch = (U8)((r->extflags & RXf_PMf_STD_PMMOD)
7316 >> RXf_PMf_STD_PMMOD_SHIFT);
7317 const char *fptr = STD_PAT_MODS; /*"msixxn"*/
7320 /* We output all the necessary flags; we never output a minus, as all
7321 * those are defaults, so are
7322 * covered by the caret */
7323 const STRLEN wraplen = plen + has_p + has_runon
7324 + has_default /* If needs a caret */
7325 + PL_bitcount[reganch] /* 1 char for each set standard flag */
7327 /* If needs a character set specifier */
7328 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
7329 + (sizeof("(?:)") - 1);
7331 /* make sure PL_bitcount bounds not exceeded */
7332 assert(sizeof(STD_PAT_MODS) <= 8);
7334 p = sv_grow(MUTABLE_SV(rx), wraplen + 1); /* +1 for the ending NUL */
7337 SvFLAGS(rx) |= SVf_UTF8;
7340 /* If a default, cover it using the caret */
7342 *p++= DEFAULT_PAT_MOD;
7346 const char* const name = get_regex_charset_name(r->extflags, &len);
7347 Copy(name, p, len, char);
7351 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
7354 while((ch = *fptr++)) {
7362 Copy(RExC_precomp, p, plen, char);
7363 assert ((RX_WRAPPED(rx) - p) < 16);
7364 r->pre_prefix = p - RX_WRAPPED(rx);
7370 SvCUR_set(rx, p - RX_WRAPPED(rx));
7374 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
7376 /* Useful during FAIL. */
7377 #ifdef RE_TRACK_PATTERN_OFFSETS
7378 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
7379 DEBUG_OFFSETS_r(Perl_re_printf( aTHX_
7380 "%s %" UVuf " bytes for offset annotations.\n",
7381 ri->u.offsets ? "Got" : "Couldn't get",
7382 (UV)((2*RExC_size+1) * sizeof(U32))));
7384 SetProgLen(ri,RExC_size);
7389 /* Second pass: emit code. */
7390 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
7391 RExC_pm_flags = pm_flags;
7393 RExC_end = exp + plen;
7395 RExC_emit_start = ri->program;
7396 RExC_emit = ri->program;
7397 RExC_emit_bound = ri->program + RExC_size + 1;
7398 pRExC_state->code_index = 0;
7400 *((char*) RExC_emit++) = (char) REG_MAGIC;
7401 /* setup various meta data about recursion, this all requires
7402 * RExC_npar to be correctly set, and a bit later on we clear it */
7403 if (RExC_seen & REG_RECURSE_SEEN) {
7404 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
7405 "%*s%*s Setting up open/close parens\n",
7406 22, "| |", (int)(0 * 2 + 1), ""));
7408 /* setup RExC_open_parens, which holds the address of each
7409 * OPEN tag, and to make things simpler for the 0 index
7410 * the start of the program - this is used later for offsets */
7411 Newxz(RExC_open_parens, RExC_npar,regnode *);
7412 SAVEFREEPV(RExC_open_parens);
7413 RExC_open_parens[0] = RExC_emit;
7415 /* setup RExC_close_parens, which holds the address of each
7416 * CLOSE tag, and to make things simpler for the 0 index
7417 * the end of the program - this is used later for offsets */
7418 Newxz(RExC_close_parens, RExC_npar,regnode *);
7419 SAVEFREEPV(RExC_close_parens);
7420 /* we dont know where end op starts yet, so we dont
7421 * need to set RExC_close_parens[0] like we do RExC_open_parens[0] above */
7423 /* Note, RExC_npar is 1 + the number of parens in a pattern.
7424 * So its 1 if there are no parens. */
7425 RExC_study_chunk_recursed_bytes= (RExC_npar >> 3) +
7426 ((RExC_npar & 0x07) != 0);
7427 Newx(RExC_study_chunk_recursed,
7428 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
7429 SAVEFREEPV(RExC_study_chunk_recursed);
7432 if (reg(pRExC_state, 0, &flags,1) == NULL) {
7434 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#" UVxf, (UV) flags);
7437 Perl_re_printf( aTHX_ "Starting post parse optimization\n");
7440 /* XXXX To minimize changes to RE engine we always allocate
7441 3-units-long substrs field. */
7442 Newx(r->substrs, 1, struct reg_substr_data);
7443 if (RExC_recurse_count) {
7444 Newx(RExC_recurse,RExC_recurse_count,regnode *);
7445 SAVEFREEPV(RExC_recurse);
7449 r->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
7451 RExC_study_chunk_recursed_count= 0;
7453 Zero(r->substrs, 1, struct reg_substr_data);
7454 if (RExC_study_chunk_recursed) {
7455 Zero(RExC_study_chunk_recursed,
7456 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
7460 #ifdef TRIE_STUDY_OPT
7462 StructCopy(&zero_scan_data, &data, scan_data_t);
7463 copyRExC_state = RExC_state;
7466 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "Restudying\n"));
7468 RExC_state = copyRExC_state;
7469 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
7470 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
7472 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
7473 StructCopy(&zero_scan_data, &data, scan_data_t);
7476 StructCopy(&zero_scan_data, &data, scan_data_t);
7479 /* Dig out information for optimizations. */
7480 r->extflags = RExC_flags; /* was pm_op */
7481 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
7484 SvUTF8_on(rx); /* Unicode in it? */
7485 ri->regstclass = NULL;
7486 if (RExC_naughty >= TOO_NAUGHTY) /* Probably an expensive pattern. */
7487 r->intflags |= PREGf_NAUGHTY;
7488 scan = ri->program + 1; /* First BRANCH. */
7490 /* testing for BRANCH here tells us whether there is "must appear"
7491 data in the pattern. If there is then we can use it for optimisations */
7492 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
7495 STRLEN longest_length[2];
7496 regnode_ssc ch_class; /* pointed to by data */
7498 SSize_t last_close = 0; /* pointed to by data */
7499 regnode *first= scan;
7500 regnode *first_next= regnext(first);
7504 * Skip introductions and multiplicators >= 1
7505 * so that we can extract the 'meat' of the pattern that must
7506 * match in the large if() sequence following.
7507 * NOTE that EXACT is NOT covered here, as it is normally
7508 * picked up by the optimiser separately.
7510 * This is unfortunate as the optimiser isnt handling lookahead
7511 * properly currently.
7514 while ((OP(first) == OPEN && (sawopen = 1)) ||
7515 /* An OR of *one* alternative - should not happen now. */
7516 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
7517 /* for now we can't handle lookbehind IFMATCH*/
7518 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
7519 (OP(first) == PLUS) ||
7520 (OP(first) == MINMOD) ||
7521 /* An {n,m} with n>0 */
7522 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
7523 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
7526 * the only op that could be a regnode is PLUS, all the rest
7527 * will be regnode_1 or regnode_2.
7529 * (yves doesn't think this is true)
7531 if (OP(first) == PLUS)
7534 if (OP(first) == MINMOD)
7536 first += regarglen[OP(first)];
7538 first = NEXTOPER(first);
7539 first_next= regnext(first);
7542 /* Starting-point info. */
7544 DEBUG_PEEP("first:", first, 0, 0);
7545 /* Ignore EXACT as we deal with it later. */
7546 if (PL_regkind[OP(first)] == EXACT) {
7547 if (OP(first) == EXACT || OP(first) == EXACTL)
7548 NOOP; /* Empty, get anchored substr later. */
7550 ri->regstclass = first;
7553 else if (PL_regkind[OP(first)] == TRIE &&
7554 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
7556 /* this can happen only on restudy */
7557 ri->regstclass = construct_ahocorasick_from_trie(pRExC_state, (regnode *)first, 0);
7560 else if (REGNODE_SIMPLE(OP(first)))
7561 ri->regstclass = first;
7562 else if (PL_regkind[OP(first)] == BOUND ||
7563 PL_regkind[OP(first)] == NBOUND)
7564 ri->regstclass = first;
7565 else if (PL_regkind[OP(first)] == BOL) {
7566 r->intflags |= (OP(first) == MBOL
7569 first = NEXTOPER(first);
7572 else if (OP(first) == GPOS) {
7573 r->intflags |= PREGf_ANCH_GPOS;
7574 first = NEXTOPER(first);
7577 else if ((!sawopen || !RExC_sawback) &&
7579 (OP(first) == STAR &&
7580 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
7581 !(r->intflags & PREGf_ANCH) && !pRExC_state->code_blocks)
7583 /* turn .* into ^.* with an implied $*=1 */
7585 (OP(NEXTOPER(first)) == REG_ANY)
7588 r->intflags |= (type | PREGf_IMPLICIT);
7589 first = NEXTOPER(first);
7592 if (sawplus && !sawminmod && !sawlookahead
7593 && (!sawopen || !RExC_sawback)
7594 && !pRExC_state->code_blocks) /* May examine pos and $& */
7595 /* x+ must match at the 1st pos of run of x's */
7596 r->intflags |= PREGf_SKIP;
7598 /* Scan is after the zeroth branch, first is atomic matcher. */
7599 #ifdef TRIE_STUDY_OPT
7602 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
7603 (IV)(first - scan + 1))
7607 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
7608 (IV)(first - scan + 1))
7614 * If there's something expensive in the r.e., find the
7615 * longest literal string that must appear and make it the
7616 * regmust. Resolve ties in favor of later strings, since
7617 * the regstart check works with the beginning of the r.e.
7618 * and avoiding duplication strengthens checking. Not a
7619 * strong reason, but sufficient in the absence of others.
7620 * [Now we resolve ties in favor of the earlier string if
7621 * it happens that c_offset_min has been invalidated, since the
7622 * earlier string may buy us something the later one won't.]
7625 data.substrs[0].str = newSVpvs("");
7626 data.substrs[1].str = newSVpvs("");
7627 data.last_found = newSVpvs("");
7628 data.cur_is_floating = 0; /* initially any found substring is fixed */
7629 ENTER_with_name("study_chunk");
7630 SAVEFREESV(data.substrs[0].str);
7631 SAVEFREESV(data.substrs[1].str);
7632 SAVEFREESV(data.last_found);
7634 if (!ri->regstclass) {
7635 ssc_init(pRExC_state, &ch_class);
7636 data.start_class = &ch_class;
7637 stclass_flag = SCF_DO_STCLASS_AND;
7638 } else /* XXXX Check for BOUND? */
7640 data.last_closep = &last_close;
7644 * MAIN ENTRY FOR study_chunk() FOR m/PATTERN/
7645 * (NO top level branches)
7647 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
7648 scan + RExC_size, /* Up to end */
7650 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
7651 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
7655 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
7658 if ( RExC_npar == 1 && !data.cur_is_floating
7659 && data.last_start_min == 0 && data.last_end > 0
7660 && !RExC_seen_zerolen
7661 && !(RExC_seen & REG_VERBARG_SEEN)
7662 && !(RExC_seen & REG_GPOS_SEEN)
7664 r->extflags |= RXf_CHECK_ALL;
7666 scan_commit(pRExC_state, &data,&minlen,0);
7669 /* XXX this is done in reverse order because that's the way the
7670 * code was before it was parameterised. Don't know whether it
7671 * actually needs doing in reverse order. DAPM */
7672 for (i = 1; i >= 0; i--) {
7673 longest_length[i] = CHR_SVLEN(data.substrs[i].str);
7676 && SvCUR(data.substrs[0].str) /* ok to leave SvCUR */
7677 && data.substrs[0].min_offset
7678 == data.substrs[1].min_offset
7679 && SvCUR(data.substrs[0].str)
7680 == SvCUR(data.substrs[1].str)
7682 && S_setup_longest (aTHX_ pRExC_state,
7683 &(r->substrs->data[i]),
7687 r->substrs->data[i].min_offset =
7688 data.substrs[i].min_offset - data.substrs[i].lookbehind;
7690 r->substrs->data[i].max_offset = data.substrs[i].max_offset;
7691 /* Don't offset infinity */
7692 if (data.substrs[i].max_offset < SSize_t_MAX)
7693 r->substrs->data[i].max_offset -= data.substrs[i].lookbehind;
7694 SvREFCNT_inc_simple_void_NN(data.substrs[i].str);
7697 r->substrs->data[i].substr = NULL;
7698 r->substrs->data[i].utf8_substr = NULL;
7699 longest_length[i] = 0;
7703 LEAVE_with_name("study_chunk");
7706 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
7707 ri->regstclass = NULL;
7709 if ((!(r->substrs->data[0].substr || r->substrs->data[0].utf8_substr)
7710 || r->substrs->data[0].min_offset)
7712 && ! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7713 && is_ssc_worth_it(pRExC_state, data.start_class))
7715 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7717 ssc_finalize(pRExC_state, data.start_class);
7719 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7720 StructCopy(data.start_class,
7721 (regnode_ssc*)RExC_rxi->data->data[n],
7723 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7724 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7725 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
7726 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7727 Perl_re_printf( aTHX_
7728 "synthetic stclass \"%s\".\n",
7729 SvPVX_const(sv));});
7730 data.start_class = NULL;
7733 /* A temporary algorithm prefers floated substr to fixed one of
7734 * same length to dig more info. */
7735 i = (longest_length[0] <= longest_length[1]);
7736 r->substrs->check_ix = i;
7737 r->check_end_shift = r->substrs->data[i].end_shift;
7738 r->check_substr = r->substrs->data[i].substr;
7739 r->check_utf8 = r->substrs->data[i].utf8_substr;
7740 r->check_offset_min = r->substrs->data[i].min_offset;
7741 r->check_offset_max = r->substrs->data[i].max_offset;
7742 if (!i && (r->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS)))
7743 r->intflags |= PREGf_NOSCAN;
7745 if ((r->check_substr || r->check_utf8) ) {
7746 r->extflags |= RXf_USE_INTUIT;
7747 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
7748 r->extflags |= RXf_INTUIT_TAIL;
7751 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
7752 if ( (STRLEN)minlen < longest_length[1] )
7753 minlen= longest_length[1];
7754 if ( (STRLEN)minlen < longest_length[0] )
7755 minlen= longest_length[0];
7759 /* Several toplevels. Best we can is to set minlen. */
7761 regnode_ssc ch_class;
7762 SSize_t last_close = 0;
7764 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "\nMulti Top Level\n"));
7766 scan = ri->program + 1;
7767 ssc_init(pRExC_state, &ch_class);
7768 data.start_class = &ch_class;
7769 data.last_closep = &last_close;
7773 * MAIN ENTRY FOR study_chunk() FOR m/P1|P2|.../
7774 * (patterns WITH top level branches)
7776 minlen = study_chunk(pRExC_state,
7777 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
7778 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
7779 ? SCF_TRIE_DOING_RESTUDY
7783 CHECK_RESTUDY_GOTO_butfirst(NOOP);
7785 r->check_substr = NULL;
7786 r->check_utf8 = NULL;
7787 r->substrs->data[0].substr = NULL;
7788 r->substrs->data[0].utf8_substr = NULL;
7789 r->substrs->data[1].substr = NULL;
7790 r->substrs->data[1].utf8_substr = NULL;
7792 if (! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7793 && is_ssc_worth_it(pRExC_state, data.start_class))
7795 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7797 ssc_finalize(pRExC_state, data.start_class);
7799 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7800 StructCopy(data.start_class,
7801 (regnode_ssc*)RExC_rxi->data->data[n],
7803 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7804 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7805 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
7806 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7807 Perl_re_printf( aTHX_
7808 "synthetic stclass \"%s\".\n",
7809 SvPVX_const(sv));});
7810 data.start_class = NULL;
7814 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
7815 r->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
7816 r->maxlen = REG_INFTY;
7819 r->maxlen = RExC_maxlen;
7822 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
7823 the "real" pattern. */
7825 Perl_re_printf( aTHX_ "minlen: %" IVdf " r->minlen:%" IVdf " maxlen:%" IVdf "\n",
7826 (IV)minlen, (IV)r->minlen, (IV)RExC_maxlen);
7828 r->minlenret = minlen;
7829 if (r->minlen < minlen)
7832 if (RExC_seen & REG_RECURSE_SEEN ) {
7833 r->intflags |= PREGf_RECURSE_SEEN;
7834 Newx(r->recurse_locinput, r->nparens + 1, char *);
7836 if (RExC_seen & REG_GPOS_SEEN)
7837 r->intflags |= PREGf_GPOS_SEEN;
7838 if (RExC_seen & REG_LOOKBEHIND_SEEN)
7839 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
7841 if (pRExC_state->code_blocks)
7842 r->extflags |= RXf_EVAL_SEEN;
7843 if (RExC_seen & REG_VERBARG_SEEN)
7845 r->intflags |= PREGf_VERBARG_SEEN;
7846 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
7848 if (RExC_seen & REG_CUTGROUP_SEEN)
7849 r->intflags |= PREGf_CUTGROUP_SEEN;
7850 if (pm_flags & PMf_USE_RE_EVAL)
7851 r->intflags |= PREGf_USE_RE_EVAL;
7852 if (RExC_paren_names)
7853 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
7855 RXp_PAREN_NAMES(r) = NULL;
7857 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
7858 * so it can be used in pp.c */
7859 if (r->intflags & PREGf_ANCH)
7860 r->extflags |= RXf_IS_ANCHORED;
7864 /* this is used to identify "special" patterns that might result
7865 * in Perl NOT calling the regex engine and instead doing the match "itself",
7866 * particularly special cases in split//. By having the regex compiler
7867 * do this pattern matching at a regop level (instead of by inspecting the pattern)
7868 * we avoid weird issues with equivalent patterns resulting in different behavior,
7869 * AND we allow non Perl engines to get the same optimizations by the setting the
7870 * flags appropriately - Yves */
7871 regnode *first = ri->program + 1;
7873 regnode *next = regnext(first);
7876 if (PL_regkind[fop] == NOTHING && nop == END)
7877 r->extflags |= RXf_NULL;
7878 else if ((fop == MBOL || (fop == SBOL && !first->flags)) && nop == END)
7879 /* when fop is SBOL first->flags will be true only when it was
7880 * produced by parsing /\A/, and not when parsing /^/. This is
7881 * very important for the split code as there we want to
7882 * treat /^/ as /^/m, but we do not want to treat /\A/ as /^/m.
7883 * See rt #122761 for more details. -- Yves */
7884 r->extflags |= RXf_START_ONLY;
7885 else if (fop == PLUS
7886 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
7888 r->extflags |= RXf_WHITE;
7889 else if ( r->extflags & RXf_SPLIT
7890 && (fop == EXACT || fop == EXACTL)
7891 && STR_LEN(first) == 1
7892 && *(STRING(first)) == ' '
7894 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
7898 if (RExC_contains_locale) {
7899 RXp_EXTFLAGS(r) |= RXf_TAINTED;
7903 if (RExC_paren_names) {
7904 ri->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
7905 ri->data->data[ri->name_list_idx]
7906 = (void*)SvREFCNT_inc(RExC_paren_name_list);
7909 ri->name_list_idx = 0;
7911 while ( RExC_recurse_count > 0 ) {
7912 const regnode *scan = RExC_recurse[ --RExC_recurse_count ];
7914 * This data structure is set up in study_chunk() and is used
7915 * to calculate the distance between a GOSUB regopcode and
7916 * the OPEN/CURLYM (CURLYM's are special and can act like OPEN's)
7919 * If for some reason someone writes code that optimises
7920 * away a GOSUB opcode then the assert should be changed to
7921 * an if(scan) to guard the ARG2L_SET() - Yves
7924 assert(scan && OP(scan) == GOSUB);
7925 ARG2L_SET( scan, RExC_open_parens[ARG(scan)] - scan );
7928 Newxz(r->offs, RExC_npar, regexp_paren_pair);
7929 /* assume we don't need to swap parens around before we match */
7931 Perl_re_printf( aTHX_ "study_chunk_recursed_count: %lu\n",
7932 (unsigned long)RExC_study_chunk_recursed_count);
7936 Perl_re_printf( aTHX_ "Final program:\n");
7939 #ifdef RE_TRACK_PATTERN_OFFSETS
7940 DEBUG_OFFSETS_r(if (ri->u.offsets) {
7941 const STRLEN len = ri->u.offsets[0];
7943 GET_RE_DEBUG_FLAGS_DECL;
7944 Perl_re_printf( aTHX_
7945 "Offsets: [%" UVuf "]\n\t", (UV)ri->u.offsets[0]);
7946 for (i = 1; i <= len; i++) {
7947 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
7948 Perl_re_printf( aTHX_ "%" UVuf ":%" UVuf "[%" UVuf "] ",
7949 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
7951 Perl_re_printf( aTHX_ "\n");
7956 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
7957 * by setting the regexp SV to readonly-only instead. If the
7958 * pattern's been recompiled, the USEDness should remain. */
7959 if (old_re && SvREADONLY(old_re))
7967 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
7970 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
7972 PERL_UNUSED_ARG(value);
7974 if (flags & RXapif_FETCH) {
7975 return reg_named_buff_fetch(rx, key, flags);
7976 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
7977 Perl_croak_no_modify();
7979 } else if (flags & RXapif_EXISTS) {
7980 return reg_named_buff_exists(rx, key, flags)
7983 } else if (flags & RXapif_REGNAMES) {
7984 return reg_named_buff_all(rx, flags);
7985 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
7986 return reg_named_buff_scalar(rx, flags);
7988 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
7994 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
7997 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
7998 PERL_UNUSED_ARG(lastkey);
8000 if (flags & RXapif_FIRSTKEY)
8001 return reg_named_buff_firstkey(rx, flags);
8002 else if (flags & RXapif_NEXTKEY)
8003 return reg_named_buff_nextkey(rx, flags);
8005 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
8012 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
8016 struct regexp *const rx = ReANY(r);
8018 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
8020 if (rx && RXp_PAREN_NAMES(rx)) {
8021 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
8024 SV* sv_dat=HeVAL(he_str);
8025 I32 *nums=(I32*)SvPVX(sv_dat);
8026 AV * const retarray = (flags & RXapif_ALL) ? newAV() : NULL;
8027 for ( i=0; i<SvIVX(sv_dat); i++ ) {
8028 if ((I32)(rx->nparens) >= nums[i]
8029 && rx->offs[nums[i]].start != -1
8030 && rx->offs[nums[i]].end != -1)
8033 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
8038 ret = newSVsv(&PL_sv_undef);
8041 av_push(retarray, ret);
8044 return newRV_noinc(MUTABLE_SV(retarray));
8051 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
8054 struct regexp *const rx = ReANY(r);
8056 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
8058 if (rx && RXp_PAREN_NAMES(rx)) {
8059 if (flags & RXapif_ALL) {
8060 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
8062 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
8064 SvREFCNT_dec_NN(sv);
8076 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
8078 struct regexp *const rx = ReANY(r);
8080 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
8082 if ( rx && RXp_PAREN_NAMES(rx) ) {
8083 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
8085 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
8092 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
8094 struct regexp *const rx = ReANY(r);
8095 GET_RE_DEBUG_FLAGS_DECL;
8097 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
8099 if (rx && RXp_PAREN_NAMES(rx)) {
8100 HV *hv = RXp_PAREN_NAMES(rx);
8102 while ( (temphe = hv_iternext_flags(hv,0)) ) {
8105 SV* sv_dat = HeVAL(temphe);
8106 I32 *nums = (I32*)SvPVX(sv_dat);
8107 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8108 if ((I32)(rx->lastparen) >= nums[i] &&
8109 rx->offs[nums[i]].start != -1 &&
8110 rx->offs[nums[i]].end != -1)
8116 if (parno || flags & RXapif_ALL) {
8117 return newSVhek(HeKEY_hek(temphe));
8125 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
8130 struct regexp *const rx = ReANY(r);
8132 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
8134 if (rx && RXp_PAREN_NAMES(rx)) {
8135 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
8136 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
8137 } else if (flags & RXapif_ONE) {
8138 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
8139 av = MUTABLE_AV(SvRV(ret));
8140 length = av_tindex(av);
8141 SvREFCNT_dec_NN(ret);
8142 return newSViv(length + 1);
8144 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
8149 return &PL_sv_undef;
8153 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
8155 struct regexp *const rx = ReANY(r);
8158 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
8160 if (rx && RXp_PAREN_NAMES(rx)) {
8161 HV *hv= RXp_PAREN_NAMES(rx);
8163 (void)hv_iterinit(hv);
8164 while ( (temphe = hv_iternext_flags(hv,0)) ) {
8167 SV* sv_dat = HeVAL(temphe);
8168 I32 *nums = (I32*)SvPVX(sv_dat);
8169 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8170 if ((I32)(rx->lastparen) >= nums[i] &&
8171 rx->offs[nums[i]].start != -1 &&
8172 rx->offs[nums[i]].end != -1)
8178 if (parno || flags & RXapif_ALL) {
8179 av_push(av, newSVhek(HeKEY_hek(temphe)));
8184 return newRV_noinc(MUTABLE_SV(av));
8188 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
8191 struct regexp *const rx = ReANY(r);
8197 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
8199 if ( n == RX_BUFF_IDX_CARET_PREMATCH
8200 || n == RX_BUFF_IDX_CARET_FULLMATCH
8201 || n == RX_BUFF_IDX_CARET_POSTMATCH
8204 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8206 /* on something like
8209 * the KEEPCOPY is set on the PMOP rather than the regex */
8210 if (PL_curpm && r == PM_GETRE(PL_curpm))
8211 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8220 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
8221 /* no need to distinguish between them any more */
8222 n = RX_BUFF_IDX_FULLMATCH;
8224 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
8225 && rx->offs[0].start != -1)
8227 /* $`, ${^PREMATCH} */
8228 i = rx->offs[0].start;
8232 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
8233 && rx->offs[0].end != -1)
8235 /* $', ${^POSTMATCH} */
8236 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
8237 i = rx->sublen + rx->suboffset - rx->offs[0].end;
8240 if ( 0 <= n && n <= (I32)rx->nparens &&
8241 (s1 = rx->offs[n].start) != -1 &&
8242 (t1 = rx->offs[n].end) != -1)
8244 /* $&, ${^MATCH}, $1 ... */
8246 s = rx->subbeg + s1 - rx->suboffset;
8251 assert(s >= rx->subbeg);
8252 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
8254 #ifdef NO_TAINT_SUPPORT
8255 sv_setpvn(sv, s, i);
8257 const int oldtainted = TAINT_get;
8259 sv_setpvn(sv, s, i);
8260 TAINT_set(oldtainted);
8262 if (RXp_MATCH_UTF8(rx))
8267 if (RXp_MATCH_TAINTED(rx)) {
8268 if (SvTYPE(sv) >= SVt_PVMG) {
8269 MAGIC* const mg = SvMAGIC(sv);
8272 SvMAGIC_set(sv, mg->mg_moremagic);
8274 if ((mgt = SvMAGIC(sv))) {
8275 mg->mg_moremagic = mgt;
8276 SvMAGIC_set(sv, mg);
8293 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
8294 SV const * const value)
8296 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
8298 PERL_UNUSED_ARG(rx);
8299 PERL_UNUSED_ARG(paren);
8300 PERL_UNUSED_ARG(value);
8303 Perl_croak_no_modify();
8307 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
8310 struct regexp *const rx = ReANY(r);
8314 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
8316 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
8317 || paren == RX_BUFF_IDX_CARET_FULLMATCH
8318 || paren == RX_BUFF_IDX_CARET_POSTMATCH
8321 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8323 /* on something like
8326 * the KEEPCOPY is set on the PMOP rather than the regex */
8327 if (PL_curpm && r == PM_GETRE(PL_curpm))
8328 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8334 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
8336 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
8337 case RX_BUFF_IDX_PREMATCH: /* $` */
8338 if (rx->offs[0].start != -1) {
8339 i = rx->offs[0].start;
8348 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
8349 case RX_BUFF_IDX_POSTMATCH: /* $' */
8350 if (rx->offs[0].end != -1) {
8351 i = rx->sublen - rx->offs[0].end;
8353 s1 = rx->offs[0].end;
8360 default: /* $& / ${^MATCH}, $1, $2, ... */
8361 if (paren <= (I32)rx->nparens &&
8362 (s1 = rx->offs[paren].start) != -1 &&
8363 (t1 = rx->offs[paren].end) != -1)
8369 if (ckWARN(WARN_UNINITIALIZED))
8370 report_uninit((const SV *)sv);
8375 if (i > 0 && RXp_MATCH_UTF8(rx)) {
8376 const char * const s = rx->subbeg - rx->suboffset + s1;
8381 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
8388 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
8390 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
8391 PERL_UNUSED_ARG(rx);
8395 return newSVpvs("Regexp");
8398 /* Scans the name of a named buffer from the pattern.
8399 * If flags is REG_RSN_RETURN_NULL returns null.
8400 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
8401 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
8402 * to the parsed name as looked up in the RExC_paren_names hash.
8403 * If there is an error throws a vFAIL().. type exception.
8406 #define REG_RSN_RETURN_NULL 0
8407 #define REG_RSN_RETURN_NAME 1
8408 #define REG_RSN_RETURN_DATA 2
8411 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
8413 char *name_start = RExC_parse;
8415 PERL_ARGS_ASSERT_REG_SCAN_NAME;
8417 assert (RExC_parse <= RExC_end);
8418 if (RExC_parse == RExC_end) NOOP;
8419 else if (isIDFIRST_lazy_if_safe(RExC_parse, RExC_end, UTF)) {
8420 /* Note that the code here assumes well-formed UTF-8. Skip IDFIRST by
8421 * using do...while */
8424 RExC_parse += UTF8SKIP(RExC_parse);
8425 } while ( RExC_parse < RExC_end
8426 && isWORDCHAR_utf8_safe((U8*)RExC_parse, (U8*) RExC_end));
8430 } while (RExC_parse < RExC_end && isWORDCHAR(*RExC_parse));
8432 RExC_parse++; /* so the <- from the vFAIL is after the offending
8434 vFAIL("Group name must start with a non-digit word character");
8438 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
8439 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
8440 if ( flags == REG_RSN_RETURN_NAME)
8442 else if (flags==REG_RSN_RETURN_DATA) {
8445 if ( ! sv_name ) /* should not happen*/
8446 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
8447 if (RExC_paren_names)
8448 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
8450 sv_dat = HeVAL(he_str);
8452 vFAIL("Reference to nonexistent named group");
8456 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
8457 (unsigned long) flags);
8459 NOT_REACHED; /* NOTREACHED */
8464 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
8466 if (RExC_lastparse!=RExC_parse) { \
8467 Perl_re_printf( aTHX_ "%s", \
8468 Perl_pv_pretty(aTHX_ RExC_mysv1, RExC_parse, \
8469 RExC_end - RExC_parse, 16, \
8471 PERL_PV_ESCAPE_UNI_DETECT | \
8472 PERL_PV_PRETTY_ELLIPSES | \
8473 PERL_PV_PRETTY_LTGT | \
8474 PERL_PV_ESCAPE_RE | \
8475 PERL_PV_PRETTY_EXACTSIZE \
8479 Perl_re_printf( aTHX_ "%16s",""); \
8482 num = RExC_size + 1; \
8484 num=REG_NODE_NUM(RExC_emit); \
8485 if (RExC_lastnum!=num) \
8486 Perl_re_printf( aTHX_ "|%4d",num); \
8488 Perl_re_printf( aTHX_ "|%4s",""); \
8489 Perl_re_printf( aTHX_ "|%*s%-4s", \
8490 (int)((depth*2)), "", \
8494 RExC_lastparse=RExC_parse; \
8499 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
8500 DEBUG_PARSE_MSG((funcname)); \
8501 Perl_re_printf( aTHX_ "%4s","\n"); \
8503 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({\
8504 DEBUG_PARSE_MSG((funcname)); \
8505 Perl_re_printf( aTHX_ fmt "\n",args); \
8508 /* This section of code defines the inversion list object and its methods. The
8509 * interfaces are highly subject to change, so as much as possible is static to
8510 * this file. An inversion list is here implemented as a malloc'd C UV array
8511 * as an SVt_INVLIST scalar.
8513 * An inversion list for Unicode is an array of code points, sorted by ordinal
8514 * number. Each element gives the code point that begins a range that extends
8515 * up-to but not including the code point given by the next element. The final
8516 * element gives the first code point of a range that extends to the platform's
8517 * infinity. The even-numbered elements (invlist[0], invlist[2], invlist[4],
8518 * ...) give ranges whose code points are all in the inversion list. We say
8519 * that those ranges are in the set. The odd-numbered elements give ranges
8520 * whose code points are not in the inversion list, and hence not in the set.
8521 * Thus, element [0] is the first code point in the list. Element [1]
8522 * is the first code point beyond that not in the list; and element [2] is the
8523 * first code point beyond that that is in the list. In other words, the first
8524 * range is invlist[0]..(invlist[1]-1), and all code points in that range are
8525 * in the inversion list. The second range is invlist[1]..(invlist[2]-1), and
8526 * all code points in that range are not in the inversion list. The third
8527 * range invlist[2]..(invlist[3]-1) gives code points that are in the inversion
8528 * list, and so forth. Thus every element whose index is divisible by two
8529 * gives the beginning of a range that is in the list, and every element whose
8530 * index is not divisible by two gives the beginning of a range not in the
8531 * list. If the final element's index is divisible by two, the inversion list
8532 * extends to the platform's infinity; otherwise the highest code point in the
8533 * inversion list is the contents of that element minus 1.
8535 * A range that contains just a single code point N will look like
8537 * invlist[i+1] == N+1
8539 * If N is UV_MAX (the highest representable code point on the machine), N+1 is
8540 * impossible to represent, so element [i+1] is omitted. The single element
8542 * invlist[0] == UV_MAX
8543 * contains just UV_MAX, but is interpreted as matching to infinity.
8545 * Taking the complement (inverting) an inversion list is quite simple, if the
8546 * first element is 0, remove it; otherwise add a 0 element at the beginning.
8547 * This implementation reserves an element at the beginning of each inversion
8548 * list to always contain 0; there is an additional flag in the header which
8549 * indicates if the list begins at the 0, or is offset to begin at the next
8550 * element. This means that the inversion list can be inverted without any
8551 * copying; just flip the flag.
8553 * More about inversion lists can be found in "Unicode Demystified"
8554 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
8556 * The inversion list data structure is currently implemented as an SV pointing
8557 * to an array of UVs that the SV thinks are bytes. This allows us to have an
8558 * array of UV whose memory management is automatically handled by the existing
8559 * facilities for SV's.
8561 * Some of the methods should always be private to the implementation, and some
8562 * should eventually be made public */
8564 /* The header definitions are in F<invlist_inline.h> */
8566 #ifndef PERL_IN_XSUB_RE
8568 PERL_STATIC_INLINE UV*
8569 S__invlist_array_init(SV* const invlist, const bool will_have_0)
8571 /* Returns a pointer to the first element in the inversion list's array.
8572 * This is called upon initialization of an inversion list. Where the
8573 * array begins depends on whether the list has the code point U+0000 in it
8574 * or not. The other parameter tells it whether the code that follows this
8575 * call is about to put a 0 in the inversion list or not. The first
8576 * element is either the element reserved for 0, if TRUE, or the element
8577 * after it, if FALSE */
8579 bool* offset = get_invlist_offset_addr(invlist);
8580 UV* zero_addr = (UV *) SvPVX(invlist);
8582 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
8585 assert(! _invlist_len(invlist));
8589 /* 1^1 = 0; 1^0 = 1 */
8590 *offset = 1 ^ will_have_0;
8591 return zero_addr + *offset;
8596 PERL_STATIC_INLINE void
8597 S_invlist_set_len(pTHX_ SV* const invlist, const UV len, const bool offset)
8599 /* Sets the current number of elements stored in the inversion list.
8600 * Updates SvCUR correspondingly */
8601 PERL_UNUSED_CONTEXT;
8602 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
8604 assert(is_invlist(invlist));
8609 : TO_INTERNAL_SIZE(len + offset));
8610 assert(SvLEN(invlist) == 0 || SvCUR(invlist) <= SvLEN(invlist));
8613 #ifndef PERL_IN_XSUB_RE
8616 S_invlist_replace_list_destroys_src(pTHX_ SV * dest, SV * src)
8618 /* Replaces the inversion list in 'dest' with the one from 'src'. It
8619 * steals the list from 'src', so 'src' is made to have a NULL list. This
8620 * is similar to what SvSetMagicSV() would do, if it were implemented on
8621 * inversion lists, though this routine avoids a copy */
8623 const UV src_len = _invlist_len(src);
8624 const bool src_offset = *get_invlist_offset_addr(src);
8625 const STRLEN src_byte_len = SvLEN(src);
8626 char * array = SvPVX(src);
8628 const int oldtainted = TAINT_get;
8630 PERL_ARGS_ASSERT_INVLIST_REPLACE_LIST_DESTROYS_SRC;
8632 assert(is_invlist(src));
8633 assert(is_invlist(dest));
8634 assert(! invlist_is_iterating(src));
8635 assert(SvCUR(src) == 0 || SvCUR(src) < SvLEN(src));
8637 /* Make sure it ends in the right place with a NUL, as our inversion list
8638 * manipulations aren't careful to keep this true, but sv_usepvn_flags()
8640 array[src_byte_len - 1] = '\0';
8642 TAINT_NOT; /* Otherwise it breaks */
8643 sv_usepvn_flags(dest,
8647 /* This flag is documented to cause a copy to be avoided */
8648 SV_HAS_TRAILING_NUL);
8649 TAINT_set(oldtainted);
8654 /* Finish up copying over the other fields in an inversion list */
8655 *get_invlist_offset_addr(dest) = src_offset;
8656 invlist_set_len(dest, src_len, src_offset);
8657 *get_invlist_previous_index_addr(dest) = 0;
8658 invlist_iterfinish(dest);
8661 PERL_STATIC_INLINE IV*
8662 S_get_invlist_previous_index_addr(SV* invlist)
8664 /* Return the address of the IV that is reserved to hold the cached index
8666 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
8668 assert(is_invlist(invlist));
8670 return &(((XINVLIST*) SvANY(invlist))->prev_index);
8673 PERL_STATIC_INLINE IV
8674 S_invlist_previous_index(SV* const invlist)
8676 /* Returns cached index of previous search */
8678 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
8680 return *get_invlist_previous_index_addr(invlist);
8683 PERL_STATIC_INLINE void
8684 S_invlist_set_previous_index(SV* const invlist, const IV index)
8686 /* Caches <index> for later retrieval */
8688 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
8690 assert(index == 0 || index < (int) _invlist_len(invlist));
8692 *get_invlist_previous_index_addr(invlist) = index;
8695 PERL_STATIC_INLINE void
8696 S_invlist_trim(SV* invlist)
8698 /* Free the not currently-being-used space in an inversion list */
8700 /* But don't free up the space needed for the 0 UV that is always at the
8701 * beginning of the list, nor the trailing NUL */
8702 const UV min_size = TO_INTERNAL_SIZE(1) + 1;
8704 PERL_ARGS_ASSERT_INVLIST_TRIM;
8706 assert(is_invlist(invlist));
8708 SvPV_renew(invlist, MAX(min_size, SvCUR(invlist) + 1));
8711 PERL_STATIC_INLINE void
8712 S_invlist_clear(pTHX_ SV* invlist) /* Empty the inversion list */
8714 PERL_ARGS_ASSERT_INVLIST_CLEAR;
8716 assert(is_invlist(invlist));
8718 invlist_set_len(invlist, 0, 0);
8719 invlist_trim(invlist);
8722 #endif /* ifndef PERL_IN_XSUB_RE */
8724 PERL_STATIC_INLINE bool
8725 S_invlist_is_iterating(SV* const invlist)
8727 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8729 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
8732 #ifndef PERL_IN_XSUB_RE
8734 PERL_STATIC_INLINE UV
8735 S_invlist_max(SV* const invlist)
8737 /* Returns the maximum number of elements storable in the inversion list's
8738 * array, without having to realloc() */
8740 PERL_ARGS_ASSERT_INVLIST_MAX;
8742 assert(is_invlist(invlist));
8744 /* Assumes worst case, in which the 0 element is not counted in the
8745 * inversion list, so subtracts 1 for that */
8746 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
8747 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
8748 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
8751 Perl__new_invlist(pTHX_ IV initial_size)
8754 /* Return a pointer to a newly constructed inversion list, with enough
8755 * space to store 'initial_size' elements. If that number is negative, a
8756 * system default is used instead */
8760 if (initial_size < 0) {
8764 /* Allocate the initial space */
8765 new_list = newSV_type(SVt_INVLIST);
8767 /* First 1 is in case the zero element isn't in the list; second 1 is for
8769 SvGROW(new_list, TO_INTERNAL_SIZE(initial_size + 1) + 1);
8770 invlist_set_len(new_list, 0, 0);
8772 /* Force iterinit() to be used to get iteration to work */
8773 *get_invlist_iter_addr(new_list) = (STRLEN) UV_MAX;
8775 *get_invlist_previous_index_addr(new_list) = 0;
8781 Perl__new_invlist_C_array(pTHX_ const UV* const list)
8783 /* Return a pointer to a newly constructed inversion list, initialized to
8784 * point to <list>, which has to be in the exact correct inversion list
8785 * form, including internal fields. Thus this is a dangerous routine that
8786 * should not be used in the wrong hands. The passed in 'list' contains
8787 * several header fields at the beginning that are not part of the
8788 * inversion list body proper */
8790 const STRLEN length = (STRLEN) list[0];
8791 const UV version_id = list[1];
8792 const bool offset = cBOOL(list[2]);
8793 #define HEADER_LENGTH 3
8794 /* If any of the above changes in any way, you must change HEADER_LENGTH
8795 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
8796 * perl -E 'say int(rand 2**31-1)'
8798 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
8799 data structure type, so that one being
8800 passed in can be validated to be an
8801 inversion list of the correct vintage.
8804 SV* invlist = newSV_type(SVt_INVLIST);
8806 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
8808 if (version_id != INVLIST_VERSION_ID) {
8809 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
8812 /* The generated array passed in includes header elements that aren't part
8813 * of the list proper, so start it just after them */
8814 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
8816 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
8817 shouldn't touch it */
8819 *(get_invlist_offset_addr(invlist)) = offset;
8821 /* The 'length' passed to us is the physical number of elements in the
8822 * inversion list. But if there is an offset the logical number is one
8824 invlist_set_len(invlist, length - offset, offset);
8826 invlist_set_previous_index(invlist, 0);
8828 /* Initialize the iteration pointer. */
8829 invlist_iterfinish(invlist);
8831 SvREADONLY_on(invlist);
8837 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
8839 /* Grow the maximum size of an inversion list */
8841 PERL_ARGS_ASSERT_INVLIST_EXTEND;
8843 assert(is_invlist(invlist));
8845 /* Add one to account for the zero element at the beginning which may not
8846 * be counted by the calling parameters */
8847 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max + 1));
8851 S__append_range_to_invlist(pTHX_ SV* const invlist,
8852 const UV start, const UV end)
8854 /* Subject to change or removal. Append the range from 'start' to 'end' at
8855 * the end of the inversion list. The range must be above any existing
8859 UV max = invlist_max(invlist);
8860 UV len = _invlist_len(invlist);
8863 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
8865 if (len == 0) { /* Empty lists must be initialized */
8866 offset = start != 0;
8867 array = _invlist_array_init(invlist, ! offset);
8870 /* Here, the existing list is non-empty. The current max entry in the
8871 * list is generally the first value not in the set, except when the
8872 * set extends to the end of permissible values, in which case it is
8873 * the first entry in that final set, and so this call is an attempt to
8874 * append out-of-order */
8876 UV final_element = len - 1;
8877 array = invlist_array(invlist);
8878 if ( array[final_element] > start
8879 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
8881 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",
8882 array[final_element], start,
8883 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
8886 /* Here, it is a legal append. If the new range begins 1 above the end
8887 * of the range below it, it is extending the range below it, so the
8888 * new first value not in the set is one greater than the newly
8889 * extended range. */
8890 offset = *get_invlist_offset_addr(invlist);
8891 if (array[final_element] == start) {
8892 if (end != UV_MAX) {
8893 array[final_element] = end + 1;
8896 /* But if the end is the maximum representable on the machine,
8897 * assume that infinity was actually what was meant. Just let
8898 * the range that this would extend to have no end */
8899 invlist_set_len(invlist, len - 1, offset);
8905 /* Here the new range doesn't extend any existing set. Add it */
8907 len += 2; /* Includes an element each for the start and end of range */
8909 /* If wll overflow the existing space, extend, which may cause the array to
8912 invlist_extend(invlist, len);
8914 /* Have to set len here to avoid assert failure in invlist_array() */
8915 invlist_set_len(invlist, len, offset);
8917 array = invlist_array(invlist);
8920 invlist_set_len(invlist, len, offset);
8923 /* The next item on the list starts the range, the one after that is
8924 * one past the new range. */
8925 array[len - 2] = start;
8926 if (end != UV_MAX) {
8927 array[len - 1] = end + 1;
8930 /* But if the end is the maximum representable on the machine, just let
8931 * the range have no end */
8932 invlist_set_len(invlist, len - 1, offset);
8937 Perl__invlist_search(SV* const invlist, const UV cp)
8939 /* Searches the inversion list for the entry that contains the input code
8940 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
8941 * return value is the index into the list's array of the range that
8942 * contains <cp>, that is, 'i' such that
8943 * array[i] <= cp < array[i+1]
8948 IV high = _invlist_len(invlist);
8949 const IV highest_element = high - 1;
8952 PERL_ARGS_ASSERT__INVLIST_SEARCH;
8954 /* If list is empty, return failure. */
8959 /* (We can't get the array unless we know the list is non-empty) */
8960 array = invlist_array(invlist);
8962 mid = invlist_previous_index(invlist);
8964 if (mid > highest_element) {
8965 mid = highest_element;
8968 /* <mid> contains the cache of the result of the previous call to this
8969 * function (0 the first time). See if this call is for the same result,
8970 * or if it is for mid-1. This is under the theory that calls to this
8971 * function will often be for related code points that are near each other.
8972 * And benchmarks show that caching gives better results. We also test
8973 * here if the code point is within the bounds of the list. These tests
8974 * replace others that would have had to be made anyway to make sure that
8975 * the array bounds were not exceeded, and these give us extra information
8976 * at the same time */
8977 if (cp >= array[mid]) {
8978 if (cp >= array[highest_element]) {
8979 return highest_element;
8982 /* Here, array[mid] <= cp < array[highest_element]. This means that
8983 * the final element is not the answer, so can exclude it; it also
8984 * means that <mid> is not the final element, so can refer to 'mid + 1'
8986 if (cp < array[mid + 1]) {
8992 else { /* cp < aray[mid] */
8993 if (cp < array[0]) { /* Fail if outside the array */
8997 if (cp >= array[mid - 1]) {
9002 /* Binary search. What we are looking for is <i> such that
9003 * array[i] <= cp < array[i+1]
9004 * The loop below converges on the i+1. Note that there may not be an
9005 * (i+1)th element in the array, and things work nonetheless */
9006 while (low < high) {
9007 mid = (low + high) / 2;
9008 assert(mid <= highest_element);
9009 if (array[mid] <= cp) { /* cp >= array[mid] */
9012 /* We could do this extra test to exit the loop early.
9013 if (cp < array[low]) {
9018 else { /* cp < array[mid] */
9025 invlist_set_previous_index(invlist, high);
9030 Perl__invlist_populate_swatch(SV* const invlist,
9031 const UV start, const UV end, U8* swatch)
9033 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
9034 * but is used when the swash has an inversion list. This makes this much
9035 * faster, as it uses a binary search instead of a linear one. This is
9036 * intimately tied to that function, and perhaps should be in utf8.c,
9037 * except it is intimately tied to inversion lists as well. It assumes
9038 * that <swatch> is all 0's on input */
9041 const IV len = _invlist_len(invlist);
9045 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
9047 if (len == 0) { /* Empty inversion list */
9051 array = invlist_array(invlist);
9053 /* Find which element it is */
9054 i = _invlist_search(invlist, start);
9056 /* We populate from <start> to <end> */
9057 while (current < end) {
9060 /* The inversion list gives the results for every possible code point
9061 * after the first one in the list. Only those ranges whose index is
9062 * even are ones that the inversion list matches. For the odd ones,
9063 * and if the initial code point is not in the list, we have to skip
9064 * forward to the next element */
9065 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
9067 if (i >= len) { /* Finished if beyond the end of the array */
9071 if (current >= end) { /* Finished if beyond the end of what we
9073 if (LIKELY(end < UV_MAX)) {
9077 /* We get here when the upper bound is the maximum
9078 * representable on the machine, and we are looking for just
9079 * that code point. Have to special case it */
9081 goto join_end_of_list;
9084 assert(current >= start);
9086 /* The current range ends one below the next one, except don't go past
9089 upper = (i < len && array[i] < end) ? array[i] : end;
9091 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
9092 * for each code point in it */
9093 for (; current < upper; current++) {
9094 const STRLEN offset = (STRLEN)(current - start);
9095 swatch[offset >> 3] |= 1 << (offset & 7);
9100 /* Quit if at the end of the list */
9103 /* But first, have to deal with the highest possible code point on
9104 * the platform. The previous code assumes that <end> is one
9105 * beyond where we want to populate, but that is impossible at the
9106 * platform's infinity, so have to handle it specially */
9107 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
9109 const STRLEN offset = (STRLEN)(end - start);
9110 swatch[offset >> 3] |= 1 << (offset & 7);
9115 /* Advance to the next range, which will be for code points not in the
9124 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9125 const bool complement_b, SV** output)
9127 /* Take the union of two inversion lists and point '*output' to it. On
9128 * input, '*output' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9129 * even 'a' or 'b'). If to an inversion list, the contents of the original
9130 * list will be replaced by the union. The first list, 'a', may be
9131 * NULL, in which case a copy of the second list is placed in '*output'.
9132 * If 'complement_b' is TRUE, the union is taken of the complement
9133 * (inversion) of 'b' instead of b itself.
9135 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9136 * Richard Gillam, published by Addison-Wesley, and explained at some
9137 * length there. The preface says to incorporate its examples into your
9138 * code at your own risk.
9140 * The algorithm is like a merge sort. */
9142 const UV* array_a; /* a's array */
9144 UV len_a; /* length of a's array */
9147 SV* u; /* the resulting union */
9151 UV i_a = 0; /* current index into a's array */
9155 /* running count, as explained in the algorithm source book; items are
9156 * stopped accumulating and are output when the count changes to/from 0.
9157 * The count is incremented when we start a range that's in an input's set,
9158 * and decremented when we start a range that's not in a set. So this
9159 * variable can be 0, 1, or 2. When it is 0 neither input is in their set,
9160 * and hence nothing goes into the union; 1, just one of the inputs is in
9161 * its set (and its current range gets added to the union); and 2 when both
9162 * inputs are in their sets. */
9165 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
9167 assert(*output == NULL || is_invlist(*output));
9169 len_b = _invlist_len(b);
9172 /* Here, 'b' is empty, hence it's complement is all possible code
9173 * points. So if the union includes the complement of 'b', it includes
9174 * everything, and we need not even look at 'a'. It's easiest to
9175 * create a new inversion list that matches everything. */
9177 SV* everything = _add_range_to_invlist(NULL, 0, UV_MAX);
9179 if (*output == NULL) { /* If the output didn't exist, just point it
9181 *output = everything;
9183 else { /* Otherwise, replace its contents with the new list */
9184 invlist_replace_list_destroys_src(*output, everything);
9185 SvREFCNT_dec_NN(everything);
9191 /* Here, we don't want the complement of 'b', and since 'b' is empty,
9192 * the union will come entirely from 'a'. If 'a' is NULL or empty, the
9193 * output will be empty */
9195 if (a == NULL || _invlist_len(a) == 0) {
9196 if (*output == NULL) {
9197 *output = _new_invlist(0);
9200 invlist_clear(*output);
9205 /* Here, 'a' is not empty, but 'b' is, so 'a' entirely determines the
9206 * union. We can just return a copy of 'a' if '*output' doesn't point
9207 * to an existing list */
9208 if (*output == NULL) {
9209 *output = invlist_clone(a);
9213 /* If the output is to overwrite 'a', we have a no-op, as it's
9219 /* Here, '*output' is to be overwritten by 'a' */
9220 u = invlist_clone(a);
9221 invlist_replace_list_destroys_src(*output, u);
9227 /* Here 'b' is not empty. See about 'a' */
9229 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
9231 /* Here, 'a' is empty (and b is not). That means the union will come
9232 * entirely from 'b'. If '*output' is NULL, we can directly return a
9233 * clone of 'b'. Otherwise, we replace the contents of '*output' with
9236 SV ** dest = (*output == NULL) ? output : &u;
9237 *dest = invlist_clone(b);
9239 _invlist_invert(*dest);
9243 invlist_replace_list_destroys_src(*output, u);
9250 /* Here both lists exist and are non-empty */
9251 array_a = invlist_array(a);
9252 array_b = invlist_array(b);
9254 /* If are to take the union of 'a' with the complement of b, set it
9255 * up so are looking at b's complement. */
9258 /* To complement, we invert: if the first element is 0, remove it. To
9259 * do this, we just pretend the array starts one later */
9260 if (array_b[0] == 0) {
9266 /* But if the first element is not zero, we pretend the list starts
9267 * at the 0 that is always stored immediately before the array. */
9273 /* Size the union for the worst case: that the sets are completely
9275 u = _new_invlist(len_a + len_b);
9277 /* Will contain U+0000 if either component does */
9278 array_u = _invlist_array_init(u, ( len_a > 0 && array_a[0] == 0)
9279 || (len_b > 0 && array_b[0] == 0));
9281 /* Go through each input list item by item, stopping when have exhausted
9283 while (i_a < len_a && i_b < len_b) {
9284 UV cp; /* The element to potentially add to the union's array */
9285 bool cp_in_set; /* is it in the the input list's set or not */
9287 /* We need to take one or the other of the two inputs for the union.
