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 *copy_start; /* start of copy of input within
137 constructed parse string */
138 char *copy_start_in_input; /* Position in input string
139 corresponding to copy_start */
140 SSize_t whilem_seen; /* number of WHILEM in this expr */
141 regnode *emit_start; /* Start of emitted-code area */
142 regnode_offset emit; /* Code-emit pointer */
143 I32 naughty; /* How bad is this pattern? */
144 I32 sawback; /* Did we see \1, ...? */
146 SSize_t size; /* Number of regnode equivalents in
149 /* position beyond 'precomp' of the warning message furthest away from
150 * 'precomp'. During the parse, no warnings are raised for any problems
151 * earlier in the parse than this position. This works if warnings are
152 * raised the first time a given spot is parsed, and if only one
153 * independent warning is raised for any given spot */
154 Size_t latest_warn_offset;
156 I32 npar; /* Capture buffer count so far in the
157 parse, (OPEN) plus one. ("par" 0 is
159 I32 total_par; /* During initial parse, is either 0,
160 or -1; the latter indicating a
161 reparse is needed. After that pass,
162 it is what 'npar' became after the
163 pass. Hence, it being > 0 indicates
164 we are in a reparse situation */
165 I32 nestroot; /* root parens we are in - used by
168 regnode_offset *open_parens; /* offsets to open parens */
169 regnode_offset *close_parens; /* offsets to close parens */
170 regnode *end_op; /* END node in program */
171 I32 utf8; /* whether the pattern is utf8 or not */
172 I32 orig_utf8; /* whether the pattern was originally in utf8 */
173 /* XXX use this for future optimisation of case
174 * where pattern must be upgraded to utf8. */
175 I32 uni_semantics; /* If a d charset modifier should use unicode
176 rules, even if the pattern is not in
178 HV *paren_names; /* Paren names */
180 regnode **recurse; /* Recurse regops */
181 I32 recurse_count; /* Number of recurse regops we have generated */
182 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
184 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
187 I32 override_recoding;
189 I32 recode_x_to_native;
191 I32 in_multi_char_class;
192 struct reg_code_blocks *code_blocks;/* positions of literal (?{})
194 int code_index; /* next code_blocks[] slot */
195 SSize_t maxlen; /* mininum possible number of chars in string to match */
196 scan_frame *frame_head;
197 scan_frame *frame_last;
200 #ifdef ADD_TO_REGEXEC
201 char *starttry; /* -Dr: where regtry was called. */
202 #define RExC_starttry (pRExC_state->starttry)
204 SV *runtime_code_qr; /* qr with the runtime code blocks */
206 const char *lastparse;
208 AV *paren_name_list; /* idx -> name */
209 U32 study_chunk_recursed_count;
213 #define RExC_lastparse (pRExC_state->lastparse)
214 #define RExC_lastnum (pRExC_state->lastnum)
215 #define RExC_paren_name_list (pRExC_state->paren_name_list)
216 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
217 #define RExC_mysv (pRExC_state->mysv1)
218 #define RExC_mysv1 (pRExC_state->mysv1)
219 #define RExC_mysv2 (pRExC_state->mysv2)
229 #define RExC_flags (pRExC_state->flags)
230 #define RExC_pm_flags (pRExC_state->pm_flags)
231 #define RExC_precomp (pRExC_state->precomp)
232 #define RExC_copy_start_in_input (pRExC_state->copy_start_in_input)
233 #define RExC_copy_start_in_constructed (pRExC_state->copy_start)
234 #define RExC_precomp_end (pRExC_state->precomp_end)
235 #define RExC_rx_sv (pRExC_state->rx_sv)
236 #define RExC_rx (pRExC_state->rx)
237 #define RExC_rxi (pRExC_state->rxi)
238 #define RExC_start (pRExC_state->start)
239 #define RExC_end (pRExC_state->end)
240 #define RExC_parse (pRExC_state->parse)
241 #define RExC_latest_warn_offset (pRExC_state->latest_warn_offset )
242 #define RExC_whilem_seen (pRExC_state->whilem_seen)
243 #define RExC_seen_d_op (pRExC_state->seen_d_op) /* Seen something that differs
244 under /d from /u ? */
247 #ifdef RE_TRACK_PATTERN_OFFSETS
248 # define RExC_offsets (RExC_rxi->u.offsets) /* I am not like the
251 #define RExC_emit (pRExC_state->emit)
252 #define RExC_emit_start (pRExC_state->emit_start)
253 #define RExC_sawback (pRExC_state->sawback)
254 #define RExC_seen (pRExC_state->seen)
255 #define RExC_size (pRExC_state->size)
256 #define RExC_maxlen (pRExC_state->maxlen)
257 #define RExC_npar (pRExC_state->npar)
258 #define RExC_total_parens (pRExC_state->total_par)
259 #define RExC_nestroot (pRExC_state->nestroot)
260 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
261 #define RExC_utf8 (pRExC_state->utf8)
262 #define RExC_uni_semantics (pRExC_state->uni_semantics)
263 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
264 #define RExC_open_parens (pRExC_state->open_parens)
265 #define RExC_close_parens (pRExC_state->close_parens)
266 #define RExC_end_op (pRExC_state->end_op)
267 #define RExC_paren_names (pRExC_state->paren_names)
268 #define RExC_recurse (pRExC_state->recurse)
269 #define RExC_recurse_count (pRExC_state->recurse_count)
270 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
271 #define RExC_study_chunk_recursed_bytes \
272 (pRExC_state->study_chunk_recursed_bytes)
273 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
274 #define RExC_contains_locale (pRExC_state->contains_locale)
276 # define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
278 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
279 #define RExC_frame_head (pRExC_state->frame_head)
280 #define RExC_frame_last (pRExC_state->frame_last)
281 #define RExC_frame_count (pRExC_state->frame_count)
282 #define RExC_strict (pRExC_state->strict)
283 #define RExC_study_started (pRExC_state->study_started)
284 #define RExC_warn_text (pRExC_state->warn_text)
285 #define RExC_in_script_run (pRExC_state->in_script_run)
286 #define RExC_use_BRANCHJ (pRExC_state->use_BRANCHJ)
288 /* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
289 * a flag to disable back-off on the fixed/floating substrings - if it's
290 * a high complexity pattern we assume the benefit of avoiding a full match
291 * is worth the cost of checking for the substrings even if they rarely help.
293 #define RExC_naughty (pRExC_state->naughty)
294 #define TOO_NAUGHTY (10)
295 #define MARK_NAUGHTY(add) \
296 if (RExC_naughty < TOO_NAUGHTY) \
297 RExC_naughty += (add)
298 #define MARK_NAUGHTY_EXP(exp, add) \
299 if (RExC_naughty < TOO_NAUGHTY) \
300 RExC_naughty += RExC_naughty / (exp) + (add)
302 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
303 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
304 ((*s) == '{' && regcurly(s)))
307 * Flags to be passed up and down.
309 #define WORST 0 /* Worst case. */
310 #define HASWIDTH 0x01 /* Known to not match null strings, could match
313 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
314 * character. (There needs to be a case: in the switch statement in regexec.c
315 * for any node marked SIMPLE.) Note that this is not the same thing as
318 #define SPSTART 0x04 /* Starts with * or + */
319 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
320 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
321 #define RESTART_PARSE 0x20 /* Need to redo the parse */
322 #define NEED_UTF8 0x40 /* In conjunction with RESTART_PARSE, need to
323 calcuate sizes as UTF-8 */
325 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
327 /* whether trie related optimizations are enabled */
328 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
329 #define TRIE_STUDY_OPT
330 #define FULL_TRIE_STUDY
336 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
337 #define PBITVAL(paren) (1 << ((paren) & 7))
338 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
339 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
340 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
342 #define REQUIRE_UTF8(flagp) STMT_START { \
344 *flagp = RESTART_PARSE|NEED_UTF8; \
349 /* Change from /d into /u rules, and restart the parse. RExC_uni_semantics is
350 * a flag that indicates we need to override /d with /u as a result of
351 * something in the pattern. It should only be used in regards to calling
352 * set_regex_charset() or get_regex_charse() */
353 #define REQUIRE_UNI_RULES(flagp, restart_retval) \
355 if (DEPENDS_SEMANTICS) { \
356 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
357 RExC_uni_semantics = 1; \
358 if (RExC_seen_d_op && LIKELY(RExC_total_parens >= 0)) { \
359 /* No need to restart the parse if we haven't seen \
360 * anything that differs between /u and /d, and no need \
361 * to restart immediately if we're going to reparse \
362 * anyway to count parens */ \
363 *flagp |= RESTART_PARSE; \
364 return restart_retval; \
369 #define BRANCH_MAX_OFFSET U16_MAX
370 #define REQUIRE_BRANCHJ(flagp, restart_retval) \
372 RExC_use_BRANCHJ = 1; \
373 if (LIKELY(RExC_total_parens >= 0)) { \
374 /* No need to restart the parse immediately if we're \
375 * going to reparse anyway to count parens */ \
376 *flagp |= RESTART_PARSE; \
377 return restart_retval; \
381 #define REQUIRE_PARENS_PASS \
383 if (RExC_total_parens == 0) RExC_total_parens = -1; \
386 /* This is used to return failure (zero) early from the calling function if
387 * various flags in 'flags' are set. Two flags always cause a return:
388 * 'RESTART_PARSE' and 'NEED_UTF8'. 'extra' can be used to specify any
389 * additional flags that should cause a return; 0 if none. If the return will
390 * be done, '*flagp' is first set to be all of the flags that caused the
392 #define RETURN_FAIL_ON_RESTART_OR_FLAGS(flags,flagp,extra) \
394 if ((flags) & (RESTART_PARSE|NEED_UTF8|(extra))) { \
395 *(flagp) = (flags) & (RESTART_PARSE|NEED_UTF8|(extra)); \
400 #define MUST_RESTART(flags) ((flags) & (RESTART_PARSE))
402 #define RETURN_FAIL_ON_RESTART(flags,flagp) \
403 RETURN_FAIL_ON_RESTART_OR_FLAGS( flags, flagp, 0)
404 #define RETURN_FAIL_ON_RESTART_FLAGP(flagp) \
405 if (MUST_RESTART(*(flagp))) return 0
407 /* This converts the named class defined in regcomp.h to its equivalent class
408 * number defined in handy.h. */
409 #define namedclass_to_classnum(class) ((int) ((class) / 2))
410 #define classnum_to_namedclass(classnum) ((classnum) * 2)
412 #define _invlist_union_complement_2nd(a, b, output) \
413 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
414 #define _invlist_intersection_complement_2nd(a, b, output) \
415 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
417 /* About scan_data_t.
419 During optimisation we recurse through the regexp program performing
420 various inplace (keyhole style) optimisations. In addition study_chunk
421 and scan_commit populate this data structure with information about
422 what strings MUST appear in the pattern. We look for the longest
423 string that must appear at a fixed location, and we look for the
424 longest string that may appear at a floating location. So for instance
429 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
430 strings (because they follow a .* construct). study_chunk will identify
431 both FOO and BAR as being the longest fixed and floating strings respectively.
433 The strings can be composites, for instance
437 will result in a composite fixed substring 'foo'.
439 For each string some basic information is maintained:
442 This is the position the string must appear at, or not before.
443 It also implicitly (when combined with minlenp) tells us how many
444 characters must match before the string we are searching for.
445 Likewise when combined with minlenp and the length of the string it
446 tells us how many characters must appear after the string we have
450 Only used for floating strings. This is the rightmost point that
451 the string can appear at. If set to SSize_t_MAX it indicates that the
452 string can occur infinitely far to the right.
453 For fixed strings, it is equal to min_offset.
456 A pointer to the minimum number of characters of the pattern that the
457 string was found inside. This is important as in the case of positive
458 lookahead or positive lookbehind we can have multiple patterns
463 The minimum length of the pattern overall is 3, the minimum length
464 of the lookahead part is 3, but the minimum length of the part that
465 will actually match is 1. So 'FOO's minimum length is 3, but the
466 minimum length for the F is 1. This is important as the minimum length
467 is used to determine offsets in front of and behind the string being
468 looked for. Since strings can be composites this is the length of the
469 pattern at the time it was committed with a scan_commit. Note that
470 the length is calculated by study_chunk, so that the minimum lengths
471 are not known until the full pattern has been compiled, thus the
472 pointer to the value.
476 In the case of lookbehind the string being searched for can be
477 offset past the start point of the final matching string.
478 If this value was just blithely removed from the min_offset it would
479 invalidate some of the calculations for how many chars must match
480 before or after (as they are derived from min_offset and minlen and
481 the length of the string being searched for).
482 When the final pattern is compiled and the data is moved from the
483 scan_data_t structure into the regexp structure the information
484 about lookbehind is factored in, with the information that would
485 have been lost precalculated in the end_shift field for the
488 The fields pos_min and pos_delta are used to store the minimum offset
489 and the delta to the maximum offset at the current point in the pattern.
