5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
10 /* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
12 * a regular expression.
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
20 /* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
24 /* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
29 /* The names of the functions have been changed from regcomp and
30 * regexec to pregcomp and pregexec in order to avoid conflicts
31 * with the POSIX routines of the same names.
34 #ifdef PERL_EXT_RE_BUILD
39 * pregcomp and pregexec -- regsub and regerror are not used in perl
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
59 **** Alterations to Henry's code are...
61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
74 #define PERL_IN_REGCOMP_C
77 #ifndef PERL_IN_XSUB_RE
82 #ifdef PERL_IN_XSUB_RE
84 EXTERN_C const struct regexp_engine my_reg_engine;
89 #include "dquote_inline.h"
90 #include "invlist_inline.h"
91 #include "unicode_constants.h"
93 #define HAS_NONLATIN1_FOLD_CLOSURE(i) \
94 _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
95 #define HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(i) \
96 _HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
97 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
98 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
101 #define STATIC static
104 /* this is a chain of data about sub patterns we are processing that
105 need to be handled separately/specially in study_chunk. Its so
106 we can simulate recursion without losing state. */
108 typedef struct scan_frame {
109 regnode *last_regnode; /* last node to process in this frame */
110 regnode *next_regnode; /* next node to process when last is reached */
111 U32 prev_recursed_depth;
112 I32 stopparen; /* what stopparen do we use */
114 struct scan_frame *this_prev_frame; /* this previous frame */
115 struct scan_frame *prev_frame; /* previous frame */
116 struct scan_frame *next_frame; /* next frame */
119 /* Certain characters are output as a sequence with the first being a
121 #define isBACKSLASHED_PUNCT(c) strchr("-[]\\^", c)
124 struct RExC_state_t {
125 U32 flags; /* RXf_* are we folding, multilining? */
126 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
127 char *precomp; /* uncompiled string. */
128 char *precomp_end; /* pointer to end of uncompiled string. */
129 REGEXP *rx_sv; /* The SV that is the regexp. */
130 regexp *rx; /* perl core regexp structure */
131 regexp_internal *rxi; /* internal data for regexp object
133 char *start; /* Start of input for compile */
134 char *end; /* End of input for compile */
135 char *parse; /* Input-scan pointer. */
136 char *adjusted_start; /* 'start', adjusted. See code use */
137 STRLEN precomp_adj; /* an offset beyond precomp. See code use */
138 SSize_t whilem_seen; /* number of WHILEM in this expr */
139 regnode *emit_start; /* Start of emitted-code area */
140 regnode *emit_bound; /* First regnode outside of the
142 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
143 implies compiling, so don't emit */
144 regnode_ssc emit_dummy; /* placeholder for emit to point to;
145 large enough for the largest
146 non-EXACTish node, so can use it as
148 I32 naughty; /* How bad is this pattern? */
149 I32 sawback; /* Did we see \1, ...? */
151 SSize_t size; /* Code size. */
152 I32 npar; /* Capture buffer count, (OPEN) plus
153 one. ("par" 0 is the whole
155 I32 nestroot; /* root parens we are in - used by
159 regnode **open_parens; /* pointers to open parens */
160 regnode **close_parens; /* pointers to close parens */
161 regnode *end_op; /* END node in program */
162 I32 utf8; /* whether the pattern is utf8 or not */
163 I32 orig_utf8; /* whether the pattern was originally in utf8 */
164 /* XXX use this for future optimisation of case
165 * where pattern must be upgraded to utf8. */
166 I32 uni_semantics; /* If a d charset modifier should use unicode
167 rules, even if the pattern is not in
169 HV *paren_names; /* Paren names */
171 regnode **recurse; /* Recurse regops */
172 I32 recurse_count; /* Number of recurse regops we have generated */
173 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
175 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
178 I32 override_recoding;
180 I32 recode_x_to_native;
182 I32 in_multi_char_class;
183 struct reg_code_blocks *code_blocks;/* positions of literal (?{})
185 int code_index; /* next code_blocks[] slot */
186 SSize_t maxlen; /* mininum possible number of chars in string to match */
187 scan_frame *frame_head;
188 scan_frame *frame_last;
191 #ifdef ADD_TO_REGEXEC
192 char *starttry; /* -Dr: where regtry was called. */
193 #define RExC_starttry (pRExC_state->starttry)
195 SV *runtime_code_qr; /* qr with the runtime code blocks */
197 const char *lastparse;
199 AV *paren_name_list; /* idx -> name */
200 U32 study_chunk_recursed_count;
203 #define RExC_lastparse (pRExC_state->lastparse)
204 #define RExC_lastnum (pRExC_state->lastnum)
205 #define RExC_paren_name_list (pRExC_state->paren_name_list)
206 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
207 #define RExC_mysv (pRExC_state->mysv1)
208 #define RExC_mysv1 (pRExC_state->mysv1)
209 #define RExC_mysv2 (pRExC_state->mysv2)
212 bool seen_unfolded_sharp_s;
218 #define RExC_flags (pRExC_state->flags)
219 #define RExC_pm_flags (pRExC_state->pm_flags)
220 #define RExC_precomp (pRExC_state->precomp)
221 #define RExC_precomp_adj (pRExC_state->precomp_adj)
222 #define RExC_adjusted_start (pRExC_state->adjusted_start)
223 #define RExC_precomp_end (pRExC_state->precomp_end)
224 #define RExC_rx_sv (pRExC_state->rx_sv)
225 #define RExC_rx (pRExC_state->rx)
226 #define RExC_rxi (pRExC_state->rxi)
227 #define RExC_start (pRExC_state->start)
228 #define RExC_end (pRExC_state->end)
229 #define RExC_parse (pRExC_state->parse)
230 #define RExC_whilem_seen (pRExC_state->whilem_seen)
232 /* Set during the sizing pass when there is a LATIN SMALL LETTER SHARP S in any
233 * EXACTF node, hence was parsed under /di rules. If later in the parse,
234 * something forces the pattern into using /ui rules, the sharp s should be
235 * folded into the sequence 'ss', which takes up more space than previously
236 * calculated. This means that the sizing pass needs to be restarted. (The
237 * node also becomes an EXACTFU_SS.) For all other characters, an EXACTF node
238 * that gets converted to /ui (and EXACTFU) occupies the same amount of space,
239 * so there is no need to resize [perl #125990]. */
240 #define RExC_seen_unfolded_sharp_s (pRExC_state->seen_unfolded_sharp_s)
242 #ifdef RE_TRACK_PATTERN_OFFSETS
243 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
246 #define RExC_emit (pRExC_state->emit)
247 #define RExC_emit_dummy (pRExC_state->emit_dummy)
248 #define RExC_emit_start (pRExC_state->emit_start)
249 #define RExC_emit_bound (pRExC_state->emit_bound)
250 #define RExC_sawback (pRExC_state->sawback)
251 #define RExC_seen (pRExC_state->seen)
252 #define RExC_size (pRExC_state->size)
253 #define RExC_maxlen (pRExC_state->maxlen)
254 #define RExC_npar (pRExC_state->npar)
255 #define RExC_nestroot (pRExC_state->nestroot)
256 #define RExC_extralen (pRExC_state->extralen)
257 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
258 #define RExC_utf8 (pRExC_state->utf8)
259 #define RExC_uni_semantics (pRExC_state->uni_semantics)
260 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
261 #define RExC_open_parens (pRExC_state->open_parens)
262 #define RExC_close_parens (pRExC_state->close_parens)
263 #define RExC_end_op (pRExC_state->end_op)
264 #define RExC_paren_names (pRExC_state->paren_names)
265 #define RExC_recurse (pRExC_state->recurse)
266 #define RExC_recurse_count (pRExC_state->recurse_count)
267 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
268 #define RExC_study_chunk_recursed_bytes \
269 (pRExC_state->study_chunk_recursed_bytes)
270 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
271 #define RExC_contains_locale (pRExC_state->contains_locale)
273 # define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
275 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
276 #define RExC_frame_head (pRExC_state->frame_head)
277 #define RExC_frame_last (pRExC_state->frame_last)
278 #define RExC_frame_count (pRExC_state->frame_count)
279 #define RExC_strict (pRExC_state->strict)
280 #define RExC_study_started (pRExC_state->study_started)
281 #define RExC_warn_text (pRExC_state->warn_text)
282 #define RExC_in_script_run (pRExC_state->in_script_run)
284 /* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
285 * a flag to disable back-off on the fixed/floating substrings - if it's
286 * a high complexity pattern we assume the benefit of avoiding a full match
287 * is worth the cost of checking for the substrings even if they rarely help.
289 #define RExC_naughty (pRExC_state->naughty)
290 #define TOO_NAUGHTY (10)
291 #define MARK_NAUGHTY(add) \
292 if (RExC_naughty < TOO_NAUGHTY) \
293 RExC_naughty += (add)
294 #define MARK_NAUGHTY_EXP(exp, add) \
295 if (RExC_naughty < TOO_NAUGHTY) \
296 RExC_naughty += RExC_naughty / (exp) + (add)
298 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
299 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
300 ((*s) == '{' && regcurly(s)))
303 * Flags to be passed up and down.
305 #define WORST 0 /* Worst case. */
306 #define HASWIDTH 0x01 /* Known to match non-null strings. */
308 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
309 * character. (There needs to be a case: in the switch statement in regexec.c
310 * for any node marked SIMPLE.) Note that this is not the same thing as
313 #define SPSTART 0x04 /* Starts with * or + */
314 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
315 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
316 #define RESTART_PASS1 0x20 /* Need to restart sizing pass */
317 #define NEED_UTF8 0x40 /* In conjunction with RESTART_PASS1, need to
318 calcuate sizes as UTF-8 */
320 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
322 /* whether trie related optimizations are enabled */
323 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
324 #define TRIE_STUDY_OPT
325 #define FULL_TRIE_STUDY
331 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
332 #define PBITVAL(paren) (1 << ((paren) & 7))
333 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
334 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
335 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
337 #define REQUIRE_UTF8(flagp) STMT_START { \
340 *flagp = RESTART_PASS1|NEED_UTF8; \
345 /* Change from /d into /u rules, and restart the parse if we've already seen
346 * something whose size would increase as a result, by setting *flagp and
347 * returning 'restart_retval'. RExC_uni_semantics is a flag that indicates
348 * we've changed to /u during the parse. */
349 #define REQUIRE_UNI_RULES(flagp, restart_retval) \
351 if (DEPENDS_SEMANTICS) { \
353 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
354 RExC_uni_semantics = 1; \
355 if (RExC_seen_unfolded_sharp_s) { \
356 *flagp |= RESTART_PASS1; \
357 return restart_retval; \
362 /* Executes a return statement with the value 'X', if 'flags' contains any of
363 * 'RESTART_PASS1', 'NEED_UTF8', or 'extra'. If so, *flagp is set to those
365 #define RETURN_X_ON_RESTART_OR_FLAGS(X, flags, flagp, extra) \
367 if ((flags) & (RESTART_PASS1|NEED_UTF8|(extra))) { \
368 *(flagp) = (flags) & (RESTART_PASS1|NEED_UTF8|(extra)); \
373 #define RETURN_NULL_ON_RESTART_OR_FLAGS(flags,flagp,extra) \
374 RETURN_X_ON_RESTART_OR_FLAGS(NULL,flags,flagp,extra)
376 #define RETURN_X_ON_RESTART(X, flags,flagp) \
377 RETURN_X_ON_RESTART_OR_FLAGS( X, flags, flagp, 0)
380 #define RETURN_NULL_ON_RESTART_FLAGP_OR_FLAGS(flagp,extra) \
381 if (*(flagp) & (RESTART_PASS1|(extra))) return NULL
383 #define MUST_RESTART(flags) ((flags) & (RESTART_PASS1))
385 #define RETURN_NULL_ON_RESTART(flags,flagp) \
386 RETURN_X_ON_RESTART(NULL, flags,flagp)
387 #define RETURN_NULL_ON_RESTART_FLAGP(flagp) \
388 RETURN_NULL_ON_RESTART_FLAGP_OR_FLAGS(flagp,0)
390 /* This converts the named class defined in regcomp.h to its equivalent class
391 * number defined in handy.h. */
392 #define namedclass_to_classnum(class) ((int) ((class) / 2))
393 #define classnum_to_namedclass(classnum) ((classnum) * 2)
395 #define _invlist_union_complement_2nd(a, b, output) \
396 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
397 #define _invlist_intersection_complement_2nd(a, b, output) \
398 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
400 /* About scan_data_t.
402 During optimisation we recurse through the regexp program performing
403 various inplace (keyhole style) optimisations. In addition study_chunk
404 and scan_commit populate this data structure with information about
405 what strings MUST appear in the pattern. We look for the longest
406 string that must appear at a fixed location, and we look for the
407 longest string that may appear at a floating location. So for instance
412 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
413 strings (because they follow a .* construct). study_chunk will identify
414 both FOO and BAR as being the longest fixed and floating strings respectively.
416 The strings can be composites, for instance
420 will result in a composite fixed substring 'foo'.
422 For each string some basic information is maintained:
425 This is the position the string must appear at, or not before.
426 It also implicitly (when combined with minlenp) tells us how many
427 characters must match before the string we are searching for.
428 Likewise when combined with minlenp and the length of the string it
429 tells us how many characters must appear after the string we have
433 Only used for floating strings. This is the rightmost point that
434 the string can appear at. If set to SSize_t_MAX it indicates that the
435 string can occur infinitely far to the right.
436 For fixed strings, it is equal to min_offset.
439 A pointer to the minimum number of characters of the pattern that the
440 string was found inside. This is important as in the case of positive
441 lookahead or positive lookbehind we can have multiple patterns
446 The minimum length of the pattern overall is 3, the minimum length
447 of the lookahead part is 3, but the minimum length of the part that
448 will actually match is 1. So 'FOO's minimum length is 3, but the
449 minimum length for the F is 1. This is important as the minimum length
450 is used to determine offsets in front of and behind the string being
451 looked for. Since strings can be composites this is the length of the
452 pattern at the time it was committed with a scan_commit. Note that
453 the length is calculated by study_chunk, so that the minimum lengths
454 are not known until the full pattern has been compiled, thus the
455 pointer to the value.
459 In the case of lookbehind the string being searched for can be
460 offset past the start point of the final matching string.
461 If this value was just blithely removed from the min_offset it would
462 invalidate some of the calculations for how many chars must match
463 before or after (as they are derived from min_offset and minlen and
464 the length of the string being searched for).
465 When the final pattern is compiled and the data is moved from the
466 scan_data_t structure into the regexp structure the information
467 about lookbehind is factored in, with the information that would
468 have been lost precalculated in the end_shift field for the
471 The fields pos_min and pos_delta are used to store the minimum offset
472 and the delta to the maximum offset at the current point in the pattern.
476 struct scan_data_substrs {
477 SV *str; /* longest substring found in pattern */
478 SSize_t min_offset; /* earliest point in string it can appear */
479 SSize_t max_offset; /* latest point in string it can appear */
480 SSize_t *minlenp; /* pointer to the minlen relevant to the string */
481 SSize_t lookbehind; /* is the pos of the string modified by LB */
482 I32 flags; /* per substring SF_* and SCF_* flags */
485 typedef struct scan_data_t {
486 /*I32 len_min; unused */
487 /*I32 len_delta; unused */
491 SSize_t last_end; /* min value, <0 unless valid. */
492 SSize_t last_start_min;
493 SSize_t last_start_max;
494 U8 cur_is_floating; /* whether the last_* values should be set as
495 * the next fixed (0) or floating (1)
498 /* [0] is longest fixed substring so far, [1] is longest float so far */
499 struct scan_data_substrs substrs[2];
501 I32 flags; /* common SF_* and SCF_* flags */
503 SSize_t *last_closep;
504 regnode_ssc *start_class;
508 * Forward declarations for pregcomp()'s friends.
511 static const scan_data_t zero_scan_data = {
512 0, 0, NULL, 0, 0, 0, 0,
514 { NULL, 0, 0, 0, 0, 0 },
515 { NULL, 0, 0, 0, 0, 0 },
522 #define SF_BEFORE_SEOL 0x0001
523 #define SF_BEFORE_MEOL 0x0002
524 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
526 #define SF_IS_INF 0x0040
527 #define SF_HAS_PAR 0x0080
528 #define SF_IN_PAR 0x0100
529 #define SF_HAS_EVAL 0x0200
532 /* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
533 * longest substring in the pattern. When it is not set the optimiser keeps
534 * track of position, but does not keep track of the actual strings seen,
536 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
539 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
540 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
541 * turned off because of the alternation (BRANCH). */
542 #define SCF_DO_SUBSTR 0x0400
544 #define SCF_DO_STCLASS_AND 0x0800
545 #define SCF_DO_STCLASS_OR 0x1000
546 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
547 #define SCF_WHILEM_VISITED_POS 0x2000
549 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
550 #define SCF_SEEN_ACCEPT 0x8000
551 #define SCF_TRIE_DOING_RESTUDY 0x10000
552 #define SCF_IN_DEFINE 0x20000
557 #define UTF cBOOL(RExC_utf8)
559 /* The enums for all these are ordered so things work out correctly */
560 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
561 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
562 == REGEX_DEPENDS_CHARSET)
563 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
564 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
565 >= REGEX_UNICODE_CHARSET)
566 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
567 == REGEX_ASCII_RESTRICTED_CHARSET)
568 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
569 >= REGEX_ASCII_RESTRICTED_CHARSET)
570 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
571 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
573 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
575 /* For programs that want to be strictly Unicode compatible by dying if any
576 * attempt is made to match a non-Unicode code point against a Unicode
578 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
580 #define OOB_NAMEDCLASS -1
582 /* There is no code point that is out-of-bounds, so this is problematic. But
583 * its only current use is to initialize a variable that is always set before
585 #define OOB_UNICODE 0xDEADBEEF
587 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
590 /* length of regex to show in messages that don't mark a position within */
591 #define RegexLengthToShowInErrorMessages 127
594 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
595 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
596 * op/pragma/warn/regcomp.
598 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
599 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
601 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
602 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
604 /* The code in this file in places uses one level of recursion with parsing
605 * rebased to an alternate string constructed by us in memory. This can take
606 * the form of something that is completely different from the input, or
607 * something that uses the input as part of the alternate. In the first case,
608 * there should be no possibility of an error, as we are in complete control of
609 * the alternate string. But in the second case we don't control the input
610 * portion, so there may be errors in that. Here's an example:
612 * is handled specially because \x{df} folds to a sequence of more than one
613 * character, 'ss'. What is done is to create and parse an alternate string,
614 * which looks like this:
615 * /(?:\x{DF}|[abc\x{DF}def])/ui
616 * where it uses the input unchanged in the middle of something it constructs,
617 * which is a branch for the DF outside the character class, and clustering
618 * parens around the whole thing. (It knows enough to skip the DF inside the
619 * class while in this substitute parse.) 'abc' and 'def' may have errors that
620 * need to be reported. The general situation looks like this:
623 * Input: ----------------------------------------------------
624 * Constructed: ---------------------------------------------------
627 * The input string sI..eI is the input pattern. The string sC..EC is the
628 * constructed substitute parse string. The portions sC..tC and eC..EC are
629 * constructed by us. The portion tC..eC is an exact duplicate of the input
630 * pattern tI..eI. In the diagram, these are vertically aligned. Suppose that
631 * while parsing, we find an error at xC. We want to display a message showing
632 * the real input string. Thus we need to find the point xI in it which
633 * corresponds to xC. xC >= tC, since the portion of the string sC..tC has
634 * been constructed by us, and so shouldn't have errors. We get:
636 * xI = sI + (tI - sI) + (xC - tC)
638 * and, the offset into sI is:
640 * (xI - sI) = (tI - sI) + (xC - tC)
642 * When the substitute is constructed, we save (tI -sI) as RExC_precomp_adj,
643 * and we save tC as RExC_adjusted_start.
645 * During normal processing of the input pattern, everything points to that,
646 * with RExC_precomp_adj set to 0, and RExC_adjusted_start set to sI.
649 #define tI_sI RExC_precomp_adj
650 #define tC RExC_adjusted_start
651 #define sC RExC_precomp
652 #define xI_offset(xC) ((IV) (tI_sI + (xC - tC)))
653 #define xI(xC) (sC + xI_offset(xC))
654 #define eC RExC_precomp_end
656 #define REPORT_LOCATION_ARGS(xC) \
658 (xI(xC) > eC) /* Don't run off end */ \
659 ? eC - sC /* Length before the <--HERE */ \
660 : ((xI_offset(xC) >= 0) \
662 : (Perl_croak(aTHX_ "panic: %s: %d: negative offset: %" \
663 IVdf " trying to output message for " \
665 __FILE__, __LINE__, xI_offset(xC), \
666 ((int) (eC - sC)), sC), 0)), \
667 sC), /* The input pattern printed up to the <--HERE */ \
669 (xI(xC) > eC) ? 0 : eC - xI(xC), /* Length after <--HERE */ \
670 (xI(xC) > eC) ? eC : xI(xC)) /* pattern after <--HERE */
672 /* Used to point after bad bytes for an error message, but avoid skipping
673 * past a nul byte. */
674 #define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
677 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
678 * arg. Show regex, up to a maximum length. If it's too long, chop and add
681 #define _FAIL(code) STMT_START { \
682 const char *ellipses = ""; \
683 IV len = RExC_precomp_end - RExC_precomp; \
686 SAVEFREESV(RExC_rx_sv); \
687 if (len > RegexLengthToShowInErrorMessages) { \
688 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
689 len = RegexLengthToShowInErrorMessages - 10; \
695 #define FAIL(msg) _FAIL( \
696 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
697 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
699 #define FAIL2(msg,arg) _FAIL( \
700 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
701 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
704 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
706 #define Simple_vFAIL(m) STMT_START { \
707 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
708 m, REPORT_LOCATION_ARGS(RExC_parse)); \
712 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
714 #define vFAIL(m) STMT_START { \
716 SAVEFREESV(RExC_rx_sv); \
721 * Like Simple_vFAIL(), but accepts two arguments.
723 #define Simple_vFAIL2(m,a1) STMT_START { \
724 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
725 REPORT_LOCATION_ARGS(RExC_parse)); \
729 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
731 #define vFAIL2(m,a1) STMT_START { \
733 SAVEFREESV(RExC_rx_sv); \
734 Simple_vFAIL2(m, a1); \
739 * Like Simple_vFAIL(), but accepts three arguments.
741 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
742 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
743 REPORT_LOCATION_ARGS(RExC_parse)); \
747 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
749 #define vFAIL3(m,a1,a2) STMT_START { \
751 SAVEFREESV(RExC_rx_sv); \
752 Simple_vFAIL3(m, a1, a2); \
756 * Like Simple_vFAIL(), but accepts four arguments.
758 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
759 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
760 REPORT_LOCATION_ARGS(RExC_parse)); \
763 #define vFAIL4(m,a1,a2,a3) STMT_START { \
765 SAVEFREESV(RExC_rx_sv); \
766 Simple_vFAIL4(m, a1, a2, a3); \
769 /* A specialized version of vFAIL2 that works with UTF8f */
770 #define vFAIL2utf8f(m, a1) STMT_START { \
772 SAVEFREESV(RExC_rx_sv); \
773 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
774 REPORT_LOCATION_ARGS(RExC_parse)); \
777 #define vFAIL3utf8f(m, a1, a2) STMT_START { \
779 SAVEFREESV(RExC_rx_sv); \
780 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
781 REPORT_LOCATION_ARGS(RExC_parse)); \
784 /* These have asserts in them because of [perl #122671] Many warnings in
785 * regcomp.c can occur twice. If they get output in pass1 and later in that
786 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
787 * would get output again. So they should be output in pass2, and these
788 * asserts make sure new warnings follow that paradigm. */
790 /* m is not necessarily a "literal string", in this macro */
791 #define reg_warn_non_literal_string(loc, m) STMT_START { \
792 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
793 "%s" REPORT_LOCATION, \
794 m, REPORT_LOCATION_ARGS(loc)); \
797 #define ckWARNreg(loc,m) STMT_START { \
798 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
800 REPORT_LOCATION_ARGS(loc)); \
803 #define vWARN(loc, m) STMT_START { \
804 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
806 REPORT_LOCATION_ARGS(loc)); \
809 #define vWARN_dep(loc, m) STMT_START { \
810 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
812 REPORT_LOCATION_ARGS(loc)); \
815 #define ckWARNdep(loc,m) STMT_START { \
816 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
818 REPORT_LOCATION_ARGS(loc)); \
821 #define ckWARNregdep(loc,m) STMT_START { \
822 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
825 REPORT_LOCATION_ARGS(loc)); \
828 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
829 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
831 a1, REPORT_LOCATION_ARGS(loc)); \
834 #define ckWARN2reg(loc, m, a1) STMT_START { \
835 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
837 a1, REPORT_LOCATION_ARGS(loc)); \
840 #define vWARN3(loc, m, a1, a2) STMT_START { \
841 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
843 a1, a2, REPORT_LOCATION_ARGS(loc)); \
846 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
847 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
850 REPORT_LOCATION_ARGS(loc)); \
853 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
854 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
857 REPORT_LOCATION_ARGS(loc)); \
860 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
861 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
864 REPORT_LOCATION_ARGS(loc)); \
867 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
868 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
871 REPORT_LOCATION_ARGS(loc)); \
874 /* Macros for recording node offsets. 20001227 mjd@plover.com
875 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
876 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
877 * Element 0 holds the number n.
878 * Position is 1 indexed.
880 #ifndef RE_TRACK_PATTERN_OFFSETS
881 #define Set_Node_Offset_To_R(node,byte)
882 #define Set_Node_Offset(node,byte)
883 #define Set_Cur_Node_Offset
884 #define Set_Node_Length_To_R(node,len)
885 #define Set_Node_Length(node,len)
886 #define Set_Node_Cur_Length(node,start)
887 #define Node_Offset(n)
888 #define Node_Length(n)
889 #define Set_Node_Offset_Length(node,offset,len)
890 #define ProgLen(ri) ri->u.proglen
891 #define SetProgLen(ri,x) ri->u.proglen = x
893 #define ProgLen(ri) ri->u.offsets[0]
894 #define SetProgLen(ri,x) ri->u.offsets[0] = x
895 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
897 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
898 __LINE__, (int)(node), (int)(byte))); \
900 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
903 RExC_offsets[2*(node)-1] = (byte); \
908 #define Set_Node_Offset(node,byte) \
909 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
910 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
912 #define Set_Node_Length_To_R(node,len) STMT_START { \
914 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
915 __LINE__, (int)(node), (int)(len))); \
917 Perl_croak(aTHX_ "value of node is %d in Length macro", \
920 RExC_offsets[2*(node)] = (len); \
925 #define Set_Node_Length(node,len) \
926 Set_Node_Length_To_R((node)-RExC_emit_start, len)
927 #define Set_Node_Cur_Length(node, start) \
928 Set_Node_Length(node, RExC_parse - start)
930 /* Get offsets and lengths */
931 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
932 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
934 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
935 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
936 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
940 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
941 #define EXPERIMENTAL_INPLACESCAN
942 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
946 Perl_re_printf(pTHX_ const char *fmt, ...)
950 PerlIO *f= Perl_debug_log;
951 PERL_ARGS_ASSERT_RE_PRINTF;
953 result = PerlIO_vprintf(f, fmt, ap);
959 Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
963 PerlIO *f= Perl_debug_log;
964 PERL_ARGS_ASSERT_RE_INDENTF;
966 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
967 result = PerlIO_vprintf(f, fmt, ap);
971 #endif /* DEBUGGING */
973 #define DEBUG_RExC_seen() \
974 DEBUG_OPTIMISE_MORE_r({ \
975 Perl_re_printf( aTHX_ "RExC_seen: "); \
977 if (RExC_seen & REG_ZERO_LEN_SEEN) \
978 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
980 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
981 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
983 if (RExC_seen & REG_GPOS_SEEN) \
984 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
986 if (RExC_seen & REG_RECURSE_SEEN) \
987 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
989 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
990 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
992 if (RExC_seen & REG_VERBARG_SEEN) \
993 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
995 if (RExC_seen & REG_CUTGROUP_SEEN) \
996 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
998 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
999 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
1001 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
1002 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
1004 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
1005 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
1007 Perl_re_printf( aTHX_ "\n"); \
1010 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
1011 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
1016 S_debug_show_study_flags(pTHX_ U32 flags, const char *open_str,
1017 const char *close_str)
1022 Perl_re_printf( aTHX_ "%s", open_str);
1023 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_SEOL);
1024 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_MEOL);
1025 DEBUG_SHOW_STUDY_FLAG(flags, SF_IS_INF);
1026 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_PAR);
1027 DEBUG_SHOW_STUDY_FLAG(flags, SF_IN_PAR);
1028 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_EVAL);
1029 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_SUBSTR);
1030 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_AND);
1031 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_OR);
1032 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS);
1033 DEBUG_SHOW_STUDY_FLAG(flags, SCF_WHILEM_VISITED_POS);
1034 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_RESTUDY);
1035 DEBUG_SHOW_STUDY_FLAG(flags, SCF_SEEN_ACCEPT);
1036 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_DOING_RESTUDY);
1037 DEBUG_SHOW_STUDY_FLAG(flags, SCF_IN_DEFINE);
1038 Perl_re_printf( aTHX_ "%s", close_str);
1043 S_debug_studydata(pTHX_ const char *where, scan_data_t *data,
1044 U32 depth, int is_inf)
1046 GET_RE_DEBUG_FLAGS_DECL;
1048 DEBUG_OPTIMISE_MORE_r({
1051 Perl_re_indentf(aTHX_ "%s: Pos:%" IVdf "/%" IVdf " Flags: 0x%" UVXf,
1055 (IV)data->pos_delta,
1059 S_debug_show_study_flags(aTHX_ data->flags," [","]");
1061 Perl_re_printf( aTHX_
1062 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s",
1064 (IV)(data->last_closep ? *((data)->last_closep) : -1),
1065 is_inf ? "INF " : ""
1068 if (data->last_found) {
1070 Perl_re_printf(aTHX_
1071 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf,
1072 SvPVX_const(data->last_found),
1074 (IV)data->last_start_min,
1075 (IV)data->last_start_max
1078 for (i = 0; i < 2; i++) {
1079 Perl_re_printf(aTHX_
1080 " %s%s: '%s' @ %" IVdf "/%" IVdf,
1081 data->cur_is_floating == i ? "*" : "",
1082 i ? "Float" : "Fixed",
1083 SvPVX_const(data->substrs[i].str),
1084 (IV)data->substrs[i].min_offset,
1085 (IV)data->substrs[i].max_offset
1087 S_debug_show_study_flags(aTHX_ data->substrs[i].flags," [","]");
1091 Perl_re_printf( aTHX_ "\n");
1097 S_debug_peep(pTHX_ const char *str, const RExC_state_t *pRExC_state,
1098 regnode *scan, U32 depth, U32 flags)
1100 GET_RE_DEBUG_FLAGS_DECL;
1107 Next = regnext(scan);
1108 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
1109 Perl_re_indentf( aTHX_ "%s>%3d: %s (%d)",
1112 REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),
1113 Next ? (REG_NODE_NUM(Next)) : 0 );
1114 S_debug_show_study_flags(aTHX_ flags," [ ","]");
1115 Perl_re_printf( aTHX_ "\n");
1120 # define DEBUG_STUDYDATA(where, data, depth, is_inf) \
1121 S_debug_studydata(aTHX_ where, data, depth, is_inf)
1123 # define DEBUG_PEEP(str, scan, depth, flags) \
1124 S_debug_peep(aTHX_ str, pRExC_state, scan, depth, flags)
1127 # define DEBUG_STUDYDATA(where, data, depth, is_inf) NOOP
1128 # define DEBUG_PEEP(str, scan, depth, flags) NOOP
1132 /* =========================================================
1133 * BEGIN edit_distance stuff.
1135 * This calculates how many single character changes of any type are needed to
1136 * transform a string into another one. It is taken from version 3.1 of
1138 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1141 /* Our unsorted dictionary linked list. */
1142 /* Note we use UVs, not chars. */
1147 struct dictionary* next;
1149 typedef struct dictionary item;
1152 PERL_STATIC_INLINE item*
1153 push(UV key,item* curr)
1156 Newx(head, 1, item);
1164 PERL_STATIC_INLINE item*
1165 find(item* head, UV key)
1167 item* iterator = head;
1169 if (iterator->key == key){
1172 iterator = iterator->next;
1178 PERL_STATIC_INLINE item*
1179 uniquePush(item* head,UV key)
1181 item* iterator = head;
1184 if (iterator->key == key) {
1187 iterator = iterator->next;
1190 return push(key,head);
1193 PERL_STATIC_INLINE void
1194 dict_free(item* head)
1196 item* iterator = head;
1199 item* temp = iterator;
1200 iterator = iterator->next;
1207 /* End of Dictionary Stuff */
1209 /* All calculations/work are done here */
1211 S_edit_distance(const UV* src,
1213 const STRLEN x, /* length of src[] */
1214 const STRLEN y, /* length of tgt[] */
1215 const SSize_t maxDistance
1219 UV swapCount,swapScore,targetCharCount,i,j;
1221 UV score_ceil = x + y;
1223 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1225 /* intialize matrix start values */
1226 Newx(scores, ( (x + 2) * (y + 2)), UV);
1227 scores[0] = score_ceil;
1228 scores[1 * (y + 2) + 0] = score_ceil;
1229 scores[0 * (y + 2) + 1] = score_ceil;
1230 scores[1 * (y + 2) + 1] = 0;
1231 head = uniquePush(uniquePush(head,src[0]),tgt[0]);
1236 for (i=1;i<=x;i++) {
1238 head = uniquePush(head,src[i]);
1239 scores[(i+1) * (y + 2) + 1] = i;
1240 scores[(i+1) * (y + 2) + 0] = score_ceil;
1243 for (j=1;j<=y;j++) {
1246 head = uniquePush(head,tgt[j]);
1247 scores[1 * (y + 2) + (j + 1)] = j;
1248 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1251 targetCharCount = find(head,tgt[j-1])->value;
1252 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1254 if (src[i-1] != tgt[j-1]){
1255 scores[(i+1) * (y + 2) + (j + 1)] = MIN(swapScore,(MIN(scores[i * (y + 2) + j], MIN(scores[(i+1) * (y + 2) + j], scores[i * (y + 2) + (j + 1)])) + 1));
1259 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1263 find(head,src[i-1])->value = i;
1267 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1270 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1274 /* END of edit_distance() stuff
1275 * ========================================================= */
1277 /* is c a control character for which we have a mnemonic? */
1278 #define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
1281 S_cntrl_to_mnemonic(const U8 c)
1283 /* Returns the mnemonic string that represents character 'c', if one
1284 * exists; NULL otherwise. The only ones that exist for the purposes of
1285 * this routine are a few control characters */
1288 case '\a': return "\\a";
1289 case '\b': return "\\b";
1290 case ESC_NATIVE: return "\\e";
1291 case '\f': return "\\f";
1292 case '\n': return "\\n";
1293 case '\r': return "\\r";
1294 case '\t': return "\\t";
1300 /* Mark that we cannot extend a found fixed substring at this point.
1301 Update the longest found anchored substring or the longest found
1302 floating substrings if needed. */
1305 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1306 SSize_t *minlenp, int is_inf)
1308 const STRLEN l = CHR_SVLEN(data->last_found);
1309 SV * const longest_sv = data->substrs[data->cur_is_floating].str;
1310 const STRLEN old_l = CHR_SVLEN(longest_sv);
1311 GET_RE_DEBUG_FLAGS_DECL;
1313 PERL_ARGS_ASSERT_SCAN_COMMIT;
1315 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
1316 const U8 i = data->cur_is_floating;
1317 SvSetMagicSV(longest_sv, data->last_found);
1318 data->substrs[i].min_offset = l ? data->last_start_min : data->pos_min;
1321 data->substrs[0].max_offset = data->substrs[0].min_offset;
1323 data->substrs[1].max_offset = (l
1324 ? data->last_start_max
1325 : (data->pos_delta > SSize_t_MAX - data->pos_min
1327 : data->pos_min + data->pos_delta));
1329 || (STRLEN)data->substrs[1].max_offset > (STRLEN)SSize_t_MAX)
1330 data->substrs[1].max_offset = SSize_t_MAX;
1333 if (data->flags & SF_BEFORE_EOL)
1334 data->substrs[i].flags |= (data->flags & SF_BEFORE_EOL);
1336 data->substrs[i].flags &= ~SF_BEFORE_EOL;
1337 data->substrs[i].minlenp = minlenp;
1338 data->substrs[i].lookbehind = 0;
1341 SvCUR_set(data->last_found, 0);
1343 SV * const sv = data->last_found;
1344 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1345 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1350 data->last_end = -1;
1351 data->flags &= ~SF_BEFORE_EOL;
1352 DEBUG_STUDYDATA("commit", data, 0, is_inf);
1355 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1356 * list that describes which code points it matches */
1359 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1361 /* Set the SSC 'ssc' to match an empty string or any code point */
1363 PERL_ARGS_ASSERT_SSC_ANYTHING;
1365 assert(is_ANYOF_SYNTHETIC(ssc));
1367 /* mortalize so won't leak */
1368 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
1369 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1373 S_ssc_is_anything(const regnode_ssc *ssc)
1375 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1376 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1377 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1378 * in any way, so there's no point in using it */
1383 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1385 assert(is_ANYOF_SYNTHETIC(ssc));
1387 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1391 /* See if the list consists solely of the range 0 - Infinity */
1392 invlist_iterinit(ssc->invlist);
1393 ret = invlist_iternext(ssc->invlist, &start, &end)
1397 invlist_iterfinish(ssc->invlist);
1403 /* If e.g., both \w and \W are set, matches everything */
1404 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1406 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1407 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1417 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1419 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1420 * string, any code point, or any posix class under locale */
1422 PERL_ARGS_ASSERT_SSC_INIT;
1424 Zero(ssc, 1, regnode_ssc);
1425 set_ANYOF_SYNTHETIC(ssc);
1426 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1429 /* If any portion of the regex is to operate under locale rules that aren't
1430 * fully known at compile time, initialization includes it. The reason
1431 * this isn't done for all regexes is that the optimizer was written under
1432 * the assumption that locale was all-or-nothing. Given the complexity and
1433 * lack of documentation in the optimizer, and that there are inadequate
1434 * test cases for locale, many parts of it may not work properly, it is
1435 * safest to avoid locale unless necessary. */
1436 if (RExC_contains_locale) {
1437 ANYOF_POSIXL_SETALL(ssc);
1440 ANYOF_POSIXL_ZERO(ssc);
1445 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1446 const regnode_ssc *ssc)
1448 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1449 * to the list of code points matched, and locale posix classes; hence does
1450 * not check its flags) */
1455 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1457 assert(is_ANYOF_SYNTHETIC(ssc));
1459 invlist_iterinit(ssc->invlist);
1460 ret = invlist_iternext(ssc->invlist, &start, &end)
1464 invlist_iterfinish(ssc->invlist);
1470 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1478 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1479 const regnode_charclass* const node)
1481 /* Returns a mortal inversion list defining which code points are matched
1482 * by 'node', which is of type ANYOF. Handles complementing the result if
1483 * appropriate. If some code points aren't knowable at this time, the
1484 * returned list must, and will, contain every code point that is a
1488 SV* only_utf8_locale_invlist = NULL;
1490 const U32 n = ARG(node);
1491 bool new_node_has_latin1 = FALSE;
1493 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1495 /* Look at the data structure created by S_set_ANYOF_arg() */
1496 if (n != ANYOF_ONLY_HAS_BITMAP) {
1497 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1498 AV * const av = MUTABLE_AV(SvRV(rv));
1499 SV **const ary = AvARRAY(av);
1500 assert(RExC_rxi->data->what[n] == 's');
1502 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1503 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1505 else if (ary[0] && ary[0] != &PL_sv_undef) {
1507 /* Here, no compile-time swash, and there are things that won't be
1508 * known until runtime -- we have to assume it could be anything */
1509 invlist = sv_2mortal(_new_invlist(1));
1510 return _add_range_to_invlist(invlist, 0, UV_MAX);
1512 else if (ary[3] && ary[3] != &PL_sv_undef) {
1514 /* Here no compile-time swash, and no run-time only data. Use the
1515 * node's inversion list */
1516 invlist = sv_2mortal(invlist_clone(ary[3]));
1519 /* Get the code points valid only under UTF-8 locales */
1520 if ((ANYOF_FLAGS(node) & ANYOFL_FOLD)
1521 && ary[2] && ary[2] != &PL_sv_undef)
1523 only_utf8_locale_invlist = ary[2];
1528 invlist = sv_2mortal(_new_invlist(0));
1531 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1532 * code points, and an inversion list for the others, but if there are code
1533 * points that should match only conditionally on the target string being
1534 * UTF-8, those are placed in the inversion list, and not the bitmap.
1535 * Since there are circumstances under which they could match, they are
1536 * included in the SSC. But if the ANYOF node is to be inverted, we have
1537 * to exclude them here, so that when we invert below, the end result
1538 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1539 * have to do this here before we add the unconditionally matched code
1541 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1542 _invlist_intersection_complement_2nd(invlist,
1547 /* Add in the points from the bit map */
1548 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1549 if (ANYOF_BITMAP_TEST(node, i)) {
1550 unsigned int start = i++;
1552 for (; i < NUM_ANYOF_CODE_POINTS && ANYOF_BITMAP_TEST(node, i); ++i) {
1555 invlist = _add_range_to_invlist(invlist, start, i-1);
1556 new_node_has_latin1 = TRUE;
1560 /* If this can match all upper Latin1 code points, have to add them
1561 * as well. But don't add them if inverting, as when that gets done below,
1562 * it would exclude all these characters, including the ones it shouldn't
1563 * that were added just above */
1564 if (! (ANYOF_FLAGS(node) & ANYOF_INVERT) && OP(node) == ANYOFD
1565 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1567 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1570 /* Similarly for these */
1571 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1572 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1575 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1576 _invlist_invert(invlist);
1578 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOFL_FOLD) {
1580 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1581 * locale. We can skip this if there are no 0-255 at all. */
1582 _invlist_union(invlist, PL_Latin1, &invlist);
1585 /* Similarly add the UTF-8 locale possible matches. These have to be
1586 * deferred until after the non-UTF-8 locale ones are taken care of just
1587 * above, or it leads to wrong results under ANYOF_INVERT */
1588 if (only_utf8_locale_invlist) {
1589 _invlist_union_maybe_complement_2nd(invlist,
1590 only_utf8_locale_invlist,
1591 ANYOF_FLAGS(node) & ANYOF_INVERT,
1598 /* These two functions currently do the exact same thing */
1599 #define ssc_init_zero ssc_init
1601 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1602 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1604 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1605 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1606 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1609 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1610 const regnode_charclass *and_with)
1612 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1613 * another SSC or a regular ANYOF class. Can create false positives. */
1618 PERL_ARGS_ASSERT_SSC_AND;
1620 assert(is_ANYOF_SYNTHETIC(ssc));
1622 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1623 * the code point inversion list and just the relevant flags */
1624 if (is_ANYOF_SYNTHETIC(and_with)) {
1625 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1626 anded_flags = ANYOF_FLAGS(and_with);
1628 /* XXX This is a kludge around what appears to be deficiencies in the
1629 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1630 * there are paths through the optimizer where it doesn't get weeded
1631 * out when it should. And if we don't make some extra provision for
1632 * it like the code just below, it doesn't get added when it should.
1633 * This solution is to add it only when AND'ing, which is here, and
1634 * only when what is being AND'ed is the pristine, original node
1635 * matching anything. Thus it is like adding it to ssc_anything() but
1636 * only when the result is to be AND'ed. Probably the same solution
1637 * could be adopted for the same problem we have with /l matching,
1638 * which is solved differently in S_ssc_init(), and that would lead to
1639 * fewer false positives than that solution has. But if this solution
1640 * creates bugs, the consequences are only that a warning isn't raised
1641 * that should be; while the consequences for having /l bugs is
1642 * incorrect matches */
1643 if (ssc_is_anything((regnode_ssc *)and_with)) {
1644 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1648 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1649 if (OP(and_with) == ANYOFD) {
1650 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1653 anded_flags = ANYOF_FLAGS(and_with)
1654 &( ANYOF_COMMON_FLAGS
1655 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1656 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1657 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(and_with))) {
1659 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1664 ANYOF_FLAGS(ssc) &= anded_flags;
1666 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1667 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1668 * 'and_with' may be inverted. When not inverted, we have the situation of
1670 * (C1 | P1) & (C2 | P2)
1671 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1672 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1673 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1674 * <= ((C1 & C2) | P1 | P2)
1675 * Alternatively, the last few steps could be:
1676 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1677 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1678 * <= (C1 | C2 | (P1 & P2))
1679 * We favor the second approach if either P1 or P2 is non-empty. This is
1680 * because these components are a barrier to doing optimizations, as what
1681 * they match cannot be known until the moment of matching as they are
1682 * dependent on the current locale, 'AND"ing them likely will reduce or
1684 * But we can do better if we know that C1,P1 are in their initial state (a
1685 * frequent occurrence), each matching everything:
1686 * (<everything>) & (C2 | P2) = C2 | P2
1687 * Similarly, if C2,P2 are in their initial state (again a frequent
1688 * occurrence), the result is a no-op
1689 * (C1 | P1) & (<everything>) = C1 | P1
1692 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1693 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1694 * <= (C1 & ~C2) | (P1 & ~P2)
1697 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1698 && ! is_ANYOF_SYNTHETIC(and_with))
1702 ssc_intersection(ssc,
1704 FALSE /* Has already been inverted */
1707 /* If either P1 or P2 is empty, the intersection will be also; can skip
1709 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
1710 ANYOF_POSIXL_ZERO(ssc);
1712 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1714 /* Note that the Posix class component P from 'and_with' actually
1716 * P = Pa | Pb | ... | Pn
1717 * where each component is one posix class, such as in [\w\s].
1719 * ~P = ~(Pa | Pb | ... | Pn)
1720 * = ~Pa & ~Pb & ... & ~Pn
1721 * <= ~Pa | ~Pb | ... | ~Pn
1722 * The last is something we can easily calculate, but unfortunately
1723 * is likely to have many false positives. We could do better
1724 * in some (but certainly not all) instances if two classes in
1725 * P have known relationships. For example
1726 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1728 * :lower: & :print: = :lower:
1729 * And similarly for classes that must be disjoint. For example,
1730 * since \s and \w can have no elements in common based on rules in
1731 * the POSIX standard,
1732 * \w & ^\S = nothing
1733 * Unfortunately, some vendor locales do not meet the Posix
1734 * standard, in particular almost everything by Microsoft.
1735 * The loop below just changes e.g., \w into \W and vice versa */
1737 regnode_charclass_posixl temp;
1738 int add = 1; /* To calculate the index of the complement */
1740 Zero(&temp, 1, regnode_charclass_posixl);
1741 ANYOF_POSIXL_ZERO(&temp);
1742 for (i = 0; i < ANYOF_MAX; i++) {
1744 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1745 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1747 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1748 ANYOF_POSIXL_SET(&temp, i + add);
1750 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1752 ANYOF_POSIXL_AND(&temp, ssc);
1754 } /* else ssc already has no posixes */
1755 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1756 in its initial state */
1757 else if (! is_ANYOF_SYNTHETIC(and_with)
1758 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1760 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1761 * copy it over 'ssc' */
1762 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1763 if (is_ANYOF_SYNTHETIC(and_with)) {
1764 StructCopy(and_with, ssc, regnode_ssc);
1767 ssc->invlist = anded_cp_list;
1768 ANYOF_POSIXL_ZERO(ssc);
1769 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1770 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1774 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1775 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
1777 /* One or the other of P1, P2 is non-empty. */
1778 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1779 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1781 ssc_union(ssc, anded_cp_list, FALSE);
1783 else { /* P1 = P2 = empty */
1784 ssc_intersection(ssc, anded_cp_list, FALSE);
1790 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1791 const regnode_charclass *or_with)
1793 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1794 * another SSC or a regular ANYOF class. Can create false positives if
1795 * 'or_with' is to be inverted. */
1800 PERL_ARGS_ASSERT_SSC_OR;
1802 assert(is_ANYOF_SYNTHETIC(ssc));
1804 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1805 * the code point inversion list and just the relevant flags */
1806 if (is_ANYOF_SYNTHETIC(or_with)) {
1807 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1808 ored_flags = ANYOF_FLAGS(or_with);
1811 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1812 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1813 if (OP(or_with) != ANYOFD) {
1815 |= ANYOF_FLAGS(or_with)
1816 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1817 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1818 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(or_with))) {
1820 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1825 ANYOF_FLAGS(ssc) |= ored_flags;
1827 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1828 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1829 * 'or_with' may be inverted. When not inverted, we have the simple
1830 * situation of computing:
1831 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1832 * If P1|P2 yields a situation with both a class and its complement are
1833 * set, like having both \w and \W, this matches all code points, and we
1834 * can delete these from the P component of the ssc going forward. XXX We
1835 * might be able to delete all the P components, but I (khw) am not certain
1836 * about this, and it is better to be safe.
1839 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1840 * <= (C1 | P1) | ~C2
1841 * <= (C1 | ~C2) | P1
1842 * (which results in actually simpler code than the non-inverted case)
1845 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1846 && ! is_ANYOF_SYNTHETIC(or_with))
1848 /* We ignore P2, leaving P1 going forward */
1849 } /* else Not inverted */
1850 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
1851 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1852 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1854 for (i = 0; i < ANYOF_MAX; i += 2) {
1855 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1857 ssc_match_all_cp(ssc);
1858 ANYOF_POSIXL_CLEAR(ssc, i);
1859 ANYOF_POSIXL_CLEAR(ssc, i+1);
1867 FALSE /* Already has been inverted */
1871 PERL_STATIC_INLINE void
1872 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1874 PERL_ARGS_ASSERT_SSC_UNION;
1876 assert(is_ANYOF_SYNTHETIC(ssc));
1878 _invlist_union_maybe_complement_2nd(ssc->invlist,
1884 PERL_STATIC_INLINE void
1885 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1887 const bool invert2nd)
1889 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1891 assert(is_ANYOF_SYNTHETIC(ssc));
1893 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1899 PERL_STATIC_INLINE void
1900 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1902 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1904 assert(is_ANYOF_SYNTHETIC(ssc));
1906 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1909 PERL_STATIC_INLINE void
1910 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1912 /* AND just the single code point 'cp' into the SSC 'ssc' */
1914 SV* cp_list = _new_invlist(2);
1916 PERL_ARGS_ASSERT_SSC_CP_AND;
1918 assert(is_ANYOF_SYNTHETIC(ssc));
1920 cp_list = add_cp_to_invlist(cp_list, cp);
1921 ssc_intersection(ssc, cp_list,
1922 FALSE /* Not inverted */
1924 SvREFCNT_dec_NN(cp_list);
1927 PERL_STATIC_INLINE void
1928 S_ssc_clear_locale(regnode_ssc *ssc)
1930 /* Set the SSC 'ssc' to not match any locale things */
1931 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1933 assert(is_ANYOF_SYNTHETIC(ssc));
1935 ANYOF_POSIXL_ZERO(ssc);
1936 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1939 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1942 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1944 /* The synthetic start class is used to hopefully quickly winnow down
1945 * places where a pattern could start a match in the target string. If it
1946 * doesn't really narrow things down that much, there isn't much point to
1947 * having the overhead of using it. This function uses some very crude
1948 * heuristics to decide if to use the ssc or not.
1950 * It returns TRUE if 'ssc' rules out more than half what it considers to
1951 * be the "likely" possible matches, but of course it doesn't know what the
1952 * actual things being matched are going to be; these are only guesses
1954 * For /l matches, it assumes that the only likely matches are going to be
1955 * in the 0-255 range, uniformly distributed, so half of that is 127
1956 * For /a and /d matches, it assumes that the likely matches will be just
1957 * the ASCII range, so half of that is 63
1958 * For /u and there isn't anything matching above the Latin1 range, it
1959 * assumes that that is the only range likely to be matched, and uses
1960 * half that as the cut-off: 127. If anything matches above Latin1,
1961 * it assumes that all of Unicode could match (uniformly), except for
1962 * non-Unicode code points and things in the General Category "Other"
1963 * (unassigned, private use, surrogates, controls and formats). This
1964 * is a much large number. */
1966 U32 count = 0; /* Running total of number of code points matched by
1968 UV start, end; /* Start and end points of current range in inversion
1970 const U32 max_code_points = (LOC)
1972 : (( ! UNI_SEMANTICS
1973 || invlist_highest(ssc->invlist) < 256)
1976 const U32 max_match = max_code_points / 2;
1978 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1980 invlist_iterinit(ssc->invlist);
1981 while (invlist_iternext(ssc->invlist, &start, &end)) {
1982 if (start >= max_code_points) {
1985 end = MIN(end, max_code_points - 1);
1986 count += end - start + 1;
1987 if (count >= max_match) {
1988 invlist_iterfinish(ssc->invlist);
1998 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
2000 /* The inversion list in the SSC is marked mortal; now we need a more
2001 * permanent copy, which is stored the same way that is done in a regular
2002 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
2005 SV* invlist = invlist_clone(ssc->invlist);
2007 PERL_ARGS_ASSERT_SSC_FINALIZE;
2009 assert(is_ANYOF_SYNTHETIC(ssc));
2011 /* The code in this file assumes that all but these flags aren't relevant
2012 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
2013 * by the time we reach here */
2014 assert(! (ANYOF_FLAGS(ssc)
2015 & ~( ANYOF_COMMON_FLAGS
2016 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
2017 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
2019 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
2021 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
2022 NULL, NULL, NULL, FALSE);
2024 /* Make sure is clone-safe */
2025 ssc->invlist = NULL;
2027 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
2028 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
2031 if (RExC_contains_locale) {
2035 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
2038 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
2039 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
2040 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
2041 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
2042 ? (TRIE_LIST_CUR( idx ) - 1) \
2048 dump_trie(trie,widecharmap,revcharmap)
2049 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
2050 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
2052 These routines dump out a trie in a somewhat readable format.
2053 The _interim_ variants are used for debugging the interim
2054 tables that are used to generate the final compressed
2055 representation which is what dump_trie expects.
2057 Part of the reason for their existence is to provide a form
2058 of documentation as to how the different representations function.
2063 Dumps the final compressed table form of the trie to Perl_debug_log.
2064 Used for debugging make_trie().
2068 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
2069 AV *revcharmap, U32 depth)
2072 SV *sv=sv_newmortal();
2073 int colwidth= widecharmap ? 6 : 4;
2075 GET_RE_DEBUG_FLAGS_DECL;
2077 PERL_ARGS_ASSERT_DUMP_TRIE;
2079 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
2080 depth+1, "Match","Base","Ofs" );
2082 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2083 SV ** const tmp = av_fetch( revcharmap, state, 0);
2085 Perl_re_printf( aTHX_ "%*s",
2087 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2088 PL_colors[0], PL_colors[1],
2089 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2090 PERL_PV_ESCAPE_FIRSTCHAR
2095 Perl_re_printf( aTHX_ "\n");
2096 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
2098 for( state = 0 ; state < trie->uniquecharcount ; state++ )
2099 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2100 Perl_re_printf( aTHX_ "\n");
2102 for( state = 1 ; state < trie->statecount ; state++ ) {
2103 const U32 base = trie->states[ state ].trans.base;
2105 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
2107 if ( trie->states[ state ].wordnum ) {
2108 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
2110 Perl_re_printf( aTHX_ "%6s", "" );
2113 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
2118 while( ( base + ofs < trie->uniquecharcount ) ||
2119 ( base + ofs - trie->uniquecharcount < trie->lasttrans
2120 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2124 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
2126 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2127 if ( ( base + ofs >= trie->uniquecharcount )
2128 && ( base + ofs - trie->uniquecharcount
2130 && trie->trans[ base + ofs
2131 - trie->uniquecharcount ].check == state )
2133 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
2134 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2137 Perl_re_printf( aTHX_ "%*s",colwidth," ." );
2141 Perl_re_printf( aTHX_ "]");
2144 Perl_re_printf( aTHX_ "\n" );
2146 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
2148 for (word=1; word <= trie->wordcount; word++) {
2149 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2150 (int)word, (int)(trie->wordinfo[word].prev),
2151 (int)(trie->wordinfo[word].len));
2153 Perl_re_printf( aTHX_ "\n" );
2156 Dumps a fully constructed but uncompressed trie in list form.
2157 List tries normally only are used for construction when the number of
2158 possible chars (trie->uniquecharcount) is very high.
2159 Used for debugging make_trie().
2162 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2163 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2167 SV *sv=sv_newmortal();
2168 int colwidth= widecharmap ? 6 : 4;
2169 GET_RE_DEBUG_FLAGS_DECL;
2171 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2173 /* print out the table precompression. */
2174 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
2176 Perl_re_indentf( aTHX_ "%s",
2177 depth+1, "------:-----+-----------------\n" );
2179 for( state=1 ; state < next_alloc ; state ++ ) {
2182 Perl_re_indentf( aTHX_ " %4" UVXf " :",
2183 depth+1, (UV)state );
2184 if ( ! trie->states[ state ].wordnum ) {
2185 Perl_re_printf( aTHX_ "%5s| ","");
2187 Perl_re_printf( aTHX_ "W%4x| ",
2188 trie->states[ state ].wordnum
2191 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2192 SV ** const tmp = av_fetch( revcharmap,
2193 TRIE_LIST_ITEM(state,charid).forid, 0);
2195 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
2197 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2199 PL_colors[0], PL_colors[1],
2200 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2201 | PERL_PV_ESCAPE_FIRSTCHAR
2203 TRIE_LIST_ITEM(state,charid).forid,
2204 (UV)TRIE_LIST_ITEM(state,charid).newstate
2207 Perl_re_printf( aTHX_ "\n%*s| ",
2208 (int)((depth * 2) + 14), "");
2211 Perl_re_printf( aTHX_ "\n");
2216 Dumps a fully constructed but uncompressed trie in table form.
2217 This is the normal DFA style state transition table, with a few
2218 twists to facilitate compression later.
2219 Used for debugging make_trie().
2222 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2223 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2228 SV *sv=sv_newmortal();
2229 int colwidth= widecharmap ? 6 : 4;
2230 GET_RE_DEBUG_FLAGS_DECL;
2232 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2235 print out the table precompression so that we can do a visual check
2236 that they are identical.
2239 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
2241 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2242 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2244 Perl_re_printf( aTHX_ "%*s",
2246 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2247 PL_colors[0], PL_colors[1],
2248 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2249 PERL_PV_ESCAPE_FIRSTCHAR
2255 Perl_re_printf( aTHX_ "\n");
2256 Perl_re_indentf( aTHX_ "State+-", depth+1 );
2258 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2259 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
2262 Perl_re_printf( aTHX_ "\n" );
2264 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2266 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
2268 (UV)TRIE_NODENUM( state ) );
2270 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2271 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2273 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
2275 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
2277 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2278 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
2279 (UV)trie->trans[ state ].check );
2281 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
2282 (UV)trie->trans[ state ].check,
2283 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2291 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2292 startbranch: the first branch in the whole branch sequence
2293 first : start branch of sequence of branch-exact nodes.
2294 May be the same as startbranch
2295 last : Thing following the last branch.
2296 May be the same as tail.
2297 tail : item following the branch sequence
2298 count : words in the sequence
2299 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2300 depth : indent depth
2302 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2304 A trie is an N'ary tree where the branches are determined by digital
2305 decomposition of the key. IE, at the root node you look up the 1st character and
2306 follow that branch repeat until you find the end of the branches. Nodes can be
2307 marked as "accepting" meaning they represent a complete word. Eg:
2311 would convert into the following structure. Numbers represent states, letters
2312 following numbers represent valid transitions on the letter from that state, if
2313 the number is in square brackets it represents an accepting state, otherwise it
2314 will be in parenthesis.
2316 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2320 (1) +-i->(6)-+-s->[7]
2322 +-s->(3)-+-h->(4)-+-e->[5]
2324 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2326 This shows that when matching against the string 'hers' we will begin at state 1
2327 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2328 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2329 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2330 single traverse. We store a mapping from accepting to state to which word was
2331 matched, and then when we have multiple possibilities we try to complete the
2332 rest of the regex in the order in which they occurred in the alternation.
2334 The only prior NFA like behaviour that would be changed by the TRIE support is
2335 the silent ignoring of duplicate alternations which are of the form:
2337 / (DUPE|DUPE) X? (?{ ... }) Y /x
2339 Thus EVAL blocks following a trie may be called a different number of times with
2340 and without the optimisation. With the optimisations dupes will be silently
2341 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2342 the following demonstrates:
2344 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2346 which prints out 'word' three times, but
2348 'words'=~/(word|word|word)(?{ print $1 })S/
2350 which doesnt print it out at all. This is due to other optimisations kicking in.
2352 Example of what happens on a structural level:
2354 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2356 1: CURLYM[1] {1,32767}(18)
2367 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2368 and should turn into:
2370 1: CURLYM[1] {1,32767}(18)
2372 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2380 Cases where tail != last would be like /(?foo|bar)baz/:
2390 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2391 and would end up looking like:
2394 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2401 d = uvchr_to_utf8_flags(d, uv, 0);
2403 is the recommended Unicode-aware way of saying
2408 #define TRIE_STORE_REVCHAR(val) \
2411 SV *zlopp = newSV(UTF8_MAXBYTES); \
2412 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2413 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2414 SvCUR_set(zlopp, kapow - flrbbbbb); \
2417 av_push(revcharmap, zlopp); \
2419 char ooooff = (char)val; \
2420 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2424 /* This gets the next character from the input, folding it if not already
2426 #define TRIE_READ_CHAR STMT_START { \
2429 /* if it is UTF then it is either already folded, or does not need \
2431 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2433 else if (folder == PL_fold_latin1) { \
2434 /* This folder implies Unicode rules, which in the range expressible \
2435 * by not UTF is the lower case, with the two exceptions, one of \
2436 * which should have been taken care of before calling this */ \
2437 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2438 uvc = toLOWER_L1(*uc); \
2439 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2442 /* raw data, will be folded later if needed */ \
2450 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2451 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2452 U32 ging = TRIE_LIST_LEN( state ) * 2; \
2453 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2454 TRIE_LIST_LEN( state ) = ging; \
2456 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2457 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2458 TRIE_LIST_CUR( state )++; \
2461 #define TRIE_LIST_NEW(state) STMT_START { \
2462 Newx( trie->states[ state ].trans.list, \
2463 4, reg_trie_trans_le ); \
2464 TRIE_LIST_CUR( state ) = 1; \
2465 TRIE_LIST_LEN( state ) = 4; \
2468 #define TRIE_HANDLE_WORD(state) STMT_START { \
2469 U16 dupe= trie->states[ state ].wordnum; \
2470 regnode * const noper_next = regnext( noper ); \
2473 /* store the word for dumping */ \
2475 if (OP(noper) != NOTHING) \
2476 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2478 tmp = newSVpvn_utf8( "", 0, UTF ); \
2479 av_push( trie_words, tmp ); \
2483 trie->wordinfo[curword].prev = 0; \
2484 trie->wordinfo[curword].len = wordlen; \
2485 trie->wordinfo[curword].accept = state; \
2487 if ( noper_next < tail ) { \
2489 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2491 trie->jump[curword] = (U16)(noper_next - convert); \
2493 jumper = noper_next; \
2495 nextbranch= regnext(cur); \
2499 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2500 /* chain, so that when the bits of chain are later */\
2501 /* linked together, the dups appear in the chain */\
2502 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2503 trie->wordinfo[dupe].prev = curword; \
2505 /* we haven't inserted this word yet. */ \
2506 trie->states[ state ].wordnum = curword; \
2511 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2512 ( ( base + charid >= ucharcount \
2513 && base + charid < ubound \
2514 && state == trie->trans[ base - ucharcount + charid ].check \
2515 && trie->trans[ base - ucharcount + charid ].next ) \
2516 ? trie->trans[ base - ucharcount + charid ].next \
2517 : ( state==1 ? special : 0 ) \
2520 #define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2522 TRIE_BITMAP_SET(trie, uvc); \
2523 /* store the folded codepoint */ \
2525 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2528 /* store first byte of utf8 representation of */ \
2529 /* variant codepoints */ \
2530 if (! UVCHR_IS_INVARIANT(uvc)) { \
2531 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2536 #define MADE_JUMP_TRIE 2
2537 #define MADE_EXACT_TRIE 4
2540 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2541 regnode *first, regnode *last, regnode *tail,
2542 U32 word_count, U32 flags, U32 depth)
2544 /* first pass, loop through and scan words */
2545 reg_trie_data *trie;
2546 HV *widecharmap = NULL;
2547 AV *revcharmap = newAV();
2553 regnode *jumper = NULL;
2554 regnode *nextbranch = NULL;
2555 regnode *convert = NULL;
2556 U32 *prev_states; /* temp array mapping each state to previous one */
2557 /* we just use folder as a flag in utf8 */
2558 const U8 * folder = NULL;
2560 /* in the below add_data call we are storing either 'tu' or 'tuaa'
2561 * which stands for one trie structure, one hash, optionally followed
2564 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuaa"));
2565 AV *trie_words = NULL;
2566 /* along with revcharmap, this only used during construction but both are
2567 * useful during debugging so we store them in the struct when debugging.
2570 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2571 STRLEN trie_charcount=0;
2573 SV *re_trie_maxbuff;
2574 GET_RE_DEBUG_FLAGS_DECL;
2576 PERL_ARGS_ASSERT_MAKE_TRIE;
2578 PERL_UNUSED_ARG(depth);
2582 case EXACT: case EXACTL: break;
2586 case EXACTFLU8: folder = PL_fold_latin1; break;
2587 case EXACTF: folder = PL_fold; break;
2588 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2591 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2593 trie->startstate = 1;
2594 trie->wordcount = word_count;
2595 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2596 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2597 if (flags == EXACT || flags == EXACTL)
2598 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2599 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2600 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2603 trie_words = newAV();
2606 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2607 assert(re_trie_maxbuff);
2608 if (!SvIOK(re_trie_maxbuff)) {
2609 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2611 DEBUG_TRIE_COMPILE_r({
2612 Perl_re_indentf( aTHX_
2613 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2615 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2616 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2619 /* Find the node we are going to overwrite */
2620 if ( first == startbranch && OP( last ) != BRANCH ) {
2621 /* whole branch chain */
2624 /* branch sub-chain */
2625 convert = NEXTOPER( first );
2628 /* -- First loop and Setup --
2630 We first traverse the branches and scan each word to determine if it
2631 contains widechars, and how many unique chars there are, this is
2632 important as we have to build a table with at least as many columns as we
2635 We use an array of integers to represent the character codes 0..255
2636 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2637 the native representation of the character value as the key and IV's for
2640 *TODO* If we keep track of how many times each character is used we can
2641 remap the columns so that the table compression later on is more
2642 efficient in terms of memory by ensuring the most common value is in the
2643 middle and the least common are on the outside. IMO this would be better
2644 than a most to least common mapping as theres a decent chance the most
2645 common letter will share a node with the least common, meaning the node
2646 will not be compressible. With a middle is most common approach the worst
2647 case is when we have the least common nodes twice.
2651 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2652 regnode *noper = NEXTOPER( cur );
2656 U32 wordlen = 0; /* required init */
2657 STRLEN minchars = 0;
2658 STRLEN maxchars = 0;
2659 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2662 if (OP(noper) == NOTHING) {
2663 /* skip past a NOTHING at the start of an alternation
2664 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2666 regnode *noper_next= regnext(noper);
2667 if (noper_next < tail)
2671 if ( noper < tail &&
2673 OP(noper) == flags ||
2676 OP(noper) == EXACTFU_SS
2680 uc= (U8*)STRING(noper);
2681 e= uc + STR_LEN(noper);
2688 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2689 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2690 regardless of encoding */
2691 if (OP( noper ) == EXACTFU_SS) {
2692 /* false positives are ok, so just set this */
2693 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2697 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2699 TRIE_CHARCOUNT(trie)++;
2702 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2703 * is in effect. Under /i, this character can match itself, or
2704 * anything that folds to it. If not under /i, it can match just
2705 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2706 * all fold to k, and all are single characters. But some folds
2707 * expand to more than one character, so for example LATIN SMALL
2708 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2709 * the string beginning at 'uc' is 'ffi', it could be matched by
2710 * three characters, or just by the one ligature character. (It
2711 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2712 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2713 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2714 * match.) The trie needs to know the minimum and maximum number
2715 * of characters that could match so that it can use size alone to
2716 * quickly reject many match attempts. The max is simple: it is
2717 * the number of folded characters in this branch (since a fold is
2718 * never shorter than what folds to it. */
2722 /* And the min is equal to the max if not under /i (indicated by
2723 * 'folder' being NULL), or there are no multi-character folds. If
2724 * there is a multi-character fold, the min is incremented just
2725 * once, for the character that folds to the sequence. Each
2726 * character in the sequence needs to be added to the list below of
2727 * characters in the trie, but we count only the first towards the
2728 * min number of characters needed. This is done through the
2729 * variable 'foldlen', which is returned by the macros that look
2730 * for these sequences as the number of bytes the sequence
2731 * occupies. Each time through the loop, we decrement 'foldlen' by
2732 * how many bytes the current char occupies. Only when it reaches
2733 * 0 do we increment 'minchars' or look for another multi-character
2735 if (folder == NULL) {
2738 else if (foldlen > 0) {
2739 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2744 /* See if *uc is the beginning of a multi-character fold. If
2745 * so, we decrement the length remaining to look at, to account
2746 * for the current character this iteration. (We can use 'uc'
2747 * instead of the fold returned by TRIE_READ_CHAR because for
2748 * non-UTF, the latin1_safe macro is smart enough to account
2749 * for all the unfolded characters, and because for UTF, the
2750 * string will already have been folded earlier in the
2751 * compilation process */
2753 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2754 foldlen -= UTF8SKIP(uc);
2757 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2762 /* The current character (and any potential folds) should be added
2763 * to the possible matching characters for this position in this
2767 U8 folded= folder[ (U8) uvc ];
2768 if ( !trie->charmap[ folded ] ) {
2769 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2770 TRIE_STORE_REVCHAR( folded );
2773 if ( !trie->charmap[ uvc ] ) {
2774 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2775 TRIE_STORE_REVCHAR( uvc );
2778 /* store the codepoint in the bitmap, and its folded
2780 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
2781 set_bit = 0; /* We've done our bit :-) */
2785 /* XXX We could come up with the list of code points that fold
2786 * to this using PL_utf8_foldclosures, except not for
2787 * multi-char folds, as there may be multiple combinations
2788 * there that could work, which needs to wait until runtime to
2789 * resolve (The comment about LIGATURE FFI above is such an
2794 widecharmap = newHV();
2796 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2799 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
2801 if ( !SvTRUE( *svpp ) ) {
2802 sv_setiv( *svpp, ++trie->uniquecharcount );
2803 TRIE_STORE_REVCHAR(uvc);
2806 } /* end loop through characters in this branch of the trie */
2808 /* We take the min and max for this branch and combine to find the min
2809 * and max for all branches processed so far */
2810 if( cur == first ) {
2811 trie->minlen = minchars;
2812 trie->maxlen = maxchars;
2813 } else if (minchars < trie->minlen) {
2814 trie->minlen = minchars;
2815 } else if (maxchars > trie->maxlen) {
2816 trie->maxlen = maxchars;
2818 } /* end first pass */
2819 DEBUG_TRIE_COMPILE_r(
2820 Perl_re_indentf( aTHX_
2821 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2823 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2824 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2825 (int)trie->minlen, (int)trie->maxlen )
2829 We now know what we are dealing with in terms of unique chars and
2830 string sizes so we can calculate how much memory a naive
2831 representation using a flat table will take. If it's over a reasonable
2832 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2833 conservative but potentially much slower representation using an array
2836 At the end we convert both representations into the same compressed
2837 form that will be used in regexec.c for matching with. The latter
2838 is a form that cannot be used to construct with but has memory
2839 properties similar to the list form and access properties similar
2840 to the table form making it both suitable for fast searches and
2841 small enough that its feasable to store for the duration of a program.
2843 See the comment in the code where the compressed table is produced
2844 inplace from the flat tabe representation for an explanation of how
2845 the compression works.
2850 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2853 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2854 > SvIV(re_trie_maxbuff) )
2857 Second Pass -- Array Of Lists Representation
2859 Each state will be represented by a list of charid:state records
2860 (reg_trie_trans_le) the first such element holds the CUR and LEN
2861 points of the allocated array. (See defines above).
2863 We build the initial structure using the lists, and then convert
2864 it into the compressed table form which allows faster lookups
2865 (but cant be modified once converted).
2868 STRLEN transcount = 1;
2870 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
2873 trie->states = (reg_trie_state *)
2874 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2875 sizeof(reg_trie_state) );
2879 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2881 regnode *noper = NEXTOPER( cur );
2882 U32 state = 1; /* required init */
2883 U16 charid = 0; /* sanity init */
2884 U32 wordlen = 0; /* required init */
2886 if (OP(noper) == NOTHING) {
2887 regnode *noper_next= regnext(noper);
2888 if (noper_next < tail)
2892 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
2893 const U8 *uc= (U8*)STRING(noper);
2894 const U8 *e= uc + STR_LEN(noper);
2896 for ( ; uc < e ; uc += len ) {
2901 charid = trie->charmap[ uvc ];
2903 SV** const svpp = hv_fetch( widecharmap,
2910 charid=(U16)SvIV( *svpp );
2913 /* charid is now 0 if we dont know the char read, or
2914 * nonzero if we do */
2921 if ( !trie->states[ state ].trans.list ) {
2922 TRIE_LIST_NEW( state );
2925 check <= TRIE_LIST_USED( state );
2928 if ( TRIE_LIST_ITEM( state, check ).forid
2931 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2936 newstate = next_alloc++;
2937 prev_states[newstate] = state;
2938 TRIE_LIST_PUSH( state, charid, newstate );
2943 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
2947 TRIE_HANDLE_WORD(state);
2949 } /* end second pass */
2951 /* next alloc is the NEXT state to be allocated */
2952 trie->statecount = next_alloc;
2953 trie->states = (reg_trie_state *)
2954 PerlMemShared_realloc( trie->states,
2956 * sizeof(reg_trie_state) );
2958 /* and now dump it out before we compress it */
2959 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2960 revcharmap, next_alloc,
2964 trie->trans = (reg_trie_trans *)
2965 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2972 for( state=1 ; state < next_alloc ; state ++ ) {
2976 DEBUG_TRIE_COMPILE_MORE_r(
2977 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
2981 if (trie->states[state].trans.list) {
2982 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2986 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2987 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2988 if ( forid < minid ) {
2990 } else if ( forid > maxid ) {
2994 if ( transcount < tp + maxid - minid + 1) {
2996 trie->trans = (reg_trie_trans *)
2997 PerlMemShared_realloc( trie->trans,
2999 * sizeof(reg_trie_trans) );
3000 Zero( trie->trans + (transcount / 2),
3004 base = trie->uniquecharcount + tp - minid;
3005 if ( maxid == minid ) {
3007 for ( ; zp < tp ; zp++ ) {
3008 if ( ! trie->trans[ zp ].next ) {
3009 base = trie->uniquecharcount + zp - minid;
3010 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
3012 trie->trans[ zp ].check = state;
3018 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
3020 trie->trans[ tp ].check = state;
3025 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3026 const U32 tid = base
3027 - trie->uniquecharcount
3028 + TRIE_LIST_ITEM( state, idx ).forid;
3029 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
3031 trie->trans[ tid ].check = state;
3033 tp += ( maxid - minid + 1 );
3035 Safefree(trie->states[ state ].trans.list);
3038 DEBUG_TRIE_COMPILE_MORE_r(
3039 Perl_re_printf( aTHX_ " base: %d\n",base);
3042 trie->states[ state ].trans.base=base;
3044 trie->lasttrans = tp + 1;
3048 Second Pass -- Flat Table Representation.
3050 we dont use the 0 slot of either trans[] or states[] so we add 1 to
3051 each. We know that we will need Charcount+1 trans at most to store
3052 the data (one row per char at worst case) So we preallocate both
3053 structures assuming worst case.
3055 We then construct the trie using only the .next slots of the entry
3058 We use the .check field of the first entry of the node temporarily
3059 to make compression both faster and easier by keeping track of how
3060 many non zero fields are in the node.
3062 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
3065 There are two terms at use here: state as a TRIE_NODEIDX() which is
3066 a number representing the first entry of the node, and state as a
3067 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
3068 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
3069 if there are 2 entrys per node. eg:
3077 The table is internally in the right hand, idx form. However as we
3078 also have to deal with the states array which is indexed by nodenum
3079 we have to use TRIE_NODENUM() to convert.
3082 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
3085 trie->trans = (reg_trie_trans *)
3086 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
3087 * trie->uniquecharcount + 1,
3088 sizeof(reg_trie_trans) );
3089 trie->states = (reg_trie_state *)
3090 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3091 sizeof(reg_trie_state) );
3092 next_alloc = trie->uniquecharcount + 1;
3095 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3097 regnode *noper = NEXTOPER( cur );
3099 U32 state = 1; /* required init */
3101 U16 charid = 0; /* sanity init */
3102 U32 accept_state = 0; /* sanity init */
3104 U32 wordlen = 0; /* required init */
3106 if (OP(noper) == NOTHING) {
3107 regnode *noper_next= regnext(noper);
3108 if (noper_next < tail)
3112 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
3113 const U8 *uc= (U8*)STRING(noper);
3114 const U8 *e= uc + STR_LEN(noper);
3116 for ( ; uc < e ; uc += len ) {
3121 charid = trie->charmap[ uvc ];
3123 SV* const * const svpp = hv_fetch( widecharmap,
3127 charid = svpp ? (U16)SvIV(*svpp) : 0;
3131 if ( !trie->trans[ state + charid ].next ) {
3132 trie->trans[ state + charid ].next = next_alloc;
3133 trie->trans[ state ].check++;
3134 prev_states[TRIE_NODENUM(next_alloc)]
3135 = TRIE_NODENUM(state);
3136 next_alloc += trie->uniquecharcount;
3138 state = trie->trans[ state + charid ].next;
3140 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3142 /* charid is now 0 if we dont know the char read, or
3143 * nonzero if we do */
3146 accept_state = TRIE_NODENUM( state );
3147 TRIE_HANDLE_WORD(accept_state);
3149 } /* end second pass */
3151 /* and now dump it out before we compress it */
3152 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3154 next_alloc, depth+1));
3158 * Inplace compress the table.*
3160 For sparse data sets the table constructed by the trie algorithm will
3161 be mostly 0/FAIL transitions or to put it another way mostly empty.
3162 (Note that leaf nodes will not contain any transitions.)
3164 This algorithm compresses the tables by eliminating most such
3165 transitions, at the cost of a modest bit of extra work during lookup:
3167 - Each states[] entry contains a .base field which indicates the
3168 index in the state[] array wheres its transition data is stored.
3170 - If .base is 0 there are no valid transitions from that node.
3172 - If .base is nonzero then charid is added to it to find an entry in
3175 -If trans[states[state].base+charid].check!=state then the
3176 transition is taken to be a 0/Fail transition. Thus if there are fail
3177 transitions at the front of the node then the .base offset will point
3178 somewhere inside the previous nodes data (or maybe even into a node
3179 even earlier), but the .check field determines if the transition is
3183 The following process inplace converts the table to the compressed
3184 table: We first do not compress the root node 1,and mark all its
3185 .check pointers as 1 and set its .base pointer as 1 as well. This
3186 allows us to do a DFA construction from the compressed table later,
3187 and ensures that any .base pointers we calculate later are greater
3190 - We set 'pos' to indicate the first entry of the second node.
3192 - We then iterate over the columns of the node, finding the first and
3193 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3194 and set the .check pointers accordingly, and advance pos
3195 appropriately and repreat for the next node. Note that when we copy
3196 the next pointers we have to convert them from the original
3197 NODEIDX form to NODENUM form as the former is not valid post
3200 - If a node has no transitions used we mark its base as 0 and do not
3201 advance the pos pointer.
3203 - If a node only has one transition we use a second pointer into the
3204 structure to fill in allocated fail transitions from other states.
3205 This pointer is independent of the main pointer and scans forward
3206 looking for null transitions that are allocated to a state. When it
3207 finds one it writes the single transition into the "hole". If the
3208 pointer doesnt find one the single transition is appended as normal.
3210 - Once compressed we can Renew/realloc the structures to release the
3213 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3214 specifically Fig 3.47 and the associated pseudocode.
3218 const U32 laststate = TRIE_NODENUM( next_alloc );
3221 trie->statecount = laststate;
3223 for ( state = 1 ; state < laststate ; state++ ) {
3225 const U32 stateidx = TRIE_NODEIDX( state );
3226 const U32 o_used = trie->trans[ stateidx ].check;
3227 U32 used = trie->trans[ stateidx ].check;
3228 trie->trans[ stateidx ].check = 0;
3231 used && charid < trie->uniquecharcount;
3234 if ( flag || trie->trans[ stateidx + charid ].next ) {
3235 if ( trie->trans[ stateidx + charid ].next ) {
3237 for ( ; zp < pos ; zp++ ) {
3238 if ( ! trie->trans[ zp ].next ) {
3242 trie->states[ state ].trans.base
3244 + trie->uniquecharcount
3246 trie->trans[ zp ].next
3247 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3249 trie->trans[ zp ].check = state;
3250 if ( ++zp > pos ) pos = zp;
3257 trie->states[ state ].trans.base
3258 = pos + trie->uniquecharcount - charid ;
3260 trie->trans[ pos ].next
3261 = SAFE_TRIE_NODENUM(
3262 trie->trans[ stateidx + charid ].next );
3263 trie->trans[ pos ].check = state;
3268 trie->lasttrans = pos + 1;
3269 trie->states = (reg_trie_state *)
3270 PerlMemShared_realloc( trie->states, laststate
3271 * sizeof(reg_trie_state) );
3272 DEBUG_TRIE_COMPILE_MORE_r(
3273 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
3275 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3279 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3282 } /* end table compress */
3284 DEBUG_TRIE_COMPILE_MORE_r(
3285 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
3287 (UV)trie->statecount,
3288 (UV)trie->lasttrans)
3290 /* resize the trans array to remove unused space */
3291 trie->trans = (reg_trie_trans *)
3292 PerlMemShared_realloc( trie->trans, trie->lasttrans
3293 * sizeof(reg_trie_trans) );
3295 { /* Modify the program and insert the new TRIE node */
3296 U8 nodetype =(U8)(flags & 0xFF);
3300 regnode *optimize = NULL;
3301 #ifdef RE_TRACK_PATTERN_OFFSETS
3304 U32 mjd_nodelen = 0;
3305 #endif /* RE_TRACK_PATTERN_OFFSETS */
3306 #endif /* DEBUGGING */
3308 This means we convert either the first branch or the first Exact,
3309 depending on whether the thing following (in 'last') is a branch
3310 or not and whther first is the startbranch (ie is it a sub part of
3311 the alternation or is it the whole thing.)
3312 Assuming its a sub part we convert the EXACT otherwise we convert
3313 the whole branch sequence, including the first.
3315 /* Find the node we are going to overwrite */
3316 if ( first != startbranch || OP( last ) == BRANCH ) {
3317 /* branch sub-chain */
3318 NEXT_OFF( first ) = (U16)(last - first);
3319 #ifdef RE_TRACK_PATTERN_OFFSETS
3321 mjd_offset= Node_Offset((convert));
3322 mjd_nodelen= Node_Length((convert));
3325 /* whole branch chain */
3327 #ifdef RE_TRACK_PATTERN_OFFSETS
3330 const regnode *nop = NEXTOPER( convert );
3331 mjd_offset= Node_Offset((nop));
3332 mjd_nodelen= Node_Length((nop));
3336 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
3338 (UV)mjd_offset, (UV)mjd_nodelen)
3341 /* But first we check to see if there is a common prefix we can
3342 split out as an EXACT and put in front of the TRIE node. */
3343 trie->startstate= 1;
3344 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3345 /* we want to find the first state that has more than
3346 * one transition, if that state is not the first state
3347 * then we have a common prefix which we can remove.
3350 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3352 I32 first_ofs = -1; /* keeps track of the ofs of the first
3353 transition, -1 means none */
3355 const U32 base = trie->states[ state ].trans.base;
3357 /* does this state terminate an alternation? */
3358 if ( trie->states[state].wordnum )
3361 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3362 if ( ( base + ofs >= trie->uniquecharcount ) &&
3363 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3364 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3366 if ( ++count > 1 ) {
3367 /* we have more than one transition */
3370 /* if this is the first state there is no common prefix
3371 * to extract, so we can exit */
3372 if ( state == 1 ) break;
3373 tmp = av_fetch( revcharmap, ofs, 0);
3374 ch = (U8*)SvPV_nolen_const( *tmp );
3376 /* if we are on count 2 then we need to initialize the
3377 * bitmap, and store the previous char if there was one
3380 /* clear the bitmap */
3381 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3383 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
3386 if (first_ofs >= 0) {
3387 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
3388 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3390 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
3392 Perl_re_printf( aTHX_ "%s", (char*)ch)
3396 /* store the current firstchar in the bitmap */
3397 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
3398 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
3404 /* This state has only one transition, its transition is part
3405 * of a common prefix - we need to concatenate the char it
3406 * represents to what we have so far. */
3407 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
3409 char *ch = SvPV( *tmp, len );
3411 SV *sv=sv_newmortal();
3412 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
3414 (UV)state, (UV)first_ofs,
3415 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3416 PL_colors[0], PL_colors[1],
3417 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3418 PERL_PV_ESCAPE_FIRSTCHAR
3423 OP( convert ) = nodetype;
3424 str=STRING(convert);
3427 STR_LEN(convert) += len;
3433 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
3438 trie->prefixlen = (state-1);
3440 regnode *n = convert+NODE_SZ_STR(convert);
3441 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3442 trie->startstate = state;
3443 trie->minlen -= (state - 1);
3444 trie->maxlen -= (state - 1);
3446 /* At least the UNICOS C compiler choked on this
3447 * being argument to DEBUG_r(), so let's just have
3450 #ifdef PERL_EXT_RE_BUILD
3456 regnode *fix = convert;
3457 U32 word = trie->wordcount;
3459 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3460 while( ++fix < n ) {
3461 Set_Node_Offset_Length(fix, 0, 0);
3464 SV ** const tmp = av_fetch( trie_words, word, 0 );
3466 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3467 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3469 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3477 NEXT_OFF(convert) = (U16)(tail - convert);
3478 DEBUG_r(optimize= n);
3484 if ( trie->maxlen ) {
3485 NEXT_OFF( convert ) = (U16)(tail - convert);
3486 ARG_SET( convert, data_slot );
3487 /* Store the offset to the first unabsorbed branch in
3488 jump[0], which is otherwise unused by the jump logic.
3489 We use this when dumping a trie and during optimisation. */
3491 trie->jump[0] = (U16)(nextbranch - convert);
3493 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3494 * and there is a bitmap
3495 * and the first "jump target" node we found leaves enough room
3496 * then convert the TRIE node into a TRIEC node, with the bitmap
3497 * embedded inline in the opcode - this is hypothetically faster.
3499 if ( !trie->states[trie->startstate].wordnum
3501 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3503 OP( convert ) = TRIEC;
3504 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3505 PerlMemShared_free(trie->bitmap);
3508 OP( convert ) = TRIE;
3510 /* store the type in the flags */
3511 convert->flags = nodetype;
3515 + regarglen[ OP( convert ) ];
3517 /* XXX We really should free up the resource in trie now,
3518 as we won't use them - (which resources?) dmq */
3520 /* needed for dumping*/
3521 DEBUG_r(if (optimize) {
3522 regnode *opt = convert;
3524 while ( ++opt < optimize) {
3525 Set_Node_Offset_Length(opt,0,0);
3528 Try to clean up some of the debris left after the
3531 while( optimize < jumper ) {
3532 mjd_nodelen += Node_Length((optimize));
3533 OP( optimize ) = OPTIMIZED;
3534 Set_Node_Offset_Length(optimize,0,0);
3537 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
3539 } /* end node insert */
3541 /* Finish populating the prev field of the wordinfo array. Walk back
3542 * from each accept state until we find another accept state, and if
3543 * so, point the first word's .prev field at the second word. If the
3544 * second already has a .prev field set, stop now. This will be the
3545 * case either if we've already processed that word's accept state,
3546 * or that state had multiple words, and the overspill words were
3547 * already linked up earlier.
3554 for (word=1; word <= trie->wordcount; word++) {
3556 if (trie->wordinfo[word].prev)
3558 state = trie->wordinfo[word].accept;
3560 state = prev_states[state];
3563 prev = trie->states[state].wordnum;
3567 trie->wordinfo[word].prev = prev;
3569 Safefree(prev_states);
3573 /* and now dump out the compressed format */
3574 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3576 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3578 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3579 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3581 SvREFCNT_dec_NN(revcharmap);
3585 : trie->startstate>1
3591 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3593 /* The Trie is constructed and compressed now so we can build a fail array if
3596 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3598 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3602 We find the fail state for each state in the trie, this state is the longest
3603 proper suffix of the current state's 'word' that is also a proper prefix of
3604 another word in our trie. State 1 represents the word '' and is thus the
3605 default fail state. This allows the DFA not to have to restart after its
3606 tried and failed a word at a given point, it simply continues as though it
3607 had been matching the other word in the first place.
3609 'abcdgu'=~/abcdefg|cdgu/
3610 When we get to 'd' we are still matching the first word, we would encounter
3611 'g' which would fail, which would bring us to the state representing 'd' in
3612 the second word where we would try 'g' and succeed, proceeding to match
3615 /* add a fail transition */
3616 const U32 trie_offset = ARG(source);
3617 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3619 const U32 ucharcount = trie->uniquecharcount;
3620 const U32 numstates = trie->statecount;
3621 const U32 ubound = trie->lasttrans + ucharcount;
3625 U32 base = trie->states[ 1 ].trans.base;
3628 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3630 GET_RE_DEBUG_FLAGS_DECL;
3632 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3633 PERL_UNUSED_CONTEXT;
3635 PERL_UNUSED_ARG(depth);
3638 if ( OP(source) == TRIE ) {
3639 struct regnode_1 *op = (struct regnode_1 *)
3640 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3641 StructCopy(source,op,struct regnode_1);
3642 stclass = (regnode *)op;
3644 struct regnode_charclass *op = (struct regnode_charclass *)
3645 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3646 StructCopy(source,op,struct regnode_charclass);
3647 stclass = (regnode *)op;
3649 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3651 ARG_SET( stclass, data_slot );
3652 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3653 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3654 aho->trie=trie_offset;
3655 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3656 Copy( trie->states, aho->states, numstates, reg_trie_state );
3657 Newx( q, numstates, U32);
3658 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3661 /* initialize fail[0..1] to be 1 so that we always have
3662 a valid final fail state */
3663 fail[ 0 ] = fail[ 1 ] = 1;
3665 for ( charid = 0; charid < ucharcount ; charid++ ) {
3666 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3668 q[ q_write ] = newstate;
3669 /* set to point at the root */
3670 fail[ q[ q_write++ ] ]=1;
3673 while ( q_read < q_write) {
3674 const U32 cur = q[ q_read++ % numstates ];
3675 base = trie->states[ cur ].trans.base;
3677 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3678 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3680 U32 fail_state = cur;
3683 fail_state = fail[ fail_state ];
3684 fail_base = aho->states[ fail_state ].trans.base;
3685 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3687 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3688 fail[ ch_state ] = fail_state;
3689 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3691 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3693 q[ q_write++ % numstates] = ch_state;
3697 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3698 when we fail in state 1, this allows us to use the
3699 charclass scan to find a valid start char. This is based on the principle
3700 that theres a good chance the string being searched contains lots of stuff
3701 that cant be a start char.
3703 fail[ 0 ] = fail[ 1 ] = 0;
3704 DEBUG_TRIE_COMPILE_r({
3705 Perl_re_indentf( aTHX_ "Stclass Failtable (%" UVuf " states): 0",
3706 depth, (UV)numstates
3708 for( q_read=1; q_read<numstates; q_read++ ) {
3709 Perl_re_printf( aTHX_ ", %" UVuf, (UV)fail[q_read]);
3711 Perl_re_printf( aTHX_ "\n");
3714 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3719 /* The below joins as many adjacent EXACTish nodes as possible into a single
3720 * one. The regop may be changed if the node(s) contain certain sequences that
3721 * require special handling. The joining is only done if:
3722 * 1) there is room in the current conglomerated node to entirely contain the
3724 * 2) they are the exact same node type
3726 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3727 * these get optimized out
3729 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3730 * as possible, even if that means splitting an existing node so that its first
3731 * part is moved to the preceeding node. This would maximise the efficiency of
3732 * memEQ during matching.
3734 * If a node is to match under /i (folded), the number of characters it matches
3735 * can be different than its character length if it contains a multi-character
3736 * fold. *min_subtract is set to the total delta number of characters of the
3739 * And *unfolded_multi_char is set to indicate whether or not the node contains
3740 * an unfolded multi-char fold. This happens when it won't be known until
3741 * runtime whether the fold is valid or not; namely
3742 * 1) for EXACTF nodes that contain LATIN SMALL LETTER SHARP S, as only if the
3743 * target string being matched against turns out to be UTF-8 is that fold
3745 * 2) for EXACTFL nodes whose folding rules depend on the locale in force at
3747 * (Multi-char folds whose components are all above the Latin1 range are not
3748 * run-time locale dependent, and have already been folded by the time this
3749 * function is called.)
3751 * This is as good a place as any to discuss the design of handling these
3752 * multi-character fold sequences. It's been wrong in Perl for a very long
3753 * time. There are three code points in Unicode whose multi-character folds
3754 * were long ago discovered to mess things up. The previous designs for
3755 * dealing with these involved assigning a special node for them. This
3756 * approach doesn't always work, as evidenced by this example:
3757 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3758 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3759 * would match just the \xDF, it won't be able to handle the case where a
3760 * successful match would have to cross the node's boundary. The new approach
3761 * that hopefully generally solves the problem generates an EXACTFU_SS node
3762 * that is "sss" in this case.
3764 * It turns out that there are problems with all multi-character folds, and not
3765 * just these three. Now the code is general, for all such cases. The
3766 * approach taken is:
3767 * 1) This routine examines each EXACTFish node that could contain multi-
3768 * character folded sequences. Since a single character can fold into
3769 * such a sequence, the minimum match length for this node is less than
3770 * the number of characters in the node. This routine returns in
3771 * *min_subtract how many characters to subtract from the the actual
3772 * length of the string to get a real minimum match length; it is 0 if
3773 * there are no multi-char foldeds. This delta is used by the caller to
3774 * adjust the min length of the match, and the delta between min and max,
3775 * so that the optimizer doesn't reject these possibilities based on size
3777 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3778 * is used for an EXACTFU node that contains at least one "ss" sequence in
3779 * it. For non-UTF-8 patterns and strings, this is the only case where
3780 * there is a possible fold length change. That means that a regular
3781 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3782 * with length changes, and so can be processed faster. regexec.c takes
3783 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3784 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3785 * known until runtime). This saves effort in regex matching. However,
3786 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3787 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3788 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3789 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3790 * possibilities for the non-UTF8 patterns are quite simple, except for
3791 * the sharp s. All the ones that don't involve a UTF-8 target string are
3792 * members of a fold-pair, and arrays are set up for all of them so that
3793 * the other member of the pair can be found quickly. Code elsewhere in
3794 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3795 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3796 * described in the next item.
3797 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3798 * validity of the fold won't be known until runtime, and so must remain
3799 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFAA
3800 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3801 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3802 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3803 * The reason this is a problem is that the optimizer part of regexec.c
3804 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3805 * that a character in the pattern corresponds to at most a single
3806 * character in the target string. (And I do mean character, and not byte
3807 * here, unlike other parts of the documentation that have never been
3808 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3809 * EXACTFL nodes can match the two character string 'ss'; in EXACTFAA
3810 * nodes it can match "\x{17F}\x{17F}". These, along with other ones in
3811 * EXACTFL nodes, violate the assumption, and they are the only instances
3812 * where it is violated. I'm reluctant to try to change the assumption,
3813 * as the code involved is impenetrable to me (khw), so instead the code
3814 * here punts. This routine examines EXACTFL nodes, and (when the pattern
3815 * isn't UTF-8) EXACTF and EXACTFAA for such unfolded folds, and returns a
3816 * boolean indicating whether or not the node contains such a fold. When
3817 * it is true, the caller sets a flag that later causes the optimizer in
3818 * this file to not set values for the floating and fixed string lengths,
3819 * and thus avoids the optimizer code in regexec.c that makes the invalid
3820 * assumption. Thus, there is no optimization based on string lengths for
3821 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3822 * EXACTF and EXACTFAA nodes that contain the sharp s. (The reason the
3823 * assumption is wrong only in these cases is that all other non-UTF-8
3824 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3825 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3826 * EXACTF nodes because we don't know at compile time if it actually
3827 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3828 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3829 * always matches; and EXACTFAA where it never does. In an EXACTFAA node
3830 * in a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3831 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3832 * string would require the pattern to be forced into UTF-8, the overhead
3833 * of which we want to avoid. Similarly the unfolded multi-char folds in
3834 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3837 * Similarly, the code that generates tries doesn't currently handle
3838 * not-already-folded multi-char folds, and it looks like a pain to change
3839 * that. Therefore, trie generation of EXACTFAA nodes with the sharp s
3840 * doesn't work. Instead, such an EXACTFAA is turned into a new regnode,
3841 * EXACTFAA_NO_TRIE, which the trie code knows not to handle. Most people
3842 * using /iaa matching will be doing so almost entirely with ASCII
3843 * strings, so this should rarely be encountered in practice */
3845 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3846 if (PL_regkind[OP(scan)] == EXACT) \
3847 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3850 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3851 UV *min_subtract, bool *unfolded_multi_char,
3852 U32 flags,regnode *val, U32 depth)
3854 /* Merge several consecutive EXACTish nodes into one. */
3855 regnode *n = regnext(scan);
3857 regnode *next = scan + NODE_SZ_STR(scan);
3861 regnode *stop = scan;
3862 GET_RE_DEBUG_FLAGS_DECL;
3864 PERL_UNUSED_ARG(depth);
3867 PERL_ARGS_ASSERT_JOIN_EXACT;
3868 #ifndef EXPERIMENTAL_INPLACESCAN
3869 PERL_UNUSED_ARG(flags);
3870 PERL_UNUSED_ARG(val);
3872 DEBUG_PEEP("join", scan, depth, 0);
3874 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3875 * EXACT ones that are mergeable to the current one. */
3877 && (PL_regkind[OP(n)] == NOTHING
3878 || (stringok && OP(n) == OP(scan)))
3880 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3883 if (OP(n) == TAIL || n > next)
3885 if (PL_regkind[OP(n)] == NOTHING) {
3886 DEBUG_PEEP("skip:", n, depth, 0);
3887 NEXT_OFF(scan) += NEXT_OFF(n);
3888 next = n + NODE_STEP_REGNODE;
3895 else if (stringok) {
3896 const unsigned int oldl = STR_LEN(scan);
3897 regnode * const nnext = regnext(n);
3899 /* XXX I (khw) kind of doubt that this works on platforms (should
3900 * Perl ever run on one) where U8_MAX is above 255 because of lots
3901 * of other assumptions */
3902 /* Don't join if the sum can't fit into a single node */
3903 if (oldl + STR_LEN(n) > U8_MAX)
3906 DEBUG_PEEP("merg", n, depth, 0);
3909 NEXT_OFF(scan) += NEXT_OFF(n);
3910 STR_LEN(scan) += STR_LEN(n);
3911 next = n + NODE_SZ_STR(n);
3912 /* Now we can overwrite *n : */
3913 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3921 #ifdef EXPERIMENTAL_INPLACESCAN
3922 if (flags && !NEXT_OFF(n)) {
3923 DEBUG_PEEP("atch", val, depth, 0);
3924 if (reg_off_by_arg[OP(n)]) {
3925 ARG_SET(n, val - n);
3928 NEXT_OFF(n) = val - n;
3936 *unfolded_multi_char = FALSE;
3938 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3939 * can now analyze for sequences of problematic code points. (Prior to
3940 * this final joining, sequences could have been split over boundaries, and
3941 * hence missed). The sequences only happen in folding, hence for any
3942 * non-EXACT EXACTish node */
3943 if (OP(scan) != EXACT && OP(scan) != EXACTL) {
3944 U8* s0 = (U8*) STRING(scan);
3946 U8* s_end = s0 + STR_LEN(scan);
3948 int total_count_delta = 0; /* Total delta number of characters that
3949 multi-char folds expand to */
3951 /* One pass is made over the node's string looking for all the
3952 * possibilities. To avoid some tests in the loop, there are two main
3953 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3958 if (OP(scan) == EXACTFL) {
3961 /* An EXACTFL node would already have been changed to another
3962 * node type unless there is at least one character in it that
3963 * is problematic; likely a character whose fold definition
3964 * won't be known until runtime, and so has yet to be folded.
3965 * For all but the UTF-8 locale, folds are 1-1 in length, but
3966 * to handle the UTF-8 case, we need to create a temporary
3967 * folded copy using UTF-8 locale rules in order to analyze it.
3968 * This is because our macros that look to see if a sequence is
3969 * a multi-char fold assume everything is folded (otherwise the
3970 * tests in those macros would be too complicated and slow).
3971 * Note that here, the non-problematic folds will have already
3972 * been done, so we can just copy such characters. We actually
3973 * don't completely fold the EXACTFL string. We skip the
3974 * unfolded multi-char folds, as that would just create work
3975 * below to figure out the size they already are */
3977 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3980 STRLEN s_len = UTF8SKIP(s);
3981 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3982 Copy(s, d, s_len, U8);
3985 else if (is_FOLDS_TO_MULTI_utf8(s)) {
3986 *unfolded_multi_char = TRUE;
3987 Copy(s, d, s_len, U8);
3990 else if (isASCII(*s)) {
3991 *(d++) = toFOLD(*s);
3995 _toFOLD_utf8_flags(s, s_end, d, &len, FOLD_FLAGS_FULL);
4001 /* Point the remainder of the routine to look at our temporary
4005 } /* End of creating folded copy of EXACTFL string */
4007 /* Examine the string for a multi-character fold sequence. UTF-8
4008 * patterns have all characters pre-folded by the time this code is
4010 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
4011 length sequence we are looking for is 2 */
4013 int count = 0; /* How many characters in a multi-char fold */
4014 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
4015 if (! len) { /* Not a multi-char fold: get next char */
4020 /* Nodes with 'ss' require special handling, except for
4021 * EXACTFAA-ish for which there is no multi-char fold to this */
4022 if (len == 2 && *s == 's' && *(s+1) == 's'
4023 && OP(scan) != EXACTFAA
4024 && OP(scan) != EXACTFAA_NO_TRIE)
4027 if (OP(scan) != EXACTFL) {
4028 OP(scan) = EXACTFU_SS;