9288 * Since we are merging two sorted lists, we take the smaller of the
9289 * next items. In case of a tie, we take first the one that is in its
9290 * set. If we first took the one not in its set, it would decrement
9291 * the count, possibly to 0 which would cause it to be output as ending
9292 * the range, and the next time through we would take the same number,
9293 * and output it again as beginning the next range. By doing it the
9294 * opposite way, there is no possibility that the count will be
9295 * momentarily decremented to 0, and thus the two adjoining ranges will
9296 * be seamlessly merged. (In a tie and both are in the set or both not
9297 * in the set, it doesn't matter which we take first.) */
9298 if ( array_a[i_a] < array_b[i_b]
9299 || ( array_a[i_a] == array_b[i_b]
9300 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9302 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9303 cp = array_a[i_a++];
9306 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9307 cp = array_b[i_b++];
9310 /* Here, have chosen which of the two inputs to look at. Only output
9311 * if the running count changes to/from 0, which marks the
9312 * beginning/end of a range that's in the set */
9315 array_u[i_u++] = cp;
9322 array_u[i_u++] = cp;
9328 /* The loop above increments the index into exactly one of the input lists
9329 * each iteration, and ends when either index gets to its list end. That
9330 * means the other index is lower than its end, and so something is
9331 * remaining in that one. We decrement 'count', as explained below, if
9332 * that list is in its set. (i_a and i_b each currently index the element
9333 * beyond the one we care about.) */
9334 if ( (i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9335 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9340 /* Above we decremented 'count' if the list that had unexamined elements in
9341 * it was in its set. This has made it so that 'count' being non-zero
9342 * means there isn't anything left to output; and 'count' equal to 0 means
9343 * that what is left to output is precisely that which is left in the
9344 * non-exhausted input list.
9346 * To see why, note first that the exhausted input obviously has nothing
9347 * left to add to the union. If it was in its set at its end, that means
9348 * the set extends from here to the platform's infinity, and hence so does
9349 * the union and the non-exhausted set is irrelevant. The exhausted set
9350 * also contributed 1 to 'count'. If 'count' was 2, it got decremented to
9351 * 1, but if it was 1, the non-exhausted set wasn't in its set, and so
9352 * 'count' remains at 1. This is consistent with the decremented 'count'
9353 * != 0 meaning there's nothing left to add to the union.
9355 * But if the exhausted input wasn't in its set, it contributed 0 to
9356 * 'count', and the rest of the union will be whatever the other input is.
9357 * If 'count' was 0, neither list was in its set, and 'count' remains 0;
9358 * otherwise it gets decremented to 0. This is consistent with 'count'
9359 * == 0 meaning the remainder of the union is whatever is left in the
9360 * non-exhausted list. */
9365 IV copy_count = len_a - i_a;
9366 if (copy_count > 0) { /* The non-exhausted input is 'a' */
9367 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
9369 else { /* The non-exhausted input is b */
9370 copy_count = len_b - i_b;
9371 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
9373 len_u = i_u + copy_count;
9376 /* Set the result to the final length, which can change the pointer to
9377 * array_u, so re-find it. (Note that it is unlikely that this will
9378 * change, as we are shrinking the space, not enlarging it) */
9379 if (len_u != _invlist_len(u)) {
9380 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
9382 array_u = invlist_array(u);
9385 if (*output == NULL) { /* Simply return the new inversion list */
9389 /* Otherwise, overwrite the inversion list that was in '*output'. We
9390 * could instead free '*output', and then set it to 'u', but experience
9391 * has shown [perl #127392] that if the input is a mortal, we can get a
9392 * huge build-up of these during regex compilation before they get
9394 invlist_replace_list_destroys_src(*output, u);
9402 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9403 const bool complement_b, SV** i)
9405 /* Take the intersection of two inversion lists and point '*i' to it. On
9406 * input, '*i' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9407 * even 'a' or 'b'). If to an inversion list, the contents of the original
9408 * list will be replaced by the intersection. The first list, 'a', may be
9409 * NULL, in which case '*i' will be an empty list. If 'complement_b' is
9410 * TRUE, the result will be the intersection of 'a' and the complement (or
9411 * inversion) of 'b' instead of 'b' directly.
9413 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9414 * Richard Gillam, published by Addison-Wesley, and explained at some
9415 * length there. The preface says to incorporate its examples into your
9416 * code at your own risk. In fact, it had bugs
9418 * The algorithm is like a merge sort, and is essentially the same as the
9422 const UV* array_a; /* a's array */
9424 UV len_a; /* length of a's array */
9427 SV* r; /* the resulting intersection */
9431 UV i_a = 0; /* current index into a's array */
9435 /* running count of how many of the two inputs are postitioned at ranges
9436 * that are in their sets. As explained in the algorithm source book,
9437 * items are stopped accumulating and are output when the count changes
9438 * to/from 2. The count is incremented when we start a range that's in an
9439 * input's set, and decremented when we start a range that's not in a set.
9440 * Only when it is 2 are we in the intersection. */
9443 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
9445 assert(*i == NULL || is_invlist(*i));
9447 /* Special case if either one is empty */
9448 len_a = (a == NULL) ? 0 : _invlist_len(a);
9449 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
9450 if (len_a != 0 && complement_b) {
9452 /* Here, 'a' is not empty, therefore from the enclosing 'if', 'b'
9453 * must be empty. Here, also we are using 'b's complement, which
9454 * hence must be every possible code point. Thus the intersection
9457 if (*i == a) { /* No-op */
9462 *i = invlist_clone(a);
9466 r = invlist_clone(a);
9467 invlist_replace_list_destroys_src(*i, r);
9472 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
9473 * intersection must be empty */
9475 *i = _new_invlist(0);
9483 /* Here both lists exist and are non-empty */
9484 array_a = invlist_array(a);
9485 array_b = invlist_array(b);
9487 /* If are to take the intersection of 'a' with the complement of b, set it
9488 * up so are looking at b's complement. */
9491 /* To complement, we invert: if the first element is 0, remove it. To
9492 * do this, we just pretend the array starts one later */
9493 if (array_b[0] == 0) {
9499 /* But if the first element is not zero, we pretend the list starts
9500 * at the 0 that is always stored immediately before the array. */
9506 /* Size the intersection for the worst case: that the intersection ends up
9507 * fragmenting everything to be completely disjoint */
9508 r= _new_invlist(len_a + len_b);
9510 /* Will contain U+0000 iff both components do */
9511 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
9512 && len_b > 0 && array_b[0] == 0);
9514 /* Go through each list item by item, stopping when have exhausted one of
9516 while (i_a < len_a && i_b < len_b) {
9517 UV cp; /* The element to potentially add to the intersection's
9519 bool cp_in_set; /* Is it in the input list's set or not */
9521 /* We need to take one or the other of the two inputs for the
9522 * intersection. Since we are merging two sorted lists, we take the
9523 * smaller of the next items. In case of a tie, we take first the one
9524 * that is not in its set (a difference from the union algorithm). If
9525 * we first took the one in its set, it would increment the count,
9526 * possibly to 2 which would cause it to be output as starting a range
9527 * in the intersection, and the next time through we would take that
9528 * same number, and output it again as ending the set. By doing the
9529 * opposite of this, there is no possibility that the count will be
9530 * momentarily incremented to 2. (In a tie and both are in the set or
9531 * both not in the set, it doesn't matter which we take first.) */
9532 if ( array_a[i_a] < array_b[i_b]
9533 || ( array_a[i_a] == array_b[i_b]
9534 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9536 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9537 cp = array_a[i_a++];
9540 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9544 /* Here, have chosen which of the two inputs to look at. Only output
9545 * if the running count changes to/from 2, which marks the
9546 * beginning/end of a range that's in the intersection */
9550 array_r[i_r++] = cp;
9555 array_r[i_r++] = cp;
9562 /* The loop above increments the index into exactly one of the input lists
9563 * each iteration, and ends when either index gets to its list end. That
9564 * means the other index is lower than its end, and so something is
9565 * remaining in that one. We increment 'count', as explained below, if the
9566 * exhausted list was in its set. (i_a and i_b each currently index the
9567 * element beyond the one we care about.) */
9568 if ( (i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9569 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9574 /* Above we incremented 'count' if the exhausted list was in its set. This
9575 * has made it so that 'count' being below 2 means there is nothing left to
9576 * output; otheriwse what's left to add to the intersection is precisely
9577 * that which is left in the non-exhausted input list.
9579 * To see why, note first that the exhausted input obviously has nothing
9580 * left to affect the intersection. If it was in its set at its end, that
9581 * means the set extends from here to the platform's infinity, and hence
9582 * anything in the non-exhausted's list will be in the intersection, and
9583 * anything not in it won't be. Hence, the rest of the intersection is
9584 * precisely what's in the non-exhausted list The exhausted set also
9585 * contributed 1 to 'count', meaning 'count' was at least 1. Incrementing
9586 * it means 'count' is now at least 2. This is consistent with the
9587 * incremented 'count' being >= 2 means to add the non-exhausted list to
9590 * But if the exhausted input wasn't in its set, it contributed 0 to
9591 * 'count', and the intersection can't include anything further; the
9592 * non-exhausted set is irrelevant. 'count' was at most 1, and doesn't get
9593 * incremented. This is consistent with 'count' being < 2 meaning nothing
9594 * further to add to the intersection. */
9595 if (count < 2) { /* Nothing left to put in the intersection. */
9598 else { /* copy the non-exhausted list, unchanged. */
9599 IV copy_count = len_a - i_a;
9600 if (copy_count > 0) { /* a is the one with stuff left */
9601 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
9603 else { /* b is the one with stuff left */
9604 copy_count = len_b - i_b;
9605 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
9607 len_r = i_r + copy_count;
9610 /* Set the result to the final length, which can change the pointer to
9611 * array_r, so re-find it. (Note that it is unlikely that this will
9612 * change, as we are shrinking the space, not enlarging it) */
9613 if (len_r != _invlist_len(r)) {
9614 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
9616 array_r = invlist_array(r);
9619 if (*i == NULL) { /* Simply return the calculated intersection */
9622 else { /* Otherwise, replace the existing inversion list in '*i'. We could
9623 instead free '*i', and then set it to 'r', but experience has
9624 shown [perl #127392] that if the input is a mortal, we can get a
9625 huge build-up of these during regex compilation before they get
9628 invlist_replace_list_destroys_src(*i, r);
9640 Perl__add_range_to_invlist(pTHX_ SV* invlist, UV start, UV end)
9642 /* Add the range from 'start' to 'end' inclusive to the inversion list's
9643 * set. A pointer to the inversion list is returned. This may actually be
9644 * a new list, in which case the passed in one has been destroyed. The
9645 * passed-in inversion list can be NULL, in which case a new one is created
9646 * with just the one range in it. The new list is not necessarily
9647 * NUL-terminated. Space is not freed if the inversion list shrinks as a
9648 * result of this function. The gain would not be large, and in many
9649 * cases, this is called multiple times on a single inversion list, so
9650 * anything freed may almost immediately be needed again.
9652 * This used to mostly call the 'union' routine, but that is much more
9653 * heavyweight than really needed for a single range addition */
9655 UV* array; /* The array implementing the inversion list */
9656 UV len; /* How many elements in 'array' */
9657 SSize_t i_s; /* index into the invlist array where 'start'
9659 SSize_t i_e = 0; /* And the index where 'end' should go */
9660 UV cur_highest; /* The highest code point in the inversion list
9661 upon entry to this function */
9663 /* This range becomes the whole inversion list if none already existed */
9664 if (invlist == NULL) {
9665 invlist = _new_invlist(2);
9666 _append_range_to_invlist(invlist, start, end);
9670 /* Likewise, if the inversion list is currently empty */
9671 len = _invlist_len(invlist);
9673 _append_range_to_invlist(invlist, start, end);
9677 /* Starting here, we have to know the internals of the list */
9678 array = invlist_array(invlist);
9680 /* If the new range ends higher than the current highest ... */
9681 cur_highest = invlist_highest(invlist);
9682 if (end > cur_highest) {
9684 /* If the whole range is higher, we can just append it */
9685 if (start > cur_highest) {
9686 _append_range_to_invlist(invlist, start, end);
9690 /* Otherwise, add the portion that is higher ... */
9691 _append_range_to_invlist(invlist, cur_highest + 1, end);
9693 /* ... and continue on below to handle the rest. As a result of the
9694 * above append, we know that the index of the end of the range is the
9695 * final even numbered one of the array. Recall that the final element
9696 * always starts a range that extends to infinity. If that range is in
9697 * the set (meaning the set goes from here to infinity), it will be an
9698 * even index, but if it isn't in the set, it's odd, and the final
9699 * range in the set is one less, which is even. */
9700 if (end == UV_MAX) {
9708 /* We have dealt with appending, now see about prepending. If the new
9709 * range starts lower than the current lowest ... */
9710 if (start < array[0]) {
9712 /* Adding something which has 0 in it is somewhat tricky, and uncommon.
9713 * Let the union code handle it, rather than having to know the
9714 * trickiness in two code places. */
9715 if (UNLIKELY(start == 0)) {
9718 range_invlist = _new_invlist(2);
9719 _append_range_to_invlist(range_invlist, start, end);
9721 _invlist_union(invlist, range_invlist, &invlist);
9723 SvREFCNT_dec_NN(range_invlist);
9728 /* If the whole new range comes before the first entry, and doesn't
9729 * extend it, we have to insert it as an additional range */
9730 if (end < array[0] - 1) {
9732 goto splice_in_new_range;
9735 /* Here the new range adjoins the existing first range, extending it
9739 /* And continue on below to handle the rest. We know that the index of
9740 * the beginning of the range is the first one of the array */
9743 else { /* Not prepending any part of the new range to the existing list.
9744 * Find where in the list it should go. This finds i_s, such that:
9745 * invlist[i_s] <= start < array[i_s+1]
9747 i_s = _invlist_search(invlist, start);
9750 /* At this point, any extending before the beginning of the inversion list
9751 * and/or after the end has been done. This has made it so that, in the
9752 * code below, each endpoint of the new range is either in a range that is
9753 * in the set, or is in a gap between two ranges that are. This means we
9754 * don't have to worry about exceeding the array bounds.
9756 * Find where in the list the new range ends (but we can skip this if we
9757 * have already determined what it is, or if it will be the same as i_s,
9758 * which we already have computed) */
9760 i_e = (start == end)
9762 : _invlist_search(invlist, end);
9765 /* Here generally invlist[i_e] <= end < array[i_e+1]. But if invlist[i_e]
9766 * is a range that goes to infinity there is no element at invlist[i_e+1],
9767 * so only the first relation holds. */
9769 if ( ! ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
9771 /* Here, the ranges on either side of the beginning of the new range
9772 * are in the set, and this range starts in the gap between them.
9774 * The new range extends the range above it downwards if the new range
9775 * ends at or above that range's start */
9776 const bool extends_the_range_above = ( end == UV_MAX
9777 || end + 1 >= array[i_s+1]);
9779 /* The new range extends the range below it upwards if it begins just
9780 * after where that range ends */
9781 if (start == array[i_s]) {
9783 /* If the new range fills the entire gap between the other ranges,
9784 * they will get merged together. Other ranges may also get
9785 * merged, depending on how many of them the new range spans. In
9786 * the general case, we do the merge later, just once, after we
9787 * figure out how many to merge. But in the case where the new
9788 * range exactly spans just this one gap (possibly extending into
9789 * the one above), we do the merge here, and an early exit. This
9790 * is done here to avoid having to special case later. */
9791 if (i_e - i_s <= 1) {
9793 /* If i_e - i_s == 1, it means that the new range terminates
9794 * within the range above, and hence 'extends_the_range_above'
9795 * must be true. (If the range above it extends to infinity,
9796 * 'i_s+2' will be above the array's limit, but 'len-i_s-2'
9797 * will be 0, so no harm done.) */
9798 if (extends_the_range_above) {
9799 Move(array + i_s + 2, array + i_s, len - i_s - 2, UV);
9800 invlist_set_len(invlist,
9802 *(get_invlist_offset_addr(invlist)));
9806 /* Here, i_e must == i_s. We keep them in sync, as they apply
9807 * to the same range, and below we are about to decrement i_s
9812 /* Here, the new range is adjacent to the one below. (It may also
9813 * span beyond the range above, but that will get resolved later.)
9814 * Extend the range below to include this one. */
9815 array[i_s] = (end == UV_MAX) ? UV_MAX : end + 1;
9819 else if (extends_the_range_above) {
9821 /* Here the new range only extends the range above it, but not the
9822 * one below. It merges with the one above. Again, we keep i_e
9823 * and i_s in sync if they point to the same range */
9832 /* Here, we've dealt with the new range start extending any adjoining
9835 * If the new range extends to infinity, it is now the final one,
9836 * regardless of what was there before */
9837 if (UNLIKELY(end == UV_MAX)) {
9838 invlist_set_len(invlist, i_s + 1, *(get_invlist_offset_addr(invlist)));
9842 /* If i_e started as == i_s, it has also been dealt with,
9843 * and been updated to the new i_s, which will fail the following if */
9844 if (! ELEMENT_RANGE_MATCHES_INVLIST(i_e)) {
9846 /* Here, the ranges on either side of the end of the new range are in
9847 * the set, and this range ends in the gap between them.
9849 * If this range is adjacent to (hence extends) the range above it, it
9850 * becomes part of that range; likewise if it extends the range below,
9851 * it becomes part of that range */
9852 if (end + 1 == array[i_e+1]) {
9856 else if (start <= array[i_e]) {
9857 array[i_e] = end + 1;
9864 /* If the range fits entirely in an existing range (as possibly already
9865 * extended above), it doesn't add anything new */
9866 if (ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
9870 /* Here, no part of the range is in the list. Must add it. It will
9871 * occupy 2 more slots */
9872 splice_in_new_range:
9874 invlist_extend(invlist, len + 2);
9875 array = invlist_array(invlist);
9876 /* Move the rest of the array down two slots. Don't include any
9878 Move(array + i_e + 1, array + i_e + 3, len - i_e - 1, UV);
9880 /* Do the actual splice */
9881 array[i_e+1] = start;
9882 array[i_e+2] = end + 1;
9883 invlist_set_len(invlist, len + 2, *(get_invlist_offset_addr(invlist)));
9887 /* Here the new range crossed the boundaries of a pre-existing range. The
9888 * code above has adjusted things so that both ends are in ranges that are
9889 * in the set. This means everything in between must also be in the set.
9890 * Just squash things together */
9891 Move(array + i_e + 1, array + i_s + 1, len - i_e - 1, UV);
9892 invlist_set_len(invlist,
9894 *(get_invlist_offset_addr(invlist)));
9900 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
9901 UV** other_elements_ptr)
9903 /* Create and return an inversion list whose contents are to be populated
9904 * by the caller. The caller gives the number of elements (in 'size') and
9905 * the very first element ('element0'). This function will set
9906 * '*other_elements_ptr' to an array of UVs, where the remaining elements
9909 * Obviously there is some trust involved that the caller will properly
9910 * fill in the other elements of the array.
9912 * (The first element needs to be passed in, as the underlying code does
9913 * things differently depending on whether it is zero or non-zero) */
9915 SV* invlist = _new_invlist(size);
9918 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
9920 invlist = add_cp_to_invlist(invlist, element0);
9921 offset = *get_invlist_offset_addr(invlist);
9923 invlist_set_len(invlist, size, offset);
9924 *other_elements_ptr = invlist_array(invlist) + 1;
9930 PERL_STATIC_INLINE SV*
9931 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
9932 return _add_range_to_invlist(invlist, cp, cp);
9935 #ifndef PERL_IN_XSUB_RE
9937 Perl__invlist_invert(pTHX_ SV* const invlist)
9939 /* Complement the input inversion list. This adds a 0 if the list didn't
9940 * have a zero; removes it otherwise. As described above, the data
9941 * structure is set up so that this is very efficient */
9943 PERL_ARGS_ASSERT__INVLIST_INVERT;
9945 assert(! invlist_is_iterating(invlist));
9947 /* The inverse of matching nothing is matching everything */
9948 if (_invlist_len(invlist) == 0) {
9949 _append_range_to_invlist(invlist, 0, UV_MAX);
9953 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
9958 PERL_STATIC_INLINE SV*
9959 S_invlist_clone(pTHX_ SV* const invlist)
9962 /* Return a new inversion list that is a copy of the input one, which is
9963 * unchanged. The new list will not be mortal even if the old one was. */
9965 /* Need to allocate extra space to accommodate Perl's addition of a
9966 * trailing NUL to SvPV's, since it thinks they are always strings */
9967 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
9968 STRLEN physical_length = SvCUR(invlist);
9969 bool offset = *(get_invlist_offset_addr(invlist));
9971 PERL_ARGS_ASSERT_INVLIST_CLONE;
9973 *(get_invlist_offset_addr(new_invlist)) = offset;
9974 invlist_set_len(new_invlist, _invlist_len(invlist), offset);
9975 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
9980 PERL_STATIC_INLINE STRLEN*
9981 S_get_invlist_iter_addr(SV* invlist)
9983 /* Return the address of the UV that contains the current iteration
9986 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
9988 assert(is_invlist(invlist));
9990 return &(((XINVLIST*) SvANY(invlist))->iterator);
9993 PERL_STATIC_INLINE void
9994 S_invlist_iterinit(SV* invlist) /* Initialize iterator for invlist */
9996 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
9998 *get_invlist_iter_addr(invlist) = 0;
10001 PERL_STATIC_INLINE void
10002 S_invlist_iterfinish(SV* invlist)
10004 /* Terminate iterator for invlist. This is to catch development errors.
10005 * Any iteration that is interrupted before completed should call this
10006 * function. Functions that add code points anywhere else but to the end
10007 * of an inversion list assert that they are not in the middle of an
10008 * iteration. If they were, the addition would make the iteration
10009 * problematical: if the iteration hadn't reached the place where things
10010 * were being added, it would be ok */
10012 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
10014 *get_invlist_iter_addr(invlist) = (STRLEN) UV_MAX;
10018 S_invlist_iternext(SV* invlist, UV* start, UV* end)
10020 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
10021 * This call sets in <*start> and <*end>, the next range in <invlist>.
10022 * Returns <TRUE> if successful and the next call will return the next
10023 * range; <FALSE> if was already at the end of the list. If the latter,
10024 * <*start> and <*end> are unchanged, and the next call to this function
10025 * will start over at the beginning of the list */
10027 STRLEN* pos = get_invlist_iter_addr(invlist);
10028 UV len = _invlist_len(invlist);
10031 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
10034 *pos = (STRLEN) UV_MAX; /* Force iterinit() to be required next time */
10038 array = invlist_array(invlist);
10040 *start = array[(*pos)++];
10046 *end = array[(*pos)++] - 1;
10052 PERL_STATIC_INLINE UV
10053 S_invlist_highest(SV* const invlist)
10055 /* Returns the highest code point that matches an inversion list. This API
10056 * has an ambiguity, as it returns 0 under either the highest is actually
10057 * 0, or if the list is empty. If this distinction matters to you, check
10058 * for emptiness before calling this function */
10060 UV len = _invlist_len(invlist);
10063 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
10069 array = invlist_array(invlist);
10071 /* The last element in the array in the inversion list always starts a
10072 * range that goes to infinity. That range may be for code points that are
10073 * matched in the inversion list, or it may be for ones that aren't
10074 * matched. In the latter case, the highest code point in the set is one
10075 * less than the beginning of this range; otherwise it is the final element
10076 * of this range: infinity */
10077 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
10079 : array[len - 1] - 1;
10083 S_invlist_contents(pTHX_ SV* const invlist, const bool traditional_style)
10085 /* Get the contents of an inversion list into a string SV so that they can
10086 * be printed out. If 'traditional_style' is TRUE, it uses the format
10087 * traditionally done for debug tracing; otherwise it uses a format
10088 * suitable for just copying to the output, with blanks between ranges and
10089 * a dash between range components */
10093 const char intra_range_delimiter = (traditional_style ? '\t' : '-');
10094 const char inter_range_delimiter = (traditional_style ? '\n' : ' ');
10096 if (traditional_style) {
10097 output = newSVpvs("\n");
10100 output = newSVpvs("");
10103 PERL_ARGS_ASSERT_INVLIST_CONTENTS;
10105 assert(! invlist_is_iterating(invlist));
10107 invlist_iterinit(invlist);
10108 while (invlist_iternext(invlist, &start, &end)) {
10109 if (end == UV_MAX) {
10110 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%cINFINITY%c",
10111 start, intra_range_delimiter,
10112 inter_range_delimiter);
10114 else if (end != start) {
10115 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c%04" UVXf "%c",
10117 intra_range_delimiter,
10118 end, inter_range_delimiter);
10121 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c",
10122 start, inter_range_delimiter);
10126 if (SvCUR(output) && ! traditional_style) {/* Get rid of trailing blank */
10127 SvCUR_set(output, SvCUR(output) - 1);
10133 #ifndef PERL_IN_XSUB_RE
10135 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
10136 const char * const indent, SV* const invlist)
10138 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
10139 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
10140 * the string 'indent'. The output looks like this:
10141 [0] 0x000A .. 0x000D
10143 [4] 0x2028 .. 0x2029
10144 [6] 0x3104 .. INFINITY
10145 * This means that the first range of code points matched by the list are
10146 * 0xA through 0xD; the second range contains only the single code point
10147 * 0x85, etc. An inversion list is an array of UVs. Two array elements
10148 * are used to define each range (except if the final range extends to
10149 * infinity, only a single element is needed). The array index of the
10150 * first element for the corresponding range is given in brackets. */
10155 PERL_ARGS_ASSERT__INVLIST_DUMP;
10157 if (invlist_is_iterating(invlist)) {
10158 Perl_dump_indent(aTHX_ level, file,
10159 "%sCan't dump inversion list because is in middle of iterating\n",
10164 invlist_iterinit(invlist);
10165 while (invlist_iternext(invlist, &start, &end)) {
10166 if (end == UV_MAX) {
10167 Perl_dump_indent(aTHX_ level, file,
10168 "%s[%" UVuf "] 0x%04" UVXf " .. INFINITY\n",
10169 indent, (UV)count, start);
10171 else if (end != start) {
10172 Perl_dump_indent(aTHX_ level, file,
10173 "%s[%" UVuf "] 0x%04" UVXf " .. 0x%04" UVXf "\n",
10174 indent, (UV)count, start, end);
10177 Perl_dump_indent(aTHX_ level, file, "%s[%" UVuf "] 0x%04" UVXf "\n",
10178 indent, (UV)count, start);
10186 #if defined(PERL_ARGS_ASSERT__INVLISTEQ) && !defined(PERL_IN_XSUB_RE)
10188 Perl__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
10190 /* Return a boolean as to if the two passed in inversion lists are
10191 * identical. The final argument, if TRUE, says to take the complement of
10192 * the second inversion list before doing the comparison */
10194 const UV* array_a = invlist_array(a);
10195 const UV* array_b = invlist_array(b);
10196 UV len_a = _invlist_len(a);
10197 UV len_b = _invlist_len(b);
10199 PERL_ARGS_ASSERT__INVLISTEQ;
10201 /* If are to compare 'a' with the complement of b, set it
10202 * up so are looking at b's complement. */
10203 if (complement_b) {
10205 /* The complement of nothing is everything, so <a> would have to have
10206 * just one element, starting at zero (ending at infinity) */
10208 return (len_a == 1 && array_a[0] == 0);
10210 else if (array_b[0] == 0) {
10212 /* Otherwise, to complement, we invert. Here, the first element is
10213 * 0, just remove it. To do this, we just pretend the array starts
10221 /* But if the first element is not zero, we pretend the list starts
10222 * at the 0 that is always stored immediately before the array. */
10228 return len_a == len_b
10229 && memEQ(array_a, array_b, len_a * sizeof(array_a[0]));
10235 * As best we can, determine the characters that can match the start of
10236 * the given EXACTF-ish node.
10238 * Returns the invlist as a new SV*; it is the caller's responsibility to
10239 * call SvREFCNT_dec() when done with it.
10242 S__make_exactf_invlist(pTHX_ RExC_state_t *pRExC_state, regnode *node)
10244 const U8 * s = (U8*)STRING(node);
10245 SSize_t bytelen = STR_LEN(node);
10247 /* Start out big enough for 2 separate code points */
10248 SV* invlist = _new_invlist(4);
10250 PERL_ARGS_ASSERT__MAKE_EXACTF_INVLIST;
10255 /* We punt and assume can match anything if the node begins
10256 * with a multi-character fold. Things are complicated. For
10257 * example, /ffi/i could match any of:
10258 * "\N{LATIN SMALL LIGATURE FFI}"
10259 * "\N{LATIN SMALL LIGATURE FF}I"
10260 * "F\N{LATIN SMALL LIGATURE FI}"
10261 * plus several other things; and making sure we have all the
10262 * possibilities is hard. */
10263 if (is_MULTI_CHAR_FOLD_latin1_safe(s, s + bytelen)) {
10264 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10267 /* Any Latin1 range character can potentially match any
10268 * other depending on the locale */
10269 if (OP(node) == EXACTFL) {
10270 _invlist_union(invlist, PL_Latin1, &invlist);
10273 /* But otherwise, it matches at least itself. We can
10274 * quickly tell if it has a distinct fold, and if so,
10275 * it matches that as well */
10276 invlist = add_cp_to_invlist(invlist, uc);
10277 if (IS_IN_SOME_FOLD_L1(uc))
10278 invlist = add_cp_to_invlist(invlist, PL_fold_latin1[uc]);
10281 /* Some characters match above-Latin1 ones under /i. This
10282 * is true of EXACTFL ones when the locale is UTF-8 */
10283 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(uc)
10284 && (! isASCII(uc) || (OP(node) != EXACTFAA
10285 && OP(node) != EXACTFAA_NO_TRIE)))
10287 add_above_Latin1_folds(pRExC_state, (U8) uc, &invlist);
10291 else { /* Pattern is UTF-8 */
10292 U8 folded[UTF8_MAX_FOLD_CHAR_EXPAND * UTF8_MAXBYTES_CASE + 1] = { '\0' };
10293 const U8* e = s + bytelen;
10296 fc = uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
10298 /* The only code points that aren't folded in a UTF EXACTFish
10299 * node are are the problematic ones in EXACTFL nodes */
10300 if (OP(node) == EXACTFL && is_PROBLEMATIC_LOCALE_FOLDEDS_START_cp(uc)) {
10301 /* We need to check for the possibility that this EXACTFL
10302 * node begins with a multi-char fold. Therefore we fold
10303 * the first few characters of it so that we can make that
10309 for (i = 0; i < UTF8_MAX_FOLD_CHAR_EXPAND && s < e; i++) {
10311 *(d++) = (U8) toFOLD(*s);
10312 if (fc < 0) { /* Save the first fold */
10319 UV fold = toFOLD_utf8_safe(s, e, d, &len);
10320 if (fc < 0) { /* Save the first fold */
10328 /* And set up so the code below that looks in this folded
10329 * buffer instead of the node's string */
10334 /* When we reach here 's' points to the fold of the first
10335 * character(s) of the node; and 'e' points to far enough along
10336 * the folded string to be just past any possible multi-char
10339 * Unlike the non-UTF-8 case, the macro for determining if a
10340 * string is a multi-char fold requires all the characters to
10341 * already be folded. This is because of all the complications
10342 * if not. Note that they are folded anyway, except in EXACTFL
10343 * nodes. Like the non-UTF case above, we punt if the node
10344 * begins with a multi-char fold */
10346 if (is_MULTI_CHAR_FOLD_utf8_safe(s, e)) {
10347 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10349 else { /* Single char fold */
10351 unsigned int first_folds_to;
10352 const unsigned int * remaining_folds_to_list;
10353 Size_t folds_to_count;
10355 /* It matches itself */
10356 invlist = add_cp_to_invlist(invlist, fc);
10358 /* ... plus all the things that fold to it, which are found in
10359 * PL_utf8_foldclosures */
10360 folds_to_count = _inverse_folds(fc, &first_folds_to,
10361 &remaining_folds_to_list);
10362 for (k = 0; k < folds_to_count; k++) {
10363 UV c = (k == 0) ? first_folds_to : remaining_folds_to_list[k-1];
10365 /* /aa doesn't allow folds between ASCII and non- */
10366 if ( (OP(node) == EXACTFAA || OP(node) == EXACTFAA_NO_TRIE)
10367 && isASCII(c) != isASCII(fc))
10372 invlist = add_cp_to_invlist(invlist, c);
10380 #undef HEADER_LENGTH
10381 #undef TO_INTERNAL_SIZE
10382 #undef FROM_INTERNAL_SIZE
10383 #undef INVLIST_VERSION_ID
10385 /* End of inversion list object */
10388 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
10390 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
10391 * constructs, and updates RExC_flags with them. On input, RExC_parse
10392 * should point to the first flag; it is updated on output to point to the
10393 * final ')' or ':'. There needs to be at least one flag, or this will
10396 /* for (?g), (?gc), and (?o) warnings; warning
10397 about (?c) will warn about (?g) -- japhy */
10399 #define WASTED_O 0x01
10400 #define WASTED_G 0x02
10401 #define WASTED_C 0x04
10402 #define WASTED_GC (WASTED_G|WASTED_C)
10403 I32 wastedflags = 0x00;
10404 U32 posflags = 0, negflags = 0;
10405 U32 *flagsp = &posflags;
10406 char has_charset_modifier = '\0';
10408 bool has_use_defaults = FALSE;
10409 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
10410 int x_mod_count = 0;
10412 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
10414 /* '^' as an initial flag sets certain defaults */
10415 if (UCHARAT(RExC_parse) == '^') {
10417 has_use_defaults = TRUE;
10418 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
10419 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
10420 ? REGEX_UNICODE_CHARSET
10421 : REGEX_DEPENDS_CHARSET);
10424 cs = get_regex_charset(RExC_flags);
10425 if (cs == REGEX_DEPENDS_CHARSET
10426 && (RExC_utf8 || RExC_uni_semantics))
10428 cs = REGEX_UNICODE_CHARSET;
10431 while (RExC_parse < RExC_end) {
10432 /* && strchr("iogcmsx", *RExC_parse) */
10433 /* (?g), (?gc) and (?o) are useless here
10434 and must be globally applied -- japhy */
10435 switch (*RExC_parse) {
10437 /* Code for the imsxn flags */
10438 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp, x_mod_count);
10440 case LOCALE_PAT_MOD:
10441 if (has_charset_modifier) {
10442 goto excess_modifier;
10444 else if (flagsp == &negflags) {
10447 cs = REGEX_LOCALE_CHARSET;
10448 has_charset_modifier = LOCALE_PAT_MOD;
10450 case UNICODE_PAT_MOD:
10451 if (has_charset_modifier) {
10452 goto excess_modifier;
10454 else if (flagsp == &negflags) {
10457 cs = REGEX_UNICODE_CHARSET;
10458 has_charset_modifier = UNICODE_PAT_MOD;
10460 case ASCII_RESTRICT_PAT_MOD:
10461 if (flagsp == &negflags) {
10464 if (has_charset_modifier) {
10465 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
10466 goto excess_modifier;
10468 /* Doubled modifier implies more restricted */
10469 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
10472 cs = REGEX_ASCII_RESTRICTED_CHARSET;
10474 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
10476 case DEPENDS_PAT_MOD:
10477 if (has_use_defaults) {
10478 goto fail_modifiers;
10480 else if (flagsp == &negflags) {
10483 else if (has_charset_modifier) {
10484 goto excess_modifier;
10487 /* The dual charset means unicode semantics if the
10488 * pattern (or target, not known until runtime) are
10489 * utf8, or something in the pattern indicates unicode
10491 cs = (RExC_utf8 || RExC_uni_semantics)
10492 ? REGEX_UNICODE_CHARSET
10493 : REGEX_DEPENDS_CHARSET;
10494 has_charset_modifier = DEPENDS_PAT_MOD;
10498 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
10499 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
10501 else if (has_charset_modifier == *(RExC_parse - 1)) {
10502 vFAIL2("Regexp modifier \"%c\" may not appear twice",
10503 *(RExC_parse - 1));
10506 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
10508 NOT_REACHED; /*NOTREACHED*/
10511 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
10512 *(RExC_parse - 1));
10513 NOT_REACHED; /*NOTREACHED*/
10514 case ONCE_PAT_MOD: /* 'o' */
10515 case GLOBAL_PAT_MOD: /* 'g' */
10516 if (PASS2 && ckWARN(WARN_REGEXP)) {
10517 const I32 wflagbit = *RExC_parse == 'o'
10520 if (! (wastedflags & wflagbit) ) {
10521 wastedflags |= wflagbit;
10522 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10525 "Useless (%s%c) - %suse /%c modifier",
10526 flagsp == &negflags ? "?-" : "?",
10528 flagsp == &negflags ? "don't " : "",
10535 case CONTINUE_PAT_MOD: /* 'c' */
10536 if (PASS2 && ckWARN(WARN_REGEXP)) {
10537 if (! (wastedflags & WASTED_C) ) {
10538 wastedflags |= WASTED_GC;
10539 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10542 "Useless (%sc) - %suse /gc modifier",
10543 flagsp == &negflags ? "?-" : "?",
10544 flagsp == &negflags ? "don't " : ""
10549 case KEEPCOPY_PAT_MOD: /* 'p' */
10550 if (flagsp == &negflags) {
10552 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
10554 *flagsp |= RXf_PMf_KEEPCOPY;
10558 /* A flag is a default iff it is following a minus, so
10559 * if there is a minus, it means will be trying to
10560 * re-specify a default which is an error */
10561 if (has_use_defaults || flagsp == &negflags) {
10562 goto fail_modifiers;
10564 flagsp = &negflags;
10565 wastedflags = 0; /* reset so (?g-c) warns twice */
10571 if ((posflags & (RXf_PMf_EXTENDED|RXf_PMf_EXTENDED_MORE)) == RXf_PMf_EXTENDED) {
10572 negflags |= RXf_PMf_EXTENDED_MORE;
10574 RExC_flags |= posflags;
10576 if (negflags & RXf_PMf_EXTENDED) {
10577 negflags |= RXf_PMf_EXTENDED_MORE;
10579 RExC_flags &= ~negflags;
10580 set_regex_charset(&RExC_flags, cs);
10585 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10586 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10587 vFAIL2utf8f("Sequence (%" UTF8f "...) not recognized",
10588 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
10589 NOT_REACHED; /*NOTREACHED*/
10592 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10595 vFAIL("Sequence (?... not terminated");
10599 - reg - regular expression, i.e. main body or parenthesized thing
10601 * Caller must absorb opening parenthesis.
10603 * Combining parenthesis handling with the base level of regular expression
10604 * is a trifle forced, but the need to tie the tails of the branches to what
10605 * follows makes it hard to avoid.
10607 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
10609 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
10611 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
10614 PERL_STATIC_INLINE regnode *
10615 S_handle_named_backref(pTHX_ RExC_state_t *pRExC_state,
10617 char * parse_start,
10622 char* name_start = RExC_parse;
10624 SV *sv_dat = reg_scan_name(pRExC_state, SIZE_ONLY
10625 ? REG_RSN_RETURN_NULL
10626 : REG_RSN_RETURN_DATA);
10627 GET_RE_DEBUG_FLAGS_DECL;
10629 PERL_ARGS_ASSERT_HANDLE_NAMED_BACKREF;
10631 if (RExC_parse == name_start || *RExC_parse != ch) {
10632 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
10633 vFAIL2("Sequence %.3s... not terminated",parse_start);
10637 num = add_data( pRExC_state, STR_WITH_LEN("S"));
10638 RExC_rxi->data->data[num]=(void*)sv_dat;
10639 SvREFCNT_inc_simple_void(sv_dat);
10642 ret = reganode(pRExC_state,
10645 : (ASCII_FOLD_RESTRICTED)
10647 : (AT_LEAST_UNI_SEMANTICS)
10653 *flagp |= HASWIDTH;
10655 Set_Node_Offset(ret, parse_start+1);
10656 Set_Node_Cur_Length(ret, parse_start);
10658 nextchar(pRExC_state);
10662 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
10663 flags. Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan
10664 needs to be restarted, or'd with NEED_UTF8 if the pattern needs to be
10665 upgraded to UTF-8. Otherwise would only return NULL if regbranch() returns
10666 NULL, which cannot happen. */
10668 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
10669 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
10670 * 2 is like 1, but indicates that nextchar() has been called to advance
10671 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
10672 * this flag alerts us to the need to check for that */
10674 regnode *ret = NULL; /* Will be the head of the group. */
10677 regnode *ender = NULL;
10680 U32 oregflags = RExC_flags;
10681 bool have_branch = 0;
10683 I32 freeze_paren = 0;
10684 I32 after_freeze = 0;
10685 I32 num; /* numeric backreferences */
10687 char * parse_start = RExC_parse; /* MJD */
10688 char * const oregcomp_parse = RExC_parse;
10690 GET_RE_DEBUG_FLAGS_DECL;
10692 PERL_ARGS_ASSERT_REG;
10693 DEBUG_PARSE("reg ");
10695 *flagp = 0; /* Tentatively. */
10697 /* Having this true makes it feasible to have a lot fewer tests for the
10698 * parse pointer being in scope. For example, we can write
10699 * while(isFOO(*RExC_parse)) RExC_parse++;
10701 * while(RExC_parse < RExC_end && isFOO(*RExC_parse)) RExC_parse++;
10703 assert(*RExC_end == '\0');
10705 /* Make an OPEN node, if parenthesized. */
10708 /* Under /x, space and comments can be gobbled up between the '(' and
10709 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
10710 * intervening space, as the sequence is a token, and a token should be
10712 bool has_intervening_patws = (paren == 2)
10713 && *(RExC_parse - 1) != '(';
10715 if (RExC_parse >= RExC_end) {
10716 vFAIL("Unmatched (");
10719 if (paren == 'r') { /* Atomic script run */
10723 else if ( *RExC_parse == '*') { /* (*VERB:ARG), (*construct:...) */
10724 char *start_verb = RExC_parse + 1;
10726 char *start_arg = NULL;
10727 unsigned char op = 0;
10728 int arg_required = 0;
10729 int internal_argval = -1; /* if >-1 we are not allowed an argument*/
10730 bool has_upper = FALSE;
10732 if (has_intervening_patws) {
10733 RExC_parse++; /* past the '*' */
10735 /* For strict backwards compatibility, don't change the message
10736 * now that we also have lowercase operands */
10737 if (isUPPER(*RExC_parse)) {
10738 vFAIL("In '(*VERB...)', the '(' and '*' must be adjacent");
10741 vFAIL("In '(*...)', the '(' and '*' must be adjacent");
10744 while (RExC_parse < RExC_end && *RExC_parse != ')' ) {
10745 if ( *RExC_parse == ':' ) {
10746 start_arg = RExC_parse + 1;
10750 if (isUPPER(*RExC_parse)) {
10756 RExC_parse += UTF8SKIP(RExC_parse);
10759 verb_len = RExC_parse - start_verb;
10761 if (RExC_parse >= RExC_end) {
10762 goto unterminated_verb_pattern;
10765 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10766 while ( RExC_parse < RExC_end && *RExC_parse != ')' ) {
10767 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10769 if ( RExC_parse >= RExC_end || *RExC_parse != ')' ) {
10770 unterminated_verb_pattern:
10772 vFAIL("Unterminated verb pattern argument");
10775 vFAIL("Unterminated '(*...' argument");
10779 if ( RExC_parse >= RExC_end || *RExC_parse != ')' ) {
10781 vFAIL("Unterminated verb pattern");
10784 vFAIL("Unterminated '(*...' construct");
10789 /* Here, we know that RExC_parse < RExC_end */
10791 switch ( *start_verb ) {
10792 case 'A': /* (*ACCEPT) */
10793 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
10795 internal_argval = RExC_nestroot;
10798 case 'C': /* (*COMMIT) */
10799 if ( memEQs(start_verb,verb_len,"COMMIT") )
10802 case 'F': /* (*FAIL) */
10803 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
10807 case ':': /* (*:NAME) */
10808 case 'M': /* (*MARK:NAME) */
10809 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
10814 case 'P': /* (*PRUNE) */
10815 if ( memEQs(start_verb,verb_len,"PRUNE") )
10818 case 'S': /* (*SKIP) */
10819 if ( memEQs(start_verb,verb_len,"SKIP") )
10822 case 'T': /* (*THEN) */
10823 /* [19:06] <TimToady> :: is then */
10824 if ( memEQs(start_verb,verb_len,"THEN") ) {
10826 RExC_seen |= REG_CUTGROUP_SEEN;
10830 if ( memEQs(start_verb, verb_len, "asr")
10831 || memEQs(start_verb, verb_len, "atomic_script_run"))
10833 paren = 'r'; /* Mnemonic: recursed run */
10836 else if (memEQs(start_verb, verb_len, "atomic")) {
10837 paren = 't'; /* AtOMIC */
10838 goto alpha_assertions;
10842 if ( memEQs(start_verb, verb_len, "plb")
10843 || memEQs(start_verb, verb_len, "positive_lookbehind"))
10846 goto lookbehind_alpha_assertions;
10848 else if ( memEQs(start_verb, verb_len, "pla")
10849 || memEQs(start_verb, verb_len, "positive_lookahead"))
10852 goto alpha_assertions;
10856 if ( memEQs(start_verb, verb_len, "nlb")
10857 || memEQs(start_verb, verb_len, "negative_lookbehind"))
10860 goto lookbehind_alpha_assertions;
10862 else if ( memEQs(start_verb, verb_len, "nla")
10863 || memEQs(start_verb, verb_len, "negative_lookahead"))
10866 goto alpha_assertions;
10870 if ( memEQs(start_verb, verb_len, "sr")
10871 || memEQs(start_verb, verb_len, "script_run"))
10879 /* This indicates Unicode rules. */
10880 REQUIRE_UNI_RULES(flagp, NULL);
10886 RExC_parse = start_arg;
10888 if (RExC_in_script_run) {
10890 /* Nested script runs are treated as no-ops, because
10891 * if the nested one fails, the outer one must as
10892 * well. It could fail sooner, and avoid (??{} with
10893 * side effects, but that is explicitly documented as
10894 * undefined behavior. */
10898 if (paren == 's') {
10903 /* But, the atomic part of a nested atomic script run
10904 * isn't a no-op, but can be treated just like a '(?>'
10910 /* By doing this here, we avoid extra warnings for nested
10913 Perl_ck_warner_d(aTHX_
10914 packWARN(WARN_EXPERIMENTAL__SCRIPT_RUN),
10915 "The script_run feature is experimental"
10916 REPORT_LOCATION, REPORT_LOCATION_ARGS(RExC_parse));
10920 if (paren == 's') {
10921 /* Here, we're starting a new regular script run */
10922 ret = reg_node(pRExC_state, SROPEN);
10923 RExC_in_script_run = 1;
10928 /* Here, we are starting an atomic script run. This is
10929 * handled by recursing to deal with the atomic portion
10930 * separately, enclosed in SROPEN ... SRCLOSE nodes */
10932 ret = reg_node(pRExC_state, SROPEN);
10934 RExC_in_script_run = 1;
10936 atomic = reg(pRExC_state, 'r', &flags, depth);
10937 if (flags & (RESTART_PASS1|NEED_UTF8)) {
10938 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
10942 REGTAIL(pRExC_state, ret, atomic);
10944 REGTAIL(pRExC_state, atomic,
10945 reg_node(pRExC_state, SRCLOSE));
10947 RExC_in_script_run = 0;
10953 lookbehind_alpha_assertions:
10954 RExC_seen |= REG_LOOKBEHIND_SEEN;
10955 RExC_in_lookbehind++;
10961 Perl_ck_warner_d(aTHX_
10962 packWARN(WARN_EXPERIMENTAL__ALPHA_ASSERTIONS),
10963 "The alpha_assertions feature is experimental"
10964 REPORT_LOCATION, REPORT_LOCATION_ARGS(RExC_parse));
10967 RExC_seen_zerolen++;
10973 /* An empty negative lookahead assertion simply is failure */
10974 if (paren == 'A' && RExC_parse == start_arg) {
10975 ret=reganode(pRExC_state, OPFAIL, 0);
10976 nextchar(pRExC_state);
10980 RExC_parse = start_arg;
10985 "'(*%" UTF8f "' requires a terminating ':'",
10986 UTF8fARG(UTF, verb_len, start_verb));
10987 NOT_REACHED; /*NOTREACHED*/
10989 } /* End of switch */
10991 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10992 if (has_upper || verb_len == 0) {
10994 "Unknown verb pattern '%" UTF8f "'",
10995 UTF8fARG(UTF, verb_len, start_verb));
10999 "Unknown '(*...)' construct '%" UTF8f "'",
11000 UTF8fARG(UTF, verb_len, start_verb));
11003 if ( RExC_parse == start_arg ) {
11006 if ( arg_required && !start_arg ) {
11007 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
11008 verb_len, start_verb);
11010 if (internal_argval == -1) {
11011 ret = reganode(pRExC_state, op, 0);
11013 ret = reg2Lanode(pRExC_state, op, 0, internal_argval);
11015 RExC_seen |= REG_VERBARG_SEEN;
11016 if ( ! SIZE_ONLY ) {
11018 SV *sv = newSVpvn( start_arg,
11019 RExC_parse - start_arg);
11020 ARG(ret) = add_data( pRExC_state,
11021 STR_WITH_LEN("S"));
11022 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
11027 if ( internal_argval != -1 )
11028 ARG2L_SET(ret, internal_argval);
11030 nextchar(pRExC_state);
11033 else if (*RExC_parse == '?') { /* (?...) */
11034 bool is_logical = 0;
11035 const char * const seqstart = RExC_parse;
11036 const char * endptr;
11037 if (has_intervening_patws) {
11039 vFAIL("In '(?...)', the '(' and '?' must be adjacent");
11042 RExC_parse++; /* past the '?' */
11043 paren = *RExC_parse; /* might be a trailing NUL, if not
11045 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11046 if (RExC_parse > RExC_end) {
11049 ret = NULL; /* For look-ahead/behind. */
11052 case 'P': /* (?P...) variants for those used to PCRE/Python */
11053 paren = *RExC_parse;
11054 if ( paren == '<') { /* (?P<...>) named capture */
11056 if (RExC_parse >= RExC_end) {
11057 vFAIL("Sequence (?P<... not terminated");
11059 goto named_capture;
11061 else if (paren == '>') { /* (?P>name) named recursion */
11063 if (RExC_parse >= RExC_end) {
11064 vFAIL("Sequence (?P>... not terminated");
11066 goto named_recursion;
11068 else if (paren == '=') { /* (?P=...) named backref */
11070 return handle_named_backref(pRExC_state, flagp,
11073 RExC_parse += SKIP_IF_CHAR(RExC_parse);
11074 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11075 vFAIL3("Sequence (%.*s...) not recognized",
11076 RExC_parse-seqstart, seqstart);
11077 NOT_REACHED; /*NOTREACHED*/
11078 case '<': /* (?<...) */
11079 if (*RExC_parse == '!')