493 struct scan_data_substrs {
494 SV *str; /* longest substring found in pattern */
495 SSize_t min_offset; /* earliest point in string it can appear */
496 SSize_t max_offset; /* latest point in string it can appear */
497 SSize_t *minlenp; /* pointer to the minlen relevant to the string */
498 SSize_t lookbehind; /* is the pos of the string modified by LB */
499 I32 flags; /* per substring SF_* and SCF_* flags */
502 typedef struct scan_data_t {
503 /*I32 len_min; unused */
504 /*I32 len_delta; unused */
508 SSize_t last_end; /* min value, <0 unless valid. */
509 SSize_t last_start_min;
510 SSize_t last_start_max;
511 U8 cur_is_floating; /* whether the last_* values should be set as
512 * the next fixed (0) or floating (1)
515 /* [0] is longest fixed substring so far, [1] is longest float so far */
516 struct scan_data_substrs substrs[2];
518 I32 flags; /* common SF_* and SCF_* flags */
520 SSize_t *last_closep;
521 regnode_ssc *start_class;
525 * Forward declarations for pregcomp()'s friends.
528 static const scan_data_t zero_scan_data = {
529 0, 0, NULL, 0, 0, 0, 0,
531 { NULL, 0, 0, 0, 0, 0 },
532 { NULL, 0, 0, 0, 0, 0 },
539 #define SF_BEFORE_SEOL 0x0001
540 #define SF_BEFORE_MEOL 0x0002
541 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
543 #define SF_IS_INF 0x0040
544 #define SF_HAS_PAR 0x0080
545 #define SF_IN_PAR 0x0100
546 #define SF_HAS_EVAL 0x0200
549 /* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
550 * longest substring in the pattern. When it is not set the optimiser keeps
551 * track of position, but does not keep track of the actual strings seen,
553 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
556 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
557 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
558 * turned off because of the alternation (BRANCH). */
559 #define SCF_DO_SUBSTR 0x0400
561 #define SCF_DO_STCLASS_AND 0x0800
562 #define SCF_DO_STCLASS_OR 0x1000
563 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
564 #define SCF_WHILEM_VISITED_POS 0x2000
566 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
567 #define SCF_SEEN_ACCEPT 0x8000
568 #define SCF_TRIE_DOING_RESTUDY 0x10000
569 #define SCF_IN_DEFINE 0x20000
574 #define UTF cBOOL(RExC_utf8)
576 /* The enums for all these are ordered so things work out correctly */
577 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
578 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
579 == REGEX_DEPENDS_CHARSET)
580 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
581 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
582 >= REGEX_UNICODE_CHARSET)
583 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
584 == REGEX_ASCII_RESTRICTED_CHARSET)
585 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
586 >= REGEX_ASCII_RESTRICTED_CHARSET)
587 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
588 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
590 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
592 /* For programs that want to be strictly Unicode compatible by dying if any
593 * attempt is made to match a non-Unicode code point against a Unicode
595 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
597 #define OOB_NAMEDCLASS -1
599 /* There is no code point that is out-of-bounds, so this is problematic. But
600 * its only current use is to initialize a variable that is always set before
602 #define OOB_UNICODE 0xDEADBEEF
604 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
607 /* length of regex to show in messages that don't mark a position within */
608 #define RegexLengthToShowInErrorMessages 127
611 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
612 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
613 * op/pragma/warn/regcomp.
615 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
616 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
618 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
619 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
621 /* The code in this file in places uses one level of recursion with parsing
622 * rebased to an alternate string constructed by us in memory. This can take
623 * the form of something that is completely different from the input, or
624 * something that uses the input as part of the alternate. In the first case,
625 * there should be no possibility of an error, as we are in complete control of
626 * the alternate string. But in the second case we don't completely control
627 * the input portion, so there may be errors in that. Here's an example:
629 * is handled specially because \x{df} folds to a sequence of more than one
630 * character: 'ss'. What is done is to create and parse an alternate string,
631 * which looks like this:
632 * /(?:\x{DF}|[abc\x{DF}def])/ui
633 * where it uses the input unchanged in the middle of something it constructs,
634 * which is a branch for the DF outside the character class, and clustering
635 * parens around the whole thing. (It knows enough to skip the DF inside the
636 * class while in this substitute parse.) 'abc' and 'def' may have errors that
637 * need to be reported. The general situation looks like this:
639 * |<------- identical ------>|
641 * Input: ---------------------------------------------------------------
642 * Constructed: ---------------------------------------------------
644 * |<------- identical ------>|
646 * sI..eI is the portion of the input pattern we are concerned with here.
647 * sC..EC is the constructed substitute parse string.
648 * sC..tC is constructed by us
649 * tC..eC is an exact duplicate of the portion of the input pattern tI..eI.
650 * In the diagram, these are vertically aligned.
651 * eC..EC is also constructed by us.
652 * xC is the position in the substitute parse string where we found a
654 * xI is the position in the original pattern corresponding to xC.
656 * We want to display a message showing the real input string. Thus we need to
657 * translate from xC to xI. We know that xC >= tC, since the portion of the
658 * string sC..tC has been constructed by us, and so shouldn't have errors. We
660 * xI = tI + (xC - tC)
662 * When the substitute parse is constructed, the code needs to set:
665 * RExC_copy_start_in_input (tI)
666 * RExC_copy_start_in_constructed (tC)
667 * and restore them when done.
669 * During normal processing of the input pattern, both
670 * 'RExC_copy_start_in_input' and 'RExC_copy_start_in_constructed' are set to
671 * sI, so that xC equals xI.
674 #define sI RExC_precomp
675 #define eI RExC_precomp_end
676 #define sC RExC_start
678 #define tI RExC_copy_start_in_input
679 #define tC RExC_copy_start_in_constructed
680 #define xI(xC) (tI + (xC - tC))
681 #define xI_offset(xC) (xI(xC) - sI)
683 #define REPORT_LOCATION_ARGS(xC) \
685 (xI(xC) > eI) /* Don't run off end */ \
686 ? eI - sI /* Length before the <--HERE */ \
687 : ((xI_offset(xC) >= 0) \
689 : (Perl_croak(aTHX_ "panic: %s: %d: negative offset: %" \
690 IVdf " trying to output message for " \
692 __FILE__, __LINE__, (IV) xI_offset(xC), \
693 ((int) (eC - sC)), sC), 0)), \
694 sI), /* The input pattern printed up to the <--HERE */ \
696 (xI(xC) > eI) ? 0 : eI - xI(xC), /* Length after <--HERE */ \
697 (xI(xC) > eI) ? eI : xI(xC)) /* pattern after <--HERE */
699 /* Used to point after bad bytes for an error message, but avoid skipping
700 * past a nul byte. */
701 #define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
703 /* Set up to clean up after our imminent demise */
704 #define PREPARE_TO_DIE \
707 SAVEFREESV(RExC_rx_sv); \
708 if (RExC_open_parens) \
709 SAVEFREEPV(RExC_open_parens); \
710 if (RExC_close_parens) \
711 SAVEFREEPV(RExC_close_parens); \
715 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
716 * arg. Show regex, up to a maximum length. If it's too long, chop and add
719 #define _FAIL(code) STMT_START { \
720 const char *ellipses = ""; \
721 IV len = RExC_precomp_end - RExC_precomp; \
724 if (len > RegexLengthToShowInErrorMessages) { \
725 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
726 len = RegexLengthToShowInErrorMessages - 10; \
732 #define FAIL(msg) _FAIL( \
733 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
734 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
736 #define FAIL2(msg,arg) _FAIL( \
737 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
738 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
741 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
743 #define Simple_vFAIL(m) STMT_START { \
744 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
745 m, REPORT_LOCATION_ARGS(RExC_parse)); \
749 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
751 #define vFAIL(m) STMT_START { \
757 * Like Simple_vFAIL(), but accepts two arguments.
759 #define Simple_vFAIL2(m,a1) STMT_START { \
760 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
761 REPORT_LOCATION_ARGS(RExC_parse)); \
765 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
767 #define vFAIL2(m,a1) STMT_START { \
769 Simple_vFAIL2(m, a1); \
774 * Like Simple_vFAIL(), but accepts three arguments.
776 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
777 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
778 REPORT_LOCATION_ARGS(RExC_parse)); \
782 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
784 #define vFAIL3(m,a1,a2) STMT_START { \
786 Simple_vFAIL3(m, a1, a2); \
790 * Like Simple_vFAIL(), but accepts four arguments.
792 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
793 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
794 REPORT_LOCATION_ARGS(RExC_parse)); \
797 #define vFAIL4(m,a1,a2,a3) STMT_START { \
799 Simple_vFAIL4(m, a1, a2, a3); \
802 /* A specialized version of vFAIL2 that works with UTF8f */
803 #define vFAIL2utf8f(m, a1) STMT_START { \
805 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
806 REPORT_LOCATION_ARGS(RExC_parse)); \
809 #define vFAIL3utf8f(m, a1, a2) STMT_START { \
811 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
812 REPORT_LOCATION_ARGS(RExC_parse)); \
815 /* Setting this to NULL is a signal to not output warnings */
816 #define TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE RExC_copy_start_in_constructed = NULL
817 #define RESTORE_WARNINGS RExC_copy_start_in_constructed = RExC_precomp
819 /* Since a warning can be generated multiple times as the input is reparsed, we
820 * output it the first time we come to that point in the parse, but suppress it
821 * otherwise. 'RExC_copy_start_in_constructed' being NULL is a flag to not
822 * generate any warnings */
823 #define TO_OUTPUT_WARNINGS(loc) \
824 ( RExC_copy_start_in_constructed \
825 && ((xI(loc)) - RExC_precomp) > (Ptrdiff_t) RExC_latest_warn_offset)
827 /* After we've emitted a warning, we save the position in the input so we don't
829 #define UPDATE_WARNINGS_LOC(loc) \
831 if (TO_OUTPUT_WARNINGS(loc)) { \
832 RExC_latest_warn_offset = (xI(loc)) - RExC_precomp; \
836 /* 'warns' is the output of the packWARNx macro used in 'code' */
837 #define _WARN_HELPER(loc, warns, code) \
839 if (! RExC_copy_start_in_constructed) { \
840 Perl_croak( aTHX_ "panic! %s: %d: Tried to warn when none" \
841 " expected at '%s'", \
842 __FILE__, __LINE__, loc); \
844 if (TO_OUTPUT_WARNINGS(loc)) { \
848 UPDATE_WARNINGS_LOC(loc); \
852 /* m is not necessarily a "literal string", in this macro */
853 #define reg_warn_non_literal_string(loc, m) \
854 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
855 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
856 "%s" REPORT_LOCATION, \
857 m, REPORT_LOCATION_ARGS(loc)))
859 #define ckWARNreg(loc,m) \
860 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
861 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
863 REPORT_LOCATION_ARGS(loc)))
865 #define vWARN(loc, m) \
866 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
867 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
869 REPORT_LOCATION_ARGS(loc))) \
871 #define vWARN_dep(loc, m) \
872 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
873 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
875 REPORT_LOCATION_ARGS(loc)))
877 #define ckWARNdep(loc,m) \
878 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
879 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
881 REPORT_LOCATION_ARGS(loc)))
883 #define ckWARNregdep(loc,m) \
884 _WARN_HELPER(loc, packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
885 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
888 REPORT_LOCATION_ARGS(loc)))
890 #define ckWARN2reg_d(loc,m, a1) \
891 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
892 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
894 a1, REPORT_LOCATION_ARGS(loc)))
896 #define ckWARN2reg(loc, m, a1) \
897 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
898 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
900 a1, REPORT_LOCATION_ARGS(loc)))
902 #define vWARN3(loc, m, a1, a2) \
903 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
904 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
906 a1, a2, REPORT_LOCATION_ARGS(loc)))
908 #define ckWARN3reg(loc, m, a1, a2) \
909 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
910 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
913 REPORT_LOCATION_ARGS(loc)))
915 #define vWARN4(loc, m, a1, a2, a3) \
916 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
917 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
920 REPORT_LOCATION_ARGS(loc)))
922 #define ckWARN4reg(loc, m, a1, a2, a3) \
923 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
924 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
927 REPORT_LOCATION_ARGS(loc)))
929 #define vWARN5(loc, m, a1, a2, a3, a4) \
930 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
931 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
934 REPORT_LOCATION_ARGS(loc)))
936 #define ckWARNexperimental(loc, class, m) \
937 _WARN_HELPER(loc, packWARN(class), \
938 Perl_ck_warner_d(aTHX_ packWARN(class), \
940 REPORT_LOCATION_ARGS(loc)))
942 /* Convert between a pointer to a node and its offset from the beginning of the
944 #define REGNODE_p(offset) (RExC_emit_start + (offset))
945 #define REGNODE_OFFSET(node) ((node) - RExC_emit_start)
947 /* Macros for recording node offsets. 20001227 mjd@plover.com
948 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
949 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
950 * Element 0 holds the number n.
951 * Position is 1 indexed.
