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
78 #ifdef PERL_IN_XSUB_RE
80 EXTERN_C const struct regexp_engine my_reg_engine;
85 #include "dquote_inline.h"
86 #include "invlist_inline.h"
87 #include "unicode_constants.h"
89 #define HAS_NONLATIN1_FOLD_CLOSURE(i) \
90 _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
91 #define HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(i) \
92 _HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
93 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
94 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
100 /* this is a chain of data about sub patterns we are processing that
101 need to be handled separately/specially in study_chunk. Its so
102 we can simulate recursion without losing state. */
104 typedef struct scan_frame {
105 regnode *last_regnode; /* last node to process in this frame */
106 regnode *next_regnode; /* next node to process when last is reached */
107 U32 prev_recursed_depth;
108 I32 stopparen; /* what stopparen do we use */
110 struct scan_frame *this_prev_frame; /* this previous frame */
111 struct scan_frame *prev_frame; /* previous frame */
112 struct scan_frame *next_frame; /* next frame */
115 /* Certain characters are output as a sequence with the first being a
117 #define isBACKSLASHED_PUNCT(c) strchr("-[]\\^", c)
120 struct RExC_state_t {
121 U32 flags; /* RXf_* are we folding, multilining? */
122 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
123 char *precomp; /* uncompiled string. */
124 char *precomp_end; /* pointer to end of uncompiled string. */
125 REGEXP *rx_sv; /* The SV that is the regexp. */
126 regexp *rx; /* perl core regexp structure */
127 regexp_internal *rxi; /* internal data for regexp object
129 char *start; /* Start of input for compile */
130 char *end; /* End of input for compile */
131 char *parse; /* Input-scan pointer. */
132 char *copy_start; /* start of copy of input within
133 constructed parse string */
134 char *copy_start_in_input; /* Position in input string
135 corresponding to copy_start */
136 SSize_t whilem_seen; /* number of WHILEM in this expr */
137 regnode *emit_start; /* Start of emitted-code area */
138 regnode_offset emit; /* Code-emit pointer */
139 I32 naughty; /* How bad is this pattern? */
140 I32 sawback; /* Did we see \1, ...? */
142 SSize_t size; /* Number of regnode equivalents in
145 /* position beyond 'precomp' of the warning message furthest away from
146 * 'precomp'. During the parse, no warnings are raised for any problems
147 * earlier in the parse than this position. This works if warnings are
148 * raised the first time a given spot is parsed, and if only one
149 * independent warning is raised for any given spot */
150 Size_t latest_warn_offset;
152 I32 npar; /* Capture buffer count so far in the
153 parse, (OPEN) plus one. ("par" 0 is
155 I32 total_par; /* During initial parse, is either 0,
156 or -1; the latter indicating a
157 reparse is needed. After that pass,
158 it is what 'npar' became after the
159 pass. Hence, it being > 0 indicates
160 we are in a reparse situation */
161 I32 nestroot; /* root parens we are in - used by
164 regnode_offset *open_parens; /* offsets to open parens */
165 regnode_offset *close_parens; /* offsets to close parens */
166 I32 parens_buf_size; /* #slots malloced open/close_parens */
167 regnode *end_op; /* END node in program */
168 I32 utf8; /* whether the pattern is utf8 or not */
169 I32 orig_utf8; /* whether the pattern was originally in utf8 */
170 /* XXX use this for future optimisation of case
171 * where pattern must be upgraded to utf8. */
172 I32 uni_semantics; /* If a d charset modifier should use unicode
173 rules, even if the pattern is not in
175 HV *paren_names; /* Paren names */
177 regnode **recurse; /* Recurse regops */
178 I32 recurse_count; /* Number of recurse regops we have generated */
179 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
181 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
184 I32 override_recoding;
186 I32 recode_x_to_native;
188 I32 in_multi_char_class;
189 struct reg_code_blocks *code_blocks;/* positions of literal (?{})
191 int code_index; /* next code_blocks[] slot */
192 SSize_t maxlen; /* mininum possible number of chars in string to match */
193 scan_frame *frame_head;
194 scan_frame *frame_last;
198 #ifdef ADD_TO_REGEXEC
199 char *starttry; /* -Dr: where regtry was called. */
200 #define RExC_starttry (pRExC_state->starttry)
202 SV *runtime_code_qr; /* qr with the runtime code blocks */
204 const char *lastparse;
206 AV *paren_name_list; /* idx -> name */
207 U32 study_chunk_recursed_count;
211 #define RExC_lastparse (pRExC_state->lastparse)
212 #define RExC_lastnum (pRExC_state->lastnum)
213 #define RExC_paren_name_list (pRExC_state->paren_name_list)
214 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
215 #define RExC_mysv (pRExC_state->mysv1)
216 #define RExC_mysv1 (pRExC_state->mysv1)
217 #define RExC_mysv2 (pRExC_state->mysv2)
227 #define RExC_flags (pRExC_state->flags)
228 #define RExC_pm_flags (pRExC_state->pm_flags)
229 #define RExC_precomp (pRExC_state->precomp)
230 #define RExC_copy_start_in_input (pRExC_state->copy_start_in_input)
231 #define RExC_copy_start_in_constructed (pRExC_state->copy_start)
232 #define RExC_precomp_end (pRExC_state->precomp_end)
233 #define RExC_rx_sv (pRExC_state->rx_sv)
234 #define RExC_rx (pRExC_state->rx)
235 #define RExC_rxi (pRExC_state->rxi)
236 #define RExC_start (pRExC_state->start)
237 #define RExC_end (pRExC_state->end)
238 #define RExC_parse (pRExC_state->parse)
239 #define RExC_latest_warn_offset (pRExC_state->latest_warn_offset )
240 #define RExC_whilem_seen (pRExC_state->whilem_seen)
241 #define RExC_seen_d_op (pRExC_state->seen_d_op) /* Seen something that differs
242 under /d from /u ? */
245 #ifdef RE_TRACK_PATTERN_OFFSETS
246 # define RExC_offsets (RExC_rxi->u.offsets) /* I am not like the
249 #define RExC_emit (pRExC_state->emit)
250 #define RExC_emit_start (pRExC_state->emit_start)
251 #define RExC_sawback (pRExC_state->sawback)
252 #define RExC_seen (pRExC_state->seen)
253 #define RExC_size (pRExC_state->size)
254 #define RExC_maxlen (pRExC_state->maxlen)
255 #define RExC_npar (pRExC_state->npar)
256 #define RExC_total_parens (pRExC_state->total_par)
257 #define RExC_parens_buf_size (pRExC_state->parens_buf_size)
258 #define RExC_nestroot (pRExC_state->nestroot)
259 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
260 #define RExC_utf8 (pRExC_state->utf8)
261 #define RExC_uni_semantics (pRExC_state->uni_semantics)
262 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
263 #define RExC_open_parens (pRExC_state->open_parens)
264 #define RExC_close_parens (pRExC_state->close_parens)
265 #define RExC_end_op (pRExC_state->end_op)
266 #define RExC_paren_names (pRExC_state->paren_names)
267 #define RExC_recurse (pRExC_state->recurse)
268 #define RExC_recurse_count (pRExC_state->recurse_count)
269 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
270 #define RExC_study_chunk_recursed_bytes \
271 (pRExC_state->study_chunk_recursed_bytes)
272 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
273 #define RExC_contains_locale (pRExC_state->contains_locale)
275 # define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
277 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
278 #define RExC_frame_head (pRExC_state->frame_head)
279 #define RExC_frame_last (pRExC_state->frame_last)
280 #define RExC_frame_count (pRExC_state->frame_count)
281 #define RExC_strict (pRExC_state->strict)
282 #define RExC_study_started (pRExC_state->study_started)
283 #define RExC_warn_text (pRExC_state->warn_text)
284 #define RExC_in_script_run (pRExC_state->in_script_run)
285 #define RExC_use_BRANCHJ (pRExC_state->use_BRANCHJ)
286 #define RExC_unlexed_names (pRExC_state->unlexed_names)
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(! IN_PARENS_PASS)) { \
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 REQUIRE_BRANCHJ(flagp, restart_retval) \
371 RExC_use_BRANCHJ = 1; \
372 if (LIKELY(! IN_PARENS_PASS)) { \
373 /* No need to restart the parse immediately if we're \
374 * going to reparse anyway to count parens */ \
375 *flagp |= RESTART_PARSE; \
376 return restart_retval; \
380 /* Until we have completed the parse, we leave RExC_total_parens at 0 or
381 * less. After that, it must always be positive, because the whole re is
382 * considered to be surrounded by virtual parens. Setting it to negative
383 * indicates there is some construct that needs to know the actual number of
384 * parens to be properly handled. And that means an extra pass will be
385 * required after we've counted them all */
386 #define ALL_PARENS_COUNTED (RExC_total_parens > 0)
387 #define REQUIRE_PARENS_PASS \
388 STMT_START { /* No-op if have completed a pass */ \
389 if (! ALL_PARENS_COUNTED) RExC_total_parens = -1; \
391 #define IN_PARENS_PASS (RExC_total_parens < 0)
394 /* This is used to return failure (zero) early from the calling function if
395 * various flags in 'flags' are set. Two flags always cause a return:
396 * 'RESTART_PARSE' and 'NEED_UTF8'. 'extra' can be used to specify any
397 * additional flags that should cause a return; 0 if none. If the return will
398 * be done, '*flagp' is first set to be all of the flags that caused the
400 #define RETURN_FAIL_ON_RESTART_OR_FLAGS(flags,flagp,extra) \
402 if ((flags) & (RESTART_PARSE|NEED_UTF8|(extra))) { \
403 *(flagp) = (flags) & (RESTART_PARSE|NEED_UTF8|(extra)); \
408 #define MUST_RESTART(flags) ((flags) & (RESTART_PARSE))
410 #define RETURN_FAIL_ON_RESTART(flags,flagp) \
411 RETURN_FAIL_ON_RESTART_OR_FLAGS( flags, flagp, 0)
412 #define RETURN_FAIL_ON_RESTART_FLAGP(flagp) \
413 if (MUST_RESTART(*(flagp))) return 0
415 /* This converts the named class defined in regcomp.h to its equivalent class
416 * number defined in handy.h. */
417 #define namedclass_to_classnum(class) ((int) ((class) / 2))
418 #define classnum_to_namedclass(classnum) ((classnum) * 2)
420 #define _invlist_union_complement_2nd(a, b, output) \
421 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
422 #define _invlist_intersection_complement_2nd(a, b, output) \
423 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
425 /* About scan_data_t.
427 During optimisation we recurse through the regexp program performing
428 various inplace (keyhole style) optimisations. In addition study_chunk
429 and scan_commit populate this data structure with information about
430 what strings MUST appear in the pattern. We look for the longest
431 string that must appear at a fixed location, and we look for the
432 longest string that may appear at a floating location. So for instance
437 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
438 strings (because they follow a .* construct). study_chunk will identify
439 both FOO and BAR as being the longest fixed and floating strings respectively.
441 The strings can be composites, for instance
445 will result in a composite fixed substring 'foo'.
447 For each string some basic information is maintained:
450 This is the position the string must appear at, or not before.
451 It also implicitly (when combined with minlenp) tells us how many
452 characters must match before the string we are searching for.
453 Likewise when combined with minlenp and the length of the string it
454 tells us how many characters must appear after the string we have
458 Only used for floating strings. This is the rightmost point that
459 the string can appear at. If set to SSize_t_MAX it indicates that the
460 string can occur infinitely far to the right.
461 For fixed strings, it is equal to min_offset.
464 A pointer to the minimum number of characters of the pattern that the
465 string was found inside. This is important as in the case of positive
466 lookahead or positive lookbehind we can have multiple patterns
471 The minimum length of the pattern overall is 3, the minimum length
472 of the lookahead part is 3, but the minimum length of the part that
473 will actually match is 1. So 'FOO's minimum length is 3, but the
474 minimum length for the F is 1. This is important as the minimum length
475 is used to determine offsets in front of and behind the string being
476 looked for. Since strings can be composites this is the length of the
477 pattern at the time it was committed with a scan_commit. Note that
478 the length is calculated by study_chunk, so that the minimum lengths
479 are not known until the full pattern has been compiled, thus the
480 pointer to the value.
484 In the case of lookbehind the string being searched for can be
485 offset past the start point of the final matching string.
486 If this value was just blithely removed from the min_offset it would
487 invalidate some of the calculations for how many chars must match
488 before or after (as they are derived from min_offset and minlen and
489 the length of the string being searched for).
490 When the final pattern is compiled and the data is moved from the
491 scan_data_t structure into the regexp structure the information
492 about lookbehind is factored in, with the information that would
493 have been lost precalculated in the end_shift field for the
496 The fields pos_min and pos_delta are used to store the minimum offset
497 and the delta to the maximum offset at the current point in the pattern.