11081 else if (*RExC_parse != '=')
11088 case '\'': /* (?'...') */
11089 name_start = RExC_parse;
11090 svname = reg_scan_name(pRExC_state,
11091 SIZE_ONLY /* reverse test from the others */
11092 ? REG_RSN_RETURN_NAME
11093 : REG_RSN_RETURN_NULL);
11094 if ( RExC_parse == name_start
11095 || RExC_parse >= RExC_end
11096 || *RExC_parse != paren)
11098 vFAIL2("Sequence (?%c... not terminated",
11099 paren=='>' ? '<' : paren);
11104 if (!svname) /* shouldn't happen */
11106 "panic: reg_scan_name returned NULL");
11107 if (!RExC_paren_names) {
11108 RExC_paren_names= newHV();
11109 sv_2mortal(MUTABLE_SV(RExC_paren_names));
11111 RExC_paren_name_list= newAV();
11112 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
11115 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
11117 sv_dat = HeVAL(he_str);
11119 /* croak baby croak */
11121 "panic: paren_name hash element allocation failed");
11122 } else if ( SvPOK(sv_dat) ) {
11123 /* (?|...) can mean we have dupes so scan to check
11124 its already been stored. Maybe a flag indicating
11125 we are inside such a construct would be useful,
11126 but the arrays are likely to be quite small, so
11127 for now we punt -- dmq */
11128 IV count = SvIV(sv_dat);
11129 I32 *pv = (I32*)SvPVX(sv_dat);
11131 for ( i = 0 ; i < count ; i++ ) {
11132 if ( pv[i] == RExC_npar ) {
11138 pv = (I32*)SvGROW(sv_dat,
11139 SvCUR(sv_dat) + sizeof(I32)+1);
11140 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
11141 pv[count] = RExC_npar;
11142 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
11145 (void)SvUPGRADE(sv_dat,SVt_PVNV);
11146 sv_setpvn(sv_dat, (char *)&(RExC_npar),
11149 SvIV_set(sv_dat, 1);
11152 /* Yes this does cause a memory leak in debugging Perls
11154 if (!av_store(RExC_paren_name_list,
11155 RExC_npar, SvREFCNT_inc(svname)))
11156 SvREFCNT_dec_NN(svname);
11159 /*sv_dump(sv_dat);*/
11161 nextchar(pRExC_state);
11163 goto capturing_parens;
11166 RExC_seen |= REG_LOOKBEHIND_SEEN;
11167 RExC_in_lookbehind++;
11169 if (RExC_parse >= RExC_end) {
11170 vFAIL("Sequence (?... not terminated");
11174 case '=': /* (?=...) */
11175 RExC_seen_zerolen++;
11177 case '!': /* (?!...) */
11178 RExC_seen_zerolen++;
11179 /* check if we're really just a "FAIL" assertion */
11180 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
11181 FALSE /* Don't force to /x */ );
11182 if (*RExC_parse == ')') {
11183 ret=reganode(pRExC_state, OPFAIL, 0);
11184 nextchar(pRExC_state);
11188 case '|': /* (?|...) */
11189 /* branch reset, behave like a (?:...) except that
11190 buffers in alternations share the same numbers */
11192 after_freeze = freeze_paren = RExC_npar;
11194 case ':': /* (?:...) */
11195 case '>': /* (?>...) */
11197 case '$': /* (?$...) */
11198 case '@': /* (?@...) */
11199 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
11201 case '0' : /* (?0) */
11202 case 'R' : /* (?R) */
11203 if (RExC_parse == RExC_end || *RExC_parse != ')')
11204 FAIL("Sequence (?R) not terminated");
11206 RExC_seen |= REG_RECURSE_SEEN;
11207 *flagp |= POSTPONED;
11208 goto gen_recurse_regop;
11210 /* named and numeric backreferences */
11211 case '&': /* (?&NAME) */
11212 parse_start = RExC_parse - 1;
11215 SV *sv_dat = reg_scan_name(pRExC_state,
11216 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
11217 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
11219 if (RExC_parse >= RExC_end || *RExC_parse != ')')
11220 vFAIL("Sequence (?&... not terminated");
11221 goto gen_recurse_regop;
11224 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
11226 vFAIL("Illegal pattern");
11228 goto parse_recursion;
11230 case '-': /* (?-1) */
11231 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
11232 RExC_parse--; /* rewind to let it be handled later */
11236 case '1': case '2': case '3': case '4': /* (?1) */
11237 case '5': case '6': case '7': case '8': case '9':
11238 RExC_parse = (char *) seqstart + 1; /* Point to the digit */
11241 bool is_neg = FALSE;
11243 parse_start = RExC_parse - 1; /* MJD */
11244 if (*RExC_parse == '-') {
11249 if (grok_atoUV(RExC_parse, &unum, &endptr)
11253 RExC_parse = (char*)endptr;
11257 /* Some limit for num? */
11261 if (*RExC_parse!=')')
11262 vFAIL("Expecting close bracket");
11265 if ( paren == '-' ) {
11267 Diagram of capture buffer numbering.
11268 Top line is the normal capture buffer numbers
11269 Bottom line is the negative indexing as from
11273 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
11277 num = RExC_npar + num;
11280 vFAIL("Reference to nonexistent group");
11282 } else if ( paren == '+' ) {
11283 num = RExC_npar + num - 1;
11285 /* We keep track how many GOSUB items we have produced.
11286 To start off the ARG2L() of the GOSUB holds its "id",
11287 which is used later in conjunction with RExC_recurse
11288 to calculate the offset we need to jump for the GOSUB,
11289 which it will store in the final representation.
11290 We have to defer the actual calculation until much later
11291 as the regop may move.
11294 ret = reg2Lanode(pRExC_state, GOSUB, num, RExC_recurse_count);
11296 if (num > (I32)RExC_rx->nparens) {
11298 vFAIL("Reference to nonexistent group");
11300 RExC_recurse_count++;
11301 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11302 "%*s%*s Recurse #%" UVuf " to %" IVdf "\n",
11303 22, "| |", (int)(depth * 2 + 1), "",
11304 (UV)ARG(ret), (IV)ARG2L(ret)));
11306 RExC_seen |= REG_RECURSE_SEEN;
11308 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
11309 Set_Node_Offset(ret, parse_start); /* MJD */
11311 *flagp |= POSTPONED;
11312 assert(*RExC_parse == ')');
11313 nextchar(pRExC_state);
11318 case '?': /* (??...) */
11320 if (*RExC_parse != '{') {
11321 RExC_parse += SKIP_IF_CHAR(RExC_parse);
11322 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11324 "Sequence (%" UTF8f "...) not recognized",
11325 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
11326 NOT_REACHED; /*NOTREACHED*/
11328 *flagp |= POSTPONED;
11332 case '{': /* (?{...}) */
11335 struct reg_code_block *cb;
11337 RExC_seen_zerolen++;
11339 if ( !pRExC_state->code_blocks
11340 || pRExC_state->code_index
11341 >= pRExC_state->code_blocks->count
11342 || pRExC_state->code_blocks->cb[pRExC_state->code_index].start
11343 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
11346 if (RExC_pm_flags & PMf_USE_RE_EVAL)
11347 FAIL("panic: Sequence (?{...}): no code block found\n");
11348 FAIL("Eval-group not allowed at runtime, use re 'eval'");
11350 /* this is a pre-compiled code block (?{...}) */
11351 cb = &pRExC_state->code_blocks->cb[pRExC_state->code_index];
11352 RExC_parse = RExC_start + cb->end;
11355 if (cb->src_regex) {
11356 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
11357 RExC_rxi->data->data[n] =
11358 (void*)SvREFCNT_inc((SV*)cb->src_regex);
11359 RExC_rxi->data->data[n+1] = (void*)o;
11362 n = add_data(pRExC_state,
11363 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
11364 RExC_rxi->data->data[n] = (void*)o;
11367 pRExC_state->code_index++;
11368 nextchar(pRExC_state);
11372 ret = reg_node(pRExC_state, LOGICAL);
11374 eval = reg2Lanode(pRExC_state, EVAL,
11377 /* for later propagation into (??{})
11379 RExC_flags & RXf_PMf_COMPILETIME
11384 REGTAIL(pRExC_state, ret, eval);
11385 /* deal with the length of this later - MJD */
11388 ret = reg2Lanode(pRExC_state, EVAL, n, 0);
11389 Set_Node_Length(ret, RExC_parse - parse_start + 1);
11390 Set_Node_Offset(ret, parse_start);
11393 case '(': /* (?(?{...})...) and (?(?=...)...) */
11396 const int DEFINE_len = sizeof("DEFINE") - 1;
11397 if ( RExC_parse < RExC_end - 1
11398 && ( ( RExC_parse[0] == '?' /* (?(?...)) */
11399 && ( RExC_parse[1] == '='
11400 || RExC_parse[1] == '!'
11401 || RExC_parse[1] == '<'
11402 || RExC_parse[1] == '{'))
11403 || ( RExC_parse[0] == '*' /* (?(*...)) */
11404 && ( memBEGINs(RExC_parse + 1,
11405 (Size_t) (RExC_end - (RExC_parse + 1)),
11407 || memBEGINs(RExC_parse + 1,
11408 (Size_t) (RExC_end - (RExC_parse + 1)),
11410 || memBEGINs(RExC_parse + 1,
11411 (Size_t) (RExC_end - (RExC_parse + 1)),
11413 || memBEGINs(RExC_parse + 1,
11414 (Size_t) (RExC_end - (RExC_parse + 1)),
11416 || memBEGINs(RExC_parse + 1,
11417 (Size_t) (RExC_end - (RExC_parse + 1)),
11418 "positive_lookahead:")
11419 || memBEGINs(RExC_parse + 1,
11420 (Size_t) (RExC_end - (RExC_parse + 1)),
11421 "positive_lookbehind:")
11422 || memBEGINs(RExC_parse + 1,
11423 (Size_t) (RExC_end - (RExC_parse + 1)),
11424 "negative_lookahead:")
11425 || memBEGINs(RExC_parse + 1,
11426 (Size_t) (RExC_end - (RExC_parse + 1)),
11427 "negative_lookbehind:"))))
11428 ) { /* Lookahead or eval. */
11432 ret = reg_node(pRExC_state, LOGICAL);
11436 tail = reg(pRExC_state, 1, &flag, depth+1);
11437 RETURN_NULL_ON_RESTART(flag,flagp);
11438 REGTAIL(pRExC_state, ret, tail);
11441 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
11442 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
11444 char ch = RExC_parse[0] == '<' ? '>' : '\'';
11445 char *name_start= RExC_parse++;
11447 SV *sv_dat=reg_scan_name(pRExC_state,
11448 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
11449 if ( RExC_parse == name_start
11450 || RExC_parse >= RExC_end
11451 || *RExC_parse != ch)
11453 vFAIL2("Sequence (?(%c... not terminated",
11454 (ch == '>' ? '<' : ch));
11458 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11459 RExC_rxi->data->data[num]=(void*)sv_dat;
11460 SvREFCNT_inc_simple_void(sv_dat);
11462 ret = reganode(pRExC_state,NGROUPP,num);
11463 goto insert_if_check_paren;
11465 else if (memBEGINs(RExC_parse,
11466 (STRLEN) (RExC_end - RExC_parse),
11469 ret = reganode(pRExC_state,DEFINEP,0);
11470 RExC_parse += DEFINE_len;
11472 goto insert_if_check_paren;
11474 else if (RExC_parse[0] == 'R') {
11476 /* parno == 0 => /(?(R)YES|NO)/ "in any form of recursion OR eval"
11477 * parno == 1 => /(?(R0)YES|NO)/ "in GOSUB (?0) / (?R)"
11478 * parno == 2 => /(?(R1)YES|NO)/ "in GOSUB (?1) (parno-1)"
11481 if (RExC_parse[0] == '0') {
11485 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
11488 if (grok_atoUV(RExC_parse, &uv, &endptr)
11491 parno = (I32)uv + 1;
11492 RExC_parse = (char*)endptr;
11494 /* else "Switch condition not recognized" below */
11495 } else if (RExC_parse[0] == '&') {
11498 sv_dat = reg_scan_name(pRExC_state,
11500 ? REG_RSN_RETURN_NULL
11501 : REG_RSN_RETURN_DATA);
11503 /* we should only have a false sv_dat when
11504 * SIZE_ONLY is true, and we always have false
11505 * sv_dat when SIZE_ONLY is true.
11506 * reg_scan_name() will VFAIL() if the name is
11507 * unknown when SIZE_ONLY is false, and otherwise
11508 * will return something, and when SIZE_ONLY is
11509 * true, reg_scan_name() just parses the string,
11510 * and doesnt return anything. (in theory) */
11511 assert(SIZE_ONLY ? !sv_dat : !!sv_dat);
11514 parno = 1 + *((I32 *)SvPVX(sv_dat));
11516 ret = reganode(pRExC_state,INSUBP,parno);
11517 goto insert_if_check_paren;
11519 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
11524 if (grok_atoUV(RExC_parse, &uv, &endptr)
11528 RExC_parse = (char*)endptr;
11531 vFAIL("panic: grok_atoUV returned FALSE");
11533 ret = reganode(pRExC_state, GROUPP, parno);
11535 insert_if_check_paren:
11536 if (UCHARAT(RExC_parse) != ')') {
11537 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11538 vFAIL("Switch condition not recognized");
11540 nextchar(pRExC_state);
11542 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
11543 br = regbranch(pRExC_state, &flags, 1,depth+1);
11545 RETURN_NULL_ON_RESTART(flags,flagp);
11546 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf,
11549 REGTAIL(pRExC_state, br, reganode(pRExC_state,
11551 c = UCHARAT(RExC_parse);
11552 nextchar(pRExC_state);
11553 if (flags&HASWIDTH)
11554 *flagp |= HASWIDTH;
11557 vFAIL("(?(DEFINE)....) does not allow branches");
11559 /* Fake one for optimizer. */
11560 lastbr = reganode(pRExC_state, IFTHEN, 0);
11562 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
11563 RETURN_NULL_ON_RESTART(flags,flagp);
11564 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf,
11567 REGTAIL(pRExC_state, ret, lastbr);
11568 if (flags&HASWIDTH)
11569 *flagp |= HASWIDTH;
11570 c = UCHARAT(RExC_parse);
11571 nextchar(pRExC_state);
11576 if (RExC_parse >= RExC_end)
11577 vFAIL("Switch (?(condition)... not terminated");
11579 vFAIL("Switch (?(condition)... contains too many branches");
11581 ender = reg_node(pRExC_state, TAIL);
11582 REGTAIL(pRExC_state, br, ender);
11584 REGTAIL(pRExC_state, lastbr, ender);
11585 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
11588 REGTAIL(pRExC_state, ret, ender);
11589 RExC_size++; /* XXX WHY do we need this?!!
11590 For large programs it seems to be required
11591 but I can't figure out why. -- dmq*/
11594 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11595 vFAIL("Unknown switch condition (?(...))");
11597 case '[': /* (?[ ... ]) */
11598 return handle_regex_sets(pRExC_state, NULL, flagp, depth+1,
11600 case 0: /* A NUL */
11601 RExC_parse--; /* for vFAIL to print correctly */
11602 vFAIL("Sequence (? incomplete");
11604 default: /* e.g., (?i) */
11605 RExC_parse = (char *) seqstart + 1;
11607 parse_lparen_question_flags(pRExC_state);
11608 if (UCHARAT(RExC_parse) != ':') {
11609 if (RExC_parse < RExC_end)
11610 nextchar(pRExC_state);
11615 nextchar(pRExC_state);
11621 if (*RExC_parse == '{' && PASS2) {
11622 ckWARNregdep(RExC_parse + 1,
11623 "Unescaped left brace in regex is "
11624 "deprecated here (and will be fatal "
11625 "in Perl 5.32), passed through");
11627 /* Not bothering to indent here, as the above 'else' is temporary
11629 if (!(RExC_flags & RXf_PMf_NOCAPTURE)) { /* (...) */
11634 ret = reganode(pRExC_state, OPEN, parno);
11636 if (!RExC_nestroot)
11637 RExC_nestroot = parno;
11638 if (RExC_open_parens && !RExC_open_parens[parno])
11640 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11641 "%*s%*s Setting open paren #%" IVdf " to %d\n",
11642 22, "| |", (int)(depth * 2 + 1), "",
11643 (IV)parno, REG_NODE_NUM(ret)));
11644 RExC_open_parens[parno]= ret;
11647 Set_Node_Length(ret, 1); /* MJD */
11648 Set_Node_Offset(ret, RExC_parse); /* MJD */
11651 /* with RXf_PMf_NOCAPTURE treat (...) as (?:...) */
11661 /* Pick up the branches, linking them together. */
11662 parse_start = RExC_parse; /* MJD */
11663 br = regbranch(pRExC_state, &flags, 1,depth+1);
11665 /* branch_len = (paren != 0); */
11668 RETURN_NULL_ON_RESTART(flags,flagp);
11669 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf, (UV) flags);
11671 if (*RExC_parse == '|') {
11672 if (!SIZE_ONLY && RExC_extralen) {
11673 reginsert(pRExC_state, BRANCHJ, br, depth+1);
11676 reginsert(pRExC_state, BRANCH, br, depth+1);
11677 Set_Node_Length(br, paren != 0);
11678 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
11682 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
11684 else if (paren == ':') {
11685 *flagp |= flags&SIMPLE;
11687 if (is_open) { /* Starts with OPEN. */
11688 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
11690 else if (paren != '?') /* Not Conditional */
11692 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
11694 while (*RExC_parse == '|') {
11695 if (!SIZE_ONLY && RExC_extralen) {
11696 ender = reganode(pRExC_state, LONGJMP,0);
11698 /* Append to the previous. */
11699 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
11702 RExC_extralen += 2; /* Account for LONGJMP. */
11703 nextchar(pRExC_state);
11704 if (freeze_paren) {
11705 if (RExC_npar > after_freeze)
11706 after_freeze = RExC_npar;
11707 RExC_npar = freeze_paren;
11709 br = regbranch(pRExC_state, &flags, 0, depth+1);
11712 RETURN_NULL_ON_RESTART(flags,flagp);
11713 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf, (UV) flags);
11715 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
11717 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
11720 if (have_branch || paren != ':') {
11721 /* Make a closing node, and hook it on the end. */
11724 ender = reg_node(pRExC_state, TAIL);
11727 ender = reganode(pRExC_state, CLOSE, parno);
11728 if ( RExC_close_parens ) {
11729 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11730 "%*s%*s Setting close paren #%" IVdf " to %d\n",
11731 22, "| |", (int)(depth * 2 + 1), "", (IV)parno, REG_NODE_NUM(ender)));
11732 RExC_close_parens[parno]= ender;
11733 if (RExC_nestroot == parno)
11736 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
11737 Set_Node_Length(ender,1); /* MJD */
11740 ender = reg_node(pRExC_state, SRCLOSE);
11741 RExC_in_script_run = 0;
11751 *flagp &= ~HASWIDTH;
11753 case 't': /* aTomic */
11755 ender = reg_node(pRExC_state, SUCCEED);
11758 ender = reg_node(pRExC_state, END);
11760 assert(!RExC_end_op); /* there can only be one! */
11761 RExC_end_op = ender;
11762 if (RExC_close_parens) {
11763 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11764 "%*s%*s Setting close paren #0 (END) to %d\n",
11765 22, "| |", (int)(depth * 2 + 1), "", REG_NODE_NUM(ender)));
11767 RExC_close_parens[0]= ender;
11772 DEBUG_PARSE_r(if (!SIZE_ONLY) {
11773 DEBUG_PARSE_MSG("lsbr");
11774 regprop(RExC_rx, RExC_mysv1, lastbr, NULL, pRExC_state);
11775 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
11776 Perl_re_printf( aTHX_ "~ tying lastbr %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
11777 SvPV_nolen_const(RExC_mysv1),
11778 (IV)REG_NODE_NUM(lastbr),
11779 SvPV_nolen_const(RExC_mysv2),
11780 (IV)REG_NODE_NUM(ender),
11781 (IV)(ender - lastbr)
11784 REGTAIL(pRExC_state, lastbr, ender);
11786 if (have_branch && !SIZE_ONLY) {
11787 char is_nothing= 1;
11789 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
11791 /* Hook the tails of the branches to the closing node. */
11792 for (br = ret; br; br = regnext(br)) {
11793 const U8 op = PL_regkind[OP(br)];
11794 if (op == BRANCH) {
11795 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
11796 if ( OP(NEXTOPER(br)) != NOTHING
11797 || regnext(NEXTOPER(br)) != ender)
11800 else if (op == BRANCHJ) {
11801 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
11802 /* for now we always disable this optimisation * /
11803 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
11804 || regnext(NEXTOPER(NEXTOPER(br))) != ender)
11810 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
11811 DEBUG_PARSE_r(if (!SIZE_ONLY) {
11812 DEBUG_PARSE_MSG("NADA");
11813 regprop(RExC_rx, RExC_mysv1, ret, NULL, pRExC_state);
11814 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
11815 Perl_re_printf( aTHX_ "~ converting ret %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
11816 SvPV_nolen_const(RExC_mysv1),
11817 (IV)REG_NODE_NUM(ret),
11818 SvPV_nolen_const(RExC_mysv2),
11819 (IV)REG_NODE_NUM(ender),
11824 if (OP(ender) == TAIL) {
11829 for ( opt= br + 1; opt < ender ; opt++ )
11830 OP(opt)= OPTIMIZED;
11831 NEXT_OFF(br)= ender - br;
11839 /* Even/odd or x=don't care: 010101x10x */
11840 static const char parens[] = "=!aA<,>Bbt";
11841 /* flag below is set to 0 up through 'A'; 1 for larger */
11843 if (paren && (p = strchr(parens, paren))) {
11844 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
11845 int flag = (p - parens) > 3;
11847 if (paren == '>' || paren == 't') {
11848 node = SUSPEND, flag = 0;
11851 reginsert(pRExC_state, node,ret, depth+1);
11852 Set_Node_Cur_Length(ret, parse_start);
11853 Set_Node_Offset(ret, parse_start + 1);
11855 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
11859 /* Check for proper termination. */
11861 /* restore original flags, but keep (?p) and, if we've changed from /d
11862 * rules to /u, keep the /u */
11863 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
11864 if (DEPENDS_SEMANTICS && RExC_uni_semantics) {
11865 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
11867 if (RExC_parse >= RExC_end || UCHARAT(RExC_parse) != ')') {
11868 RExC_parse = oregcomp_parse;
11869 vFAIL("Unmatched (");
11871 nextchar(pRExC_state);
11873 else if (!paren && RExC_parse < RExC_end) {
11874 if (*RExC_parse == ')') {
11876 vFAIL("Unmatched )");
11879 FAIL("Junk on end of regexp"); /* "Can't happen". */
11880 NOT_REACHED; /* NOTREACHED */
11883 if (RExC_in_lookbehind) {
11884 RExC_in_lookbehind--;
11886 if (after_freeze > RExC_npar)
11887 RExC_npar = after_freeze;
11892 - regbranch - one alternative of an | operator
11894 * Implements the concatenation operator.
11896 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
11897 * restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
11900 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
11903 regnode *chain = NULL;
11905 I32 flags = 0, c = 0;
11906 GET_RE_DEBUG_FLAGS_DECL;
11908 PERL_ARGS_ASSERT_REGBRANCH;
11910 DEBUG_PARSE("brnc");
11915 if (!SIZE_ONLY && RExC_extralen)
11916 ret = reganode(pRExC_state, BRANCHJ,0);
11918 ret = reg_node(pRExC_state, BRANCH);
11919 Set_Node_Length(ret, 1);
11923 if (!first && SIZE_ONLY)
11924 RExC_extralen += 1; /* BRANCHJ */
11926 *flagp = WORST; /* Tentatively. */
11928 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
11929 FALSE /* Don't force to /x */ );
11930 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
11931 flags &= ~TRYAGAIN;
11932 latest = regpiece(pRExC_state, &flags,depth+1);
11933 if (latest == NULL) {
11934 if (flags & TRYAGAIN)
11936 RETURN_NULL_ON_RESTART(flags,flagp);
11937 FAIL2("panic: regpiece returned NULL, flags=%#" UVxf, (UV) flags);
11939 else if (ret == NULL)
11941 *flagp |= flags&(HASWIDTH|POSTPONED);
11942 if (chain == NULL) /* First piece. */
11943 *flagp |= flags&SPSTART;
11945 /* FIXME adding one for every branch after the first is probably
11946 * excessive now we have TRIE support. (hv) */
11948 REGTAIL(pRExC_state, chain, latest);
11953 if (chain == NULL) { /* Loop ran zero times. */
11954 chain = reg_node(pRExC_state, NOTHING);
11959 *flagp |= flags&SIMPLE;
11966 - regpiece - something followed by possible quantifier * + ? {n,m}
11968 * Note that the branching code sequences used for ? and the general cases
11969 * of * and + are somewhat optimized: they use the same NOTHING node as
11970 * both the endmarker for their branch list and the body of the last branch.
11971 * It might seem that this node could be dispensed with entirely, but the
11972 * endmarker role is not redundant.
11974 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
11976 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
11977 * restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
11980 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
11986 const char * const origparse = RExC_parse;
11988 I32 max = REG_INFTY;
11989 #ifdef RE_TRACK_PATTERN_OFFSETS
11992 const char *maxpos = NULL;
11995 /* Save the original in case we change the emitted regop to a FAIL. */
11996 regnode * const orig_emit = RExC_emit;
11998 GET_RE_DEBUG_FLAGS_DECL;
12000 PERL_ARGS_ASSERT_REGPIECE;
12002 DEBUG_PARSE("piec");
12004 ret = regatom(pRExC_state, &flags,depth+1);
12006 RETURN_NULL_ON_RESTART_OR_FLAGS(flags,flagp,TRYAGAIN);
12007 FAIL2("panic: regatom returned NULL, flags=%#" UVxf, (UV) flags);
12012 if (op == '{' && regcurly(RExC_parse)) {
12014 #ifdef RE_TRACK_PATTERN_OFFSETS
12015 parse_start = RExC_parse; /* MJD */
12017 next = RExC_parse + 1;
12018 while (isDIGIT(*next) || *next == ',') {
12019 if (*next == ',') {
12027 if (*next == '}') { /* got one */
12028 const char* endptr;
12032 if (isDIGIT(*RExC_parse)) {
12034 if (!grok_atoUV(RExC_parse, &uv, &endptr))
12035 vFAIL("Invalid quantifier in {,}");
12036 if (uv >= REG_INFTY)
12037 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
12042 if (*maxpos == ',')
12045 maxpos = RExC_parse;
12046 if (isDIGIT(*maxpos)) {
12048 if (!grok_atoUV(maxpos, &uv, &endptr))
12049 vFAIL("Invalid quantifier in {,}");
12050 if (uv >= REG_INFTY)
12051 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
12054 max = REG_INFTY; /* meaning "infinity" */
12057 nextchar(pRExC_state);
12058 if (max < min) { /* If can't match, warn and optimize to fail
12060 reginsert(pRExC_state, OPFAIL, orig_emit, depth+1);
12062 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
12063 NEXT_OFF(orig_emit)= regarglen[OPFAIL] + NODE_STEP_REGNODE;
12067 else if (min == max && *RExC_parse == '?')
12070 ckWARN2reg(RExC_parse + 1,
12071 "Useless use of greediness modifier '%c'",
12077 if ((flags&SIMPLE)) {
12078 if (min == 0 && max == REG_INFTY) {
12079 reginsert(pRExC_state, STAR, ret, depth+1);
12081 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
12084 if (min == 1 && max == REG_INFTY) {
12085 reginsert(pRExC_state, PLUS, ret, depth+1);
12087 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
12090 MARK_NAUGHTY_EXP(2, 2);
12091 reginsert(pRExC_state, CURLY, ret, depth+1);
12092 Set_Node_Offset(ret, parse_start+1); /* MJD */
12093 Set_Node_Cur_Length(ret, parse_start);
12096 regnode * const w = reg_node(pRExC_state, WHILEM);
12099 REGTAIL(pRExC_state, ret, w);
12100 if (!SIZE_ONLY && RExC_extralen) {
12101 reginsert(pRExC_state, LONGJMP,ret, depth+1);
12102 reginsert(pRExC_state, NOTHING,ret, depth+1);
12103 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
12105 reginsert(pRExC_state, CURLYX,ret, depth+1);
12107 Set_Node_Offset(ret, parse_start+1);
12108 Set_Node_Length(ret,
12109 op == '{' ? (RExC_parse - parse_start) : 1);
12111 if (!SIZE_ONLY && RExC_extralen)
12112 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
12113 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
12115 RExC_whilem_seen++, RExC_extralen += 3;
12116 MARK_NAUGHTY_EXP(1, 4); /* compound interest */
12123 *flagp |= HASWIDTH;
12125 ARG1_SET(ret, (U16)min);
12126 ARG2_SET(ret, (U16)max);
12128 if (max == REG_INFTY)
12129 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
12135 if (!ISMULT1(op)) {
12140 #if 0 /* Now runtime fix should be reliable. */
12142 /* if this is reinstated, don't forget to put this back into perldiag:
12144 =item Regexp *+ operand could be empty at {#} in regex m/%s/
12146 (F) The part of the regexp subject to either the * or + quantifier
12147 could match an empty string. The {#} shows in the regular
12148 expression about where the problem was discovered.
12152 if (!(flags&HASWIDTH) && op != '?')
12153 vFAIL("Regexp *+ operand could be empty");
12156 #ifdef RE_TRACK_PATTERN_OFFSETS
12157 parse_start = RExC_parse;
12159 nextchar(pRExC_state);
12161 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
12167 else if (op == '+') {
12171 else if (op == '?') {
12176 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
12177 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
12178 ckWARN2reg(RExC_parse,
12179 "%" UTF8f " matches null string many times",
12180 UTF8fARG(UTF, (RExC_parse >= origparse
12181 ? RExC_parse - origparse
12184 (void)ReREFCNT_inc(RExC_rx_sv);
12187 if (*RExC_parse == '?') {
12188 nextchar(pRExC_state);
12189 reginsert(pRExC_state, MINMOD, ret, depth+1);
12190 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
12192 else if (*RExC_parse == '+') {
12194 nextchar(pRExC_state);
12195 ender = reg_node(pRExC_state, SUCCEED);
12196 REGTAIL(pRExC_state, ret, ender);
12197 reginsert(pRExC_state, SUSPEND, ret, depth+1);
12198 ender = reg_node(pRExC_state, TAIL);
12199 REGTAIL(pRExC_state, ret, ender);
12202 if (ISMULT2(RExC_parse)) {
12204 vFAIL("Nested quantifiers");
12211 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state,
12220 /* This routine teases apart the various meanings of \N and returns
12221 * accordingly. The input parameters constrain which meaning(s) is/are valid
12222 * in the current context.
12224 * Exactly one of <node_p> and <code_point_p> must be non-NULL.
12226 * If <code_point_p> is not NULL, the context is expecting the result to be a
12227 * single code point. If this \N instance turns out to a single code point,
12228 * the function returns TRUE and sets *code_point_p to that code point.
12230 * If <node_p> is not NULL, the context is expecting the result to be one of
12231 * the things representable by a regnode. If this \N instance turns out to be
12232 * one such, the function generates the regnode, returns TRUE and sets *node_p
12233 * to point to that regnode.
12235 * If this instance of \N isn't legal in any context, this function will
12236 * generate a fatal error and not return.
12238 * On input, RExC_parse should point to the first char following the \N at the
12239 * time of the call. On successful return, RExC_parse will have been updated
12240 * to point to just after the sequence identified by this routine. Also
12241 * *flagp has been updated as needed.
12243 * When there is some problem with the current context and this \N instance,
12244 * the function returns FALSE, without advancing RExC_parse, nor setting
12245 * *node_p, nor *code_point_p, nor *flagp.
12247 * If <cp_count> is not NULL, the caller wants to know the length (in code
12248 * points) that this \N sequence matches. This is set, and the input is
12249 * parsed for errors, even if the function returns FALSE, as detailed below.
12251 * There are 5 possibilities here, as detailed in the next 5 paragraphs.
12253 * Probably the most common case is for the \N to specify a single code point.
12254 * *cp_count will be set to 1, and *code_point_p will be set to that code
12257 * Another possibility is for the input to be an empty \N{}, which for
12258 * backwards compatibility we accept. *cp_count will be set to 0. *node_p
12259 * will be set to a generated NOTHING node.
12261 * Still another possibility is for the \N to mean [^\n]. *cp_count will be
12262 * set to 0. *node_p will be set to a generated REG_ANY node.
12264 * The fourth possibility is that \N resolves to a sequence of more than one
12265 * code points. *cp_count will be set to the number of code points in the
12266 * sequence. *node_p * will be set to a generated node returned by this
12267 * function calling S_reg().
12269 * The final possibility is that it is premature to be calling this function;
12270 * that pass1 needs to be restarted. This can happen when this changes from
12271 * /d to /u rules, or when the pattern needs to be upgraded to UTF-8. The
12272 * latter occurs only when the fourth possibility would otherwise be in
12273 * effect, and is because one of those code points requires the pattern to be
12274 * recompiled as UTF-8. The function returns FALSE, and sets the
12275 * RESTART_PASS1 and NEED_UTF8 flags in *flagp, as appropriate. When this
12276 * happens, the caller needs to desist from continuing parsing, and return
12277 * this information to its caller. This is not set for when there is only one
12278 * code point, as this can be called as part of an ANYOF node, and they can
12279 * store above-Latin1 code points without the pattern having to be in UTF-8.
12281 * For non-single-quoted regexes, the tokenizer has resolved character and
12282 * sequence names inside \N{...} into their Unicode values, normalizing the
12283 * result into what we should see here: '\N{U+c1.c2...}', where c1... are the
12284 * hex-represented code points in the sequence. This is done there because
12285 * the names can vary based on what charnames pragma is in scope at the time,
12286 * so we need a way to take a snapshot of what they resolve to at the time of
12287 * the original parse. [perl #56444].
12289 * That parsing is skipped for single-quoted regexes, so we may here get
12290 * '\N{NAME}'. This is a fatal error. These names have to be resolved by the
12291 * parser. But if the single-quoted regex is something like '\N{U+41}', that
12292 * is legal and handled here. The code point is Unicode, and has to be
12293 * translated into the native character set for non-ASCII platforms.
12296 char * endbrace; /* points to '}' following the name */
12297 char* p = RExC_parse; /* Temporary */
12299 SV * substitute_parse = NULL;
12303 Size_t count = 0; /* code point count kept internally by this function */
12305 GET_RE_DEBUG_FLAGS_DECL;
12307 PERL_ARGS_ASSERT_GROK_BSLASH_N;
12309 GET_RE_DEBUG_FLAGS;
12311 assert(cBOOL(node_p) ^ cBOOL(code_point_p)); /* Exactly one should be set */
12312 assert(! (node_p && cp_count)); /* At most 1 should be set */
12314 if (cp_count) { /* Initialize return for the most common case */
12318 /* The [^\n] meaning of \N ignores spaces and comments under the /x
12319 * modifier. The other meanings do not, so use a temporary until we find
12320 * out which we are being called with */
12321 skip_to_be_ignored_text(pRExC_state, &p,
12322 FALSE /* Don't force to /x */ );
12324 /* Disambiguate between \N meaning a named character versus \N meaning
12325 * [^\n]. The latter is assumed when the {...} following the \N is a legal
12326 * quantifier, or there is no '{' at all */
12327 if (*p != '{' || regcurly(p)) {
12337 *node_p = reg_node(pRExC_state, REG_ANY);
12338 *flagp |= HASWIDTH|SIMPLE;
12340 Set_Node_Length(*node_p, 1); /* MJD */
12344 /* The test above made sure that the next real character is a '{', but
12345 * under the /x modifier, it could be separated by space (or a comment and
12346 * \n) and this is not allowed (for consistency with \x{...} and the
12347 * tokenizer handling of \N{NAME}). */
12348 if (*RExC_parse != '{') {
12349 vFAIL("Missing braces on \\N{}");
12352 RExC_parse++; /* Skip past the '{' */
12354 endbrace = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
12355 if (! endbrace) { /* no trailing brace */
12356 vFAIL2("Missing right brace on \\%c{}", 'N');
12359 /* Here, we have decided it should be a named character or sequence */
12360 REQUIRE_UNI_RULES(flagp, FALSE); /* Unicode named chars imply Unicode
12363 if (endbrace == RExC_parse) { /* empty: \N{} */
12365 RExC_parse++; /* Position after the "}" */
12366 vFAIL("Zero length \\N{}");
12371 nextchar(pRExC_state);
12376 *node_p = reg_node(pRExC_state,NOTHING);
12380 /* If we haven't got something that begins with 'U+', then it didn't get lexed. */
12381 if ( endbrace - RExC_parse < 2
12382 || strnNE(RExC_parse, "U+", 2))
12384 RExC_parse = endbrace; /* position msg's '<--HERE' */
12385 vFAIL("\\N{NAME} must be resolved by the lexer");
12388 /* This code purposely indented below because of future changes coming */
12390 /* We can get to here when the input is \N{U+...} or when toke.c has
12391 * converted a name to the \N{U+...} form. This include changing a
12392 * name that evaluates to multiple code points to \N{U+c1.c2.c3 ...} */
12394 RExC_parse += 2; /* Skip past the 'U+' */
12396 /* Code points are separated by dots. The '}' terminates the whole
12399 do { /* Loop until the ending brace */
12401 char * start_digit; /* The first of the current code point */
12402 if (! isXDIGIT(*RExC_parse)) {
12404 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12407 start_digit = RExC_parse;
12410 /* Loop through the hex digits of the current code point */
12412 /* Adding this digit will shift the result 4 bits. If that
12413 * result would be above the legal max, it's overflow */
12414 if (cp > MAX_LEGAL_CP >> 4) {
12416 /* Find the end of the code point */
12419 } while (isXDIGIT(*RExC_parse) || *RExC_parse == '_');
12421 /* Be sure to synchronize this message with the similar one
12423 vFAIL4("Use of code point 0x%.*s is not allowed; the"
12424 " permissible max is 0x%" UVxf,
12425 (int) (RExC_parse - start_digit), start_digit,
12429 /* Accumulate this (valid) digit into the running total */
12430 cp = (cp << 4) + READ_XDIGIT(RExC_parse);
12432 /* READ_XDIGIT advanced the input pointer. Ignore a single
12433 * underscore separator */
12434 if (*RExC_parse == '_' && isXDIGIT(RExC_parse[1])) {
12437 } while (isXDIGIT(*RExC_parse));
12439 /* Here, have accumulated the next code point */
12440 if (RExC_parse >= endbrace) { /* If done ... */
12445 /* Here, is a single code point; fail if doesn't want that */
12446 if (! code_point_p) {
12451 /* A single code point is easy to handle; just return it */
12452 *code_point_p = UNI_TO_NATIVE(cp);
12453 RExC_parse = endbrace;
12454 nextchar(pRExC_state);
12458 /* Here, the only legal thing would be a multiple character
12459 * sequence (of the form "\N{U+c1.c2. ... }". So the next
12460 * character must be a dot (and the one after that can't be the
12461 * endbrace, or we'd have something like \N{U+100.} ) */
12462 if (*RExC_parse != '.' || RExC_parse + 1 >= endbrace) {
12463 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
12464 ? UTF8SKIP(RExC_parse)
12466 if (RExC_parse >= endbrace) { /* Guard against malformed utf8 */
12467 RExC_parse = endbrace;
12469 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12472 /* Here, looks like its really a multiple character sequence. Fail
12473 * if that's not what the caller wants. But continue with counting
12474 * and error checking if they still want a count */
12475 if (! node_p && ! cp_count) {
12479 /* What is done here is to convert this to a sub-pattern of the
12480 * form \x{char1}\x{char2}... and then call reg recursively to
12481 * parse it (enclosing in "(?: ... )" ). That way, it retains its
12482 * atomicness, while not having to worry about special handling
12483 * that some code points may have. We don't create a subpattern,
12484 * but go through the motions of code point counting and error
12485 * checking, if the caller doesn't want a node returned. */
12487 if (node_p && count == 1) {
12488 substitute_parse = newSVpvs("?:");
12494 /* Convert to notation the rest of the code understands */
12495 sv_catpvs(substitute_parse, "\\x{");
12496 sv_catpvn(substitute_parse, start_digit,
12497 RExC_parse - start_digit);
12498 sv_catpvs(substitute_parse, "}");
12501 /* Move to after the dot (or ending brace the final time through.)
12506 } while (RExC_parse < endbrace);
12508 if (! node_p) { /* Doesn't want the node */
12515 sv_catpvs(substitute_parse, ")");
12518 /* The values are Unicode, and therefore have to be converted to native
12519 * on a non-Unicode (meaning non-ASCII) platform. */
12520 RExC_recode_x_to_native = 1;
12523 /* Here, we have the string the name evaluates to, ready to be parsed,
12524 * stored in 'substitute_parse' as a series of valid "\x{...}\x{...}"
12525 * constructs. This can be called from within a substitute parse already.
12526 * The error reporting mechanism doesn't work for 2 levels of this, but the
12527 * code above has validated this new construct, so there should be no
12528 * errors generated by the below.*/
12529 save_start = RExC_start;
12530 orig_end = RExC_end;
12532 RExC_parse = RExC_start = SvPVX(substitute_parse);
12533 RExC_end = RExC_parse + SvCUR(substitute_parse);
12535 *node_p = reg(pRExC_state, 1, &flags, depth+1);
12537 /* Restore the saved values */
12538 RExC_start = save_start;
12539 RExC_parse = endbrace;
12540 RExC_end = orig_end;
12542 RExC_recode_x_to_native = 0;
12545 SvREFCNT_dec_NN(substitute_parse);
12548 RETURN_X_ON_RESTART(FALSE, flags,flagp);
12549 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#" UVxf,
12552 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
12554 nextchar(pRExC_state);
12560 PERL_STATIC_INLINE U8
12561 S_compute_EXACTish(RExC_state_t *pRExC_state)
12565 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
12573 op = get_regex_charset(RExC_flags);
12574 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
12575 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
12576 been, so there is no hole */
12579 return op + EXACTF;
12582 PERL_STATIC_INLINE void
12583 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state,
12584 regnode *node, I32* flagp, STRLEN len, UV code_point,
12587 /* This knows the details about sizing an EXACTish node, setting flags for
12588 * it (by setting <*flagp>, and potentially populating it with a single
12591 * If <len> (the length in bytes) is non-zero, this function assumes that
12592 * the node has already been populated, and just does the sizing. In this
12593 * case <code_point> should be the final code point that has already been
12594 * placed into the node. This value will be ignored except that under some
12595 * circumstances <*flagp> is set based on it.
12597 * If <len> is zero, the function assumes that the node is to contain only
12598 * the single character given by <code_point> and calculates what <len>
12599 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
12600 * additionally will populate the node's STRING with <code_point> or its
12603 * In both cases <*flagp> is appropriately set
12605 * It knows that under FOLD, the Latin Sharp S and UTF characters above
12606 * 255, must be folded (the former only when the rules indicate it can
12609 * When it does the populating, it looks at the flag 'downgradable'. If
12610 * true with a node that folds, it checks if the single code point
12611 * participates in a fold, and if not downgrades the node to an EXACT.
12612 * This helps the optimizer */
12614 bool len_passed_in = cBOOL(len != 0);
12615 U8 character[UTF8_MAXBYTES_CASE+1];
12617 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
12619 /* Don't bother to check for downgrading in PASS1, as it doesn't make any
12620 * sizing difference, and is extra work that is thrown away */
12621 if (downgradable && ! PASS2) {
12622 downgradable = FALSE;
12625 if (! len_passed_in) {
12627 if (UVCHR_IS_INVARIANT(code_point)) {
12628 if (LOC || ! FOLD) { /* /l defers folding until runtime */
12629 *character = (U8) code_point;
12631 else { /* Here is /i and not /l. (toFOLD() is defined on just
12632 ASCII, which isn't the same thing as INVARIANT on
12633 EBCDIC, but it works there, as the extra invariants
12634 fold to themselves) */
12635 *character = toFOLD((U8) code_point);
12637 /* We can downgrade to an EXACT node if this character
12638 * isn't a folding one. Note that this assumes that
12639 * nothing above Latin1 folds to some other invariant than
12640 * one of these alphabetics; otherwise we would also have
12642 * && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
12643 * || ASCII_FOLD_RESTRICTED))
12645 if (downgradable && PL_fold[code_point] == code_point) {
12651 else if (FOLD && (! LOC
12652 || ! is_PROBLEMATIC_LOCALE_FOLD_cp(code_point)))
12653 { /* Folding, and ok to do so now */
12654 UV folded = _to_uni_fold_flags(
12658 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
12659 ? FOLD_FLAGS_NOMIX_ASCII
12662 && folded == code_point /* This quickly rules out many
12663 cases, avoiding the
12664 _invlist_contains_cp() overhead
12666 && ! _invlist_contains_cp(PL_utf8_foldable, code_point))
12673 else if (code_point <= MAX_UTF8_TWO_BYTE) {
12675 /* Not folding this cp, and can output it directly */
12676 *character = UTF8_TWO_BYTE_HI(code_point);
12677 *(character + 1) = UTF8_TWO_BYTE_LO(code_point);
12681 uvchr_to_utf8( character, code_point);
12682 len = UTF8SKIP(character);
12684 } /* Else pattern isn't UTF8. */
12686 *character = (U8) code_point;
12688 } /* Else is folded non-UTF8 */
12689 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
12690 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
12691 || UNICODE_DOT_DOT_VERSION > 0)
12692 else if (LIKELY(code_point != LATIN_SMALL_LETTER_SHARP_S)) {
12696 /* We don't fold any non-UTF8 except possibly the Sharp s (see
12697 * comments at join_exact()); */
12698 *character = (U8) code_point;
12701 /* Can turn into an EXACT node if we know the fold at compile time,
12702 * and it folds to itself and doesn't particpate in other folds */
12705 && PL_fold_latin1[code_point] == code_point
12706 && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
12707 || (isASCII(code_point) && ASCII_FOLD_RESTRICTED)))
12711 } /* else is Sharp s. May need to fold it */
12712 else if (AT_LEAST_UNI_SEMANTICS && ! ASCII_FOLD_RESTRICTED) {
12714 *(character + 1) = 's';
12718 *character = LATIN_SMALL_LETTER_SHARP_S;
12724 RExC_size += STR_SZ(len);
12727 RExC_emit += STR_SZ(len);
12728 STR_LEN(node) = len;
12729 if (! len_passed_in) {
12730 Copy((char *) character, STRING(node), len, char);
12734 *flagp |= HASWIDTH;
12736 /* A single character node is SIMPLE, except for the special-cased SHARP S
12738 if ((len == 1 || (UTF && len == UVCHR_SKIP(code_point)))
12739 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
12740 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
12741 || UNICODE_DOT_DOT_VERSION > 0)
12742 && ( code_point != LATIN_SMALL_LETTER_SHARP_S
12743 || ! FOLD || ! DEPENDS_SEMANTICS)
12749 /* The OP may not be well defined in PASS1 */
12750 if (PASS2 && OP(node) == EXACTFL) {
12751 RExC_contains_locale = 1;
12756 S_new_regcurly(const char *s, const char *e)
12758 /* This is a temporary function designed to match the most lenient form of
12759 * a {m,n} quantifier we ever envision, with either number omitted, and
12760 * spaces anywhere between/before/after them.
12762 * If this function fails, then the string it matches is very unlikely to
12763 * ever be considered a valid quantifier, so we can allow the '{' that
12764 * begins it to be considered as a literal */
12766 bool has_min = FALSE;
12767 bool has_max = FALSE;
12769 PERL_ARGS_ASSERT_NEW_REGCURLY;
12771 if (s >= e || *s++ != '{')
12774 while (s < e && isSPACE(*s)) {
12777 while (s < e && isDIGIT(*s)) {
12781 while (s < e && isSPACE(*s)) {
12787 while (s < e && isSPACE(*s)) {
12790 while (s < e && isDIGIT(*s)) {
12794 while (s < e && isSPACE(*s)) {
12799 return s < e && *s == '}' && (has_min || has_max);
12802 /* Parse backref decimal value, unless it's too big to sensibly be a backref,
12803 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
12806 S_backref_value(char *p, char *e)
12808 const char* endptr = e;
12810 if (grok_atoUV(p, &val, &endptr) && val <= I32_MAX)
12817 - regatom - the lowest level
12819 Try to identify anything special at the start of the current parse position.
12820 If there is, then handle it as required. This may involve generating a
12821 single regop, such as for an assertion; or it may involve recursing, such as
12822 to handle a () structure.
12824 If the string doesn't start with something special then we gobble up
12825 as much literal text as we can. If we encounter a quantifier, we have to
12826 back off the final literal character, as that quantifier applies to just it
12827 and not to the whole string of literals.
12829 Once we have been able to handle whatever type of thing started the
12830 sequence, we return.
12832 Note: we have to be careful with escapes, as they can be both literal
12833 and special, and in the case of \10 and friends, context determines which.
12835 A summary of the code structure is:
12837 switch (first_byte) {
12838 cases for each special:
12839 handle this special;
12842 switch (2nd byte) {
12843 cases for each unambiguous special:
12844 handle this special;
12846 cases for each ambigous special/literal:
12848 if (special) handle here
12850 default: // unambiguously literal:
12853 default: // is a literal char
12856 create EXACTish node for literal;
12857 while (more input and node isn't full) {
12858 switch (input_byte) {
12859 cases for each special;
12860 make sure parse pointer is set so that the next call to
12861 regatom will see this special first
12862 goto loopdone; // EXACTish node terminated by prev. char
12864 append char to EXACTISH node;
12866 get next input byte;
12870 return the generated node;
12872 Specifically there are two separate switches for handling
12873 escape sequences, with the one for handling literal escapes requiring
12874 a dummy entry for all of the special escapes that are actually handled
12877 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
12879 Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
12880 restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
12881 Otherwise does not return NULL.
12885 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
12887 regnode *ret = NULL;
12894 GET_RE_DEBUG_FLAGS_DECL;
12896 *flagp = WORST; /* Tentatively. */
12898 DEBUG_PARSE("atom");
12900 PERL_ARGS_ASSERT_REGATOM;
12903 parse_start = RExC_parse;
12904 assert(RExC_parse < RExC_end);
12905 switch ((U8)*RExC_parse) {
12907 RExC_seen_zerolen++;
12908 nextchar(pRExC_state);
12909 if (RExC_flags & RXf_PMf_MULTILINE)
12910 ret = reg_node(pRExC_state, MBOL);
12912 ret = reg_node(pRExC_state, SBOL);
12913 Set_Node_Length(ret, 1); /* MJD */
12916 nextchar(pRExC_state);
12918 RExC_seen_zerolen++;
12919 if (RExC_flags & RXf_PMf_MULTILINE)
12920 ret = reg_node(pRExC_state, MEOL);
12922 ret = reg_node(pRExC_state, SEOL);
12923 Set_Node_Length(ret, 1); /* MJD */
12926 nextchar(pRExC_state);
12927 if (RExC_flags & RXf_PMf_SINGLELINE)
12928 ret = reg_node(pRExC_state, SANY);
12930 ret = reg_node(pRExC_state, REG_ANY);
12931 *flagp |= HASWIDTH|SIMPLE;
12933 Set_Node_Length(ret, 1); /* MJD */
12937 char * const oregcomp_parse = ++RExC_parse;
12938 ret = regclass(pRExC_state, flagp,depth+1,
12939 FALSE, /* means parse the whole char class */
12940 TRUE, /* allow multi-char folds */
12941 FALSE, /* don't silence non-portable warnings. */
12942 (bool) RExC_strict,
12943 TRUE, /* Allow an optimized regnode result */
12947 RETURN_NULL_ON_RESTART_FLAGP_OR_FLAGS(flagp,NEED_UTF8);
12948 FAIL2("panic: regclass returned NULL to regatom, flags=%#" UVxf,
12951 if (*RExC_parse != ']') {
12952 RExC_parse = oregcomp_parse;
12953 vFAIL("Unmatched [");
12955 nextchar(pRExC_state);
12956 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
12960 nextchar(pRExC_state);
12961 ret = reg(pRExC_state, 2, &flags,depth+1);
12963 if (flags & TRYAGAIN) {
12964 if (RExC_parse >= RExC_end) {
12965 /* Make parent create an empty node if needed. */
12966 *flagp |= TRYAGAIN;
12971 RETURN_NULL_ON_RESTART(flags,flagp);
12972 FAIL2("panic: reg returned NULL to regatom, flags=%#" UVxf,
12975 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
12979 if (flags & TRYAGAIN) {
12980 *flagp |= TRYAGAIN;
12983 vFAIL("Internal urp");
12984 /* Supposed to be caught earlier. */
12990 vFAIL("Quantifier follows nothing");
12995 This switch handles escape sequences that resolve to some kind
12996 of special regop and not to literal text. Escape sequnces that
12997 resolve to literal text are handled below in the switch marked
13000 Every entry in this switch *must* have a corresponding entry
13001 in the literal escape switch. However, the opposite is not
13002 required, as the default for this switch is to jump to the
13003 literal text handling code.