953 #ifndef RE_TRACK_PATTERN_OFFSETS
954 #define Set_Node_Offset_To_R(offset,byte)
955 #define Set_Node_Offset(node,byte)
956 #define Set_Cur_Node_Offset
957 #define Set_Node_Length_To_R(node,len)
958 #define Set_Node_Length(node,len)
959 #define Set_Node_Cur_Length(node,start)
960 #define Node_Offset(n)
961 #define Node_Length(n)
962 #define Set_Node_Offset_Length(node,offset,len)
963 #define ProgLen(ri) ri->u.proglen
964 #define SetProgLen(ri,x) ri->u.proglen = x
965 #define Track_Code(code)
967 #define ProgLen(ri) ri->u.offsets[0]
968 #define SetProgLen(ri,x) ri->u.offsets[0] = x
969 #define Set_Node_Offset_To_R(offset,byte) STMT_START { \
970 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
971 __LINE__, (int)(offset), (int)(byte))); \
973 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
976 RExC_offsets[2*(offset)-1] = (byte); \
980 #define Set_Node_Offset(node,byte) \
981 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (byte)-RExC_start)
982 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
984 #define Set_Node_Length_To_R(node,len) STMT_START { \
985 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
986 __LINE__, (int)(node), (int)(len))); \
988 Perl_croak(aTHX_ "value of node is %d in Length macro", \
991 RExC_offsets[2*(node)] = (len); \
995 #define Set_Node_Length(node,len) \
996 Set_Node_Length_To_R(REGNODE_OFFSET(node), len)
997 #define Set_Node_Cur_Length(node, start) \
998 Set_Node_Length(node, RExC_parse - start)
1000 /* Get offsets and lengths */
1001 #define Node_Offset(n) (RExC_offsets[2*(REGNODE_OFFSET(n))-1])
1002 #define Node_Length(n) (RExC_offsets[2*(REGNODE_OFFSET(n))])
1004 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
1005 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (offset)); \
1006 Set_Node_Length_To_R(REGNODE_OFFSET(node), (len)); \
1009 #define Track_Code(code) STMT_START { code } STMT_END
1012 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
1013 #define EXPERIMENTAL_INPLACESCAN
1014 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
1018 Perl_re_printf(pTHX_ const char *fmt, ...)
1022 PerlIO *f= Perl_debug_log;
1023 PERL_ARGS_ASSERT_RE_PRINTF;
1025 result = PerlIO_vprintf(f, fmt, ap);
1031 Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
1035 PerlIO *f= Perl_debug_log;
1036 PERL_ARGS_ASSERT_RE_INDENTF;
1037 va_start(ap, depth);
1038 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
1039 result = PerlIO_vprintf(f, fmt, ap);
1043 #endif /* DEBUGGING */
1045 #define DEBUG_RExC_seen() \
1046 DEBUG_OPTIMISE_MORE_r({ \
1047 Perl_re_printf( aTHX_ "RExC_seen: "); \
1049 if (RExC_seen & REG_ZERO_LEN_SEEN) \
1050 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
1052 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
1053 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
1055 if (RExC_seen & REG_GPOS_SEEN) \
1056 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
1058 if (RExC_seen & REG_RECURSE_SEEN) \
1059 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
1061 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
1062 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
1064 if (RExC_seen & REG_VERBARG_SEEN) \
1065 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
1067 if (RExC_seen & REG_CUTGROUP_SEEN) \
1068 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
1070 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
1071 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
1073 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
1074 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
1076 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
1077 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
1079 Perl_re_printf( aTHX_ "\n"); \
1082 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
1083 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
1088 S_debug_show_study_flags(pTHX_ U32 flags, const char *open_str,
1089 const char *close_str)
1094 Perl_re_printf( aTHX_ "%s", open_str);
1095 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_SEOL);
1096 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_MEOL);
1097 DEBUG_SHOW_STUDY_FLAG(flags, SF_IS_INF);
1098 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_PAR);
1099 DEBUG_SHOW_STUDY_FLAG(flags, SF_IN_PAR);
1100 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_EVAL);
1101 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_SUBSTR);
1102 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_AND);
1103 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_OR);
1104 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS);
1105 DEBUG_SHOW_STUDY_FLAG(flags, SCF_WHILEM_VISITED_POS);
1106 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_RESTUDY);
1107 DEBUG_SHOW_STUDY_FLAG(flags, SCF_SEEN_ACCEPT);
1108 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_DOING_RESTUDY);
1109 DEBUG_SHOW_STUDY_FLAG(flags, SCF_IN_DEFINE);
1110 Perl_re_printf( aTHX_ "%s", close_str);
1115 S_debug_studydata(pTHX_ const char *where, scan_data_t *data,
1116 U32 depth, int is_inf)
1118 GET_RE_DEBUG_FLAGS_DECL;
1120 DEBUG_OPTIMISE_MORE_r({
1123 Perl_re_indentf(aTHX_ "%s: Pos:%" IVdf "/%" IVdf " Flags: 0x%" UVXf,
1127 (IV)data->pos_delta,
1131 S_debug_show_study_flags(aTHX_ data->flags," [","]");
1133 Perl_re_printf( aTHX_
1134 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s",
1136 (IV)(data->last_closep ? *((data)->last_closep) : -1),
1137 is_inf ? "INF " : ""
1140 if (data->last_found) {
1142 Perl_re_printf(aTHX_
1143 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf,
1144 SvPVX_const(data->last_found),
1146 (IV)data->last_start_min,
1147 (IV)data->last_start_max
1150 for (i = 0; i < 2; i++) {
1151 Perl_re_printf(aTHX_
1152 " %s%s: '%s' @ %" IVdf "/%" IVdf,
1153 data->cur_is_floating == i ? "*" : "",
1154 i ? "Float" : "Fixed",
1155 SvPVX_const(data->substrs[i].str),
1156 (IV)data->substrs[i].min_offset,
1157 (IV)data->substrs[i].max_offset
1159 S_debug_show_study_flags(aTHX_ data->substrs[i].flags," [","]");
1163 Perl_re_printf( aTHX_ "\n");
1169 S_debug_peep(pTHX_ const char *str, const RExC_state_t *pRExC_state,
1170 regnode *scan, U32 depth, U32 flags)
1172 GET_RE_DEBUG_FLAGS_DECL;
1179 Next = regnext(scan);
1180 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
1181 Perl_re_indentf( aTHX_ "%s>%3d: %s (%d)",
1184 REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),
1185 Next ? (REG_NODE_NUM(Next)) : 0 );
1186 S_debug_show_study_flags(aTHX_ flags," [ ","]");
1187 Perl_re_printf( aTHX_ "\n");
1192 # define DEBUG_STUDYDATA(where, data, depth, is_inf) \
1193 S_debug_studydata(aTHX_ where, data, depth, is_inf)
1195 # define DEBUG_PEEP(str, scan, depth, flags) \
1196 S_debug_peep(aTHX_ str, pRExC_state, scan, depth, flags)
1199 # define DEBUG_STUDYDATA(where, data, depth, is_inf) NOOP
1200 # define DEBUG_PEEP(str, scan, depth, flags) NOOP
1204 /* =========================================================
1205 * BEGIN edit_distance stuff.
1207 * This calculates how many single character changes of any type are needed to
1208 * transform a string into another one. It is taken from version 3.1 of
1210 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1213 /* Our unsorted dictionary linked list. */
1214 /* Note we use UVs, not chars. */
1219 struct dictionary* next;
1221 typedef struct dictionary item;
1224 PERL_STATIC_INLINE item*
1225 push(UV key, item* curr)
1228 Newx(head, 1, item);
1236 PERL_STATIC_INLINE item*
1237 find(item* head, UV key)
1239 item* iterator = head;
1241 if (iterator->key == key){
1244 iterator = iterator->next;
1250 PERL_STATIC_INLINE item*
1251 uniquePush(item* head, UV key)
1253 item* iterator = head;
1256 if (iterator->key == key) {
1259 iterator = iterator->next;
1262 return push(key, head);
1265 PERL_STATIC_INLINE void
1266 dict_free(item* head)
1268 item* iterator = head;
1271 item* temp = iterator;
1272 iterator = iterator->next;
1279 /* End of Dictionary Stuff */
1281 /* All calculations/work are done here */
1283 S_edit_distance(const UV* src,
1285 const STRLEN x, /* length of src[] */
1286 const STRLEN y, /* length of tgt[] */
1287 const SSize_t maxDistance
1291 UV swapCount, swapScore, targetCharCount, i, j;
1293 UV score_ceil = x + y;
1295 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1297 /* intialize matrix start values */
1298 Newx(scores, ( (x + 2) * (y + 2)), UV);
1299 scores[0] = score_ceil;
1300 scores[1 * (y + 2) + 0] = score_ceil;
1301 scores[0 * (y + 2) + 1] = score_ceil;
1302 scores[1 * (y + 2) + 1] = 0;
1303 head = uniquePush(uniquePush(head, src[0]), tgt[0]);
1308 for (i=1;i<=x;i++) {
1310 head = uniquePush(head, src[i]);
1311 scores[(i+1) * (y + 2) + 1] = i;
1312 scores[(i+1) * (y + 2) + 0] = score_ceil;
1315 for (j=1;j<=y;j++) {
1318 head = uniquePush(head, tgt[j]);
1319 scores[1 * (y + 2) + (j + 1)] = j;
1320 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1323 targetCharCount = find(head, tgt[j-1])->value;
1324 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1326 if (src[i-1] != tgt[j-1]){
1327 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));
1331 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1335 find(head, src[i-1])->value = i;
1339 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1342 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1346 /* END of edit_distance() stuff
1347 * ========================================================= */
1349 /* is c a control character for which we have a mnemonic? */
1350 #define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
1353 S_cntrl_to_mnemonic(const U8 c)
1355 /* Returns the mnemonic string that represents character 'c', if one
1356 * exists; NULL otherwise. The only ones that exist for the purposes of
1357 * this routine are a few control characters */
1360 case '\a': return "\\a";
1361 case '\b': return "\\b";
1362 case ESC_NATIVE: return "\\e";
1363 case '\f': return "\\f";
1364 case '\n': return "\\n";
1365 case '\r': return "\\r";
1366 case '\t': return "\\t";
1372 /* Mark that we cannot extend a found fixed substring at this point.