501 struct scan_data_substrs {
502 SV *str; /* longest substring found in pattern */
503 SSize_t min_offset; /* earliest point in string it can appear */
504 SSize_t max_offset; /* latest point in string it can appear */
505 SSize_t *minlenp; /* pointer to the minlen relevant to the string */
506 SSize_t lookbehind; /* is the pos of the string modified by LB */
507 I32 flags; /* per substring SF_* and SCF_* flags */
510 typedef struct scan_data_t {
511 /*I32 len_min; unused */
512 /*I32 len_delta; unused */
516 SSize_t last_end; /* min value, <0 unless valid. */
517 SSize_t last_start_min;
518 SSize_t last_start_max;
519 U8 cur_is_floating; /* whether the last_* values should be set as
520 * the next fixed (0) or floating (1)
523 /* [0] is longest fixed substring so far, [1] is longest float so far */
524 struct scan_data_substrs substrs[2];
526 I32 flags; /* common SF_* and SCF_* flags */
528 SSize_t *last_closep;
529 regnode_ssc *start_class;
533 * Forward declarations for pregcomp()'s friends.
536 static const scan_data_t zero_scan_data = {
537 0, 0, NULL, 0, 0, 0, 0,
539 { NULL, 0, 0, 0, 0, 0 },
540 { NULL, 0, 0, 0, 0, 0 },
547 #define SF_BEFORE_SEOL 0x0001
548 #define SF_BEFORE_MEOL 0x0002
549 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
551 #define SF_IS_INF 0x0040
552 #define SF_HAS_PAR 0x0080
553 #define SF_IN_PAR 0x0100
554 #define SF_HAS_EVAL 0x0200
557 /* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
558 * longest substring in the pattern. When it is not set the optimiser keeps
559 * track of position, but does not keep track of the actual strings seen,
561 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
564 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
565 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
566 * turned off because of the alternation (BRANCH). */
567 #define SCF_DO_SUBSTR 0x0400
569 #define SCF_DO_STCLASS_AND 0x0800
570 #define SCF_DO_STCLASS_OR 0x1000
571 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
572 #define SCF_WHILEM_VISITED_POS 0x2000
574 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
575 #define SCF_SEEN_ACCEPT 0x8000
576 #define SCF_TRIE_DOING_RESTUDY 0x10000
577 #define SCF_IN_DEFINE 0x20000
582 #define UTF cBOOL(RExC_utf8)
584 /* The enums for all these are ordered so things work out correctly */
585 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
586 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
587 == REGEX_DEPENDS_CHARSET)
588 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
589 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
590 >= REGEX_UNICODE_CHARSET)
591 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
592 == REGEX_ASCII_RESTRICTED_CHARSET)
593 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
594 >= REGEX_ASCII_RESTRICTED_CHARSET)
595 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
596 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
598 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
600 /* For programs that want to be strictly Unicode compatible by dying if any
601 * attempt is made to match a non-Unicode code point against a Unicode
603 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
605 #define OOB_NAMEDCLASS -1
607 /* There is no code point that is out-of-bounds, so this is problematic. But
608 * its only current use is to initialize a variable that is always set before
610 #define OOB_UNICODE 0xDEADBEEF
612 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
615 /* length of regex to show in messages that don't mark a position within */
616 #define RegexLengthToShowInErrorMessages 127
619 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
620 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
621 * op/pragma/warn/regcomp.
623 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
624 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
626 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
627 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
629 /* The code in this file in places uses one level of recursion with parsing
630 * rebased to an alternate string constructed by us in memory. This can take
631 * the form of something that is completely different from the input, or
632 * something that uses the input as part of the alternate. In the first case,
633 * there should be no possibility of an error, as we are in complete control of
634 * the alternate string. But in the second case we don't completely control
635 * the input portion, so there may be errors in that. Here's an example:
637 * is handled specially because \x{df} folds to a sequence of more than one
638 * character: 'ss'. What is done is to create and parse an alternate string,
639 * which looks like this:
640 * /(?:\x{DF}|[abc\x{DF}def])/ui
641 * where it uses the input unchanged in the middle of something it constructs,
642 * which is a branch for the DF outside the character class, and clustering
643 * parens around the whole thing. (It knows enough to skip the DF inside the
644 * class while in this substitute parse.) 'abc' and 'def' may have errors that
645 * need to be reported. The general situation looks like this:
647 * |<------- identical ------>|
649 * Input: ---------------------------------------------------------------
650 * Constructed: ---------------------------------------------------
652 * |<------- identical ------>|
654 * sI..eI is the portion of the input pattern we are concerned with here.
655 * sC..EC is the constructed substitute parse string.
656 * sC..tC is constructed by us
657 * tC..eC is an exact duplicate of the portion of the input pattern tI..eI.
658 * In the diagram, these are vertically aligned.
659 * eC..EC is also constructed by us.
660 * xC is the position in the substitute parse string where we found a
662 * xI is the position in the original pattern corresponding to xC.
664 * We want to display a message showing the real input string. Thus we need to
665 * translate from xC to xI. We know that xC >= tC, since the portion of the
666 * string sC..tC has been constructed by us, and so shouldn't have errors. We
668 * xI = tI + (xC - tC)
670 * When the substitute parse is constructed, the code needs to set:
673 * RExC_copy_start_in_input (tI)
674 * RExC_copy_start_in_constructed (tC)
675 * and restore them when done.
677 * During normal processing of the input pattern, both
678 * 'RExC_copy_start_in_input' and 'RExC_copy_start_in_constructed' are set to
679 * sI, so that xC equals xI.
682 #define sI RExC_precomp
683 #define eI RExC_precomp_end
684 #define sC RExC_start
686 #define tI RExC_copy_start_in_input
687 #define tC RExC_copy_start_in_constructed
688 #define xI(xC) (tI + (xC - tC))
689 #define xI_offset(xC) (xI(xC) - sI)
691 #define REPORT_LOCATION_ARGS(xC) \
693 (xI(xC) > eI) /* Don't run off end */ \
694 ? eI - sI /* Length before the <--HERE */ \
695 : ((xI_offset(xC) >= 0) \
697 : (Perl_croak(aTHX_ "panic: %s: %d: negative offset: %" \
698 IVdf " trying to output message for " \
700 __FILE__, __LINE__, (IV) xI_offset(xC), \
701 ((int) (eC - sC)), sC), 0)), \
702 sI), /* The input pattern printed up to the <--HERE */ \
704 (xI(xC) > eI) ? 0 : eI - xI(xC), /* Length after <--HERE */ \
705 (xI(xC) > eI) ? eI : xI(xC)) /* pattern after <--HERE */
707 /* Used to point after bad bytes for an error message, but avoid skipping
708 * past a nul byte. */
709 #define SKIP_IF_CHAR(s, e) (!*(s) ? 0 : UTF ? UTF8_SAFE_SKIP(s, e) : 1)
711 /* Set up to clean up after our imminent demise */
712 #define PREPARE_TO_DIE \
715 SAVEFREESV(RExC_rx_sv); \
716 if (RExC_open_parens) \
717 SAVEFREEPV(RExC_open_parens); \
718 if (RExC_close_parens) \
719 SAVEFREEPV(RExC_close_parens); \
723 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
724 * arg. Show regex, up to a maximum length. If it's too long, chop and add
727 #define _FAIL(code) STMT_START { \
728 const char *ellipses = ""; \
729 IV len = RExC_precomp_end - RExC_precomp; \
732 if (len > RegexLengthToShowInErrorMessages) { \
733 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
734 len = RegexLengthToShowInErrorMessages - 10; \
740 #define FAIL(msg) _FAIL( \
741 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
742 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
744 #define FAIL2(msg,arg) _FAIL( \
745 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
746 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
749 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
751 #define Simple_vFAIL(m) STMT_START { \
752 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
753 m, REPORT_LOCATION_ARGS(RExC_parse)); \
757 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
759 #define vFAIL(m) STMT_START { \
765 * Like Simple_vFAIL(), but accepts two arguments.
767 #define Simple_vFAIL2(m,a1) STMT_START { \
768 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
769 REPORT_LOCATION_ARGS(RExC_parse)); \
773 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
775 #define vFAIL2(m,a1) STMT_START { \
777 Simple_vFAIL2(m, a1); \
782 * Like Simple_vFAIL(), but accepts three arguments.
784 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
785 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
786 REPORT_LOCATION_ARGS(RExC_parse)); \
790 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
792 #define vFAIL3(m,a1,a2) STMT_START { \
794 Simple_vFAIL3(m, a1, a2); \
798 * Like Simple_vFAIL(), but accepts four arguments.
800 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
801 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
802 REPORT_LOCATION_ARGS(RExC_parse)); \
805 #define vFAIL4(m,a1,a2,a3) STMT_START { \
807 Simple_vFAIL4(m, a1, a2, a3); \
810 /* A specialized version of vFAIL2 that works with UTF8f */
811 #define vFAIL2utf8f(m, a1) STMT_START { \
813 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
814 REPORT_LOCATION_ARGS(RExC_parse)); \
817 #define vFAIL3utf8f(m, a1, a2) STMT_START { \
819 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
820 REPORT_LOCATION_ARGS(RExC_parse)); \
823 /* Setting this to NULL is a signal to not output warnings */
824 #define TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE RExC_copy_start_in_constructed = NULL
825 #define RESTORE_WARNINGS RExC_copy_start_in_constructed = RExC_precomp
827 /* Since a warning can be generated multiple times as the input is reparsed, we
828 * output it the first time we come to that point in the parse, but suppress it
829 * otherwise. 'RExC_copy_start_in_constructed' being NULL is a flag to not
830 * generate any warnings */
831 #define TO_OUTPUT_WARNINGS(loc) \
832 ( RExC_copy_start_in_constructed \
833 && ((xI(loc)) - RExC_precomp) > (Ptrdiff_t) RExC_latest_warn_offset)
835 /* After we've emitted a warning, we save the position in the input so we don't
837 #define UPDATE_WARNINGS_LOC(loc) \
839 if (TO_OUTPUT_WARNINGS(loc)) { \
840 RExC_latest_warn_offset = (xI(loc)) - RExC_precomp; \
844 /* 'warns' is the output of the packWARNx macro used in 'code' */
845 #define _WARN_HELPER(loc, warns, code) \
847 if (! RExC_copy_start_in_constructed) { \
848 Perl_croak( aTHX_ "panic! %s: %d: Tried to warn when none" \
849 " expected at '%s'", \
850 __FILE__, __LINE__, loc); \
852 if (TO_OUTPUT_WARNINGS(loc)) { \
856 UPDATE_WARNINGS_LOC(loc); \
860 /* m is not necessarily a "literal string", in this macro */
861 #define reg_warn_non_literal_string(loc, m) \
862 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
863 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
864 "%s" REPORT_LOCATION, \
865 m, REPORT_LOCATION_ARGS(loc)))
867 #define ckWARNreg(loc,m) \
868 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
869 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
871 REPORT_LOCATION_ARGS(loc)))
873 #define vWARN(loc, m) \
874 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
875 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
877 REPORT_LOCATION_ARGS(loc))) \
879 #define vWARN_dep(loc, m) \
880 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
881 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
883 REPORT_LOCATION_ARGS(loc)))
885 #define ckWARNdep(loc,m) \
886 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
887 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
889 REPORT_LOCATION_ARGS(loc)))
891 #define ckWARNregdep(loc,m) \
892 _WARN_HELPER(loc, packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
893 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
896 REPORT_LOCATION_ARGS(loc)))
898 #define ckWARN2reg_d(loc,m, a1) \
899 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
900 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
902 a1, REPORT_LOCATION_ARGS(loc)))
904 #define ckWARN2reg(loc, m, a1) \
905 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
906 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
908 a1, REPORT_LOCATION_ARGS(loc)))
910 #define vWARN3(loc, m, a1, a2) \
911 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
912 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
914 a1, a2, REPORT_LOCATION_ARGS(loc)))
916 #define ckWARN3reg(loc, m, a1, a2) \
917 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
918 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
921 REPORT_LOCATION_ARGS(loc)))
923 #define vWARN4(loc, m, a1, a2, a3) \
924 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
925 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
928 REPORT_LOCATION_ARGS(loc)))
930 #define ckWARN4reg(loc, m, a1, a2, a3) \
931 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
932 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
935 REPORT_LOCATION_ARGS(loc)))
937 #define vWARN5(loc, m, a1, a2, a3, a4) \
938 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
939 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
942 REPORT_LOCATION_ARGS(loc)))
944 #define ckWARNexperimental(loc, class, m) \
945 _WARN_HELPER(loc, packWARN(class), \
946 Perl_ck_warner_d(aTHX_ packWARN(class), \
948 REPORT_LOCATION_ARGS(loc)))
950 /* Convert between a pointer to a node and its offset from the beginning of the
952 #define REGNODE_p(offset) (RExC_emit_start + (offset))
953 #define REGNODE_OFFSET(node) ((node) - RExC_emit_start)
955 /* Macros for recording node offsets. 20001227 mjd@plover.com
956 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
957 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
958 * Element 0 holds the number n.
959 * Position is 1 indexed.