13006 switch ((U8)*RExC_parse) {
13007 /* Special Escapes */
13009 RExC_seen_zerolen++;
13010 ret = reg_node(pRExC_state, SBOL);
13011 /* SBOL is shared with /^/ so we set the flags so we can tell
13012 * /\A/ from /^/ in split. We check ret because first pass we
13013 * have no regop struct to set the flags on. */
13017 goto finish_meta_pat;
13019 ret = reg_node(pRExC_state, GPOS);
13020 RExC_seen |= REG_GPOS_SEEN;
13022 goto finish_meta_pat;
13024 RExC_seen_zerolen++;
13025 ret = reg_node(pRExC_state, KEEPS);
13027 /* XXX:dmq : disabling in-place substitution seems to
13028 * be necessary here to avoid cases of memory corruption, as
13029 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
13031 RExC_seen |= REG_LOOKBEHIND_SEEN;
13032 goto finish_meta_pat;
13034 ret = reg_node(pRExC_state, SEOL);
13036 RExC_seen_zerolen++; /* Do not optimize RE away */
13037 goto finish_meta_pat;
13039 ret = reg_node(pRExC_state, EOS);
13041 RExC_seen_zerolen++; /* Do not optimize RE away */
13042 goto finish_meta_pat;
13044 vFAIL("\\C no longer supported");
13046 ret = reg_node(pRExC_state, CLUMP);
13047 *flagp |= HASWIDTH;
13048 goto finish_meta_pat;
13054 arg = ANYOF_WORDCHAR;
13062 regex_charset charset = get_regex_charset(RExC_flags);
13064 RExC_seen_zerolen++;
13065 RExC_seen |= REG_LOOKBEHIND_SEEN;
13066 op = BOUND + charset;
13068 if (op == BOUNDL) {
13069 RExC_contains_locale = 1;
13072 ret = reg_node(pRExC_state, op);
13074 if (RExC_parse >= RExC_end || *(RExC_parse + 1) != '{') {
13075 FLAGS(ret) = TRADITIONAL_BOUND;
13076 if (PASS2 && op > BOUNDA) { /* /aa is same as /a */
13082 char name = *RExC_parse;
13083 char * endbrace = NULL;
13085 if (RExC_parse < RExC_end) {
13086 endbrace = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
13090 vFAIL2("Missing right brace on \\%c{}", name);
13092 /* XXX Need to decide whether to take spaces or not. Should be
13093 * consistent with \p{}, but that currently is SPACE, which
13094 * means vertical too, which seems wrong
13095 * while (isBLANK(*RExC_parse)) {
13098 if (endbrace == RExC_parse) {
13099 RExC_parse++; /* After the '}' */
13100 vFAIL2("Empty \\%c{}", name);
13102 length = endbrace - RExC_parse;
13103 /*while (isBLANK(*(RExC_parse + length - 1))) {
13106 switch (*RExC_parse) {
13109 && (memNEs(RExC_parse + 1, length - 1, "cb")))
13111 goto bad_bound_type;
13113 FLAGS(ret) = GCB_BOUND;
13116 if (length != 2 || *(RExC_parse + 1) != 'b') {
13117 goto bad_bound_type;
13119 FLAGS(ret) = LB_BOUND;
13122 if (length != 2 || *(RExC_parse + 1) != 'b') {
13123 goto bad_bound_type;
13125 FLAGS(ret) = SB_BOUND;
13128 if (length != 2 || *(RExC_parse + 1) != 'b') {
13129 goto bad_bound_type;
13131 FLAGS(ret) = WB_BOUND;
13135 RExC_parse = endbrace;
13137 "'%" UTF8f "' is an unknown bound type",
13138 UTF8fARG(UTF, length, endbrace - length));
13139 NOT_REACHED; /*NOTREACHED*/
13141 RExC_parse = endbrace;
13142 REQUIRE_UNI_RULES(flagp, NULL);
13144 if (PASS2 && op >= BOUNDA) { /* /aa is same as /a */
13148 /* Don't have to worry about UTF-8, in this message because
13149 * to get here the contents of the \b must be ASCII */
13150 ckWARN4reg(RExC_parse + 1, /* Include the '}' in msg */
13151 "Using /u for '%.*s' instead of /%s",
13153 endbrace - length + 1,
13154 (charset == REGEX_ASCII_RESTRICTED_CHARSET)
13155 ? ASCII_RESTRICT_PAT_MODS
13156 : ASCII_MORE_RESTRICT_PAT_MODS);
13160 if (PASS2 && invert) {
13161 OP(ret) += NBOUND - BOUND;
13163 goto finish_meta_pat;
13171 if (! DEPENDS_SEMANTICS) {
13175 /* \d doesn't have any matches in the upper Latin1 range, hence /d
13176 * is equivalent to /u. Changing to /u saves some branches at
13179 goto join_posix_op_known;
13182 ret = reg_node(pRExC_state, LNBREAK);
13183 *flagp |= HASWIDTH|SIMPLE;
13184 goto finish_meta_pat;
13192 goto join_posix_op_known;
13198 arg = ANYOF_VERTWS;
13200 goto join_posix_op_known;
13210 op = POSIXD + get_regex_charset(RExC_flags);
13211 if (op > POSIXA) { /* /aa is same as /a */
13214 else if (op == POSIXL) {
13215 RExC_contains_locale = 1;
13218 join_posix_op_known:
13221 op += NPOSIXD - POSIXD;
13224 ret = reg_node(pRExC_state, op);
13226 FLAGS(ret) = namedclass_to_classnum(arg);
13229 *flagp |= HASWIDTH|SIMPLE;
13233 if ( UCHARAT(RExC_parse + 1) == '{'
13234 && UNLIKELY(! new_regcurly(RExC_parse + 1, RExC_end)))
13237 vFAIL("Unescaped left brace in regex is illegal here");
13239 nextchar(pRExC_state);
13240 Set_Node_Length(ret, 2); /* MJD */
13246 ret = regclass(pRExC_state, flagp,depth+1,
13247 TRUE, /* means just parse this element */
13248 FALSE, /* don't allow multi-char folds */
13249 FALSE, /* don't silence non-portable warnings. It
13250 would be a bug if these returned
13252 (bool) RExC_strict,
13253 TRUE, /* Allow an optimized regnode result */
13256 RETURN_NULL_ON_RESTART_FLAGP(flagp);
13257 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
13258 * multi-char folds are allowed. */
13260 FAIL2("panic: regclass returned NULL to regatom, flags=%#" UVxf,
13265 Set_Node_Offset(ret, parse_start);
13266 Set_Node_Cur_Length(ret, parse_start - 2);
13267 nextchar(pRExC_state);
13270 /* Handle \N, \N{} and \N{NAMED SEQUENCE} (the latter meaning the
13271 * \N{...} evaluates to a sequence of more than one code points).
13272 * The function call below returns a regnode, which is our result.
13273 * The parameters cause it to fail if the \N{} evaluates to a
13274 * single code point; we handle those like any other literal. The
13275 * reason that the multicharacter case is handled here and not as
13276 * part of the EXACtish code is because of quantifiers. In
13277 * /\N{BLAH}+/, the '+' applies to the whole thing, and doing it
13278 * this way makes that Just Happen. dmq.
13279 * join_exact() will join this up with adjacent EXACTish nodes
13280 * later on, if appropriate. */
13282 if (grok_bslash_N(pRExC_state,
13283 &ret, /* Want a regnode returned */
13284 NULL, /* Fail if evaluates to a single code
13286 NULL, /* Don't need a count of how many code
13295 RETURN_NULL_ON_RESTART_FLAGP(flagp);
13297 /* Here, evaluates to a single code point. Go get that */
13298 RExC_parse = parse_start;
13301 case 'k': /* Handle \k<NAME> and \k'NAME' */
13305 if ( RExC_parse >= RExC_end - 1
13306 || (( ch = RExC_parse[1]) != '<'
13311 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
13312 vFAIL2("Sequence %.2s... not terminated",parse_start);
13315 ret = handle_named_backref(pRExC_state,
13327 case '1': case '2': case '3': case '4':
13328 case '5': case '6': case '7': case '8': case '9':
13333 if (*RExC_parse == 'g') {
13337 if (*RExC_parse == '{') {
13341 if (*RExC_parse == '-') {
13345 if (hasbrace && !isDIGIT(*RExC_parse)) {
13346 if (isrel) RExC_parse--;
13348 goto parse_named_seq;
13351 if (RExC_parse >= RExC_end) {
13352 goto unterminated_g;
13354 num = S_backref_value(RExC_parse, RExC_end);
13356 vFAIL("Reference to invalid group 0");
13357 else if (num == I32_MAX) {
13358 if (isDIGIT(*RExC_parse))
13359 vFAIL("Reference to nonexistent group");
13362 vFAIL("Unterminated \\g... pattern");
13366 num = RExC_npar - num;
13368 vFAIL("Reference to nonexistent or unclosed group");
13372 num = S_backref_value(RExC_parse, RExC_end);
13373 /* bare \NNN might be backref or octal - if it is larger
13374 * than or equal RExC_npar then it is assumed to be an
13375 * octal escape. Note RExC_npar is +1 from the actual
13376 * number of parens. */
13377 /* Note we do NOT check if num == I32_MAX here, as that is
13378 * handled by the RExC_npar check */
13381 /* any numeric escape < 10 is always a backref */
13383 /* any numeric escape < RExC_npar is a backref */
13384 && num >= RExC_npar
13385 /* cannot be an octal escape if it starts with 8 */
13386 && *RExC_parse != '8'
13387 /* cannot be an octal escape it it starts with 9 */
13388 && *RExC_parse != '9'
13391 /* Probably not a backref, instead likely to be an
13392 * octal character escape, e.g. \35 or \777.
13393 * The above logic should make it obvious why using
13394 * octal escapes in patterns is problematic. - Yves */
13395 RExC_parse = parse_start;
13400 /* At this point RExC_parse points at a numeric escape like
13401 * \12 or \88 or something similar, which we should NOT treat
13402 * as an octal escape. It may or may not be a valid backref
13403 * escape. For instance \88888888 is unlikely to be a valid
13405 while (isDIGIT(*RExC_parse))
13408 if (*RExC_parse != '}')
13409 vFAIL("Unterminated \\g{...} pattern");
13413 if (num > (I32)RExC_rx->nparens)
13414 vFAIL("Reference to nonexistent group");
13417 ret = reganode(pRExC_state,
13420 : (ASCII_FOLD_RESTRICTED)
13422 : (AT_LEAST_UNI_SEMANTICS)
13428 *flagp |= HASWIDTH;
13430 /* override incorrect value set in reganode MJD */
13431 Set_Node_Offset(ret, parse_start);
13432 Set_Node_Cur_Length(ret, parse_start-1);
13433 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13434 FALSE /* Don't force to /x */ );
13438 if (RExC_parse >= RExC_end)
13439 FAIL("Trailing \\");
13442 /* Do not generate "unrecognized" warnings here, we fall
13443 back into the quick-grab loop below */
13444 RExC_parse = parse_start;
13446 } /* end of switch on a \foo sequence */
13451 /* '#' comments should have been spaced over before this function was
13453 assert((RExC_flags & RXf_PMf_EXTENDED) == 0);
13455 if (RExC_flags & RXf_PMf_EXTENDED) {
13456 RExC_parse = reg_skipcomment( pRExC_state, RExC_parse );
13457 if (RExC_parse < RExC_end)
13467 /* Here, we have determined that the next thing is probably a
13468 * literal character. RExC_parse points to the first byte of its
13469 * definition. (It still may be an escape sequence that evaluates
13470 * to a single character) */
13477 /* This allows us to fill a node with just enough spare so that if the final
13478 * character folds, its expansion is guaranteed to fit */
13479 #define MAX_NODE_STRING_SIZE (255-UTF8_MAXBYTES_CASE)
13480 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE+1];
13483 U8 upper_parse = MAX_NODE_STRING_SIZE;
13485 /* We start out as an EXACT node, even if under /i, until we find a
13486 * character which is in a fold. The algorithm now segregates into
13487 * separate nodes, characters that fold from those that don't under
13488 * /i. (This hopefull will create nodes that are fixed strings
13489 * even under /i, giving the optimizer something to grab onto to.)
13490 * So, if a node has something in it and the next character is in
13491 * the opposite category, that node is closed up, and the function
13492 * returns. Then regatom is called again, and a new node is
13493 * created for the new category. */
13494 U8 node_type = EXACT;
13496 bool next_is_quantifier;
13497 char * oldp = NULL;
13499 /* We can convert EXACTF nodes to EXACTFU if they contain only
13500 * characters that match identically regardless of the target
13501 * string's UTF8ness. The reason to do this is that EXACTF is not
13502 * trie-able, EXACTFU is.
13504 * Similarly, we can convert EXACTFL nodes to EXACTFLU8 if they
13505 * contain only above-Latin1 characters (hence must be in UTF8),
13506 * which don't participate in folds with Latin1-range characters,
13507 * as the latter's folds aren't known until runtime. (We don't
13508 * need to figure this out until pass 2) */
13509 bool maybe_exactfu = PASS2;
13511 /* To see if RExC_uni_semantics changes during parsing of the node.
13513 bool uni_semantics_at_node_start;
13515 /* The node_type may change below, but since the size of the node
13516 * doesn't change, it works */
13517 ret = reg_node(pRExC_state, node_type);
13519 /* In pass1, folded, we use a temporary buffer instead of the
13520 * actual node, as the node doesn't exist yet */
13521 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
13527 /* This breaks under rare circumstances. If folding, we do not
13528 * want to split a node at a character that is a non-final in a
13529 * multi-char fold, as an input string could just happen to want to
13530 * match across the node boundary. The code at the end of the loop
13531 * looks for this, and backs off until it finds not such a
13532 * character, but it is possible (though extremely, extremely
13533 * unlikely) for all characters in the node to be non-final fold
13534 * ones, in which case we just leave the node fully filled, and
13535 * hope that it doesn't match the string in just the wrong place */
13537 assert( ! UTF /* Is at the beginning of a character */
13538 || UTF8_IS_INVARIANT(UCHARAT(RExC_parse))
13539 || UTF8_IS_START(UCHARAT(RExC_parse)));
13541 uni_semantics_at_node_start = RExC_uni_semantics;
13543 /* Here, we have a literal character. Find the maximal string of
13544 * them in the input that we can fit into a single EXACTish node.
13545 * We quit at the first non-literal or when the node gets full, or
13546 * under /i the categorization of folding/non-folding character
13548 for (p = RExC_parse; len < upper_parse && p < RExC_end; ) {
13550 /* In most cases each iteration adds one byte to the output.
13551 * The exceptions override this */
13552 Size_t added_len = 1;
13556 /* White space has already been ignored */
13557 assert( (RExC_flags & RXf_PMf_EXTENDED) == 0
13558 || ! is_PATWS_safe((p), RExC_end, UTF));
13570 /* Literal Escapes Switch
13572 This switch is meant to handle escape sequences that
13573 resolve to a literal character.
13575 Every escape sequence that represents something
13576 else, like an assertion or a char class, is handled
13577 in the switch marked 'Special Escapes' above in this
13578 routine, but also has an entry here as anything that
13579 isn't explicitly mentioned here will be treated as
13580 an unescaped equivalent literal.
13583 switch ((U8)*++p) {
13584 /* These are all the special escapes. */
13585 case 'A': /* Start assertion */
13586 case 'b': case 'B': /* Word-boundary assertion*/
13587 case 'C': /* Single char !DANGEROUS! */
13588 case 'd': case 'D': /* digit class */
13589 case 'g': case 'G': /* generic-backref, pos assertion */
13590 case 'h': case 'H': /* HORIZWS */
13591 case 'k': case 'K': /* named backref, keep marker */
13592 case 'p': case 'P': /* Unicode property */
13593 case 'R': /* LNBREAK */
13594 case 's': case 'S': /* space class */
13595 case 'v': case 'V': /* VERTWS */
13596 case 'w': case 'W': /* word class */
13597 case 'X': /* eXtended Unicode "combining
13598 character sequence" */
13599 case 'z': case 'Z': /* End of line/string assertion */
13603 /* Anything after here is an escape that resolves to a
13604 literal. (Except digits, which may or may not)
13610 case 'N': /* Handle a single-code point named character. */
13611 RExC_parse = p + 1;
13612 if (! grok_bslash_N(pRExC_state,
13613 NULL, /* Fail if evaluates to
13614 anything other than a
13615 single code point */
13616 &ender, /* The returned single code
13618 NULL, /* Don't need a count of
13619 how many code points */
13624 if (*flagp & NEED_UTF8)
13625 FAIL("panic: grok_bslash_N set NEED_UTF8");
13626 RETURN_NULL_ON_RESTART_FLAGP(flagp);
13628 /* Here, it wasn't a single code point. Go close
13629 * up this EXACTish node. The switch() prior to
13630 * this switch handles the other cases */
13631 RExC_parse = p = oldp;
13635 RExC_parse = parse_start;
13636 if (ender > 0xff) {
13637 REQUIRE_UTF8(flagp);
13653 ender = ESC_NATIVE;
13663 const char* error_msg;
13665 bool valid = grok_bslash_o(&p,
13669 PASS2, /* out warnings */
13670 (bool) RExC_strict,
13671 TRUE, /* Output warnings
13676 RExC_parse = p; /* going to die anyway; point
13677 to exact spot of failure */
13681 if (ender > 0xff) {
13682 REQUIRE_UTF8(flagp);
13688 UV result = UV_MAX; /* initialize to erroneous
13690 const char* error_msg;
13692 bool valid = grok_bslash_x(&p,
13696 PASS2, /* out warnings */
13697 (bool) RExC_strict,
13698 TRUE, /* Silence warnings
13703 RExC_parse = p; /* going to die anyway; point
13704 to exact spot of failure */
13709 if (ender < 0x100) {
13711 if (RExC_recode_x_to_native) {
13712 ender = LATIN1_TO_NATIVE(ender);
13717 REQUIRE_UTF8(flagp);
13723 ender = grok_bslash_c(*p++, PASS2);
13725 case '8': case '9': /* must be a backreference */
13727 /* we have an escape like \8 which cannot be an octal escape
13728 * so we exit the loop, and let the outer loop handle this
13729 * escape which may or may not be a legitimate backref. */
13731 case '1': case '2': case '3':case '4':
13732 case '5': case '6': case '7':
13733 /* When we parse backslash escapes there is ambiguity
13734 * between backreferences and octal escapes. Any escape
13735 * from \1 - \9 is a backreference, any multi-digit
13736 * escape which does not start with 0 and which when
13737 * evaluated as decimal could refer to an already
13738 * parsed capture buffer is a back reference. Anything
13741 * Note this implies that \118 could be interpreted as
13742 * 118 OR as "\11" . "8" depending on whether there
13743 * were 118 capture buffers defined already in the
13746 /* NOTE, RExC_npar is 1 more than the actual number of
13747 * parens we have seen so far, hence the < RExC_npar below. */
13749 if ( !isDIGIT(p[1]) || S_backref_value(p, RExC_end) < RExC_npar)
13750 { /* Not to be treated as an octal constant, go
13758 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
13760 ender = grok_oct(p, &numlen, &flags, NULL);
13761 if (ender > 0xff) {
13762 REQUIRE_UTF8(flagp);
13765 if (PASS2 /* like \08, \178 */
13767 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
13769 reg_warn_non_literal_string(
13771 form_short_octal_warning(p, numlen));
13777 FAIL("Trailing \\");
13780 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
13781 /* Include any left brace following the alpha to emphasize
13782 * that it could be part of an escape at some point
13784 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
13785 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
13787 goto normal_default;
13788 } /* End of switch on '\' */
13791 /* Trying to gain new uses for '{' without breaking too
13792 * much existing code is hard. The solution currently
13794 * 1) If there is no ambiguity that a '{' should always
13795 * be taken literally, at the start of a construct, we
13797 * 2) If the literal '{' conflicts with our desired use
13798 * of it as a metacharacter, we die. The deprecation
13799 * cycles for this have come and gone.
13800 * 3) If there is ambiguity, we raise a simple warning.
13801 * This could happen, for example, if the user
13802 * intended it to introduce a quantifier, but slightly
13803 * misspelled the quantifier. Without this warning,
13804 * the quantifier would silently be taken as a literal
13805 * string of characters instead of a meta construct */
13806 if (len || (p > RExC_start && isALPHA_A(*(p - 1)))) {
13808 || ( p > parse_start + 1
13809 && isALPHA_A(*(p - 1))
13810 && *(p - 2) == '\\')
13811 || new_regcurly(p, RExC_end))
13813 RExC_parse = p + 1;
13814 vFAIL("Unescaped left brace in regex is "
13818 ckWARNreg(p + 1, "Unescaped left brace in regex is"
13819 " passed through");
13822 goto normal_default;
13825 if (PASS2 && p > RExC_parse && RExC_strict) {
13826 ckWARN2reg(p + 1, "Unescaped literal '%c'", *p);
13829 default: /* A literal character */
13831 if (! UTF8_IS_INVARIANT(*p) && UTF) {
13833 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
13834 &numlen, UTF8_ALLOW_DEFAULT);
13840 } /* End of switch on the literal */
13842 /* Here, have looked at the literal character, and <ender>
13843 * contains its ordinal; <p> points to the character after it.
13844 * We need to check if the next non-ignored thing is a
13845 * quantifier. Move <p> to after anything that should be
13846 * ignored, which, as a side effect, positions <p> for the next
13847 * loop iteration */
13848 skip_to_be_ignored_text(pRExC_state, &p,
13849 FALSE /* Don't force to /x */ );
13851 /* If the next thing is a quantifier, it applies to this
13852 * character only, which means that this character has to be in
13853 * its own node and can't just be appended to the string in an
13854 * existing node, so if there are already other characters in
13855 * the node, close the node with just them, and set up to do
13856 * this character again next time through, when it will be the
13857 * only thing in its new node */
13859 next_is_quantifier = LIKELY(p < RExC_end)
13860 && UNLIKELY(ISMULT2(p));
13862 if (next_is_quantifier && LIKELY(len)) {
13867 /* Ready to add 'ender' to the node */
13869 if (! FOLD) { /* The simple case, just append the literal */
13871 /* In the sizing pass, we need only the size of the
13872 * character we are appending, hence we can delay getting
13873 * its representation until PASS2. */
13875 if (UTF && ! UVCHR_IS_INVARIANT(ender)) {
13876 const STRLEN unilen = UVCHR_SKIP(ender);
13878 added_len = unilen;
13883 } else { /* PASS2 */
13885 if (UTF && ! UVCHR_IS_INVARIANT(ender)) {
13886 U8 * new_s = uvchr_to_utf8((U8*)s, ender);
13887 added_len = (char *) new_s - s;
13888 s = (char *) new_s;
13891 *(s++) = (char) ender;
13895 else if (LOC && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)) {
13897 /* Here are folding under /l, and the code point is
13898 * problematic. If this is the first character in the
13899 * node, change the node type to folding. Otherwise, if
13900 * this is the first problematic character, close up the
13901 * existing node, so can start a new node with this one */
13903 node_type = EXACTFL;
13905 else if (node_type == EXACT) {
13910 /* This code point means we can't simplify things */
13911 maybe_exactfu = FALSE;
13913 /* A problematic code point in this context means that its
13914 * fold isn't known until runtime, so we can't fold it now.
13915 * (The non-problematic code points are the above-Latin1
13916 * ones that fold to also all above-Latin1. Their folds
13917 * don't vary no matter what the locale is.) But here we
13918 * have characters whose fold depends on the locale.
13919 * Unlike the non-folding case above, we have to keep track
13920 * of these in the sizing pass, so that we can make sure we
13921 * don't split too-long nodes in the middle of a potential
13922 * multi-char fold. And unlike the regular fold case
13923 * handled in the else clauses below, we don't actually
13924 * fold and don't have special cases to consider. What we
13925 * do for both passes is the PASS2 code for non-folding */
13926 goto not_fold_common;
13928 else /* A regular FOLD code point */
13931 /* Here, are folding and are not UTF-8 encoded; therefore
13932 * the character must be in the range 0-255, and is not /l.
13933 * (Not /l because we already handled these under /l in
13934 * is_PROBLEMATIC_LOCALE_FOLD_cp) */
13935 if (! IS_IN_SOME_FOLD_L1(ender)) {
13937 /* Start a new node for this non-folding character if
13938 * previous ones in the node were folded */
13939 if (len && node_type != EXACT) {
13944 *(s++) = (char) ender;
13946 else { /* Here, does participate in some fold */
13948 /* if this is the first character in the node, change
13949 * its type to folding. Otherwise, if this is the
13950 * first folding character in the node, close up the
13951 * existing node, so can start a new node with this
13954 node_type = compute_EXACTish(pRExC_state);
13956 else if (node_type == EXACT) {
13961 /* See if the character's fold differs between /d and
13962 * /u. On non-ancient Unicode versions, this includes
13963 * the multi-char fold SHARP S to 'ss' */
13965 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
13966 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
13967 || UNICODE_DOT_DOT_VERSION > 0)
13969 if (UNLIKELY(ender == LATIN_SMALL_LETTER_SHARP_S)) {
13971 /* See comments for join_exact() as to why we fold
13972 * this non-UTF at compile time */
13973 if (node_type == EXACTFU) {
13976 /* Let the code below add in the extra 's' */
13980 else if ( uni_semantics_at_node_start
13981 != RExC_uni_semantics)
13983 /* Here, we are supossed to be using Unicode
13984 * rules, but this folding node is not. This
13985 * happens during pass 1 when the node started
13986 * out not under Unicode rules, but a \N{} was
13987 * encountered during the processing of it,
13988 * causing Unicode rules to be switched into.
13989 * Pass 1 continues uninterrupted, as by the
13990 * time we get to pass 2, we will know enough
13991 * to generate the correct folds. Except in
13992 * this one case, we need to restart the node,
13993 * because the fold of the sharp s requires 2
13994 * characters, and the sizing needs to account
14000 RExC_seen_unfolded_sharp_s = 1;
14001 maybe_exactfu = FALSE;
14005 && isALPHA_FOLD_EQ(ender, 's')
14006 && isALPHA_FOLD_EQ(*(s-1), 's'))
14008 maybe_exactfu = FALSE;
14013 if (PL_fold[ender] != PL_fold_latin1[ender]) {
14014 maybe_exactfu = FALSE;
14017 /* Even when folding, we store just the input
14018 * character, as we have an array that finds its fold
14020 *(s++) = (char) ender;
14023 else { /* FOLD, and UTF */
14024 /* Unlike the non-fold case, we do actually have to
14025 * calculate the fold in pass 1. This is for two reasons,
14026 * the folded length may be longer than the unfolded, and
14027 * we have to calculate how many EXACTish nodes it will
14028 * take; and we may run out of room in a node in the middle
14029 * of a potential multi-char fold, and have to back off
14032 if (isASCII_uni(ender)) {
14034 /* As above, we close up and start a new node if the
14035 * previous characters don't match the fold/non-fold
14036 * state of this one. And if this is the first
14037 * character in the node, and it folds, we change the
14038 * node away from being EXACT */
14039 if (! IS_IN_SOME_FOLD_L1(ender)) {
14040 if (len && node_type != EXACT) {
14045 *(s)++ = (U8) ender;
14047 else { /* Is in a fold */
14050 node_type = compute_EXACTish(pRExC_state);
14052 else if (node_type == EXACT) {
14057 *(s)++ = (U8) toFOLD(ender);
14060 else { /* Not ASCII */
14063 /* As above, we close up and start a new node if the
14064 * previous characters don't match the fold/non-fold
14065 * state of this one. And if this is the first
14066 * character in the node, and it folds, we change the
14067 * node away from being EXACT */
14068 if (! _invlist_contains_cp(PL_utf8_foldable, ender)) {
14069 if (len && node_type != EXACT) {
14074 s = (char *) uvchr_to_utf8((U8 *) s, ender);
14075 added_len = UVCHR_SKIP(ender);
14080 node_type = compute_EXACTish(pRExC_state);
14082 else if (node_type == EXACT) {
14087 ender = _to_uni_fold_flags(
14091 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
14092 ? FOLD_FLAGS_NOMIX_ASCII
14095 added_len = foldlen;
14102 if (next_is_quantifier) {
14104 /* Here, the next input is a quantifier, and to get here,
14105 * the current character is the only one in the node. */
14109 } /* End of loop through literal characters */
14111 /* Here we have either exhausted the input or ran out of room in
14112 * the node. (If we encountered a character that can't be in the
14113 * node, transfer is made directly to <loopdone>, and so we
14114 * wouldn't have fallen off the end of the loop.) In the latter
14115 * case, we artificially have to split the node into two, because
14116 * we just don't have enough space to hold everything. This
14117 * creates a problem if the final character participates in a
14118 * multi-character fold in the non-final position, as a match that
14119 * should have occurred won't, due to the way nodes are matched,
14120 * and our artificial boundary. So back off until we find a non-
14121 * problematic character -- one that isn't at the beginning or
14122 * middle of such a fold. (Either it doesn't participate in any
14123 * folds, or appears only in the final position of all the folds it
14124 * does participate in.) A better solution with far fewer false
14125 * positives, and that would fill the nodes more completely, would
14126 * be to actually have available all the multi-character folds to
14127 * test against, and to back-off only far enough to be sure that
14128 * this node isn't ending with a partial one. <upper_parse> is set
14129 * further below (if we need to reparse the node) to include just
14130 * up through that final non-problematic character that this code
14131 * identifies, so when it is set to less than the full node, we can
14132 * skip the rest of this */
14133 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
14135 const STRLEN full_len = len;
14137 assert(len >= MAX_NODE_STRING_SIZE);
14139 /* Here, <s> points to the final byte of the final character.
14140 * Look backwards through the string until find a non-
14141 * problematic character */
14145 /* This has no multi-char folds to non-UTF characters */
14146 if (ASCII_FOLD_RESTRICTED) {
14150 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
14155 /* Point to the first byte of the final character */
14156 s = (char *) utf8_hop((U8 *) s, -1);
14158 while (s >= s0) { /* Search backwards until find
14159 a non-problematic char */
14160 if (UTF8_IS_INVARIANT(*s)) {
14162 /* There are no ascii characters that participate
14163 * in multi-char folds under /aa. In EBCDIC, the
14164 * non-ascii invariants are all control characters,
14165 * so don't ever participate in any folds. */
14166 if (ASCII_FOLD_RESTRICTED
14167 || ! IS_NON_FINAL_FOLD(*s))
14172 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
14173 if (! IS_NON_FINAL_FOLD(EIGHT_BIT_UTF8_TO_NATIVE(
14179 else if (! _invlist_contains_cp(
14180 PL_NonL1NonFinalFold,
14181 valid_utf8_to_uvchr((U8 *) s, NULL)))
14186 /* Here, the current character is problematic in that
14187 * it does occur in the non-final position of some
14188 * fold, so try the character before it, but have to
14189 * special case the very first byte in the string, so
14190 * we don't read outside the string */
14191 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
14192 } /* End of loop backwards through the string */
14194 /* If there were only problematic characters in the string,
14195 * <s> will point to before s0, in which case the length
14196 * should be 0, otherwise include the length of the
14197 * non-problematic character just found */
14198 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
14201 /* Here, have found the final character, if any, that is
14202 * non-problematic as far as ending the node without splitting
14203 * it across a potential multi-char fold. <len> contains the
14204 * number of bytes in the node up-to and including that
14205 * character, or is 0 if there is no such character, meaning
14206 * the whole node contains only problematic characters. In
14207 * this case, give up and just take the node as-is. We can't
14212 /* If the node ends in an 's' we make sure it stays EXACTF,
14213 * as if it turns into an EXACTFU, it could later get
14214 * joined with another 's' that would then wrongly match
14216 if (maybe_exactfu && isALPHA_FOLD_EQ(ender, 's'))
14218 maybe_exactfu = FALSE;
14222 /* Here, the node does contain some characters that aren't
14223 * problematic. If one such is the final character in the
14224 * node, we are done */
14225 if (len == full_len) {
14228 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
14230 /* If the final character is problematic, but the
14231 * penultimate is not, back-off that last character to
14232 * later start a new node with it */
14237 /* Here, the final non-problematic character is earlier
14238 * in the input than the penultimate character. What we do
14239 * is reparse from the beginning, going up only as far as
14240 * this final ok one, thus guaranteeing that the node ends
14241 * in an acceptable character. The reason we reparse is
14242 * that we know how far in the character is, but we don't
14243 * know how to correlate its position with the input parse.
14244 * An alternate implementation would be to build that
14245 * correlation as we go along during the original parse,
14246 * but that would entail extra work for every node, whereas
14247 * this code gets executed only when the string is too
14248 * large for the node, and the final two characters are
14249 * problematic, an infrequent occurrence. Yet another
14250 * possible strategy would be to save the tail of the
14251 * string, and the next time regatom is called, initialize
14252 * with that. The problem with this is that unless you
14253 * back off one more character, you won't be guaranteed
14254 * regatom will get called again, unless regbranch,
14255 * regpiece ... are also changed. If you do back off that
14256 * extra character, so that there is input guaranteed to
14257 * force calling regatom, you can't handle the case where
14258 * just the first character in the node is acceptable. I
14259 * (khw) decided to try this method which doesn't have that
14260 * pitfall; if performance issues are found, we can do a
14261 * combination of the current approach plus that one */
14267 } /* End of verifying node ends with an appropriate char */
14269 loopdone: /* Jumped to when encounters something that shouldn't be
14272 /* I (khw) don't know if you can get here with zero length, but the
14273 * old code handled this situation by creating a zero-length EXACT
14274 * node. Might as well be NOTHING instead */
14279 OP(ret) = node_type;
14281 /* If the node type is EXACT here, check to see if it
14282 * should be EXACTL. */
14283 if (node_type == EXACT) {
14290 /* If 'maybe_exactfu' is set, then there are no code points
14291 * that match differently depending on UTF8ness of the
14292 * target string (for /u), or depending on locale for /l */
14293 if (maybe_exactfu) {
14294 if (node_type == EXACTF) {
14297 else if (node_type == EXACTFL) {
14298 OP(ret) = EXACTFLU8;
14303 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender,
14304 FALSE /* Don't look to see if could
14305 be turned into an EXACT
14306 node, as we have already
14311 RExC_parse = p - 1;
14312 Set_Node_Cur_Length(ret, parse_start);
14315 /* len is STRLEN which is unsigned, need to copy to signed */
14318 vFAIL("Internal disaster");
14321 } /* End of label 'defchar:' */
14323 } /* End of giant switch on input character */
14325 /* Position parse to next real character */
14326 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14327 FALSE /* Don't force to /x */ );
14328 if ( PASS2 && *RExC_parse == '{'
14329 && OP(ret) != SBOL && ! regcurly(RExC_parse))
14331 if (RExC_strict || new_regcurly(RExC_parse, RExC_end)) {
14333 vFAIL("Unescaped left brace in regex is illegal here");
14335 ckWARNreg(RExC_parse + 1, "Unescaped left brace in regex is"
14336 " passed through");
14344 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
14346 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
14347 * sets up the bitmap and any flags, removing those code points from the
14348 * inversion list, setting it to NULL should it become completely empty */
14350 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
14351 assert(PL_regkind[OP(node)] == ANYOF);
14353 ANYOF_BITMAP_ZERO(node);
14354 if (*invlist_ptr) {
14356 /* This gets set if we actually need to modify things */
14357 bool change_invlist = FALSE;
14361 /* Start looking through *invlist_ptr */
14362 invlist_iterinit(*invlist_ptr);
14363 while (invlist_iternext(*invlist_ptr, &start, &end)) {
14367 if (end == UV_MAX && start <= NUM_ANYOF_CODE_POINTS) {
14368 ANYOF_FLAGS(node) |= ANYOF_MATCHES_ALL_ABOVE_BITMAP;
14371 /* Quit if are above what we should change */
14372 if (start >= NUM_ANYOF_CODE_POINTS) {
14376 change_invlist = TRUE;
14378 /* Set all the bits in the range, up to the max that we are doing */
14379 high = (end < NUM_ANYOF_CODE_POINTS - 1)
14381 : NUM_ANYOF_CODE_POINTS - 1;
14382 for (i = start; i <= (int) high; i++) {
14383 if (! ANYOF_BITMAP_TEST(node, i)) {
14384 ANYOF_BITMAP_SET(node, i);
14388 invlist_iterfinish(*invlist_ptr);
14390 /* Done with loop; remove any code points that are in the bitmap from
14391 * *invlist_ptr; similarly for code points above the bitmap if we have
14392 * a flag to match all of them anyways */
14393 if (change_invlist) {
14394 _invlist_subtract(*invlist_ptr, PL_InBitmap, invlist_ptr);
14396 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
14397 _invlist_intersection(*invlist_ptr, PL_InBitmap, invlist_ptr);
14400 /* If have completely emptied it, remove it completely */
14401 if (_invlist_len(*invlist_ptr) == 0) {
14402 SvREFCNT_dec_NN(*invlist_ptr);
14403 *invlist_ptr = NULL;
14408 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
14409 Character classes ([:foo:]) can also be negated ([:^foo:]).
14410 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
14411 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
14412 but trigger failures because they are currently unimplemented. */
14414 #define POSIXCC_DONE(c) ((c) == ':')
14415 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
14416 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
14417 #define MAYBE_POSIXCC(c) (POSIXCC(c) || (c) == '^' || (c) == ';')
14419 #define WARNING_PREFIX "Assuming NOT a POSIX class since "
14420 #define NO_BLANKS_POSIX_WARNING "no blanks are allowed in one"
14421 #define SEMI_COLON_POSIX_WARNING "a semi-colon was found instead of a colon"
14423 #define NOT_MEANT_TO_BE_A_POSIX_CLASS (OOB_NAMEDCLASS - 1)
14425 /* 'posix_warnings' and 'warn_text' are names of variables in the following
14427 #define ADD_POSIX_WARNING(p, text) STMT_START { \
14428 if (posix_warnings) { \
14429 if (! RExC_warn_text ) RExC_warn_text = (AV *) sv_2mortal((SV *) newAV()); \
14430 av_push(RExC_warn_text, Perl_newSVpvf(aTHX_ \
14434 REPORT_LOCATION_ARGS(p))); \
14437 #define CLEAR_POSIX_WARNINGS() \
14439 if (posix_warnings && RExC_warn_text) \
14440 av_clear(RExC_warn_text); \
14443 #define CLEAR_POSIX_WARNINGS_AND_RETURN(ret) \
14445 CLEAR_POSIX_WARNINGS(); \
14450 S_handle_possible_posix(pTHX_ RExC_state_t *pRExC_state,
14452 const char * const s, /* Where the putative posix class begins.
14453 Normally, this is one past the '['. This
14454 parameter exists so it can be somewhere
14455 besides RExC_parse. */
14456 char ** updated_parse_ptr, /* Where to set the updated parse pointer, or
14458 AV ** posix_warnings, /* Where to place any generated warnings, or
14460 const bool check_only /* Don't die if error */
14463 /* This parses what the caller thinks may be one of the three POSIX
14465 * 1) a character class, like [:blank:]
14466 * 2) a collating symbol, like [. .]
14467 * 3) an equivalence class, like [= =]
14468 * In the latter two cases, it croaks if it finds a syntactically legal
14469 * one, as these are not handled by Perl.
14471 * The main purpose is to look for a POSIX character class. It returns:
14472 * a) the class number
14473 * if it is a completely syntactically and semantically legal class.
14474 * 'updated_parse_ptr', if not NULL, is set to point to just after the
14475 * closing ']' of the class
14476 * b) OOB_NAMEDCLASS
14477 * if it appears that one of the three POSIX constructs was meant, but
14478 * its specification was somehow defective. 'updated_parse_ptr', if
14479 * not NULL, is set to point to the character just after the end
14480 * character of the class. See below for handling of warnings.
14481 * c) NOT_MEANT_TO_BE_A_POSIX_CLASS
14482 * if it doesn't appear that a POSIX construct was intended.
14483 * 'updated_parse_ptr' is not changed. No warnings nor errors are
14486 * In b) there may be errors or warnings generated. If 'check_only' is
14487 * TRUE, then any errors are discarded. Warnings are returned to the
14488 * caller via an AV* created into '*posix_warnings' if it is not NULL. If
14489 * instead it is NULL, warnings are suppressed. This is done in all
14490 * passes. The reason for this is that the rest of the parsing is heavily
14491 * dependent on whether this routine found a valid posix class or not. If
14492 * it did, the closing ']' is absorbed as part of the class. If no class,
14493 * or an invalid one is found, any ']' will be considered the terminator of
14494 * the outer bracketed character class, leading to very different results.
14495 * In particular, a '(?[ ])' construct will likely have a syntax error if
14496 * the class is parsed other than intended, and this will happen in pass1,
14497 * before the warnings would normally be output. This mechanism allows the
14498 * caller to output those warnings in pass1 just before dieing, giving a
14499 * much better clue as to what is wrong.
14501 * The reason for this function, and its complexity is that a bracketed
14502 * character class can contain just about anything. But it's easy to
14503 * mistype the very specific posix class syntax but yielding a valid
14504 * regular bracketed class, so it silently gets compiled into something
14505 * quite unintended.
14507 * The solution adopted here maintains backward compatibility except that
14508 * it adds a warning if it looks like a posix class was intended but
14509 * improperly specified. The warning is not raised unless what is input
14510 * very closely resembles one of the 14 legal posix classes. To do this,
14511 * it uses fuzzy parsing. It calculates how many single-character edits it
14512 * would take to transform what was input into a legal posix class. Only
14513 * if that number is quite small does it think that the intention was a
14514 * posix class. Obviously these are heuristics, and there will be cases
14515 * where it errs on one side or another, and they can be tweaked as
14516 * experience informs.
14518 * The syntax for a legal posix class is:
14520 * qr/(?xa: \[ : \^? [[:lower:]]{4,6} : \] )/
14522 * What this routine considers syntactically to be an intended posix class
14523 * is this (the comments indicate some restrictions that the pattern
14526 * qr/(?x: \[? # The left bracket, possibly
14528 * \h* # possibly followed by blanks
14529 * (?: \^ \h* )? # possibly a misplaced caret
14530 * [:;]? # The opening class character,
14531 * # possibly omitted. A typo
14532 * # semi-colon can also be used.
14534 * \^? # possibly a correctly placed
14535 * # caret, but not if there was also
14536 * # a misplaced one
14538 * .{3,15} # The class name. If there are
14539 * # deviations from the legal syntax,
14540 * # its edit distance must be close
14541 * # to a real class name in order
14542 * # for it to be considered to be
14543 * # an intended posix class.
14545 * [[:punct:]]? # The closing class character,
14546 * # possibly omitted. If not a colon
14547 * # nor semi colon, the class name
14548 * # must be even closer to a valid
14551 * \]? # The right bracket, possibly
14555 * In the above, \h must be ASCII-only.
14557 * These are heuristics, and can be tweaked as field experience dictates.
14558 * There will be cases when someone didn't intend to specify a posix class
14559 * that this warns as being so. The goal is to minimize these, while
14560 * maximizing the catching of things intended to be a posix class that
14561 * aren't parsed as such.
14565 const char * const e = RExC_end;
14566 unsigned complement = 0; /* If to complement the class */
14567 bool found_problem = FALSE; /* Assume OK until proven otherwise */
14568 bool has_opening_bracket = FALSE;
14569 bool has_opening_colon = FALSE;
14570 int class_number = OOB_NAMEDCLASS; /* Out-of-bounds until find
14572 const char * possible_end = NULL; /* used for a 2nd parse pass */
14573 const char* name_start; /* ptr to class name first char */
14575 /* If the number of single-character typos the input name is away from a
14576 * legal name is no more than this number, it is considered to have meant
14577 * the legal name */
14578 int max_distance = 2;
14580 /* to store the name. The size determines the maximum length before we
14581 * decide that no posix class was intended. Should be at least
14582 * sizeof("alphanumeric") */
14584 STATIC_ASSERT_DECL(C_ARRAY_LENGTH(input_text) >= sizeof "alphanumeric");
14586 PERL_ARGS_ASSERT_HANDLE_POSSIBLE_POSIX;
14588 CLEAR_POSIX_WARNINGS();
14591 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14594 if (*(p - 1) != '[') {
14595 ADD_POSIX_WARNING(p, "it doesn't start with a '['");
14596 found_problem = TRUE;
14599 has_opening_bracket = TRUE;
14602 /* They could be confused and think you can put spaces between the
14605 found_problem = TRUE;
14609 } while (p < e && isBLANK(*p));
14611 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14614 /* For [. .] and [= =]. These are quite different internally from [: :],
14615 * so they are handled separately. */
14616 if (POSIXCC_NOTYET(*p) && p < e - 3) /* 1 for the close, and 1 for the ']'
14617 and 1 for at least one char in it
14620 const char open_char = *p;
14621 const char * temp_ptr = p + 1;
14623 /* These two constructs are not handled by perl, and if we find a
14624 * syntactically valid one, we croak. khw, who wrote this code, finds
14625 * this explanation of them very unclear:
14626 * http://pubs.opengroup.org/onlinepubs/009696899/basedefs/xbd_chap09.html
14627 * And searching the rest of the internet wasn't very helpful either.
14628 * It looks like just about any byte can be in these constructs,
14629 * depending on the locale. But unless the pattern is being compiled
14630 * under /l, which is very rare, Perl runs under the C or POSIX locale.
14631 * In that case, it looks like [= =] isn't allowed at all, and that
14632 * [. .] could be any single code point, but for longer strings the
14633 * constituent characters would have to be the ASCII alphabetics plus
14634 * the minus-hyphen. Any sensible locale definition would limit itself
14635 * to these. And any portable one definitely should. Trying to parse
14636 * the general case is a nightmare (see [perl #127604]). So, this code
14637 * looks only for interiors of these constructs that match:
14639 * Using \w relaxes the apparent rules a little, without adding much
14640 * danger of mistaking something else for one of these constructs.
14642 * [. .] in some implementations described on the internet is usable to
14643 * escape a character that otherwise is special in bracketed character
14644 * classes. For example [.].] means a literal right bracket instead of
14645 * the ending of the class
14647 * [= =] can legitimately contain a [. .] construct, but we don't
14648 * handle this case, as that [. .] construct will later get parsed
14649 * itself and croak then. And [= =] is checked for even when not under
14650 * /l, as Perl has long done so.
14652 * The code below relies on there being a trailing NUL, so it doesn't
14653 * have to keep checking if the parse ptr < e.
14655 if (temp_ptr[1] == open_char) {
14658 else while ( temp_ptr < e
14659 && (isWORDCHAR(*temp_ptr) || *temp_ptr == '-'))
14664 if (*temp_ptr == open_char) {
14666 if (*temp_ptr == ']') {
14668 if (! found_problem && ! check_only) {
14669 RExC_parse = (char *) temp_ptr;
14670 vFAIL3("POSIX syntax [%c %c] is reserved for future "
14671 "extensions", open_char, open_char);
14674 /* Here, the syntax wasn't completely valid, or else the call
14675 * is to check-only */
14676 if (updated_parse_ptr) {
14677 *updated_parse_ptr = (char *) temp_ptr;
14680 CLEAR_POSIX_WARNINGS_AND_RETURN(OOB_NAMEDCLASS);
14684 /* If we find something that started out to look like one of these
14685 * constructs, but isn't, we continue below so that it can be checked
14686 * for being a class name with a typo of '.' or '=' instead of a colon.
14690 /* Here, we think there is a possibility that a [: :] class was meant, and
14691 * we have the first real character. It could be they think the '^' comes
14694 found_problem = TRUE;
14695 ADD_POSIX_WARNING(p + 1, "the '^' must come after the colon");
14700 found_problem = TRUE;
14704 } while (p < e && isBLANK(*p));
14706 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14710 /* But the first character should be a colon, which they could have easily
14711 * mistyped on a qwerty keyboard as a semi-colon (and which may be hard to
14712 * distinguish from a colon, so treat that as a colon). */
14715 has_opening_colon = TRUE;
14717 else if (*p == ';') {
14718 found_problem = TRUE;
14720 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14721 has_opening_colon = TRUE;
14724 found_problem = TRUE;
14725 ADD_POSIX_WARNING(p, "there must be a starting ':'");
14727 /* Consider an initial punctuation (not one of the recognized ones) to
14728 * be a left terminator */
14729 if (*p != '^' && *p != ']' && isPUNCT(*p)) {
14734 /* They may think that you can put spaces between the components */
14736 found_problem = TRUE;
14740 } while (p < e && isBLANK(*p));
14742 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14747 /* We consider something like [^:^alnum:]] to not have been intended to
14748 * be a posix class, but XXX maybe we should */
14750 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
14757 /* Again, they may think that you can put spaces between the components */
14759 found_problem = TRUE;
14763 } while (p < e && isBLANK(*p));
14765 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14770 /* XXX This ']' may be a typo, and something else was meant. But
14771 * treating it as such creates enough complications, that that
14772 * possibility isn't currently considered here. So we assume that the
14773 * ']' is what is intended, and if we've already found an initial '[',
14774 * this leaves this construct looking like [:] or [:^], which almost
14775 * certainly weren't intended to be posix classes */
14776 if (has_opening_bracket) {
14777 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
14780 /* But this function can be called when we parse the colon for
14781 * something like qr/[alpha:]]/, so we back up to look for the
14786 found_problem = TRUE;
14787 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14789 else if (*p != ':') {
14791 /* XXX We are currently very restrictive here, so this code doesn't
14792 * consider the possibility that, say, /[alpha.]]/ was intended to
14793 * be a posix class. */
14794 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
14797 /* Here we have something like 'foo:]'. There was no initial colon,
14798 * and we back up over 'foo. XXX Unlike the going forward case, we
14799 * don't handle typos of non-word chars in the middle */
14800 has_opening_colon = FALSE;
14803 while (p > RExC_start && isWORDCHAR(*p)) {
14808 /* Here, we have positioned ourselves to where we think the first
14809 * character in the potential class is */
14812 /* Now the interior really starts. There are certain key characters that
14813 * can end the interior, or these could just be typos. To catch both
14814 * cases, we may have to do two passes. In the first pass, we keep on
14815 * going unless we come to a sequence that matches
14816 * qr/ [[:punct:]] [[:blank:]]* \] /xa
14817 * This means it takes a sequence to end the pass, so two typos in a row if
14818 * that wasn't what was intended. If the class is perfectly formed, just
14819 * this one pass is needed. We also stop if there are too many characters
14820 * being accumulated, but this number is deliberately set higher than any
14821 * real class. It is set high enough so that someone who thinks that
14822 * 'alphanumeric' is a correct name would get warned that it wasn't.
14823 * While doing the pass, we keep track of where the key characters were in
14824 * it. If we don't find an end to the class, and one of the key characters
14825 * was found, we redo the pass, but stop when we get to that character.
14826 * Thus the key character was considered a typo in the first pass, but a
14827 * terminator in the second. If two key characters are found, we stop at
14828 * the second one in the first pass. Again this can miss two typos, but
14829 * catches a single one
14831 * In the first pass, 'possible_end' starts as NULL, and then gets set to
14832 * point to the first key character. For the second pass, it starts as -1.
14838 bool has_blank = FALSE;
14839 bool has_upper = FALSE;
14840 bool has_terminating_colon = FALSE;
14841 bool has_terminating_bracket = FALSE;
14842 bool has_semi_colon = FALSE;
14843 unsigned int name_len = 0;
14844 int punct_count = 0;
14848 /* Squeeze out blanks when looking up the class name below */
14849 if (isBLANK(*p) ) {
14851 found_problem = TRUE;
14856 /* The name will end with a punctuation */
14858 const char * peek = p + 1;
14860 /* Treat any non-']' punctuation followed by a ']' (possibly
14861 * with intervening blanks) as trying to terminate the class.
14862 * ']]' is very likely to mean a class was intended (but
14863 * missing the colon), but the warning message that gets
14864 * generated shows the error position better if we exit the
14865 * loop at the bottom (eventually), so skip it here. */
14867 if (peek < e && isBLANK(*peek)) {
14869 found_problem = TRUE;
14872 } while (peek < e && isBLANK(*peek));
14875 if (peek < e && *peek == ']') {
14876 has_terminating_bracket = TRUE;
14878 has_terminating_colon = TRUE;
14880 else if (*p == ';') {
14881 has_semi_colon = TRUE;
14882 has_terminating_colon = TRUE;
14885 found_problem = TRUE;
14892 /* Here we have punctuation we thought didn't end the class.
