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;
1578 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1580 /* Look at the data structure created by S_set_ANYOF_arg() */
1581 if (n != ANYOF_ONLY_HAS_BITMAP) {
1582 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1583 AV * const av = MUTABLE_AV(SvRV(rv));
1584 SV **const ary = AvARRAY(av);
1585 assert(RExC_rxi->data->what[n] == 's');
1587 if (av_tindex_skip_len_mg(av) >= DEFERRED_USER_DEFINED_INDEX) {
1589 /* Here there are things that won't be known until runtime -- we
1590 * have to assume it could be anything */
1591 invlist = sv_2mortal(_new_invlist(1));
1592 return _add_range_to_invlist(invlist, 0, UV_MAX);
1594 else if (ary[INVLIST_INDEX]) {
1596 /* Use the node's inversion list */
1597 invlist = sv_2mortal(invlist_clone(ary[INVLIST_INDEX], NULL));
1600 /* Get the code points valid only under UTF-8 locales */
1601 if ( (ANYOF_FLAGS(node) & ANYOFL_FOLD)
1602 && av_tindex_skip_len_mg(av) >= ONLY_LOCALE_MATCHES_INDEX)
1604 only_utf8_locale_invlist = ary[ONLY_LOCALE_MATCHES_INDEX];
1609 invlist = sv_2mortal(_new_invlist(0));
1612 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1613 * code points, and an inversion list for the others, but if there are code
1614 * points that should match only conditionally on the target string being
1615 * UTF-8, those are placed in the inversion list, and not the bitmap.
1616 * Since there are circumstances under which they could match, they are
1617 * included in the SSC. But if the ANYOF node is to be inverted, we have
1618 * to exclude them here, so that when we invert below, the end result
1619 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1620 * have to do this here before we add the unconditionally matched code
1622 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1623 _invlist_intersection_complement_2nd(invlist,
1628 /* Add in the points from the bit map */
1629 if (OP(node) != ANYOFH) {
1630 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1631 if (ANYOF_BITMAP_TEST(node, i)) {
1632 unsigned int start = i++;
1634 for (; i < NUM_ANYOF_CODE_POINTS
1635 && ANYOF_BITMAP_TEST(node, i); ++i)
1639 invlist = _add_range_to_invlist(invlist, start, i-1);
1640 new_node_has_latin1 = TRUE;
1645 /* If this can match all upper Latin1 code points, have to add them
1646 * as well. But don't add them if inverting, as when that gets done below,
1647 * it would exclude all these characters, including the ones it shouldn't
1648 * that were added just above */
1649 if (! (ANYOF_FLAGS(node) & ANYOF_INVERT) && OP(node) == ANYOFD
1650 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1652 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1655 /* Similarly for these */
1656 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1657 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1660 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1661 _invlist_invert(invlist);
1663 else if (ANYOF_FLAGS(node) & ANYOFL_FOLD) {
1664 if (new_node_has_latin1) {
1666 /* Under /li, any 0-255 could fold to any other 0-255, depending on
1667 * the locale. We can skip this if there are no 0-255 at all. */
1668 _invlist_union(invlist, PL_Latin1, &invlist);
1670 invlist = add_cp_to_invlist(invlist, LATIN_SMALL_LETTER_DOTLESS_I);
1671 invlist = add_cp_to_invlist(invlist, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
1674 if (_invlist_contains_cp(invlist, LATIN_SMALL_LETTER_DOTLESS_I)) {
1675 invlist = add_cp_to_invlist(invlist, 'I');
1677 if (_invlist_contains_cp(invlist,
1678 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE))
1680 invlist = add_cp_to_invlist(invlist, 'i');
1685 /* Similarly add the UTF-8 locale possible matches. These have to be
1686 * deferred until after the non-UTF-8 locale ones are taken care of just
1687 * above, or it leads to wrong results under ANYOF_INVERT */
1688 if (only_utf8_locale_invlist) {
1689 _invlist_union_maybe_complement_2nd(invlist,
1690 only_utf8_locale_invlist,
1691 ANYOF_FLAGS(node) & ANYOF_INVERT,
1698 /* These two functions currently do the exact same thing */
1699 #define ssc_init_zero ssc_init
1701 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1702 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1704 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1705 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1706 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1709 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1710 const regnode_charclass *and_with)
1712 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1713 * another SSC or a regular ANYOF class. Can create false positives. */
1716 U8 and_with_flags = (OP(and_with) == ANYOFH) ? 0 : ANYOF_FLAGS(and_with);
1719 PERL_ARGS_ASSERT_SSC_AND;
1721 assert(is_ANYOF_SYNTHETIC(ssc));
1723 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1724 * the code point inversion list and just the relevant flags */
1725 if (is_ANYOF_SYNTHETIC(and_with)) {
1726 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1727 anded_flags = and_with_flags;
1729 /* XXX This is a kludge around what appears to be deficiencies in the
1730 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1731 * there are paths through the optimizer where it doesn't get weeded
1732 * out when it should. And if we don't make some extra provision for
1733 * it like the code just below, it doesn't get added when it should.
1734 * This solution is to add it only when AND'ing, which is here, and
1735 * only when what is being AND'ed is the pristine, original node
1736 * matching anything. Thus it is like adding it to ssc_anything() but
1737 * only when the result is to be AND'ed. Probably the same solution
1738 * could be adopted for the same problem we have with /l matching,
1739 * which is solved differently in S_ssc_init(), and that would lead to
1740 * fewer false positives than that solution has. But if this solution
1741 * creates bugs, the consequences are only that a warning isn't raised
1742 * that should be; while the consequences for having /l bugs is
1743 * incorrect matches */
1744 if (ssc_is_anything((regnode_ssc *)and_with)) {
1745 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1749 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1750 if (OP(and_with) == ANYOFD) {
1751 anded_flags = and_with_flags & ANYOF_COMMON_FLAGS;
1754 anded_flags = and_with_flags
1755 &( ANYOF_COMMON_FLAGS
1756 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1757 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1758 if (ANYOFL_UTF8_LOCALE_REQD(and_with_flags)) {
1760 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1765 ANYOF_FLAGS(ssc) &= anded_flags;
1767 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1768 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1769 * 'and_with' may be inverted. When not inverted, we have the situation of
1771 * (C1 | P1) & (C2 | P2)
1772 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1773 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1774 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1775 * <= ((C1 & C2) | P1 | P2)
1776 * Alternatively, the last few steps could be:
1777 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1778 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1779 * <= (C1 | C2 | (P1 & P2))
1780 * We favor the second approach if either P1 or P2 is non-empty. This is
1781 * because these components are a barrier to doing optimizations, as what
1782 * they match cannot be known until the moment of matching as they are
1783 * dependent on the current locale, 'AND"ing them likely will reduce or
1785 * But we can do better if we know that C1,P1 are in their initial state (a
1786 * frequent occurrence), each matching everything:
1787 * (<everything>) & (C2 | P2) = C2 | P2
1788 * Similarly, if C2,P2 are in their initial state (again a frequent
1789 * occurrence), the result is a no-op
1790 * (C1 | P1) & (<everything>) = C1 | P1
1793 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1794 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1795 * <= (C1 & ~C2) | (P1 & ~P2)
1798 if ((and_with_flags & ANYOF_INVERT)
1799 && ! is_ANYOF_SYNTHETIC(and_with))
1803 ssc_intersection(ssc,
1805 FALSE /* Has already been inverted */
1808 /* If either P1 or P2 is empty, the intersection will be also; can skip
1810 if (! (and_with_flags & ANYOF_MATCHES_POSIXL)) {
1811 ANYOF_POSIXL_ZERO(ssc);
1813 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1815 /* Note that the Posix class component P from 'and_with' actually
1817 * P = Pa | Pb | ... | Pn
1818 * where each component is one posix class, such as in [\w\s].