1373 Update the longest found anchored substring or the longest found
1374 floating substrings if needed. */
1377 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1378 SSize_t *minlenp, int is_inf)
1380 const STRLEN l = CHR_SVLEN(data->last_found);
1381 SV * const longest_sv = data->substrs[data->cur_is_floating].str;
1382 const STRLEN old_l = CHR_SVLEN(longest_sv);
1383 GET_RE_DEBUG_FLAGS_DECL;
1385 PERL_ARGS_ASSERT_SCAN_COMMIT;
1387 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
1388 const U8 i = data->cur_is_floating;
1389 SvSetMagicSV(longest_sv, data->last_found);
1390 data->substrs[i].min_offset = l ? data->last_start_min : data->pos_min;
1393 data->substrs[0].max_offset = data->substrs[0].min_offset;
1395 data->substrs[1].max_offset = (l
1396 ? data->last_start_max
1397 : (data->pos_delta > SSize_t_MAX - data->pos_min
1399 : data->pos_min + data->pos_delta));
1401 || (STRLEN)data->substrs[1].max_offset > (STRLEN)SSize_t_MAX)
1402 data->substrs[1].max_offset = SSize_t_MAX;
1405 if (data->flags & SF_BEFORE_EOL)
1406 data->substrs[i].flags |= (data->flags & SF_BEFORE_EOL);
1408 data->substrs[i].flags &= ~SF_BEFORE_EOL;
1409 data->substrs[i].minlenp = minlenp;
1410 data->substrs[i].lookbehind = 0;
1413 SvCUR_set(data->last_found, 0);
1415 SV * const sv = data->last_found;
1416 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1417 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1422 data->last_end = -1;
1423 data->flags &= ~SF_BEFORE_EOL;
1424 DEBUG_STUDYDATA("commit", data, 0, is_inf);
1427 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1428 * list that describes which code points it matches */
1431 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1433 /* Set the SSC 'ssc' to match an empty string or any code point */
1435 PERL_ARGS_ASSERT_SSC_ANYTHING;
1437 assert(is_ANYOF_SYNTHETIC(ssc));
1439 /* mortalize so won't leak */
1440 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
1441 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1445 S_ssc_is_anything(const regnode_ssc *ssc)
1447 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1448 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1449 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1450 * in any way, so there's no point in using it */
1455 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1457 assert(is_ANYOF_SYNTHETIC(ssc));
1459 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1463 /* See if the list consists solely of the range 0 - Infinity */
1464 invlist_iterinit(ssc->invlist);
1465 ret = invlist_iternext(ssc->invlist, &start, &end)
1469 invlist_iterfinish(ssc->invlist);
1475 /* If e.g., both \w and \W are set, matches everything */
1476 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1478 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1479 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1489 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1491 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1492 * string, any code point, or any posix class under locale */
1494 PERL_ARGS_ASSERT_SSC_INIT;
1496 Zero(ssc, 1, regnode_ssc);
1497 set_ANYOF_SYNTHETIC(ssc);
1498 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1501 /* If any portion of the regex is to operate under locale rules that aren't
1502 * fully known at compile time, initialization includes it. The reason
1503 * this isn't done for all regexes is that the optimizer was written under
1504 * the assumption that locale was all-or-nothing. Given the complexity and
1505 * lack of documentation in the optimizer, and that there are inadequate
1506 * test cases for locale, many parts of it may not work properly, it is
1507 * safest to avoid locale unless necessary. */
1508 if (RExC_contains_locale) {
1509 ANYOF_POSIXL_SETALL(ssc);
1512 ANYOF_POSIXL_ZERO(ssc);
1517 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1518 const regnode_ssc *ssc)
1520 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1521 * to the list of code points matched, and locale posix classes; hence does
1522 * not check its flags) */
1527 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1529 assert(is_ANYOF_SYNTHETIC(ssc));
1531 invlist_iterinit(ssc->invlist);
1532 ret = invlist_iternext(ssc->invlist, &start, &end)
1536 invlist_iterfinish(ssc->invlist);
1542 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1549 #define INVLIST_INDEX 0
1550 #define ONLY_LOCALE_MATCHES_INDEX 1
1551 #define DEFERRED_USER_DEFINED_INDEX 2
1554 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1555 const regnode_charclass* const node)
1557 /* Returns a mortal inversion list defining which code points are matched
1558 * by 'node', which is of type ANYOF. Handles complementing the result if
1559 * appropriate. If some code points aren't knowable at this time, the
1560 * returned list must, and will, contain every code point that is a
1564 SV* only_utf8_locale_invlist = NULL;
1566 const U32 n = ARG(node);
1567 bool new_node_has_latin1 = FALSE;
1569 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1571 /* Look at the data structure created by S_set_ANYOF_arg() */
1572 if (n != ANYOF_ONLY_HAS_BITMAP) {
1573 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1574 AV * const av = MUTABLE_AV(SvRV(rv));
1575 SV **const ary = AvARRAY(av);
1576 assert(RExC_rxi->data->what[n] == 's');
1578 if (av_tindex_skip_len_mg(av) >= DEFERRED_USER_DEFINED_INDEX) {
1580 /* Here there are things that won't be known until runtime -- we
1581 * have to assume it could be anything */
1582 invlist = sv_2mortal(_new_invlist(1));
1583 return _add_range_to_invlist(invlist, 0, UV_MAX);
1585 else if (ary[INVLIST_INDEX]) {
1587 /* Use the node's inversion list */
1588 invlist = sv_2mortal(invlist_clone(ary[INVLIST_INDEX], NULL));
1591 /* Get the code points valid only under UTF-8 locales */
1592 if ( (ANYOF_FLAGS(node) & ANYOFL_FOLD)
1593 && av_tindex_skip_len_mg(av) >= ONLY_LOCALE_MATCHES_INDEX)
1595 only_utf8_locale_invlist = ary[ONLY_LOCALE_MATCHES_INDEX];
1600 invlist = sv_2mortal(_new_invlist(0));
1603 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1604 * code points, and an inversion list for the others, but if there are code
1605 * points that should match only conditionally on the target string being
1606 * UTF-8, those are placed in the inversion list, and not the bitmap.
1607 * Since there are circumstances under which they could match, they are
1608 * included in the SSC. But if the ANYOF node is to be inverted, we have
1609 * to exclude them here, so that when we invert below, the end result
1610 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1611 * have to do this here before we add the unconditionally matched code
1613 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1614 _invlist_intersection_complement_2nd(invlist,
1619 /* Add in the points from the bit map */
1620 if (OP(node) != ANYOFH) {
1621 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1622 if (ANYOF_BITMAP_TEST(node, i)) {
1623 unsigned int start = i++;
1625 for (; i < NUM_ANYOF_CODE_POINTS
1626 && ANYOF_BITMAP_TEST(node, i); ++i)
1630 invlist = _add_range_to_invlist(invlist, start, i-1);
1631 new_node_has_latin1 = TRUE;
1636 /* If this can match all upper Latin1 code points, have to add them
1637 * as well. But don't add them if inverting, as when that gets done below,
1638 * it would exclude all these characters, including the ones it shouldn't
1639 * that were added just above */
1640 if (! (ANYOF_FLAGS(node) & ANYOF_INVERT) && OP(node) == ANYOFD
1641 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1643 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1646 /* Similarly for these */
1647 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1648 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1651 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1652 _invlist_invert(invlist);
1654 else if (ANYOF_FLAGS(node) & ANYOFL_FOLD) {
1655 if (new_node_has_latin1) {
1657 /* Under /li, any 0-255 could fold to any other 0-255, depending on
1658 * the locale. We can skip this if there are no 0-255 at all. */
1659 _invlist_union(invlist, PL_Latin1, &invlist);
1661 invlist = add_cp_to_invlist(invlist, LATIN_SMALL_LETTER_DOTLESS_I);
1662 invlist = add_cp_to_invlist(invlist, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
1665 if (_invlist_contains_cp(invlist, LATIN_SMALL_LETTER_DOTLESS_I)) {
1666 invlist = add_cp_to_invlist(invlist, 'I');
1668 if (_invlist_contains_cp(invlist,
1669 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE))
1671 invlist = add_cp_to_invlist(invlist, 'i');
1676 /* Similarly add the UTF-8 locale possible matches. These have to be
1677 * deferred until after the non-UTF-8 locale ones are taken care of just
1678 * above, or it leads to wrong results under ANYOF_INVERT */
1679 if (only_utf8_locale_invlist) {
1680 _invlist_union_maybe_complement_2nd(invlist,
1681 only_utf8_locale_invlist,
1682 ANYOF_FLAGS(node) & ANYOF_INVERT,
1689 /* These two functions currently do the exact same thing */
1690 #define ssc_init_zero ssc_init
1692 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1693 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1695 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1696 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1697 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1700 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1701 const regnode_charclass *and_with)
1703 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1704 * another SSC or a regular ANYOF class. Can create false positives. */
1709 PERL_ARGS_ASSERT_SSC_AND;
1711 assert(is_ANYOF_SYNTHETIC(ssc));
1713 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1714 * the code point inversion list and just the relevant flags */
1715 if (is_ANYOF_SYNTHETIC(and_with)) {
1716 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1717 anded_flags = ANYOF_FLAGS(and_with);
1719 /* XXX This is a kludge around what appears to be deficiencies in the
1720 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1721 * there are paths through the optimizer where it doesn't get weeded
1722 * out when it should. And if we don't make some extra provision for
1723 * it like the code just below, it doesn't get added when it should.
1724 * This solution is to add it only when AND'ing, which is here, and
1725 * only when what is being AND'ed is the pristine, original node
1726 * matching anything. Thus it is like adding it to ssc_anything() but
1727 * only when the result is to be AND'ed. Probably the same solution
1728 * could be adopted for the same problem we have with /l matching,
1729 * which is solved differently in S_ssc_init(), and that would lead to
1730 * fewer false positives than that solution has. But if this solution
1731 * creates bugs, the consequences are only that a warning isn't raised
1732 * that should be; while the consequences for having /l bugs is
1733 * incorrect matches */
1734 if (ssc_is_anything((regnode_ssc *)and_with)) {
1735 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1739 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1740 if (OP(and_with) == ANYOFD) {
1741 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1744 anded_flags = ANYOF_FLAGS(and_with)
1745 &( ANYOF_COMMON_FLAGS
1746 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1747 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1748 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(and_with))) {
1750 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1755 ANYOF_FLAGS(ssc) &= anded_flags;
1757 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1758 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1759 * 'and_with' may be inverted. When not inverted, we have the situation of
1761 * (C1 | P1) & (C2 | P2)
1762 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1763 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1764 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1765 * <= ((C1 & C2) | P1 | P2)
1766 * Alternatively, the last few steps could be:
1767 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1768 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1769 * <= (C1 | C2 | (P1 & P2))
1770 * We favor the second approach if either P1 or P2 is non-empty. This is
1771 * because these components are a barrier to doing optimizations, as what
1772 * they match cannot be known until the moment of matching as they are
1773 * dependent on the current locale, 'AND"ing them likely will reduce or
1775 * But we can do better if we know that C1,P1 are in their initial state (a
1776 * frequent occurrence), each matching everything:
1777 * (<everything>) & (C2 | P2) = C2 | P2
1778 * Similarly, if C2,P2 are in their initial state (again a frequent
1779 * occurrence), the result is a no-op
1780 * (C1 | P1) & (<everything>) = C1 | P1
1783 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1784 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1785 * <= (C1 & ~C2) | (P1 & ~P2)
1788 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1789 && ! is_ANYOF_SYNTHETIC(and_with))
1793 ssc_intersection(ssc,
1795 FALSE /* Has already been inverted */
1798 /* If either P1 or P2 is empty, the intersection will be also; can skip
1800 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
1801 ANYOF_POSIXL_ZERO(ssc);
1803 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1805 /* Note that the Posix class component P from 'and_with' actually
1807 * P = Pa | Pb | ... | Pn
1808 * where each component is one posix class, such as in [\w\s].
1810 * ~P = ~(Pa | Pb | ... | Pn)
1811 * = ~Pa & ~Pb & ... & ~Pn
1812 * <= ~Pa | ~Pb | ... | ~Pn
1813 * The last is something we can easily calculate, but unfortunately
1814 * is likely to have many false positives. We could do better
1815 * in some (but certainly not all) instances if two classes in
1816 * P have known relationships. For example
1817 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1819 * :lower: & :print: = :lower:
1820 * And similarly for classes that must be disjoint. For example,
1821 * since \s and \w can have no elements in common based on rules in
1822 * the POSIX standard,
1823 * \w & ^\S = nothing
1824 * Unfortunately, some vendor locales do not meet the Posix
1825 * standard, in particular almost everything by Microsoft.
1826 * The loop below just changes e.g., \w into \W and vice versa */
1828 regnode_charclass_posixl temp;
1829 int add = 1; /* To calculate the index of the complement */
1831 Zero(&temp, 1, regnode_charclass_posixl);
1832 ANYOF_POSIXL_ZERO(&temp);
1833 for (i = 0; i < ANYOF_MAX; i++) {
1835 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1836 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1838 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1839 ANYOF_POSIXL_SET(&temp, i + add);
1841 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1843 ANYOF_POSIXL_AND(&temp, ssc);
1845 } /* else ssc already has no posixes */
1846 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1847 in its initial state */
1848 else if (! is_ANYOF_SYNTHETIC(and_with)
1849 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1851 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1852 * copy it over 'ssc' */
1853 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1854 if (is_ANYOF_SYNTHETIC(and_with)) {
1855 StructCopy(and_with, ssc, regnode_ssc);
1858 ssc->invlist = anded_cp_list;
1859 ANYOF_POSIXL_ZERO(ssc);
1860 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1861 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1865 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1866 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
1868 /* One or the other of P1, P2 is non-empty. */
1869 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1870 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1872 ssc_union(ssc, anded_cp_list, FALSE);
1874 else { /* P1 = P2 = empty */
1875 ssc_intersection(ssc, anded_cp_list, FALSE);
1881 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1882 const regnode_charclass *or_with)
1884 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1885 * another SSC or a regular ANYOF class. Can create false positives if
1886 * 'or_with' is to be inverted. */
1891 PERL_ARGS_ASSERT_SSC_OR;
1893 assert(is_ANYOF_SYNTHETIC(ssc));
1895 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1896 * the code point inversion list and just the relevant flags */
1897 if (is_ANYOF_SYNTHETIC(or_with)) {
1898 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1899 ored_flags = ANYOF_FLAGS(or_with);
1902 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1903 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1904 if (OP(or_with) != ANYOFD) {
1906 |= ANYOF_FLAGS(or_with)
1907 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1908 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1909 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(or_with))) {
1911 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1916 ANYOF_FLAGS(ssc) |= ored_flags;
1918 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1919 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1920 * 'or_with' may be inverted. When not inverted, we have the simple
1921 * situation of computing:
1922 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1923 * If P1|P2 yields a situation with both a class and its complement are
1924 * set, like having both \w and \W, this matches all code points, and we
1925 * can delete these from the P component of the ssc going forward. XXX We
1926 * might be able to delete all the P components, but I (khw) am not certain
1927 * about this, and it is better to be safe.