14893 * Keep track of the position of the key characters that are
14894 * more likely to have been class-enders */
14895 if (*p == ']' || *p == '[' || *p == ':' || *p == ';') {
14897 /* Allow just one such possible class-ender not actually
14898 * ending the class. */
14899 if (possible_end) {
14905 /* If we have too many punctuation characters, no use in
14907 if (++punct_count > max_distance) {
14911 /* Treat the punctuation as a typo. */
14912 input_text[name_len++] = *p;
14915 else if (isUPPER(*p)) { /* Use lowercase for lookup */
14916 input_text[name_len++] = toLOWER(*p);
14918 found_problem = TRUE;
14920 } else if (! UTF || UTF8_IS_INVARIANT(*p)) {
14921 input_text[name_len++] = *p;
14925 input_text[name_len++] = utf8_to_uvchr_buf((U8 *) p, e, NULL);
14929 /* The declaration of 'input_text' is how long we allow a potential
14930 * class name to be, before saying they didn't mean a class name at
14932 if (name_len >= C_ARRAY_LENGTH(input_text)) {
14937 /* We get to here when the possible class name hasn't been properly
14938 * terminated before:
14939 * 1) we ran off the end of the pattern; or
14940 * 2) found two characters, each of which might have been intended to
14941 * be the name's terminator
14942 * 3) found so many punctuation characters in the purported name,
14943 * that the edit distance to a valid one is exceeded
14944 * 4) we decided it was more characters than anyone could have
14945 * intended to be one. */
14947 found_problem = TRUE;
14949 /* In the final two cases, we know that looking up what we've
14950 * accumulated won't lead to a match, even a fuzzy one. */
14951 if ( name_len >= C_ARRAY_LENGTH(input_text)
14952 || punct_count > max_distance)
14954 /* If there was an intermediate key character that could have been
14955 * an intended end, redo the parse, but stop there */
14956 if (possible_end && possible_end != (char *) -1) {
14957 possible_end = (char *) -1; /* Special signal value to say
14958 we've done a first pass */
14963 /* Otherwise, it can't have meant to have been a class */
14964 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
14967 /* If we ran off the end, and the final character was a punctuation
14968 * one, back up one, to look at that final one just below. Later, we
14969 * will restore the parse pointer if appropriate */
14970 if (name_len && p == e && isPUNCT(*(p-1))) {
14975 if (p < e && isPUNCT(*p)) {
14977 has_terminating_bracket = TRUE;
14979 /* If this is a 2nd ']', and the first one is just below this
14980 * one, consider that to be the real terminator. This gives a
14981 * uniform and better positioning for the warning message */
14983 && possible_end != (char *) -1
14984 && *possible_end == ']'
14985 && name_len && input_text[name_len - 1] == ']')
14990 /* And this is actually equivalent to having done the 2nd
14991 * pass now, so set it to not try again */
14992 possible_end = (char *) -1;
14997 has_terminating_colon = TRUE;
14999 else if (*p == ';') {
15000 has_semi_colon = TRUE;
15001 has_terminating_colon = TRUE;
15009 /* Here, we have a class name to look up. We can short circuit the
15010 * stuff below for short names that can't possibly be meant to be a
15011 * class name. (We can do this on the first pass, as any second pass
15012 * will yield an even shorter name) */
15013 if (name_len < 3) {
15014 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15017 /* Find which class it is. Initially switch on the length of the name.
15019 switch (name_len) {
15021 if (memEQs(name_start, 4, "word")) {
15022 /* this is not POSIX, this is the Perl \w */
15023 class_number = ANYOF_WORDCHAR;
15027 /* Names all of length 5: alnum alpha ascii blank cntrl digit
15028 * graph lower print punct space upper
15029 * Offset 4 gives the best switch position. */
15030 switch (name_start[4]) {
15032 if (memBEGINs(name_start, 5, "alph")) /* alpha */
15033 class_number = ANYOF_ALPHA;
15036 if (memBEGINs(name_start, 5, "spac")) /* space */
15037 class_number = ANYOF_SPACE;
15040 if (memBEGINs(name_start, 5, "grap")) /* graph */
15041 class_number = ANYOF_GRAPH;
15044 if (memBEGINs(name_start, 5, "asci")) /* ascii */
15045 class_number = ANYOF_ASCII;
15048 if (memBEGINs(name_start, 5, "blan")) /* blank */
15049 class_number = ANYOF_BLANK;
15052 if (memBEGINs(name_start, 5, "cntr")) /* cntrl */
15053 class_number = ANYOF_CNTRL;
15056 if (memBEGINs(name_start, 5, "alnu")) /* alnum */
15057 class_number = ANYOF_ALPHANUMERIC;
15060 if (memBEGINs(name_start, 5, "lowe")) /* lower */
15061 class_number = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
15062 else if (memBEGINs(name_start, 5, "uppe")) /* upper */
15063 class_number = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
15066 if (memBEGINs(name_start, 5, "digi")) /* digit */
15067 class_number = ANYOF_DIGIT;
15068 else if (memBEGINs(name_start, 5, "prin")) /* print */
15069 class_number = ANYOF_PRINT;
15070 else if (memBEGINs(name_start, 5, "punc")) /* punct */
15071 class_number = ANYOF_PUNCT;
15076 if (memEQs(name_start, 6, "xdigit"))
15077 class_number = ANYOF_XDIGIT;
15081 /* If the name exactly matches a posix class name the class number will
15082 * here be set to it, and the input almost certainly was meant to be a
15083 * posix class, so we can skip further checking. If instead the syntax
15084 * is exactly correct, but the name isn't one of the legal ones, we
15085 * will return that as an error below. But if neither of these apply,
15086 * it could be that no posix class was intended at all, or that one
15087 * was, but there was a typo. We tease these apart by doing fuzzy
15088 * matching on the name */
15089 if (class_number == OOB_NAMEDCLASS && found_problem) {
15090 const UV posix_names[][6] = {
15091 { 'a', 'l', 'n', 'u', 'm' },
15092 { 'a', 'l', 'p', 'h', 'a' },
15093 { 'a', 's', 'c', 'i', 'i' },
15094 { 'b', 'l', 'a', 'n', 'k' },
15095 { 'c', 'n', 't', 'r', 'l' },
15096 { 'd', 'i', 'g', 'i', 't' },
15097 { 'g', 'r', 'a', 'p', 'h' },
15098 { 'l', 'o', 'w', 'e', 'r' },
15099 { 'p', 'r', 'i', 'n', 't' },
15100 { 'p', 'u', 'n', 'c', 't' },
15101 { 's', 'p', 'a', 'c', 'e' },
15102 { 'u', 'p', 'p', 'e', 'r' },
15103 { 'w', 'o', 'r', 'd' },
15104 { 'x', 'd', 'i', 'g', 'i', 't' }
15106 /* The names of the above all have added NULs to make them the same
15107 * size, so we need to also have the real lengths */
15108 const UV posix_name_lengths[] = {
15109 sizeof("alnum") - 1,
15110 sizeof("alpha") - 1,
15111 sizeof("ascii") - 1,
15112 sizeof("blank") - 1,
15113 sizeof("cntrl") - 1,
15114 sizeof("digit") - 1,
15115 sizeof("graph") - 1,
15116 sizeof("lower") - 1,
15117 sizeof("print") - 1,
15118 sizeof("punct") - 1,
15119 sizeof("space") - 1,
15120 sizeof("upper") - 1,
15121 sizeof("word") - 1,
15122 sizeof("xdigit")- 1
15125 int temp_max = max_distance; /* Use a temporary, so if we
15126 reparse, we haven't changed the
15129 /* Use a smaller max edit distance if we are missing one of the
15131 if ( has_opening_bracket + has_opening_colon < 2
15132 || has_terminating_bracket + has_terminating_colon < 2)
15137 /* See if the input name is close to a legal one */
15138 for (i = 0; i < C_ARRAY_LENGTH(posix_names); i++) {
15140 /* Short circuit call if the lengths are too far apart to be
15142 if (abs( (int) (name_len - posix_name_lengths[i]))
15148 if (edit_distance(input_text,
15151 posix_name_lengths[i],
15155 { /* If it is close, it probably was intended to be a class */
15156 goto probably_meant_to_be;
15160 /* Here the input name is not close enough to a valid class name
15161 * for us to consider it to be intended to be a posix class. If
15162 * we haven't already done so, and the parse found a character that
15163 * could have been terminators for the name, but which we absorbed
15164 * as typos during the first pass, repeat the parse, signalling it
15165 * to stop at that character */
15166 if (possible_end && possible_end != (char *) -1) {
15167 possible_end = (char *) -1;
15172 /* Here neither pass found a close-enough class name */
15173 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15176 probably_meant_to_be:
15178 /* Here we think that a posix specification was intended. Update any
15180 if (updated_parse_ptr) {
15181 *updated_parse_ptr = (char *) p;
15184 /* If a posix class name was intended but incorrectly specified, we
15185 * output or return the warnings */
15186 if (found_problem) {
15188 /* We set flags for these issues in the parse loop above instead of
15189 * adding them to the list of warnings, because we can parse it
15190 * twice, and we only want one warning instance */
15192 ADD_POSIX_WARNING(p, "the name must be all lowercase letters");
15195 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15197 if (has_semi_colon) {
15198 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
15200 else if (! has_terminating_colon) {
15201 ADD_POSIX_WARNING(p, "there is no terminating ':'");
15203 if (! has_terminating_bracket) {
15204 ADD_POSIX_WARNING(p, "there is no terminating ']'");
15207 if (posix_warnings && RExC_warn_text && av_top_index(RExC_warn_text) > -1) {
15208 *posix_warnings = RExC_warn_text;
15211 else if (class_number != OOB_NAMEDCLASS) {
15212 /* If it is a known class, return the class. The class number
15213 * #defines are structured so each complement is +1 to the normal
15215 CLEAR_POSIX_WARNINGS_AND_RETURN(class_number + complement);
15217 else if (! check_only) {
15219 /* Here, it is an unrecognized class. This is an error (unless the
15220 * call is to check only, which we've already handled above) */
15221 const char * const complement_string = (complement)
15224 RExC_parse = (char *) p;
15225 vFAIL3utf8f("POSIX class [:%s%" UTF8f ":] unknown",
15227 UTF8fARG(UTF, RExC_parse - name_start - 2, name_start));
15231 return OOB_NAMEDCLASS;
15233 #undef ADD_POSIX_WARNING
15235 STATIC unsigned int
15236 S_regex_set_precedence(const U8 my_operator) {
15238 /* Returns the precedence in the (?[...]) construct of the input operator,
15239 * specified by its character representation. The precedence follows
15240 * general Perl rules, but it extends this so that ')' and ']' have (low)
15241 * precedence even though they aren't really operators */
15243 switch (my_operator) {
15259 NOT_REACHED; /* NOTREACHED */
15260 return 0; /* Silence compiler warning */
15264 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
15265 I32 *flagp, U32 depth,
15266 char * const oregcomp_parse)
15268 /* Handle the (?[...]) construct to do set operations */
15270 U8 curchar; /* Current character being parsed */
15271 UV start, end; /* End points of code point ranges */
15272 SV* final = NULL; /* The end result inversion list */
15273 SV* result_string; /* 'final' stringified */
15274 AV* stack; /* stack of operators and operands not yet
15276 AV* fence_stack = NULL; /* A stack containing the positions in
15277 'stack' of where the undealt-with left
15278 parens would be if they were actually
15280 /* The 'volatile' is a workaround for an optimiser bug
15281 * in Solaris Studio 12.3. See RT #127455 */
15282 volatile IV fence = 0; /* Position of where most recent undealt-
15283 with left paren in stack is; -1 if none.
15285 STRLEN len; /* Temporary */
15286 regnode* node; /* Temporary, and final regnode returned by
15288 const bool save_fold = FOLD; /* Temporary */
15289 char *save_end, *save_parse; /* Temporaries */
15290 const bool in_locale = LOC; /* we turn off /l during processing */
15291 AV* posix_warnings = NULL;
15293 GET_RE_DEBUG_FLAGS_DECL;
15295 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
15297 DEBUG_PARSE("xcls");
15300 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
15303 REQUIRE_UNI_RULES(flagp, NULL); /* The use of this operator implies /u.
15304 This is required so that the compile
15305 time values are valid in all runtime
15308 /* This will return only an ANYOF regnode, or (unlikely) something smaller
15309 * (such as EXACT). Thus we can skip most everything if just sizing. We
15310 * call regclass to handle '[]' so as to not have to reinvent its parsing
15311 * rules here (throwing away the size it computes each time). And, we exit
15312 * upon an unescaped ']' that isn't one ending a regclass. To do both
15313 * these things, we need to realize that something preceded by a backslash
15314 * is escaped, so we have to keep track of backslashes */
15316 UV nest_depth = 0; /* how many nested (?[...]) constructs */
15318 while (RExC_parse < RExC_end) {
15319 SV* current = NULL;
15321 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
15322 TRUE /* Force /x */ );
15324 switch (*RExC_parse) {
15326 if (RExC_parse[1] == '?' && RExC_parse[2] == '[')
15327 nest_depth++, RExC_parse+=2;
15332 /* Skip past this, so the next character gets skipped, after
15335 if (*RExC_parse == 'c') {
15336 /* Skip the \cX notation for control characters */
15337 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
15343 /* See if this is a [:posix:] class. */
15344 bool is_posix_class = (OOB_NAMEDCLASS
15345 < handle_possible_posix(pRExC_state,
15349 TRUE /* checking only */));
15350 /* If it is a posix class, leave the parse pointer at the
15351 * '[' to fool regclass() into thinking it is part of a
15352 * '[[:posix:]]'. */
15353 if (! is_posix_class) {
15357 /* regclass() can only return RESTART_PASS1 and NEED_UTF8
15358 * if multi-char folds are allowed. */
15359 if (!regclass(pRExC_state, flagp,depth+1,
15360 is_posix_class, /* parse the whole char
15361 class only if not a
15363 FALSE, /* don't allow multi-char folds */
15364 TRUE, /* silence non-portable warnings. */
15366 FALSE, /* Require return to be an ANYOF */
15370 FAIL2("panic: regclass returned NULL to handle_sets, "
15371 "flags=%#" UVxf, (UV) *flagp);
15373 /* function call leaves parse pointing to the ']', except
15374 * if we faked it */
15375 if (is_posix_class) {
15379 SvREFCNT_dec(current); /* In case it returned something */
15384 if (RExC_parse[1] == ')') {
15386 if (nest_depth--) break;
15387 node = reganode(pRExC_state, ANYOF, 0);
15388 RExC_size += ANYOF_SKIP;
15389 nextchar(pRExC_state);
15390 Set_Node_Length(node,
15391 RExC_parse - oregcomp_parse + 1); /* MJD */
15393 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
15398 /* We output the messages even if warnings are off, because we'll fail
15399 * the very next thing, and these give a likely diagnosis for that */
15400 if (posix_warnings && av_tindex_skip_len_mg(posix_warnings) >= 0) {
15401 output_or_return_posix_warnings(pRExC_state, posix_warnings, NULL);
15404 vFAIL("Unexpected ']' with no following ')' in (?[...");
15407 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
15410 /* We output the messages even if warnings are off, because we'll fail
15411 * the very next thing, and these give a likely diagnosis for that */
15412 if (posix_warnings && av_tindex_skip_len_mg(posix_warnings) >= 0) {
15413 output_or_return_posix_warnings(pRExC_state, posix_warnings, NULL);
15416 vFAIL("Syntax error in (?[...])");
15419 /* Pass 2 only after this. */
15420 Perl_ck_warner_d(aTHX_
15421 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
15422 "The regex_sets feature is experimental" REPORT_LOCATION,
15423 REPORT_LOCATION_ARGS(RExC_parse));
15425 /* Everything in this construct is a metacharacter. Operands begin with
15426 * either a '\' (for an escape sequence), or a '[' for a bracketed
15427 * character class. Any other character should be an operator, or
15428 * parenthesis for grouping. Both types of operands are handled by calling
15429 * regclass() to parse them. It is called with a parameter to indicate to
15430 * return the computed inversion list. The parsing here is implemented via
15431 * a stack. Each entry on the stack is a single character representing one
15432 * of the operators; or else a pointer to an operand inversion list. */
15434 #define IS_OPERATOR(a) SvIOK(a)
15435 #define IS_OPERAND(a) (! IS_OPERATOR(a))
15437 /* The stack is kept in Łukasiewicz order. (That's pronounced similar
15438 * to luke-a-shave-itch (or -itz), but people who didn't want to bother
15439 * with pronouncing it called it Reverse Polish instead, but now that YOU
15440 * know how to pronounce it you can use the correct term, thus giving due
15441 * credit to the person who invented it, and impressing your geek friends.
15442 * Wikipedia says that the pronounciation of "Ł" has been changing so that
15443 * it is now more like an English initial W (as in wonk) than an L.)
15445 * This means that, for example, 'a | b & c' is stored on the stack as
15453 * where the numbers in brackets give the stack [array] element number.
15454 * In this implementation, parentheses are not stored on the stack.
15455 * Instead a '(' creates a "fence" so that the part of the stack below the
15456 * fence is invisible except to the corresponding ')' (this allows us to
15457 * replace testing for parens, by using instead subtraction of the fence
15458 * position). As new operands are processed they are pushed onto the stack
15459 * (except as noted in the next paragraph). New operators of higher
15460 * precedence than the current final one are inserted on the stack before
15461 * the lhs operand (so that when the rhs is pushed next, everything will be
15462 * in the correct positions shown above. When an operator of equal or
15463 * lower precedence is encountered in parsing, all the stacked operations
15464 * of equal or higher precedence are evaluated, leaving the result as the
15465 * top entry on the stack. This makes higher precedence operations
15466 * evaluate before lower precedence ones, and causes operations of equal
15467 * precedence to left associate.
15469 * The only unary operator '!' is immediately pushed onto the stack when
15470 * encountered. When an operand is encountered, if the top of the stack is
15471 * a '!", the complement is immediately performed, and the '!' popped. The
15472 * resulting value is treated as a new operand, and the logic in the
15473 * previous paragraph is executed. Thus in the expression
15475 * the stack looks like
15481 * as 'b' gets parsed, the latter gets evaluated to '!b', and the stack
15488 * A ')' is treated as an operator with lower precedence than all the
15489 * aforementioned ones, which causes all operations on the stack above the
15490 * corresponding '(' to be evaluated down to a single resultant operand.
15491 * Then the fence for the '(' is removed, and the operand goes through the
15492 * algorithm above, without the fence.
15494 * A separate stack is kept of the fence positions, so that the position of
15495 * the latest so-far unbalanced '(' is at the top of it.
15497 * The ']' ending the construct is treated as the lowest operator of all,
15498 * so that everything gets evaluated down to a single operand, which is the
15501 sv_2mortal((SV *)(stack = newAV()));
15502 sv_2mortal((SV *)(fence_stack = newAV()));
15504 while (RExC_parse < RExC_end) {
15505 I32 top_index; /* Index of top-most element in 'stack' */
15506 SV** top_ptr; /* Pointer to top 'stack' element */
15507 SV* current = NULL; /* To contain the current inversion list
15509 SV* only_to_avoid_leaks;
15511 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
15512 TRUE /* Force /x */ );
15513 if (RExC_parse >= RExC_end) {
15514 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
15517 curchar = UCHARAT(RExC_parse);
15521 #ifdef ENABLE_REGEX_SETS_DEBUGGING
15522 /* Enable with -Accflags=-DENABLE_REGEX_SETS_DEBUGGING */
15523 DEBUG_U(dump_regex_sets_structures(pRExC_state,
15524 stack, fence, fence_stack));
15527 top_index = av_tindex_skip_len_mg(stack);
15530 SV** stacked_ptr; /* Ptr to something already on 'stack' */
15531 char stacked_operator; /* The topmost operator on the 'stack'. */
15532 SV* lhs; /* Operand to the left of the operator */
15533 SV* rhs; /* Operand to the right of the operator */
15534 SV* fence_ptr; /* Pointer to top element of the fence
15539 if ( RExC_parse < RExC_end - 1
15540 && (UCHARAT(RExC_parse + 1) == '?'))
15542 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
15543 * This happens when we have some thing like
15545 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
15547 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
15549 * Here we would be handling the interpolated
15550 * '$thai_or_lao'. We handle this by a recursive call to
15551 * ourselves which returns the inversion list the
15552 * interpolated expression evaluates to. We use the flags
15553 * from the interpolated pattern. */
15554 U32 save_flags = RExC_flags;
15555 const char * save_parse;
15557 RExC_parse += 2; /* Skip past the '(?' */
15558 save_parse = RExC_parse;
15560 /* Parse any flags for the '(?' */
15561 parse_lparen_question_flags(pRExC_state);
15563 if (RExC_parse == save_parse /* Makes sure there was at
15564 least one flag (or else
15565 this embedding wasn't
15567 || RExC_parse >= RExC_end - 4
15568 || UCHARAT(RExC_parse) != ':'
15569 || UCHARAT(++RExC_parse) != '('
15570 || UCHARAT(++RExC_parse) != '?'
15571 || UCHARAT(++RExC_parse) != '[')
15574 /* In combination with the above, this moves the
15575 * pointer to the point just after the first erroneous
15576 * character (or if there are no flags, to where they
15577 * should have been) */
15578 if (RExC_parse >= RExC_end - 4) {
15579 RExC_parse = RExC_end;
15581 else if (RExC_parse != save_parse) {
15582 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15584 vFAIL("Expecting '(?flags:(?[...'");
15587 /* Recurse, with the meat of the embedded expression */
15589 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
15590 depth+1, oregcomp_parse);
15592 /* Here, 'current' contains the embedded expression's
15593 * inversion list, and RExC_parse points to the trailing
15594 * ']'; the next character should be the ')' */
15596 if (UCHARAT(RExC_parse) != ')')
15597 vFAIL("Expecting close paren for nested extended charclass");
15599 /* Then the ')' matching the original '(' handled by this
15600 * case: statement */
15602 if (UCHARAT(RExC_parse) != ')')
15603 vFAIL("Expecting close paren for wrapper for nested extended charclass");
15606 RExC_flags = save_flags;
15607 goto handle_operand;
15610 /* A regular '('. Look behind for illegal syntax */
15611 if (top_index - fence >= 0) {
15612 /* If the top entry on the stack is an operator, it had
15613 * better be a '!', otherwise the entry below the top
15614 * operand should be an operator */
15615 if ( ! (top_ptr = av_fetch(stack, top_index, FALSE))
15616 || (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) != '!')
15617 || ( IS_OPERAND(*top_ptr)
15618 && ( top_index - fence < 1
15619 || ! (stacked_ptr = av_fetch(stack,
15622 || ! IS_OPERATOR(*stacked_ptr))))
15625 vFAIL("Unexpected '(' with no preceding operator");
15629 /* Stack the position of this undealt-with left paren */
15630 av_push(fence_stack, newSViv(fence));
15631 fence = top_index + 1;
15635 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
15636 * multi-char folds are allowed. */
15637 if (!regclass(pRExC_state, flagp,depth+1,
15638 TRUE, /* means parse just the next thing */
15639 FALSE, /* don't allow multi-char folds */
15640 FALSE, /* don't silence non-portable warnings. */
15642 FALSE, /* Require return to be an ANYOF */
15646 FAIL2("panic: regclass returned NULL to handle_sets, "
15647 "flags=%#" UVxf, (UV) *flagp);
15650 /* regclass() will return with parsing just the \ sequence,
15651 * leaving the parse pointer at the next thing to parse */
15653 goto handle_operand;
15655 case '[': /* Is a bracketed character class */
15657 /* See if this is a [:posix:] class. */
15658 bool is_posix_class = (OOB_NAMEDCLASS
15659 < handle_possible_posix(pRExC_state,
15663 TRUE /* checking only */));
15664 /* If it is a posix class, leave the parse pointer at the '['
15665 * to fool regclass() into thinking it is part of a
15666 * '[[:posix:]]'. */
15667 if (! is_posix_class) {
15671 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
15672 * multi-char folds are allowed. */
15673 if (!regclass(pRExC_state, flagp,depth+1,
15674 is_posix_class, /* parse the whole char
15675 class only if not a
15677 FALSE, /* don't allow multi-char folds */
15678 TRUE, /* silence non-portable warnings. */
15680 FALSE, /* Require return to be an ANYOF */
15685 FAIL2("panic: regclass returned NULL to handle_sets, "
15686 "flags=%#" UVxf, (UV) *flagp);
15689 /* function call leaves parse pointing to the ']', except if we
15691 if (is_posix_class) {
15695 goto handle_operand;
15699 if (top_index >= 1) {
15700 goto join_operators;
15703 /* Only a single operand on the stack: are done */
15707 if (av_tindex_skip_len_mg(fence_stack) < 0) {
15709 vFAIL("Unexpected ')'");
15712 /* If nothing after the fence, is missing an operand */
15713 if (top_index - fence < 0) {
15717 /* If at least two things on the stack, treat this as an
15719 if (top_index - fence >= 1) {
15720 goto join_operators;
15723 /* Here only a single thing on the fenced stack, and there is a
15724 * fence. Get rid of it */
15725 fence_ptr = av_pop(fence_stack);
15727 fence = SvIV(fence_ptr);
15728 SvREFCNT_dec_NN(fence_ptr);
15735 /* Having gotten rid of the fence, we pop the operand at the
15736 * stack top and process it as a newly encountered operand */
15737 current = av_pop(stack);
15738 if (IS_OPERAND(current)) {
15739 goto handle_operand;
15751 /* These binary operators should have a left operand already
15753 if ( top_index - fence < 0
15754 || top_index - fence == 1
15755 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
15756 || ! IS_OPERAND(*top_ptr))
15758 goto unexpected_binary;
15761 /* If only the one operand is on the part of the stack visible
15762 * to us, we just place this operator in the proper position */
15763 if (top_index - fence < 2) {
15765 /* Place the operator before the operand */
15767 SV* lhs = av_pop(stack);
15768 av_push(stack, newSVuv(curchar));
15769 av_push(stack, lhs);
15773 /* But if there is something else on the stack, we need to
15774 * process it before this new operator if and only if the
15775 * stacked operation has equal or higher precedence than the
15780 /* The operator on the stack is supposed to be below both its
15782 if ( ! (stacked_ptr = av_fetch(stack, top_index - 2, FALSE))
15783 || IS_OPERAND(*stacked_ptr))
15785 /* But if not, it's legal and indicates we are completely
15786 * done if and only if we're currently processing a ']',
15787 * which should be the final thing in the expression */
15788 if (curchar == ']') {
15794 vFAIL2("Unexpected binary operator '%c' with no "
15795 "preceding operand", curchar);
15797 stacked_operator = (char) SvUV(*stacked_ptr);
15799 if (regex_set_precedence(curchar)
15800 > regex_set_precedence(stacked_operator))
15802 /* Here, the new operator has higher precedence than the
15803 * stacked one. This means we need to add the new one to
15804 * the stack to await its rhs operand (and maybe more
15805 * stuff). We put it before the lhs operand, leaving
15806 * untouched the stacked operator and everything below it
15808 lhs = av_pop(stack);
15809 assert(IS_OPERAND(lhs));
15811 av_push(stack, newSVuv(curchar));
15812 av_push(stack, lhs);
15816 /* Here, the new operator has equal or lower precedence than
15817 * what's already there. This means the operation already
15818 * there should be performed now, before the new one. */
15820 rhs = av_pop(stack);
15821 if (! IS_OPERAND(rhs)) {
15823 /* This can happen when a ! is not followed by an operand,
15824 * like in /(?[\t &!])/ */
15828 lhs = av_pop(stack);
15830 if (! IS_OPERAND(lhs)) {
15832 /* This can happen when there is an empty (), like in
15833 * /(?[[0]+()+])/ */
15837 switch (stacked_operator) {
15839 _invlist_intersection(lhs, rhs, &rhs);
15844 _invlist_union(lhs, rhs, &rhs);
15848 _invlist_subtract(lhs, rhs, &rhs);
15851 case '^': /* The union minus the intersection */
15856 _invlist_union(lhs, rhs, &u);
15857 _invlist_intersection(lhs, rhs, &i);
15858 _invlist_subtract(u, i, &rhs);
15859 SvREFCNT_dec_NN(i);
15860 SvREFCNT_dec_NN(u);
15866 /* Here, the higher precedence operation has been done, and the
15867 * result is in 'rhs'. We overwrite the stacked operator with
15868 * the result. Then we redo this code to either push the new
15869 * operator onto the stack or perform any higher precedence
15870 * stacked operation */
15871 only_to_avoid_leaks = av_pop(stack);
15872 SvREFCNT_dec(only_to_avoid_leaks);
15873 av_push(stack, rhs);
15876 case '!': /* Highest priority, right associative */
15878 /* If what's already at the top of the stack is another '!",
15879 * they just cancel each other out */
15880 if ( (top_ptr = av_fetch(stack, top_index, FALSE))
15881 && (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) == '!'))
15883 only_to_avoid_leaks = av_pop(stack);
15884 SvREFCNT_dec(only_to_avoid_leaks);
15886 else { /* Otherwise, since it's right associative, just push
15888 av_push(stack, newSVuv(curchar));
15893 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15894 vFAIL("Unexpected character");
15898 /* Here 'current' is the operand. If something is already on the
15899 * stack, we have to check if it is a !. But first, the code above
15900 * may have altered the stack in the time since we earlier set
15903 top_index = av_tindex_skip_len_mg(stack);
15904 if (top_index - fence >= 0) {
15905 /* If the top entry on the stack is an operator, it had better
15906 * be a '!', otherwise the entry below the top operand should
15907 * be an operator */
15908 top_ptr = av_fetch(stack, top_index, FALSE);
15910 if (IS_OPERATOR(*top_ptr)) {
15912 /* The only permissible operator at the top of the stack is
15913 * '!', which is applied immediately to this operand. */
15914 curchar = (char) SvUV(*top_ptr);
15915 if (curchar != '!') {
15916 SvREFCNT_dec(current);
15917 vFAIL2("Unexpected binary operator '%c' with no "
15918 "preceding operand", curchar);
15921 _invlist_invert(current);
15923 only_to_avoid_leaks = av_pop(stack);
15924 SvREFCNT_dec(only_to_avoid_leaks);
15926 /* And we redo with the inverted operand. This allows
15927 * handling multiple ! in a row */
15928 goto handle_operand;
15930 /* Single operand is ok only for the non-binary ')'
15932 else if ((top_index - fence == 0 && curchar != ')')
15933 || (top_index - fence > 0
15934 && (! (stacked_ptr = av_fetch(stack,
15937 || IS_OPERAND(*stacked_ptr))))
15939 SvREFCNT_dec(current);
15940 vFAIL("Operand with no preceding operator");
15944 /* Here there was nothing on the stack or the top element was
15945 * another operand. Just add this new one */
15946 av_push(stack, current);
15948 } /* End of switch on next parse token */
15950 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15951 } /* End of loop parsing through the construct */
15954 if (av_tindex_skip_len_mg(fence_stack) >= 0) {
15955 vFAIL("Unmatched (");
15958 if (av_tindex_skip_len_mg(stack) < 0 /* Was empty */
15959 || ((final = av_pop(stack)) == NULL)
15960 || ! IS_OPERAND(final)
15961 || ! is_invlist(final)
15962 || av_tindex_skip_len_mg(stack) >= 0) /* More left on stack */
15965 SvREFCNT_dec(final);
15966 vFAIL("Incomplete expression within '(?[ ])'");
15969 /* Here, 'final' is the resultant inversion list from evaluating the
15970 * expression. Return it if so requested */
15971 if (return_invlist) {
15972 *return_invlist = final;
15976 /* Otherwise generate a resultant node, based on 'final'. regclass() is
15977 * expecting a string of ranges and individual code points */
15978 invlist_iterinit(final);
15979 result_string = newSVpvs("");
15980 while (invlist_iternext(final, &start, &end)) {
15981 if (start == end) {
15982 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}", start);
15985 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}-\\x{%" UVXf "}",
15990 /* About to generate an ANYOF (or similar) node from the inversion list we
15991 * have calculated */
15992 save_parse = RExC_parse;
15993 RExC_parse = SvPV(result_string, len);
15994 save_end = RExC_end;
15995 RExC_end = RExC_parse + len;
15997 /* We turn off folding around the call, as the class we have constructed
15998 * already has all folding taken into consideration, and we don't want
15999 * regclass() to add to that */
16000 RExC_flags &= ~RXf_PMf_FOLD;
16001 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if multi-char
16002 * folds are allowed. */
16003 node = regclass(pRExC_state, flagp,depth+1,
16004 FALSE, /* means parse the whole char class */
16005 FALSE, /* don't allow multi-char folds */
16006 TRUE, /* silence non-portable warnings. The above may very
16007 well have generated non-portable code points, but
16008 they're valid on this machine */
16009 FALSE, /* similarly, no need for strict */
16010 FALSE, /* Require return to be an ANYOF */
16015 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#" UVxf,
16018 /* Fix up the node type if we are in locale. (We have pretended we are
16019 * under /u for the purposes of regclass(), as this construct will only
16020 * work under UTF-8 locales. But now we change the opcode to be ANYOFL (so
16021 * as to cause any warnings about bad locales to be output in regexec.c),
16022 * and add the flag that indicates to check if not in a UTF-8 locale. The
16023 * reason we above forbid optimization into something other than an ANYOF
16024 * node is simply to minimize the number of code changes in regexec.c.
16025 * Otherwise we would have to create new EXACTish node types and deal with
16026 * them. This decision could be revisited should this construct become
16029 * (One might think we could look at the resulting ANYOF node and suppress
16030 * the flag if everything is above 255, as those would be UTF-8 only,
16031 * but this isn't true, as the components that led to that result could
16032 * have been locale-affected, and just happen to cancel each other out
16033 * under UTF-8 locales.) */
16035 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
16037 assert(OP(node) == ANYOF);
16041 |= ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
16045 RExC_flags |= RXf_PMf_FOLD;
16048 RExC_parse = save_parse + 1;
16049 RExC_end = save_end;
16050 SvREFCNT_dec_NN(final);
16051 SvREFCNT_dec_NN(result_string);
16053 nextchar(pRExC_state);
16054 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
16058 #ifdef ENABLE_REGEX_SETS_DEBUGGING
16061 S_dump_regex_sets_structures(pTHX_ RExC_state_t *pRExC_state,
16062 AV * stack, const IV fence, AV * fence_stack)
16063 { /* Dumps the stacks in handle_regex_sets() */
16065 const SSize_t stack_top = av_tindex_skip_len_mg(stack);
16066 const SSize_t fence_stack_top = av_tindex_skip_len_mg(fence_stack);
16069 PERL_ARGS_ASSERT_DUMP_REGEX_SETS_STRUCTURES;
16071 PerlIO_printf(Perl_debug_log, "\nParse position is:%s\n", RExC_parse);
16073 if (stack_top < 0) {
16074 PerlIO_printf(Perl_debug_log, "Nothing on stack\n");
16077 PerlIO_printf(Perl_debug_log, "Stack: (fence=%d)\n", (int) fence);
16078 for (i = stack_top; i >= 0; i--) {
16079 SV ** element_ptr = av_fetch(stack, i, FALSE);
16080 if (! element_ptr) {
16083 if (IS_OPERATOR(*element_ptr)) {
16084 PerlIO_printf(Perl_debug_log, "[%d]: %c\n",
16085 (int) i, (int) SvIV(*element_ptr));
16088 PerlIO_printf(Perl_debug_log, "[%d] ", (int) i);
16089 sv_dump(*element_ptr);
16094 if (fence_stack_top < 0) {
16095 PerlIO_printf(Perl_debug_log, "Nothing on fence_stack\n");
16098 PerlIO_printf(Perl_debug_log, "Fence_stack: \n");
16099 for (i = fence_stack_top; i >= 0; i--) {
16100 SV ** element_ptr = av_fetch(fence_stack, i, FALSE);
16101 if (! element_ptr) {
16104 PerlIO_printf(Perl_debug_log, "[%d]: %d\n",
16105 (int) i, (int) SvIV(*element_ptr));
16116 S_add_above_Latin1_folds(pTHX_ RExC_state_t *pRExC_state, const U8 cp, SV** invlist)
16118 /* This adds the Latin1/above-Latin1 folding rules.
16120 * This should be called only for a Latin1-range code points, cp, which is
16121 * known to be involved in a simple fold with other code points above
16122 * Latin1. It would give false results if /aa has been specified.
16123 * Multi-char folds are outside the scope of this, and must be handled
16126 PERL_ARGS_ASSERT_ADD_ABOVE_LATIN1_FOLDS;
16128 assert(HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(cp));
16130 /* The rules that are valid for all Unicode versions are hard-coded in */
16135 add_cp_to_invlist(*invlist, KELVIN_SIGN);
16139 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_LONG_S);
16142 *invlist = add_cp_to_invlist(*invlist, GREEK_CAPITAL_LETTER_MU);
16143 *invlist = add_cp_to_invlist(*invlist, GREEK_SMALL_LETTER_MU);
16145 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
16146 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
16147 *invlist = add_cp_to_invlist(*invlist, ANGSTROM_SIGN);
16149 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
16150 *invlist = add_cp_to_invlist(*invlist,
16151 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
16154 default: /* Other code points are checked against the data for the
16155 current Unicode version */
16157 Size_t folds_to_count;
16158 unsigned int first_folds_to;
16159 const unsigned int * remaining_folds_to_list;
16163 folded_cp = toFOLD(cp);
16166 U8 dummy_fold[UTF8_MAXBYTES_CASE+1];
16168 folded_cp = _to_fold_latin1(cp, dummy_fold, &dummy_len, 0);
16171 if (folded_cp > 255) {
16172 *invlist = add_cp_to_invlist(*invlist, folded_cp);
16175 folds_to_count = _inverse_folds(folded_cp, &first_folds_to,
16176 &remaining_folds_to_list);
16177 if (folds_to_count == 0) {
16179 /* Use deprecated warning to increase the chances of this being
16182 ckWARN2reg_d(RExC_parse,
16183 "Perl folding rules are not up-to-date for 0x%02X;"
16184 " please use the perlbug utility to report;", cp);
16190 if (first_folds_to > 255) {
16191 *invlist = add_cp_to_invlist(*invlist, first_folds_to);
16193 for (i = 0; i < folds_to_count - 1; i++) {
16194 if (remaining_folds_to_list[i] > 255) {
16195 *invlist = add_cp_to_invlist(*invlist,
16196 remaining_folds_to_list[i]);
16206 S_output_or_return_posix_warnings(pTHX_ RExC_state_t *pRExC_state, AV* posix_warnings, AV** return_posix_warnings)
16208 /* If the final parameter is NULL, output the elements of the array given
16209 * by '*posix_warnings' as REGEXP warnings. Otherwise, the elements are
16210 * pushed onto it, (creating if necessary) */
16213 const bool first_is_fatal = ! return_posix_warnings
16214 && ckDEAD(packWARN(WARN_REGEXP));
16216 PERL_ARGS_ASSERT_OUTPUT_OR_RETURN_POSIX_WARNINGS;
16218 while ((msg = av_shift(posix_warnings)) != &PL_sv_undef) {
16219 if (return_posix_warnings) {
16220 if (! *return_posix_warnings) { /* mortalize to not leak if
16221 warnings are fatal */
16222 *return_posix_warnings = (AV *) sv_2mortal((SV *) newAV());
16224 av_push(*return_posix_warnings, msg);
16227 if (first_is_fatal) { /* Avoid leaking this */
16228 av_undef(posix_warnings); /* This isn't necessary if the
16229 array is mortal, but is a
16231 (void) sv_2mortal(msg);
16233 SAVEFREESV(RExC_rx_sv);
16236 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s", SvPVX(msg));
16237 SvREFCNT_dec_NN(msg);
16243 S_add_multi_match(pTHX_ AV* multi_char_matches, SV* multi_string, const STRLEN cp_count)
16245 /* This adds the string scalar <multi_string> to the array
16246 * <multi_char_matches>. <multi_string> is known to have exactly
16247 * <cp_count> code points in it. This is used when constructing a
16248 * bracketed character class and we find something that needs to match more
16249 * than a single character.
16251 * <multi_char_matches> is actually an array of arrays. Each top-level
16252 * element is an array that contains all the strings known so far that are
16253 * the same length. And that length (in number of code points) is the same
16254 * as the index of the top-level array. Hence, the [2] element is an
16255 * array, each element thereof is a string containing TWO code points;
16256 * while element [3] is for strings of THREE characters, and so on. Since
16257 * this is for multi-char strings there can never be a [0] nor [1] element.
16259 * When we rewrite the character class below, we will do so such that the
16260 * longest strings are written first, so that it prefers the longest
16261 * matching strings first. This is done even if it turns out that any
16262 * quantifier is non-greedy, out of this programmer's (khw) laziness. Tom
16263 * Christiansen has agreed that this is ok. This makes the test for the
16264 * ligature 'ffi' come before the test for 'ff', for example */
16267 AV** this_array_ptr;
16269 PERL_ARGS_ASSERT_ADD_MULTI_MATCH;
16271 if (! multi_char_matches) {
16272 multi_char_matches = newAV();
16275 if (av_exists(multi_char_matches, cp_count)) {
16276 this_array_ptr = (AV**) av_fetch(multi_char_matches, cp_count, FALSE);
16277 this_array = *this_array_ptr;
16280 this_array = newAV();
16281 av_store(multi_char_matches, cp_count,
16284 av_push(this_array, multi_string);
16286 return multi_char_matches;
16289 /* The names of properties whose definitions are not known at compile time are
16290 * stored in this SV, after a constant heading. So if the length has been
16291 * changed since initialization, then there is a run-time definition. */
16292 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
16293 (SvCUR(listsv) != initial_listsv_len)
16295 /* There is a restricted set of white space characters that are legal when
16296 * ignoring white space in a bracketed character class. This generates the
16297 * code to skip them.
16299 * There is a line below that uses the same white space criteria but is outside
16300 * this macro. Both here and there must use the same definition */
16301 #define SKIP_BRACKETED_WHITE_SPACE(do_skip, p) \
16304 while (isBLANK_A(UCHARAT(p))) \
16312 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
16313 const bool stop_at_1, /* Just parse the next thing, don't
16314 look for a full character class */
16315 bool allow_multi_folds,
16316 const bool silence_non_portable, /* Don't output warnings
16320 bool optimizable, /* ? Allow a non-ANYOF return
16322 SV** ret_invlist, /* Return an inversion list, not a node */
16323 AV** return_posix_warnings
16326 /* parse a bracketed class specification. Most of these will produce an
16327 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
16328 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
16329 * under /i with multi-character folds: it will be rewritten following the
16330 * paradigm of this example, where the <multi-fold>s are characters which
16331 * fold to multiple character sequences:
16332 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
16333 * gets effectively rewritten as:
16334 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
16335 * reg() gets called (recursively) on the rewritten version, and this
16336 * function will return what it constructs. (Actually the <multi-fold>s
16337 * aren't physically removed from the [abcdefghi], it's just that they are
16338 * ignored in the recursion by means of a flag:
16339 * <RExC_in_multi_char_class>.)
16341 * ANYOF nodes contain a bit map for the first NUM_ANYOF_CODE_POINTS
16342 * characters, with the corresponding bit set if that character is in the
16343 * list. For characters above this, a range list or swash is used. There
16344 * are extra bits for \w, etc. in locale ANYOFs, as what these match is not
16345 * determinable at compile time
16347 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs
16348 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded
16349 * to UTF-8. This can only happen if ret_invlist is non-NULL.
16352 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
16354 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
16357 int namedclass = OOB_NAMEDCLASS;
16358 char *rangebegin = NULL;
16359 bool need_class = 0;
16361 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
16362 than just initialized. */
16363 SV* properties = NULL; /* Code points that match \p{} \P{} */
16364 SV* posixes = NULL; /* Code points that match classes like [:word:],
16365 extended beyond the Latin1 range. These have to
16366 be kept separate from other code points for much
16367 of this function because their handling is
16368 different under /i, and for most classes under
16370 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
16371 separate for a while from the non-complemented
16372 versions because of complications with /d
16374 SV* simple_posixes = NULL; /* But under some conditions, the classes can be
16375 treated more simply than the general case,
16376 leading to less compilation and execution
16378 UV element_count = 0; /* Number of distinct elements in the class.
16379 Optimizations may be possible if this is tiny */
16380 AV * multi_char_matches = NULL; /* Code points that fold to more than one
16381 character; used under /i */
16383 char * stop_ptr = RExC_end; /* where to stop parsing */
16385 /* ignore unescaped whitespace? */
16386 const bool skip_white = cBOOL( ret_invlist
16387 || (RExC_flags & RXf_PMf_EXTENDED_MORE));
16389 /* Unicode properties are stored in a swash; this holds the current one
16390 * being parsed. If this swash is the only above-latin1 component of the
16391 * character class, an optimization is to pass it directly on to the
16392 * execution engine. Otherwise, it is set to NULL to indicate that there
16393 * are other things in the class that have to be dealt with at execution
16395 SV* swash = NULL; /* Code points that match \p{} \P{} */
16397 /* Set if a component of this character class is user-defined; just passed
16398 * on to the engine */
16399 bool has_user_defined_property = FALSE;
16401 /* inversion list of code points this node matches only when the target
16402 * string is in UTF-8. These are all non-ASCII, < 256. (Because is under
16404 SV* has_upper_latin1_only_utf8_matches = NULL;
16406 /* Inversion list of code points this node matches regardless of things
16407 * like locale, folding, utf8ness of the target string */
16408 SV* cp_list = NULL;
16410 /* Like cp_list, but code points on this list need to be checked for things
16411 * that fold to/from them under /i */
16412 SV* cp_foldable_list = NULL;
16414 /* Like cp_list, but code points on this list are valid only when the
16415 * runtime locale is UTF-8 */
16416 SV* only_utf8_locale_list = NULL;
16418 /* In a range, if one of the endpoints is non-character-set portable,
16419 * meaning that it hard-codes a code point that may mean a different
16420 * charactger in ASCII vs. EBCDIC, as opposed to, say, a literal 'A' or a
16421 * mnemonic '\t' which each mean the same character no matter which
16422 * character set the platform is on. */
16423 unsigned int non_portable_endpoint = 0;
16425 /* Is the range unicode? which means on a platform that isn't 1-1 native
16426 * to Unicode (i.e. non-ASCII), each code point in it should be considered
16427 * to be a Unicode value. */
16428 bool unicode_range = FALSE;
16429 bool invert = FALSE; /* Is this class to be complemented */
16431 bool warn_super = ALWAYS_WARN_SUPER;
16433 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
16434 case we need to change the emitted regop to an EXACT. */
16435 const char * orig_parse = RExC_parse;
16436 const SSize_t orig_size = RExC_size;
16437 bool posixl_matches_all = FALSE; /* Does /l class have both e.g. \W,\w ? */
16439 /* This variable is used to mark where the end in the input is of something
16440 * that looks like a POSIX construct but isn't. During the parse, when
16441 * something looks like it could be such a construct is encountered, it is
16442 * checked for being one, but not if we've already checked this area of the
16443 * input. Only after this position is reached do we check again */
16444 char *not_posix_region_end = RExC_parse - 1;
16446 AV* posix_warnings = NULL;
16447 const bool do_posix_warnings = return_posix_warnings
16448 || (PASS2 && ckWARN(WARN_REGEXP));
16450 GET_RE_DEBUG_FLAGS_DECL;
16452 PERL_ARGS_ASSERT_REGCLASS;
16454 PERL_UNUSED_ARG(depth);
16457 DEBUG_PARSE("clas");
16459 #if UNICODE_MAJOR_VERSION < 3 /* no multifolds in early Unicode */ \
16460 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0 \
16461 && UNICODE_DOT_DOT_VERSION == 0)
16462 allow_multi_folds = FALSE;
16465 /* Assume we are going to generate an ANYOF node. */
16466 ret = reganode(pRExC_state,
16473 RExC_size += ANYOF_SKIP;
16474 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
16477 ANYOF_FLAGS(ret) = 0;
16479 RExC_emit += ANYOF_SKIP;
16480 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
16481 initial_listsv_len = SvCUR(listsv);
16482 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
16485 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16487 assert(RExC_parse <= RExC_end);
16489 if (UCHARAT(RExC_parse) == '^') { /* Complement the class */
16492 allow_multi_folds = FALSE;
16494 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16497 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
16498 if (! ret_invlist && MAYBE_POSIXCC(UCHARAT(RExC_parse))) {
16499 int maybe_class = handle_possible_posix(pRExC_state,
16501 ¬_posix_region_end,
16503 TRUE /* checking only */);
16504 if (PASS2 && maybe_class >= OOB_NAMEDCLASS && do_posix_warnings) {
16505 SAVEFREESV(RExC_rx_sv);
16506 ckWARN4reg(not_posix_region_end,
16507 "POSIX syntax [%c %c] belongs inside character classes%s",
16508 *RExC_parse, *RExC_parse,
16509 (maybe_class == OOB_NAMEDCLASS)
16510 ? ((POSIXCC_NOTYET(*RExC_parse))
16511 ? " (but this one isn't implemented)"
16512 : " (but this one isn't fully valid)")
16515 (void)ReREFCNT_inc(RExC_rx_sv);
16519 /* If the caller wants us to just parse a single element, accomplish this
16520 * by faking the loop ending condition */
16521 if (stop_at_1 && RExC_end > RExC_parse) {
16522 stop_ptr = RExC_parse + 1;
16525 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
16526 if (UCHARAT(RExC_parse) == ']')
16527 goto charclassloop;
16531 if ( posix_warnings
16532 && av_tindex_skip_len_mg(posix_warnings) >= 0
16533 && RExC_parse > not_posix_region_end)
16535 /* Warnings about posix class issues are considered tentative until
16536 * we are far enough along in the parse that we can no longer
16537 * change our mind, at which point we either output them or add
16538 * them, if it has so specified, to what gets returned to the
16539 * caller. This is done each time through the loop so that a later
16540 * class won't zap them before they have been dealt with. */
16541 output_or_return_posix_warnings(pRExC_state, posix_warnings,
16542 return_posix_warnings);
16545 if (RExC_parse >= stop_ptr) {
16549 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16551 if (UCHARAT(RExC_parse) == ']') {
16557 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
16558 save_value = value;
16559 save_prevvalue = prevvalue;
16562 rangebegin = RExC_parse;
16564 non_portable_endpoint = 0;
16566 if (UTF && ! UTF8_IS_INVARIANT(* RExC_parse)) {
16567 value = utf8n_to_uvchr((U8*)RExC_parse,
16568 RExC_end - RExC_parse,
16569 &numlen, UTF8_ALLOW_DEFAULT);
16570 RExC_parse += numlen;
16573 value = UCHARAT(RExC_parse++);
16575 if (value == '[') {
16576 char * posix_class_end;
16577 namedclass = handle_possible_posix(pRExC_state,
16580 do_posix_warnings ? &posix_warnings : NULL,
16581 FALSE /* die if error */);
16582 if (namedclass > OOB_NAMEDCLASS) {
16584 /* If there was an earlier attempt to parse this particular
16585 * posix class, and it failed, it was a false alarm, as this
16586 * successful one proves */
16587 if ( posix_warnings
16588 && av_tindex_skip_len_mg(posix_warnings) >= 0
16589 && not_posix_region_end >= RExC_parse
16590 && not_posix_region_end <= posix_class_end)
16592 av_undef(posix_warnings);
16595 RExC_parse = posix_class_end;
16597 else if (namedclass == OOB_NAMEDCLASS) {
16598 not_posix_region_end = posix_class_end;
16601 namedclass = OOB_NAMEDCLASS;
16604 else if ( RExC_parse - 1 > not_posix_region_end
16605 && MAYBE_POSIXCC(value))
16607 (void) handle_possible_posix(
16609 RExC_parse - 1, /* -1 because parse has already been
16611 ¬_posix_region_end,
16612 do_posix_warnings ? &posix_warnings : NULL,
16613 TRUE /* checking only */);
16615 else if ( strict && ! skip_white
16616 && ( _generic_isCC(value, _CC_VERTSPACE)
16617 || is_VERTWS_cp_high(value)))
16619 vFAIL("Literal vertical space in [] is illegal except under /x");
16621 else if (value == '\\') {
16622 /* Is a backslash; get the code point of the char after it */
16624 if (RExC_parse >= RExC_end) {
16625 vFAIL("Unmatched [");
16628 if (UTF && ! UTF8_IS_INVARIANT(UCHARAT(RExC_parse))) {
16629 value = utf8n_to_uvchr((U8*)RExC_parse,
16630 RExC_end - RExC_parse,
16631 &numlen, UTF8_ALLOW_DEFAULT);
16632 RExC_parse += numlen;
16635 value = UCHARAT(RExC_parse++);
16637 /* Some compilers cannot handle switching on 64-bit integer
16638 * values, therefore value cannot be an UV. Yes, this will
16639 * be a problem later if we want switch on Unicode.