1820 * ~P = ~(Pa | Pb | ... | Pn)
1821 * = ~Pa & ~Pb & ... & ~Pn
1822 * <= ~Pa | ~Pb | ... | ~Pn
1823 * The last is something we can easily calculate, but unfortunately
1824 * is likely to have many false positives. We could do better
1825 * in some (but certainly not all) instances if two classes in
1826 * P have known relationships. For example
1827 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1829 * :lower: & :print: = :lower:
1830 * And similarly for classes that must be disjoint. For example,
1831 * since \s and \w can have no elements in common based on rules in
1832 * the POSIX standard,
1833 * \w & ^\S = nothing
1834 * Unfortunately, some vendor locales do not meet the Posix
1835 * standard, in particular almost everything by Microsoft.
1836 * The loop below just changes e.g., \w into \W and vice versa */
1838 regnode_charclass_posixl temp;
1839 int add = 1; /* To calculate the index of the complement */
1841 Zero(&temp, 1, regnode_charclass_posixl);
1842 ANYOF_POSIXL_ZERO(&temp);
1843 for (i = 0; i < ANYOF_MAX; i++) {
1845 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1846 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1848 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1849 ANYOF_POSIXL_SET(&temp, i + add);
1851 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1853 ANYOF_POSIXL_AND(&temp, ssc);
1855 } /* else ssc already has no posixes */
1856 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1857 in its initial state */
1858 else if (! is_ANYOF_SYNTHETIC(and_with)
1859 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1861 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1862 * copy it over 'ssc' */
1863 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1864 if (is_ANYOF_SYNTHETIC(and_with)) {
1865 StructCopy(and_with, ssc, regnode_ssc);
1868 ssc->invlist = anded_cp_list;
1869 ANYOF_POSIXL_ZERO(ssc);
1870 if (and_with_flags & ANYOF_MATCHES_POSIXL) {
1871 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1875 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1876 || (and_with_flags & ANYOF_MATCHES_POSIXL))
1878 /* One or the other of P1, P2 is non-empty. */
1879 if (and_with_flags & ANYOF_MATCHES_POSIXL) {
1880 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1882 ssc_union(ssc, anded_cp_list, FALSE);
1884 else { /* P1 = P2 = empty */
1885 ssc_intersection(ssc, anded_cp_list, FALSE);
1891 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1892 const regnode_charclass *or_with)
1894 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1895 * another SSC or a regular ANYOF class. Can create false positives if
1896 * 'or_with' is to be inverted. */
1900 U8 or_with_flags = (OP(or_with) == ANYOFH) ? 0 : ANYOF_FLAGS(or_with);
1902 PERL_ARGS_ASSERT_SSC_OR;
1904 assert(is_ANYOF_SYNTHETIC(ssc));
1906 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1907 * the code point inversion list and just the relevant flags */
1908 if (is_ANYOF_SYNTHETIC(or_with)) {
1909 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1910 ored_flags = or_with_flags;
1913 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1914 ored_flags = or_with_flags & ANYOF_COMMON_FLAGS;
1915 if (OP(or_with) != ANYOFD) {
1918 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1919 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1920 if (ANYOFL_UTF8_LOCALE_REQD(or_with_flags)) {
1922 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1927 ANYOF_FLAGS(ssc) |= ored_flags;
1929 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1930 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1931 * 'or_with' may be inverted. When not inverted, we have the simple
1932 * situation of computing:
1933 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1934 * If P1|P2 yields a situation with both a class and its complement are
1935 * set, like having both \w and \W, this matches all code points, and we
1936 * can delete these from the P component of the ssc going forward. XXX We
1937 * might be able to delete all the P components, but I (khw) am not certain
1938 * about this, and it is better to be safe.
1941 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1942 * <= (C1 | P1) | ~C2
1943 * <= (C1 | ~C2) | P1
1944 * (which results in actually simpler code than the non-inverted case)
1947 if ((or_with_flags & ANYOF_INVERT)
1948 && ! is_ANYOF_SYNTHETIC(or_with))
1950 /* We ignore P2, leaving P1 going forward */
1951 } /* else Not inverted */
1952 else if (or_with_flags & ANYOF_MATCHES_POSIXL) {
1953 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1954 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1956 for (i = 0; i < ANYOF_MAX; i += 2) {
1957 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1959 ssc_match_all_cp(ssc);
1960 ANYOF_POSIXL_CLEAR(ssc, i);
1961 ANYOF_POSIXL_CLEAR(ssc, i+1);
1969 FALSE /* Already has been inverted */
1973 PERL_STATIC_INLINE void
1974 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1976 PERL_ARGS_ASSERT_SSC_UNION;
1978 assert(is_ANYOF_SYNTHETIC(ssc));
1980 _invlist_union_maybe_complement_2nd(ssc->invlist,
1986 PERL_STATIC_INLINE void
1987 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1989 const bool invert2nd)
1991 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1993 assert(is_ANYOF_SYNTHETIC(ssc));
1995 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
2001 PERL_STATIC_INLINE void
2002 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
2004 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
2006 assert(is_ANYOF_SYNTHETIC(ssc));
2008 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
2011 PERL_STATIC_INLINE void
2012 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
2014 /* AND just the single code point 'cp' into the SSC 'ssc' */
2016 SV* cp_list = _new_invlist(2);
2018 PERL_ARGS_ASSERT_SSC_CP_AND;
2020 assert(is_ANYOF_SYNTHETIC(ssc));
2022 cp_list = add_cp_to_invlist(cp_list, cp);
2023 ssc_intersection(ssc, cp_list,
2024 FALSE /* Not inverted */
2026 SvREFCNT_dec_NN(cp_list);
2029 PERL_STATIC_INLINE void
2030 S_ssc_clear_locale(regnode_ssc *ssc)
2032 /* Set the SSC 'ssc' to not match any locale things */
2033 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
2035 assert(is_ANYOF_SYNTHETIC(ssc));
2037 ANYOF_POSIXL_ZERO(ssc);
2038 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
2041 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
2044 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
2046 /* The synthetic start class is used to hopefully quickly winnow down
2047 * places where a pattern could start a match in the target string. If it
2048 * doesn't really narrow things down that much, there isn't much point to
2049 * having the overhead of using it. This function uses some very crude
2050 * heuristics to decide if to use the ssc or not.
2052 * It returns TRUE if 'ssc' rules out more than half what it considers to
2053 * be the "likely" possible matches, but of course it doesn't know what the
2054 * actual things being matched are going to be; these are only guesses
2056 * For /l matches, it assumes that the only likely matches are going to be
2057 * in the 0-255 range, uniformly distributed, so half of that is 127
2058 * For /a and /d matches, it assumes that the likely matches will be just
2059 * the ASCII range, so half of that is 63
2060 * For /u and there isn't anything matching above the Latin1 range, it
2061 * assumes that that is the only range likely to be matched, and uses
2062 * half that as the cut-off: 127. If anything matches above Latin1,
2063 * it assumes that all of Unicode could match (uniformly), except for
2064 * non-Unicode code points and things in the General Category "Other"
2065 * (unassigned, private use, surrogates, controls and formats). This
2066 * is a much large number. */
2068 U32 count = 0; /* Running total of number of code points matched by
2070 UV start, end; /* Start and end points of current range in inversion
2071 XXX outdated. UTF-8 locales are common, what about invert? list */
2072 const U32 max_code_points = (LOC)
2074 : (( ! UNI_SEMANTICS
2075 || invlist_highest(ssc->invlist) < 256)
2078 const U32 max_match = max_code_points / 2;
2080 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
2082 invlist_iterinit(ssc->invlist);
2083 while (invlist_iternext(ssc->invlist, &start, &end)) {
2084 if (start >= max_code_points) {
2087 end = MIN(end, max_code_points - 1);
2088 count += end - start + 1;
2089 if (count >= max_match) {
2090 invlist_iterfinish(ssc->invlist);
2100 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
2102 /* The inversion list in the SSC is marked mortal; now we need a more
2103 * permanent copy, which is stored the same way that is done in a regular
2104 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
2107 SV* invlist = invlist_clone(ssc->invlist, NULL);
2109 PERL_ARGS_ASSERT_SSC_FINALIZE;
2111 assert(is_ANYOF_SYNTHETIC(ssc));
2113 /* The code in this file assumes that all but these flags aren't relevant
2114 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
2115 * by the time we reach here */
2116 assert(! (ANYOF_FLAGS(ssc)
2117 & ~( ANYOF_COMMON_FLAGS
2118 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
2119 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
2121 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
2123 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist, NULL, NULL);
2125 /* Make sure is clone-safe */
2126 ssc->invlist = NULL;
2128 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
2129 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
2130 OP(ssc) = ANYOFPOSIXL;
2132 else if (RExC_contains_locale) {
2136 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
2139 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
2140 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
2141 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
2142 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
2143 ? (TRIE_LIST_CUR( idx ) - 1) \
2149 dump_trie(trie,widecharmap,revcharmap)
2150 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
2151 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
2153 These routines dump out a trie in a somewhat readable format.
2154 The _interim_ variants are used for debugging the interim
2155 tables that are used to generate the final compressed
2156 representation which is what dump_trie expects.
2158 Part of the reason for their existence is to provide a form
2159 of documentation as to how the different representations function.
2164 Dumps the final compressed table form of the trie to Perl_debug_log.
2165 Used for debugging make_trie().
2169 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
2170 AV *revcharmap, U32 depth)
2173 SV *sv=sv_newmortal();
2174 int colwidth= widecharmap ? 6 : 4;
2176 GET_RE_DEBUG_FLAGS_DECL;
2178 PERL_ARGS_ASSERT_DUMP_TRIE;
2180 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
2181 depth+1, "Match","Base","Ofs" );
2183 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2184 SV ** const tmp = av_fetch( revcharmap, state, 0);
2186 Perl_re_printf( aTHX_ "%*s",
2188 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2189 PL_colors[0], PL_colors[1],
2190 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2191 PERL_PV_ESCAPE_FIRSTCHAR
2196 Perl_re_printf( aTHX_ "\n");
2197 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
2199 for( state = 0 ; state < trie->uniquecharcount ; state++ )
2200 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2201 Perl_re_printf( aTHX_ "\n");
2203 for( state = 1 ; state < trie->statecount ; state++ ) {
2204 const U32 base = trie->states[ state ].trans.base;
2206 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
2208 if ( trie->states[ state ].wordnum ) {
2209 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
2211 Perl_re_printf( aTHX_ "%6s", "" );
2214 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
2219 while( ( base + ofs < trie->uniquecharcount ) ||
2220 ( base + ofs - trie->uniquecharcount < trie->lasttrans
2221 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2225 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
2227 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2228 if ( ( base + ofs >= trie->uniquecharcount )
2229 && ( base + ofs - trie->uniquecharcount
2231 && trie->trans[ base + ofs
2232 - trie->uniquecharcount ].check == state )
2234 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
2235 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2238 Perl_re_printf( aTHX_ "%*s", colwidth," ." );
2242 Perl_re_printf( aTHX_ "]");
2245 Perl_re_printf( aTHX_ "\n" );
2247 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
2249 for (word=1; word <= trie->wordcount; word++) {
2250 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2251 (int)word, (int)(trie->wordinfo[word].prev),
2252 (int)(trie->wordinfo[word].len));
2254 Perl_re_printf( aTHX_ "\n" );
2257 Dumps a fully constructed but uncompressed trie in list form.
2258 List tries normally only are used for construction when the number of
2259 possible chars (trie->uniquecharcount) is very high.
2260 Used for debugging make_trie().
2263 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2264 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2268 SV *sv=sv_newmortal();
2269 int colwidth= widecharmap ? 6 : 4;
2270 GET_RE_DEBUG_FLAGS_DECL;
2272 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2274 /* print out the table precompression. */
2275 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
2277 Perl_re_indentf( aTHX_ "%s",
2278 depth+1, "------:-----+-----------------\n" );
2280 for( state=1 ; state < next_alloc ; state ++ ) {
2283 Perl_re_indentf( aTHX_ " %4" UVXf " :",
2284 depth+1, (UV)state );
2285 if ( ! trie->states[ state ].wordnum ) {
2286 Perl_re_printf( aTHX_ "%5s| ","");
2288 Perl_re_printf( aTHX_ "W%4x| ",
2289 trie->states[ state ].wordnum
2292 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2293 SV ** const tmp = av_fetch( revcharmap,
2294 TRIE_LIST_ITEM(state, charid).forid, 0);
2296 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
2298 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2300 PL_colors[0], PL_colors[1],
2301 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2302 | PERL_PV_ESCAPE_FIRSTCHAR
2304 TRIE_LIST_ITEM(state, charid).forid,
2305 (UV)TRIE_LIST_ITEM(state, charid).newstate
2308 Perl_re_printf( aTHX_ "\n%*s| ",
2309 (int)((depth * 2) + 14), "");
2312 Perl_re_printf( aTHX_ "\n");
2317 Dumps a fully constructed but uncompressed trie in table form.
2318 This is the normal DFA style state transition table, with a few
2319 twists to facilitate compression later.
2320 Used for debugging make_trie().
2323 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2324 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2329 SV *sv=sv_newmortal();
2330 int colwidth= widecharmap ? 6 : 4;
2331 GET_RE_DEBUG_FLAGS_DECL;
2333 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2336 print out the table precompression so that we can do a visual check
2337 that they are identical.
2340 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
2342 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2343 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2345 Perl_re_printf( aTHX_ "%*s",
2347 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2348 PL_colors[0], PL_colors[1],
2349 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2350 PERL_PV_ESCAPE_FIRSTCHAR
2356 Perl_re_printf( aTHX_ "\n");
2357 Perl_re_indentf( aTHX_ "State+-", depth+1 );
2359 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2360 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
2363 Perl_re_printf( aTHX_ "\n" );
2365 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2367 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
2369 (UV)TRIE_NODENUM( state ) );
2371 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2372 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2374 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
2376 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
2378 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2379 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
2380 (UV)trie->trans[ state ].check );
2382 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
2383 (UV)trie->trans[ state ].check,
2384 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2392 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2393 startbranch: the first branch in the whole branch sequence
2394 first : start branch of sequence of branch-exact nodes.
2395 May be the same as startbranch
2396 last : Thing following the last branch.
2397 May be the same as tail.
2398 tail : item following the branch sequence
2399 count : words in the sequence
2400 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2401 depth : indent depth
2403 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2405 A trie is an N'ary tree where the branches are determined by digital
2406 decomposition of the key. IE, at the root node you look up the 1st character and
2407 follow that branch repeat until you find the end of the branches. Nodes can be
2408 marked as "accepting" meaning they represent a complete word. Eg:
2412 would convert into the following structure. Numbers represent states, letters
2413 following numbers represent valid transitions on the letter from that state, if
2414 the number is in square brackets it represents an accepting state, otherwise it
2415 will be in parenthesis.
2417 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2421 (1) +-i->(6)-+-s->[7]
2423 +-s->(3)-+-h->(4)-+-e->[5]
2425 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2427 This shows that when matching against the string 'hers' we will begin at state 1
2428 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2429 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2430 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2431 single traverse. We store a mapping from accepting to state to which word was
2432 matched, and then when we have multiple possibilities we try to complete the
2433 rest of the regex in the order in which they occurred in the alternation.
2435 The only prior NFA like behaviour that would be changed by the TRIE support is
2436 the silent ignoring of duplicate alternations which are of the form:
2438 / (DUPE|DUPE) X? (?{ ... }) Y /x
2440 Thus EVAL blocks following a trie may be called a different number of times with
2441 and without the optimisation. With the optimisations dupes will be silently
2442 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2443 the following demonstrates:
2445 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2447 which prints out 'word' three times, but
2449 'words'=~/(word|word|word)(?{ print $1 })S/
2451 which doesnt print it out at all. This is due to other optimisations kicking in.
2453 Example of what happens on a structural level:
2455 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2457 1: CURLYM[1] {1,32767}(18)
2468 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2469 and should turn into:
2471 1: CURLYM[1] {1,32767}(18)
2473 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2481 Cases where tail != last would be like /(?foo|bar)baz/:
2491 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2492 and would end up looking like:
2495 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2502 d = uvchr_to_utf8_flags(d, uv, 0);
2504 is the recommended Unicode-aware way of saying
2509 #define TRIE_STORE_REVCHAR(val) \
2512 SV *zlopp = newSV(UTF8_MAXBYTES); \
2513 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2514 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2515 SvCUR_set(zlopp, kapow - flrbbbbb); \
2518 av_push(revcharmap, zlopp); \
2520 char ooooff = (char)val; \
2521 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2525 /* This gets the next character from the input, folding it if not already
2527 #define TRIE_READ_CHAR STMT_START { \
2530 /* if it is UTF then it is either already folded, or does not need \
2532 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2534 else if (folder == PL_fold_latin1) { \
2535 /* This folder implies Unicode rules, which in the range expressible \
2536 * by not UTF is the lower case, with the two exceptions, one of \
2537 * which should have been taken care of before calling this */ \
2538 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2539 uvc = toLOWER_L1(*uc); \
2540 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2543 /* raw data, will be folded later if needed */ \
2551 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2552 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2553 U32 ging = TRIE_LIST_LEN( state ) * 2; \
2554 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2555 TRIE_LIST_LEN( state ) = ging; \
2557 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2558 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2559 TRIE_LIST_CUR( state )++; \
2562 #define TRIE_LIST_NEW(state) STMT_START { \
2563 Newx( trie->states[ state ].trans.list, \
2564 4, reg_trie_trans_le ); \
2565 TRIE_LIST_CUR( state ) = 1; \
2566 TRIE_LIST_LEN( state ) = 4; \
2569 #define TRIE_HANDLE_WORD(state) STMT_START { \
2570 U16 dupe= trie->states[ state ].wordnum; \
2571 regnode * const noper_next = regnext( noper ); \
2574 /* store the word for dumping */ \
2576 if (OP(noper) != NOTHING) \
2577 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2579 tmp = newSVpvn_utf8( "", 0, UTF ); \
2580 av_push( trie_words, tmp ); \
2584 trie->wordinfo[curword].prev = 0; \
2585 trie->wordinfo[curword].len = wordlen; \
2586 trie->wordinfo[curword].accept = state; \
2588 if ( noper_next < tail ) { \
2590 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2592 trie->jump[curword] = (U16)(noper_next - convert); \
2594 jumper = noper_next; \
2596 nextbranch= regnext(cur); \
2600 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2601 /* chain, so that when the bits of chain are later */\
2602 /* linked together, the dups appear in the chain */\
2603 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2604 trie->wordinfo[dupe].prev = curword; \
2606 /* we haven't inserted this word yet. */ \
2607 trie->states[ state ].wordnum = curword; \
2612 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2613 ( ( base + charid >= ucharcount \
2614 && base + charid < ubound \
2615 && state == trie->trans[ base - ucharcount + charid ].check \
2616 && trie->trans[ base - ucharcount + charid ].next ) \
2617 ? trie->trans[ base - ucharcount + charid ].next \
2618 : ( state==1 ? special : 0 ) \
2621 #define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2623 TRIE_BITMAP_SET(trie, uvc); \
2624 /* store the folded codepoint */ \
2626 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2629 /* store first byte of utf8 representation of */ \
2630 /* variant codepoints */ \
2631 if (! UVCHR_IS_INVARIANT(uvc)) { \
2632 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2637 #define MADE_JUMP_TRIE 2
2638 #define MADE_EXACT_TRIE 4
2641 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2642 regnode *first, regnode *last, regnode *tail,
2643 U32 word_count, U32 flags, U32 depth)
2645 /* first pass, loop through and scan words */
2646 reg_trie_data *trie;
2647 HV *widecharmap = NULL;
2648 AV *revcharmap = newAV();
2654 regnode *jumper = NULL;
2655 regnode *nextbranch = NULL;
2656 regnode *convert = NULL;
2657 U32 *prev_states; /* temp array mapping each state to previous one */
2658 /* we just use folder as a flag in utf8 */
2659 const U8 * folder = NULL;
2661 /* in the below add_data call we are storing either 'tu' or 'tuaa'
2662 * which stands for one trie structure, one hash, optionally followed
2665 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuaa"));
2666 AV *trie_words = NULL;
2667 /* along with revcharmap, this only used during construction but both are
2668 * useful during debugging so we store them in the struct when debugging.
2671 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2672 STRLEN trie_charcount=0;
2674 SV *re_trie_maxbuff;
2675 GET_RE_DEBUG_FLAGS_DECL;
2677 PERL_ARGS_ASSERT_MAKE_TRIE;
2679 PERL_UNUSED_ARG(depth);
2683 case EXACT: case EXACT_ONLY8: case EXACTL: break;
2687 case EXACTFLU8: folder = PL_fold_latin1; break;
2688 case EXACTF: folder = PL_fold; break;
2689 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2692 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2694 trie->startstate = 1;
2695 trie->wordcount = word_count;
2696 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2697 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2698 if (flags == EXACT || flags == EXACT_ONLY8 || flags == EXACTL)
2699 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2700 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2701 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2704 trie_words = newAV();
2707 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, GV_ADD);
2708 assert(re_trie_maxbuff);
2709 if (!SvIOK(re_trie_maxbuff)) {
2710 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2712 DEBUG_TRIE_COMPILE_r({
2713 Perl_re_indentf( aTHX_
2714 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2716 REG_NODE_NUM(startbranch), REG_NODE_NUM(first),
2717 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2720 /* Find the node we are going to overwrite */
2721 if ( first == startbranch && OP( last ) != BRANCH ) {
2722 /* whole branch chain */
2725 /* branch sub-chain */
2726 convert = NEXTOPER( first );
2729 /* -- First loop and Setup --
2731 We first traverse the branches and scan each word to determine if it
2732 contains widechars, and how many unique chars there are, this is
2733 important as we have to build a table with at least as many columns as we
2736 We use an array of integers to represent the character codes 0..255
2737 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2738 the native representation of the character value as the key and IV's for
2741 *TODO* If we keep track of how many times each character is used we can
2742 remap the columns so that the table compression later on is more
2743 efficient in terms of memory by ensuring the most common value is in the
2744 middle and the least common are on the outside. IMO this would be better
2745 than a most to least common mapping as theres a decent chance the most
2746 common letter will share a node with the least common, meaning the node
2747 will not be compressible. With a middle is most common approach the worst
2748 case is when we have the least common nodes twice.
2752 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2753 regnode *noper = NEXTOPER( cur );
2757 U32 wordlen = 0; /* required init */
2758 STRLEN minchars = 0;
2759 STRLEN maxchars = 0;
2760 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2763 if (OP(noper) == NOTHING) {
2764 /* skip past a NOTHING at the start of an alternation
2765 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2767 regnode *noper_next= regnext(noper);
2768 if (noper_next < tail)
2773 && ( OP(noper) == flags
2774 || (flags == EXACT && OP(noper) == EXACT_ONLY8)
2775 || (flags == EXACTFU && ( OP(noper) == EXACTFU_ONLY8
2776 || OP(noper) == EXACTFUP))))
2778 uc= (U8*)STRING(noper);
2779 e= uc + STR_LEN(noper);
2786 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2787 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2788 regardless of encoding */
2789 if (OP( noper ) == EXACTFUP) {
2790 /* false positives are ok, so just set this */
2791 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2795 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2797 TRIE_CHARCOUNT(trie)++;
2800 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2801 * is in effect. Under /i, this character can match itself, or
2802 * anything that folds to it. If not under /i, it can match just
2803 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2804 * all fold to k, and all are single characters. But some folds
2805 * expand to more than one character, so for example LATIN SMALL
2806 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2807 * the string beginning at 'uc' is 'ffi', it could be matched by
2808 * three characters, or just by the one ligature character. (It
2809 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2810 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2811 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2812 * match.) The trie needs to know the minimum and maximum number
2813 * of characters that could match so that it can use size alone to
2814 * quickly reject many match attempts. The max is simple: it is
2815 * the number of folded characters in this branch (since a fold is
2816 * never shorter than what folds to it. */
2820 /* And the min is equal to the max if not under /i (indicated by
2821 * 'folder' being NULL), or there are no multi-character folds. If
2822 * there is a multi-character fold, the min is incremented just
2823 * once, for the character that folds to the sequence. Each
2824 * character in the sequence needs to be added to the list below of
2825 * characters in the trie, but we count only the first towards the
2826 * min number of characters needed. This is done through the
2827 * variable 'foldlen', which is returned by the macros that look
2828 * for these sequences as the number of bytes the sequence
2829 * occupies. Each time through the loop, we decrement 'foldlen' by
2830 * how many bytes the current char occupies. Only when it reaches
2831 * 0 do we increment 'minchars' or look for another multi-character
2833 if (folder == NULL) {
2836 else if (foldlen > 0) {
2837 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2842 /* See if *uc is the beginning of a multi-character fold. If
2843 * so, we decrement the length remaining to look at, to account
2844 * for the current character this iteration. (We can use 'uc'
2845 * instead of the fold returned by TRIE_READ_CHAR because for
2846 * non-UTF, the latin1_safe macro is smart enough to account
2847 * for all the unfolded characters, and because for UTF, the
2848 * string will already have been folded earlier in the
2849 * compilation process */
2851 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2852 foldlen -= UTF8SKIP(uc);
2855 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2860 /* The current character (and any potential folds) should be added
2861 * to the possible matching characters for this position in this
2865 U8 folded= folder[ (U8) uvc ];
2866 if ( !trie->charmap[ folded ] ) {
2867 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2868 TRIE_STORE_REVCHAR( folded );
2871 if ( !trie->charmap[ uvc ] ) {
2872 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2873 TRIE_STORE_REVCHAR( uvc );
2876 /* store the codepoint in the bitmap, and its folded
2878 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
2879 set_bit = 0; /* We've done our bit :-) */
2883 /* XXX We could come up with the list of code points that fold
2884 * to this using PL_utf8_foldclosures, except not for
2885 * multi-char folds, as there may be multiple combinations
2886 * there that could work, which needs to wait until runtime to
2887 * resolve (The comment about LIGATURE FFI above is such an
2892 widecharmap = newHV();
2894 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2897 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
2899 if ( !SvTRUE( *svpp ) ) {
2900 sv_setiv( *svpp, ++trie->uniquecharcount );
2901 TRIE_STORE_REVCHAR(uvc);
2904 } /* end loop through characters in this branch of the trie */
2906 /* We take the min and max for this branch and combine to find the min
2907 * and max for all branches processed so far */
2908 if( cur == first ) {
2909 trie->minlen = minchars;
2910 trie->maxlen = maxchars;
2911 } else if (minchars < trie->minlen) {
2912 trie->minlen = minchars;
2913 } else if (maxchars > trie->maxlen) {
2914 trie->maxlen = maxchars;
2916 } /* end first pass */
2917 DEBUG_TRIE_COMPILE_r(
2918 Perl_re_indentf( aTHX_
2919 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2921 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2922 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2923 (int)trie->minlen, (int)trie->maxlen )
2927 We now know what we are dealing with in terms of unique chars and
2928 string sizes so we can calculate how much memory a naive
2929 representation using a flat table will take. If it's over a reasonable
2930 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2931 conservative but potentially much slower representation using an array
2934 At the end we convert both representations into the same compressed
2935 form that will be used in regexec.c for matching with. The latter
2936 is a form that cannot be used to construct with but has memory
2937 properties similar to the list form and access properties similar
2938 to the table form making it both suitable for fast searches and
2939 small enough that its feasable to store for the duration of a program.
2941 See the comment in the code where the compressed table is produced
2942 inplace from the flat tabe representation for an explanation of how
2943 the compression works.
2948 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2951 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2952 > SvIV(re_trie_maxbuff) )
2955 Second Pass -- Array Of Lists Representation
2957 Each state will be represented by a list of charid:state records
2958 (reg_trie_trans_le) the first such element holds the CUR and LEN
2959 points of the allocated array. (See defines above).
2961 We build the initial structure using the lists, and then convert
2962 it into the compressed table form which allows faster lookups
2963 (but cant be modified once converted).
2966 STRLEN transcount = 1;
2968 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
2971 trie->states = (reg_trie_state *)
2972 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2973 sizeof(reg_trie_state) );
2977 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2979 regnode *noper = NEXTOPER( cur );
2980 U32 state = 1; /* required init */
2981 U16 charid = 0; /* sanity init */
2982 U32 wordlen = 0; /* required init */
2984 if (OP(noper) == NOTHING) {
2985 regnode *noper_next= regnext(noper);
2986 if (noper_next < tail)
2991 && ( OP(noper) == flags
2992 || (flags == EXACT && OP(noper) == EXACT_ONLY8)
2993 || (flags == EXACTFU && ( OP(noper) == EXACTFU_ONLY8
2994 || OP(noper) == EXACTFUP))))
2996 const U8 *uc= (U8*)STRING(noper);
2997 const U8 *e= uc + STR_LEN(noper);
2999 for ( ; uc < e ; uc += len ) {
3004 charid = trie->charmap[ uvc ];
3006 SV** const svpp = hv_fetch( widecharmap,
3013 charid=(U16)SvIV( *svpp );
3016 /* charid is now 0 if we dont know the char read, or
3017 * nonzero if we do */
3024 if ( !trie->states[ state ].trans.list ) {
3025 TRIE_LIST_NEW( state );
3028 check <= TRIE_LIST_USED( state );
3031 if ( TRIE_LIST_ITEM( state, check ).forid
3034 newstate = TRIE_LIST_ITEM( state, check ).newstate;
3039 newstate = next_alloc++;
3040 prev_states[newstate] = state;
3041 TRIE_LIST_PUSH( state, charid, newstate );
3046 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3050 TRIE_HANDLE_WORD(state);
3052 } /* end second pass */
3054 /* next alloc is the NEXT state to be allocated */
3055 trie->statecount = next_alloc;
3056 trie->states = (reg_trie_state *)
3057 PerlMemShared_realloc( trie->states,
3059 * sizeof(reg_trie_state) );
3061 /* and now dump it out before we compress it */
3062 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
3063 revcharmap, next_alloc,
3067 trie->trans = (reg_trie_trans *)
3068 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
3075 for( state=1 ; state < next_alloc ; state ++ ) {
3079 DEBUG_TRIE_COMPILE_MORE_r(
3080 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
3084 if (trie->states[state].trans.list) {
3085 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
3089 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3090 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
3091 if ( forid < minid ) {
3093 } else if ( forid > maxid ) {
3097 if ( transcount < tp + maxid - minid + 1) {
3099 trie->trans = (reg_trie_trans *)
3100 PerlMemShared_realloc( trie->trans,
3102 * sizeof(reg_trie_trans) );
3103 Zero( trie->trans + (transcount / 2),
3107 base = trie->uniquecharcount + tp - minid;
3108 if ( maxid == minid ) {
3110 for ( ; zp < tp ; zp++ ) {
3111 if ( ! trie->trans[ zp ].next ) {
3112 base = trie->uniquecharcount + zp - minid;
3113 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
3115 trie->trans[ zp ].check = state;
3121 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
3123 trie->trans[ tp ].check = state;
3128 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3129 const U32 tid = base
3130 - trie->uniquecharcount
3131 + TRIE_LIST_ITEM( state, idx ).forid;
3132 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
3134 trie->trans[ tid ].check = state;
3136 tp += ( maxid - minid + 1 );
3138 Safefree(trie->states[ state ].trans.list);
3141 DEBUG_TRIE_COMPILE_MORE_r(
3142 Perl_re_printf( aTHX_ " base: %d\n",base);
3145 trie->states[ state ].trans.base=base;
3147 trie->lasttrans = tp + 1;
3151 Second Pass -- Flat Table Representation.
3153 we dont use the 0 slot of either trans[] or states[] so we add 1 to
3154 each. We know that we will need Charcount+1 trans at most to store
3155 the data (one row per char at worst case) So we preallocate both
3156 structures assuming worst case.
3158 We then construct the trie using only the .next slots of the entry
3161 We use the .check field of the first entry of the node temporarily
3162 to make compression both faster and easier by keeping track of how
3163 many non zero fields are in the node.
3165 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
3168 There are two terms at use here: state as a TRIE_NODEIDX() which is
3169 a number representing the first entry of the node, and state as a
3170 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
3171 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
3172 if there are 2 entrys per node. eg:
3180 The table is internally in the right hand, idx form. However as we
3181 also have to deal with the states array which is indexed by nodenum
3182 we have to use TRIE_NODENUM() to convert.
3185 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
3188 trie->trans = (reg_trie_trans *)
3189 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
3190 * trie->uniquecharcount + 1,
3191 sizeof(reg_trie_trans) );
3192 trie->states = (reg_trie_state *)
3193 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3194 sizeof(reg_trie_state) );
3195 next_alloc = trie->uniquecharcount + 1;
3198 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3200 regnode *noper = NEXTOPER( cur );
3202 U32 state = 1; /* required init */
3204 U16 charid = 0; /* sanity init */
3205 U32 accept_state = 0; /* sanity init */
3207 U32 wordlen = 0; /* required init */
3209 if (OP(noper) == NOTHING) {
3210 regnode *noper_next= regnext(noper);
3211 if (noper_next < tail)
3216 && ( OP(noper) == flags
3217 || (flags == EXACT && OP(noper) == EXACT_ONLY8)
3218 || (flags == EXACTFU && ( OP(noper) == EXACTFU_ONLY8
3219 || OP(noper) == EXACTFUP))))
3221 const U8 *uc= (U8*)STRING(noper);
3222 const U8 *e= uc + STR_LEN(noper);
3224 for ( ; uc < e ; uc += len ) {
3229 charid = trie->charmap[ uvc ];
3231 SV* const * const svpp = hv_fetch( widecharmap,
3235 charid = svpp ? (U16)SvIV(*svpp) : 0;
3239 if ( !trie->trans[ state + charid ].next ) {
3240 trie->trans[ state + charid ].next = next_alloc;
3241 trie->trans[ state ].check++;
3242 prev_states[TRIE_NODENUM(next_alloc)]
3243 = TRIE_NODENUM(state);
3244 next_alloc += trie->uniquecharcount;
3246 state = trie->trans[ state + charid ].next;
3248 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3250 /* charid is now 0 if we dont know the char read, or
3251 * nonzero if we do */
3254 accept_state = TRIE_NODENUM( state );
3255 TRIE_HANDLE_WORD(accept_state);
3257 } /* end second pass */
3259 /* and now dump it out before we compress it */
3260 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3262 next_alloc, depth+1));
3266 * Inplace compress the table.*
3268 For sparse data sets the table constructed by the trie algorithm will
3269 be mostly 0/FAIL transitions or to put it another way mostly empty.
3270 (Note that leaf nodes will not contain any transitions.)
3272 This algorithm compresses the tables by eliminating most such
3273 transitions, at the cost of a modest bit of extra work during lookup:
3275 - Each states[] entry contains a .base field which indicates the
3276 index in the state[] array wheres its transition data is stored.
3278 - If .base is 0 there are no valid transitions from that node.
3280 - If .base is nonzero then charid is added to it to find an entry in
3283 -If trans[states[state].base+charid].check!=state then the
3284 transition is taken to be a 0/Fail transition. Thus if there are fail
3285 transitions at the front of the node then the .base offset will point
3286 somewhere inside the previous nodes data (or maybe even into a node
3287 even earlier), but the .check field determines if the transition is
3291 The following process inplace converts the table to the compressed
3292 table: We first do not compress the root node 1,and mark all its
3293 .check pointers as 1 and set its .base pointer as 1 as well. This
3294 allows us to do a DFA construction from the compressed table later,
3295 and ensures that any .base pointers we calculate later are greater
3298 - We set 'pos' to indicate the first entry of the second node.
3300 - We then iterate over the columns of the node, finding the first and
3301 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3302 and set the .check pointers accordingly, and advance pos
3303 appropriately and repreat for the next node. Note that when we copy
3304 the next pointers we have to convert them from the original
3305 NODEIDX form to NODENUM form as the former is not valid post
3308 - If a node has no transitions used we mark its base as 0 and do not
3309 advance the pos pointer.
3311 - If a node only has one transition we use a second pointer into the
3312 structure to fill in allocated fail transitions from other states.
3313 This pointer is independent of the main pointer and scans forward
3314 looking for null transitions that are allocated to a state. When it
3315 finds one it writes the single transition into the "hole". If the
3316 pointer doesnt find one the single transition is appended as normal.
3318 - Once compressed we can Renew/realloc the structures to release the
3321 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3322 specifically Fig 3.47 and the associated pseudocode.
3326 const U32 laststate = TRIE_NODENUM( next_alloc );
3329 trie->statecount = laststate;
3331 for ( state = 1 ; state < laststate ; state++ ) {
3333 const U32 stateidx = TRIE_NODEIDX( state );
3334 const U32 o_used = trie->trans[ stateidx ].check;
3335 U32 used = trie->trans[ stateidx ].check;
3336 trie->trans[ stateidx ].check = 0;
3339 used && charid < trie->uniquecharcount;
3342 if ( flag || trie->trans[ stateidx + charid ].next ) {
3343 if ( trie->trans[ stateidx + charid ].next ) {
3345 for ( ; zp < pos ; zp++ ) {
3346 if ( ! trie->trans[ zp ].next ) {
3350 trie->states[ state ].trans.base
3352 + trie->uniquecharcount
3354 trie->trans[ zp ].next
3355 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3357 trie->trans[ zp ].check = state;
3358 if ( ++zp > pos ) pos = zp;
3365 trie->states[ state ].trans.base
3366 = pos + trie->uniquecharcount - charid ;
3368 trie->trans[ pos ].next
3369 = SAFE_TRIE_NODENUM(
3370 trie->trans[ stateidx + charid ].next );
3371 trie->trans[ pos ].check = state;
3376 trie->lasttrans = pos + 1;
3377 trie->states = (reg_trie_state *)
3378 PerlMemShared_realloc( trie->states, laststate
3379 * sizeof(reg_trie_state) );
3380 DEBUG_TRIE_COMPILE_MORE_r(
3381 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
3383 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3387 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3390 } /* end table compress */
3392 DEBUG_TRIE_COMPILE_MORE_r(
3393 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
3395 (UV)trie->statecount,
3396 (UV)trie->lasttrans)
3398 /* resize the trans array to remove unused space */
3399 trie->trans = (reg_trie_trans *)
3400 PerlMemShared_realloc( trie->trans, trie->lasttrans
3401 * sizeof(reg_trie_trans) );
3403 { /* Modify the program and insert the new TRIE node */
3404 U8 nodetype =(U8)(flags & 0xFF);
3408 regnode *optimize = NULL;
3409 #ifdef RE_TRACK_PATTERN_OFFSETS
3412 U32 mjd_nodelen = 0;
3413 #endif /* RE_TRACK_PATTERN_OFFSETS */
3414 #endif /* DEBUGGING */
3416 This means we convert either the first branch or the first Exact,
3417 depending on whether the thing following (in 'last') is a branch
3418 or not and whther first is the startbranch (ie is it a sub part of
3419 the alternation or is it the whole thing.)
3420 Assuming its a sub part we convert the EXACT otherwise we convert
3421 the whole branch sequence, including the first.
3423 /* Find the node we are going to overwrite */
3424 if ( first != startbranch || OP( last ) == BRANCH ) {
3425 /* branch sub-chain */
3426 NEXT_OFF( first ) = (U16)(last - first);
3427 #ifdef RE_TRACK_PATTERN_OFFSETS
3429 mjd_offset= Node_Offset((convert));
3430 mjd_nodelen= Node_Length((convert));
3433 /* whole branch chain */
3435 #ifdef RE_TRACK_PATTERN_OFFSETS
3438 const regnode *nop = NEXTOPER( convert );
3439 mjd_offset= Node_Offset((nop));
3440 mjd_nodelen= Node_Length((nop));
3444 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
3446 (UV)mjd_offset, (UV)mjd_nodelen)
3449 /* But first we check to see if there is a common prefix we can
3450 split out as an EXACT and put in front of the TRIE node. */
3451 trie->startstate= 1;
3452 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3453 /* we want to find the first state that has more than
3454 * one transition, if that state is not the first state
3455 * then we have a common prefix which we can remove.
3458 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3460 I32 first_ofs = -1; /* keeps track of the ofs of the first
3461 transition, -1 means none */
3463 const U32 base = trie->states[ state ].trans.base;
3465 /* does this state terminate an alternation? */
3466 if ( trie->states[state].wordnum )
3469 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3470 if ( ( base + ofs >= trie->uniquecharcount ) &&
3471 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3472 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3474 if ( ++count > 1 ) {
3475 /* we have more than one transition */
3478 /* if this is the first state there is no common prefix
3479 * to extract, so we can exit */
3480 if ( state == 1 ) break;
3481 tmp = av_fetch( revcharmap, ofs, 0);
3482 ch = (U8*)SvPV_nolen_const( *tmp );
3484 /* if we are on count 2 then we need to initialize the
3485 * bitmap, and store the previous char if there was one
3488 /* clear the bitmap */
3489 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3491 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
3494 if (first_ofs >= 0) {
3495 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
3496 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3498 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3500 Perl_re_printf( aTHX_ "%s", (char*)ch)
3504 /* store the current firstchar in the bitmap */
3505 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3506 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
3512 /* This state has only one transition, its transition is part
3513 * of a common prefix - we need to concatenate the char it
3514 * represents to what we have so far. */
3515 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
3517 char *ch = SvPV( *tmp, len );
3519 SV *sv=sv_newmortal();
3520 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
3522 (UV)state, (UV)first_ofs,
3523 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3524 PL_colors[0], PL_colors[1],
3525 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3526 PERL_PV_ESCAPE_FIRSTCHAR
3531 OP( convert ) = nodetype;
3532 str=STRING(convert);
3535 STR_LEN(convert) += len;
3541 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
3546 trie->prefixlen = (state-1);
3548 regnode *n = convert+NODE_SZ_STR(convert);
3549 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3550 trie->startstate = state;
3551 trie->minlen -= (state - 1);
3552 trie->maxlen -= (state - 1);
3554 /* At least the UNICOS C compiler choked on this
3555 * being argument to DEBUG_r(), so let's just have
3558 #ifdef PERL_EXT_RE_BUILD
3564 regnode *fix = convert;
3565 U32 word = trie->wordcount;
3566 #ifdef RE_TRACK_PATTERN_OFFSETS
3569 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3570 while( ++fix < n ) {
3571 Set_Node_Offset_Length(fix, 0, 0);
3574 SV ** const tmp = av_fetch( trie_words, word, 0 );
3576 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3577 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3579 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3587 NEXT_OFF(convert) = (U16)(tail - convert);
3588 DEBUG_r(optimize= n);
3594 if ( trie->maxlen ) {
3595 NEXT_OFF( convert ) = (U16)(tail - convert);
3596 ARG_SET( convert, data_slot );
3597 /* Store the offset to the first unabsorbed branch in
3598 jump[0], which is otherwise unused by the jump logic.
3599 We use this when dumping a trie and during optimisation. */
3601 trie->jump[0] = (U16)(nextbranch - convert);
3603 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3604 * and there is a bitmap
3605 * and the first "jump target" node we found leaves enough room
3606 * then convert the TRIE node into a TRIEC node, with the bitmap
3607 * embedded inline in the opcode - this is hypothetically faster.
3609 if ( !trie->states[trie->startstate].wordnum
3611 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3613 OP( convert ) = TRIEC;
3614 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3615 PerlMemShared_free(trie->bitmap);
3618 OP( convert ) = TRIE;
3620 /* store the type in the flags */
3621 convert->flags = nodetype;
3625 + regarglen[ OP( convert ) ];
3627 /* XXX We really should free up the resource in trie now,
3628 as we won't use them - (which resources?) dmq */
3630 /* needed for dumping*/
3631 DEBUG_r(if (optimize) {
3632 regnode *opt = convert;
3634 while ( ++opt < optimize) {
3635 Set_Node_Offset_Length(opt, 0, 0);
3638 Try to clean up some of the debris left after the
3641 while( optimize < jumper ) {
3642 Track_Code( mjd_nodelen += Node_Length((optimize)); );
3643 OP( optimize ) = OPTIMIZED;
3644 Set_Node_Offset_Length(optimize, 0, 0);
3647 Set_Node_Offset_Length(convert, mjd_offset, mjd_nodelen);
3649 } /* end node insert */
3651 /* Finish populating the prev field of the wordinfo array. Walk back
3652 * from each accept state until we find another accept state, and if
3653 * so, point the first word's .prev field at the second word. If the
3654 * second already has a .prev field set, stop now. This will be the
3655 * case either if we've already processed that word's accept state,
3656 * or that state had multiple words, and the overspill words were
3657 * already linked up earlier.
3664 for (word=1; word <= trie->wordcount; word++) {
3666 if (trie->wordinfo[word].prev)
3668 state = trie->wordinfo[word].accept;
3670 state = prev_states[state];
3673 prev = trie->states[state].wordnum;
3677 trie->wordinfo[word].prev = prev;
3679 Safefree(prev_states);
3683 /* and now dump out the compressed format */
3684 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3686 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3688 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3689 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3691 SvREFCNT_dec_NN(revcharmap);
3695 : trie->startstate>1
3701 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3703 /* The Trie is constructed and compressed now so we can build a fail array if
3706 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3708 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3712 We find the fail state for each state in the trie, this state is the longest
3713 proper suffix of the current state's 'word' that is also a proper prefix of
3714 another word in our trie. State 1 represents the word '' and is thus the
3715 default fail state. This allows the DFA not to have to restart after its
3716 tried and failed a word at a given point, it simply continues as though it
3717 had been matching the other word in the first place.
3719 'abcdgu'=~/abcdefg|cdgu/
3720 When we get to 'd' we are still matching the first word, we would encounter
3721 'g' which would fail, which would bring us to the state representing 'd' in
3722 the second word where we would try 'g' and succeed, proceeding to match
3725 /* add a fail transition */
3726 const U32 trie_offset = ARG(source);
3727 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3729 const U32 ucharcount = trie->uniquecharcount;
3730 const U32 numstates = trie->statecount;
3731 const U32 ubound = trie->lasttrans + ucharcount;
3735 U32 base = trie->states[ 1 ].trans.base;
3738 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3740 GET_RE_DEBUG_FLAGS_DECL;
3742 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3743 PERL_UNUSED_CONTEXT;
3745 PERL_UNUSED_ARG(depth);
3748 if ( OP(source) == TRIE ) {
3749 struct regnode_1 *op = (struct regnode_1 *)
3750 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3751 StructCopy(source, op, struct regnode_1);
3752 stclass = (regnode *)op;
3754 struct regnode_charclass *op = (struct regnode_charclass *)
3755 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3756 StructCopy(source, op, struct regnode_charclass);
3757 stclass = (regnode *)op;
3759 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3761 ARG_SET( stclass, data_slot );
3762 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3763 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3764 aho->trie=trie_offset;
3765 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3766 Copy( trie->states, aho->states, numstates, reg_trie_state );
3767 Newx( q, numstates, U32);
3768 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3771 /* initialize fail[0..1] to be 1 so that we always have
3772 a valid final fail state */
3773 fail[ 0 ] = fail[ 1 ] = 1;
3775 for ( charid = 0; charid < ucharcount ; charid++ ) {
3776 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3778 q[ q_write ] = newstate;
3779 /* set to point at the root */
3780 fail[ q[ q_write++ ] ]=1;
3783 while ( q_read < q_write) {
3784 const U32 cur = q[ q_read++ % numstates ];
3785 base = trie->states[ cur ].trans.base;
3787 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3788 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3790 U32 fail_state = cur;
3793 fail_state = fail[ fail_state ];
3794 fail_base = aho->states[ fail_state ].trans.base;
3795 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3797 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3798 fail[ ch_state ] = fail_state;
3799 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3801 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3803 q[ q_write++ % numstates] = ch_state;
3807 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3808 when we fail in state 1, this allows us to use the
3809 charclass scan to find a valid start char. This is based on the principle
3810 that theres a good chance the string being searched contains lots of stuff
3811 that cant be a start char.