961 #ifndef RE_TRACK_PATTERN_OFFSETS
962 #define Set_Node_Offset_To_R(offset,byte)
963 #define Set_Node_Offset(node,byte)
964 #define Set_Cur_Node_Offset
965 #define Set_Node_Length_To_R(node,len)
966 #define Set_Node_Length(node,len)
967 #define Set_Node_Cur_Length(node,start)
968 #define Node_Offset(n)
969 #define Node_Length(n)
970 #define Set_Node_Offset_Length(node,offset,len)
971 #define ProgLen(ri) ri->u.proglen
972 #define SetProgLen(ri,x) ri->u.proglen = x
973 #define Track_Code(code)
975 #define ProgLen(ri) ri->u.offsets[0]
976 #define SetProgLen(ri,x) ri->u.offsets[0] = x
977 #define Set_Node_Offset_To_R(offset,byte) STMT_START { \
978 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
979 __LINE__, (int)(offset), (int)(byte))); \
981 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
984 RExC_offsets[2*(offset)-1] = (byte); \
988 #define Set_Node_Offset(node,byte) \
989 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (byte)-RExC_start)
990 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
992 #define Set_Node_Length_To_R(node,len) STMT_START { \
993 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
994 __LINE__, (int)(node), (int)(len))); \
996 Perl_croak(aTHX_ "value of node is %d in Length macro", \
999 RExC_offsets[2*(node)] = (len); \
1003 #define Set_Node_Length(node,len) \
1004 Set_Node_Length_To_R(REGNODE_OFFSET(node), len)
1005 #define Set_Node_Cur_Length(node, start) \
1006 Set_Node_Length(node, RExC_parse - start)
1008 /* Get offsets and lengths */
1009 #define Node_Offset(n) (RExC_offsets[2*(REGNODE_OFFSET(n))-1])
1010 #define Node_Length(n) (RExC_offsets[2*(REGNODE_OFFSET(n))])
1012 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
1013 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (offset)); \
1014 Set_Node_Length_To_R(REGNODE_OFFSET(node), (len)); \
1017 #define Track_Code(code) STMT_START { code } STMT_END
1020 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
1021 #define EXPERIMENTAL_INPLACESCAN
1022 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
1026 Perl_re_printf(pTHX_ const char *fmt, ...)
1030 PerlIO *f= Perl_debug_log;
1031 PERL_ARGS_ASSERT_RE_PRINTF;
1033 result = PerlIO_vprintf(f, fmt, ap);
1039 Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
1043 PerlIO *f= Perl_debug_log;
1044 PERL_ARGS_ASSERT_RE_INDENTF;
1045 va_start(ap, depth);
1046 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
1047 result = PerlIO_vprintf(f, fmt, ap);
1051 #endif /* DEBUGGING */
1053 #define DEBUG_RExC_seen() \
1054 DEBUG_OPTIMISE_MORE_r({ \
1055 Perl_re_printf( aTHX_ "RExC_seen: "); \
1057 if (RExC_seen & REG_ZERO_LEN_SEEN) \
1058 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
1060 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
1061 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
1063 if (RExC_seen & REG_GPOS_SEEN) \
1064 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
1066 if (RExC_seen & REG_RECURSE_SEEN) \
1067 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
1069 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
1070 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
1072 if (RExC_seen & REG_VERBARG_SEEN) \
1073 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
1075 if (RExC_seen & REG_CUTGROUP_SEEN) \
1076 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
1078 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
1079 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
1081 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
1082 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
1084 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
1085 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
1087 Perl_re_printf( aTHX_ "\n"); \
1090 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
1091 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
1096 S_debug_show_study_flags(pTHX_ U32 flags, const char *open_str,
1097 const char *close_str)
1102 Perl_re_printf( aTHX_ "%s", open_str);
1103 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_SEOL);
1104 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_MEOL);
1105 DEBUG_SHOW_STUDY_FLAG(flags, SF_IS_INF);
1106 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_PAR);
1107 DEBUG_SHOW_STUDY_FLAG(flags, SF_IN_PAR);
1108 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_EVAL);
1109 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_SUBSTR);
1110 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_AND);
1111 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_OR);
1112 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS);
1113 DEBUG_SHOW_STUDY_FLAG(flags, SCF_WHILEM_VISITED_POS);
1114 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_RESTUDY);
1115 DEBUG_SHOW_STUDY_FLAG(flags, SCF_SEEN_ACCEPT);
1116 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_DOING_RESTUDY);
1117 DEBUG_SHOW_STUDY_FLAG(flags, SCF_IN_DEFINE);
1118 Perl_re_printf( aTHX_ "%s", close_str);
1123 S_debug_studydata(pTHX_ const char *where, scan_data_t *data,
1124 U32 depth, int is_inf)
1126 GET_RE_DEBUG_FLAGS_DECL;
1128 DEBUG_OPTIMISE_MORE_r({
1131 Perl_re_indentf(aTHX_ "%s: Pos:%" IVdf "/%" IVdf " Flags: 0x%" UVXf,
1135 (IV)data->pos_delta,
1139 S_debug_show_study_flags(aTHX_ data->flags," [","]");
1141 Perl_re_printf( aTHX_
1142 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s",
1144 (IV)(data->last_closep ? *((data)->last_closep) : -1),
1145 is_inf ? "INF " : ""
1148 if (data->last_found) {
1150 Perl_re_printf(aTHX_
1151 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf,
1152 SvPVX_const(data->last_found),
1154 (IV)data->last_start_min,
1155 (IV)data->last_start_max
1158 for (i = 0; i < 2; i++) {
1159 Perl_re_printf(aTHX_
1160 " %s%s: '%s' @ %" IVdf "/%" IVdf,
1161 data->cur_is_floating == i ? "*" : "",
1162 i ? "Float" : "Fixed",
1163 SvPVX_const(data->substrs[i].str),
1164 (IV)data->substrs[i].min_offset,
1165 (IV)data->substrs[i].max_offset
1167 S_debug_show_study_flags(aTHX_ data->substrs[i].flags," [","]");
1171 Perl_re_printf( aTHX_ "\n");
1177 S_debug_peep(pTHX_ const char *str, const RExC_state_t *pRExC_state,
1178 regnode *scan, U32 depth, U32 flags)
1180 GET_RE_DEBUG_FLAGS_DECL;
1187 Next = regnext(scan);
1188 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
1189 Perl_re_indentf( aTHX_ "%s>%3d: %s (%d)",
1192 REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),
1193 Next ? (REG_NODE_NUM(Next)) : 0 );
1194 S_debug_show_study_flags(aTHX_ flags," [ ","]");
1195 Perl_re_printf( aTHX_ "\n");
1200 # define DEBUG_STUDYDATA(where, data, depth, is_inf) \
1201 S_debug_studydata(aTHX_ where, data, depth, is_inf)
1203 # define DEBUG_PEEP(str, scan, depth, flags) \
1204 S_debug_peep(aTHX_ str, pRExC_state, scan, depth, flags)
1207 # define DEBUG_STUDYDATA(where, data, depth, is_inf) NOOP
1208 # define DEBUG_PEEP(str, scan, depth, flags) NOOP
1212 /* =========================================================
1213 * BEGIN edit_distance stuff.
1215 * This calculates how many single character changes of any type are needed to
1216 * transform a string into another one. It is taken from version 3.1 of
1218 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1221 /* Our unsorted dictionary linked list. */
1222 /* Note we use UVs, not chars. */
1227 struct dictionary* next;
1229 typedef struct dictionary item;
1232 PERL_STATIC_INLINE item*
1233 push(UV key, item* curr)
1236 Newx(head, 1, item);
1244 PERL_STATIC_INLINE item*
1245 find(item* head, UV key)
1247 item* iterator = head;
1249 if (iterator->key == key){
1252 iterator = iterator->next;
1258 PERL_STATIC_INLINE item*
1259 uniquePush(item* head, UV key)
1261 item* iterator = head;
1264 if (iterator->key == key) {
1267 iterator = iterator->next;
1270 return push(key, head);
1273 PERL_STATIC_INLINE void
1274 dict_free(item* head)
1276 item* iterator = head;
1279 item* temp = iterator;
1280 iterator = iterator->next;
1287 /* End of Dictionary Stuff */
1289 /* All calculations/work are done here */
1291 S_edit_distance(const UV* src,
1293 const STRLEN x, /* length of src[] */
1294 const STRLEN y, /* length of tgt[] */
1295 const SSize_t maxDistance
1299 UV swapCount, swapScore, targetCharCount, i, j;
1301 UV score_ceil = x + y;
1303 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1305 /* intialize matrix start values */
1306 Newx(scores, ( (x + 2) * (y + 2)), UV);
1307 scores[0] = score_ceil;
1308 scores[1 * (y + 2) + 0] = score_ceil;
1309 scores[0 * (y + 2) + 1] = score_ceil;
1310 scores[1 * (y + 2) + 1] = 0;
1311 head = uniquePush(uniquePush(head, src[0]), tgt[0]);
1316 for (i=1;i<=x;i++) {
1318 head = uniquePush(head, src[i]);
1319 scores[(i+1) * (y + 2) + 1] = i;
1320 scores[(i+1) * (y + 2) + 0] = score_ceil;
1323 for (j=1;j<=y;j++) {
1326 head = uniquePush(head, tgt[j]);
1327 scores[1 * (y + 2) + (j + 1)] = j;
1328 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1331 targetCharCount = find(head, tgt[j-1])->value;
1332 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1334 if (src[i-1] != tgt[j-1]){
1335 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));
1339 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1343 find(head, src[i-1])->value = i;
1347 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1350 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1354 /* END of edit_distance() stuff
1355 * ========================================================= */
1357 /* is c a control character for which we have a mnemonic? */
1358 #define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
1361 S_cntrl_to_mnemonic(const U8 c)
1363 /* Returns the mnemonic string that represents character 'c', if one
1364 * exists; NULL otherwise. The only ones that exist for the purposes of
1365 * this routine are a few control characters */
1368 case '\a': return "\\a";
1369 case '\b': return "\\b";
1370 case ESC_NATIVE: return "\\e";
1371 case '\f': return "\\f";
1372 case '\n': return "\\n";
1373 case '\r': return "\\r";
1374 case '\t': return "\\t";
1380 /* Mark that we cannot extend a found fixed substring at this point.
1381 Update the longest found anchored substring or the longest found
1382 floating substrings if needed. */
1385 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1386 SSize_t *minlenp, int is_inf)
1388 const STRLEN l = CHR_SVLEN(data->last_found);
1389 SV * const longest_sv = data->substrs[data->cur_is_floating].str;
1390 const STRLEN old_l = CHR_SVLEN(longest_sv);
1391 GET_RE_DEBUG_FLAGS_DECL;
1393 PERL_ARGS_ASSERT_SCAN_COMMIT;
1395 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
1396 const U8 i = data->cur_is_floating;
1397 SvSetMagicSV(longest_sv, data->last_found);
1398 data->substrs[i].min_offset = l ? data->last_start_min : data->pos_min;
1401 data->substrs[0].max_offset = data->substrs[0].min_offset;
1403 data->substrs[1].max_offset = (l
1404 ? data->last_start_max
1405 : (data->pos_delta > SSize_t_MAX - data->pos_min
1407 : data->pos_min + data->pos_delta));
1409 || (STRLEN)data->substrs[1].max_offset > (STRLEN)SSize_t_MAX)
1410 data->substrs[1].max_offset = SSize_t_MAX;
1413 if (data->flags & SF_BEFORE_EOL)
1414 data->substrs[i].flags |= (data->flags & SF_BEFORE_EOL);
1416 data->substrs[i].flags &= ~SF_BEFORE_EOL;
1417 data->substrs[i].minlenp = minlenp;
1418 data->substrs[i].lookbehind = 0;
1421 SvCUR_set(data->last_found, 0);
1423 SV * const sv = data->last_found;
1424 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1425 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1430 data->last_end = -1;
1431 data->flags &= ~SF_BEFORE_EOL;
1432 DEBUG_STUDYDATA("commit", data, 0, is_inf);
1435 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1436 * list that describes which code points it matches */
1439 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1441 /* Set the SSC 'ssc' to match an empty string or any code point */
1443 PERL_ARGS_ASSERT_SSC_ANYTHING;
1445 assert(is_ANYOF_SYNTHETIC(ssc));
1447 /* mortalize so won't leak */
1448 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
1449 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1453 S_ssc_is_anything(const regnode_ssc *ssc)
1455 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1456 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1457 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1458 * in any way, so there's no point in using it */
1463 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1465 assert(is_ANYOF_SYNTHETIC(ssc));
1467 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1471 /* See if the list consists solely of the range 0 - Infinity */
1472 invlist_iterinit(ssc->invlist);
1473 ret = invlist_iternext(ssc->invlist, &start, &end)
1477 invlist_iterfinish(ssc->invlist);
1483 /* If e.g., both \w and \W are set, matches everything */
1484 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1486 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1487 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1497 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1499 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1500 * string, any code point, or any posix class under locale */
1502 PERL_ARGS_ASSERT_SSC_INIT;
1504 Zero(ssc, 1, regnode_ssc);
1505 set_ANYOF_SYNTHETIC(ssc);
1506 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1509 /* If any portion of the regex is to operate under locale rules that aren't
1510 * fully known at compile time, initialization includes it. The reason
1511 * this isn't done for all regexes is that the optimizer was written under
1512 * the assumption that locale was all-or-nothing. Given the complexity and
1513 * lack of documentation in the optimizer, and that there are inadequate
1514 * test cases for locale, many parts of it may not work properly, it is
1515 * safest to avoid locale unless necessary. */
1516 if (RExC_contains_locale) {
1517 ANYOF_POSIXL_SETALL(ssc);
1520 ANYOF_POSIXL_ZERO(ssc);
1525 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1526 const regnode_ssc *ssc)
1528 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1529 * to the list of code points matched, and locale posix classes; hence does
1530 * not check its flags) */
1535 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1537 assert(is_ANYOF_SYNTHETIC(ssc));
1539 invlist_iterinit(ssc->invlist);
1540 ret = invlist_iternext(ssc->invlist, &start, &end)
1544 invlist_iterfinish(ssc->invlist);
1550 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1557 #define INVLIST_INDEX 0
1558 #define ONLY_LOCALE_MATCHES_INDEX 1
1559 #define DEFERRED_USER_DEFINED_INDEX 2
1562 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1563 const regnode_charclass* const node)
1565 /* Returns a mortal inversion list defining which code points are matched
1566 * by 'node', which is of type ANYOF. Handles complementing the result if
1567 * appropriate. If some code points aren't knowable at this time, the
1568 * returned list must, and will, contain every code point that is a
1573 SV* only_utf8_locale_invlist = NULL;
1575 const U32 n = ARG(node);
1576 bool new_node_has_latin1 = FALSE;
1577 const U8 flags = OP(node) == ANYOFH ? 0 : ANYOF_FLAGS(node);
1579 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1581 /* Look at the data structure created by S_set_ANYOF_arg() */
1582 if (n != ANYOF_ONLY_HAS_BITMAP) {
1583 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1584 AV * const av = MUTABLE_AV(SvRV(rv));
1585 SV **const ary = AvARRAY(av);
1586 assert(RExC_rxi->data->what[n] == 's');
1588 if (av_tindex_skip_len_mg(av) >= DEFERRED_USER_DEFINED_INDEX) {
1590 /* Here there are things that won't be known until runtime -- we
1591 * have to assume it could be anything */
1592 invlist = sv_2mortal(_new_invlist(1));
1593 return _add_range_to_invlist(invlist, 0, UV_MAX);
1595 else if (ary[INVLIST_INDEX]) {
1597 /* Use the node's inversion list */
1598 invlist = sv_2mortal(invlist_clone(ary[INVLIST_INDEX], NULL));
1601 /* Get the code points valid only under UTF-8 locales */
1602 if ( (flags & ANYOFL_FOLD)
1603 && av_tindex_skip_len_mg(av) >= ONLY_LOCALE_MATCHES_INDEX)
1605 only_utf8_locale_invlist = ary[ONLY_LOCALE_MATCHES_INDEX];
1610 invlist = sv_2mortal(_new_invlist(0));
1613 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1614 * code points, and an inversion list for the others, but if there are code
1615 * points that should match only conditionally on the target string being
1616 * UTF-8, those are placed in the inversion list, and not the bitmap.
1617 * Since there are circumstances under which they could match, they are
1618 * included in the SSC. But if the ANYOF node is to be inverted, we have
1619 * to exclude them here, so that when we invert below, the end result
1620 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1621 * have to do this here before we add the unconditionally matched code
1623 if (flags & ANYOF_INVERT) {
1624 _invlist_intersection_complement_2nd(invlist,
1629 /* Add in the points from the bit map */
1630 if (OP(node) != ANYOFH) {
1631 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1632 if (ANYOF_BITMAP_TEST(node, i)) {
1633 unsigned int start = i++;
1635 for (; i < NUM_ANYOF_CODE_POINTS
1636 && ANYOF_BITMAP_TEST(node, i); ++i)
1640 invlist = _add_range_to_invlist(invlist, start, i-1);
1641 new_node_has_latin1 = TRUE;
1646 /* If this can match all upper Latin1 code points, have to add them
1647 * as well. But don't add them if inverting, as when that gets done below,
1648 * it would exclude all these characters, including the ones it shouldn't
1649 * that were added just above */
1650 if (! (flags & ANYOF_INVERT) && OP(node) == ANYOFD
1651 && (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1653 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1656 /* Similarly for these */
1657 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1658 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1661 if (flags & ANYOF_INVERT) {
1662 _invlist_invert(invlist);
1664 else if (flags & ANYOFL_FOLD) {
1665 if (new_node_has_latin1) {
1667 /* Under /li, any 0-255 could fold to any other 0-255, depending on
1668 * the locale. We can skip this if there are no 0-255 at all. */
1669 _invlist_union(invlist, PL_Latin1, &invlist);
1671 invlist = add_cp_to_invlist(invlist, LATIN_SMALL_LETTER_DOTLESS_I);
1672 invlist = add_cp_to_invlist(invlist, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
1675 if (_invlist_contains_cp(invlist, LATIN_SMALL_LETTER_DOTLESS_I)) {
1676 invlist = add_cp_to_invlist(invlist, 'I');
1678 if (_invlist_contains_cp(invlist,
1679 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE))
1681 invlist = add_cp_to_invlist(invlist, 'i');
1686 /* Similarly add the UTF-8 locale possible matches. These have to be
1687 * deferred until after the non-UTF-8 locale ones are taken care of just
1688 * above, or it leads to wrong results under ANYOF_INVERT */
1689 if (only_utf8_locale_invlist) {
1690 _invlist_union_maybe_complement_2nd(invlist,
1691 only_utf8_locale_invlist,
1692 flags & ANYOF_INVERT,
1699 /* These two functions currently do the exact same thing */
1700 #define ssc_init_zero ssc_init
1702 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1703 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1705 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1706 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1707 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1710 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1711 const regnode_charclass *and_with)
1713 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1714 * another SSC or a regular ANYOF class. Can create false positives. */
1717 U8 and_with_flags = (OP(and_with) == ANYOFH) ? 0 : ANYOF_FLAGS(and_with);
1720 PERL_ARGS_ASSERT_SSC_AND;
1722 assert(is_ANYOF_SYNTHETIC(ssc));
1724 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1725 * the code point inversion list and just the relevant flags */
1726 if (is_ANYOF_SYNTHETIC(and_with)) {
1727 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1728 anded_flags = and_with_flags;
1730 /* XXX This is a kludge around what appears to be deficiencies in the
1731 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1732 * there are paths through the optimizer where it doesn't get weeded
1733 * out when it should. And if we don't make some extra provision for
1734 * it like the code just below, it doesn't get added when it should.
1735 * This solution is to add it only when AND'ing, which is here, and
1736 * only when what is being AND'ed is the pristine, original node
1737 * matching anything. Thus it is like adding it to ssc_anything() but
1738 * only when the result is to be AND'ed. Probably the same solution
1739 * could be adopted for the same problem we have with /l matching,
1740 * which is solved differently in S_ssc_init(), and that would lead to
1741 * fewer false positives than that solution has. But if this solution
1742 * creates bugs, the consequences are only that a warning isn't raised
1743 * that should be; while the consequences for having /l bugs is
1744 * incorrect matches */
1745 if (ssc_is_anything((regnode_ssc *)and_with)) {
1746 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1750 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1751 if (OP(and_with) == ANYOFD) {
1752 anded_flags = and_with_flags & ANYOF_COMMON_FLAGS;
1755 anded_flags = and_with_flags
1756 &( ANYOF_COMMON_FLAGS
1757 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1758 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1759 if (ANYOFL_UTF8_LOCALE_REQD(and_with_flags)) {
1761 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1766 ANYOF_FLAGS(ssc) &= anded_flags;
1768 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1769 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1770 * 'and_with' may be inverted. When not inverted, we have the situation of
1772 * (C1 | P1) & (C2 | P2)
1773 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1774 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1775 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1776 * <= ((C1 & C2) | P1 | P2)
1777 * Alternatively, the last few steps could be:
1778 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1779 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1780 * <= (C1 | C2 | (P1 & P2))
1781 * We favor the second approach if either P1 or P2 is non-empty. This is
1782 * because these components are a barrier to doing optimizations, as what
1783 * they match cannot be known until the moment of matching as they are
1784 * dependent on the current locale, 'AND"ing them likely will reduce or
1786 * But we can do better if we know that C1,P1 are in their initial state (a
1787 * frequent occurrence), each matching everything:
1788 * (<everything>) & (C2 | P2) = C2 | P2
1789 * Similarly, if C2,P2 are in their initial state (again a frequent
1790 * occurrence), the result is a no-op
1791 * (C1 | P1) & (<everything>) = C1 | P1
1794 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1795 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1796 * <= (C1 & ~C2) | (P1 & ~P2)
1799 if ((and_with_flags & ANYOF_INVERT)
1800 && ! is_ANYOF_SYNTHETIC(and_with))
1804 ssc_intersection(ssc,
1806 FALSE /* Has already been inverted */
1809 /* If either P1 or P2 is empty, the intersection will be also; can skip
1811 if (! (and_with_flags & ANYOF_MATCHES_POSIXL)) {
1812 ANYOF_POSIXL_ZERO(ssc);
1814 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1816 /* Note that the Posix class component P from 'and_with' actually
1818 * P = Pa | Pb | ... | Pn
1819 * where each component is one posix class, such as in [\w\s].
1821 * ~P = ~(Pa | Pb | ... | Pn)
1822 * = ~Pa & ~Pb & ... & ~Pn
1823 * <= ~Pa | ~Pb | ... | ~Pn
1824 * The last is something we can easily calculate, but unfortunately
1825 * is likely to have many false positives. We could do better
1826 * in some (but certainly not all) instances if two classes in
1827 * P have known relationships. For example
1828 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1830 * :lower: & :print: = :lower:
1831 * And similarly for classes that must be disjoint. For example,
1832 * since \s and \w can have no elements in common based on rules in
1833 * the POSIX standard,
1834 * \w & ^\S = nothing
1835 * Unfortunately, some vendor locales do not meet the Posix
1836 * standard, in particular almost everything by Microsoft.
1837 * The loop below just changes e.g., \w into \W and vice versa */
1839 regnode_charclass_posixl temp;
1840 int add = 1; /* To calculate the index of the complement */
1842 Zero(&temp, 1, regnode_charclass_posixl);
1843 ANYOF_POSIXL_ZERO(&temp);
1844 for (i = 0; i < ANYOF_MAX; i++) {
1846 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1847 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1849 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1850 ANYOF_POSIXL_SET(&temp, i + add);
1852 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1854 ANYOF_POSIXL_AND(&temp, ssc);
1856 } /* else ssc already has no posixes */
1857 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1858 in its initial state */
1859 else if (! is_ANYOF_SYNTHETIC(and_with)
1860 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1862 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1863 * copy it over 'ssc' */
1864 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1865 if (is_ANYOF_SYNTHETIC(and_with)) {
1866 StructCopy(and_with, ssc, regnode_ssc);
1869 ssc->invlist = anded_cp_list;
1870 ANYOF_POSIXL_ZERO(ssc);
1871 if (and_with_flags & ANYOF_MATCHES_POSIXL) {
1872 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1876 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1877 || (and_with_flags & ANYOF_MATCHES_POSIXL))
1879 /* One or the other of P1, P2 is non-empty. */
1880 if (and_with_flags & ANYOF_MATCHES_POSIXL) {
1881 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1883 ssc_union(ssc, anded_cp_list, FALSE);
1885 else { /* P1 = P2 = empty */
1886 ssc_intersection(ssc, anded_cp_list, FALSE);
1892 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1893 const regnode_charclass *or_with)
1895 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1896 * another SSC or a regular ANYOF class. Can create false positives if
1897 * 'or_with' is to be inverted. */
1901 U8 or_with_flags = (OP(or_with) == ANYOFH) ? 0 : ANYOF_FLAGS(or_with);
1903 PERL_ARGS_ASSERT_SSC_OR;
1905 assert(is_ANYOF_SYNTHETIC(ssc));
1907 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1908 * the code point inversion list and just the relevant flags */
1909 if (is_ANYOF_SYNTHETIC(or_with)) {
1910 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1911 ored_flags = or_with_flags;
1914 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1915 ored_flags = or_with_flags & ANYOF_COMMON_FLAGS;
1916 if (OP(or_with) != ANYOFD) {
1919 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1920 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1921 if (ANYOFL_UTF8_LOCALE_REQD(or_with_flags)) {
1923 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1928 ANYOF_FLAGS(ssc) |= ored_flags;
1930 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1931 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1932 * 'or_with' may be inverted. When not inverted, we have the simple
1933 * situation of computing:
1934 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1935 * If P1|P2 yields a situation with both a class and its complement are
1936 * set, like having both \w and \W, this matches all code points, and we
1937 * can delete these from the P component of the ssc going forward. XXX We
1938 * might be able to delete all the P components, but I (khw) am not certain
1939 * about this, and it is better to be safe.
1942 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1943 * <= (C1 | P1) | ~C2
1944 * <= (C1 | ~C2) | P1
1945 * (which results in actually simpler code than the non-inverted case)
1948 if ((or_with_flags & ANYOF_INVERT)
1949 && ! is_ANYOF_SYNTHETIC(or_with))
1951 /* We ignore P2, leaving P1 going forward */
1952 } /* else Not inverted */
1953 else if (or_with_flags & ANYOF_MATCHES_POSIXL) {
1954 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1955 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1957 for (i = 0; i < ANYOF_MAX; i += 2) {
1958 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1960 ssc_match_all_cp(ssc);
1961 ANYOF_POSIXL_CLEAR(ssc, i);
1962 ANYOF_POSIXL_CLEAR(ssc, i+1);
1970 FALSE /* Already has been inverted */
1974 PERL_STATIC_INLINE void
1975 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1977 PERL_ARGS_ASSERT_SSC_UNION;
1979 assert(is_ANYOF_SYNTHETIC(ssc));
1981 _invlist_union_maybe_complement_2nd(ssc->invlist,
1987 PERL_STATIC_INLINE void
1988 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1990 const bool invert2nd)
1992 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1994 assert(is_ANYOF_SYNTHETIC(ssc));
1996 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
2002 PERL_STATIC_INLINE void
2003 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
2005 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
2007 assert(is_ANYOF_SYNTHETIC(ssc));
2009 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
2012 PERL_STATIC_INLINE void
2013 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
2015 /* AND just the single code point 'cp' into the SSC 'ssc' */
2017 SV* cp_list = _new_invlist(2);
2019 PERL_ARGS_ASSERT_SSC_CP_AND;
2021 assert(is_ANYOF_SYNTHETIC(ssc));
2023 cp_list = add_cp_to_invlist(cp_list, cp);
2024 ssc_intersection(ssc, cp_list,
2025 FALSE /* Not inverted */
2027 SvREFCNT_dec_NN(cp_list);
2030 PERL_STATIC_INLINE void
2031 S_ssc_clear_locale(regnode_ssc *ssc)
2033 /* Set the SSC 'ssc' to not match any locale things */
2034 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
2036 assert(is_ANYOF_SYNTHETIC(ssc));
2038 ANYOF_POSIXL_ZERO(ssc);
2039 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
2042 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
2045 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
2047 /* The synthetic start class is used to hopefully quickly winnow down
2048 * places where a pattern could start a match in the target string. If it
2049 * doesn't really narrow things down that much, there isn't much point to
2050 * having the overhead of using it. This function uses some very crude
2051 * heuristics to decide if to use the ssc or not.
2053 * It returns TRUE if 'ssc' rules out more than half what it considers to
2054 * be the "likely" possible matches, but of course it doesn't know what the
2055 * actual things being matched are going to be; these are only guesses
2057 * For /l matches, it assumes that the only likely matches are going to be
2058 * in the 0-255 range, uniformly distributed, so half of that is 127
2059 * For /a and /d matches, it assumes that the likely matches will be just
2060 * the ASCII range, so half of that is 63
2061 * For /u and there isn't anything matching above the Latin1 range, it
2062 * assumes that that is the only range likely to be matched, and uses
2063 * half that as the cut-off: 127. If anything matches above Latin1,
2064 * it assumes that all of Unicode could match (uniformly), except for
2065 * non-Unicode code points and things in the General Category "Other"
2066 * (unassigned, private use, surrogates, controls and formats). This
2067 * is a much large number. */
2069 U32 count = 0; /* Running total of number of code points matched by
2071 UV start, end; /* Start and end points of current range in inversion
2072 XXX outdated. UTF-8 locales are common, what about invert? list */
2073 const U32 max_code_points = (LOC)
2075 : (( ! UNI_SEMANTICS
2076 || invlist_highest(ssc->invlist) < 256)
2079 const U32 max_match = max_code_points / 2;
2081 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
2083 invlist_iterinit(ssc->invlist);
2084 while (invlist_iternext(ssc->invlist, &start, &end)) {
2085 if (start >= max_code_points) {
2088 end = MIN(end, max_code_points - 1);
2089 count += end - start + 1;
2090 if (count >= max_match) {
2091 invlist_iterfinish(ssc->invlist);
2101 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
2103 /* The inversion list in the SSC is marked mortal; now we need a more
2104 * permanent copy, which is stored the same way that is done in a regular
2105 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
2108 SV* invlist = invlist_clone(ssc->invlist, NULL);
2110 PERL_ARGS_ASSERT_SSC_FINALIZE;
2112 assert(is_ANYOF_SYNTHETIC(ssc));
2114 /* The code in this file assumes that all but these flags aren't relevant
2115 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
2116 * by the time we reach here */
2117 assert(! (ANYOF_FLAGS(ssc)
2118 & ~( ANYOF_COMMON_FLAGS
2119 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
2120 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
2122 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
2124 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist, NULL, NULL);
2126 /* Make sure is clone-safe */
2127 ssc->invlist = NULL;
2129 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
2130 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
2131 OP(ssc) = ANYOFPOSIXL;
2133 else if (RExC_contains_locale) {
2137 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
2140 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
2141 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
2142 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
2143 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
2144 ? (TRIE_LIST_CUR( idx ) - 1) \
2150 dump_trie(trie,widecharmap,revcharmap)
2151 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
2152 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
2154 These routines dump out a trie in a somewhat readable format.
2155 The _interim_ variants are used for debugging the interim
2156 tables that are used to generate the final compressed
2157 representation which is what dump_trie expects.
2159 Part of the reason for their existence is to provide a form
2160 of documentation as to how the different representations function.
2165 Dumps the final compressed table form of the trie to Perl_debug_log.
2166 Used for debugging make_trie().
2170 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
2171 AV *revcharmap, U32 depth)
2174 SV *sv=sv_newmortal();
2175 int colwidth= widecharmap ? 6 : 4;
2177 GET_RE_DEBUG_FLAGS_DECL;
2179 PERL_ARGS_ASSERT_DUMP_TRIE;
2181 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
2182 depth+1, "Match","Base","Ofs" );
2184 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2185 SV ** const tmp = av_fetch( revcharmap, state, 0);
2187 Perl_re_printf( aTHX_ "%*s",
2189 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2190 PL_colors[0], PL_colors[1],
2191 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2192 PERL_PV_ESCAPE_FIRSTCHAR
2197 Perl_re_printf( aTHX_ "\n");
2198 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
2200 for( state = 0 ; state < trie->uniquecharcount ; state++ )
2201 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2202 Perl_re_printf( aTHX_ "\n");
2204 for( state = 1 ; state < trie->statecount ; state++ ) {
2205 const U32 base = trie->states[ state ].trans.base;
2207 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
2209 if ( trie->states[ state ].wordnum ) {
2210 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
2212 Perl_re_printf( aTHX_ "%6s", "" );
2215 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
2220 while( ( base + ofs < trie->uniquecharcount ) ||
2221 ( base + ofs - trie->uniquecharcount < trie->lasttrans
2222 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2226 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
2228 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2229 if ( ( base + ofs >= trie->uniquecharcount )
2230 && ( base + ofs - trie->uniquecharcount
2232 && trie->trans[ base + ofs
2233 - trie->uniquecharcount ].check == state )
2235 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
2236 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2239 Perl_re_printf( aTHX_ "%*s", colwidth," ." );
2243 Perl_re_printf( aTHX_ "]");
2246 Perl_re_printf( aTHX_ "\n" );
2248 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
2250 for (word=1; word <= trie->wordcount; word++) {
2251 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2252 (int)word, (int)(trie->wordinfo[word].prev),
2253 (int)(trie->wordinfo[word].len));
2255 Perl_re_printf( aTHX_ "\n" );
2258 Dumps a fully constructed but uncompressed trie in list form.
2259 List tries normally only are used for construction when the number of
2260 possible chars (trie->uniquecharcount) is very high.
2261 Used for debugging make_trie().
2264 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2265 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2269 SV *sv=sv_newmortal();
2270 int colwidth= widecharmap ? 6 : 4;
2271 GET_RE_DEBUG_FLAGS_DECL;
2273 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2275 /* print out the table precompression. */
2276 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
2278 Perl_re_indentf( aTHX_ "%s",
2279 depth+1, "------:-----+-----------------\n" );
2281 for( state=1 ; state < next_alloc ; state ++ ) {
2284 Perl_re_indentf( aTHX_ " %4" UVXf " :",
2285 depth+1, (UV)state );
2286 if ( ! trie->states[ state ].wordnum ) {
2287 Perl_re_printf( aTHX_ "%5s| ","");
2289 Perl_re_printf( aTHX_ "W%4x| ",
2290 trie->states[ state ].wordnum
2293 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2294 SV ** const tmp = av_fetch( revcharmap,
2295 TRIE_LIST_ITEM(state, charid).forid, 0);
2297 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
2299 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2301 PL_colors[0], PL_colors[1],
2302 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2303 | PERL_PV_ESCAPE_FIRSTCHAR
2305 TRIE_LIST_ITEM(state, charid).forid,
2306 (UV)TRIE_LIST_ITEM(state, charid).newstate
2309 Perl_re_printf( aTHX_ "\n%*s| ",
2310 (int)((depth * 2) + 14), "");
2313 Perl_re_printf( aTHX_ "\n");
2318 Dumps a fully constructed but uncompressed trie in table form.
2319 This is the normal DFA style state transition table, with a few
2320 twists to facilitate compression later.
2321 Used for debugging make_trie().
2324 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2325 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2330 SV *sv=sv_newmortal();
2331 int colwidth= widecharmap ? 6 : 4;
2332 GET_RE_DEBUG_FLAGS_DECL;
2334 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2337 print out the table precompression so that we can do a visual check
2338 that they are identical.
2341 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
2343 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2344 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2346 Perl_re_printf( aTHX_ "%*s",
2348 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2349 PL_colors[0], PL_colors[1],
2350 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2351 PERL_PV_ESCAPE_FIRSTCHAR
2357 Perl_re_printf( aTHX_ "\n");
2358 Perl_re_indentf( aTHX_ "State+-", depth+1 );
2360 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2361 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
2364 Perl_re_printf( aTHX_ "\n" );
2366 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2368 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
2370 (UV)TRIE_NODENUM( state ) );
2372 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2373 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2375 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
2377 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
2379 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2380 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
2381 (UV)trie->trans[ state ].check );
2383 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
2384 (UV)trie->trans[ state ].check,
2385 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2393 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2394 startbranch: the first branch in the whole branch sequence
2395 first : start branch of sequence of branch-exact nodes.
2396 May be the same as startbranch
2397 last : Thing following the last branch.
2398 May be the same as tail.
2399 tail : item following the branch sequence
2400 count : words in the sequence
2401 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2402 depth : indent depth
2404 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2406 A trie is an N'ary tree where the branches are determined by digital
2407 decomposition of the key. IE, at the root node you look up the 1st character and
2408 follow that branch repeat until you find the end of the branches. Nodes can be
2409 marked as "accepting" meaning they represent a complete word. Eg:
2413 would convert into the following structure. Numbers represent states, letters
2414 following numbers represent valid transitions on the letter from that state, if
2415 the number is in square brackets it represents an accepting state, otherwise it
2416 will be in parenthesis.
2418 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2422 (1) +-i->(6)-+-s->[7]
2424 +-s->(3)-+-h->(4)-+-e->[5]
2426 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2428 This shows that when matching against the string 'hers' we will begin at state 1
2429 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2430 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2431 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2432 single traverse. We store a mapping from accepting to state to which word was
2433 matched, and then when we have multiple possibilities we try to complete the
2434 rest of the regex in the order in which they occurred in the alternation.
2436 The only prior NFA like behaviour that would be changed by the TRIE support is
2437 the silent ignoring of duplicate alternations which are of the form:
2439 / (DUPE|DUPE) X? (?{ ... }) Y /x
2441 Thus EVAL blocks following a trie may be called a different number of times with
2442 and without the optimisation. With the optimisations dupes will be silently
2443 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2444 the following demonstrates:
2446 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2448 which prints out 'word' three times, but
2450 'words'=~/(word|word|word)(?{ print $1 })S/
2452 which doesnt print it out at all. This is due to other optimisations kicking in.
2454 Example of what happens on a structural level:
2456 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2458 1: CURLYM[1] {1,32767}(18)
2469 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2470 and should turn into:
2472 1: CURLYM[1] {1,32767}(18)
2474 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2482 Cases where tail != last would be like /(?foo|bar)baz/:
2492 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2493 and would end up looking like:
2496 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2503 d = uvchr_to_utf8_flags(d, uv, 0);
2505 is the recommended Unicode-aware way of saying
2510 #define TRIE_STORE_REVCHAR(val) \
2513 SV *zlopp = newSV(UTF8_MAXBYTES); \
2514 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2515 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2516 SvCUR_set(zlopp, kapow - flrbbbbb); \
2519 av_push(revcharmap, zlopp); \
2521 char ooooff = (char)val; \
2522 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2526 /* This gets the next character from the input, folding it if not already
2528 #define TRIE_READ_CHAR STMT_START { \
2531 /* if it is UTF then it is either already folded, or does not need \
2533 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2535 else if (folder == PL_fold_latin1) { \
2536 /* This folder implies Unicode rules, which in the range expressible \
2537 * by not UTF is the lower case, with the two exceptions, one of \
2538 * which should have been taken care of before calling this */ \
2539 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2540 uvc = toLOWER_L1(*uc); \
2541 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2544 /* raw data, will be folded later if needed */ \
2552 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2553 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2554 U32 ging = TRIE_LIST_LEN( state ) * 2; \
2555 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2556 TRIE_LIST_LEN( state ) = ging; \
2558 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2559 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2560 TRIE_LIST_CUR( state )++; \
2563 #define TRIE_LIST_NEW(state) STMT_START { \
2564 Newx( trie->states[ state ].trans.list, \
2565 4, reg_trie_trans_le ); \
2566 TRIE_LIST_CUR( state ) = 1; \
2567 TRIE_LIST_LEN( state ) = 4; \
2570 #define TRIE_HANDLE_WORD(state) STMT_START { \
2571 U16 dupe= trie->states[ state ].wordnum; \
2572 regnode * const noper_next = regnext( noper ); \
2575 /* store the word for dumping */ \
2577 if (OP(noper) != NOTHING) \
2578 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2580 tmp = newSVpvn_utf8( "", 0, UTF ); \
2581 av_push( trie_words, tmp ); \
2585 trie->wordinfo[curword].prev = 0; \
2586 trie->wordinfo[curword].len = wordlen; \
2587 trie->wordinfo[curword].accept = state; \
2589 if ( noper_next < tail ) { \
2591 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2593 trie->jump[curword] = (U16)(noper_next - convert); \
2595 jumper = noper_next; \
2597 nextbranch= regnext(cur); \
2601 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2602 /* chain, so that when the bits of chain are later */\
2603 /* linked together, the dups appear in the chain */\
2604 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2605 trie->wordinfo[dupe].prev = curword; \
2607 /* we haven't inserted this word yet. */ \
2608 trie->states[ state ].wordnum = curword; \
2613 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2614 ( ( base + charid >= ucharcount \
2615 && base + charid < ubound \
2616 && state == trie->trans[ base - ucharcount + charid ].check \
2617 && trie->trans[ base - ucharcount + charid ].next ) \
2618 ? trie->trans[ base - ucharcount + charid ].next \
2619 : ( state==1 ? special : 0 ) \
2622 #define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2624 TRIE_BITMAP_SET(trie, uvc); \
2625 /* store the folded codepoint */ \
2627 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2630 /* store first byte of utf8 representation of */ \
2631 /* variant codepoints */ \
2632 if (! UVCHR_IS_INVARIANT(uvc)) { \
2633 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2638 #define MADE_JUMP_TRIE 2
2639 #define MADE_EXACT_TRIE 4
2642 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2643 regnode *first, regnode *last, regnode *tail,
2644 U32 word_count, U32 flags, U32 depth)
2646 /* first pass, loop through and scan words */
2647 reg_trie_data *trie;
2648 HV *widecharmap = NULL;
2649 AV *revcharmap = newAV();
2655 regnode *jumper = NULL;
2656 regnode *nextbranch = NULL;
2657 regnode *convert = NULL;
2658 U32 *prev_states; /* temp array mapping each state to previous one */
2659 /* we just use folder as a flag in utf8 */
2660 const U8 * folder = NULL;
2662 /* in the below add_data call we are storing either 'tu' or 'tuaa'
2663 * which stands for one trie structure, one hash, optionally followed
2666 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuaa"));
2667 AV *trie_words = NULL;
2668 /* along with revcharmap, this only used during construction but both are
2669 * useful during debugging so we store them in the struct when debugging.
2672 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2673 STRLEN trie_charcount=0;
2675 SV *re_trie_maxbuff;
2676 GET_RE_DEBUG_FLAGS_DECL;
2678 PERL_ARGS_ASSERT_MAKE_TRIE;
2680 PERL_UNUSED_ARG(depth);
2684 case EXACT: case EXACT_ONLY8: case EXACTL: break;
2688 case EXACTFLU8: folder = PL_fold_latin1; break;
2689 case EXACTF: folder = PL_fold; break;
2690 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2693 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2695 trie->startstate = 1;
2696 trie->wordcount = word_count;
2697 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2698 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2699 if (flags == EXACT || flags == EXACT_ONLY8 || flags == EXACTL)
2700 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2701 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2702 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2705 trie_words = newAV();
2708 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, GV_ADD);
2709 assert(re_trie_maxbuff);
2710 if (!SvIOK(re_trie_maxbuff)) {
2711 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2713 DEBUG_TRIE_COMPILE_r({
2714 Perl_re_indentf( aTHX_
2715 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2717 REG_NODE_NUM(startbranch), REG_NODE_NUM(first),
2718 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2721 /* Find the node we are going to overwrite */
2722 if ( first == startbranch && OP( last ) != BRANCH ) {
2723 /* whole branch chain */
2726 /* branch sub-chain */
2727 convert = NEXTOPER( first );
2730 /* -- First loop and Setup --
2732 We first traverse the branches and scan each word to determine if it
2733 contains widechars, and how many unique chars there are, this is
2734 important as we have to build a table with at least as many columns as we
2737 We use an array of integers to represent the character codes 0..255
2738 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2739 the native representation of the character value as the key and IV's for
2742 *TODO* If we keep track of how many times each character is used we can
2743 remap the columns so that the table compression later on is more
2744 efficient in terms of memory by ensuring the most common value is in the
2745 middle and the least common are on the outside. IMO this would be better
2746 than a most to least common mapping as theres a decent chance the most
2747 common letter will share a node with the least common, meaning the node
2748 will not be compressible. With a middle is most common approach the worst
2749 case is when we have the least common nodes twice.
2753 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2754 regnode *noper = NEXTOPER( cur );
2758 U32 wordlen = 0; /* required init */
2759 STRLEN minchars = 0;
2760 STRLEN maxchars = 0;
2761 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2764 if (OP(noper) == NOTHING) {
2765 /* skip past a NOTHING at the start of an alternation
2766 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2768 regnode *noper_next= regnext(noper);
2769 if (noper_next < tail)
2774 && ( OP(noper) == flags
2775 || (flags == EXACT && OP(noper) == EXACT_ONLY8)
2776 || (flags == EXACTFU && ( OP(noper) == EXACTFU_ONLY8
2777 || OP(noper) == EXACTFUP))))
2779 uc= (U8*)STRING(noper);
2780 e= uc + STR_LEN(noper);
2787 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2788 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2789 regardless of encoding */
2790 if (OP( noper ) == EXACTFUP) {
2791 /* false positives are ok, so just set this */
2792 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2796 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2798 TRIE_CHARCOUNT(trie)++;
2801 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2802 * is in effect. Under /i, this character can match itself, or
2803 * anything that folds to it. If not under /i, it can match just
2804 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2805 * all fold to k, and all are single characters. But some folds
2806 * expand to more than one character, so for example LATIN SMALL
2807 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2808 * the string beginning at 'uc' is 'ffi', it could be matched by
2809 * three characters, or just by the one ligature character. (It
2810 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2811 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2812 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2813 * match.) The trie needs to know the minimum and maximum number
2814 * of characters that could match so that it can use size alone to
2815 * quickly reject many match attempts. The max is simple: it is
2816 * the number of folded characters in this branch (since a fold is
2817 * never shorter than what folds to it. */
2821 /* And the min is equal to the max if not under /i (indicated by
2822 * 'folder' being NULL), or there are no multi-character folds. If
2823 * there is a multi-character fold, the min is incremented just
2824 * once, for the character that folds to the sequence. Each
2825 * character in the sequence needs to be added to the list below of
2826 * characters in the trie, but we count only the first towards the
2827 * min number of characters needed. This is done through the
2828 * variable 'foldlen', which is returned by the macros that look
2829 * for these sequences as the number of bytes the sequence
2830 * occupies. Each time through the loop, we decrement 'foldlen' by
2831 * how many bytes the current char occupies. Only when it reaches
2832 * 0 do we increment 'minchars' or look for another multi-character
2834 if (folder == NULL) {
2837 else if (foldlen > 0) {
2838 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2843 /* See if *uc is the beginning of a multi-character fold. If
2844 * so, we decrement the length remaining to look at, to account
2845 * for the current character this iteration. (We can use 'uc'
2846 * instead of the fold returned by TRIE_READ_CHAR because for
2847 * non-UTF, the latin1_safe macro is smart enough to account
2848 * for all the unfolded characters, and because for UTF, the
2849 * string will already have been folded earlier in the
2850 * compilation process */
2852 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2853 foldlen -= UTF8SKIP(uc);
2856 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2861 /* The current character (and any potential folds) should be added
2862 * to the possible matching characters for this position in this
2866 U8 folded= folder[ (U8) uvc ];
2867 if ( !trie->charmap[ folded ] ) {
2868 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2869 TRIE_STORE_REVCHAR( folded );
2872 if ( !trie->charmap[ uvc ] ) {
2873 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2874 TRIE_STORE_REVCHAR( uvc );
2877 /* store the codepoint in the bitmap, and its folded
2879 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
2880 set_bit = 0; /* We've done our bit :-) */
2884 /* XXX We could come up with the list of code points that fold
2885 * to this using PL_utf8_foldclosures, except not for
2886 * multi-char folds, as there may be multiple combinations
2887 * there that could work, which needs to wait until runtime to
2888 * resolve (The comment about LIGATURE FFI above is such an
2893 widecharmap = newHV();
2895 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2898 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
2900 if ( !SvTRUE( *svpp ) ) {
2901 sv_setiv( *svpp, ++trie->uniquecharcount );
2902 TRIE_STORE_REVCHAR(uvc);
2905 } /* end loop through characters in this branch of the trie */
2907 /* We take the min and max for this branch and combine to find the min
2908 * and max for all branches processed so far */
2909 if( cur == first ) {
2910 trie->minlen = minchars;
2911 trie->maxlen = maxchars;
2912 } else if (minchars < trie->minlen) {
2913 trie->minlen = minchars;
2914 } else if (maxchars > trie->maxlen) {
2915 trie->maxlen = maxchars;
2917 } /* end first pass */
2918 DEBUG_TRIE_COMPILE_r(
2919 Perl_re_indentf( aTHX_
2920 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2922 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2923 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2924 (int)trie->minlen, (int)trie->maxlen )
2928 We now know what we are dealing with in terms of unique chars and
2929 string sizes so we can calculate how much memory a naive
2930 representation using a flat table will take. If it's over a reasonable
2931 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2932 conservative but potentially much slower representation using an array
2935 At the end we convert both representations into the same compressed
2936 form that will be used in regexec.c for matching with. The latter
2937 is a form that cannot be used to construct with but has memory
2938 properties similar to the list form and access properties similar
2939 to the table form making it both suitable for fast searches and
2940 small enough that its feasable to store for the duration of a program.
2942 See the comment in the code where the compressed table is produced
2943 inplace from the flat tabe representation for an explanation of how
2944 the compression works.
2949 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2952 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2953 > SvIV(re_trie_maxbuff) )
2956 Second Pass -- Array Of Lists Representation
2958 Each state will be represented by a list of charid:state records
2959 (reg_trie_trans_le) the first such element holds the CUR and LEN
2960 points of the allocated array. (See defines above).
2962 We build the initial structure using the lists, and then convert
2963 it into the compressed table form which allows faster lookups
2964 (but cant be modified once converted).
2967 STRLEN transcount = 1;
2969 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
2972 trie->states = (reg_trie_state *)
2973 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2974 sizeof(reg_trie_state) );
2978 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2980 regnode *noper = NEXTOPER( cur );
2981 U32 state = 1; /* required init */
2982 U16 charid = 0; /* sanity init */
2983 U32 wordlen = 0; /* required init */
2985 if (OP(noper) == NOTHING) {
2986 regnode *noper_next= regnext(noper);
2987 if (noper_next < tail)
2992 && ( OP(noper) == flags
2993 || (flags == EXACT && OP(noper) == EXACT_ONLY8)
2994 || (flags == EXACTFU && ( OP(noper) == EXACTFU_ONLY8
2995 || OP(noper) == EXACTFUP))))
2997 const U8 *uc= (U8*)STRING(noper);
2998 const U8 *e= uc + STR_LEN(noper);
3000 for ( ; uc < e ; uc += len ) {
3005 charid = trie->charmap[ uvc ];
3007 SV** const svpp = hv_fetch( widecharmap,
3014 charid=(U16)SvIV( *svpp );
3017 /* charid is now 0 if we dont know the char read, or
3018 * nonzero if we do */
3025 if ( !trie->states[ state ].trans.list ) {
3026 TRIE_LIST_NEW( state );
3029 check <= TRIE_LIST_USED( state );
3032 if ( TRIE_LIST_ITEM( state, check ).forid
3035 newstate = TRIE_LIST_ITEM( state, check ).newstate;
3040 newstate = next_alloc++;
3041 prev_states[newstate] = state;
3042 TRIE_LIST_PUSH( state, charid, newstate );
3047 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3051 TRIE_HANDLE_WORD(state);
3053 } /* end second pass */
3055 /* next alloc is the NEXT state to be allocated */
3056 trie->statecount = next_alloc;
3057 trie->states = (reg_trie_state *)
3058 PerlMemShared_realloc( trie->states,
3060 * sizeof(reg_trie_state) );
3062 /* and now dump it out before we compress it */
3063 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
3064 revcharmap, next_alloc,
3068 trie->trans = (reg_trie_trans *)
3069 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
3076 for( state=1 ; state < next_alloc ; state ++ ) {
3080 DEBUG_TRIE_COMPILE_MORE_r(
3081 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
3085 if (trie->states[state].trans.list) {
3086 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
3090 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3091 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
3092 if ( forid < minid ) {
3094 } else if ( forid > maxid ) {
3098 if ( transcount < tp + maxid - minid + 1) {
3100 trie->trans = (reg_trie_trans *)
3101 PerlMemShared_realloc( trie->trans,
3103 * sizeof(reg_trie_trans) );
3104 Zero( trie->trans + (transcount / 2),
3108 base = trie->uniquecharcount + tp - minid;
3109 if ( maxid == minid ) {
3111 for ( ; zp < tp ; zp++ ) {
3112 if ( ! trie->trans[ zp ].next ) {
3113 base = trie->uniquecharcount + zp - minid;
3114 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
3116 trie->trans[ zp ].check = state;
3122 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
3124 trie->trans[ tp ].check = state;
3129 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3130 const U32 tid = base
3131 - trie->uniquecharcount
3132 + TRIE_LIST_ITEM( state, idx ).forid;
3133 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
3135 trie->trans[ tid ].check = state;
3137 tp += ( maxid - minid + 1 );
3139 Safefree(trie->states[ state ].trans.list);
3142 DEBUG_TRIE_COMPILE_MORE_r(
3143 Perl_re_printf( aTHX_ " base: %d\n",base);
3146 trie->states[ state ].trans.base=base;
3148 trie->lasttrans = tp + 1;
3152 Second Pass -- Flat Table Representation.
3154 we dont use the 0 slot of either trans[] or states[] so we add 1 to
3155 each. We know that we will need Charcount+1 trans at most to store
3156 the data (one row per char at worst case) So we preallocate both
3157 structures assuming worst case.
3159 We then construct the trie using only the .next slots of the entry
3162 We use the .check field of the first entry of the node temporarily
3163 to make compression both faster and easier by keeping track of how
3164 many non zero fields are in the node.
3166 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
3169 There are two terms at use here: state as a TRIE_NODEIDX() which is
3170 a number representing the first entry of the node, and state as a
3171 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
3172 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
3173 if there are 2 entrys per node. eg:
3181 The table is internally in the right hand, idx form. However as we
3182 also have to deal with the states array which is indexed by nodenum
3183 we have to use TRIE_NODENUM() to convert.
3186 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
3189 trie->trans = (reg_trie_trans *)
3190 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
3191 * trie->uniquecharcount + 1,
3192 sizeof(reg_trie_trans) );
3193 trie->states = (reg_trie_state *)
3194 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3195 sizeof(reg_trie_state) );
3196 next_alloc = trie->uniquecharcount + 1;
3199 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3201 regnode *noper = NEXTOPER( cur );
3203 U32 state = 1; /* required init */
3205 U16 charid = 0; /* sanity init */
3206 U32 accept_state = 0; /* sanity init */
3208 U32 wordlen = 0; /* required init */
3210 if (OP(noper) == NOTHING) {
3211 regnode *noper_next= regnext(noper);
3212 if (noper_next < tail)
3217 && ( OP(noper) == flags
3218 || (flags == EXACT && OP(noper) == EXACT_ONLY8)
3219 || (flags == EXACTFU && ( OP(noper) == EXACTFU_ONLY8
3220 || OP(noper) == EXACTFUP))))
3222 const U8 *uc= (U8*)STRING(noper);
3223 const U8 *e= uc + STR_LEN(noper);
3225 for ( ; uc < e ; uc += len ) {
3230 charid = trie->charmap[ uvc ];
3232 SV* const * const svpp = hv_fetch( widecharmap,
3236 charid = svpp ? (U16)SvIV(*svpp) : 0;
3240 if ( !trie->trans[ state + charid ].next ) {
3241 trie->trans[ state + charid ].next = next_alloc;
3242 trie->trans[ state ].check++;
3243 prev_states[TRIE_NODENUM(next_alloc)]
3244 = TRIE_NODENUM(state);
3245 next_alloc += trie->uniquecharcount;
3247 state = trie->trans[ state + charid ].next;
3249 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3251 /* charid is now 0 if we dont know the char read, or
3252 * nonzero if we do */
3255 accept_state = TRIE_NODENUM( state );
3256 TRIE_HANDLE_WORD(accept_state);
3258 } /* end second pass */
3260 /* and now dump it out before we compress it */
3261 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3263 next_alloc, depth+1));
3267 * Inplace compress the table.*
3269 For sparse data sets the table constructed by the trie algorithm will
3270 be mostly 0/FAIL transitions or to put it another way mostly empty.
3271 (Note that leaf nodes will not contain any transitions.)
3273 This algorithm compresses the tables by eliminating most such
3274 transitions, at the cost of a modest bit of extra work during lookup:
3276 - Each states[] entry contains a .base field which indicates the
3277 index in the state[] array wheres its transition data is stored.
3279 - If .base is 0 there are no valid transitions from that node.
3281 - If .base is nonzero then charid is added to it to find an entry in
3284 -If trans[states[state].base+charid].check!=state then the
3285 transition is taken to be a 0/Fail transition. Thus if there are fail
3286 transitions at the front of the node then the .base offset will point
3287 somewhere inside the previous nodes data (or maybe even into a node
3288 even earlier), but the .check field determines if the transition is
3292 The following process inplace converts the table to the compressed
3293 table: We first do not compress the root node 1,and mark all its
3294 .check pointers as 1 and set its .base pointer as 1 as well. This
3295 allows us to do a DFA construction from the compressed table later,
3296 and ensures that any .base pointers we calculate later are greater
3299 - We set 'pos' to indicate the first entry of the second node.
3301 - We then iterate over the columns of the node, finding the first and
3302 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3303 and set the .check pointers accordingly, and advance pos
3304 appropriately and repreat for the next node. Note that when we copy
3305 the next pointers we have to convert them from the original
3306 NODEIDX form to NODENUM form as the former is not valid post
3309 - If a node has no transitions used we mark its base as 0 and do not
3310 advance the pos pointer.
3312 - If a node only has one transition we use a second pointer into the
3313 structure to fill in allocated fail transitions from other states.
3314 This pointer is independent of the main pointer and scans forward
3315 looking for null transitions that are allocated to a state. When it
3316 finds one it writes the single transition into the "hole". If the
3317 pointer doesnt find one the single transition is appended as normal.
3319 - Once compressed we can Renew/realloc the structures to release the
3322 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3323 specifically Fig 3.47 and the associated pseudocode.
3327 const U32 laststate = TRIE_NODENUM( next_alloc );
3330 trie->statecount = laststate;
3332 for ( state = 1 ; state < laststate ; state++ ) {
3334 const U32 stateidx = TRIE_NODEIDX( state );
3335 const U32 o_used = trie->trans[ stateidx ].check;
3336 U32 used = trie->trans[ stateidx ].check;
3337 trie->trans[ stateidx ].check = 0;
3340 used && charid < trie->uniquecharcount;
3343 if ( flag || trie->trans[ stateidx + charid ].next ) {
3344 if ( trie->trans[ stateidx + charid ].next ) {
3346 for ( ; zp < pos ; zp++ ) {
3347 if ( ! trie->trans[ zp ].next ) {
3351 trie->states[ state ].trans.base
3353 + trie->uniquecharcount
3355 trie->trans[ zp ].next
3356 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3358 trie->trans[ zp ].check = state;
3359 if ( ++zp > pos ) pos = zp;
3366 trie->states[ state ].trans.base
3367 = pos + trie->uniquecharcount - charid ;
3369 trie->trans[ pos ].next
3370 = SAFE_TRIE_NODENUM(
3371 trie->trans[ stateidx + charid ].next );
3372 trie->trans[ pos ].check = state;
3377 trie->lasttrans = pos + 1;
3378 trie->states = (reg_trie_state *)
3379 PerlMemShared_realloc( trie->states, laststate
3380 * sizeof(reg_trie_state) );
3381 DEBUG_TRIE_COMPILE_MORE_r(
3382 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
3384 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3388 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3391 } /* end table compress */
3393 DEBUG_TRIE_COMPILE_MORE_r(
3394 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
3396 (UV)trie->statecount,
3397 (UV)trie->lasttrans)
3399 /* resize the trans array to remove unused space */
3400 trie->trans = (reg_trie_trans *)
3401 PerlMemShared_realloc( trie->trans, trie->lasttrans
3402 * sizeof(reg_trie_trans) );
3404 { /* Modify the program and insert the new TRIE node */
3405 U8 nodetype =(U8)(flags & 0xFF);
3409 regnode *optimize = NULL;
3410 #ifdef RE_TRACK_PATTERN_OFFSETS
3413 U32 mjd_nodelen = 0;
3414 #endif /* RE_TRACK_PATTERN_OFFSETS */
3415 #endif /* DEBUGGING */
3417 This means we convert either the first branch or the first Exact,
3418 depending on whether the thing following (in 'last') is a branch
3419 or not and whther first is the startbranch (ie is it a sub part of
3420 the alternation or is it the whole thing.)
3421 Assuming its a sub part we convert the EXACT otherwise we convert
3422 the whole branch sequence, including the first.
3424 /* Find the node we are going to overwrite */
3425 if ( first != startbranch || OP( last ) == BRANCH ) {
3426 /* branch sub-chain */
3427 NEXT_OFF( first ) = (U16)(last - first);
3428 #ifdef RE_TRACK_PATTERN_OFFSETS
3430 mjd_offset= Node_Offset((convert));
3431 mjd_nodelen= Node_Length((convert));
3434 /* whole branch chain */
3436 #ifdef RE_TRACK_PATTERN_OFFSETS
3439 const regnode *nop = NEXTOPER( convert );
3440 mjd_offset= Node_Offset((nop));
3441 mjd_nodelen= Node_Length((nop));
3445 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
3447 (UV)mjd_offset, (UV)mjd_nodelen)
3450 /* But first we check to see if there is a common prefix we can
3451 split out as an EXACT and put in front of the TRIE node. */
3452 trie->startstate= 1;
3453 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3454 /* we want to find the first state that has more than
3455 * one transition, if that state is not the first state
3456 * then we have a common prefix which we can remove.
3459 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3461 I32 first_ofs = -1; /* keeps track of the ofs of the first
3462 transition, -1 means none */
3464 const U32 base = trie->states[ state ].trans.base;
3466 /* does this state terminate an alternation? */
3467 if ( trie->states[state].wordnum )
3470 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3471 if ( ( base + ofs >= trie->uniquecharcount ) &&
3472 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3473 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3475 if ( ++count > 1 ) {
3476 /* we have more than one transition */
3479 /* if this is the first state there is no common prefix
3480 * to extract, so we can exit */
3481 if ( state == 1 ) break;
3482 tmp = av_fetch( revcharmap, ofs, 0);
3483 ch = (U8*)SvPV_nolen_const( *tmp );
3485 /* if we are on count 2 then we need to initialize the
3486 * bitmap, and store the previous char if there was one
3489 /* clear the bitmap */
3490 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3492 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
3495 if (first_ofs >= 0) {
3496 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
3497 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3499 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3501 Perl_re_printf( aTHX_ "%s", (char*)ch)
3505 /* store the current firstchar in the bitmap */
3506 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3507 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
3513 /* This state has only one transition, its transition is part
3514 * of a common prefix - we need to concatenate the char it
3515 * represents to what we have so far. */
3516 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
3518 char *ch = SvPV( *tmp, len );
3520 SV *sv=sv_newmortal();
3521 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
3523 (UV)state, (UV)first_ofs,
3524 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3525 PL_colors[0], PL_colors[1],
3526 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3527 PERL_PV_ESCAPE_FIRSTCHAR
3532 OP( convert ) = nodetype;
3533 str=STRING(convert);
3536 STR_LEN(convert) += len;
3542 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
3547 trie->prefixlen = (state-1);
3549 regnode *n = convert+NODE_SZ_STR(convert);
3550 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3551 trie->startstate = state;
3552 trie->minlen -= (state - 1);
3553 trie->maxlen -= (state - 1);
3555 /* At least the UNICOS C compiler choked on this
3556 * being argument to DEBUG_r(), so let's just have
3559 #ifdef PERL_EXT_RE_BUILD
3565 regnode *fix = convert;
3566 U32 word = trie->wordcount;
3567 #ifdef RE_TRACK_PATTERN_OFFSETS
3570 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3571 while( ++fix < n ) {
3572 Set_Node_Offset_Length(fix, 0, 0);
3575 SV ** const tmp = av_fetch( trie_words, word, 0 );
3577 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3578 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3580 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3588 NEXT_OFF(convert) = (U16)(tail - convert);
3589 DEBUG_r(optimize= n);
3595 if ( trie->maxlen ) {
3596 NEXT_OFF( convert ) = (U16)(tail - convert);
3597 ARG_SET( convert, data_slot );
3598 /* Store the offset to the first unabsorbed branch in
3599 jump[0], which is otherwise unused by the jump logic.
3600 We use this when dumping a trie and during optimisation. */
3602 trie->jump[0] = (U16)(nextbranch - convert);
3604 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3605 * and there is a bitmap
3606 * and the first "jump target" node we found leaves enough room
3607 * then convert the TRIE node into a TRIEC node, with the bitmap
3608 * embedded inline in the opcode - this is hypothetically faster.
3610 if ( !trie->states[trie->startstate].wordnum
3612 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3614 OP( convert ) = TRIEC;
3615 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3616 PerlMemShared_free(trie->bitmap);
3619 OP( convert ) = TRIE;
3621 /* store the type in the flags */
3622 convert->flags = nodetype;
3626 + regarglen[ OP( convert ) ];
3628 /* XXX We really should free up the resource in trie now,
3629 as we won't use them - (which resources?) dmq */
3631 /* needed for dumping*/
3632 DEBUG_r(if (optimize) {
3633 regnode *opt = convert;
3635 while ( ++opt < optimize) {
3636 Set_Node_Offset_Length(opt, 0, 0);
3639 Try to clean up some of the debris left after the
3642 while( optimize < jumper ) {
3643 Track_Code( mjd_nodelen += Node_Length((optimize)); );
3644 OP( optimize ) = OPTIMIZED;
3645 Set_Node_Offset_Length(optimize, 0, 0);
3648 Set_Node_Offset_Length(convert, mjd_offset, mjd_nodelen);
3650 } /* end node insert */
3652 /* Finish populating the prev field of the wordinfo array. Walk back
3653 * from each accept state until we find another accept state, and if
3654 * so, point the first word's .prev field at the second word. If the
3655 * second already has a .prev field set, stop now. This will be the
3656 * case either if we've already processed that word's accept state,
3657 * or that state had multiple words, and the overspill words were
3658 * already linked up earlier.
3665 for (word=1; word <= trie->wordcount; word++) {
3667 if (trie->wordinfo[word].prev)
3669 state = trie->wordinfo[word].accept;
3671 state = prev_states[state];
3674 prev = trie->states[state].wordnum;
3678 trie->wordinfo[word].prev = prev;
3680 Safefree(prev_states);
3684 /* and now dump out the compressed format */
3685 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3687 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3689 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3690 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3692 SvREFCNT_dec_NN(revcharmap);
3696 : trie->startstate>1
3702 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3704 /* The Trie is constructed and compressed now so we can build a fail array if
3707 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3709 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3713 We find the fail state for each state in the trie, this state is the longest
3714 proper suffix of the current state's 'word' that is also a proper prefix of
3715 another word in our trie. State 1 represents the word '' and is thus the
3716 default fail state. This allows the DFA not to have to restart after its
3717 tried and failed a word at a given point, it simply continues as though it
3718 had been matching the other word in the first place.
3720 'abcdgu'=~/abcdefg|cdgu/
3721 When we get to 'd' we are still matching the first word, we would encounter
3722 'g' which would fail, which would bring us to the state representing 'd' in
3723 the second word where we would try 'g' and succeed, proceeding to match
3726 /* add a fail transition */
3727 const U32 trie_offset = ARG(source);
3728 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3730 const U32 ucharcount = trie->uniquecharcount;
3731 const U32 numstates = trie->statecount;
3732 const U32 ubound = trie->lasttrans + ucharcount;
3736 U32 base = trie->states[ 1 ].trans.base;
3739 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3741 GET_RE_DEBUG_FLAGS_DECL;
3743 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3744 PERL_UNUSED_CONTEXT;
3746 PERL_UNUSED_ARG(depth);
3749 if ( OP(source) == TRIE ) {
3750 struct regnode_1 *op = (struct regnode_1 *)
3751 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3752 StructCopy(source, op, struct regnode_1);
3753 stclass = (regnode *)op;
3755 struct regnode_charclass *op = (struct regnode_charclass *)
3756 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3757 StructCopy(source, op, struct regnode_charclass);
3758 stclass = (regnode *)op;
3760 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3762 ARG_SET( stclass, data_slot );
3763 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3764 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3765 aho->trie=trie_offset;
3766 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3767 Copy( trie->states, aho->states, numstates, reg_trie_state );
3768 Newx( q, numstates, U32);
3769 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3772 /* initialize fail[0..1] to be 1 so that we always have
3773 a valid final fail state */
3774 fail[ 0 ] = fail[ 1 ] = 1;
3776 for ( charid = 0; charid < ucharcount ; charid++ ) {
3777 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3779 q[ q_write ] = newstate;
3780 /* set to point at the root */
3781 fail[ q[ q_write++ ] ]=1;
3784 while ( q_read < q_write) {
3785 const U32 cur = q[ q_read++ % numstates ];
3786 base = trie->states[ cur ].trans.base;
3788 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3789 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3791 U32 fail_state = cur;
3794 fail_state = fail[ fail_state ];
3795 fail_base = aho->states[ fail_state ].trans.base;
3796 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3798 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3799 fail[ ch_state ] = fail_state;
3800 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3802 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3804 q[ q_write++ % numstates] = ch_state;
3808 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3809 when we fail in state 1, this allows us to use the
3810 charclass scan to find a valid start char. This is based on the principle
3811 that theres a good chance the string being searched contains lots of stuff
3812 that cant be a start char.
3814 fail[ 0 ] = fail[ 1 ] = 0;
3815 DEBUG_TRIE_COMPILE_r({
3816 Perl_re_indentf( aTHX_ "Stclass Failtable (%" UVuf " states): 0",
3817 depth, (UV)numstates
3819 for( q_read=1; q_read<numstates; q_read++ ) {
3820 Perl_re_printf( aTHX_ ", %" UVuf, (UV)fail[q_read]);
3822 Perl_re_printf( aTHX_ "\n");
3825 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3830 /* The below joins as many adjacent EXACTish nodes as possible into a single
3831 * one. The regop may be changed if the node(s) contain certain sequences that
3832 * require special handling. The joining is only done if:
3833 * 1) there is room in the current conglomerated node to entirely contain the
3835 * 2) they are compatible node types
3837 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3838 * these get optimized out
3840 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3841 * as possible, even if that means splitting an existing node so that its first
3842 * part is moved to the preceeding node. This would maximise the efficiency of
3843 * memEQ during matching.
3845 * If a node is to match under /i (folded), the number of characters it matches
3846 * can be different than its character length if it contains a multi-character
3847 * fold. *min_subtract is set to the total delta number of characters of the
3850 * And *unfolded_multi_char is set to indicate whether or not the node contains
3851 * an unfolded multi-char fold. This happens when it won't be known until
3852 * runtime whether the fold is valid or not; namely
3853 * 1) for EXACTF nodes that contain LATIN SMALL LETTER SHARP S, as only if the
3854 * target string being matched against turns out to be UTF-8 is that fold
3856 * 2) for EXACTFL nodes whose folding rules depend on the locale in force at
3858 * (Multi-char folds whose components are all above the Latin1 range are not
3859 * run-time locale dependent, and have already been folded by the time this
3860 * function is called.)
3862 * This is as good a place as any to discuss the design of handling these
3863 * multi-character fold sequences. It's been wrong in Perl for a very long
3864 * time. There are three code points in Unicode whose multi-character folds
3865 * were long ago discovered to mess things up. The previous designs for
3866 * dealing with these involved assigning a special node for them. This
3867 * approach doesn't always work, as evidenced by this example:
3868 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3869 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3870 * would match just the \xDF, it won't be able to handle the case where a
3871 * successful match would have to cross the node's boundary. The new approach
3872 * that hopefully generally solves the problem generates an EXACTFUP node
3873 * that is "sss" in this case.
3875 * It turns out that there are problems with all multi-character folds, and not
3876 * just these three. Now the code is general, for all such cases. The
3877 * approach taken is:
3878 * 1) This routine examines each EXACTFish node that could contain multi-
3879 * character folded sequences. Since a single character can fold into
3880 * such a sequence, the minimum match length for this node is less than
3881 * the number of characters in the node. This routine returns in
3882 * *min_subtract how many characters to subtract from the the actual
3883 * length of the string to get a real minimum match length; it is 0 if
3884 * there are no multi-char foldeds. This delta is used by the caller to
3885 * adjust the min length of the match, and the delta between min and max,
3886 * so that the optimizer doesn't reject these possibilities based on size
3889 * 2) For the sequence involving the LATIN SMALL LETTER SHARP S (U+00DF)
3890 * under /u, we fold it to 'ss' in regatom(), and in this routine, after
3891 * joining, we scan for occurrences of the sequence 'ss' in non-UTF-8
3892 * EXACTFU nodes. The node type of such nodes is then changed to
3893 * EXACTFUP, indicating it is problematic, and needs careful handling.
3894 * (The procedures in step 1) above are sufficient to handle this case in
3895 * UTF-8 encoded nodes.) The reason this is problematic is that this is
3896 * the only case where there is a possible fold length change in non-UTF-8
3897 * patterns. By reserving a special node type for problematic cases, the
3898 * far more common regular EXACTFU nodes can be processed faster.
3899 * regexec.c takes advantage of this.
3901 * EXACTFUP has been created as a grab-bag for (hopefully uncommon)
3902 * problematic cases. These all only occur when the pattern is not
3903 * UTF-8. In addition to the 'ss' sequence where there is a possible fold
3904 * length change, it handles the situation where the string cannot be
3905 * entirely folded. The strings in an EXACTFish node are folded as much
3906 * as possible during compilation in regcomp.c. This saves effort in
3907 * regex matching. By using an EXACTFUP node when it is not possible to
3908 * fully fold at compile time, regexec.c can know that everything in an
3909 * EXACTFU node is folded, so folding can be skipped at runtime. The only
3910 * case where folding in EXACTFU nodes can't be done at compile time is
3911 * the presumably uncommon MICRO SIGN, when the pattern isn't UTF-8. This
3912 * is because its fold requires UTF-8 to represent. Thus EXACTFUP nodes
3913 * handle two very different cases. Alternatively, there could have been
3914 * a node type where there are length changes, one for unfolded, and one
3915 * for both. If yet another special case needed to be created, the number
3916 * of required node types would have to go to 7. khw figures that even
3917 * though there are plenty of node types to spare, that the maintenance
3918 * cost wasn't worth the small speedup of doing it that way, especially
3919 * since he thinks the MICRO SIGN is rarely encountered in practice.
3921 * There are other cases where folding isn't done at compile time, but
3922 * none of them are under /u, and hence not for EXACTFU nodes. The folds
3923 * in EXACTFL nodes aren't known until runtime, and vary as the locale
3924 * changes. Some folds in EXACTF depend on if the runtime target string
3925 * is UTF-8 or not. (regatom() will create an EXACTFU node even under /di
3926 * when no fold in it depends on the UTF-8ness of the target string.)
3928 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3929 * validity of the fold won't be known until runtime, and so must remain
3930 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFAA
3931 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3932 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3933 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3934 * The reason this is a problem is that the optimizer part of regexec.c
3935 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3936 * that a character in the pattern corresponds to at most a single
3937 * character in the target string. (And I do mean character, and not byte
3938 * here, unlike other parts of the documentation that have never been
3939 * updated to account for multibyte Unicode.) Sharp s in EXACTF and
3940 * EXACTFL nodes can match the two character string 'ss'; in EXACTFAA
3941 * nodes it can match "\x{17F}\x{17F}". These, along with other ones in
3942 * EXACTFL nodes, violate the assumption, and they are the only instances
3943 * where it is violated. I'm reluctant to try to change the assumption,
3944 * as the code involved is impenetrable to me (khw), so instead the code
3945 * here punts. This routine examines EXACTFL nodes, and (when the pattern
3946 * isn't UTF-8) EXACTF and EXACTFAA for such unfolded folds, and returns a
3947 * boolean indicating whether or not the node contains such a fold. When
3948 * it is true, the caller sets a flag that later causes the optimizer in
3949 * this file to not set values for the floating and fixed string lengths,
3950 * and thus avoids the optimizer code in regexec.c that makes the invalid
3951 * assumption. Thus, there is no optimization based on string lengths for
3952 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3953 * EXACTF and EXACTFAA nodes that contain the sharp s. (The reason the
3954 * assumption is wrong only in these cases is that all other non-UTF-8
3955 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3956 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3957 * EXACTF nodes because we don't know at compile time if it actually
3958 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3959 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3960 * always matches; and EXACTFAA where it never does. In an EXACTFAA node
3961 * in a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3962 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3963 * string would require the pattern to be forced into UTF-8, the overhead
3964 * of which we want to avoid. Similarly the unfolded multi-char folds in
3965 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3968 * Similarly, the code that generates tries doesn't currently handle
3969 * not-already-folded multi-char folds, and it looks like a pain to change
3970 * that. Therefore, trie generation of EXACTFAA nodes with the sharp s
3971 * doesn't work. Instead, such an EXACTFAA is turned into a new regnode,
3972 * EXACTFAA_NO_TRIE, which the trie code knows not to handle. Most people
3973 * using /iaa matching will be doing so almost entirely with ASCII
3974 * strings, so this should rarely be encountered in practice */
3976 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3977 if (PL_regkind[OP(scan)] == EXACT) \
3978 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags), NULL, depth+1)
3981 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3982 UV *min_subtract, bool *unfolded_multi_char,
3983 U32 flags, regnode *val, U32 depth)
3985 /* Merge several consecutive EXACTish nodes into one. */
3987 regnode *n = regnext(scan);
3989 regnode *next = scan + NODE_SZ_STR(scan);
https://perl5.git.perl.org / perl5.git / blob