16640 * A similar issue a little bit later when switching on
16641 * namedclass. --jhi */
16643 /* If the \ is escaping white space when white space is being
16644 * skipped, it means that that white space is wanted literally, and
16645 * is already in 'value'. Otherwise, need to translate the escape
16646 * into what it signifies. */
16647 if (! skip_white || ! isBLANK_A(value)) switch ((I32)value) {
16649 case 'w': namedclass = ANYOF_WORDCHAR; break;
16650 case 'W': namedclass = ANYOF_NWORDCHAR; break;
16651 case 's': namedclass = ANYOF_SPACE; break;
16652 case 'S': namedclass = ANYOF_NSPACE; break;
16653 case 'd': namedclass = ANYOF_DIGIT; break;
16654 case 'D': namedclass = ANYOF_NDIGIT; break;
16655 case 'v': namedclass = ANYOF_VERTWS; break;
16656 case 'V': namedclass = ANYOF_NVERTWS; break;
16657 case 'h': namedclass = ANYOF_HORIZWS; break;
16658 case 'H': namedclass = ANYOF_NHORIZWS; break;
16659 case 'N': /* Handle \N{NAME} in class */
16661 const char * const backslash_N_beg = RExC_parse - 2;
16664 if (! grok_bslash_N(pRExC_state,
16665 NULL, /* No regnode */
16666 &value, /* Yes single value */
16667 &cp_count, /* Multiple code pt count */
16673 if (*flagp & NEED_UTF8)
16674 FAIL("panic: grok_bslash_N set NEED_UTF8");
16676 RETURN_NULL_ON_RESTART_FLAGP(flagp);
16678 if (cp_count < 0) {
16679 vFAIL("\\N in a character class must be a named character: \\N{...}");
16681 else if (cp_count == 0) {
16683 ckWARNreg(RExC_parse,
16684 "Ignoring zero length \\N{} in character class");
16687 else { /* cp_count > 1 */
16688 if (! RExC_in_multi_char_class) {
16689 if (invert || range || *RExC_parse == '-') {
16692 vFAIL("\\N{} in inverted character class or as a range end-point is restricted to one character");
16695 ckWARNreg(RExC_parse, "Using just the first character returned by \\N{} in character class");
16697 break; /* <value> contains the first code
16698 point. Drop out of the switch to
16702 SV * multi_char_N = newSVpvn(backslash_N_beg,
16703 RExC_parse - backslash_N_beg);
16705 = add_multi_match(multi_char_matches,
16710 } /* End of cp_count != 1 */
16712 /* This element should not be processed further in this
16715 value = save_value;
16716 prevvalue = save_prevvalue;
16717 continue; /* Back to top of loop to get next char */
16720 /* Here, is a single code point, and <value> contains it */
16721 unicode_range = TRUE; /* \N{} are Unicode */
16731 /* We will handle any undefined properties ourselves */
16732 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF
16733 /* And we actually would prefer to get
16734 * the straight inversion list of the
16735 * swash, since we will be accessing it
16736 * anyway, to save a little time */
16737 |_CORE_SWASH_INIT_ACCEPT_INVLIST;
16739 SvREFCNT_dec(swash); /* Free any left-overs */
16740 if (RExC_parse >= RExC_end)
16741 vFAIL2("Empty \\%c", (U8)value);
16742 if (*RExC_parse == '{') {
16743 const U8 c = (U8)value;
16744 e = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
16747 vFAIL2("Missing right brace on \\%c{}", c);
16752 /* White space is allowed adjacent to the braces and after
16753 * any '^', even when not under /x */
16754 while (isSPACE(*RExC_parse)) {
16758 if (UCHARAT(RExC_parse) == '^') {
16760 /* toggle. (The rhs xor gets the single bit that
16761 * differs between P and p; the other xor inverts just
16763 value ^= 'P' ^ 'p';
16766 while (isSPACE(*RExC_parse)) {
16771 if (e == RExC_parse)
16772 vFAIL2("Empty \\%c{}", c);
16774 n = e - RExC_parse;
16775 while (isSPACE(*(RExC_parse + n - 1)))
16778 } /* The \p isn't immediately followed by a '{' */
16779 else if (! isALPHA(*RExC_parse)) {
16780 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16781 vFAIL2("Character following \\%c must be '{' or a "
16782 "single-character Unicode property name",
16790 char* name = RExC_parse;
16791 char* base_name; /* name after any packages are stripped */
16792 char* lookup_name = NULL;
16793 const char * const colon_colon = "::";
16798 /* Temporary workaround for [perl #133136]. For this
16799 * precise input that is in the .t that is failing, load
16800 * utf8.pm, which is what the test wants, so that that
16802 if ( memEQs(RExC_start, e + 1 - RExC_start,
16804 && ! hv_common(GvHVn(PL_incgv),
16806 "utf8.pm", sizeof("utf8.pm") - 1,
16807 0, HV_FETCH_ISEXISTS, NULL, 0))
16809 require_pv("utf8.pm");
16811 invlist = parse_uniprop_string(name, n, FOLD, &invert);
16814 value ^= 'P' ^ 'p';
16819 /* Try to get the definition of the property into
16820 * <invlist>. If /i is in effect, the effective property
16821 * will have its name be <__NAME_i>. The design is
16822 * discussed in commit
16823 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
16824 name = savepv(Perl_form(aTHX_ "%.*s", (int)n, RExC_parse));
16827 for (i = RExC_parse; i < RExC_parse + n; i++) {
16828 if (isCNTRL(*i) && *i != '\t') {
16829 RExC_parse = e + 1;
16830 vFAIL2("Can't find Unicode property definition \"%s\"", name);
16835 lookup_name = savepv(Perl_form(aTHX_ "__%s_i", name));
16837 /* The function call just below that uses this can fail
16838 * to return, leaking memory if we don't do this */
16839 SAVEFREEPV(lookup_name);
16842 /* Look up the property name, and get its swash and
16843 * inversion list, if the property is found */
16844 swash = _core_swash_init("utf8",
16851 NULL, /* No inversion list */
16854 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
16855 HV* curpkg = (IN_PERL_COMPILETIME)
16857 : CopSTASH(PL_curcop);
16861 if (swash) { /* Got a swash but no inversion list.
16862 Something is likely wrong that will
16863 be sorted-out later */
16864 SvREFCNT_dec_NN(swash);
16868 /* Here didn't find it. It could be a an error (like a
16869 * typo) in specifying a Unicode property, or it could
16870 * be a user-defined property that will be available at
16871 * run-time. The names of these must begin with 'In'
16872 * or 'Is' (after any packages are stripped off). So
16873 * if not one of those, or if we accept only
16874 * compile-time properties, is an error; otherwise add
16875 * it to the list for run-time look up. */
16876 if ((base_name = rninstr(name, name + n,
16877 colon_colon, colon_colon + 2)))
16878 { /* Has ::. We know this must be a user-defined
16881 final_n -= base_name - name;
16890 || base_name[0] != 'I'
16891 || (base_name[1] != 's' && base_name[1] != 'n')
16894 const char * const msg
16896 ? "Illegal user-defined property name"
16897 : "Can't find Unicode property definition";
16898 RExC_parse = e + 1;
16900 /* diag_listed_as: Can't find Unicode property definition "%s" */
16901 vFAIL3utf8f("%s \"%" UTF8f "\"",
16902 msg, UTF8fARG(UTF, n, name));
16905 /* If the property name doesn't already have a package
16906 * name, add the current one to it so that it can be
16907 * referred to outside it. [perl #121777] */
16908 if (! has_pkg && curpkg) {
16909 char* pkgname = HvNAME(curpkg);
16910 if (memNEs(pkgname, HvNAMELEN(curpkg), "main")) {
16911 char* full_name = Perl_form(aTHX_
16915 n = strlen(full_name);
16916 name = savepvn(full_name, n);
16920 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%s%" UTF8f "%s\n",
16921 (value == 'p' ? '+' : '!'),
16922 (FOLD) ? "__" : "",
16923 UTF8fARG(UTF, n, name),
16924 (FOLD) ? "_i" : "");
16925 has_user_defined_property = TRUE;
16926 optimizable = FALSE; /* Will have to leave this an
16929 /* We don't know yet what this matches, so have to flag
16931 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
16935 /* Here, did get the swash and its inversion list. If
16936 * the swash is from a user-defined property, then this
16937 * whole character class should be regarded as such */
16938 if (swash_init_flags
16939 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY)
16941 has_user_defined_property = TRUE;
16946 if (! has_user_defined_property &&
16947 /* We warn on matching an above-Unicode code point
16948 * if the match would return true, except don't
16949 * warn for \p{All}, which has exactly one element
16951 (_invlist_contains_cp(invlist, 0x110000)
16952 && (! (_invlist_len(invlist) == 1
16953 && *invlist_array(invlist) == 0))))
16958 /* Invert if asking for the complement */
16959 if (value == 'P') {
16960 _invlist_union_complement_2nd(properties,
16964 /* The swash can't be used as-is, because we've
16965 * inverted things; delay removing it to here after
16966 * have copied its invlist above */
16968 SvREFCNT_dec_NN(invlist);
16970 SvREFCNT_dec(swash);
16974 _invlist_union(properties, invlist, &properties);
16976 SvREFCNT_dec_NN(invlist);
16980 } /* End of actually getting the values in pass 2 */
16982 RExC_parse = e + 1;
16983 namedclass = ANYOF_UNIPROP; /* no official name, but it's
16986 /* \p means they want Unicode semantics */
16987 REQUIRE_UNI_RULES(flagp, NULL);
16990 case 'n': value = '\n'; break;
16991 case 'r': value = '\r'; break;
16992 case 't': value = '\t'; break;
16993 case 'f': value = '\f'; break;
16994 case 'b': value = '\b'; break;
16995 case 'e': value = ESC_NATIVE; break;
16996 case 'a': value = '\a'; break;
16998 RExC_parse--; /* function expects to be pointed at the 'o' */
17000 const char* error_msg;
17001 bool valid = grok_bslash_o(&RExC_parse,
17005 PASS2, /* warnings only in
17008 silence_non_portable,
17014 non_portable_endpoint++;
17017 RExC_parse--; /* function expects to be pointed at the 'x' */
17019 const char* error_msg;
17020 bool valid = grok_bslash_x(&RExC_parse,
17024 PASS2, /* Output warnings */
17026 silence_non_portable,
17032 non_portable_endpoint++;
17035 value = grok_bslash_c(*RExC_parse++, PASS2);
17036 non_portable_endpoint++;
17038 case '0': case '1': case '2': case '3': case '4':
17039 case '5': case '6': case '7':
17041 /* Take 1-3 octal digits */
17042 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
17043 numlen = (strict) ? 4 : 3;
17044 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
17045 RExC_parse += numlen;
17048 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
17049 vFAIL("Need exactly 3 octal digits");
17051 else if (! SIZE_ONLY /* like \08, \178 */
17053 && RExC_parse < RExC_end
17054 && isDIGIT(*RExC_parse)
17055 && ckWARN(WARN_REGEXP))
17057 SAVEFREESV(RExC_rx_sv);
17058 reg_warn_non_literal_string(
17060 form_short_octal_warning(RExC_parse, numlen));
17061 (void)ReREFCNT_inc(RExC_rx_sv);
17064 non_portable_endpoint++;
17068 /* Allow \_ to not give an error */
17069 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
17071 vFAIL2("Unrecognized escape \\%c in character class",
17075 SAVEFREESV(RExC_rx_sv);
17076 ckWARN2reg(RExC_parse,
17077 "Unrecognized escape \\%c in character class passed through",
17079 (void)ReREFCNT_inc(RExC_rx_sv);
17083 } /* End of switch on char following backslash */
17084 } /* end of handling backslash escape sequences */
17086 /* Here, we have the current token in 'value' */
17088 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
17091 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
17092 * literal, as is the character that began the false range, i.e.
17093 * the 'a' in the examples */
17096 const int w = (RExC_parse >= rangebegin)
17097 ? RExC_parse - rangebegin
17101 "False [] range \"%" UTF8f "\"",
17102 UTF8fARG(UTF, w, rangebegin));
17105 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
17106 ckWARN2reg(RExC_parse,
17107 "False [] range \"%" UTF8f "\"",
17108 UTF8fARG(UTF, w, rangebegin));
17109 (void)ReREFCNT_inc(RExC_rx_sv);
17110 cp_list = add_cp_to_invlist(cp_list, '-');
17111 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
17116 range = 0; /* this was not a true range */
17117 element_count += 2; /* So counts for three values */
17120 classnum = namedclass_to_classnum(namedclass);
17122 if (LOC && namedclass < ANYOF_POSIXL_MAX
17123 #ifndef HAS_ISASCII
17124 && classnum != _CC_ASCII
17127 /* What the Posix classes (like \w, [:space:]) match in locale
17128 * isn't knowable under locale until actual match time. Room
17129 * must be reserved (one time per outer bracketed class) to
17130 * store such classes. The space will contain a bit for each
17131 * named class that is to be matched against. This isn't
17132 * needed for \p{} and pseudo-classes, as they are not affected
17133 * by locale, and hence are dealt with separately */
17134 if (! need_class) {
17137 RExC_size += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
17140 RExC_emit += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
17142 ANYOF_FLAGS(ret) |= ANYOF_MATCHES_POSIXL;
17143 ANYOF_POSIXL_ZERO(ret);
17145 /* We can't change this into some other type of node
17146 * (unless this is the only element, in which case there
17147 * are nodes that mean exactly this) as has runtime
17149 optimizable = FALSE;
17152 /* Coverity thinks it is possible for this to be negative; both
17153 * jhi and khw think it's not, but be safer */
17154 assert(! (ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
17155 || (namedclass + ((namedclass % 2) ? -1 : 1)) >= 0);
17157 /* See if it already matches the complement of this POSIX
17159 if ((ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
17160 && ANYOF_POSIXL_TEST(ret, namedclass + ((namedclass % 2)
17164 posixl_matches_all = TRUE;
17165 break; /* No need to continue. Since it matches both
17166 e.g., \w and \W, it matches everything, and the
17167 bracketed class can be optimized into qr/./s */
17170 /* Add this class to those that should be checked at runtime */
17171 ANYOF_POSIXL_SET(ret, namedclass);
17173 /* The above-Latin1 characters are not subject to locale rules.
17174 * Just add them, in the second pass, to the
17175 * unconditionally-matched list */
17177 SV* scratch_list = NULL;
17179 /* Get the list of the above-Latin1 code points this
17181 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
17182 PL_XPosix_ptrs[classnum],
17184 /* Odd numbers are complements, like
17185 * NDIGIT, NASCII, ... */
17186 namedclass % 2 != 0,
17188 /* Checking if 'cp_list' is NULL first saves an extra
17189 * clone. Its reference count will be decremented at the
17190 * next union, etc, or if this is the only instance, at the
17191 * end of the routine */
17193 cp_list = scratch_list;
17196 _invlist_union(cp_list, scratch_list, &cp_list);
17197 SvREFCNT_dec_NN(scratch_list);
17199 continue; /* Go get next character */
17202 else if (! SIZE_ONLY) {
17204 /* Here, not in pass1 (in that pass we skip calculating the
17205 * contents of this class), and is not /l, or is a POSIX class
17206 * for which /l doesn't matter (or is a Unicode property, which
17207 * is skipped here). */
17208 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
17209 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
17211 /* Here, should be \h, \H, \v, or \V. None of /d, /i
17212 * nor /l make a difference in what these match,
17213 * therefore we just add what they match to cp_list. */
17214 if (classnum != _CC_VERTSPACE) {
17215 assert( namedclass == ANYOF_HORIZWS
17216 || namedclass == ANYOF_NHORIZWS);
17218 /* It turns out that \h is just a synonym for
17220 classnum = _CC_BLANK;
17223 _invlist_union_maybe_complement_2nd(
17225 PL_XPosix_ptrs[classnum],
17226 namedclass % 2 != 0, /* Complement if odd
17227 (NHORIZWS, NVERTWS)
17232 else if ( UNI_SEMANTICS
17233 || AT_LEAST_ASCII_RESTRICTED
17234 || classnum == _CC_ASCII
17235 || (DEPENDS_SEMANTICS && ( classnum == _CC_DIGIT
17236 || classnum == _CC_XDIGIT)))
17238 /* We usually have to worry about /d a affecting what POSIX
17239 * classes match, with special code needed because we won't
17240 * know until runtime what all matches. But there is no
17241 * extra work needed under /u and /a; and [:ascii:] is
17242 * unaffected by /d; and :digit: and :xdigit: don't have
17243 * runtime differences under /d. So we can special case
17244 * these, and avoid some extra work below, and at runtime.
17246 _invlist_union_maybe_complement_2nd(
17248 ((AT_LEAST_ASCII_RESTRICTED)
17249 ? PL_Posix_ptrs[classnum]
17250 : PL_XPosix_ptrs[classnum]),
17251 namedclass % 2 != 0,
17254 else { /* Garden variety class. If is NUPPER, NALPHA, ...
17255 complement and use nposixes */
17256 SV** posixes_ptr = namedclass % 2 == 0
17259 _invlist_union_maybe_complement_2nd(
17261 PL_XPosix_ptrs[classnum],
17262 namedclass % 2 != 0,
17266 } /* end of namedclass \blah */
17268 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
17270 /* If 'range' is set, 'value' is the ending of a range--check its
17271 * validity. (If value isn't a single code point in the case of a
17272 * range, we should have figured that out above in the code that
17273 * catches false ranges). Later, we will handle each individual code
17274 * point in the range. If 'range' isn't set, this could be the
17275 * beginning of a range, so check for that by looking ahead to see if
17276 * the next real character to be processed is the range indicator--the
17281 /* For unicode ranges, we have to test that the Unicode as opposed
17282 * to the native values are not decreasing. (Above 255, there is
17283 * no difference between native and Unicode) */
17284 if (unicode_range && prevvalue < 255 && value < 255) {
17285 if (NATIVE_TO_LATIN1(prevvalue) > NATIVE_TO_LATIN1(value)) {
17286 goto backwards_range;
17291 if (prevvalue > value) /* b-a */ {
17296 w = RExC_parse - rangebegin;
17298 "Invalid [] range \"%" UTF8f "\"",
17299 UTF8fARG(UTF, w, rangebegin));
17300 NOT_REACHED; /* NOTREACHED */
17304 prevvalue = value; /* save the beginning of the potential range */
17305 if (! stop_at_1 /* Can't be a range if parsing just one thing */
17306 && *RExC_parse == '-')
17308 char* next_char_ptr = RExC_parse + 1;
17310 /* Get the next real char after the '-' */
17311 SKIP_BRACKETED_WHITE_SPACE(skip_white, next_char_ptr);
17313 /* If the '-' is at the end of the class (just before the ']',
17314 * it is a literal minus; otherwise it is a range */
17315 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
17316 RExC_parse = next_char_ptr;
17318 /* a bad range like \w-, [:word:]- ? */
17319 if (namedclass > OOB_NAMEDCLASS) {
17320 if (strict || (PASS2 && ckWARN(WARN_REGEXP))) {
17321 const int w = RExC_parse >= rangebegin
17322 ? RExC_parse - rangebegin
17325 vFAIL4("False [] range \"%*.*s\"",
17330 "False [] range \"%*.*s\"",
17335 cp_list = add_cp_to_invlist(cp_list, '-');
17339 range = 1; /* yeah, it's a range! */
17340 continue; /* but do it the next time */
17345 if (namedclass > OOB_NAMEDCLASS) {
17349 /* Here, we have a single value this time through the loop, and
17350 * <prevvalue> is the beginning of the range, if any; or <value> if
17353 /* non-Latin1 code point implies unicode semantics. Must be set in
17354 * pass1 so is there for the whole of pass 2 */
17356 REQUIRE_UNI_RULES(flagp, NULL);
17359 /* Ready to process either the single value, or the completed range.
17360 * For single-valued non-inverted ranges, we consider the possibility
17361 * of multi-char folds. (We made a conscious decision to not do this
17362 * for the other cases because it can often lead to non-intuitive
17363 * results. For example, you have the peculiar case that:
17364 * "s s" =~ /^[^\xDF]+$/i => Y
17365 * "ss" =~ /^[^\xDF]+$/i => N
17367 * See [perl #89750] */
17368 if (FOLD && allow_multi_folds && value == prevvalue) {
17369 if (value == LATIN_SMALL_LETTER_SHARP_S
17370 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
17373 /* Here <value> is indeed a multi-char fold. Get what it is */
17375 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
17378 UV folded = _to_uni_fold_flags(
17382 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
17383 ? FOLD_FLAGS_NOMIX_ASCII
17387 /* Here, <folded> should be the first character of the
17388 * multi-char fold of <value>, with <foldbuf> containing the
17389 * whole thing. But, if this fold is not allowed (because of
17390 * the flags), <fold> will be the same as <value>, and should
17391 * be processed like any other character, so skip the special
17393 if (folded != value) {
17395 /* Skip if we are recursed, currently parsing the class
17396 * again. Otherwise add this character to the list of
17397 * multi-char folds. */
17398 if (! RExC_in_multi_char_class) {
17399 STRLEN cp_count = utf8_length(foldbuf,
17400 foldbuf + foldlen);
17401 SV* multi_fold = sv_2mortal(newSVpvs(""));
17403 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%" UVXf "}", value);
17406 = add_multi_match(multi_char_matches,
17412 /* This element should not be processed further in this
17415 value = save_value;
17416 prevvalue = save_prevvalue;
17422 if (strict && PASS2 && ckWARN(WARN_REGEXP)) {
17425 /* If the range starts above 255, everything is portable and
17426 * likely to be so for any forseeable character set, so don't
17428 if (unicode_range && non_portable_endpoint && prevvalue < 256) {
17429 vWARN(RExC_parse, "Both or neither range ends should be Unicode");
17431 else if (prevvalue != value) {
17433 /* Under strict, ranges that stop and/or end in an ASCII
17434 * printable should have each end point be a portable value
17435 * for it (preferably like 'A', but we don't warn if it is
17436 * a (portable) Unicode name or code point), and the range
17437 * must be be all digits or all letters of the same case.
17438 * Otherwise, the range is non-portable and unclear as to
17439 * what it contains */
17440 if ( (isPRINT_A(prevvalue) || isPRINT_A(value))
17441 && ( non_portable_endpoint
17442 || ! ( (isDIGIT_A(prevvalue) && isDIGIT_A(value))
17443 || (isLOWER_A(prevvalue) && isLOWER_A(value))
17444 || (isUPPER_A(prevvalue) && isUPPER_A(value))
17446 vWARN(RExC_parse, "Ranges of ASCII printables should"
17447 " be some subset of \"0-9\","
17448 " \"A-Z\", or \"a-z\"");
17450 else if (prevvalue >= FIRST_NON_ASCII_DECIMAL_DIGIT) {
17451 SSize_t index_start;
17452 SSize_t index_final;
17454 /* But the nature of Unicode and languages mean we
17455 * can't do the same checks for above-ASCII ranges,
17456 * except in the case of digit ones. These should
17457 * contain only digits from the same group of 10. The
17458 * ASCII case is handled just above. Hence here, the
17459 * range could be a range of digits. First some
17460 * unlikely special cases. Grandfather in that a range
17461 * ending in 19DA (NEW TAI LUE THAM DIGIT ONE) is bad
17462 * if its starting value is one of the 10 digits prior
17463 * to it. This is because it is an alternate way of
17464 * writing 19D1, and some people may expect it to be in
17465 * that group. But it is bad, because it won't give
17466 * the expected results. In Unicode 5.2 it was
17467 * considered to be in that group (of 11, hence), but
17468 * this was fixed in the next version */
17470 if (UNLIKELY(value == 0x19DA && prevvalue >= 0x19D0)) {
17471 goto warn_bad_digit_range;
17473 else if (UNLIKELY( prevvalue >= 0x1D7CE
17474 && value <= 0x1D7FF))
17476 /* This is the only other case currently in Unicode
17477 * where the algorithm below fails. The code
17478 * points just above are the end points of a single
17479 * range containing only decimal digits. It is 5
17480 * different series of 0-9. All other ranges of
17481 * digits currently in Unicode are just a single
17482 * series. (And mktables will notify us if a later
17483 * Unicode version breaks this.)
17485 * If the range being checked is at most 9 long,
17486 * and the digit values represented are in
17487 * numerical order, they are from the same series.
17489 if ( value - prevvalue > 9
17490 || ((( value - 0x1D7CE) % 10)
17491 <= (prevvalue - 0x1D7CE) % 10))
17493 goto warn_bad_digit_range;
17498 /* For all other ranges of digits in Unicode, the
17499 * algorithm is just to check if both end points
17500 * are in the same series, which is the same range.
17502 index_start = _invlist_search(
17503 PL_XPosix_ptrs[_CC_DIGIT],
17506 /* Warn if the range starts and ends with a digit,
17507 * and they are not in the same group of 10. */
17508 if ( index_start >= 0
17509 && ELEMENT_RANGE_MATCHES_INVLIST(index_start)
17511 _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
17512 value)) != index_start
17513 && index_final >= 0
17514 && ELEMENT_RANGE_MATCHES_INVLIST(index_final))
17516 warn_bad_digit_range:
17517 vWARN(RExC_parse, "Ranges of digits should be"
17518 " from the same group of"
17525 if ((! range || prevvalue == value) && non_portable_endpoint) {
17526 if (isPRINT_A(value)) {
17529 if (isBACKSLASHED_PUNCT(value)) {
17530 literal[d++] = '\\';
17532 literal[d++] = (char) value;
17533 literal[d++] = '\0';
17536 "\"%.*s\" is more clearly written simply as \"%s\"",
17537 (int) (RExC_parse - rangebegin),
17542 else if isMNEMONIC_CNTRL(value) {
17544 "\"%.*s\" is more clearly written simply as \"%s\"",
17545 (int) (RExC_parse - rangebegin),
17547 cntrl_to_mnemonic((U8) value)
17553 /* Deal with this element of the class */
17557 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
17560 /* On non-ASCII platforms, for ranges that span all of 0..255, and
17561 * ones that don't require special handling, we can just add the
17562 * range like we do for ASCII platforms */
17563 if ((UNLIKELY(prevvalue == 0) && value >= 255)
17564 || ! (prevvalue < 256
17566 || (! non_portable_endpoint
17567 && ((isLOWER_A(prevvalue) && isLOWER_A(value))
17568 || (isUPPER_A(prevvalue)
17569 && isUPPER_A(value)))))))
17571 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
17575 /* Here, requires special handling. This can be because it is
17576 * a range whose code points are considered to be Unicode, and
17577 * so must be individually translated into native, or because
17578 * its a subrange of 'A-Z' or 'a-z' which each aren't
17579 * contiguous in EBCDIC, but we have defined them to include
17580 * only the "expected" upper or lower case ASCII alphabetics.
17581 * Subranges above 255 are the same in native and Unicode, so
17582 * can be added as a range */
17583 U8 start = NATIVE_TO_LATIN1(prevvalue);
17585 U8 end = (value < 256) ? NATIVE_TO_LATIN1(value) : 255;
17586 for (j = start; j <= end; j++) {
17587 cp_foldable_list = add_cp_to_invlist(cp_foldable_list, LATIN1_TO_NATIVE(j));
17590 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
17597 range = 0; /* this range (if it was one) is done now */
17598 } /* End of loop through all the text within the brackets */
17601 if ( posix_warnings && av_tindex_skip_len_mg(posix_warnings) >= 0) {
17602 output_or_return_posix_warnings(pRExC_state, posix_warnings,
17603 return_posix_warnings);
17606 /* If anything in the class expands to more than one character, we have to
17607 * deal with them by building up a substitute parse string, and recursively
17608 * calling reg() on it, instead of proceeding */
17609 if (multi_char_matches) {
17610 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
17613 char *save_end = RExC_end;
17614 char *save_parse = RExC_parse;
17615 char *save_start = RExC_start;
17616 STRLEN prefix_end = 0; /* We copy the character class after a
17617 prefix supplied here. This is the size
17618 + 1 of that prefix */
17619 bool first_time = TRUE; /* First multi-char occurrence doesn't get
17624 assert(RExC_precomp_adj == 0); /* Only one level of recursion allowed */
17626 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
17627 because too confusing */
17629 sv_catpvs(substitute_parse, "(?:");
17633 /* Look at the longest folds first */
17634 for (cp_count = av_tindex_skip_len_mg(multi_char_matches);
17639 if (av_exists(multi_char_matches, cp_count)) {
17640 AV** this_array_ptr;
17643 this_array_ptr = (AV**) av_fetch(multi_char_matches,
17645 while ((this_sequence = av_pop(*this_array_ptr)) !=
17648 if (! first_time) {
17649 sv_catpvs(substitute_parse, "|");
17651 first_time = FALSE;
17653 sv_catpv(substitute_parse, SvPVX(this_sequence));
17658 /* If the character class contains anything else besides these
17659 * multi-character folds, have to include it in recursive parsing */
17660 if (element_count) {
17661 sv_catpvs(substitute_parse, "|[");
17662 prefix_end = SvCUR(substitute_parse);
17663 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
17665 /* Put in a closing ']' only if not going off the end, as otherwise
17666 * we are adding something that really isn't there */
17667 if (RExC_parse < RExC_end) {
17668 sv_catpvs(substitute_parse, "]");
17672 sv_catpvs(substitute_parse, ")");
17675 /* This is a way to get the parse to skip forward a whole named
17676 * sequence instead of matching the 2nd character when it fails the
17678 sv_catpvs(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
17682 /* Set up the data structure so that any errors will be properly
17683 * reported. See the comments at the definition of
17684 * REPORT_LOCATION_ARGS for details */
17685 RExC_precomp_adj = orig_parse - RExC_precomp;
17686 RExC_start = RExC_parse = SvPV(substitute_parse, len);
17687 RExC_adjusted_start = RExC_start + prefix_end;
17688 RExC_end = RExC_parse + len;
17689 RExC_in_multi_char_class = 1;
17690 RExC_emit = (regnode *)orig_emit;
17692 ret = reg(pRExC_state, 1, ®_flags, depth+1);
17694 *flagp |= reg_flags & (HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_PASS1|NEED_UTF8);
17696 /* And restore so can parse the rest of the pattern */
17697 RExC_parse = save_parse;
17698 RExC_start = RExC_adjusted_start = save_start;
17699 RExC_precomp_adj = 0;
17700 RExC_end = save_end;
17701 RExC_in_multi_char_class = 0;
17702 SvREFCNT_dec_NN(multi_char_matches);
17706 /* Here, we've gone through the entire class and dealt with multi-char
17707 * folds. We are now in a position that we can do some checks to see if we
17708 * can optimize this ANYOF node into a simpler one, even in Pass 1.
17709 * Currently we only do two checks:
17710 * 1) is in the unlikely event that the user has specified both, eg. \w and
17711 * \W under /l, then the class matches everything. (This optimization
17712 * is done only to make the optimizer code run later work.)
17713 * 2) if the character class contains only a single element (including a
17714 * single range), we see if there is an equivalent node for it.
17715 * Other checks are possible */
17717 && ! ret_invlist /* Can't optimize if returning the constructed
17719 && (UNLIKELY(posixl_matches_all) || element_count == 1))
17724 if (UNLIKELY(posixl_matches_all)) {
17727 else if (namedclass > OOB_NAMEDCLASS) { /* this is a single named
17728 class, like \w or [:digit:]
17731 /* All named classes are mapped into POSIXish nodes, with its FLAG
17732 * argument giving which class it is */
17733 switch ((I32)namedclass) {
17734 case ANYOF_UNIPROP:
17737 /* These don't depend on the charset modifiers. They always
17738 * match under /u rules */
17739 case ANYOF_NHORIZWS:
17740 case ANYOF_HORIZWS:
17741 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
17744 case ANYOF_NVERTWS:
17749 /* The actual POSIXish node for all the rest depends on the
17750 * charset modifier. The ones in the first set depend only on
17751 * ASCII or, if available on this platform, also locale */
17762 /* (named_class - ANYOF_ASCII) is 0 or 1. xor'ing with
17763 * invert converts that to 1 or 0 */
17764 op = ASCII + ((namedclass - ANYOF_ASCII) ^ invert);
17767 /* The following don't have any matches in the upper Latin1
17768 * range, hence /d is equivalent to /u for them. Making it /u
17769 * saves some branches at runtime */
17773 case ANYOF_NXDIGIT:
17774 if (! DEPENDS_SEMANTICS) {
17775 goto treat_as_default;
17781 /* The following change to CASED under /i */
17787 namedclass = ANYOF_CASED + (namedclass % 2);
17791 /* The rest have more possibilities depending on the charset.
17792 * We take advantage of the enum ordering of the charset
17793 * modifiers to get the exact node type, */
17796 op = POSIXD + get_regex_charset(RExC_flags);
17797 if (op > POSIXA) { /* /aa is same as /a */
17802 /* The odd numbered ones are the complements of the
17803 * next-lower even number one */
17804 if (namedclass % 2 == 1) {
17808 arg = namedclass_to_classnum(namedclass);
17812 else if (value == prevvalue) {
17814 /* Here, the class consists of just a single code point */
17817 if (! LOC && value == '\n') {
17818 op = REG_ANY; /* Optimize [^\n] */
17819 *flagp |= HASWIDTH|SIMPLE;
17823 else if (value < 256 || UTF) {
17825 /* Optimize a single value into an EXACTish node, but not if it
17826 * would require converting the pattern to UTF-8. */
17827 op = compute_EXACTish(pRExC_state);
17829 } /* Otherwise is a range */
17830 else if (! LOC) { /* locale could vary these */
17831 if (prevvalue == '0') {
17832 if (value == '9') {
17837 else if (! FOLD || ASCII_FOLD_RESTRICTED) {
17838 /* We can optimize A-Z or a-z, but not if they could match
17839 * something like the KELVIN SIGN under /i. */
17840 if (prevvalue == 'A') {
17843 && ! non_portable_endpoint
17846 arg = (FOLD) ? _CC_ALPHA : _CC_UPPER;
17850 else if (prevvalue == 'a') {
17853 && ! non_portable_endpoint
17856 arg = (FOLD) ? _CC_ALPHA : _CC_LOWER;
17863 /* Here, we have changed <op> away from its initial value iff we found
17864 * an optimization */
17867 /* Throw away this ANYOF regnode, and emit the calculated one,
17868 * which should correspond to the beginning, not current, state of
17870 const char * cur_parse = RExC_parse;
17871 RExC_parse = (char *)orig_parse;
17875 /* To get locale nodes to not use the full ANYOF size would
17876 * require moving the code above that writes the portions
17877 * of it that aren't in other nodes to after this point.
17878 * e.g. ANYOF_POSIXL_SET */
17879 RExC_size = orig_size;
17883 RExC_emit = (regnode *)orig_emit;
17884 if (PL_regkind[op] == POSIXD) {
17885 if (op == POSIXL) {
17886 RExC_contains_locale = 1;
17889 op += NPOSIXD - POSIXD;
17894 ret = reg_node(pRExC_state, op);
17896 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
17900 *flagp |= HASWIDTH|SIMPLE;
17902 else if (PL_regkind[op] == EXACT) {
17903 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
17904 TRUE /* downgradable to EXACT */
17908 *flagp |= HASWIDTH|SIMPLE;
17911 RExC_parse = (char *) cur_parse;
17913 SvREFCNT_dec(posixes);
17914 SvREFCNT_dec(nposixes);
17915 SvREFCNT_dec(simple_posixes);
17916 SvREFCNT_dec(cp_list);
17917 SvREFCNT_dec(cp_foldable_list);
17924 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
17926 /* If folding, we calculate all characters that could fold to or from the
17927 * ones already on the list */
17928 if (cp_foldable_list) {
17930 UV start, end; /* End points of code point ranges */
17932 SV* fold_intersection = NULL;
17935 /* Our calculated list will be for Unicode rules. For locale
17936 * matching, we have to keep a separate list that is consulted at
17937 * runtime only when the locale indicates Unicode rules. For
17938 * non-locale, we just use the general list */
17940 use_list = &only_utf8_locale_list;
17943 use_list = &cp_list;
17946 /* Only the characters in this class that participate in folds need
17947 * be checked. Get the intersection of this class and all the
17948 * possible characters that are foldable. This can quickly narrow
17949 * down a large class */
17950 _invlist_intersection(PL_utf8_foldable, cp_foldable_list,
17951 &fold_intersection);
17953 /* Now look at the foldable characters in this class individually */
17954 invlist_iterinit(fold_intersection);
17955 while (invlist_iternext(fold_intersection, &start, &end)) {
17959 /* Look at every character in the range */
17960 for (j = start; j <= end; j++) {
17961 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
17964 Size_t folds_to_count;
17965 unsigned int first_folds_to;
17966 const unsigned int * remaining_folds_to_list;
17970 if (IS_IN_SOME_FOLD_L1(j)) {
17972 /* ASCII is always matched; non-ASCII is matched
17973 * only under Unicode rules (which could happen
17974 * under /l if the locale is a UTF-8 one */
17975 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
17976 *use_list = add_cp_to_invlist(*use_list,
17977 PL_fold_latin1[j]);
17980 has_upper_latin1_only_utf8_matches
17981 = add_cp_to_invlist(
17982 has_upper_latin1_only_utf8_matches,
17983 PL_fold_latin1[j]);
17987 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(j)
17988 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
17990 add_above_Latin1_folds(pRExC_state,
17997 /* Here is an above Latin1 character. We don't have the
17998 * rules hard-coded for it. First, get its fold. This is
17999 * the simple fold, as the multi-character folds have been
18000 * handled earlier and separated out */
18001 folded = _to_uni_fold_flags(j, foldbuf, &foldlen,
18002 (ASCII_FOLD_RESTRICTED)
18003 ? FOLD_FLAGS_NOMIX_ASCII
18006 /* Single character fold of above Latin1. Add everything
18007 * in its fold closure to the list that this node should
18009 folds_to_count = _inverse_folds(folded, &first_folds_to,
18010 &remaining_folds_to_list);
18011 for (k = 0; k <= folds_to_count; k++) {
18012 UV c = (k == 0) /* First time through use itself */
18014 : (k == 1) /* 2nd time use, the first fold */
18017 /* Then the remaining ones */
18018 : remaining_folds_to_list[k-2];
18020 /* /aa doesn't allow folds between ASCII and non- */
18021 if (( ASCII_FOLD_RESTRICTED
18022 && (isASCII(c) != isASCII(j))))
18027 /* Folds under /l which cross the 255/256 boundary are
18028 * added to a separate list. (These are valid only
18029 * when the locale is UTF-8.) */
18030 if (c < 256 && LOC) {
18031 *use_list = add_cp_to_invlist(*use_list, c);
18035 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
18037 cp_list = add_cp_to_invlist(cp_list, c);
18040 /* Similarly folds involving non-ascii Latin1
18041 * characters under /d are added to their list */
18042 has_upper_latin1_only_utf8_matches
18043 = add_cp_to_invlist(
18044 has_upper_latin1_only_utf8_matches,
18050 SvREFCNT_dec_NN(fold_intersection);
18053 /* Now that we have finished adding all the folds, there is no reason
18054 * to keep the foldable list separate */
18055 _invlist_union(cp_list, cp_foldable_list, &cp_list);
18056 SvREFCNT_dec_NN(cp_foldable_list);
18059 /* And combine the result (if any) with any inversion lists from posix
18060 * classes. The lists are kept separate up to now because we don't want to
18061 * fold the classes (folding of those is automatically handled by the swash
18062 * fetching code) */
18063 if (simple_posixes) { /* These are the classes known to be unaffected by
18066 _invlist_union(cp_list, simple_posixes, &cp_list);
18067 SvREFCNT_dec_NN(simple_posixes);
18070 cp_list = simple_posixes;
18073 if (posixes || nposixes) {
18074 if (! DEPENDS_SEMANTICS) {
18076 /* For everything but /d, we can just add the current 'posixes' and
18077 * 'nposixes' to the main list */
18080 _invlist_union(cp_list, posixes, &cp_list);
18081 SvREFCNT_dec_NN(posixes);
18089 _invlist_union(cp_list, nposixes, &cp_list);
18090 SvREFCNT_dec_NN(nposixes);
18093 cp_list = nposixes;
18098 /* Under /d, things like \w match upper Latin1 characters only if
18099 * the target string is in UTF-8. But things like \W match all the
18100 * upper Latin1 characters if the target string is not in UTF-8.
18102 * Handle the case where there something like \W separately */
18104 SV* only_non_utf8_list = invlist_clone(PL_UpperLatin1);
18106 /* A complemented posix class matches all upper Latin1
18107 * characters if not in UTF-8. And it matches just certain
18108 * ones when in UTF-8. That means those certain ones are
18109 * matched regardless, so can just be added to the
18110 * unconditional list */
18112 _invlist_union(cp_list, nposixes, &cp_list);
18113 SvREFCNT_dec_NN(nposixes);
18117 cp_list = nposixes;
18120 /* Likewise for 'posixes' */
18121 _invlist_union(posixes, cp_list, &cp_list);
18123 /* Likewise for anything else in the range that matched only
18125 if (has_upper_latin1_only_utf8_matches) {
18126 _invlist_union(cp_list,
18127 has_upper_latin1_only_utf8_matches,
18129 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
18130 has_upper_latin1_only_utf8_matches = NULL;
18133 /* If we don't match all the upper Latin1 characters regardless
18134 * of UTF-8ness, we have to set a flag to match the rest when
18136 _invlist_subtract(only_non_utf8_list, cp_list,
18137 &only_non_utf8_list);
18138 if (_invlist_len(only_non_utf8_list) != 0) {
18139 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
18141 SvREFCNT_dec_NN(only_non_utf8_list);
18144 /* Here there were no complemented posix classes. That means
18145 * the upper Latin1 characters in 'posixes' match only when the
18146 * target string is in UTF-8. So we have to add them to the
18147 * list of those types of code points, while adding the
18148 * remainder to the unconditional list.
18150 * First calculate what they are */
18151 SV* nonascii_but_latin1_properties = NULL;
18152 _invlist_intersection(posixes, PL_UpperLatin1,
18153 &nonascii_but_latin1_properties);
18155 /* And add them to the final list of such characters. */
18156 _invlist_union(has_upper_latin1_only_utf8_matches,
18157 nonascii_but_latin1_properties,
18158 &has_upper_latin1_only_utf8_matches);
18160 /* Remove them from what now becomes the unconditional list */
18161 _invlist_subtract(posixes, nonascii_but_latin1_properties,
18164 /* And add those unconditional ones to the final list */
18166 _invlist_union(cp_list, posixes, &cp_list);
18167 SvREFCNT_dec_NN(posixes);
18174 SvREFCNT_dec(nonascii_but_latin1_properties);
18176 /* Get rid of any characters that we now know are matched
18177 * unconditionally from the conditional list, which may make
18178 * that list empty */
18179 _invlist_subtract(has_upper_latin1_only_utf8_matches,
18181 &has_upper_latin1_only_utf8_matches);
18182 if (_invlist_len(has_upper_latin1_only_utf8_matches) == 0) {
18183 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
18184 has_upper_latin1_only_utf8_matches = NULL;
18190 /* And combine the result (if any) with any inversion list from properties.
18191 * The lists are kept separate up to now so that we can distinguish the two
18192 * in regards to matching above-Unicode. A run-time warning is generated
18193 * if a Unicode property is matched against a non-Unicode code point. But,
18194 * we allow user-defined properties to match anything, without any warning,
18195 * and we also suppress the warning if there is a portion of the character
18196 * class that isn't a Unicode property, and which matches above Unicode, \W
18197 * or [\x{110000}] for example.
18198 * (Note that in this case, unlike the Posix one above, there is no
18199 * <has_upper_latin1_only_utf8_matches>, because having a Unicode property
18200 * forces Unicode semantics */
18204 /* If it matters to the final outcome, see if a non-property
18205 * component of the class matches above Unicode. If so, the
18206 * warning gets suppressed. This is true even if just a single
18207 * such code point is specified, as, though not strictly correct if
18208 * another such code point is matched against, the fact that they
18209 * are using above-Unicode code points indicates they should know
18210 * the issues involved */
18212 warn_super = ! (invert
18213 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
18216 _invlist_union(properties, cp_list, &cp_list);
18217 SvREFCNT_dec_NN(properties);
18220 cp_list = properties;
18225 |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
18227 /* Because an ANYOF node is the only one that warns, this node
18228 * can't be optimized into something else */
18229 optimizable = FALSE;
18233 /* Here, we have calculated what code points should be in the character
18236 * Now we can see about various optimizations. Fold calculation (which we
18237 * did above) needs to take place before inversion. Otherwise /[^k]/i
18238 * would invert to include K, which under /i would match k, which it
18239 * shouldn't. Therefore we can't invert folded locale now, as it won't be
18240 * folded until runtime */
18242 /* If we didn't do folding, it's because some information isn't available
18243 * until runtime; set the run-time fold flag for these. (We don't have to
18244 * worry about properties folding, as that is taken care of by the swash
18245 * fetching). We know to set the flag if we have a non-NULL list for UTF-8
18246 * locales, or the class matches at least one 0-255 range code point */
18249 /* Some things on the list might be unconditionally included because of
18250 * other components. Remove them, and clean up the list if it goes to
18252 if (only_utf8_locale_list && cp_list) {
18253 _invlist_subtract(only_utf8_locale_list, cp_list,
18254 &only_utf8_locale_list);
18256 if (_invlist_len(only_utf8_locale_list) == 0) {
18257 SvREFCNT_dec_NN(only_utf8_locale_list);
18258 only_utf8_locale_list = NULL;
18261 if (only_utf8_locale_list) {
18264 |ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
18266 else if (cp_list) { /* Look to see if a 0-255 code point is in list */
18268 invlist_iterinit(cp_list);
18269 if (invlist_iternext(cp_list, &start, &end) && start < 256) {
18270 ANYOF_FLAGS(ret) |= ANYOFL_FOLD;
18272 invlist_iterfinish(cp_list);
18275 else if ( DEPENDS_SEMANTICS
18276 && ( has_upper_latin1_only_utf8_matches
18277 || (ANYOF_FLAGS(ret) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)))
18280 optimizable = FALSE;
18284 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
18285 * at compile time. Besides not inverting folded locale now, we can't
18286 * invert if there are things such as \w, which aren't known until runtime
18290 && OP(ret) != ANYOFD
18291 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOCALE_FLAGS))
18292 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
18294 _invlist_invert(cp_list);
18296 /* Any swash can't be used as-is, because we've inverted things */
18298 SvREFCNT_dec_NN(swash);
18302 /* Clear the invert flag since have just done it here */
18309 *ret_invlist = cp_list;
18310 SvREFCNT_dec(swash);
18312 /* Discard the generated node */
18314 RExC_size = orig_size;
18317 RExC_emit = orig_emit;
18322 /* Some character classes are equivalent to other nodes. Such nodes take
18323 * up less room and generally fewer operations to execute than ANYOF nodes.
18324 * Above, we checked for and optimized into some such equivalents for
18325 * certain common classes that are easy to test. Getting to this point in
18326 * the code means that the class didn't get optimized there. Since this
18327 * code is only executed in Pass 2, it is too late to save space--it has
18328 * been allocated in Pass 1, and currently isn't given back. XXX Why not?
18329 * But turning things into an EXACTish node can allow the optimizer to join
18330 * it to any adjacent such nodes. And if the class is equivalent to things
18331 * like /./, expensive run-time swashes can be avoided. Now that we have
18332 * more complete information, we can find things necessarily missed by the
18335 if (optimizable && cp_list && ! invert) {
18337 U8 op = END; /* The optimzation node-type */
18338 int posix_class = -1; /* Illegal value */
18339 const char * cur_parse= RExC_parse;
18340 U8 ANYOFM_mask = 0xFF;
18343 invlist_iterinit(cp_list);
18344 if (! invlist_iternext(cp_list, &start, &end)) {
18346 /* Here, the list is empty. This happens, for example, when a
18347 * Unicode property that doesn't match anything is the only element
18348 * in the character class (perluniprops.pod notes such properties).
18351 *flagp |= HASWIDTH|SIMPLE;
18353 else if (start == end) { /* The range is a single code point */
18354 if (! invlist_iternext(cp_list, &start, &end)
18356 /* Don't do this optimization if it would require changing
18357 * the pattern to UTF-8 */
18358 && (start < 256 || UTF))
18360 /* Here, the list contains a single code point. Can optimize
18361 * into an EXACTish node */
18372 /* A locale node under folding with one code point can be
18373 * an EXACTFL, as its fold won't be calculated until
18379 /* Here, we are generally folding, but there is only one
18380 * code point to match. If we have to, we use an EXACT
18381 * node, but it would be better for joining with adjacent
18382 * nodes in the optimization pass if we used the same
18383 * EXACTFish node that any such are likely to be. We can
18384 * do this iff the code point doesn't participate in any
18385 * folds. For example, an EXACTF of a colon is the same as
18386 * an EXACT one, since nothing folds to or from a colon. */
18388 if (IS_IN_SOME_FOLD_L1(value)) {
18393 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
18398 /* If we haven't found the node type, above, it means we
18399 * can use the prevailing one */
18401 op = compute_EXACTish(pRExC_state);
18405 } /* End of first range contains just a single code point */
18406 else if (start == 0) {
18407 if (end == UV_MAX) {
18409 *flagp |= HASWIDTH|SIMPLE;
18412 else if (end == '\n' - 1
18413 && invlist_iternext(cp_list, &start, &end)
18414 && start == '\n' + 1 && end == UV_MAX)
18417 *flagp |= HASWIDTH|SIMPLE;
18421 invlist_iterfinish(cp_list);
18425 /* Here, didn't find an optimization. See if this matches any of
18426 * the POSIX classes. First try ASCII */
18428 if (_invlistEQ(cp_list, PL_XPosix_ptrs[_CC_ASCII], 0)) {
18430 *flagp |= HASWIDTH|SIMPLE;
18432 else if (_invlistEQ(cp_list, PL_XPosix_ptrs[_CC_ASCII], 1)) {
18434 *flagp |= HASWIDTH|SIMPLE;
18436 else if (invlist_highest(cp_list) >= 0x2029) {
18438 /* Then try the other POSIX classes. The POSIXA ones are about
18439 * the same speed as ANYOF ops, but the ones that have
18440 * above-Latin1 code point matches are somewhat faster than
18441 * ANYOF. So optimize those, but don't bother with the POSIXA
18442 * ones nor [:cntrl:] which has no above-Latin1 matches. If
18443 * this ANYOF node has a lower highest possible matching code
18444 * point than any of the XPosix ones, we know that it can't
18445 * possibly be the same as any of them, so we can avoid
18446 * executing this code. The 0x2029 above for the lowest max
18447 * was determined by manual inspection of the classes, and
18448 * comes from \v. Suppose Unicode in a later version adds a
18449 * higher code point to \v. All that means is that this code
18450 * can be executed unnecessarily. It will still give the
18451 * correct answer. */
18453 for (posix_class = 0;
18454 posix_class <= _HIGHEST_REGCOMP_DOT_H_SYNC;
18459 if (posix_class == _CC_CNTRL) {
18463 for (try_inverted = 0; try_inverted < 2; try_inverted++) {
18465 /* Check if matches normal or inverted */
18466 if (_invlistEQ(cp_list,
18467 PL_XPosix_ptrs[posix_class],
18470 op = (try_inverted)
18473 *flagp |= HASWIDTH|SIMPLE;
18481 /* If it didn't match a POSIX class, it might be able to be turned
18482 * into an ANYOFM node. Compare two different bytes, bit-by-bit.
18483 * In some positions, the bits in each will be 1; and in other
18484 * positions both will be 0; and in some positions the bit will be
18485 * 1 in one byte, and 0 in the other. Let 'n' be the number of
18486 * positions where the bits differ. We create a mask which has
18487 * exactly 'n' 0 bits, each in a position where the two bytes
18488 * differ. Now take the set of all bytes that when ANDed with the
18489 * mask yield the same result. That set has 2**n elements, and is
18490 * representable by just two 8 bit numbers: the result and the
18491 * mask. Importantly, matching the set can be vectorized by
18492 * creating a word full of the result bytes, and a word full of the
18493 * mask bytes, yielding a significant speed up. Here, see if this
18494 * node matches such a set. As a concrete example consider [01],
18495 * and the byte representing '0' which is 0x30 on ASCII machines.