3813 fail[ 0 ] = fail[ 1 ] = 0;
3814 DEBUG_TRIE_COMPILE_r({
3815 Perl_re_indentf( aTHX_ "Stclass Failtable (%" UVuf " states): 0",
3816 depth, (UV)numstates
3818 for( q_read=1; q_read<numstates; q_read++ ) {
3819 Perl_re_printf( aTHX_ ", %" UVuf, (UV)fail[q_read]);
3821 Perl_re_printf( aTHX_ "\n");
3824 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3829 /* The below joins as many adjacent EXACTish nodes as possible into a single
3830 * one. The regop may be changed if the node(s) contain certain sequences that
3831 * require special handling. The joining is only done if:
3832 * 1) there is room in the current conglomerated node to entirely contain the
3834 * 2) they are compatible node types
3836 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3837 * these get optimized out
3839 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3840 * as possible, even if that means splitting an existing node so that its first
3841 * part is moved to the preceeding node. This would maximise the efficiency of
3842 * memEQ during matching.
3844 * If a node is to match under /i (folded), the number of characters it matches
3845 * can be different than its character length if it contains a multi-character
3846 * fold. *min_subtract is set to the total delta number of characters of the
3849 * And *unfolded_multi_char is set to indicate whether or not the node contains
3850 * an unfolded multi-char fold. This happens when it won't be known until
3851 * runtime whether the fold is valid or not; namely
3852 * 1) for EXACTF nodes that contain LATIN SMALL LETTER SHARP S, as only if the
3853 * target string being matched against turns out to be UTF-8 is that fold
3855 * 2) for EXACTFL nodes whose folding rules depend on the locale in force at
3857 * (Multi-char folds whose components are all above the Latin1 range are not
3858 * run-time locale dependent, and have already been folded by the time this
3859 * function is called.)
3861 * This is as good a place as any to discuss the design of handling these
3862 * multi-character fold sequences. It's been wrong in Perl for a very long
3863 * time. There are three code points in Unicode whose multi-character folds
3864 * were long ago discovered to mess things up. The previous designs for
3865 * dealing with these involved assigning a special node for them. This
3866 * approach doesn't always work, as evidenced by this example:
3867 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3868 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3869 * would match just the \xDF, it won't be able to handle the case where a
3870 * successful match would have to cross the node's boundary. The new approach
3871 * that hopefully generally solves the problem generates an EXACTFUP node
3872 * that is "sss" in this case.
3874 * It turns out that there are problems with all multi-character folds, and not
3875 * just these three. Now the code is general, for all such cases. The
3876 * approach taken is:
3877 * 1) This routine examines each EXACTFish node that could contain multi-
3878 * character folded sequences. Since a single character can fold into
3879 * such a sequence, the minimum match length for this node is less than
3880 * the number of characters in the node. This routine returns in
3881 * *min_subtract how many characters to subtract from the the actual
3882 * length of the string to get a real minimum match length; it is 0 if
3883 * there are no multi-char foldeds. This delta is used by the caller to
3884 * adjust the min length of the match, and the delta between min and max,
3885 * so that the optimizer doesn't reject these possibilities based on size
3888 * 2) For the sequence involving the LATIN SMALL LETTER SHARP S (U+00DF)
3889 * under /u, we fold it to 'ss' in regatom(), and in this routine, after
3890 * joining, we scan for occurrences of the sequence 'ss' in non-UTF-8
3891 * EXACTFU nodes. The node type of such nodes is then changed to
3892 * EXACTFUP, indicating it is problematic, and needs careful handling.
3893 * (The procedures in step 1) above are sufficient to handle this case in
3894 * UTF-8 encoded nodes.) The reason this is problematic is that this is
3895 * the only case where there is a possible fold length change in non-UTF-8
3896 * patterns. By reserving a special node type for problematic cases, the
3897 * far more common regular EXACTFU nodes can be processed faster.
3898 * regexec.c takes advantage of this.
3900 * EXACTFUP has been created as a grab-bag for (hopefully uncommon)
3901 * problematic cases. These all only occur when the pattern is not
3902 * UTF-8. In addition to the 'ss' sequence where there is a possible fold
3903 * length change, it handles the situation where the string cannot be
3904 * entirely folded. The strings in an EXACTFish node are folded as much
3905 * as possible during compilation in regcomp.c. This saves effort in
3906 * regex matching. By using an EXACTFUP node when it is not possible to
3907 * fully fold at compile time, regexec.c can know that everything in an
3908 * EXACTFU node is folded, so folding can be skipped at runtime. The only
3909 * case where folding in EXACTFU nodes can't be done at compile time is
3910 * the presumably uncommon MICRO SIGN, when the pattern isn't UTF-8. This
3911 * is because its fold requires UTF-8 to represent. Thus EXACTFUP nodes
3912 * handle two very different cases. Alternatively, there could have been
3913 * a node type where there are length changes, one for unfolded, and one
3914 * for both. If yet another special case needed to be created, the number
3915 * of required node types would have to go to 7. khw figures that even
3916 * though there are plenty of node types to spare, that the maintenance
3917 * cost wasn't worth the small speedup of doing it that way, especially
3918 * since he thinks the MICRO SIGN is rarely encountered in practice.
3920 * There are other cases where folding isn't done at compile time, but
3921 * none of them are under /u, and hence not for EXACTFU nodes. The folds
3922 * in EXACTFL nodes aren't known until runtime, and vary as the locale
3923 * changes. Some folds in EXACTF depend on if the runtime target string
3924 * is UTF-8 or not. (regatom() will create an EXACTFU node even under /di
3925 * when no fold in it depends on the UTF-8ness of the target string.)
3927 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3928 * validity of the fold won't be known until runtime, and so must remain
3929 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFAA
3930 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3931 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3932 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3933 * The reason this is a problem is that the optimizer part of regexec.c
3934 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3935 * that a character in the pattern corresponds to at most a single
3936 * character in the target string. (And I do mean character, and not byte
3937 * here, unlike other parts of the documentation that have never been
3938 * updated to account for multibyte Unicode.) Sharp s in EXACTF and
3939 * EXACTFL nodes can match the two character string 'ss'; in EXACTFAA
3940 * nodes it can match "\x{17F}\x{17F}". These, along with other ones in
3941 * EXACTFL nodes, violate the assumption, and they are the only instances
3942 * where it is violated. I'm reluctant to try to change the assumption,
3943 * as the code involved is impenetrable to me (khw), so instead the code
3944 * here punts. This routine examines EXACTFL nodes, and (when the pattern
3945 * isn't UTF-8) EXACTF and EXACTFAA for such unfolded folds, and returns a
3946 * boolean indicating whether or not the node contains such a fold. When
3947 * it is true, the caller sets a flag that later causes the optimizer in
3948 * this file to not set values for the floating and fixed string lengths,
3949 * and thus avoids the optimizer code in regexec.c that makes the invalid
3950 * assumption. Thus, there is no optimization based on string lengths for
3951 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3952 * EXACTF and EXACTFAA nodes that contain the sharp s. (The reason the
3953 * assumption is wrong only in these cases is that all other non-UTF-8
3954 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3955 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3956 * EXACTF nodes because we don't know at compile time if it actually
3957 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3958 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3959 * always matches; and EXACTFAA where it never does. In an EXACTFAA node
3960 * in a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3961 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3962 * string would require the pattern to be forced into UTF-8, the overhead
3963 * of which we want to avoid. Similarly the unfolded multi-char folds in
3964 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3967 * Similarly, the code that generates tries doesn't currently handle
3968 * not-already-folded multi-char folds, and it looks like a pain to change
3969 * that. Therefore, trie generation of EXACTFAA nodes with the sharp s
3970 * doesn't work. Instead, such an EXACTFAA is turned into a new regnode,
3971 * EXACTFAA_NO_TRIE, which the trie code knows not to handle. Most people
3972 * using /iaa matching will be doing so almost entirely with ASCII
3973 * strings, so this should rarely be encountered in practice */
3975 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3976 if (PL_regkind[OP(scan)] == EXACT) \
3977 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags), NULL, depth+1)
3980 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3981 UV *min_subtract, bool *unfolded_multi_char,
3982 U32 flags, regnode *val, U32 depth)
3984 /* Merge several consecutive EXACTish nodes into one. */
3986 regnode *n = regnext(scan);
3988 regnode *next = scan + NODE_SZ_STR(scan);
https://perl5.git.perl.org / perl5.git / blob