1930 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1931 * <= (C1 | P1) | ~C2
1932 * <= (C1 | ~C2) | P1
1933 * (which results in actually simpler code than the non-inverted case)
1936 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1937 && ! is_ANYOF_SYNTHETIC(or_with))
1939 /* We ignore P2, leaving P1 going forward */
1940 } /* else Not inverted */
1941 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
1942 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1943 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1945 for (i = 0; i < ANYOF_MAX; i += 2) {
1946 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1948 ssc_match_all_cp(ssc);
1949 ANYOF_POSIXL_CLEAR(ssc, i);
1950 ANYOF_POSIXL_CLEAR(ssc, i+1);
1958 FALSE /* Already has been inverted */
1962 PERL_STATIC_INLINE void
1963 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1965 PERL_ARGS_ASSERT_SSC_UNION;
1967 assert(is_ANYOF_SYNTHETIC(ssc));
1969 _invlist_union_maybe_complement_2nd(ssc->invlist,
1975 PERL_STATIC_INLINE void
1976 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1978 const bool invert2nd)
1980 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1982 assert(is_ANYOF_SYNTHETIC(ssc));
1984 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1990 PERL_STATIC_INLINE void
1991 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1993 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1995 assert(is_ANYOF_SYNTHETIC(ssc));
1997 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
2000 PERL_STATIC_INLINE void
2001 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
2003 /* AND just the single code point 'cp' into the SSC 'ssc' */
2005 SV* cp_list = _new_invlist(2);
2007 PERL_ARGS_ASSERT_SSC_CP_AND;
2009 assert(is_ANYOF_SYNTHETIC(ssc));
2011 cp_list = add_cp_to_invlist(cp_list, cp);
2012 ssc_intersection(ssc, cp_list,
2013 FALSE /* Not inverted */
2015 SvREFCNT_dec_NN(cp_list);
2018 PERL_STATIC_INLINE void
2019 S_ssc_clear_locale(regnode_ssc *ssc)
2021 /* Set the SSC 'ssc' to not match any locale things */
2022 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
2024 assert(is_ANYOF_SYNTHETIC(ssc));
2026 ANYOF_POSIXL_ZERO(ssc);
2027 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
2030 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
2033 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
2035 /* The synthetic start class is used to hopefully quickly winnow down
2036 * places where a pattern could start a match in the target string. If it
2037 * doesn't really narrow things down that much, there isn't much point to
2038 * having the overhead of using it. This function uses some very crude
2039 * heuristics to decide if to use the ssc or not.
2041 * It returns TRUE if 'ssc' rules out more than half what it considers to
2042 * be the "likely" possible matches, but of course it doesn't know what the
2043 * actual things being matched are going to be; these are only guesses
2045 * For /l matches, it assumes that the only likely matches are going to be
2046 * in the 0-255 range, uniformly distributed, so half of that is 127
2047 * For /a and /d matches, it assumes that the likely matches will be just
2048 * the ASCII range, so half of that is 63
2049 * For /u and there isn't anything matching above the Latin1 range, it
2050 * assumes that that is the only range likely to be matched, and uses
2051 * half that as the cut-off: 127. If anything matches above Latin1,
2052 * it assumes that all of Unicode could match (uniformly), except for
2053 * non-Unicode code points and things in the General Category "Other"
2054 * (unassigned, private use, surrogates, controls and formats). This
2055 * is a much large number. */
2057 U32 count = 0; /* Running total of number of code points matched by
2059 UV start, end; /* Start and end points of current range in inversion
2060 XXX outdated. UTF-8 locales are common, what about invert? list */
2061 const U32 max_code_points = (LOC)
2063 : (( ! UNI_SEMANTICS
2064 || invlist_highest(ssc->invlist) < 256)
2067 const U32 max_match = max_code_points / 2;
2069 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
2071 invlist_iterinit(ssc->invlist);
2072 while (invlist_iternext(ssc->invlist, &start, &end)) {
2073 if (start >= max_code_points) {
2076 end = MIN(end, max_code_points - 1);
2077 count += end - start + 1;
2078 if (count >= max_match) {
2079 invlist_iterfinish(ssc->invlist);
2089 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
2091 /* The inversion list in the SSC is marked mortal; now we need a more
2092 * permanent copy, which is stored the same way that is done in a regular
2093 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
2096 SV* invlist = invlist_clone(ssc->invlist, NULL);
2098 PERL_ARGS_ASSERT_SSC_FINALIZE;
2100 assert(is_ANYOF_SYNTHETIC(ssc));
2102 /* The code in this file assumes that all but these flags aren't relevant
2103 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
2104 * by the time we reach here */
2105 assert(! (ANYOF_FLAGS(ssc)
2106 & ~( ANYOF_COMMON_FLAGS
2107 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
2108 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
2110 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
2112 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist, NULL, NULL);
2114 /* Make sure is clone-safe */
2115 ssc->invlist = NULL;
2117 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
2118 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
2119 OP(ssc) = ANYOFPOSIXL;
2121 else if (RExC_contains_locale) {
2125 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
2128 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
2129 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
2130 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
2131 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
2132 ? (TRIE_LIST_CUR( idx ) - 1) \
2138 dump_trie(trie,widecharmap,revcharmap)
2139 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
2140 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
2142 These routines dump out a trie in a somewhat readable format.
2143 The _interim_ variants are used for debugging the interim
2144 tables that are used to generate the final compressed
2145 representation which is what dump_trie expects.
2147 Part of the reason for their existence is to provide a form
2148 of documentation as to how the different representations function.
2153 Dumps the final compressed table form of the trie to Perl_debug_log.
2154 Used for debugging make_trie().
2158 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
2159 AV *revcharmap, U32 depth)
2162 SV *sv=sv_newmortal();
2163 int colwidth= widecharmap ? 6 : 4;
2165 GET_RE_DEBUG_FLAGS_DECL;
2167 PERL_ARGS_ASSERT_DUMP_TRIE;
2169 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
2170 depth+1, "Match","Base","Ofs" );
2172 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2173 SV ** const tmp = av_fetch( revcharmap, state, 0);
2175 Perl_re_printf( aTHX_ "%*s",
2177 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2178 PL_colors[0], PL_colors[1],
2179 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2180 PERL_PV_ESCAPE_FIRSTCHAR
2185 Perl_re_printf( aTHX_ "\n");
2186 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
2188 for( state = 0 ; state < trie->uniquecharcount ; state++ )
2189 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2190 Perl_re_printf( aTHX_ "\n");
2192 for( state = 1 ; state < trie->statecount ; state++ ) {
2193 const U32 base = trie->states[ state ].trans.base;
2195 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
2197 if ( trie->states[ state ].wordnum ) {
2198 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
2200 Perl_re_printf( aTHX_ "%6s", "" );
2203 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
2208 while( ( base + ofs < trie->uniquecharcount ) ||
2209 ( base + ofs - trie->uniquecharcount < trie->lasttrans
2210 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2214 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
2216 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2217 if ( ( base + ofs >= trie->uniquecharcount )
2218 && ( base + ofs - trie->uniquecharcount
2220 && trie->trans[ base + ofs
2221 - trie->uniquecharcount ].check == state )
2223 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
2224 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2227 Perl_re_printf( aTHX_ "%*s", colwidth," ." );
2231 Perl_re_printf( aTHX_ "]");
2234 Perl_re_printf( aTHX_ "\n" );
2236 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
2238 for (word=1; word <= trie->wordcount; word++) {
2239 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2240 (int)word, (int)(trie->wordinfo[word].prev),
2241 (int)(trie->wordinfo[word].len));
2243 Perl_re_printf( aTHX_ "\n" );
2246 Dumps a fully constructed but uncompressed trie in list form.
2247 List tries normally only are used for construction when the number of
2248 possible chars (trie->uniquecharcount) is very high.
2249 Used for debugging make_trie().
2252 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2253 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2257 SV *sv=sv_newmortal();
2258 int colwidth= widecharmap ? 6 : 4;
2259 GET_RE_DEBUG_FLAGS_DECL;
2261 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2263 /* print out the table precompression. */
2264 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
2266 Perl_re_indentf( aTHX_ "%s",
2267 depth+1, "------:-----+-----------------\n" );
2269 for( state=1 ; state < next_alloc ; state ++ ) {
2272 Perl_re_indentf( aTHX_ " %4" UVXf " :",
2273 depth+1, (UV)state );
2274 if ( ! trie->states[ state ].wordnum ) {
2275 Perl_re_printf( aTHX_ "%5s| ","");
2277 Perl_re_printf( aTHX_ "W%4x| ",
2278 trie->states[ state ].wordnum
2281 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2282 SV ** const tmp = av_fetch( revcharmap,
2283 TRIE_LIST_ITEM(state, charid).forid, 0);
2285 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
2287 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2289 PL_colors[0], PL_colors[1],
2290 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2291 | PERL_PV_ESCAPE_FIRSTCHAR
2293 TRIE_LIST_ITEM(state, charid).forid,
2294 (UV)TRIE_LIST_ITEM(state, charid).newstate
2297 Perl_re_printf( aTHX_ "\n%*s| ",
2298 (int)((depth * 2) + 14), "");
2301 Perl_re_printf( aTHX_ "\n");
2306 Dumps a fully constructed but uncompressed trie in table form.
2307 This is the normal DFA style state transition table, with a few
2308 twists to facilitate compression later.
2309 Used for debugging make_trie().
2312 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2313 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2318 SV *sv=sv_newmortal();
2319 int colwidth= widecharmap ? 6 : 4;
2320 GET_RE_DEBUG_FLAGS_DECL;
2322 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2325 print out the table precompression so that we can do a visual check
2326 that they are identical.
2329 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
2331 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2332 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2334 Perl_re_printf( aTHX_ "%*s",
2336 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2337 PL_colors[0], PL_colors[1],
2338 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2339 PERL_PV_ESCAPE_FIRSTCHAR
2345 Perl_re_printf( aTHX_ "\n");
2346 Perl_re_indentf( aTHX_ "State+-", depth+1 );
2348 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2349 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
2352 Perl_re_printf( aTHX_ "\n" );
2354 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2356 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
2358 (UV)TRIE_NODENUM( state ) );
2360 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2361 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2363 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
2365 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
2367 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2368 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
2369 (UV)trie->trans[ state ].check );
2371 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
2372 (UV)trie->trans[ state ].check,
2373 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2381 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2382 startbranch: the first branch in the whole branch sequence
2383 first : start branch of sequence of branch-exact nodes.
2384 May be the same as startbranch
2385 last : Thing following the last branch.
2386 May be the same as tail.
2387 tail : item following the branch sequence
2388 count : words in the sequence
2389 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2390 depth : indent depth
2392 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2394 A trie is an N'ary tree where the branches are determined by digital
2395 decomposition of the key. IE, at the root node you look up the 1st character and
2396 follow that branch repeat until you find the end of the branches. Nodes can be
2397 marked as "accepting" meaning they represent a complete word. Eg:
2401 would convert into the following structure. Numbers represent states, letters
2402 following numbers represent valid transitions on the letter from that state, if
2403 the number is in square brackets it represents an accepting state, otherwise it
2404 will be in parenthesis.
2406 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2410 (1) +-i->(6)-+-s->[7]
2412 +-s->(3)-+-h->(4)-+-e->[5]
2414 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2416 This shows that when matching against the string 'hers' we will begin at state 1
2417 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2418 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2419 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2420 single traverse. We store a mapping from accepting to state to which word was
2421 matched, and then when we have multiple possibilities we try to complete the
2422 rest of the regex in the order in which they occurred in the alternation.
2424 The only prior NFA like behaviour that would be changed by the TRIE support is
2425 the silent ignoring of duplicate alternations which are of the form:
2427 / (DUPE|DUPE) X? (?{ ... }) Y /x
2429 Thus EVAL blocks following a trie may be called a different number of times with
2430 and without the optimisation. With the optimisations dupes will be silently
2431 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2432 the following demonstrates:
2434 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2436 which prints out 'word' three times, but
2438 'words'=~/(word|word|word)(?{ print $1 })S/
2440 which doesnt print it out at all. This is due to other optimisations kicking in.
2442 Example of what happens on a structural level:
2444 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2446 1: CURLYM[1] {1,32767}(18)
2457 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2458 and should turn into:
2460 1: CURLYM[1] {1,32767}(18)
2462 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2470 Cases where tail != last would be like /(?foo|bar)baz/:
2480 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2481 and would end up looking like:
2484 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2491 d = uvchr_to_utf8_flags(d, uv, 0);
2493 is the recommended Unicode-aware way of saying
2498 #define TRIE_STORE_REVCHAR(val) \
2501 SV *zlopp = newSV(UTF8_MAXBYTES); \
2502 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2503 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2504 SvCUR_set(zlopp, kapow - flrbbbbb); \
2507 av_push(revcharmap, zlopp); \
2509 char ooooff = (char)val; \
2510 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2514 /* This gets the next character from the input, folding it if not already
2516 #define TRIE_READ_CHAR STMT_START { \
2519 /* if it is UTF then it is either already folded, or does not need \
2521 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2523 else if (folder == PL_fold_latin1) { \
2524 /* This folder implies Unicode rules, which in the range expressible \
2525 * by not UTF is the lower case, with the two exceptions, one of \
2526 * which should have been taken care of before calling this */ \
2527 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2528 uvc = toLOWER_L1(*uc); \
2529 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2532 /* raw data, will be folded later if needed */ \
2540 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2541 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2542 U32 ging = TRIE_LIST_LEN( state ) * 2; \
2543 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2544 TRIE_LIST_LEN( state ) = ging; \
2546 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2547 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2548 TRIE_LIST_CUR( state )++; \
2551 #define TRIE_LIST_NEW(state) STMT_START { \
2552 Newx( trie->states[ state ].trans.list, \
2553 4, reg_trie_trans_le ); \
2554 TRIE_LIST_CUR( state ) = 1; \
2555 TRIE_LIST_LEN( state ) = 4; \
2558 #define TRIE_HANDLE_WORD(state) STMT_START { \
2559 U16 dupe= trie->states[ state ].wordnum; \
2560 regnode * const noper_next = regnext( noper ); \
2563 /* store the word for dumping */ \
2565 if (OP(noper) != NOTHING) \
2566 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2568 tmp = newSVpvn_utf8( "", 0, UTF ); \
2569 av_push( trie_words, tmp ); \
2573 trie->wordinfo[curword].prev = 0; \
2574 trie->wordinfo[curword].len = wordlen; \
2575 trie->wordinfo[curword].accept = state; \
2577 if ( noper_next < tail ) { \
2579 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2581 trie->jump[curword] = (U16)(noper_next - convert); \
2583 jumper = noper_next; \
2585 nextbranch= regnext(cur); \
2589 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2590 /* chain, so that when the bits of chain are later */\
2591 /* linked together, the dups appear in the chain */\
2592 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2593 trie->wordinfo[dupe].prev = curword; \
2595 /* we haven't inserted this word yet. */ \
2596 trie->states[ state ].wordnum = curword; \
2601 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2602 ( ( base + charid >= ucharcount \
2603 && base + charid < ubound \
2604 && state == trie->trans[ base - ucharcount + charid ].check \
2605 && trie->trans[ base - ucharcount + charid ].next ) \
2606 ? trie->trans[ base - ucharcount + charid ].next \
2607 : ( state==1 ? special : 0 ) \
2610 #define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2612 TRIE_BITMAP_SET(trie, uvc); \
2613 /* store the folded codepoint */ \
2615 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2618 /* store first byte of utf8 representation of */ \
2619 /* variant codepoints */ \
2620 if (! UVCHR_IS_INVARIANT(uvc)) { \
2621 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2626 #define MADE_JUMP_TRIE 2
2627 #define MADE_EXACT_TRIE 4
2630 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2631 regnode *first, regnode *last, regnode *tail,
2632 U32 word_count, U32 flags, U32 depth)
2634 /* first pass, loop through and scan words */
2635 reg_trie_data *trie;
2636 HV *widecharmap = NULL;
2637 AV *revcharmap = newAV();
2643 regnode *jumper = NULL;
2644 regnode *nextbranch = NULL;
2645 regnode *convert = NULL;
2646 U32 *prev_states; /* temp array mapping each state to previous one */
2647 /* we just use folder as a flag in utf8 */
2648 const U8 * folder = NULL;
2650 /* in the below add_data call we are storing either 'tu' or 'tuaa'
2651 * which stands for one trie structure, one hash, optionally followed
2654 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuaa"));
2655 AV *trie_words = NULL;
2656 /* along with revcharmap, this only used during construction but both are
2657 * useful during debugging so we store them in the struct when debugging.
2660 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2661 STRLEN trie_charcount=0;
2663 SV *re_trie_maxbuff;
2664 GET_RE_DEBUG_FLAGS_DECL;
2666 PERL_ARGS_ASSERT_MAKE_TRIE;
2668 PERL_UNUSED_ARG(depth);
2672 case EXACT: case EXACT_ONLY8: case EXACTL: break;
2676 case EXACTFLU8: folder = PL_fold_latin1; break;
2677 case EXACTF: folder = PL_fold; break;
2678 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2681 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2683 trie->startstate = 1;
2684 trie->wordcount = word_count;
2685 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2686 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2687 if (flags == EXACT || flags == EXACT_ONLY8 || flags == EXACTL)
2688 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2689 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2690 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2693 trie_words = newAV();
2696 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2697 assert(re_trie_maxbuff);
2698 if (!SvIOK(re_trie_maxbuff)) {
2699 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2701 DEBUG_TRIE_COMPILE_r({
2702 Perl_re_indentf( aTHX_
2703 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2705 REG_NODE_NUM(startbranch), REG_NODE_NUM(first),
2706 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2709 /* Find the node we are going to overwrite */
2710 if ( first == startbranch && OP( last ) != BRANCH ) {
2711 /* whole branch chain */
2714 /* branch sub-chain */
2715 convert = NEXTOPER( first );
2718 /* -- First loop and Setup --
2720 We first traverse the branches and scan each word to determine if it
2721 contains widechars, and how many unique chars there are, this is
2722 important as we have to build a table with at least as many columns as we
2725 We use an array of integers to represent the character codes 0..255
2726 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2727 the native representation of the character value as the key and IV's for
2730 *TODO* If we keep track of how many times each character is used we can
2731 remap the columns so that the table compression later on is more
2732 efficient in terms of memory by ensuring the most common value is in the
2733 middle and the least common are on the outside. IMO this would be better
2734 than a most to least common mapping as theres a decent chance the most
2735 common letter will share a node with the least common, meaning the node
2736 will not be compressible. With a middle is most common approach the worst
2737 case is when we have the least common nodes twice.
2741 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2742 regnode *noper = NEXTOPER( cur );
2746 U32 wordlen = 0; /* required init */
2747 STRLEN minchars = 0;
2748 STRLEN maxchars = 0;
2749 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2752 if (OP(noper) == NOTHING) {
2753 /* skip past a NOTHING at the start of an alternation
2754 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2756 regnode *noper_next= regnext(noper);
2757 if (noper_next < tail)
2762 && ( OP(noper) == flags
2763 || (flags == EXACT && OP(noper) == EXACT_ONLY8)
2764 || (flags == EXACTFU && ( OP(noper) == EXACTFU_ONLY8
2765 || OP(noper) == EXACTFUP))))
2767 uc= (U8*)STRING(noper);
2768 e= uc + STR_LEN(noper);
2775 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2776 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2777 regardless of encoding */
2778 if (OP( noper ) == EXACTFUP) {
2779 /* false positives are ok, so just set this */
2780 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2784 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2786 TRIE_CHARCOUNT(trie)++;
2789 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2790 * is in effect. Under /i, this character can match itself, or
2791 * anything that folds to it. If not under /i, it can match just
2792 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2793 * all fold to k, and all are single characters. But some folds
2794 * expand to more than one character, so for example LATIN SMALL
2795 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2796 * the string beginning at 'uc' is 'ffi', it could be matched by
2797 * three characters, or just by the one ligature character. (It
2798 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2799 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2800 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2801 * match.) The trie needs to know the minimum and maximum number
2802 * of characters that could match so that it can use size alone to
2803 * quickly reject many match attempts. The max is simple: it is
2804 * the number of folded characters in this branch (since a fold is
2805 * never shorter than what folds to it. */
2809 /* And the min is equal to the max if not under /i (indicated by
2810 * 'folder' being NULL), or there are no multi-character folds. If
2811 * there is a multi-character fold, the min is incremented just
2812 * once, for the character that folds to the sequence. Each
2813 * character in the sequence needs to be added to the list below of
2814 * characters in the trie, but we count only the first towards the
2815 * min number of characters needed. This is done through the
2816 * variable 'foldlen', which is returned by the macros that look
2817 * for these sequences as the number of bytes the sequence
2818 * occupies. Each time through the loop, we decrement 'foldlen' by
2819 * how many bytes the current char occupies. Only when it reaches
2820 * 0 do we increment 'minchars' or look for another multi-character
2822 if (folder == NULL) {
2825 else if (foldlen > 0) {
2826 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2831 /* See if *uc is the beginning of a multi-character fold. If
2832 * so, we decrement the length remaining to look at, to account
2833 * for the current character this iteration. (We can use 'uc'
2834 * instead of the fold returned by TRIE_READ_CHAR because for
2835 * non-UTF, the latin1_safe macro is smart enough to account
2836 * for all the unfolded characters, and because for UTF, the
2837 * string will already have been folded earlier in the
2838 * compilation process */
2840 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2841 foldlen -= UTF8SKIP(uc);
2844 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2849 /* The current character (and any potential folds) should be added
2850 * to the possible matching characters for this position in this
2854 U8 folded= folder[ (U8) uvc ];
2855 if ( !trie->charmap[ folded ] ) {
2856 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2857 TRIE_STORE_REVCHAR( folded );
2860 if ( !trie->charmap[ uvc ] ) {
2861 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2862 TRIE_STORE_REVCHAR( uvc );
2865 /* store the codepoint in the bitmap, and its folded
2867 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
2868 set_bit = 0; /* We've done our bit :-) */
2872 /* XXX We could come up with the list of code points that fold
2873 * to this using PL_utf8_foldclosures, except not for
2874 * multi-char folds, as there may be multiple combinations
2875 * there that could work, which needs to wait until runtime to
2876 * resolve (The comment about LIGATURE FFI above is such an
2881 widecharmap = newHV();
2883 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2886 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
2888 if ( !SvTRUE( *svpp ) ) {
2889 sv_setiv( *svpp, ++trie->uniquecharcount );
2890 TRIE_STORE_REVCHAR(uvc);
2893 } /* end loop through characters in this branch of the trie */
2895 /* We take the min and max for this branch and combine to find the min
2896 * and max for all branches processed so far */
2897 if( cur == first ) {
2898 trie->minlen = minchars;
2899 trie->maxlen = maxchars;
2900 } else if (minchars < trie->minlen) {
2901 trie->minlen = minchars;
2902 } else if (maxchars > trie->maxlen) {
2903 trie->maxlen = maxchars;
2905 } /* end first pass */
2906 DEBUG_TRIE_COMPILE_r(
2907 Perl_re_indentf( aTHX_
2908 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2910 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2911 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2912 (int)trie->minlen, (int)trie->maxlen )
2916 We now know what we are dealing with in terms of unique chars and
2917 string sizes so we can calculate how much memory a naive
2918 representation using a flat table will take. If it's over a reasonable
2919 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2920 conservative but potentially much slower representation using an array
2923 At the end we convert both representations into the same compressed
2924 form that will be used in regexec.c for matching with. The latter
2925 is a form that cannot be used to construct with but has memory
2926 properties similar to the list form and access properties similar
2927 to the table form making it both suitable for fast searches and
2928 small enough that its feasable to store for the duration of a program.
2930 See the comment in the code where the compressed table is produced
2931 inplace from the flat tabe representation for an explanation of how
2932 the compression works.
2937 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2940 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2941 > SvIV(re_trie_maxbuff) )
2944 Second Pass -- Array Of Lists Representation
2946 Each state will be represented by a list of charid:state records
2947 (reg_trie_trans_le) the first such element holds the CUR and LEN
2948 points of the allocated array. (See defines above).
2950 We build the initial structure using the lists, and then convert
2951 it into the compressed table form which allows faster lookups
2952 (but cant be modified once converted).
2955 STRLEN transcount = 1;
2957 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
2960 trie->states = (reg_trie_state *)
2961 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2962 sizeof(reg_trie_state) );
2966 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2968 regnode *noper = NEXTOPER( cur );
2969 U32 state = 1; /* required init */
2970 U16 charid = 0; /* sanity init */
2971 U32 wordlen = 0; /* required init */
2973 if (OP(noper) == NOTHING) {
2974 regnode *noper_next= regnext(noper);
2975 if (noper_next < tail)
2980 && ( OP(noper) == flags
2981 || (flags == EXACT && OP(noper) == EXACT_ONLY8)
2982 || (flags == EXACTFU && ( OP(noper) == EXACTFU_ONLY8
2983 || OP(noper) == EXACTFUP))))
2985 const U8 *uc= (U8*)STRING(noper);
2986 const U8 *e= uc + STR_LEN(noper);
2988 for ( ; uc < e ; uc += len ) {
2993 charid = trie->charmap[ uvc ];
2995 SV** const svpp = hv_fetch( widecharmap,
3002 charid=(U16)SvIV( *svpp );
3005 /* charid is now 0 if we dont know the char read, or
3006 * nonzero if we do */
3013 if ( !trie->states[ state ].trans.list ) {
3014 TRIE_LIST_NEW( state );
3017 check <= TRIE_LIST_USED( state );
3020 if ( TRIE_LIST_ITEM( state, check ).forid
3023 newstate = TRIE_LIST_ITEM( state, check ).newstate;
3028 newstate = next_alloc++;
3029 prev_states[newstate] = state;
3030 TRIE_LIST_PUSH( state, charid, newstate );
3035 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3039 TRIE_HANDLE_WORD(state);
3041 } /* end second pass */
3043 /* next alloc is the NEXT state to be allocated */
3044 trie->statecount = next_alloc;
3045 trie->states = (reg_trie_state *)
3046 PerlMemShared_realloc( trie->states,
3048 * sizeof(reg_trie_state) );
3050 /* and now dump it out before we compress it */
3051 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
3052 revcharmap, next_alloc,
3056 trie->trans = (reg_trie_trans *)
3057 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
3064 for( state=1 ; state < next_alloc ; state ++ ) {
3068 DEBUG_TRIE_COMPILE_MORE_r(
3069 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
3073 if (trie->states[state].trans.list) {
3074 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
3078 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3079 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
3080 if ( forid < minid ) {
3082 } else if ( forid > maxid ) {
3086 if ( transcount < tp + maxid - minid + 1) {
3088 trie->trans = (reg_trie_trans *)
3089 PerlMemShared_realloc( trie->trans,
3091 * sizeof(reg_trie_trans) );
3092 Zero( trie->trans + (transcount / 2),
3096 base = trie->uniquecharcount + tp - minid;
3097 if ( maxid == minid ) {
3099 for ( ; zp < tp ; zp++ ) {
3100 if ( ! trie->trans[ zp ].next ) {
3101 base = trie->uniquecharcount + zp - minid;
3102 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
3104 trie->trans[ zp ].check = state;
3110 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
3112 trie->trans[ tp ].check = state;
3117 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3118 const U32 tid = base
3119 - trie->uniquecharcount
3120 + TRIE_LIST_ITEM( state, idx ).forid;
3121 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
3123 trie->trans[ tid ].check = state;
3125 tp += ( maxid - minid + 1 );
3127 Safefree(trie->states[ state ].trans.list);
3130 DEBUG_TRIE_COMPILE_MORE_r(
3131 Perl_re_printf( aTHX_ " base: %d\n",base);
3134 trie->states[ state ].trans.base=base;
3136 trie->lasttrans = tp + 1;
3140 Second Pass -- Flat Table Representation.
3142 we dont use the 0 slot of either trans[] or states[] so we add 1 to
3143 each. We know that we will need Charcount+1 trans at most to store
3144 the data (one row per char at worst case) So we preallocate both
3145 structures assuming worst case.
3147 We then construct the trie using only the .next slots of the entry
3150 We use the .check field of the first entry of the node temporarily
3151 to make compression both faster and easier by keeping track of how
3152 many non zero fields are in the node.
3154 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
3157 There are two terms at use here: state as a TRIE_NODEIDX() which is
3158 a number representing the first entry of the node, and state as a
3159 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
3160 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
3161 if there are 2 entrys per node. eg:
3169 The table is internally in the right hand, idx form. However as we
3170 also have to deal with the states array which is indexed by nodenum
3171 we have to use TRIE_NODENUM() to convert.
3174 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
3177 trie->trans = (reg_trie_trans *)
3178 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
3179 * trie->uniquecharcount + 1,
3180 sizeof(reg_trie_trans) );
3181 trie->states = (reg_trie_state *)
3182 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3183 sizeof(reg_trie_state) );
3184 next_alloc = trie->uniquecharcount + 1;
3187 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3189 regnode *noper = NEXTOPER( cur );
3191 U32 state = 1; /* required init */
3193 U16 charid = 0; /* sanity init */
3194 U32 accept_state = 0; /* sanity init */
3196 U32 wordlen = 0; /* required init */
3198 if (OP(noper) == NOTHING) {
3199 regnode *noper_next= regnext(noper);
3200 if (noper_next < tail)
3205 && ( OP(noper) == flags
3206 || (flags == EXACT && OP(noper) == EXACT_ONLY8)
3207 || (flags == EXACTFU && ( OP(noper) == EXACTFU_ONLY8
3208 || OP(noper) == EXACTFUP))))
3210 const U8 *uc= (U8*)STRING(noper);
3211 const U8 *e= uc + STR_LEN(noper);
3213 for ( ; uc < e ; uc += len ) {
3218 charid = trie->charmap[ uvc ];
3220 SV* const * const svpp = hv_fetch( widecharmap,
3224 charid = svpp ? (U16)SvIV(*svpp) : 0;
3228 if ( !trie->trans[ state + charid ].next ) {
3229 trie->trans[ state + charid ].next = next_alloc;
3230 trie->trans[ state ].check++;
3231 prev_states[TRIE_NODENUM(next_alloc)]
3232 = TRIE_NODENUM(state);
3233 next_alloc += trie->uniquecharcount;
3235 state = trie->trans[ state + charid ].next;
3237 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3239 /* charid is now 0 if we dont know the char read, or
3240 * nonzero if we do */
3243 accept_state = TRIE_NODENUM( state );
3244 TRIE_HANDLE_WORD(accept_state);
3246 } /* end second pass */
3248 /* and now dump it out before we compress it */
3249 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3251 next_alloc, depth+1));
3255 * Inplace compress the table.*
3257 For sparse data sets the table constructed by the trie algorithm will
3258 be mostly 0/FAIL transitions or to put it another way mostly empty.
3259 (Note that leaf nodes will not contain any transitions.)
3261 This algorithm compresses the tables by eliminating most such
3262 transitions, at the cost of a modest bit of extra work during lookup:
3264 - Each states[] entry contains a .base field which indicates the
3265 index in the state[] array wheres its transition data is stored.
3267 - If .base is 0 there are no valid transitions from that node.
3269 - If .base is nonzero then charid is added to it to find an entry in
3272 -If trans[states[state].base+charid].check!=state then the
3273 transition is taken to be a 0/Fail transition. Thus if there are fail
3274 transitions at the front of the node then the .base offset will point
3275 somewhere inside the previous nodes data (or maybe even into a node
3276 even earlier), but the .check field determines if the transition is
3280 The following process inplace converts the table to the compressed
3281 table: We first do not compress the root node 1,and mark all its
3282 .check pointers as 1 and set its .base pointer as 1 as well. This
3283 allows us to do a DFA construction from the compressed table later,
3284 and ensures that any .base pointers we calculate later are greater
3287 - We set 'pos' to indicate the first entry of the second node.
3289 - We then iterate over the columns of the node, finding the first and
3290 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3291 and set the .check pointers accordingly, and advance pos
3292 appropriately and repreat for the next node. Note that when we copy
3293 the next pointers we have to convert them from the original
3294 NODEIDX form to NODENUM form as the former is not valid post
3297 - If a node has no transitions used we mark its base as 0 and do not
3298 advance the pos pointer.
3300 - If a node only has one transition we use a second pointer into the
3301 structure to fill in allocated fail transitions from other states.
3302 This pointer is independent of the main pointer and scans forward
3303 looking for null transitions that are allocated to a state. When it
3304 finds one it writes the single transition into the "hole". If the
3305 pointer doesnt find one the single transition is appended as normal.
3307 - Once compressed we can Renew/realloc the structures to release the
3310 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3311 specifically Fig 3.47 and the associated pseudocode.
3315 const U32 laststate = TRIE_NODENUM( next_alloc );
3318 trie->statecount = laststate;
3320 for ( state = 1 ; state < laststate ; state++ ) {
3322 const U32 stateidx = TRIE_NODEIDX( state );
3323 const U32 o_used = trie->trans[ stateidx ].check;
3324 U32 used = trie->trans[ stateidx ].check;
3325 trie->trans[ stateidx ].check = 0;
3328 used && charid < trie->uniquecharcount;
3331 if ( flag || trie->trans[ stateidx + charid ].next ) {
3332 if ( trie->trans[ stateidx + charid ].next ) {
3334 for ( ; zp < pos ; zp++ ) {
3335 if ( ! trie->trans[ zp ].next ) {
3339 trie->states[ state ].trans.base
3341 + trie->uniquecharcount
3343 trie->trans[ zp ].next
3344 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3346 trie->trans[ zp ].check = state;
3347 if ( ++zp > pos ) pos = zp;
3354 trie->states[ state ].trans.base
3355 = pos + trie->uniquecharcount - charid ;
3357 trie->trans[ pos ].next
3358 = SAFE_TRIE_NODENUM(
3359 trie->trans[ stateidx + charid ].next );
3360 trie->trans[ pos ].check = state;
3365 trie->lasttrans = pos + 1;
3366 trie->states = (reg_trie_state *)
3367 PerlMemShared_realloc( trie->states, laststate
3368 * sizeof(reg_trie_state) );
3369 DEBUG_TRIE_COMPILE_MORE_r(
3370 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
3372 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3376 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3379 } /* end table compress */
3381 DEBUG_TRIE_COMPILE_MORE_r(
3382 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
3384 (UV)trie->statecount,
3385 (UV)trie->lasttrans)
3387 /* resize the trans array to remove unused space */
3388 trie->trans = (reg_trie_trans *)
3389 PerlMemShared_realloc( trie->trans, trie->lasttrans
3390 * sizeof(reg_trie_trans) );
3392 { /* Modify the program and insert the new TRIE node */
3393 U8 nodetype =(U8)(flags & 0xFF);
3397 regnode *optimize = NULL;
3398 #ifdef RE_TRACK_PATTERN_OFFSETS
3401 U32 mjd_nodelen = 0;
3402 #endif /* RE_TRACK_PATTERN_OFFSETS */
3403 #endif /* DEBUGGING */
3405 This means we convert either the first branch or the first Exact,
3406 depending on whether the thing following (in 'last') is a branch
3407 or not and whther first is the startbranch (ie is it a sub part of
3408 the alternation or is it the whole thing.)
3409 Assuming its a sub part we convert the EXACT otherwise we convert
3410 the whole branch sequence, including the first.
3412 /* Find the node we are going to overwrite */
3413 if ( first != startbranch || OP( last ) == BRANCH ) {
3414 /* branch sub-chain */
3415 NEXT_OFF( first ) = (U16)(last - first);
3416 #ifdef RE_TRACK_PATTERN_OFFSETS
3418 mjd_offset= Node_Offset((convert));
3419 mjd_nodelen= Node_Length((convert));
3422 /* whole branch chain */
3424 #ifdef RE_TRACK_PATTERN_OFFSETS
3427 const regnode *nop = NEXTOPER( convert );
3428 mjd_offset= Node_Offset((nop));
3429 mjd_nodelen= Node_Length((nop));
3433 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
3435 (UV)mjd_offset, (UV)mjd_nodelen)
3438 /* But first we check to see if there is a common prefix we can
3439 split out as an EXACT and put in front of the TRIE node. */
3440 trie->startstate= 1;
3441 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3442 /* we want to find the first state that has more than
3443 * one transition, if that state is not the first state
3444 * then we have a common prefix which we can remove.
3447 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3449 I32 first_ofs = -1; /* keeps track of the ofs of the first
3450 transition, -1 means none */
3452 const U32 base = trie->states[ state ].trans.base;
3454 /* does this state terminate an alternation? */
3455 if ( trie->states[state].wordnum )
3458 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3459 if ( ( base + ofs >= trie->uniquecharcount ) &&
3460 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3461 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3463 if ( ++count > 1 ) {
3464 /* we have more than one transition */
3467 /* if this is the first state there is no common prefix
3468 * to extract, so we can exit */
3469 if ( state == 1 ) break;
3470 tmp = av_fetch( revcharmap, ofs, 0);
3471 ch = (U8*)SvPV_nolen_const( *tmp );
3473 /* if we are on count 2 then we need to initialize the
3474 * bitmap, and store the previous char if there was one
3477 /* clear the bitmap */
3478 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3480 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
3483 if (first_ofs >= 0) {
3484 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
3485 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3487 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3489 Perl_re_printf( aTHX_ "%s", (char*)ch)
3493 /* store the current firstchar in the bitmap */
3494 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3495 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
3501 /* This state has only one transition, its transition is part
3502 * of a common prefix - we need to concatenate the char it
3503 * represents to what we have so far. */
3504 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
3506 char *ch = SvPV( *tmp, len );
3508 SV *sv=sv_newmortal();
3509 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
3511 (UV)state, (UV)first_ofs,
3512 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3513 PL_colors[0], PL_colors[1],
3514 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3515 PERL_PV_ESCAPE_FIRSTCHAR
3520 OP( convert ) = nodetype;
3521 str=STRING(convert);
3524 STR_LEN(convert) += len;
3530 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
3535 trie->prefixlen = (state-1);
3537 regnode *n = convert+NODE_SZ_STR(convert);
3538 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3539 trie->startstate = state;
3540 trie->minlen -= (state - 1);
3541 trie->maxlen -= (state - 1);
3543 /* At least the UNICOS C compiler choked on this
3544 * being argument to DEBUG_r(), so let's just have
3547 #ifdef PERL_EXT_RE_BUILD
3553 regnode *fix = convert;
3554 U32 word = trie->wordcount;
3555 #ifdef RE_TRACK_PATTERN_OFFSETS
3558 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3559 while( ++fix < n ) {
3560 Set_Node_Offset_Length(fix, 0, 0);
3563 SV ** const tmp = av_fetch( trie_words, word, 0 );
3565 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3566 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3568 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3576 NEXT_OFF(convert) = (U16)(tail - convert);
3577 DEBUG_r(optimize= n);
3583 if ( trie->maxlen ) {
3584 NEXT_OFF( convert ) = (U16)(tail - convert);
3585 ARG_SET( convert, data_slot );
3586 /* Store the offset to the first unabsorbed branch in
3587 jump[0], which is otherwise unused by the jump logic.
3588 We use this when dumping a trie and during optimisation. */
3590 trie->jump[0] = (U16)(nextbranch - convert);
3592 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3593 * and there is a bitmap
3594 * and the first "jump target" node we found leaves enough room
3595 * then convert the TRIE node into a TRIEC node, with the bitmap
3596 * embedded inline in the opcode - this is hypothetically faster.
3598 if ( !trie->states[trie->startstate].wordnum
3600 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3602 OP( convert ) = TRIEC;
3603 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3604 PerlMemShared_free(trie->bitmap);
3607 OP( convert ) = TRIE;
3609 /* store the type in the flags */
3610 convert->flags = nodetype;
3614 + regarglen[ OP( convert ) ];
3616 /* XXX We really should free up the resource in trie now,
3617 as we won't use them - (which resources?) dmq */
3619 /* needed for dumping*/
3620 DEBUG_r(if (optimize) {
3621 regnode *opt = convert;
3623 while ( ++opt < optimize) {
3624 Set_Node_Offset_Length(opt, 0, 0);
3627 Try to clean up some of the debris left after the
3630 while( optimize < jumper ) {
3631 Track_Code( mjd_nodelen += Node_Length((optimize)); );
3632 OP( optimize ) = OPTIMIZED;
3633 Set_Node_Offset_Length(optimize, 0, 0);
3636 Set_Node_Offset_Length(convert, mjd_offset, mjd_nodelen);
3638 } /* end node insert */
3640 /* Finish populating the prev field of the wordinfo array. Walk back
3641 * from each accept state until we find another accept state, and if
3642 * so, point the first word's .prev field at the second word. If the
3643 * second already has a .prev field set, stop now. This will be the
3644 * case either if we've already processed that word's accept state,
3645 * or that state had multiple words, and the overspill words were
3646 * already linked up earlier.
3653 for (word=1; word <= trie->wordcount; word++) {
3655 if (trie->wordinfo[word].prev)
3657 state = trie->wordinfo[word].accept;
3659 state = prev_states[state];
3662 prev = trie->states[state].wordnum;
3666 trie->wordinfo[word].prev = prev;
3668 Safefree(prev_states);
3672 /* and now dump out the compressed format */
3673 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3675 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3677 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3678 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3680 SvREFCNT_dec_NN(revcharmap);
3684 : trie->startstate>1
3690 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3692 /* The Trie is constructed and compressed now so we can build a fail array if
3695 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3697 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3701 We find the fail state for each state in the trie, this state is the longest
3702 proper suffix of the current state's 'word' that is also a proper prefix of
3703 another word in our trie. State 1 represents the word '' and is thus the
3704 default fail state. This allows the DFA not to have to restart after its
3705 tried and failed a word at a given point, it simply continues as though it
3706 had been matching the other word in the first place.
3708 'abcdgu'=~/abcdefg|cdgu/
3709 When we get to 'd' we are still matching the first word, we would encounter
3710 'g' which would fail, which would bring us to the state representing 'd' in
3711 the second word where we would try 'g' and succeed, proceeding to match
3714 /* add a fail transition */
3715 const U32 trie_offset = ARG(source);
3716 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3718 const U32 ucharcount = trie->uniquecharcount;
3719 const U32 numstates = trie->statecount;
3720 const U32 ubound = trie->lasttrans + ucharcount;
3724 U32 base = trie->states[ 1 ].trans.base;
3727 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3729 GET_RE_DEBUG_FLAGS_DECL;
3731 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3732 PERL_UNUSED_CONTEXT;
3734 PERL_UNUSED_ARG(depth);
3737 if ( OP(source) == TRIE ) {
3738 struct regnode_1 *op = (struct regnode_1 *)
3739 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3740 StructCopy(source, op, struct regnode_1);
3741 stclass = (regnode *)op;
3743 struct regnode_charclass *op = (struct regnode_charclass *)
3744 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3745 StructCopy(source, op, struct regnode_charclass);
3746 stclass = (regnode *)op;
3748 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3750 ARG_SET( stclass, data_slot );
3751 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3752 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3753 aho->trie=trie_offset;
3754 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3755 Copy( trie->states, aho->states, numstates, reg_trie_state );
3756 Newx( q, numstates, U32);
3757 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3760 /* initialize fail[0..1] to be 1 so that we always have
3761 a valid final fail state */
3762 fail[ 0 ] = fail[ 1 ] = 1;
3764 for ( charid = 0; charid < ucharcount ; charid++ ) {
3765 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3767 q[ q_write ] = newstate;
3768 /* set to point at the root */
3769 fail[ q[ q_write++ ] ]=1;
3772 while ( q_read < q_write) {
3773 const U32 cur = q[ q_read++ % numstates ];
3774 base = trie->states[ cur ].trans.base;
3776 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3777 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3779 U32 fail_state = cur;
3782 fail_state = fail[ fail_state ];
3783 fail_base = aho->states[ fail_state ].trans.base;
3784 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3786 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3787 fail[ ch_state ] = fail_state;
3788 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3790 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3792 q[ q_write++ % numstates] = ch_state;
3796 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3797 when we fail in state 1, this allows us to use the
3798 charclass scan to find a valid start char. This is based on the principle
3799 that theres a good chance the string being searched contains lots of stuff
3800 that cant be a start char.
3802 fail[ 0 ] = fail[ 1 ] = 0;
3803 DEBUG_TRIE_COMPILE_r({
3804 Perl_re_indentf( aTHX_ "Stclass Failtable (%" UVuf " states): 0",
3805 depth, (UV)numstates
3807 for( q_read=1; q_read<numstates; q_read++ ) {
3808 Perl_re_printf( aTHX_ ", %" UVuf, (UV)fail[q_read]);
3810 Perl_re_printf( aTHX_ "\n");
3813 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3818 /* The below joins as many adjacent EXACTish nodes as possible into a single
3819 * one. The regop may be changed if the node(s) contain certain sequences that
3820 * require special handling. The joining is only done if:
3821 * 1) there is room in the current conglomerated node to entirely contain the
3823 * 2) they are compatible node types
3825 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3826 * these get optimized out
3828 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3829 * as possible, even if that means splitting an existing node so that its first
3830 * part is moved to the preceeding node. This would maximise the efficiency of
3831 * memEQ during matching.
3833 * If a node is to match under /i (folded), the number of characters it matches
3834 * can be different than its character length if it contains a multi-character
3835 * fold. *min_subtract is set to the total delta number of characters of the
3838 * And *unfolded_multi_char is set to indicate whether or not the node contains
3839 * an unfolded multi-char fold. This happens when it won't be known until
3840 * runtime whether the fold is valid or not; namely
3841 * 1) for EXACTF nodes that contain LATIN SMALL LETTER SHARP S, as only if the
3842 * target string being matched against turns out to be UTF-8 is that fold
3844 * 2) for EXACTFL nodes whose folding rules depend on the locale in force at
3846 * (Multi-char folds whose components are all above the Latin1 range are not
3847 * run-time locale dependent, and have already been folded by the time this
3848 * function is called.)
3850 * This is as good a place as any to discuss the design of handling these
3851 * multi-character fold sequences. It's been wrong in Perl for a very long
3852 * time. There are three code points in Unicode whose multi-character folds
3853 * were long ago discovered to mess things up. The previous designs for
3854 * dealing with these involved assigning a special node for them. This
3855 * approach doesn't always work, as evidenced by this example:
3856 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3857 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3858 * would match just the \xDF, it won't be able to handle the case where a
3859 * successful match would have to cross the node's boundary. The new approach
3860 * that hopefully generally solves the problem generates an EXACTFUP node
3861 * that is "sss" in this case.
3863 * It turns out that there are problems with all multi-character folds, and not
3864 * just these three. Now the code is general, for all such cases. The
3865 * approach taken is:
3866 * 1) This routine examines each EXACTFish node that could contain multi-
3867 * character folded sequences. Since a single character can fold into
3868 * such a sequence, the minimum match length for this node is less than
3869 * the number of characters in the node. This routine returns in
3870 * *min_subtract how many characters to subtract from the the actual
3871 * length of the string to get a real minimum match length; it is 0 if
3872 * there are no multi-char foldeds. This delta is used by the caller to
3873 * adjust the min length of the match, and the delta between min and max,
3874 * so that the optimizer doesn't reject these possibilities based on size
3877 * 2) For the sequence involving the LATIN SMALL LETTER SHARP S (U+00DF)
3878 * under /u, we fold it to 'ss' in regatom(), and in this routine, after
3879 * joining, we scan for occurrences of the sequence 'ss' in non-UTF-8
3880 * EXACTFU nodes. The node type of such nodes is then changed to
3881 * EXACTFUP, indicating it is problematic, and needs careful handling.
3882 * (The procedures in step 1) above are sufficient to handle this case in
3883 * UTF-8 encoded nodes.) The reason this is problematic is that this is
3884 * the only case where there is a possible fold length change in non-UTF-8
3885 * patterns. By reserving a special node type for problematic cases, the
3886 * far more common regular EXACTFU nodes can be processed faster.
3887 * regexec.c takes advantage of this.
3889 * EXACTFUP has been created as a grab-bag for (hopefully uncommon)
3890 * problematic cases. These all only occur when the pattern is not
3891 * UTF-8. In addition to the 'ss' sequence where there is a possible fold
3892 * length change, it handles the situation where the string cannot be
3893 * entirely folded. The strings in an EXACTFish node are folded as much
3894 * as possible during compilation in regcomp.c. This saves effort in
3895 * regex matching. By using an EXACTFUP node when it is not possible to
3896 * fully fold at compile time, regexec.c can know that everything in an
3897 * EXACTFU node is folded, so folding can be skipped at runtime. The only
3898 * case where folding in EXACTFU nodes can't be done at compile time is
3899 * the presumably uncommon MICRO SIGN, when the pattern isn't UTF-8. This
3900 * is because its fold requires UTF-8 to represent. Thus EXACTFUP nodes
3901 * handle two very different cases. Alternatively, there could have been
3902 * a node type where there are length changes, one for unfolded, and one
3903 * for both. If yet another special case needed to be created, the number
3904 * of required node types would have to go to 7. khw figures that even
3905 * though there are plenty of node types to spare, that the maintenance
3906 * cost wasn't worth the small speedup of doing it that way, especially
3907 * since he thinks the MICRO SIGN is rarely encountered in practice.
3909 * There are other cases where folding isn't done at compile time, but
3910 * none of them are under /u, and hence not for EXACTFU nodes. The folds
3911 * in EXACTFL nodes aren't known until runtime, and vary as the locale
3912 * changes. Some folds in EXACTF depend on if the runtime target string
3913 * is UTF-8 or not. (regatom() will create an EXACTFU node even under /di
3914 * when no fold in it depends on the UTF-8ness of the target string.)
3916 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3917 * validity of the fold won't be known until runtime, and so must remain
3918 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFAA
3919 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3920 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3921 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3922 * The reason this is a problem is that the optimizer part of regexec.c
3923 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3924 * that a character in the pattern corresponds to at most a single
3925 * character in the target string. (And I do mean character, and not byte
3926 * here, unlike other parts of the documentation that have never been
3927 * updated to account for multibyte Unicode.) Sharp s in EXACTF and
3928 * EXACTFL nodes can match the two character string 'ss'; in EXACTFAA
3929 * nodes it can match "\x{17F}\x{17F}". These, along with other ones in
3930 * EXACTFL nodes, violate the assumption, and they are the only instances
3931 * where it is violated. I'm reluctant to try to change the assumption,
3932 * as the code involved is impenetrable to me (khw), so instead the code
3933 * here punts. This routine examines EXACTFL nodes, and (when the pattern
3934 * isn't UTF-8) EXACTF and EXACTFAA for such unfolded folds, and returns a
3935 * boolean indicating whether or not the node contains such a fold. When
3936 * it is true, the caller sets a flag that later causes the optimizer in
3937 * this file to not set values for the floating and fixed string lengths,
3938 * and thus avoids the optimizer code in regexec.c that makes the invalid
3939 * assumption. Thus, there is no optimization based on string lengths for
3940 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3941 * EXACTF and EXACTFAA nodes that contain the sharp s. (The reason the
3942 * assumption is wrong only in these cases is that all other non-UTF-8
3943 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3944 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3945 * EXACTF nodes because we don't know at compile time if it actually
3946 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3947 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3948 * always matches; and EXACTFAA where it never does. In an EXACTFAA node
3949 * in a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3950 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3951 * string would require the pattern to be forced into UTF-8, the overhead
3952 * of which we want to avoid. Similarly the unfolded multi-char folds in
3953 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3956 * Similarly, the code that generates tries doesn't currently handle
3957 * not-already-folded multi-char folds, and it looks like a pain to change
3958 * that. Therefore, trie generation of EXACTFAA nodes with the sharp s
3959 * doesn't work. Instead, such an EXACTFAA is turned into a new regnode,
3960 * EXACTFAA_NO_TRIE, which the trie code knows not to handle. Most people
3961 * using /iaa matching will be doing so almost entirely with ASCII
3962 * strings, so this should rarely be encountered in practice */
3964 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3965 if (PL_regkind[OP(scan)] == EXACT) \
3966 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags), NULL, depth+1)
3969 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3970 UV *min_subtract, bool *unfolded_multi_char,
3971 U32 flags, regnode *val, U32 depth)
3973 /* Merge several consecutive EXACTish nodes into one. */
3975 regnode *n = regnext(scan);
3977 regnode *next = scan + NODE_SZ_STR(scan);
https://perl5.git.perl.org / perl5.git / blob