18496 * It has the bits 0011 0000. Take the mask 1111 1110. If we AND
18497 * 0x31 and 0x30 with that mask we get 0x30. Any other bytes ANDed
18498 * yield something else. So [01], which is a common usage, is
18499 * optimizable into ANYOFM, and can benefit from the speed up. We
18500 * can only do this on UTF-8 invariant bytes, because the variance
18501 * would throw this off. */
18503 && invlist_highest(cp_list) <=
18510 Size_t cp_count = 0;
18511 bool first_time = TRUE;
18512 unsigned int lowest_cp = 0xFF;
18513 U8 bits_differing = 0;
18515 /* Only needed on EBCDIC, as there, variants and non- are mixed
18516 * together. Could #ifdef it out on ASCII, but probably the
18517 * compiler will optimize it out */
18518 bool has_variant = FALSE;
18520 /* Go through the bytes and find the bit positions that differ */
18521 invlist_iterinit(cp_list);
18522 while (invlist_iternext(cp_list, &start, &end)) {
18523 unsigned int i = start;
18525 cp_count += end - start + 1;
18528 if (! UVCHR_IS_INVARIANT(i)) {
18529 has_variant = TRUE;
18533 first_time = FALSE;
18539 /* Find the bit positions that differ from the lowest code
18540 * point in the node. Keep track of all such positions by
18542 for (; i <= end; i++) {
18543 if (! UVCHR_IS_INVARIANT(i)) {
18544 has_variant = TRUE;
18548 bits_differing |= i ^ lowest_cp;
18551 invlist_iterfinish(cp_list);
18553 /* At the end of the loop, we count how many bits differ from
18554 * the bits in lowest code point, call the count 'd'. If the
18555 * set we found contains 2**d elements, it is the closure of
18556 * all code points that differ only in those bit positions. To
18557 * convince yourself of that, first note that the number in the
18558 * closure must be a power of 2, which we test for. The only
18559 * way we could have that count and it be some differing set,
18560 * is if we got some code points that don't differ from the
18561 * lowest code point in any position, but do differ from each
18562 * other in some other position. That means one code point has
18563 * a 1 in that position, and another has a 0. But that would
18564 * mean that one of them differs from the lowest code point in
18565 * that position, which possibility we've already excluded. */
18567 && cp_count == 1U << PL_bitcount[bits_differing])
18569 assert(cp_count > 1);
18572 /* We need to make the bits that differ be 0's */
18573 ANYOFM_mask = ~ bits_differing; /* This goes into FLAGS */
18575 /* The argument is the lowest code point */
18576 anode_arg = lowest_cp;
18577 *flagp |= HASWIDTH|SIMPLE;
18583 RExC_parse = (char *)orig_parse;
18584 RExC_emit = (regnode *)orig_emit;
18586 if (regarglen[op]) {
18587 ret = reganode(pRExC_state, op, anode_arg);
18589 ret = reg_node(pRExC_state, op);
18592 RExC_parse = (char *)cur_parse;
18594 if (PL_regkind[op] == EXACT) {
18595 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
18596 TRUE /* downgradable to EXACT */
18599 else if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
18600 FLAGS(ret) = posix_class;
18602 else if (PL_regkind[op] == ANYOFM) {
18603 FLAGS(ret) = ANYOFM_mask;
18606 SvREFCNT_dec_NN(cp_list);
18611 /* Here, <cp_list> contains all the code points we can determine at
18612 * compile time that match under all conditions. Go through it, and
18613 * for things that belong in the bitmap, put them there, and delete from
18614 * <cp_list>. While we are at it, see if everything above 255 is in the
18615 * list, and if so, set a flag to speed up execution */
18617 populate_ANYOF_from_invlist(ret, &cp_list);
18620 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
18623 /* Here, the bitmap has been populated with all the Latin1 code points that
18624 * always match. Can now add to the overall list those that match only
18625 * when the target string is UTF-8 (<has_upper_latin1_only_utf8_matches>).
18627 if (has_upper_latin1_only_utf8_matches) {
18629 _invlist_union(cp_list,
18630 has_upper_latin1_only_utf8_matches,
18632 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
18635 cp_list = has_upper_latin1_only_utf8_matches;
18637 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
18640 /* If there is a swash and more than one element, we can't use the swash in
18641 * the optimization below. */
18642 if (swash && element_count > 1) {
18643 SvREFCNT_dec_NN(swash);
18647 /* Note that the optimization of using 'swash' if it is the only thing in
18648 * the class doesn't have us change swash at all, so it can include things
18649 * that are also in the bitmap; otherwise we have purposely deleted that
18650 * duplicate information */
18651 set_ANYOF_arg(pRExC_state, ret, cp_list,
18652 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
18654 only_utf8_locale_list,
18655 swash, has_user_defined_property);
18657 *flagp |= HASWIDTH|SIMPLE;
18659 if (ANYOF_FLAGS(ret) & ANYOF_LOCALE_FLAGS) {
18660 RExC_contains_locale = 1;
18666 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
18669 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
18670 regnode* const node,
18672 SV* const runtime_defns,
18673 SV* const only_utf8_locale_list,
18675 const bool has_user_defined_property)
18677 /* Sets the arg field of an ANYOF-type node 'node', using information about
18678 * the node passed-in. If there is nothing outside the node's bitmap, the
18679 * arg is set to ANYOF_ONLY_HAS_BITMAP. Otherwise, it sets the argument to
18680 * the count returned by add_data(), having allocated and stored an array,
18681 * av, that that count references, as follows:
18682 * av[0] stores the character class description in its textual form.
18683 * This is used later (regexec.c:Perl_regclass_swash()) to
18684 * initialize the appropriate swash, and is also useful for dumping
18685 * the regnode. This is set to &PL_sv_undef if the textual
18686 * description is not needed at run-time (as happens if the other
18687 * elements completely define the class)
18688 * av[1] if &PL_sv_undef, is a placeholder to later contain the swash
18689 * computed from av[0]. But if no further computation need be done,
18690 * the swash is stored here now (and av[0] is &PL_sv_undef).
18691 * av[2] stores the inversion list of code points that match only if the
18692 * current locale is UTF-8
18693 * av[3] stores the cp_list inversion list for use in addition or instead
18694 * of av[0]; used only if cp_list exists and av[1] is &PL_sv_undef.
18695 * (Otherwise everything needed is already in av[0] and av[1])
18696 * av[4] is set if any component of the class is from a user-defined
18697 * property; used only if av[3] exists */
18701 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
18703 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
18704 assert(! (ANYOF_FLAGS(node)
18705 & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP));
18706 ARG_SET(node, ANYOF_ONLY_HAS_BITMAP);
18709 AV * const av = newAV();
18712 av_store(av, 0, (runtime_defns)
18713 ? SvREFCNT_inc(runtime_defns) : &PL_sv_undef);
18716 av_store(av, 1, swash);
18717 SvREFCNT_dec_NN(cp_list);
18720 av_store(av, 1, &PL_sv_undef);
18722 av_store(av, 3, cp_list);
18723 av_store(av, 4, newSVuv(has_user_defined_property));
18727 if (only_utf8_locale_list) {
18728 av_store(av, 2, only_utf8_locale_list);
18731 av_store(av, 2, &PL_sv_undef);
18734 rv = newRV_noinc(MUTABLE_SV(av));
18735 n = add_data(pRExC_state, STR_WITH_LEN("s"));
18736 RExC_rxi->data->data[n] = (void*)rv;
18741 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
18743 Perl__get_regclass_nonbitmap_data(pTHX_ const regexp *prog,
18744 const regnode* node,
18747 SV** only_utf8_locale_ptr,
18748 SV** output_invlist)
18751 /* For internal core use only.
18752 * Returns the swash for the input 'node' in the regex 'prog'.
18753 * If <doinit> is 'true', will attempt to create the swash if not already
18755 * If <listsvp> is non-null, will return the printable contents of the
18756 * swash. This can be used to get debugging information even before the
18757 * swash exists, by calling this function with 'doinit' set to false, in
18758 * which case the components that will be used to eventually create the
18759 * swash are returned (in a printable form).
18760 * If <only_utf8_locale_ptr> is not NULL, it is where this routine is to
18761 * store an inversion list of code points that should match only if the
18762 * execution-time locale is a UTF-8 one.
18763 * If <output_invlist> is not NULL, it is where this routine is to store an
18764 * inversion list of the code points that would be instead returned in
18765 * <listsvp> if this were NULL. Thus, what gets output in <listsvp>
18766 * when this parameter is used, is just the non-code point data that
18767 * will go into creating the swash. This currently should be just
18768 * user-defined properties whose definitions were not known at compile
18769 * time. Using this parameter allows for easier manipulation of the
18770 * swash's data by the caller. It is illegal to call this function with
18771 * this parameter set, but not <listsvp>
18773 * Tied intimately to how S_set_ANYOF_arg sets up the data structure. Note
18774 * that, in spite of this function's name, the swash it returns may include
18775 * the bitmap data as well */
18778 SV *si = NULL; /* Input swash initialization string */
18779 SV* invlist = NULL;
18781 RXi_GET_DECL(prog,progi);
18782 const struct reg_data * const data = prog ? progi->data : NULL;
18784 PERL_ARGS_ASSERT__GET_REGCLASS_NONBITMAP_DATA;
18785 assert(! output_invlist || listsvp);
18787 if (data && data->count) {
18788 const U32 n = ARG(node);
18790 if (data->what[n] == 's') {
18791 SV * const rv = MUTABLE_SV(data->data[n]);
18792 AV * const av = MUTABLE_AV(SvRV(rv));
18793 SV **const ary = AvARRAY(av);
18794 U8 swash_init_flags = _CORE_SWASH_INIT_ACCEPT_INVLIST;
18796 si = *ary; /* ary[0] = the string to initialize the swash with */
18798 if (av_tindex_skip_len_mg(av) >= 2) {
18799 if (only_utf8_locale_ptr
18801 && ary[2] != &PL_sv_undef)
18803 *only_utf8_locale_ptr = ary[2];
18806 assert(only_utf8_locale_ptr);
18807 *only_utf8_locale_ptr = NULL;
18810 /* Elements 3 and 4 are either both present or both absent. [3]
18811 * is any inversion list generated at compile time; [4]
18812 * indicates if that inversion list has any user-defined
18813 * properties in it. */
18814 if (av_tindex_skip_len_mg(av) >= 3) {
18816 if (SvUV(ary[4])) {
18817 swash_init_flags |= _CORE_SWASH_INIT_USER_DEFINED_PROPERTY;
18825 /* Element [1] is reserved for the set-up swash. If already there,
18826 * return it; if not, create it and store it there */
18827 if (ary[1] && SvROK(ary[1])) {
18830 else if (doinit && ((si && si != &PL_sv_undef)
18831 || (invlist && invlist != &PL_sv_undef))) {
18833 sw = _core_swash_init("utf8", /* the utf8 package */
18837 0, /* not from tr/// */
18839 &swash_init_flags);
18840 (void)av_store(av, 1, sw);
18845 /* If requested, return a printable version of what this swash matches */
18847 SV* matches_string = NULL;
18849 /* The swash should be used, if possible, to get the data, as it
18850 * contains the resolved data. But this function can be called at
18851 * compile-time, before everything gets resolved, in which case we
18852 * return the currently best available information, which is the string
18853 * that will eventually be used to do that resolving, 'si' */
18854 if ((! sw || (invlist = _get_swash_invlist(sw)) == NULL)
18855 && (si && si != &PL_sv_undef))
18857 /* Here, we only have 'si' (and possibly some passed-in data in
18858 * 'invlist', which is handled below) If the caller only wants
18859 * 'si', use that. */
18860 if (! output_invlist) {
18861 matches_string = newSVsv(si);
18864 /* But if the caller wants an inversion list of the node, we
18865 * need to parse 'si' and place as much as possible in the
18866 * desired output inversion list, making 'matches_string' only
18867 * contain the currently unresolvable things */
18868 const char *si_string = SvPVX(si);
18869 STRLEN remaining = SvCUR(si);
18873 /* Ignore everything before the first new-line */
18874 while (*si_string != '\n' && remaining > 0) {
18878 assert(remaining > 0);
18883 while (remaining > 0) {
18885 /* The data consists of just strings defining user-defined
18886 * property names, but in prior incarnations, and perhaps
18887 * somehow from pluggable regex engines, it could still
18888 * hold hex code point definitions. Each component of a
18889 * range would be separated by a tab, and each range by a
18890 * new-line. If these are found, instead add them to the
18891 * inversion list */
18892 I32 grok_flags = PERL_SCAN_SILENT_ILLDIGIT
18893 |PERL_SCAN_SILENT_NON_PORTABLE;
18894 STRLEN len = remaining;
18895 UV cp = grok_hex(si_string, &len, &grok_flags, NULL);
18897 /* If the hex decode routine found something, it should go
18898 * up to the next \n */
18899 if ( *(si_string + len) == '\n') {
18900 if (count) { /* 2nd code point on line */
18901 *output_invlist = _add_range_to_invlist(*output_invlist, prev_cp, cp);
18904 *output_invlist = add_cp_to_invlist(*output_invlist, cp);
18907 goto prepare_for_next_iteration;
18910 /* If the hex decode was instead for the lower range limit,
18911 * save it, and go parse the upper range limit */
18912 if (*(si_string + len) == '\t') {
18913 assert(count == 0);
18917 prepare_for_next_iteration:
18918 si_string += len + 1;
18919 remaining -= len + 1;
18923 /* Here, didn't find a legal hex number. Just add it from
18924 * here to the next \n */
18927 while (*(si_string + len) != '\n' && remaining > 0) {
18931 if (*(si_string + len) == '\n') {
18935 if (matches_string) {
18936 sv_catpvn(matches_string, si_string, len - 1);
18939 matches_string = newSVpvn(si_string, len - 1);
18942 sv_catpvs(matches_string, " ");
18943 } /* end of loop through the text */
18945 assert(matches_string);
18946 if (SvCUR(matches_string)) { /* Get rid of trailing blank */
18947 SvCUR_set(matches_string, SvCUR(matches_string) - 1);
18949 } /* end of has an 'si' but no swash */
18952 /* If we have a swash in place, its equivalent inversion list was above
18953 * placed into 'invlist'. If not, this variable may contain a stored
18954 * inversion list which is information beyond what is in 'si' */
18957 /* Again, if the caller doesn't want the output inversion list, put
18958 * everything in 'matches-string' */
18959 if (! output_invlist) {
18960 if ( ! matches_string) {
18961 matches_string = newSVpvs("\n");
18963 sv_catsv(matches_string, invlist_contents(invlist,
18964 TRUE /* traditional style */
18967 else if (! *output_invlist) {
18968 *output_invlist = invlist_clone(invlist);
18971 _invlist_union(*output_invlist, invlist, output_invlist);
18975 *listsvp = matches_string;
18980 #endif /* !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION) */
18982 /* reg_skipcomment()
18984 Absorbs an /x style # comment from the input stream,
18985 returning a pointer to the first character beyond the comment, or if the
18986 comment terminates the pattern without anything following it, this returns
18987 one past the final character of the pattern (in other words, RExC_end) and
18988 sets the REG_RUN_ON_COMMENT_SEEN flag.
18990 Note it's the callers responsibility to ensure that we are
18991 actually in /x mode
18995 PERL_STATIC_INLINE char*
18996 S_reg_skipcomment(RExC_state_t *pRExC_state, char* p)
18998 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
19002 while (p < RExC_end) {
19003 if (*(++p) == '\n') {
19008 /* we ran off the end of the pattern without ending the comment, so we have
19009 * to add an \n when wrapping */
19010 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
19015 S_skip_to_be_ignored_text(pTHX_ RExC_state_t *pRExC_state,
19017 const bool force_to_xmod
19020 /* If the text at the current parse position '*p' is a '(?#...)' comment,
19021 * or if we are under /x or 'force_to_xmod' is TRUE, and the text at '*p'
19022 * is /x whitespace, advance '*p' so that on exit it points to the first
19023 * byte past all such white space and comments */
19025 const bool use_xmod = force_to_xmod || (RExC_flags & RXf_PMf_EXTENDED);
19027 PERL_ARGS_ASSERT_SKIP_TO_BE_IGNORED_TEXT;
19029 assert( ! UTF || UTF8_IS_INVARIANT(**p) || UTF8_IS_START(**p));
19032 if (RExC_end - (*p) >= 3
19034 && *(*p + 1) == '?'
19035 && *(*p + 2) == '#')
19037 while (*(*p) != ')') {
19038 if ((*p) == RExC_end)
19039 FAIL("Sequence (?#... not terminated");
19047 const char * save_p = *p;
19048 while ((*p) < RExC_end) {
19050 if ((len = is_PATWS_safe((*p), RExC_end, UTF))) {
19053 else if (*(*p) == '#') {
19054 (*p) = reg_skipcomment(pRExC_state, (*p));
19060 if (*p != save_p) {
19073 Advances the parse position by one byte, unless that byte is the beginning
19074 of a '(?#...)' style comment, or is /x whitespace and /x is in effect. In
19075 those two cases, the parse position is advanced beyond all such comments and
19078 This is the UTF, (?#...), and /x friendly way of saying RExC_parse++.
19082 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
19084 PERL_ARGS_ASSERT_NEXTCHAR;
19086 if (RExC_parse < RExC_end) {
19088 || UTF8_IS_INVARIANT(*RExC_parse)
19089 || UTF8_IS_START(*RExC_parse));
19091 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
19093 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
19094 FALSE /* Don't force /x */ );
19099 S_regnode_guts(pTHX_ RExC_state_t *pRExC_state, const U8 op, const STRLEN extra_size, const char* const name)
19101 /* Allocate a regnode for 'op' and returns it, with 'extra_size' extra
19102 * space. In pass1, it aligns and increments RExC_size; in pass2,
19105 regnode * const ret = RExC_emit;
19106 GET_RE_DEBUG_FLAGS_DECL;
19108 PERL_ARGS_ASSERT_REGNODE_GUTS;
19110 assert(extra_size >= regarglen[op]);
19113 SIZE_ALIGN(RExC_size);
19114 RExC_size += 1 + extra_size;
19117 if (RExC_emit >= RExC_emit_bound)
19118 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
19119 op, (void*)RExC_emit, (void*)RExC_emit_bound);
19121 NODE_ALIGN_FILL(ret);
19122 #ifndef RE_TRACK_PATTERN_OFFSETS
19123 PERL_UNUSED_ARG(name);
19125 if (RExC_offsets) { /* MJD */
19127 ("%s:%d: (op %s) %s %" UVuf " (len %" UVuf ") (max %" UVuf ").\n",
19130 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
19131 ? "Overwriting end of array!\n" : "OK",
19132 (UV)(RExC_emit - RExC_emit_start),
19133 (UV)(RExC_parse - RExC_start),
19134 (UV)RExC_offsets[0]));
19135 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
19142 - reg_node - emit a node
19144 STATIC regnode * /* Location. */
19145 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
19147 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg_node");
19149 PERL_ARGS_ASSERT_REG_NODE;
19151 assert(regarglen[op] == 0);
19154 regnode *ptr = ret;
19155 FILL_ADVANCE_NODE(ptr, op);
19162 - reganode - emit a node with an argument
19164 STATIC regnode * /* Location. */
19165 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
19167 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reganode");
19169 PERL_ARGS_ASSERT_REGANODE;
19171 assert(regarglen[op] == 1);
19174 regnode *ptr = ret;
19175 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
19182 S_reg2Lanode(pTHX_ RExC_state_t *pRExC_state, const U8 op, const U32 arg1, const I32 arg2)
19184 /* emit a node with U32 and I32 arguments */
19186 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg2Lanode");
19188 PERL_ARGS_ASSERT_REG2LANODE;
19190 assert(regarglen[op] == 2);
19193 regnode *ptr = ret;
19194 FILL_ADVANCE_NODE_2L_ARG(ptr, op, arg1, arg2);
19201 - reginsert - insert an operator in front of already-emitted operand
19203 * Means relocating the operand.
19205 * IMPORTANT NOTE - it is the *callers* responsibility to correctly
19206 * set up NEXT_OFF() of the inserted node if needed. Something like this:
19208 * reginsert(pRExC, OPFAIL, orig_emit, depth+1);
19210 * NEXT_OFF(orig_emit) = regarglen[OPFAIL] + NODE_STEP_REGNODE;
19212 * ALSO NOTE - operand->flags will be set to 0 as well.
19215 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *operand, U32 depth)
19220 const int offset = regarglen[(U8)op];
19221 const int size = NODE_STEP_REGNODE + offset;
19222 GET_RE_DEBUG_FLAGS_DECL;
19224 PERL_ARGS_ASSERT_REGINSERT;
19225 PERL_UNUSED_CONTEXT;
19226 PERL_UNUSED_ARG(depth);
19227 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
19228 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
19233 assert(!RExC_study_started); /* I believe we should never use reginsert once we have started
19234 studying. If this is wrong then we need to adjust RExC_recurse
19235 below like we do with RExC_open_parens/RExC_close_parens. */
19239 if (RExC_open_parens) {
19241 /*DEBUG_PARSE_FMT("inst"," - %" IVdf, (IV)RExC_npar);*/
19242 /* remember that RExC_npar is rex->nparens + 1,
19243 * iow it is 1 more than the number of parens seen in
19244 * the pattern so far. */
19245 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
19246 /* note, RExC_open_parens[0] is the start of the
19247 * regex, it can't move. RExC_close_parens[0] is the end
19248 * of the regex, it *can* move. */
19249 if ( paren && RExC_open_parens[paren] >= operand ) {
19250 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
19251 RExC_open_parens[paren] += size;
19253 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
19255 if ( RExC_close_parens[paren] >= operand ) {
19256 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
19257 RExC_close_parens[paren] += size;
19259 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
19264 RExC_end_op += size;
19266 while (src > operand) {
19267 StructCopy(--src, --dst, regnode);
19268 #ifdef RE_TRACK_PATTERN_OFFSETS
19269 if (RExC_offsets) { /* MJD 20010112 */
19271 ("%s(%d): (op %s) %s copy %" UVuf " -> %" UVuf " (max %" UVuf ").\n",
19275 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
19276 ? "Overwriting end of array!\n" : "OK",
19277 (UV)(src - RExC_emit_start),
19278 (UV)(dst - RExC_emit_start),
19279 (UV)RExC_offsets[0]));
19280 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
19281 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
19286 place = operand; /* Op node, where operand used to be. */
19287 #ifdef RE_TRACK_PATTERN_OFFSETS
19288 if (RExC_offsets) { /* MJD */
19290 ("%s(%d): (op %s) %s %" UVuf " <- %" UVuf " (max %" UVuf ").\n",
19294 (UV)(place - RExC_emit_start) > RExC_offsets[0]
19295 ? "Overwriting end of array!\n" : "OK",
19296 (UV)(place - RExC_emit_start),
19297 (UV)(RExC_parse - RExC_start),
19298 (UV)RExC_offsets[0]));
19299 Set_Node_Offset(place, RExC_parse);
19300 Set_Node_Length(place, 1);
19303 src = NEXTOPER(place);
19305 FILL_ADVANCE_NODE(place, op);
19306 Zero(src, offset, regnode);
19310 - regtail - set the next-pointer at the end of a node chain of p to val.
19311 - SEE ALSO: regtail_study
19314 S_regtail(pTHX_ RExC_state_t * pRExC_state,
19315 const regnode * const p,
19316 const regnode * const val,
19320 GET_RE_DEBUG_FLAGS_DECL;
19322 PERL_ARGS_ASSERT_REGTAIL;
19324 PERL_UNUSED_ARG(depth);
19330 /* Find last node. */
19331 scan = (regnode *) p;
19333 regnode * const temp = regnext(scan);
19335 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
19336 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
19337 Perl_re_printf( aTHX_ "~ %s (%d) %s %s\n",
19338 SvPV_nolen_const(RExC_mysv), REG_NODE_NUM(scan),
19339 (temp == NULL ? "->" : ""),
19340 (temp == NULL ? PL_reg_name[OP(val)] : "")
19348 if (reg_off_by_arg[OP(scan)]) {
19349 ARG_SET(scan, val - scan);
19352 NEXT_OFF(scan) = val - scan;
19358 - regtail_study - set the next-pointer at the end of a node chain of p to val.
19359 - Look for optimizable sequences at the same time.
19360 - currently only looks for EXACT chains.
19362 This is experimental code. The idea is to use this routine to perform
19363 in place optimizations on branches and groups as they are constructed,
19364 with the long term intention of removing optimization from study_chunk so
19365 that it is purely analytical.
19367 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
19368 to control which is which.
19371 /* TODO: All four parms should be const */
19374 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p,
19375 const regnode *val,U32 depth)
19379 #ifdef EXPERIMENTAL_INPLACESCAN
19382 GET_RE_DEBUG_FLAGS_DECL;
19384 PERL_ARGS_ASSERT_REGTAIL_STUDY;
19390 /* Find last node. */
19394 regnode * const temp = regnext(scan);
19395 #ifdef EXPERIMENTAL_INPLACESCAN
19396 if (PL_regkind[OP(scan)] == EXACT) {
19397 bool unfolded_multi_char; /* Unexamined in this routine */
19398 if (join_exact(pRExC_state, scan, &min,
19399 &unfolded_multi_char, 1, val, depth+1))
19404 switch (OP(scan)) {
19408 case EXACTFAA_NO_TRIE:
19414 if( exact == PSEUDO )
19416 else if ( exact != OP(scan) )
19425 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
19426 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
19427 Perl_re_printf( aTHX_ "~ %s (%d) -> %s\n",
19428 SvPV_nolen_const(RExC_mysv),
19429 REG_NODE_NUM(scan),
19430 PL_reg_name[exact]);
19437 DEBUG_PARSE_MSG("");
19438 regprop(RExC_rx, RExC_mysv, val, NULL, pRExC_state);
19439 Perl_re_printf( aTHX_
19440 "~ attach to %s (%" IVdf ") offset to %" IVdf "\n",
19441 SvPV_nolen_const(RExC_mysv),
19442 (IV)REG_NODE_NUM(val),
19446 if (reg_off_by_arg[OP(scan)]) {
19447 ARG_SET(scan, val - scan);
19450 NEXT_OFF(scan) = val - scan;
19458 S_get_ANYOFM_contents(pTHX_ const regnode * n) {
19460 /* Returns an inversion list of all the code points matched by the ANYOFM
19463 SV * cp_list = _new_invlist(-1);
19464 const U8 lowest = (U8) ARG(n);
19467 U8 needed = 1U << PL_bitcount[ (U8) ~ FLAGS(n)];
19469 PERL_ARGS_ASSERT_GET_ANYOFM_CONTENTS;
19471 /* Starting with the lowest code point, any code point that ANDed with the
19472 * mask yields the lowest code point is in the set */
19473 for (i = lowest; i <= 0xFF; i++) {
19474 if ((i & FLAGS(n)) == ARG(n)) {
19475 cp_list = add_cp_to_invlist(cp_list, i);
19478 /* We know how many code points (a power of two) that are in the
19479 * set. No use looking once we've got that number */
19480 if (count >= needed) break;
19488 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
19493 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
19498 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
19500 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
19501 if (flags & (1<<bit)) {
19502 if (!set++ && lead)
19503 Perl_re_printf( aTHX_ "%s",lead);
19504 Perl_re_printf( aTHX_ "%s ",PL_reg_intflags_name[bit]);
19509 Perl_re_printf( aTHX_ "\n");
19511 Perl_re_printf( aTHX_ "%s[none-set]\n",lead);
19516 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
19522 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
19524 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
19525 if (flags & (1<<bit)) {
19526 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
19529 if (!set++ && lead)
19530 Perl_re_printf( aTHX_ "%s",lead);
19531 Perl_re_printf( aTHX_ "%s ",PL_reg_extflags_name[bit]);
19534 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
19535 if (!set++ && lead) {
19536 Perl_re_printf( aTHX_ "%s",lead);
19539 case REGEX_UNICODE_CHARSET:
19540 Perl_re_printf( aTHX_ "UNICODE");
19542 case REGEX_LOCALE_CHARSET:
19543 Perl_re_printf( aTHX_ "LOCALE");
19545 case REGEX_ASCII_RESTRICTED_CHARSET:
19546 Perl_re_printf( aTHX_ "ASCII-RESTRICTED");
19548 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
19549 Perl_re_printf( aTHX_ "ASCII-MORE_RESTRICTED");
19552 Perl_re_printf( aTHX_ "UNKNOWN CHARACTER SET");
19558 Perl_re_printf( aTHX_ "\n");
19560 Perl_re_printf( aTHX_ "%s[none-set]\n",lead);
19566 Perl_regdump(pTHX_ const regexp *r)
19570 SV * const sv = sv_newmortal();
19571 SV *dsv= sv_newmortal();
19572 RXi_GET_DECL(r,ri);
19573 GET_RE_DEBUG_FLAGS_DECL;
19575 PERL_ARGS_ASSERT_REGDUMP;
19577 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
19579 /* Header fields of interest. */
19580 for (i = 0; i < 2; i++) {
19581 if (r->substrs->data[i].substr) {
19582 RE_PV_QUOTED_DECL(s, 0, dsv,
19583 SvPVX_const(r->substrs->data[i].substr),
19584 RE_SV_DUMPLEN(r->substrs->data[i].substr),
19585 PL_dump_re_max_len);
19586 Perl_re_printf( aTHX_
19587 "%s %s%s at %" IVdf "..%" UVuf " ",
19588 i ? "floating" : "anchored",
19590 RE_SV_TAIL(r->substrs->data[i].substr),
19591 (IV)r->substrs->data[i].min_offset,
19592 (UV)r->substrs->data[i].max_offset);
19594 else if (r->substrs->data[i].utf8_substr) {
19595 RE_PV_QUOTED_DECL(s, 1, dsv,
19596 SvPVX_const(r->substrs->data[i].utf8_substr),
19597 RE_SV_DUMPLEN(r->substrs->data[i].utf8_substr),
19599 Perl_re_printf( aTHX_
19600 "%s utf8 %s%s at %" IVdf "..%" UVuf " ",
19601 i ? "floating" : "anchored",
19603 RE_SV_TAIL(r->substrs->data[i].utf8_substr),
19604 (IV)r->substrs->data[i].min_offset,
19605 (UV)r->substrs->data[i].max_offset);
19609 if (r->check_substr || r->check_utf8)
19610 Perl_re_printf( aTHX_
19612 ( r->check_substr == r->substrs->data[1].substr
19613 && r->check_utf8 == r->substrs->data[1].utf8_substr
19614 ? "(checking floating" : "(checking anchored"));
19615 if (r->intflags & PREGf_NOSCAN)
19616 Perl_re_printf( aTHX_ " noscan");
19617 if (r->extflags & RXf_CHECK_ALL)
19618 Perl_re_printf( aTHX_ " isall");
19619 if (r->check_substr || r->check_utf8)
19620 Perl_re_printf( aTHX_ ") ");
19622 if (ri->regstclass) {
19623 regprop(r, sv, ri->regstclass, NULL, NULL);
19624 Perl_re_printf( aTHX_ "stclass %s ", SvPVX_const(sv));
19626 if (r->intflags & PREGf_ANCH) {
19627 Perl_re_printf( aTHX_ "anchored");
19628 if (r->intflags & PREGf_ANCH_MBOL)
19629 Perl_re_printf( aTHX_ "(MBOL)");
19630 if (r->intflags & PREGf_ANCH_SBOL)
19631 Perl_re_printf( aTHX_ "(SBOL)");
19632 if (r->intflags & PREGf_ANCH_GPOS)
19633 Perl_re_printf( aTHX_ "(GPOS)");
19634 Perl_re_printf( aTHX_ " ");
19636 if (r->intflags & PREGf_GPOS_SEEN)
19637 Perl_re_printf( aTHX_ "GPOS:%" UVuf " ", (UV)r->gofs);
19638 if (r->intflags & PREGf_SKIP)
19639 Perl_re_printf( aTHX_ "plus ");
19640 if (r->intflags & PREGf_IMPLICIT)
19641 Perl_re_printf( aTHX_ "implicit ");
19642 Perl_re_printf( aTHX_ "minlen %" IVdf " ", (IV)r->minlen);
19643 if (r->extflags & RXf_EVAL_SEEN)
19644 Perl_re_printf( aTHX_ "with eval ");
19645 Perl_re_printf( aTHX_ "\n");
19647 regdump_extflags("r->extflags: ",r->extflags);
19648 regdump_intflags("r->intflags: ",r->intflags);
19651 PERL_ARGS_ASSERT_REGDUMP;
19652 PERL_UNUSED_CONTEXT;
19653 PERL_UNUSED_ARG(r);
19654 #endif /* DEBUGGING */
19657 /* Should be synchronized with ANYOF_ #defines in regcomp.h */
19660 # if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 \
19661 || _CC_LOWER != 3 || _CC_UPPER != 4 || _CC_PUNCT != 5 \
19662 || _CC_PRINT != 6 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 \
19663 || _CC_CASED != 9 || _CC_SPACE != 10 || _CC_BLANK != 11 \
19664 || _CC_XDIGIT != 12 || _CC_CNTRL != 13 || _CC_ASCII != 14 \
19665 || _CC_VERTSPACE != 15
19666 # error Need to adjust order of anyofs[]
19668 static const char * const anyofs[] = {
19705 - regprop - printable representation of opcode, with run time support
19709 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo, const RExC_state_t *pRExC_state)
19713 RXi_GET_DECL(prog,progi);
19714 GET_RE_DEBUG_FLAGS_DECL;
19716 PERL_ARGS_ASSERT_REGPROP;
19720 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
19721 /* It would be nice to FAIL() here, but this may be called from
19722 regexec.c, and it would be hard to supply pRExC_state. */
19723 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
19724 (int)OP(o), (int)REGNODE_MAX);
19725 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
19727 k = PL_regkind[OP(o)];
19730 sv_catpvs(sv, " ");
19731 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
19732 * is a crude hack but it may be the best for now since
19733 * we have no flag "this EXACTish node was UTF-8"
19735 pv_pretty(sv, STRING(o), STR_LEN(o), PL_dump_re_max_len,
19736 PL_colors[0], PL_colors[1],
19737 PERL_PV_ESCAPE_UNI_DETECT |
19738 PERL_PV_ESCAPE_NONASCII |
19739 PERL_PV_PRETTY_ELLIPSES |
19740 PERL_PV_PRETTY_LTGT |
19741 PERL_PV_PRETTY_NOCLEAR
19743 } else if (k == TRIE) {
19744 /* print the details of the trie in dumpuntil instead, as
19745 * progi->data isn't available here */
19746 const char op = OP(o);
19747 const U32 n = ARG(o);
19748 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
19749 (reg_ac_data *)progi->data->data[n] :
19751 const reg_trie_data * const trie
19752 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
19754 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
19755 DEBUG_TRIE_COMPILE_r({
19757 sv_catpvs(sv, "(JUMP)");
19758 Perl_sv_catpvf(aTHX_ sv,
19759 "<S:%" UVuf "/%" IVdf " W:%" UVuf " L:%" UVuf "/%" UVuf " C:%" UVuf "/%" UVuf ">",
19760 (UV)trie->startstate,
19761 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
19762 (UV)trie->wordcount,
19765 (UV)TRIE_CHARCOUNT(trie),
19766 (UV)trie->uniquecharcount
19769 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
19770 sv_catpvs(sv, "[");
19771 (void) put_charclass_bitmap_innards(sv,
19772 ((IS_ANYOF_TRIE(op))
19774 : TRIE_BITMAP(trie)),
19780 sv_catpvs(sv, "]");
19782 } else if (k == CURLY) {
19783 U32 lo = ARG1(o), hi = ARG2(o);
19784 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
19785 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
19786 Perl_sv_catpvf(aTHX_ sv, "{%u,", (unsigned) lo);
19787 if (hi == REG_INFTY)
19788 sv_catpvs(sv, "INFTY");
19790 Perl_sv_catpvf(aTHX_ sv, "%u", (unsigned) hi);
19791 sv_catpvs(sv, "}");
19793 else if (k == WHILEM && o->flags) /* Ordinal/of */
19794 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
19795 else if (k == REF || k == OPEN || k == CLOSE
19796 || k == GROUPP || OP(o)==ACCEPT)
19798 AV *name_list= NULL;
19799 U32 parno= OP(o) == ACCEPT ? (U32)ARG2L(o) : ARG(o);
19800 Perl_sv_catpvf(aTHX_ sv, "%" UVuf, (UV)parno); /* Parenth number */
19801 if ( RXp_PAREN_NAMES(prog) ) {
19802 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
19803 } else if ( pRExC_state ) {
19804 name_list= RExC_paren_name_list;
19807 if ( k != REF || (OP(o) < NREF)) {
19808 SV **name= av_fetch(name_list, parno, 0 );
19810 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19813 SV *sv_dat= MUTABLE_SV(progi->data->data[ parno ]);
19814 I32 *nums=(I32*)SvPVX(sv_dat);
19815 SV **name= av_fetch(name_list, nums[0], 0 );
19818 for ( n=0; n<SvIVX(sv_dat); n++ ) {
19819 Perl_sv_catpvf(aTHX_ sv, "%s%" IVdf,
19820 (n ? "," : ""), (IV)nums[n]);
19822 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19826 if ( k == REF && reginfo) {
19827 U32 n = ARG(o); /* which paren pair */
19828 I32 ln = prog->offs[n].start;
19829 if (prog->lastparen < n || ln == -1 || prog->offs[n].end == -1)
19830 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
19831 else if (ln == prog->offs[n].end)
19832 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
19834 const char *s = reginfo->strbeg + ln;
19835 Perl_sv_catpvf(aTHX_ sv, ": ");
19836 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
19837 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
19840 } else if (k == GOSUB) {
19841 AV *name_list= NULL;
19842 if ( RXp_PAREN_NAMES(prog) ) {
19843 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
19844 } else if ( pRExC_state ) {
19845 name_list= RExC_paren_name_list;
19848 /* Paren and offset */
19849 Perl_sv_catpvf(aTHX_ sv, "%d[%+d:%d]", (int)ARG(o),(int)ARG2L(o),
19850 (int)((o + (int)ARG2L(o)) - progi->program) );
19852 SV **name= av_fetch(name_list, ARG(o), 0 );
19854 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19857 else if (k == LOGICAL)
19858 /* 2: embedded, otherwise 1 */
19859 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
19860 else if (k == ANYOF) {
19861 const U8 flags = ANYOF_FLAGS(o);
19862 bool do_sep = FALSE; /* Do we need to separate various components of
19864 /* Set if there is still an unresolved user-defined property */
19865 SV *unresolved = NULL;
19867 /* Things that are ignored except when the runtime locale is UTF-8 */
19868 SV *only_utf8_locale_invlist = NULL;
19870 /* Code points that don't fit in the bitmap */
19871 SV *nonbitmap_invlist = NULL;
19873 /* And things that aren't in the bitmap, but are small enough to be */
19874 SV* bitmap_range_not_in_bitmap = NULL;
19876 const bool inverted = flags & ANYOF_INVERT;
19878 if (OP(o) == ANYOFL) {
19879 if (ANYOFL_UTF8_LOCALE_REQD(flags)) {
19880 sv_catpvs(sv, "{utf8-locale-reqd}");
19882 if (flags & ANYOFL_FOLD) {
19883 sv_catpvs(sv, "{i}");
19887 /* If there is stuff outside the bitmap, get it */
19888 if (ARG(o) != ANYOF_ONLY_HAS_BITMAP) {
19889 (void) _get_regclass_nonbitmap_data(prog, o, FALSE,
19891 &only_utf8_locale_invlist,
19892 &nonbitmap_invlist);
19893 /* The non-bitmap data may contain stuff that could fit in the
19894 * bitmap. This could come from a user-defined property being
19895 * finally resolved when this call was done; or much more likely
19896 * because there are matches that require UTF-8 to be valid, and so
19897 * aren't in the bitmap. This is teased apart later */
19898 _invlist_intersection(nonbitmap_invlist,
19900 &bitmap_range_not_in_bitmap);
19901 /* Leave just the things that don't fit into the bitmap */
19902 _invlist_subtract(nonbitmap_invlist,
19904 &nonbitmap_invlist);
19907 /* Obey this flag to add all above-the-bitmap code points */
19908 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
19909 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
19910 NUM_ANYOF_CODE_POINTS,
19914 /* Ready to start outputting. First, the initial left bracket */
19915 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
19917 /* Then all the things that could fit in the bitmap */
19918 do_sep = put_charclass_bitmap_innards(sv,
19920 bitmap_range_not_in_bitmap,
19921 only_utf8_locale_invlist,
19924 /* Can't try inverting for a
19925 * better display if there are
19926 * things that haven't been
19928 unresolved != NULL);
19929 SvREFCNT_dec(bitmap_range_not_in_bitmap);
19931 /* If there are user-defined properties which haven't been defined yet,
19932 * output them. If the result is not to be inverted, it is clearest to
19933 * output them in a separate [] from the bitmap range stuff. If the
19934 * result is to be complemented, we have to show everything in one [],
19935 * as the inversion applies to the whole thing. Use {braces} to
19936 * separate them from anything in the bitmap and anything above the
19940 if (! do_sep) { /* If didn't output anything in the bitmap */
19941 sv_catpvs(sv, "^");
19943 sv_catpvs(sv, "{");
19946 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
19948 sv_catsv(sv, unresolved);
19950 sv_catpvs(sv, "}");
19952 do_sep = ! inverted;
19955 /* And, finally, add the above-the-bitmap stuff */
19956 if (nonbitmap_invlist && _invlist_len(nonbitmap_invlist)) {
19959 /* See if truncation size is overridden */
19960 const STRLEN dump_len = (PL_dump_re_max_len > 256)
19961 ? PL_dump_re_max_len
19964 /* This is output in a separate [] */
19966 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
19969 /* And, for easy of understanding, it is shown in the
19970 * uncomplemented form if possible. The one exception being if
19971 * there are unresolved items, where the inversion has to be
19972 * delayed until runtime */
19973 if (inverted && ! unresolved) {
19974 _invlist_invert(nonbitmap_invlist);
19975 _invlist_subtract(nonbitmap_invlist, PL_InBitmap, &nonbitmap_invlist);
19978 contents = invlist_contents(nonbitmap_invlist,
19979 FALSE /* output suitable for catsv */
19982 /* If the output is shorter than the permissible maximum, just do it. */
19983 if (SvCUR(contents) <= dump_len) {
19984 sv_catsv(sv, contents);
19987 const char * contents_string = SvPVX(contents);
19988 STRLEN i = dump_len;
19990 /* Otherwise, start at the permissible max and work back to the
19991 * first break possibility */
19992 while (i > 0 && contents_string[i] != ' ') {
19995 if (i == 0) { /* Fail-safe. Use the max if we couldn't
19996 find a legal break */
20000 sv_catpvn(sv, contents_string, i);
20001 sv_catpvs(sv, "...");
20004 SvREFCNT_dec_NN(contents);
20005 SvREFCNT_dec_NN(nonbitmap_invlist);
20008 /* And finally the matching, closing ']' */
20009 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
20011 SvREFCNT_dec(unresolved);
20013 else if (k == ANYOFM) {
20014 SV * cp_list = get_ANYOFM_contents(o);
20016 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
20017 put_charclass_bitmap_innards(sv, NULL, cp_list, NULL, NULL, TRUE);
20018 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
20020 SvREFCNT_dec(cp_list);
20022 else if (k == POSIXD || k == NPOSIXD) {
20023 U8 index = FLAGS(o) * 2;
20024 if (index < C_ARRAY_LENGTH(anyofs)) {
20025 if (*anyofs[index] != '[') {
20026 sv_catpvs(sv, "[");
20028 sv_catpv(sv, anyofs[index]);
20029 if (*anyofs[index] != '[') {
20030 sv_catpvs(sv, "]");
20034 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
20037 else if (k == BOUND || k == NBOUND) {
20038 /* Must be synced with order of 'bound_type' in regcomp.h */
20039 const char * const bounds[] = {
20040 "", /* Traditional */
20046 assert(FLAGS(o) < C_ARRAY_LENGTH(bounds));
20047 sv_catpv(sv, bounds[FLAGS(o)]);
20049 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
20050 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
20051 else if (OP(o) == SBOL)
20052 Perl_sv_catpvf(aTHX_ sv, " /%s/", o->flags ? "\\A" : "^");
20054 /* add on the verb argument if there is one */
20055 if ( ( k == VERB || OP(o) == ACCEPT || OP(o) == OPFAIL ) && o->flags) {
20057 Perl_sv_catpvf(aTHX_ sv, ":%" SVf,
20058 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
20060 sv_catpvs(sv, ":NULL");
20063 PERL_UNUSED_CONTEXT;
20064 PERL_UNUSED_ARG(sv);
20065 PERL_UNUSED_ARG(o);
20066 PERL_UNUSED_ARG(prog);
20067 PERL_UNUSED_ARG(reginfo);
20068 PERL_UNUSED_ARG(pRExC_state);
20069 #endif /* DEBUGGING */
20075 Perl_re_intuit_string(pTHX_ REGEXP * const r)
20076 { /* Assume that RE_INTUIT is set */
20077 struct regexp *const prog = ReANY(r);
20078 GET_RE_DEBUG_FLAGS_DECL;
20080 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
20081 PERL_UNUSED_CONTEXT;
20085 const char * const s = SvPV_nolen_const(RX_UTF8(r)
20086 ? prog->check_utf8 : prog->check_substr);
20088 if (!PL_colorset) reginitcolors();
20089 Perl_re_printf( aTHX_
20090 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
20092 RX_UTF8(r) ? "utf8 " : "",
20093 PL_colors[5],PL_colors[0],
20096 (strlen(s) > PL_dump_re_max_len ? "..." : ""));
20099 /* use UTF8 check substring if regexp pattern itself is in UTF8 */
20100 return RX_UTF8(r) ? prog->check_utf8 : prog->check_substr;
20106 handles refcounting and freeing the perl core regexp structure. When
20107 it is necessary to actually free the structure the first thing it
20108 does is call the 'free' method of the regexp_engine associated to
20109 the regexp, allowing the handling of the void *pprivate; member
20110 first. (This routine is not overridable by extensions, which is why
20111 the extensions free is called first.)
20113 See regdupe and regdupe_internal if you change anything here.
20115 #ifndef PERL_IN_XSUB_RE
20117 Perl_pregfree(pTHX_ REGEXP *r)
20123 Perl_pregfree2(pTHX_ REGEXP *rx)
20125 struct regexp *const r = ReANY(rx);
20126 GET_RE_DEBUG_FLAGS_DECL;
20128 PERL_ARGS_ASSERT_PREGFREE2;
20130 if (r->mother_re) {
20131 ReREFCNT_dec(r->mother_re);
20133 CALLREGFREE_PVT(rx); /* free the private data */
20134 SvREFCNT_dec(RXp_PAREN_NAMES(r));
20138 for (i = 0; i < 2; i++) {
20139 SvREFCNT_dec(r->substrs->data[i].substr);
20140 SvREFCNT_dec(r->substrs->data[i].utf8_substr);
20142 Safefree(r->substrs);
20144 RX_MATCH_COPY_FREE(rx);
20145 #ifdef PERL_ANY_COW
20146 SvREFCNT_dec(r->saved_copy);
20149 SvREFCNT_dec(r->qr_anoncv);
20150 if (r->recurse_locinput)
20151 Safefree(r->recurse_locinput);
20157 Copy ssv to dsv, both of which should of type SVt_REGEXP or SVt_PVLV,
20158 except that dsv will be created if NULL.
20160 This function is used in two main ways. First to implement
20161 $r = qr/....; $s = $$r;
20163 Secondly, it is used as a hacky workaround to the structural issue of
20165 being stored in the regexp structure which is in turn stored in
20166 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
20167 could be PL_curpm in multiple contexts, and could require multiple
20168 result sets being associated with the pattern simultaneously, such
20169 as when doing a recursive match with (??{$qr})
20171 The solution is to make a lightweight copy of the regexp structure
20172 when a qr// is returned from the code executed by (??{$qr}) this
20173 lightweight copy doesn't actually own any of its data except for
20174 the starp/end and the actual regexp structure itself.
20180 Perl_reg_temp_copy(pTHX_ REGEXP *dsv, REGEXP *ssv)
20182 struct regexp *drx;
20183 struct regexp *const srx = ReANY(ssv);
20184 const bool islv = dsv && SvTYPE(dsv) == SVt_PVLV;
20186 PERL_ARGS_ASSERT_REG_TEMP_COPY;
20189 dsv = (REGEXP*) newSV_type(SVt_REGEXP);
20191 SvOK_off((SV *)dsv);
20193 /* For PVLVs, the head (sv_any) points to an XPVLV, while
20194 * the LV's xpvlenu_rx will point to a regexp body, which
20195 * we allocate here */
20196 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
20197 assert(!SvPVX(dsv));
20198 ((XPV*)SvANY(dsv))->xpv_len_u.xpvlenu_rx = temp->sv_any;
20199 temp->sv_any = NULL;
20200 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
20201 SvREFCNT_dec_NN(temp);
20202 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
20203 ing below will not set it. */
20204 SvCUR_set(dsv, SvCUR(ssv));
20207 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
20208 sv_force_normal(sv) is called. */
20212 SvFLAGS(dsv) |= SvFLAGS(ssv) & (SVf_POK|SVp_POK|SVf_UTF8);
20213 SvPV_set(dsv, RX_WRAPPED(ssv));
20214 /* We share the same string buffer as the original regexp, on which we
20215 hold a reference count, incremented when mother_re is set below.
20216 The string pointer is copied here, being part of the regexp struct.
20218 memcpy(&(drx->xpv_cur), &(srx->xpv_cur),
20219 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
20223 const I32 npar = srx->nparens+1;
20224 Newx(drx->offs, npar, regexp_paren_pair);
20225 Copy(srx->offs, drx->offs, npar, regexp_paren_pair);
20227 if (srx->substrs) {
20229 Newx(drx->substrs, 1, struct reg_substr_data);
20230 StructCopy(srx->substrs, drx->substrs, struct reg_substr_data);
20232 for (i = 0; i < 2; i++) {
20233 SvREFCNT_inc_void(drx->substrs->data[i].substr);
20234 SvREFCNT_inc_void(drx->substrs->data[i].utf8_substr);
20237 /* check_substr and check_utf8, if non-NULL, point to either their
20238 anchored or float namesakes, and don't hold a second reference. */
20240 RX_MATCH_COPIED_off(dsv);
20241 #ifdef PERL_ANY_COW
20242 drx->saved_copy = NULL;
20244 drx->mother_re = ReREFCNT_inc(srx->mother_re ? srx->mother_re : ssv);
20245 SvREFCNT_inc_void(drx->qr_anoncv);
20246 if (srx->recurse_locinput)
20247 Newx(drx->recurse_locinput,srx->nparens + 1,char *);
20254 /* regfree_internal()
20256 Free the private data in a regexp. This is overloadable by
20257 extensions. Perl takes care of the regexp structure in pregfree(),
20258 this covers the *pprivate pointer which technically perl doesn't
20259 know about, however of course we have to handle the
20260 regexp_internal structure when no extension is in use.
20262 Note this is called before freeing anything in the regexp
20267 Perl_regfree_internal(pTHX_ REGEXP * const rx)
20269 struct regexp *const r = ReANY(rx);
20270 RXi_GET_DECL(r,ri);
20271 GET_RE_DEBUG_FLAGS_DECL;
20273 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
20279 SV *dsv= sv_newmortal();
20280 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
20281 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), PL_dump_re_max_len);
20282 Perl_re_printf( aTHX_ "%sFreeing REx:%s %s\n",
20283 PL_colors[4],PL_colors[5],s);
20286 #ifdef RE_TRACK_PATTERN_OFFSETS
20288 Safefree(ri->u.offsets); /* 20010421 MJD */
20290 if (ri->code_blocks)
20291 S_free_codeblocks(aTHX_ ri->code_blocks);
20294 int n = ri->data->count;
20297 /* If you add a ->what type here, update the comment in regcomp.h */
20298 switch (ri->data->what[n]) {
20304 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
20307 Safefree(ri->data->data[n]);
20313 { /* Aho Corasick add-on structure for a trie node.
20314 Used in stclass optimization only */
20316 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
20317 #ifdef USE_ITHREADS
20321 refcount = --aho->refcount;
20324 PerlMemShared_free(aho->states);
20325 PerlMemShared_free(aho->fail);
20326 /* do this last!!!! */
20327 PerlMemShared_free(ri->data->data[n]);
20328 /* we should only ever get called once, so
20329 * assert as much, and also guard the free
20330 * which /might/ happen twice. At the least
20331 * it will make code anlyzers happy and it
20332 * doesn't cost much. - Yves */
20333 assert(ri->regstclass);
20334 if (ri->regstclass) {
20335 PerlMemShared_free(ri->regstclass);
20336 ri->regstclass = 0;
20343 /* trie structure. */
20345 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
20346 #ifdef USE_ITHREADS
20350 refcount = --trie->refcount;
20353 PerlMemShared_free(trie->charmap);
20354 PerlMemShared_free(trie->states);
20355 PerlMemShared_free(trie->trans);
20357 PerlMemShared_free(trie->bitmap);
20359 PerlMemShared_free(trie->jump);
20360 PerlMemShared_free(trie->wordinfo);
20361 /* do this last!!!! */
20362 PerlMemShared_free(ri->data->data[n]);
20367 Perl_croak(aTHX_ "panic: regfree data code '%c'",
20368 ri->data->what[n]);
20371 Safefree(ri->data->what);
20372 Safefree(ri->data);
20378 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
20379 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
20380 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
20383 re_dup_guts - duplicate a regexp.
20385 This routine is expected to clone a given regexp structure. It is only
20386 compiled under USE_ITHREADS.
20388 After all of the core data stored in struct regexp is duplicated
20389 the regexp_engine.dupe method is used to copy any private data
20390 stored in the *pprivate pointer. This allows extensions to handle
20391 any duplication it needs to do.
20393 See pregfree() and regfree_internal() if you change anything here.
20395 #if defined(USE_ITHREADS)
20396 #ifndef PERL_IN_XSUB_RE
20398 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
20402 const struct regexp *r = ReANY(sstr);
20403 struct regexp *ret = ReANY(dstr);
20405 PERL_ARGS_ASSERT_RE_DUP_GUTS;
20407 npar = r->nparens+1;
20408 Newx(ret->offs, npar, regexp_paren_pair);
20409 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
20411 if (ret->substrs) {
20412 /* Do it this way to avoid reading from *r after the StructCopy().
20413 That way, if any of the sv_dup_inc()s dislodge *r from the L1
20414 cache, it doesn't matter. */
20416 const bool anchored = r->check_substr
20417 ? r->check_substr == r->substrs->data[0].substr
20418 : r->check_utf8 == r->substrs->data[0].utf8_substr;
20419 Newx(ret->substrs, 1, struct reg_substr_data);
20420 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
20422 for (i = 0; i < 2; i++) {
20423 ret->substrs->data[i].substr =
20424 sv_dup_inc(ret->substrs->data[i].substr, param);
20425 ret->substrs->data[i].utf8_substr =
20426 sv_dup_inc(ret->substrs->data[i].utf8_substr, param);
20429 /* check_substr and check_utf8, if non-NULL, point to either their
20430 anchored or float namesakes, and don't hold a second reference. */
20432 if (ret->check_substr) {
20434 assert(r->check_utf8 == r->substrs->data[0].utf8_substr);
20436 ret->check_substr = ret->substrs->data[0].substr;
20437 ret->check_utf8 = ret->substrs->data[0].utf8_substr;
20439 assert(r->check_substr == r->substrs->data[1].substr);
20440 assert(r->check_utf8 == r->substrs->data[1].utf8_substr);
20442 ret->check_substr = ret->substrs->data[1].substr;
20443 ret->check_utf8 = ret->substrs->data[1].utf8_substr;
20445 } else if (ret->check_utf8) {
20447 ret->check_utf8 = ret->substrs->data[0].utf8_substr;
20449 ret->check_utf8 = ret->substrs->data[1].utf8_substr;
20454 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
20455 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
20456 if (r->recurse_locinput)
20457 Newx(ret->recurse_locinput,r->nparens + 1,char *);
20460 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
20462 if (RX_MATCH_COPIED(dstr))
20463 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
20465 ret->subbeg = NULL;
20466 #ifdef PERL_ANY_COW
20467 ret->saved_copy = NULL;
20470 /* Whether mother_re be set or no, we need to copy the string. We
20471 cannot refrain from copying it when the storage points directly to
20472 our mother regexp, because that's
20473 1: a buffer in a different thread
20474 2: something we no longer hold a reference on
20475 so we need to copy it locally. */
20476 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED_const(sstr), SvCUR(sstr)+1);
20477 ret->mother_re = NULL;
20479 #endif /* PERL_IN_XSUB_RE */
20484 This is the internal complement to regdupe() which is used to copy
20485 the structure pointed to by the *pprivate pointer in the regexp.
20486 This is the core version of the extension overridable cloning hook.
20487 The regexp structure being duplicated will be copied by perl prior
20488 to this and will be provided as the regexp *r argument, however
20489 with the /old/ structures pprivate pointer value. Thus this routine
20490 may override any copying normally done by perl.
20492 It returns a pointer to the new regexp_internal structure.
20496 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
20499 struct regexp *const r = ReANY(rx);
20500 regexp_internal *reti;
20502 RXi_GET_DECL(r,ri);
20504 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
20508 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
20509 char, regexp_internal);
20510 Copy(ri->program, reti->program, len+1, regnode);
20513 if (ri->code_blocks) {
20515 Newx(reti->code_blocks, 1, struct reg_code_blocks);
20516 Newx(reti->code_blocks->cb, ri->code_blocks->count,
20517 struct reg_code_block);
20518 Copy(ri->code_blocks->cb, reti->code_blocks->cb,
20519 ri->code_blocks->count, struct reg_code_block);
20520 for (n = 0; n < ri->code_blocks->count; n++)
20521 reti->code_blocks->cb[n].src_regex = (REGEXP*)
20522 sv_dup_inc((SV*)(ri->code_blocks->cb[n].src_regex), param);
20523 reti->code_blocks->count = ri->code_blocks->count;
20524 reti->code_blocks->refcnt = 1;
20527 reti->code_blocks = NULL;
20529 reti->regstclass = NULL;
20532 struct reg_data *d;
20533 const int count = ri->data->count;
20536 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
20537 char, struct reg_data);
20538 Newx(d->what, count, U8);
20541 for (i = 0; i < count; i++) {
20542 d->what[i] = ri->data->what[i];
20543 switch (d->what[i]) {
20544 /* see also regcomp.h and regfree_internal() */
20545 case 'a': /* actually an AV, but the dup function is identical.
20546 values seem to be "plain sv's" generally. */
20547 case 'r': /* a compiled regex (but still just another SV) */
20548 case 's': /* an RV (currently only used for an RV to an AV by the ANYOF code)
20549 this use case should go away, the code could have used
20550 'a' instead - see S_set_ANYOF_arg() for array contents. */
20551 case 'S': /* actually an SV, but the dup function is identical. */
20552 case 'u': /* actually an HV, but the dup function is identical.
20553 values are "plain sv's" */
20554 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
20557 /* Synthetic Start Class - "Fake" charclass we generate to optimize
20558 * patterns which could start with several different things. Pre-TRIE
20559 * this was more important than it is now, however this still helps
20560 * in some places, for instance /x?a+/ might produce a SSC equivalent
20561 * to [xa]. This is used by Perl_re_intuit_start() and S_find_byclass()
20564 /* This is cheating. */
20565 Newx(d->data[i], 1, regnode_ssc);
20566 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
20567 reti->regstclass = (regnode*)d->data[i];
20570 /* AHO-CORASICK fail table */
20571 /* Trie stclasses are readonly and can thus be shared
20572 * without duplication. We free the stclass in pregfree
20573 * when the corresponding reg_ac_data struct is freed.
20575 reti->regstclass= ri->regstclass;
20578 /* TRIE transition table */
20580 ((reg_trie_data*)ri->data->data[i])->refcount++;
20583 case 'l': /* (?{...}) or (??{ ... }) code (cb->block) */
20584 case 'L': /* same when RExC_pm_flags & PMf_HAS_CV and code
20585 is not from another regexp */
20586 d->data[i] = ri->data->data[i];
20589 Perl_croak(aTHX_ "panic: re_dup_guts unknown data code '%c'",
20590 ri->data->what[i]);
20599 reti->name_list_idx = ri->name_list_idx;
20601 #ifdef RE_TRACK_PATTERN_OFFSETS
20602 if (ri->u.offsets) {
20603 Newx(reti->u.offsets, 2*len+1, U32);
20604 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
20607 SetProgLen(reti,len);
20610 return (void*)reti;
20613 #endif /* USE_ITHREADS */
20615 #ifndef PERL_IN_XSUB_RE
20618 - regnext - dig the "next" pointer out of a node
20621 Perl_regnext(pTHX_ regnode *p)
20628 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
20629 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
20630 (int)OP(p), (int)REGNODE_MAX);
20633 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
20642 S_re_croak2(pTHX_ bool utf8, const char* pat1,const char* pat2,...)
20645 STRLEN l1 = strlen(pat1);
20646 STRLEN l2 = strlen(pat2);
20649 const char *message;
20651 PERL_ARGS_ASSERT_RE_CROAK2;
20657 Copy(pat1, buf, l1 , char);
20658 Copy(pat2, buf + l1, l2 , char);
20659 buf[l1 + l2] = '\n';
20660 buf[l1 + l2 + 1] = '\0';
20661 va_start(args, pat2);
20662 msv = vmess(buf, &args);
20664 message = SvPV_const(msv,l1);
20667 Copy(message, buf, l1 , char);
20668 /* l1-1 to avoid \n */
20669 Perl_croak(aTHX_ "%" UTF8f, UTF8fARG(utf8, l1-1, buf));
20672 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
20674 #ifndef PERL_IN_XSUB_RE
20676 Perl_save_re_context(pTHX)
20681 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
20684 const REGEXP * const rx = PM_GETRE(PL_curpm);
20686 nparens = RX_NPARENS(rx);
20689 /* RT #124109. This is a complete hack; in the SWASHNEW case we know
20690 * that PL_curpm will be null, but that utf8.pm and the modules it
20691 * loads will only use $1..$3.
20692 * The t/porting/re_context.t test file checks this assumption.
20697 for (i = 1; i <= nparens; i++) {
20698 char digits[TYPE_CHARS(long)];
20699 const STRLEN len = my_snprintf(digits, sizeof(digits),
20701 GV *const *const gvp
20702 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
20705 GV * const gv = *gvp;
20706 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
20716 S_put_code_point(pTHX_ SV *sv, UV c)
20718 PERL_ARGS_ASSERT_PUT_CODE_POINT;
20721 Perl_sv_catpvf(aTHX_ sv, "\\x{%04" UVXf "}", c);
20723 else if (isPRINT(c)) {
20724 const char string = (char) c;
20726 /* We use {phrase} as metanotation in the class, so also escape literal
20728 if (isBACKSLASHED_PUNCT(c) || c == '{' || c == '}')
20729 sv_catpvs(sv, "\\");
20730 sv_catpvn(sv, &string, 1);
20732 else if (isMNEMONIC_CNTRL(c)) {
20733 Perl_sv_catpvf(aTHX_ sv, "%s", cntrl_to_mnemonic((U8) c));
20736 Perl_sv_catpvf(aTHX_ sv, "\\x%02X", (U8) c);
20740 #define MAX_PRINT_A MAX_PRINT_A_FOR_USE_ONLY_BY_REGCOMP_DOT_C
20743 S_put_range(pTHX_ SV *sv, UV start, const UV end, const bool allow_literals)
20745 /* Appends to 'sv' a displayable version of the range of code points from
20746 * 'start' to 'end'. Mnemonics (like '\r') are used for the few controls
20747 * that have them, when they occur at the beginning or end of the range.
20748 * It uses hex to output the remaining code points, unless 'allow_literals'
20749 * is true, in which case the printable ASCII ones are output as-is (though
20750 * some of these will be escaped by put_code_point()).
20752 * NOTE: This is designed only for printing ranges of code points that fit
20753 * inside an ANYOF bitmap. Higher code points are simply suppressed
20756 const unsigned int min_range_count = 3;
20758 assert(start <= end);
20760 PERL_ARGS_ASSERT_PUT_RANGE;
20762 while (start <= end) {
20764 const char * format;
20766 if (end - start < min_range_count) {
20768 /* Output chars individually when they occur in short ranges */
20769 for (; start <= end; start++) {
20770 put_code_point(sv, start);
20775 /* If permitted by the input options, and there is a possibility that
20776 * this range contains a printable literal, look to see if there is
20778 if (allow_literals && start <= MAX_PRINT_A) {
20780 /* If the character at the beginning of the range isn't an ASCII
20781 * printable, effectively split the range into two parts:
20782 * 1) the portion before the first such printable,
20784 * and output them separately. */
20785 if (! isPRINT_A(start)) {
20786 UV temp_end = start + 1;
20788 /* There is no point looking beyond the final possible
20789 * printable, in MAX_PRINT_A */
20790 UV max = MIN(end, MAX_PRINT_A);
20792 while (temp_end <= max && ! isPRINT_A(temp_end)) {
20796 /* Here, temp_end points to one beyond the first printable if
20797 * found, or to one beyond 'max' if not. If none found, make
20798 * sure that we use the entire range */
20799 if (temp_end > MAX_PRINT_A) {
20800 temp_end = end + 1;
20803 /* Output the first part of the split range: the part that
20804 * doesn't have printables, with the parameter set to not look
20805 * for literals (otherwise we would infinitely recurse) */
20806 put_range(sv, start, temp_end - 1, FALSE);
20808 /* The 2nd part of the range (if any) starts here. */
20811 /* We do a continue, instead of dropping down, because even if
20812 * the 2nd part is non-empty, it could be so short that we want
20813 * to output it as individual characters, as tested for at the
20814 * top of this loop. */
20818 /* Here, 'start' is a printable ASCII. If it is an alphanumeric,
20819 * output a sub-range of just the digits or letters, then process
20820 * the remaining portion as usual. */
20821 if (isALPHANUMERIC_A(start)) {
20822 UV mask = (isDIGIT_A(start))
20827 UV temp_end = start + 1;
20829 /* Find the end of the sub-range that includes just the
20830 * characters in the same class as the first character in it */
20831 while (temp_end <= end && _generic_isCC_A(temp_end, mask)) {
20836 /* For short ranges, don't duplicate the code above to output
20837 * them; just call recursively */
20838 if (temp_end - start < min_range_count) {
20839 put_range(sv, start, temp_end, FALSE);
20841 else { /* Output as a range */
20842 put_code_point(sv, start);
20843 sv_catpvs(sv, "-");
20844 put_code_point(sv, temp_end);
20846 start = temp_end + 1;
20850 /* We output any other printables as individual characters */
20851 if (isPUNCT_A(start) || isSPACE_A(start)) {
20852 while (start <= end && (isPUNCT_A(start)
20853 || isSPACE_A(start)))
20855 put_code_point(sv, start);
20860 } /* End of looking for literals */
20862 /* Here is not to output as a literal. Some control characters have
20863 * mnemonic names. Split off any of those at the beginning and end of
20864 * the range to print mnemonically. It isn't possible for many of
20865 * these to be in a row, so this won't overwhelm with output */
20867 && (isMNEMONIC_CNTRL(start) || isMNEMONIC_CNTRL(end)))
20869 while (isMNEMONIC_CNTRL(start) && start <= end) {
20870 put_code_point(sv, start);
20874 /* If this didn't take care of the whole range ... */
20875 if (start <= end) {
20877 /* Look backwards from the end to find the final non-mnemonic
20880 while (isMNEMONIC_CNTRL(temp_end)) {
20884 /* And separately output the interior range that doesn't start
20885 * or end with mnemonics */
20886 put_range(sv, start, temp_end, FALSE);
20888 /* Then output the mnemonic trailing controls */
20889 start = temp_end + 1;
20890 while (start <= end) {
20891 put_code_point(sv, start);
20898 /* As a final resort, output the range or subrange as hex. */
20900 this_end = (end < NUM_ANYOF_CODE_POINTS)
20902 : NUM_ANYOF_CODE_POINTS - 1;
20903 #if NUM_ANYOF_CODE_POINTS > 256
20904 format = (this_end < 256)
20905 ? "\\x%02" UVXf "-\\x%02" UVXf
20906 : "\\x{%04" UVXf "}-\\x{%04" UVXf "}";
20908 format = "\\x%02" UVXf "-\\x%02" UVXf;
20910 GCC_DIAG_IGNORE_STMT(-Wformat-nonliteral);
20911 Perl_sv_catpvf(aTHX_ sv, format, start, this_end);
20912 GCC_DIAG_RESTORE_STMT;
20918 S_put_charclass_bitmap_innards_invlist(pTHX_ SV *sv, SV* invlist)
20920 /* Concatenate onto the PV in 'sv' a displayable form of the inversion list
20924 bool allow_literals = TRUE;
20926 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_INVLIST;
20928 /* Generally, it is more readable if printable characters are output as
20929 * literals, but if a range (nearly) spans all of them, it's best to output
20930 * it as a single range. This code will use a single range if all but 2
20931 * ASCII printables are in it */
20932 invlist_iterinit(invlist);
20933 while (invlist_iternext(invlist, &start, &end)) {
20935 /* If the range starts beyond the final printable, it doesn't have any
20937 if (start > MAX_PRINT_A) {
20941 /* In both ASCII and EBCDIC, a SPACE is the lowest printable. To span
20942 * all but two, the range must start and end no later than 2 from
20944 if (start < ' ' + 2 && end > MAX_PRINT_A - 2) {
20945 if (end > MAX_PRINT_A) {
20951 if (end - start >= MAX_PRINT_A - ' ' - 2) {
20952 allow_literals = FALSE;
20957 invlist_iterfinish(invlist);
20959 /* Here we have figured things out. Output each range */
20960 invlist_iterinit(invlist);
20961 while (invlist_iternext(invlist, &start, &end)) {
20962 if (start >= NUM_ANYOF_CODE_POINTS) {
20965 put_range(sv, start, end, allow_literals);
20967 invlist_iterfinish(invlist);
20973 S_put_charclass_bitmap_innards_common(pTHX_
20974 SV* invlist, /* The bitmap */
20975 SV* posixes, /* Under /l, things like [:word:], \S */
20976 SV* only_utf8, /* Under /d, matches iff the target is UTF-8 */
20977 SV* not_utf8, /* /d, matches iff the target isn't UTF-8 */
20978 SV* only_utf8_locale, /* Under /l, matches if the locale is UTF-8 */
20979 const bool invert /* Is the result to be inverted? */
20982 /* Create and return an SV containing a displayable version of the bitmap
20983 * and associated information determined by the input parameters. If the
20984 * output would have been only the inversion indicator '^', NULL is instead
20989 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_COMMON;
20992 output = newSVpvs("^");
20995 output = newSVpvs("");
20998 /* First, the code points in the bitmap that are unconditionally there */
20999 put_charclass_bitmap_innards_invlist(output, invlist);
21001 /* Traditionally, these have been placed after the main code points */
21003 sv_catsv(output, posixes);
21006 if (only_utf8 && _invlist_len(only_utf8)) {
21007 Perl_sv_catpvf(aTHX_ output, "%s{utf8}%s", PL_colors[1], PL_colors[0]);
21008 put_charclass_bitmap_innards_invlist(output, only_utf8);
21011 if (not_utf8 && _invlist_len(not_utf8)) {
21012 Perl_sv_catpvf(aTHX_ output, "%s{not utf8}%s", PL_colors[1], PL_colors[0]);
21013 put_charclass_bitmap_innards_invlist(output, not_utf8);
21016 if (only_utf8_locale && _invlist_len(only_utf8_locale)) {
21017 Perl_sv_catpvf(aTHX_ output, "%s{utf8 locale}%s", PL_colors[1], PL_colors[0]);
21018 put_charclass_bitmap_innards_invlist(output, only_utf8_locale);
21020 /* This is the only list in this routine that can legally contain code
21021 * points outside the bitmap range. The call just above to
21022 * 'put_charclass_bitmap_innards_invlist' will simply suppress them, so
21023 * output them here. There's about a half-dozen possible, and none in
21024 * contiguous ranges longer than 2 */
21025 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
21027 SV* above_bitmap = NULL;
21029 _invlist_subtract(only_utf8_locale, PL_InBitmap, &above_bitmap);
21031 invlist_iterinit(above_bitmap);
21032 while (invlist_iternext(above_bitmap, &start, &end)) {
21035 for (i = start; i <= end; i++) {
21036 put_code_point(output, i);
21039 invlist_iterfinish(above_bitmap);
21040 SvREFCNT_dec_NN(above_bitmap);
21044 if (invert && SvCUR(output) == 1) {
21052 S_put_charclass_bitmap_innards(pTHX_ SV *sv,
21054 SV *nonbitmap_invlist,
21055 SV *only_utf8_locale_invlist,
21056 const regnode * const node,
21057 const bool force_as_is_display)
21059 /* Appends to 'sv' a displayable version of the innards of the bracketed
21060 * character class defined by the other arguments:
21061 * 'bitmap' points to the bitmap, or NULL if to ignore that.
21062 * 'nonbitmap_invlist' is an inversion list of the code points that are in
21063 * the bitmap range, but for some reason aren't in the bitmap; NULL if
21064 * none. The reasons for this could be that they require some
21065 * condition such as the target string being or not being in UTF-8
21066 * (under /d), or because they came from a user-defined property that
21067 * was not resolved at the time of the regex compilation (under /u)
21068 * 'only_utf8_locale_invlist' is an inversion list of the code points that
21069 * are valid only if the runtime locale is a UTF-8 one; NULL if none
21070 * 'node' is the regex pattern ANYOF node. It is needed only when the
21071 * above two parameters are not null, and is passed so that this
21072 * routine can tease apart the various reasons for them.
21073 * 'force_as_is_display' is TRUE if this routine should definitely NOT try
21074 * to invert things to see if that leads to a cleaner display. If
21075 * FALSE, this routine is free to use its judgment about doing this.
21077 * It returns TRUE if there was actually something output. (It may be that
21078 * the bitmap, etc is empty.)
21080 * When called for outputting the bitmap of a non-ANYOF node, just pass the
21081 * bitmap, with the succeeding parameters set to NULL, and the final one to
21085 /* In general, it tries to display the 'cleanest' representation of the
21086 * innards, choosing whether to display them inverted or not, regardless of
21087 * whether the class itself is to be inverted. However, there are some
21088 * cases where it can't try inverting, as what actually matches isn't known
21089 * until runtime, and hence the inversion isn't either. */
21090 bool inverting_allowed = ! force_as_is_display;
21093 STRLEN orig_sv_cur = SvCUR(sv);
21095 SV* invlist; /* Inversion list we accumulate of code points that
21096 are unconditionally matched */
21097 SV* only_utf8 = NULL; /* Under /d, list of matches iff the target is
21099 SV* not_utf8 = NULL; /* /d, list of matches iff the target isn't UTF-8
21101 SV* posixes = NULL; /* Under /l, string of things like [:word:], \D */
21102 SV* only_utf8_locale = NULL; /* Under /l, list of matches if the locale
21105 SV* as_is_display; /* The output string when we take the inputs
21107 SV* inverted_display; /* The output string when we invert the inputs */
21109 U8 flags = (node) ? ANYOF_FLAGS(node) : 0;
21111 bool invert = cBOOL(flags & ANYOF_INVERT); /* Is the input to be inverted
21113 /* We are biased in favor of displaying things without them being inverted,
21114 * as that is generally easier to understand */
21115 const int bias = 5;
21117 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS;
21119 /* Start off with whatever code points are passed in. (We clone, so we
21120 * don't change the caller's list) */
21121 if (nonbitmap_invlist) {
21122 assert(invlist_highest(nonbitmap_invlist) < NUM_ANYOF_CODE_POINTS);
21123 invlist = invlist_clone(nonbitmap_invlist);
21125 else { /* Worst case size is every other code point is matched */
21126 invlist = _new_invlist(NUM_ANYOF_CODE_POINTS / 2);
21130 if (OP(node) == ANYOFD) {
21132 /* This flag indicates that the code points below 0x100 in the
21133 * nonbitmap list are precisely the ones that match only when the
21134 * target is UTF-8 (they should all be non-ASCII). */
21135 if (flags & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)
21137 _invlist_intersection(invlist, PL_UpperLatin1, &only_utf8);
21138 _invlist_subtract(invlist, only_utf8, &invlist);
21141 /* And this flag for matching all non-ASCII 0xFF and below */
21142 if (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)
21144 not_utf8 = invlist_clone(PL_UpperLatin1);
21147 else if (OP(node) == ANYOFL) {
21149 /* If either of these flags are set, what matches isn't
21150 * determinable except during execution, so don't know enough here
21152 if (flags & (ANYOFL_FOLD|ANYOF_MATCHES_POSIXL)) {
21153 inverting_allowed = FALSE;
21156 /* What the posix classes match also varies at runtime, so these
21157 * will be output symbolically. */
21158 if (ANYOF_POSIXL_TEST_ANY_SET(node)) {
21161 posixes = newSVpvs("");
21162 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
21163 if (ANYOF_POSIXL_TEST(node,i)) {
21164 sv_catpv(posixes, anyofs[i]);
21171 /* Accumulate the bit map into the unconditional match list */
21173 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
21174 if (BITMAP_TEST(bitmap, i)) {
21177 i < NUM_ANYOF_CODE_POINTS && BITMAP_TEST(bitmap, i);
21180 invlist = _add_range_to_invlist(invlist, start, i-1);
21185 /* Make sure that the conditional match lists don't have anything in them
21186 * that match unconditionally; otherwise the output is quite confusing.
21187 * This could happen if the code that populates these misses some
21190 _invlist_subtract(only_utf8, invlist, &only_utf8);
21193 _invlist_subtract(not_utf8, invlist, ¬_utf8);
21196 if (only_utf8_locale_invlist) {
21198 /* Since this list is passed in, we have to make a copy before
21200 only_utf8_locale = invlist_clone(only_utf8_locale_invlist);
21202 _invlist_subtract(only_utf8_locale, invlist, &only_utf8_locale);
21204 /* And, it can get really weird for us to try outputting an inverted
21205 * form of this list when it has things above the bitmap, so don't even
21207 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
21208 inverting_allowed = FALSE;
21212 /* Calculate what the output would be if we take the input as-is */
21213 as_is_display = put_charclass_bitmap_innards_common(invlist,
21220 /* If have to take the output as-is, just do that */
21221 if (! inverting_allowed) {
21222 if (as_is_display) {
21223 sv_catsv(sv, as_is_display);
21224 SvREFCNT_dec_NN(as_is_display);
21227 else { /* But otherwise, create the output again on the inverted input, and
21228 use whichever version is shorter */
21230 int inverted_bias, as_is_bias;
21232 /* We will apply our bias to whichever of the the results doesn't have
21242 inverted_bias = bias;
21245 /* Now invert each of the lists that contribute to the output,
21246 * excluding from the result things outside the possible range */
21248 /* For the unconditional inversion list, we have to add in all the
21249 * conditional code points, so that when inverted, they will be gone
21251 _invlist_union(only_utf8, invlist, &invlist);
21252 _invlist_union(not_utf8, invlist, &invlist);
21253 _invlist_union(only_utf8_locale, invlist, &invlist);
21254 _invlist_invert(invlist);
21255 _invlist_intersection(invlist, PL_InBitmap, &invlist);
21258 _invlist_invert(only_utf8);
21259 _invlist_intersection(only_utf8, PL_UpperLatin1, &only_utf8);
21261 else if (not_utf8) {
21263 /* If a code point matches iff the target string is not in UTF-8,
21264 * then complementing the result has it not match iff not in UTF-8,
21265 * which is the same thing as matching iff it is UTF-8. */
21266 only_utf8 = not_utf8;
21270 if (only_utf8_locale) {
21271 _invlist_invert(only_utf8_locale);
21272 _invlist_intersection(only_utf8_locale,
21274 &only_utf8_locale);
21277 inverted_display = put_charclass_bitmap_innards_common(
21282 only_utf8_locale, invert);
21284 /* Use the shortest representation, taking into account our bias
21285 * against showing it inverted */
21286 if ( inverted_display
21287 && ( ! as_is_display
21288 || ( SvCUR(inverted_display) + inverted_bias
21289 < SvCUR(as_is_display) + as_is_bias)))
21291 sv_catsv(sv, inverted_display);
21293 else if (as_is_display) {
21294 sv_catsv(sv, as_is_display);
21297 SvREFCNT_dec(as_is_display);
21298 SvREFCNT_dec(inverted_display);
21301 SvREFCNT_dec_NN(invlist);
21302 SvREFCNT_dec(only_utf8);
21303 SvREFCNT_dec(not_utf8);
21304 SvREFCNT_dec(posixes);
21305 SvREFCNT_dec(only_utf8_locale);
21307 return SvCUR(sv) > orig_sv_cur;
21310 #define CLEAR_OPTSTART \
21311 if (optstart) STMT_START { \
21312 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ \
21313 " (%" IVdf " nodes)\n", (IV)(node - optstart))); \
21317 #define DUMPUNTIL(b,e) \
21319 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
21321 STATIC const regnode *
21322 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
21323 const regnode *last, const regnode *plast,
21324 SV* sv, I32 indent, U32 depth)
21326 U8 op = PSEUDO; /* Arbitrary non-END op. */
21327 const regnode *next;
21328 const regnode *optstart= NULL;
21330 RXi_GET_DECL(r,ri);
21331 GET_RE_DEBUG_FLAGS_DECL;
21333 PERL_ARGS_ASSERT_DUMPUNTIL;
21335 #ifdef DEBUG_DUMPUNTIL
21336 Perl_re_printf( aTHX_ "--- %d : %d - %d - %d\n",indent,node-start,
21337 last ? last-start : 0,plast ? plast-start : 0);
21340 if (plast && plast < last)
21343 while (PL_regkind[op] != END && (!last || node < last)) {
21345 /* While that wasn't END last time... */
21348 if (op == CLOSE || op == SRCLOSE || op == WHILEM)
21350 next = regnext((regnode *)node);
21353 if (OP(node) == OPTIMIZED) {
21354 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
21361 regprop(r, sv, node, NULL, NULL);
21362 Perl_re_printf( aTHX_ "%4" IVdf ":%*s%s", (IV)(node - start),
21363 (int)(2*indent + 1), "", SvPVX_const(sv));
21365 if (OP(node) != OPTIMIZED) {
21366 if (next == NULL) /* Next ptr. */
21367 Perl_re_printf( aTHX_ " (0)");
21368 else if (PL_regkind[(U8)op] == BRANCH
21369 && PL_regkind[OP(next)] != BRANCH )
21370 Perl_re_printf( aTHX_ " (FAIL)");
21372 Perl_re_printf( aTHX_ " (%" IVdf ")", (IV)(next - start));
21373 Perl_re_printf( aTHX_ "\n");
21377 if (PL_regkind[(U8)op] == BRANCHJ) {
21380 const regnode *nnode = (OP(next) == LONGJMP
21381 ? regnext((regnode *)next)
21383 if (last && nnode > last)
21385 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
21388 else if (PL_regkind[(U8)op] == BRANCH) {
21390 DUMPUNTIL(NEXTOPER(node), next);
21392 else if ( PL_regkind[(U8)op] == TRIE ) {
21393 const regnode *this_trie = node;
21394 const char op = OP(node);
21395 const U32 n = ARG(node);
21396 const reg_ac_data * const ac = op>=AHOCORASICK ?
21397 (reg_ac_data *)ri->data->data[n] :
21399 const reg_trie_data * const trie =
21400 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
21402 AV *const trie_words
21403 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
21405 const regnode *nextbranch= NULL;
21408 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
21409 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
21411 Perl_re_indentf( aTHX_ "%s ",
21414 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
21415 SvCUR(*elem_ptr), PL_dump_re_max_len,
21416 PL_colors[0], PL_colors[1],
21418 ? PERL_PV_ESCAPE_UNI
21420 | PERL_PV_PRETTY_ELLIPSES
21421 | PERL_PV_PRETTY_LTGT
21426 U16 dist= trie->jump[word_idx+1];
21427 Perl_re_printf( aTHX_ "(%" UVuf ")\n",
21428 (UV)((dist ? this_trie + dist : next) - start));
21431 nextbranch= this_trie + trie->jump[0];
21432 DUMPUNTIL(this_trie + dist, nextbranch);
21434 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
21435 nextbranch= regnext((regnode *)nextbranch);
21437 Perl_re_printf( aTHX_ "\n");
21440 if (last && next > last)
21445 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
21446 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
21447 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
21449 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
21451 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
21453 else if ( op == PLUS || op == STAR) {
21454 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
21456 else if (PL_regkind[(U8)op] == ANYOF) {
21457 /* arglen 1 + class block */
21458 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_MATCHES_POSIXL)
21459 ? ANYOF_POSIXL_SKIP
21461 node = NEXTOPER(node);
21463 else if (PL_regkind[(U8)op] == EXACT) {
21464 /* Literal string, where present. */
21465 node += NODE_SZ_STR(node) - 1;
21466 node = NEXTOPER(node);
21469 node = NEXTOPER(node);
21470 node += regarglen[(U8)op];
21472 if (op == CURLYX || op == OPEN || op == SROPEN)
21476 #ifdef DEBUG_DUMPUNTIL
21477 Perl_re_printf( aTHX_ "--- %d\n", (int)indent);
21482 #endif /* DEBUGGING */
21484 #ifndef PERL_IN_XSUB_RE
21486 #include "uni_keywords.h"
21489 Perl_init_uniprops(pTHX)
21491 /* Set up the inversion list global variables */
21493 PL_XPosix_ptrs[_CC_ASCII] = _new_invlist_C_array(uni_prop_ptrs[UNI_ASCII]);
21494 PL_XPosix_ptrs[_CC_ALPHANUMERIC] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXALNUM]);
21495 PL_XPosix_ptrs[_CC_ALPHA] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXALPHA]);
21496 PL_XPosix_ptrs[_CC_BLANK] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXBLANK]);
21497 PL_XPosix_ptrs[_CC_CASED] = _new_invlist_C_array(uni_prop_ptrs[UNI_CASED]);
21498 PL_XPosix_ptrs[_CC_CNTRL] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXCNTRL]);
21499 PL_XPosix_ptrs[_CC_DIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXDIGIT]);
21500 PL_XPosix_ptrs[_CC_GRAPH] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXGRAPH]);
21501 PL_XPosix_ptrs[_CC_LOWER] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXLOWER]);
21502 PL_XPosix_ptrs[_CC_PRINT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXPRINT]);
21503 PL_XPosix_ptrs[_CC_PUNCT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXPUNCT]);
21504 PL_XPosix_ptrs[_CC_SPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXSPACE]);
21505 PL_XPosix_ptrs[_CC_UPPER] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXUPPER]);
21506 PL_XPosix_ptrs[_CC_VERTSPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_VERTSPACE]);
21507 PL_XPosix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXWORD]);
21508 PL_XPosix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXXDIGIT]);
21510 PL_Posix_ptrs[_CC_ASCII] = _new_invlist_C_array(uni_prop_ptrs[UNI_ASCII]);
21511 PL_Posix_ptrs[_CC_ALPHANUMERIC] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXALNUM]);
21512 PL_Posix_ptrs[_CC_ALPHA] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXALPHA]);
21513 PL_Posix_ptrs[_CC_BLANK] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXBLANK]);
21514 PL_Posix_ptrs[_CC_CASED] = _new_invlist_C_array(uni_prop_ptrs[UNI_CASED]);
21515 PL_Posix_ptrs[_CC_CNTRL] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXCNTRL]);
21516 PL_Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXDIGIT]);
21517 PL_Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXGRAPH]);
21518 PL_Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXLOWER]);
21519 PL_Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXPRINT]);
21520 PL_Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXPUNCT]);
21521 PL_Posix_ptrs[_CC_SPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXSPACE]);
21522 PL_Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXUPPER]);
21523 PL_Posix_ptrs[_CC_VERTSPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_VERTSPACE]);
21524 PL_Posix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXWORD]);
21525 PL_Posix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXXDIGIT]);
21527 PL_GCB_invlist = _new_invlist_C_array(_Perl_GCB_invlist);
21528 PL_SB_invlist = _new_invlist_C_array(_Perl_SB_invlist);
21529 PL_WB_invlist = _new_invlist_C_array(_Perl_WB_invlist);
21530 PL_LB_invlist = _new_invlist_C_array(_Perl_LB_invlist);
21531 PL_SCX_invlist = _new_invlist_C_array(_Perl_SCX_invlist);
21533 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
21534 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
21535 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
21537 PL_Assigned_invlist = _new_invlist_C_array(uni_prop_ptrs[UNI_ASSIGNED]);
21539 PL_utf8_perl_idstart = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_IDSTART]);
21540 PL_utf8_perl_idcont = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_IDCONT]);
21542 PL_utf8_charname_begin = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_CHARNAME_BEGIN]);
21543 PL_utf8_charname_continue = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_CHARNAME_CONTINUE]);
21545 PL_utf8_foldable = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_ANY_FOLDS]);
21546 PL_HasMultiCharFold = _new_invlist_C_array(uni_prop_ptrs[
21547 UNI__PERL_FOLDS_TO_MULTI_CHAR]);
21548 PL_NonL1NonFinalFold = _new_invlist_C_array(
21549 NonL1_Perl_Non_Final_Folds_invlist);
21551 PL_utf8_toupper = _new_invlist_C_array(Uppercase_Mapping_invlist);
21552 PL_utf8_tolower = _new_invlist_C_array(Lowercase_Mapping_invlist);
21553 PL_utf8_totitle = _new_invlist_C_array(Titlecase_Mapping_invlist);
21554 PL_utf8_tofold = _new_invlist_C_array(Case_Folding_invlist);
21555 PL_utf8_tosimplefold = _new_invlist_C_array(Simple_Case_Folding_invlist);
21556 PL_utf8_foldclosures = _new_invlist_C_array(_Perl_IVCF_invlist);
21557 PL_utf8_mark = _new_invlist_C_array(uni_prop_ptrs[UNI_M]);
21559 /* The below are used only by deprecated functions. They could be removed */
21560 PL_utf8_xidcont = _new_invlist_C_array(uni_prop_ptrs[UNI_XIDC]);
21561 PL_utf8_idcont = _new_invlist_C_array(uni_prop_ptrs[UNI_IDC]);
21562 PL_utf8_xidstart = _new_invlist_C_array(uni_prop_ptrs[UNI_XIDS]);
21566 Perl_parse_uniprop_string(pTHX_ const char * const name, const Size_t name_len,
21567 const bool to_fold, bool * invert)
21569 /* Parse the interior meat of \p{} passed to this in 'name' with length
21570 * 'name_len', and return an inversion list if a property with 'name' is
21571 * found, or NULL if not. 'name' point to the input with leading and
21572 * trailing space trimmed. 'to_fold' indicates if /i is in effect.
21574 * When the return is an inversion list, '*invert' will be set to a boolean
21575 * indicating if it should be inverted or not
21577 * This currently doesn't handle all cases. A NULL return indicates the
21578 * caller should try a different approach
21582 bool stricter = FALSE;
21583 bool is_nv_type = FALSE; /* nv= or numeric_value=, or possibly one
21584 of the cjk numeric properties (though
21585 it requires extra effort to compile
21588 unsigned int j = 0, lookup_len;
21589 int equals_pos = -1; /* Where the '=' is found, or negative if none */
21590 int slash_pos = -1; /* Where the '/' is found, or negative if none */
21591 int table_index = 0;
21592 bool starts_with_In_or_Is = FALSE;
21593 Size_t lookup_offset = 0;
21595 PERL_ARGS_ASSERT_PARSE_UNIPROP_STRING;
21597 /* The input will be modified into 'lookup_name' */
21598 Newx(lookup_name, name_len, char);
21599 SAVEFREEPV(lookup_name);
21601 /* Parse the input. */
21602 for (i = 0; i < name_len; i++) {
21603 char cur = name[i];
21605 /* These characters can be freely ignored in most situations. Later it
21606 * may turn out we shouldn't have ignored them, and we have to reparse,
21607 * but we don't have enough information yet to make that decision */
21608 if (cur == '-' || cur == '_' || isSPACE_A(cur)) {
21612 /* Case differences are also ignored. Our lookup routine assumes
21613 * everything is lowercase */
21614 if (isUPPER_A(cur)) {
21615 lookup_name[j++] = toLOWER(cur);
21619 /* A double colon is either an error, or a package qualifier to a
21620 * subroutine user-defined property; neither of which do we currently
21623 * But a single colon is a synonym for '=' */
21625 if (i < name_len - 1 && name[i+1] == ':') {
21631 /* Otherwise, this character is part of the name. */
21632 lookup_name[j++] = cur;
21634 /* Only the equals sign needs further processing */
21636 equals_pos = j; /* Note where it occurred in the input */
21641 /* Here, we are either done with the whole property name, if it was simple;
21642 * or are positioned just after the '=' if it is compound. */
21644 if (equals_pos >= 0) {
21645 assert(! stricter); /* We shouldn't have set this yet */
21647 /* Space immediately after the '=' is ignored */
21649 for (; i < name_len; i++) {
21650 if (! isSPACE_A(name[i])) {
21655 /* Certain properties need special handling. They may optionally be
21656 * prefixed by 'is'. Ignore that prefix for the purposes of checking
21657 * if this is one of those properties */
21658 if (memBEGINPs(lookup_name, name_len, "is")) {
21662 /* Then check if it is one of these properties. This is hard-coded
21663 * because easier this way, and the list is unlikely to change. There
21664 * are several properties like this in the Unihan DB, which is unlikely
21665 * to be compiled, and they all end with 'numeric'. The interiors
21666 * aren't checked for the precise property. This would stop working if
21667 * a cjk property were to be created that ended with 'numeric' and
21668 * wasn't a numeric type */
21669 is_nv_type = memEQs(lookup_name + lookup_offset,
21670 j - 1 - lookup_offset, "numericvalue")
21671 || memEQs(lookup_name + lookup_offset,
21672 j - 1 - lookup_offset, "nv")
21673 || ( memENDPs(lookup_name + lookup_offset,
21674 j - 1 - lookup_offset, "numeric")
21675 && ( memBEGINPs(lookup_name + lookup_offset,
21676 j - 1 - lookup_offset, "cjk")
21677 || memBEGINPs(lookup_name + lookup_offset,
21678 j - 1 - lookup_offset, "k")));
21680 || memEQs(lookup_name + lookup_offset,
21681 j - 1 - lookup_offset, "canonicalcombiningclass")
21682 || memEQs(lookup_name + lookup_offset,
21683 j - 1 - lookup_offset, "ccc")
21684 || memEQs(lookup_name + lookup_offset,
21685 j - 1 - lookup_offset, "age")
21686 || memEQs(lookup_name + lookup_offset,
21687 j - 1 - lookup_offset, "in")
21688 || memEQs(lookup_name + lookup_offset,
21689 j - 1 - lookup_offset, "presentin"))
21693 /* What makes these properties special is that the stuff after the
21694 * '=' is a number. Therefore, we can't throw away '-'
21695 * willy-nilly, as those could be a minus sign. Other stricter
21696 * rules also apply. However, these properties all can have the
21697 * rhs not be a number, in which case they contain at least one
21698 * alphabetic. In those cases, the stricter rules don't apply.
21699 * But the numeric type properties can have the alphas [Ee] to
21700 * signify an exponent, and it is still a number with stricter
21701 * rules. So look for an alpha that signifys not-strict */
21703 for (k = i; k < name_len; k++) {
21704 if ( isALPHA_A(name[k])
21705 && (! is_nv_type || ! isALPHA_FOLD_EQ(name[k], 'E')))
21715 /* A number may have a leading '+' or '-'. The latter is retained
21717 if (name[i] == '+') {
21720 else if (name[i] == '-') {
21721 lookup_name[j++] = '-';
21725 /* Skip leading zeros including single underscores separating the
21726 * zeros, or between the final leading zero and the first other
21728 for (; i < name_len - 1; i++) {
21729 if ( name[i] != '0'
21730 && (name[i] != '_' || ! isDIGIT_A(name[i+1])))
21737 else { /* No '=' */
21739 /* We are now in a position to determine if this property should have
21740 * been parsed using stricter rules. Only a few are like that, and
21741 * unlikely to change. */
21742 if ( memBEGINPs(lookup_name, j, "perl")
21743 && memNEs(lookup_name + 4, j - 4, "space")
21744 && memNEs(lookup_name + 4, j - 4, "word"))
21748 /* We set the inputs back to 0 and the code below will reparse,
21754 /* Here, we have either finished the property, or are positioned to parse
21755 * the remainder, and we know if stricter rules apply. Finish out, if not
21757 for (; i < name_len; i++) {
21758 char cur = name[i];
21760 /* In all instances, case differences are ignored, and we normalize to
21762 if (isUPPER_A(cur)) {
21763 lookup_name[j++] = toLOWER(cur);
21767 /* An underscore is skipped, but not under strict rules unless it
21768 * separates two digits */
21771 && ( i == 0 || (int) i == equals_pos || i == name_len- 1
21772 || ! isDIGIT_A(name[i-1]) || ! isDIGIT_A(name[i+1])))
21774 lookup_name[j++] = '_';
21779 /* Hyphens are skipped except under strict */
21780 if (cur == '-' && ! stricter) {
21784 /* XXX Bug in documentation. It says white space skipped adjacent to
21785 * non-word char. Maybe we should, but shouldn't skip it next to a dot
21787 if (isSPACE_A(cur) && ! stricter) {
21791 lookup_name[j++] = cur;
21793 /* Unless this is a non-trailing slash, we are done with it */
21794 if (i >= name_len - 1 || cur != '/') {
21800 /* A slash in the 'numeric value' property indicates that what follows
21801 * is a denominator. It can have a leading '+' and '0's that should be
21802 * skipped. But we have never allowed a negative denominator, so treat
21803 * a minus like every other character. (No need to rule out a second
21804 * '/', as that won't match anything anyway */
21807 if (i < name_len && name[i] == '+') {
21811 /* Skip leading zeros including underscores separating digits */
21812 for (; i < name_len - 1; i++) {
21813 if ( name[i] != '0'
21814 && (name[i] != '_' || ! isDIGIT_A(name[i+1])))
21820 /* Store the first real character in the denominator */
21821 lookup_name[j++] = name[i];
21825 /* Here are completely done parsing the input 'name', and 'lookup_name'
21826 * contains a copy, normalized.
21828 * This special case is grandfathered in: 'L_' and 'GC=L_' are accepted and
21829 * different from without the underscores. */
21830 if ( ( UNLIKELY(memEQs(lookup_name, j, "l"))
21831 || UNLIKELY(memEQs(lookup_name, j, "gc=l")))
21832 && UNLIKELY(name[name_len-1] == '_'))
21834 lookup_name[j++] = '&';
21836 else if (name_len > 2 && name[0] == 'I' && ( name[1] == 'n'
21837 || name[1] == 's'))
21840 /* Also, if the original input began with 'In' or 'Is', it could be a
21841 * subroutine call instead of a property names, which currently isn't
21842 * handled by this function. Subroutine calls can't happen if there is
21843 * an '=' in the name */
21844 if (equals_pos < 0 && get_cvn_flags(name, name_len, GV_NOTQUAL) != NULL)
21849 starts_with_In_or_Is = TRUE;
21852 lookup_len = j; /* Use a more mnemonic name starting here */
21854 /* Get the index into our pointer table of the inversion list corresponding
21855 * to the property */
21856 table_index = match_uniprop((U8 *) lookup_name, lookup_len);
21858 /* If it didn't find the property */
21859 if (table_index == 0) {
21861 /* If didn't find the property, we try again stripping off any initial
21863 if (starts_with_In_or_Is) {
21869 table_index = match_uniprop((U8 *) lookup_name, lookup_len);
21872 if (table_index == 0) {
21875 /* If not found, and not a numeric type property, isn't a legal
21877 if (! is_nv_type) {
21881 /* But the numeric type properties need more work to decide. What
21882 * we do is make sure we have the number in canonical form and look
21885 if (slash_pos < 0) { /* No slash */
21887 /* When it isn't a rational, take the input, convert it to a
21888 * NV, then create a canonical string representation of that
21893 /* Get the value */
21894 if (my_atof3(lookup_name + equals_pos, &value,
21895 lookup_len - equals_pos)
21896 != lookup_name + lookup_len)
21901 /* If the value is an integer, the canonical value is integral */
21902 if (Perl_ceil(value) == value) {
21903 canonical = Perl_form(aTHX_ "%.*s%.0" NVff,
21904 equals_pos, lookup_name, value);
21906 else { /* Otherwise, it is %e with a known precision */
21909 canonical = Perl_form(aTHX_ "%.*s%.*" NVef,
21910 equals_pos, lookup_name,
21911 PL_E_FORMAT_PRECISION, value);
21913 /* The exponent generated is expecting two digits, whereas
21914 * %e on some systems will generate three. Remove leading
21915 * zeros in excess of 2 from the exponent. We start
21916 * looking for them after the '=' */
21917 exp_ptr = strchr(canonical + equals_pos, 'e');
21919 char * cur_ptr = exp_ptr + 2; /* past the 'e[+-]' */
21920 SSize_t excess_exponent_len = strlen(cur_ptr) - 2;
21922 assert(*(cur_ptr - 1) == '-' || *(cur_ptr - 1) == '+');
21924 if (excess_exponent_len > 0) {
21925 SSize_t leading_zeros = strspn(cur_ptr, "0");
21926 SSize_t excess_leading_zeros
21927 = MIN(leading_zeros, excess_exponent_len);
21928 if (excess_leading_zeros > 0) {
21929 Move(cur_ptr + excess_leading_zeros,
21931 strlen(cur_ptr) - excess_leading_zeros
21932 + 1, /* Copy the NUL as well */
21939 else { /* Has a slash. Create a rational in canonical form */
21940 UV numerator, denominator, gcd, trial;
21941 const char * end_ptr;
21942 const char * sign = "";
21944 /* We can't just find the numerator, denominator, and do the
21945 * division, then use the method above, because that is
21946 * inexact. And the input could be a rational that is within
21947 * epsilon (given our precision) of a valid rational, and would
21948 * then incorrectly compare valid.
21950 * We're only interested in the part after the '=' */
21951 const char * this_lookup_name = lookup_name + equals_pos;
21952 lookup_len -= equals_pos;
21953 slash_pos -= equals_pos;
21955 /* Handle any leading minus */
21956 if (this_lookup_name[0] == '-') {
21958 this_lookup_name++;
21963 /* Convert the numerator to numeric */
21964 end_ptr = this_lookup_name + slash_pos;
21965 if (! grok_atoUV(this_lookup_name, &numerator, &end_ptr)) {
21969 /* It better have included all characters before the slash */
21970 if (*end_ptr != '/') {
21974 /* Set to look at just the denominator */
21975 this_lookup_name += slash_pos;
21976 lookup_len -= slash_pos;
21977 end_ptr = this_lookup_name + lookup_len;
21979 /* Convert the denominator to numeric */
21980 if (! grok_atoUV(this_lookup_name, &denominator, &end_ptr)) {
21984 /* It better be the rest of the characters, and don't divide by
21986 if ( end_ptr != this_lookup_name + lookup_len
21987 || denominator == 0)
21992 /* Get the greatest common denominator using
21993 http://en.wikipedia.org/wiki/Euclidean_algorithm */
21995 trial = denominator;
21996 while (trial != 0) {
21998 trial = gcd % trial;
22002 /* If already in lowest possible terms, we have already tried
22003 * looking this up */
22008 /* Reduce the rational, which should put it in canonical form.
22009 * Then look it up */
22011 denominator /= gcd;
22013 canonical = Perl_form(aTHX_ "%.*s%s%" UVuf "/%" UVuf,
22014 equals_pos, lookup_name, sign, numerator, denominator);
22017 /* Here, we have the number in canonical form. Try that */
22018 table_index = match_uniprop((U8 *) canonical, strlen(canonical));
22019 if (table_index == 0) {
22025 /* The return is an index into a table of ptrs. A negative return
22026 * signifies that the real index is the absolute value, but the result
22027 * needs to be inverted */
22028 if (table_index < 0) {
22030 table_index = -table_index;
22036 /* Out-of band indices indicate a deprecated property. The proper index is
22037 * modulo it with the table size. And dividing by the table size yields
22038 * an offset into a table constructed to contain the corresponding warning
22040 if (table_index > MAX_UNI_KEYWORD_INDEX) {
22041 Size_t warning_offset = table_index / MAX_UNI_KEYWORD_INDEX;
22042 table_index %= MAX_UNI_KEYWORD_INDEX;
22043 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED),
22044 "Use of '%.*s' in \\p{} or \\P{} is deprecated because: %s",
22045 (int) name_len, name, deprecated_property_msgs[warning_offset]);
22048 /* In a few properties, a different property is used under /i. These are
22049 * unlikely to change, so are hard-coded here. */
22051 if ( table_index == UNI_XPOSIXUPPER
22052 || table_index == UNI_XPOSIXLOWER
22053 || table_index == UNI_TITLE)
22055 table_index = UNI_CASED;
22057 else if ( table_index == UNI_UPPERCASELETTER
22058 || table_index == UNI_LOWERCASELETTER
22059 # ifdef UNI_TITLECASELETTER /* Missing from early Unicodes */
22060 || table_index == UNI_TITLECASELETTER
22063 table_index = UNI_CASEDLETTER;
22065 else if ( table_index == UNI_POSIXUPPER
22066 || table_index == UNI_POSIXLOWER)
22068 table_index = UNI_POSIXALPHA;
22072 /* Create and return the inversion list */
22073 return _new_invlist_C_array(uni_prop_ptrs[table_index]);
22079 * ex: set ts=8 sts=4 sw=4 et: