5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
10 /* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
12 * a regular expression.
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
20 /* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
24 /* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
29 /* The names of the functions have been changed from regcomp and
30 * regexec to pregcomp and pregexec in order to avoid conflicts
31 * with the POSIX routines of the same names.
34 #ifdef PERL_EXT_RE_BUILD
39 * pregcomp and pregexec -- regsub and regerror are not used in perl
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
59 **** Alterations to Henry's code are...
61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
74 #define PERL_IN_REGCOMP_C
77 #ifndef PERL_IN_XSUB_RE
82 #ifdef PERL_IN_XSUB_RE
84 EXTERN_C const struct regexp_engine my_reg_engine;
89 #include "dquote_inline.h"
90 #include "invlist_inline.h"
91 #include "unicode_constants.h"
93 #define HAS_NONLATIN1_FOLD_CLOSURE(i) \
94 _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
95 #define HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(i) \
96 _HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
97 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
98 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
101 #define STATIC static
104 /* this is a chain of data about sub patterns we are processing that
105 need to be handled separately/specially in study_chunk. Its so
106 we can simulate recursion without losing state. */
108 typedef struct scan_frame {
109 regnode *last_regnode; /* last node to process in this frame */
110 regnode *next_regnode; /* next node to process when last is reached */
111 U32 prev_recursed_depth;
112 I32 stopparen; /* what stopparen do we use */
114 struct scan_frame *this_prev_frame; /* this previous frame */
115 struct scan_frame *prev_frame; /* previous frame */
116 struct scan_frame *next_frame; /* next frame */
119 /* Certain characters are output as a sequence with the first being a
121 #define isBACKSLASHED_PUNCT(c) strchr("-[]\\^", c)
124 struct RExC_state_t {
125 U32 flags; /* RXf_* are we folding, multilining? */
126 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
127 char *precomp; /* uncompiled string. */
128 char *precomp_end; /* pointer to end of uncompiled string. */
129 REGEXP *rx_sv; /* The SV that is the regexp. */
130 regexp *rx; /* perl core regexp structure */
131 regexp_internal *rxi; /* internal data for regexp object
133 char *start; /* Start of input for compile */
134 char *end; /* End of input for compile */
135 char *parse; /* Input-scan pointer. */
136 char *copy_start; /* start of copy of input within
137 constructed parse string */
138 char *copy_start_in_input; /* Position in input string
139 corresponding to copy_start */
140 SSize_t whilem_seen; /* number of WHILEM in this expr */
141 regnode *emit_start; /* Start of emitted-code area */
142 regnode *emit_bound; /* First regnode outside of the
144 regnode_offset emit; /* Code-emit pointer */
145 I32 naughty; /* How bad is this pattern? */
146 I32 sawback; /* Did we see \1, ...? */
148 SSize_t size; /* Number of regnode equivalents in
150 I32 npar; /* Capture buffer count, (OPEN) plus
151 one. ("par" 0 is the whole
153 I32 nestroot; /* root parens we are in - used by
157 regnode_offset *open_parens; /* offsets to open parens */
158 regnode_offset *close_parens; /* offsets to close parens */
159 regnode *end_op; /* END node in program */
160 I32 utf8; /* whether the pattern is utf8 or not */
161 I32 orig_utf8; /* whether the pattern was originally in utf8 */
162 /* XXX use this for future optimisation of case
163 * where pattern must be upgraded to utf8. */
164 I32 uni_semantics; /* If a d charset modifier should use unicode
165 rules, even if the pattern is not in
167 HV *paren_names; /* Paren names */
169 regnode **recurse; /* Recurse regops */
170 I32 recurse_count; /* Number of recurse regops we have generated */
171 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
173 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
176 I32 override_recoding;
178 I32 recode_x_to_native;
180 I32 in_multi_char_class;
181 struct reg_code_blocks *code_blocks;/* positions of literal (?{})
183 int code_index; /* next code_blocks[] slot */
184 SSize_t maxlen; /* mininum possible number of chars in string to match */
185 scan_frame *frame_head;
186 scan_frame *frame_last;
189 #ifdef ADD_TO_REGEXEC
190 char *starttry; /* -Dr: where regtry was called. */
191 #define RExC_starttry (pRExC_state->starttry)
193 SV *runtime_code_qr; /* qr with the runtime code blocks */
195 const char *lastparse;
197 AV *paren_name_list; /* idx -> name */
198 U32 study_chunk_recursed_count;
202 #define RExC_lastparse (pRExC_state->lastparse)
203 #define RExC_lastnum (pRExC_state->lastnum)
204 #define RExC_paren_name_list (pRExC_state->paren_name_list)
205 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
206 #define RExC_mysv (pRExC_state->mysv1)
207 #define RExC_mysv1 (pRExC_state->mysv1)
208 #define RExC_mysv2 (pRExC_state->mysv2)
211 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_copy_start_in_input (pRExC_state->copy_start_in_input)
222 #define RExC_copy_start_in_constructed (pRExC_state->copy_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 (RExC_rxi->u.offsets) /* I am not like the
246 #define RExC_emit (pRExC_state->emit)
247 #define RExC_pass1 (pRExC_state->pass1)
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_PARSE 0x20 /* Need to redo the parse */
317 #define NEED_UTF8 0x40 /* In conjunction with RESTART_PARSE, 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_PARSE|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_PARSE; \
357 return restart_retval; \
362 /* Executes a return statement with the value 'X', if 'flags' contains any of
363 * 'RESTART_PARSE', '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_PARSE|NEED_UTF8|(extra))) { \
368 *(flagp) = (flags) & (RESTART_PARSE|NEED_UTF8|(extra)); \
373 #define RETURN_FAIL_ON_RESTART_OR_FLAGS(flags,flagp,extra) \
374 RETURN_X_ON_RESTART_OR_FLAGS(0,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_FAIL_ON_RESTART_FLAGP_OR_FLAGS(flagp,extra) \
381 if (*(flagp) & (RESTART_PARSE|(extra))) return 0
383 #define MUST_RESTART(flags) ((flags) & (RESTART_PARSE))
385 #define RETURN_FAIL_ON_RESTART(flags,flagp) \
386 RETURN_X_ON_RESTART(0, flags,flagp)
387 #define RETURN_FAIL_ON_RESTART_FLAGP(flagp) \
388 RETURN_FAIL_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 completely control
610 * the input 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:
622 * |<------- identical ------>|
624 * Input: ---------------------------------------------------------------
625 * Constructed: ---------------------------------------------------
627 * |<------- identical ------>|
629 * sI..eI is the portion of the input pattern we are concerned with here.
630 * sC..EC is the constructed substitute parse string.
631 * sC..tC is constructed by us
632 * tC..eC is an exact duplicate of the portion of the input pattern tI..eI.
633 * In the diagram, these are vertically aligned.
634 * eC..EC is also constructed by us.
635 * xC is the position in the substitute parse string where we found a
637 * xI is the position in the original pattern corresponding to xC.
639 * We want to display a message showing the real input string. Thus we need to
640 * translate from xC to xI. We know that xC >= tC, since the portion of the
641 * string sC..tC has been constructed by us, and so shouldn't have errors. We
643 * xI = tI + (xC - tC)
645 * When the substitute parse is constructed, the code needs to set:
648 * RExC_copy_start_in_input (tI)
649 * RExC_copy_start_in_constructed (tC)
650 * and restore them when done.
652 * During normal processing of the input pattern, both
653 * 'RExC_copy_start_in_input' and 'RExC_copy_start_in_constructed' are set to
654 * sI, so that xC equals xI.
657 #define sI RExC_precomp
658 #define eI RExC_precomp_end
659 #define sC RExC_start
661 #define tI RExC_copy_start_in_input
662 #define tC RExC_copy_start_in_constructed
663 #define xI(xC) (tI + (xC - tC))
664 #define xI_offset(xC) (xI(xC) - sI)
666 #define REPORT_LOCATION_ARGS(xC) \
668 (xI(xC) > eI) /* Don't run off end */ \
669 ? eC - sC /* Length before the <--HERE */ \
670 : ((xI_offset(xC) >= 0) \
672 : (Perl_croak(aTHX_ "panic: %s: %d: negative offset: %" \
673 IVdf " trying to output message for " \
675 __FILE__, __LINE__, (IV) xI_offset(xC), \
676 ((int) (eC - sC)), sC), 0)), \
677 sI), /* The input pattern printed up to the <--HERE */ \
679 (xI(xC) > eI) ? 0 : eI - xI(xC), /* Length after <--HERE */ \
680 (xI(xC) > eI) ? eI : xI(xC)) /* pattern after <--HERE */
682 /* Used to point after bad bytes for an error message, but avoid skipping
683 * past a nul byte. */
684 #define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
687 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
688 * arg. Show regex, up to a maximum length. If it's too long, chop and add
691 #define _FAIL(code) STMT_START { \
692 const char *ellipses = ""; \
693 IV len = RExC_precomp_end - RExC_precomp; \
696 SAVEFREESV(RExC_rx_sv); \
697 if (len > RegexLengthToShowInErrorMessages) { \
698 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
699 len = RegexLengthToShowInErrorMessages - 10; \
705 #define FAIL(msg) _FAIL( \
706 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
707 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
709 #define FAIL2(msg,arg) _FAIL( \
710 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
711 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
714 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
716 #define Simple_vFAIL(m) STMT_START { \
717 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
718 m, REPORT_LOCATION_ARGS(RExC_parse)); \
722 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
724 #define vFAIL(m) STMT_START { \
726 SAVEFREESV(RExC_rx_sv); \
731 * Like Simple_vFAIL(), but accepts two arguments.
733 #define Simple_vFAIL2(m,a1) STMT_START { \
734 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
735 REPORT_LOCATION_ARGS(RExC_parse)); \
739 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
741 #define vFAIL2(m,a1) STMT_START { \
743 SAVEFREESV(RExC_rx_sv); \
744 Simple_vFAIL2(m, a1); \
749 * Like Simple_vFAIL(), but accepts three arguments.
751 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
752 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
753 REPORT_LOCATION_ARGS(RExC_parse)); \
757 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
759 #define vFAIL3(m,a1,a2) STMT_START { \
761 SAVEFREESV(RExC_rx_sv); \
762 Simple_vFAIL3(m, a1, a2); \
766 * Like Simple_vFAIL(), but accepts four arguments.
768 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
769 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
770 REPORT_LOCATION_ARGS(RExC_parse)); \
773 #define vFAIL4(m,a1,a2,a3) STMT_START { \
775 SAVEFREESV(RExC_rx_sv); \
776 Simple_vFAIL4(m, a1, a2, a3); \
779 /* A specialized version of vFAIL2 that works with UTF8f */
780 #define vFAIL2utf8f(m, a1) STMT_START { \
782 SAVEFREESV(RExC_rx_sv); \
783 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
784 REPORT_LOCATION_ARGS(RExC_parse)); \
787 #define vFAIL3utf8f(m, a1, a2) STMT_START { \
789 SAVEFREESV(RExC_rx_sv); \
790 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
791 REPORT_LOCATION_ARGS(RExC_parse)); \
794 /* This has an assert in it because of [perl #122671] Many warnings in
795 * regcomp.c can occur twice. If they get output in pass1 and later in that
796 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
797 * would get output again. So they should be output in pass2, and this
798 * assert makes sure new warnings follow that paradigm. */
799 #define _WARN_HELPER(loc, warns, code) \
801 __ASSERT_(PASS2) code; \
804 /* m is not necessarily a "literal string", in this macro */
805 #define reg_warn_non_literal_string(loc, m) \
806 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
807 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
808 "%s" REPORT_LOCATION, \
809 m, REPORT_LOCATION_ARGS(loc)))
811 #define ckWARNreg(loc,m) \
812 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
813 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
815 REPORT_LOCATION_ARGS(loc)))
817 #define vWARN(loc, m) \
818 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
819 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
821 REPORT_LOCATION_ARGS(loc))) \
823 #define vWARN_dep(loc, m) \
824 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
825 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
827 REPORT_LOCATION_ARGS(loc)))
829 #define ckWARNdep(loc,m) \
830 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
831 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
833 REPORT_LOCATION_ARGS(loc)))
835 #define ckWARNregdep(loc,m) \
836 _WARN_HELPER(loc, packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
837 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
840 REPORT_LOCATION_ARGS(loc)))
842 #define ckWARN2reg_d(loc,m, a1) \
843 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
844 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
846 a1, REPORT_LOCATION_ARGS(loc)))
848 #define ckWARN2reg(loc, m, a1) \
849 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
850 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
852 a1, REPORT_LOCATION_ARGS(loc)))
854 #define vWARN3(loc, m, a1, a2) \
855 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
856 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
858 a1, a2, REPORT_LOCATION_ARGS(loc)))
860 #define ckWARN3reg(loc, m, a1, a2) \
861 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
862 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
865 REPORT_LOCATION_ARGS(loc)))
867 #define vWARN4(loc, m, a1, a2, a3) \
868 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
869 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
872 REPORT_LOCATION_ARGS(loc)))
874 #define ckWARN4reg(loc, m, a1, a2, a3) \
875 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
876 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
879 REPORT_LOCATION_ARGS(loc)))
881 #define vWARN5(loc, m, a1, a2, a3, a4) \
882 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
883 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
886 REPORT_LOCATION_ARGS(loc)))
888 /* Convert between a pointer to a node and its offset from the beginning of the
890 #define REGNODE_p(offset) (RExC_emit_start + (offset))
891 #define REGNODE_OFFSET(node) ((node) - RExC_emit_start)
893 /* Macros for recording node offsets. 20001227 mjd@plover.com
894 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
895 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
896 * Element 0 holds the number n.
897 * Position is 1 indexed.
899 #ifndef RE_TRACK_PATTERN_OFFSETS
900 #define Set_Node_Offset_To_R(offset,byte)
901 #define Set_Node_Offset(node,byte)
902 #define Set_Cur_Node_Offset
903 #define Set_Node_Length_To_R(node,len)
904 #define Set_Node_Length(node,len)
905 #define Set_Node_Cur_Length(node,start)
906 #define Node_Offset(n)
907 #define Node_Length(n)
908 #define Set_Node_Offset_Length(node,offset,len)
909 #define ProgLen(ri) ri->u.proglen
910 #define SetProgLen(ri,x) ri->u.proglen = x
912 #define ProgLen(ri) ri->u.offsets[0]
913 #define SetProgLen(ri,x) ri->u.offsets[0] = x
914 #define Set_Node_Offset_To_R(offset,byte) STMT_START { \
916 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
917 __LINE__, (int)(offset), (int)(byte))); \
919 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
922 RExC_offsets[2*(offset)-1] = (byte); \
927 #define Set_Node_Offset(node,byte) \
928 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (byte)-RExC_start)
929 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
931 #define Set_Node_Length_To_R(node,len) STMT_START { \
933 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
934 __LINE__, (int)(node), (int)(len))); \
936 Perl_croak(aTHX_ "value of node is %d in Length macro", \
939 RExC_offsets[2*(node)] = (len); \
944 #define Set_Node_Length(node,len) \
945 Set_Node_Length_To_R(REGNODE_OFFSET(node), len)
946 #define Set_Node_Cur_Length(node, start) \
947 Set_Node_Length(node, RExC_parse - start)
949 /* Get offsets and lengths */
950 #define Node_Offset(n) (RExC_offsets[2*(REGNODE_OFFSET(n))-1])
951 #define Node_Length(n) (RExC_offsets[2*(REGNODE_OFFSET(n))])
953 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
954 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (offset)); \
955 Set_Node_Length_To_R(REGNODE_OFFSET(node), (len)); \
959 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
960 #define EXPERIMENTAL_INPLACESCAN
961 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
965 Perl_re_printf(pTHX_ const char *fmt, ...)
969 PerlIO *f= Perl_debug_log;
970 PERL_ARGS_ASSERT_RE_PRINTF;
972 result = PerlIO_vprintf(f, fmt, ap);
978 Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
982 PerlIO *f= Perl_debug_log;
983 PERL_ARGS_ASSERT_RE_INDENTF;
985 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
986 result = PerlIO_vprintf(f, fmt, ap);
990 #endif /* DEBUGGING */
992 #define DEBUG_RExC_seen() \
993 DEBUG_OPTIMISE_MORE_r({ \
994 Perl_re_printf( aTHX_ "RExC_seen: "); \
996 if (RExC_seen & REG_ZERO_LEN_SEEN) \
997 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
999 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
1000 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
1002 if (RExC_seen & REG_GPOS_SEEN) \
1003 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
1005 if (RExC_seen & REG_RECURSE_SEEN) \
1006 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
1008 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
1009 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
1011 if (RExC_seen & REG_VERBARG_SEEN) \
1012 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
1014 if (RExC_seen & REG_CUTGROUP_SEEN) \
1015 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
1017 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
1018 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
1020 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
1021 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
1023 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
1024 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
1026 Perl_re_printf( aTHX_ "\n"); \
1029 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
1030 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
1035 S_debug_show_study_flags(pTHX_ U32 flags, const char *open_str,
1036 const char *close_str)
1041 Perl_re_printf( aTHX_ "%s", open_str);
1042 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_SEOL);
1043 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_MEOL);
1044 DEBUG_SHOW_STUDY_FLAG(flags, SF_IS_INF);
1045 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_PAR);
1046 DEBUG_SHOW_STUDY_FLAG(flags, SF_IN_PAR);
1047 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_EVAL);
1048 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_SUBSTR);
1049 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_AND);
1050 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_OR);
1051 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS);
1052 DEBUG_SHOW_STUDY_FLAG(flags, SCF_WHILEM_VISITED_POS);
1053 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_RESTUDY);
1054 DEBUG_SHOW_STUDY_FLAG(flags, SCF_SEEN_ACCEPT);
1055 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_DOING_RESTUDY);
1056 DEBUG_SHOW_STUDY_FLAG(flags, SCF_IN_DEFINE);
1057 Perl_re_printf( aTHX_ "%s", close_str);
1062 S_debug_studydata(pTHX_ const char *where, scan_data_t *data,
1063 U32 depth, int is_inf)
1065 GET_RE_DEBUG_FLAGS_DECL;
1067 DEBUG_OPTIMISE_MORE_r({
1070 Perl_re_indentf(aTHX_ "%s: Pos:%" IVdf "/%" IVdf " Flags: 0x%" UVXf,
1074 (IV)data->pos_delta,
1078 S_debug_show_study_flags(aTHX_ data->flags," [","]");
1080 Perl_re_printf( aTHX_
1081 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s",
1083 (IV)(data->last_closep ? *((data)->last_closep) : -1),
1084 is_inf ? "INF " : ""
1087 if (data->last_found) {
1089 Perl_re_printf(aTHX_
1090 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf,
1091 SvPVX_const(data->last_found),
1093 (IV)data->last_start_min,
1094 (IV)data->last_start_max
1097 for (i = 0; i < 2; i++) {
1098 Perl_re_printf(aTHX_
1099 " %s%s: '%s' @ %" IVdf "/%" IVdf,
1100 data->cur_is_floating == i ? "*" : "",
1101 i ? "Float" : "Fixed",
1102 SvPVX_const(data->substrs[i].str),
1103 (IV)data->substrs[i].min_offset,
1104 (IV)data->substrs[i].max_offset
1106 S_debug_show_study_flags(aTHX_ data->substrs[i].flags," [","]");
1110 Perl_re_printf( aTHX_ "\n");
1116 S_debug_peep(pTHX_ const char *str, const RExC_state_t *pRExC_state,
1117 regnode *scan, U32 depth, U32 flags)
1119 GET_RE_DEBUG_FLAGS_DECL;
1126 Next = regnext(scan);
1127 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
1128 Perl_re_indentf( aTHX_ "%s>%3d: %s (%d)",
1131 REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),
1132 Next ? (REG_NODE_NUM(Next)) : 0 );
1133 S_debug_show_study_flags(aTHX_ flags," [ ","]");
1134 Perl_re_printf( aTHX_ "\n");
1139 # define DEBUG_STUDYDATA(where, data, depth, is_inf) \
1140 S_debug_studydata(aTHX_ where, data, depth, is_inf)
1142 # define DEBUG_PEEP(str, scan, depth, flags) \
1143 S_debug_peep(aTHX_ str, pRExC_state, scan, depth, flags)
1146 # define DEBUG_STUDYDATA(where, data, depth, is_inf) NOOP
1147 # define DEBUG_PEEP(str, scan, depth, flags) NOOP
1151 /* =========================================================
1152 * BEGIN edit_distance stuff.
1154 * This calculates how many single character changes of any type are needed to
1155 * transform a string into another one. It is taken from version 3.1 of
1157 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1160 /* Our unsorted dictionary linked list. */
1161 /* Note we use UVs, not chars. */
1166 struct dictionary* next;
1168 typedef struct dictionary item;
1171 PERL_STATIC_INLINE item*
1172 push(UV key, item* curr)
1175 Newx(head, 1, item);
1183 PERL_STATIC_INLINE item*
1184 find(item* head, UV key)
1186 item* iterator = head;
1188 if (iterator->key == key){
1191 iterator = iterator->next;
1197 PERL_STATIC_INLINE item*
1198 uniquePush(item* head, UV key)
1200 item* iterator = head;
1203 if (iterator->key == key) {
1206 iterator = iterator->next;
1209 return push(key, head);
1212 PERL_STATIC_INLINE void
1213 dict_free(item* head)
1215 item* iterator = head;
1218 item* temp = iterator;
1219 iterator = iterator->next;
1226 /* End of Dictionary Stuff */
1228 /* All calculations/work are done here */
1230 S_edit_distance(const UV* src,
1232 const STRLEN x, /* length of src[] */
1233 const STRLEN y, /* length of tgt[] */
1234 const SSize_t maxDistance
1238 UV swapCount, swapScore, targetCharCount, i, j;
1240 UV score_ceil = x + y;
1242 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1244 /* intialize matrix start values */
1245 Newx(scores, ( (x + 2) * (y + 2)), UV);
1246 scores[0] = score_ceil;
1247 scores[1 * (y + 2) + 0] = score_ceil;
1248 scores[0 * (y + 2) + 1] = score_ceil;
1249 scores[1 * (y + 2) + 1] = 0;
1250 head = uniquePush(uniquePush(head, src[0]), tgt[0]);
1255 for (i=1;i<=x;i++) {
1257 head = uniquePush(head, src[i]);
1258 scores[(i+1) * (y + 2) + 1] = i;
1259 scores[(i+1) * (y + 2) + 0] = score_ceil;
1262 for (j=1;j<=y;j++) {
1265 head = uniquePush(head, tgt[j]);
1266 scores[1 * (y + 2) + (j + 1)] = j;
1267 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1270 targetCharCount = find(head, tgt[j-1])->value;
1271 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1273 if (src[i-1] != tgt[j-1]){
1274 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));
1278 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1282 find(head, src[i-1])->value = i;
1286 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1289 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1293 /* END of edit_distance() stuff
1294 * ========================================================= */
1296 /* is c a control character for which we have a mnemonic? */
1297 #define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
1300 S_cntrl_to_mnemonic(const U8 c)
1302 /* Returns the mnemonic string that represents character 'c', if one
1303 * exists; NULL otherwise. The only ones that exist for the purposes of
1304 * this routine are a few control characters */
1307 case '\a': return "\\a";
1308 case '\b': return "\\b";
1309 case ESC_NATIVE: return "\\e";
1310 case '\f': return "\\f";
1311 case '\n': return "\\n";
1312 case '\r': return "\\r";
1313 case '\t': return "\\t";
1319 /* Mark that we cannot extend a found fixed substring at this point.
1320 Update the longest found anchored substring or the longest found
1321 floating substrings if needed. */
1324 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1325 SSize_t *minlenp, int is_inf)
1327 const STRLEN l = CHR_SVLEN(data->last_found);
1328 SV * const longest_sv = data->substrs[data->cur_is_floating].str;
1329 const STRLEN old_l = CHR_SVLEN(longest_sv);
1330 GET_RE_DEBUG_FLAGS_DECL;
1332 PERL_ARGS_ASSERT_SCAN_COMMIT;
1334 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
1335 const U8 i = data->cur_is_floating;
1336 SvSetMagicSV(longest_sv, data->last_found);
1337 data->substrs[i].min_offset = l ? data->last_start_min : data->pos_min;
1340 data->substrs[0].max_offset = data->substrs[0].min_offset;
1342 data->substrs[1].max_offset = (l
1343 ? data->last_start_max
1344 : (data->pos_delta > SSize_t_MAX - data->pos_min
1346 : data->pos_min + data->pos_delta));
1348 || (STRLEN)data->substrs[1].max_offset > (STRLEN)SSize_t_MAX)
1349 data->substrs[1].max_offset = SSize_t_MAX;
1352 if (data->flags & SF_BEFORE_EOL)
1353 data->substrs[i].flags |= (data->flags & SF_BEFORE_EOL);
1355 data->substrs[i].flags &= ~SF_BEFORE_EOL;
1356 data->substrs[i].minlenp = minlenp;
1357 data->substrs[i].lookbehind = 0;
1360 SvCUR_set(data->last_found, 0);
1362 SV * const sv = data->last_found;
1363 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1364 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1369 data->last_end = -1;
1370 data->flags &= ~SF_BEFORE_EOL;
1371 DEBUG_STUDYDATA("commit", data, 0, is_inf);
1374 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1375 * list that describes which code points it matches */
1378 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1380 /* Set the SSC 'ssc' to match an empty string or any code point */
1382 PERL_ARGS_ASSERT_SSC_ANYTHING;
1384 assert(is_ANYOF_SYNTHETIC(ssc));
1386 /* mortalize so won't leak */
1387 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
1388 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1392 S_ssc_is_anything(const regnode_ssc *ssc)
1394 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1395 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1396 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1397 * in any way, so there's no point in using it */
1402 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1404 assert(is_ANYOF_SYNTHETIC(ssc));
1406 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1410 /* See if the list consists solely of the range 0 - Infinity */
1411 invlist_iterinit(ssc->invlist);
1412 ret = invlist_iternext(ssc->invlist, &start, &end)
1416 invlist_iterfinish(ssc->invlist);
1422 /* If e.g., both \w and \W are set, matches everything */
1423 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1425 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1426 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1436 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1438 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1439 * string, any code point, or any posix class under locale */
1441 PERL_ARGS_ASSERT_SSC_INIT;
1443 Zero(ssc, 1, regnode_ssc);
1444 set_ANYOF_SYNTHETIC(ssc);
1445 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1448 /* If any portion of the regex is to operate under locale rules that aren't
1449 * fully known at compile time, initialization includes it. The reason
1450 * this isn't done for all regexes is that the optimizer was written under
1451 * the assumption that locale was all-or-nothing. Given the complexity and
1452 * lack of documentation in the optimizer, and that there are inadequate
1453 * test cases for locale, many parts of it may not work properly, it is
1454 * safest to avoid locale unless necessary. */
1455 if (RExC_contains_locale) {
1456 ANYOF_POSIXL_SETALL(ssc);
1459 ANYOF_POSIXL_ZERO(ssc);
1464 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1465 const regnode_ssc *ssc)
1467 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1468 * to the list of code points matched, and locale posix classes; hence does
1469 * not check its flags) */
1474 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1476 assert(is_ANYOF_SYNTHETIC(ssc));
1478 invlist_iterinit(ssc->invlist);
1479 ret = invlist_iternext(ssc->invlist, &start, &end)
1483 invlist_iterfinish(ssc->invlist);
1489 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1497 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1498 const regnode_charclass* const node)
1500 /* Returns a mortal inversion list defining which code points are matched
1501 * by 'node', which is of type ANYOF. Handles complementing the result if
1502 * appropriate. If some code points aren't knowable at this time, the
1503 * returned list must, and will, contain every code point that is a
1507 SV* only_utf8_locale_invlist = NULL;
1509 const U32 n = ARG(node);
1510 bool new_node_has_latin1 = FALSE;
1512 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1514 /* Look at the data structure created by S_set_ANYOF_arg() */
1515 if (n != ANYOF_ONLY_HAS_BITMAP) {
1516 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1517 AV * const av = MUTABLE_AV(SvRV(rv));
1518 SV **const ary = AvARRAY(av);
1519 assert(RExC_rxi->data->what[n] == 's');
1521 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1522 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1]), NULL));
1524 else if (ary[0] && ary[0] != &PL_sv_undef) {
1526 /* Here, no compile-time swash, and there are things that won't be
1527 * known until runtime -- we have to assume it could be anything */
1528 invlist = sv_2mortal(_new_invlist(1));
1529 return _add_range_to_invlist(invlist, 0, UV_MAX);
1531 else if (ary[3] && ary[3] != &PL_sv_undef) {
1533 /* Here no compile-time swash, and no run-time only data. Use the
1534 * node's inversion list */
1535 invlist = sv_2mortal(invlist_clone(ary[3], NULL));
1538 /* Get the code points valid only under UTF-8 locales */
1539 if ((ANYOF_FLAGS(node) & ANYOFL_FOLD)
1540 && ary[2] && ary[2] != &PL_sv_undef)
1542 only_utf8_locale_invlist = ary[2];
1547 invlist = sv_2mortal(_new_invlist(0));
1550 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1551 * code points, and an inversion list for the others, but if there are code
1552 * points that should match only conditionally on the target string being
1553 * UTF-8, those are placed in the inversion list, and not the bitmap.
1554 * Since there are circumstances under which they could match, they are
1555 * included in the SSC. But if the ANYOF node is to be inverted, we have
1556 * to exclude them here, so that when we invert below, the end result
1557 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1558 * have to do this here before we add the unconditionally matched code
1560 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1561 _invlist_intersection_complement_2nd(invlist,
1566 /* Add in the points from the bit map */
1567 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1568 if (ANYOF_BITMAP_TEST(node, i)) {
1569 unsigned int start = i++;
1571 for (; i < NUM_ANYOF_CODE_POINTS && ANYOF_BITMAP_TEST(node, i); ++i) {
1574 invlist = _add_range_to_invlist(invlist, start, i-1);
1575 new_node_has_latin1 = TRUE;
1579 /* If this can match all upper Latin1 code points, have to add them
1580 * as well. But don't add them if inverting, as when that gets done below,
1581 * it would exclude all these characters, including the ones it shouldn't
1582 * that were added just above */
1583 if (! (ANYOF_FLAGS(node) & ANYOF_INVERT) && OP(node) == ANYOFD
1584 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1586 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1589 /* Similarly for these */
1590 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1591 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1594 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1595 _invlist_invert(invlist);
1597 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOFL_FOLD) {
1599 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1600 * locale. We can skip this if there are no 0-255 at all. */
1601 _invlist_union(invlist, PL_Latin1, &invlist);
1604 /* Similarly add the UTF-8 locale possible matches. These have to be
1605 * deferred until after the non-UTF-8 locale ones are taken care of just
1606 * above, or it leads to wrong results under ANYOF_INVERT */
1607 if (only_utf8_locale_invlist) {
1608 _invlist_union_maybe_complement_2nd(invlist,
1609 only_utf8_locale_invlist,
1610 ANYOF_FLAGS(node) & ANYOF_INVERT,
1617 /* These two functions currently do the exact same thing */
1618 #define ssc_init_zero ssc_init
1620 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1621 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1623 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1624 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1625 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1628 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1629 const regnode_charclass *and_with)
1631 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1632 * another SSC or a regular ANYOF class. Can create false positives. */
1637 PERL_ARGS_ASSERT_SSC_AND;
1639 assert(is_ANYOF_SYNTHETIC(ssc));
1641 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1642 * the code point inversion list and just the relevant flags */
1643 if (is_ANYOF_SYNTHETIC(and_with)) {
1644 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1645 anded_flags = ANYOF_FLAGS(and_with);
1647 /* XXX This is a kludge around what appears to be deficiencies in the
1648 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1649 * there are paths through the optimizer where it doesn't get weeded
1650 * out when it should. And if we don't make some extra provision for
1651 * it like the code just below, it doesn't get added when it should.
1652 * This solution is to add it only when AND'ing, which is here, and
1653 * only when what is being AND'ed is the pristine, original node
1654 * matching anything. Thus it is like adding it to ssc_anything() but
1655 * only when the result is to be AND'ed. Probably the same solution
1656 * could be adopted for the same problem we have with /l matching,
1657 * which is solved differently in S_ssc_init(), and that would lead to
1658 * fewer false positives than that solution has. But if this solution
1659 * creates bugs, the consequences are only that a warning isn't raised
1660 * that should be; while the consequences for having /l bugs is
1661 * incorrect matches */
1662 if (ssc_is_anything((regnode_ssc *)and_with)) {
1663 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1667 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1668 if (OP(and_with) == ANYOFD) {
1669 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1672 anded_flags = ANYOF_FLAGS(and_with)
1673 &( ANYOF_COMMON_FLAGS
1674 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1675 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1676 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(and_with))) {
1678 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1683 ANYOF_FLAGS(ssc) &= anded_flags;
1685 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1686 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1687 * 'and_with' may be inverted. When not inverted, we have the situation of
1689 * (C1 | P1) & (C2 | P2)
1690 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1691 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1692 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1693 * <= ((C1 & C2) | P1 | P2)
1694 * Alternatively, the last few steps could be:
1695 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1696 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1697 * <= (C1 | C2 | (P1 & P2))
1698 * We favor the second approach if either P1 or P2 is non-empty. This is
1699 * because these components are a barrier to doing optimizations, as what
1700 * they match cannot be known until the moment of matching as they are
1701 * dependent on the current locale, 'AND"ing them likely will reduce or
1703 * But we can do better if we know that C1,P1 are in their initial state (a
1704 * frequent occurrence), each matching everything:
1705 * (<everything>) & (C2 | P2) = C2 | P2
1706 * Similarly, if C2,P2 are in their initial state (again a frequent
1707 * occurrence), the result is a no-op
1708 * (C1 | P1) & (<everything>) = C1 | P1
1711 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1712 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1713 * <= (C1 & ~C2) | (P1 & ~P2)
1716 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1717 && ! is_ANYOF_SYNTHETIC(and_with))
1721 ssc_intersection(ssc,
1723 FALSE /* Has already been inverted */
1726 /* If either P1 or P2 is empty, the intersection will be also; can skip
1728 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
1729 ANYOF_POSIXL_ZERO(ssc);
1731 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1733 /* Note that the Posix class component P from 'and_with' actually
1735 * P = Pa | Pb | ... | Pn
1736 * where each component is one posix class, such as in [\w\s].
1738 * ~P = ~(Pa | Pb | ... | Pn)
1739 * = ~Pa & ~Pb & ... & ~Pn
1740 * <= ~Pa | ~Pb | ... | ~Pn
1741 * The last is something we can easily calculate, but unfortunately
1742 * is likely to have many false positives. We could do better
1743 * in some (but certainly not all) instances if two classes in
1744 * P have known relationships. For example
1745 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1747 * :lower: & :print: = :lower:
1748 * And similarly for classes that must be disjoint. For example,
1749 * since \s and \w can have no elements in common based on rules in
1750 * the POSIX standard,
1751 * \w & ^\S = nothing
1752 * Unfortunately, some vendor locales do not meet the Posix
1753 * standard, in particular almost everything by Microsoft.
1754 * The loop below just changes e.g., \w into \W and vice versa */
1756 regnode_charclass_posixl temp;
1757 int add = 1; /* To calculate the index of the complement */
1759 Zero(&temp, 1, regnode_charclass_posixl);
1760 ANYOF_POSIXL_ZERO(&temp);
1761 for (i = 0; i < ANYOF_MAX; i++) {
1763 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1764 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1766 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1767 ANYOF_POSIXL_SET(&temp, i + add);
1769 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1771 ANYOF_POSIXL_AND(&temp, ssc);
1773 } /* else ssc already has no posixes */
1774 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1775 in its initial state */
1776 else if (! is_ANYOF_SYNTHETIC(and_with)
1777 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1779 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1780 * copy it over 'ssc' */
1781 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1782 if (is_ANYOF_SYNTHETIC(and_with)) {
1783 StructCopy(and_with, ssc, regnode_ssc);
1786 ssc->invlist = anded_cp_list;
1787 ANYOF_POSIXL_ZERO(ssc);
1788 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1789 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1793 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1794 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
1796 /* One or the other of P1, P2 is non-empty. */
1797 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1798 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1800 ssc_union(ssc, anded_cp_list, FALSE);
1802 else { /* P1 = P2 = empty */
1803 ssc_intersection(ssc, anded_cp_list, FALSE);
1809 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1810 const regnode_charclass *or_with)
1812 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1813 * another SSC or a regular ANYOF class. Can create false positives if
1814 * 'or_with' is to be inverted. */
1819 PERL_ARGS_ASSERT_SSC_OR;
1821 assert(is_ANYOF_SYNTHETIC(ssc));
1823 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1824 * the code point inversion list and just the relevant flags */
1825 if (is_ANYOF_SYNTHETIC(or_with)) {
1826 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1827 ored_flags = ANYOF_FLAGS(or_with);
1830 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1831 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1832 if (OP(or_with) != ANYOFD) {
1834 |= ANYOF_FLAGS(or_with)
1835 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1836 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1837 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(or_with))) {
1839 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1844 ANYOF_FLAGS(ssc) |= ored_flags;
1846 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1847 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1848 * 'or_with' may be inverted. When not inverted, we have the simple
1849 * situation of computing:
1850 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1851 * If P1|P2 yields a situation with both a class and its complement are
1852 * set, like having both \w and \W, this matches all code points, and we
1853 * can delete these from the P component of the ssc going forward. XXX We
1854 * might be able to delete all the P components, but I (khw) am not certain
1855 * about this, and it is better to be safe.
1858 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1859 * <= (C1 | P1) | ~C2
1860 * <= (C1 | ~C2) | P1
1861 * (which results in actually simpler code than the non-inverted case)
1864 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1865 && ! is_ANYOF_SYNTHETIC(or_with))
1867 /* We ignore P2, leaving P1 going forward */
1868 } /* else Not inverted */
1869 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
1870 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1871 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1873 for (i = 0; i < ANYOF_MAX; i += 2) {
1874 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1876 ssc_match_all_cp(ssc);
1877 ANYOF_POSIXL_CLEAR(ssc, i);
1878 ANYOF_POSIXL_CLEAR(ssc, i+1);
1886 FALSE /* Already has been inverted */
1890 PERL_STATIC_INLINE void
1891 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1893 PERL_ARGS_ASSERT_SSC_UNION;
1895 assert(is_ANYOF_SYNTHETIC(ssc));
1897 _invlist_union_maybe_complement_2nd(ssc->invlist,
1903 PERL_STATIC_INLINE void
1904 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1906 const bool invert2nd)
1908 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1910 assert(is_ANYOF_SYNTHETIC(ssc));
1912 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1918 PERL_STATIC_INLINE void
1919 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1921 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1923 assert(is_ANYOF_SYNTHETIC(ssc));
1925 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1928 PERL_STATIC_INLINE void
1929 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1931 /* AND just the single code point 'cp' into the SSC 'ssc' */
1933 SV* cp_list = _new_invlist(2);
1935 PERL_ARGS_ASSERT_SSC_CP_AND;
1937 assert(is_ANYOF_SYNTHETIC(ssc));
1939 cp_list = add_cp_to_invlist(cp_list, cp);
1940 ssc_intersection(ssc, cp_list,
1941 FALSE /* Not inverted */
1943 SvREFCNT_dec_NN(cp_list);
1946 PERL_STATIC_INLINE void
1947 S_ssc_clear_locale(regnode_ssc *ssc)
1949 /* Set the SSC 'ssc' to not match any locale things */
1950 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1952 assert(is_ANYOF_SYNTHETIC(ssc));
1954 ANYOF_POSIXL_ZERO(ssc);
1955 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1958 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1961 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1963 /* The synthetic start class is used to hopefully quickly winnow down
1964 * places where a pattern could start a match in the target string. If it
1965 * doesn't really narrow things down that much, there isn't much point to
1966 * having the overhead of using it. This function uses some very crude
1967 * heuristics to decide if to use the ssc or not.
1969 * It returns TRUE if 'ssc' rules out more than half what it considers to
1970 * be the "likely" possible matches, but of course it doesn't know what the
1971 * actual things being matched are going to be; these are only guesses
1973 * For /l matches, it assumes that the only likely matches are going to be
1974 * in the 0-255 range, uniformly distributed, so half of that is 127
1975 * For /a and /d matches, it assumes that the likely matches will be just
1976 * the ASCII range, so half of that is 63
1977 * For /u and there isn't anything matching above the Latin1 range, it
1978 * assumes that that is the only range likely to be matched, and uses
1979 * half that as the cut-off: 127. If anything matches above Latin1,
1980 * it assumes that all of Unicode could match (uniformly), except for
1981 * non-Unicode code points and things in the General Category "Other"
1982 * (unassigned, private use, surrogates, controls and formats). This
1983 * is a much large number. */
1985 U32 count = 0; /* Running total of number of code points matched by
1987 UV start, end; /* Start and end points of current range in inversion
1989 const U32 max_code_points = (LOC)
1991 : (( ! UNI_SEMANTICS
1992 || invlist_highest(ssc->invlist) < 256)
1995 const U32 max_match = max_code_points / 2;
1997 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1999 invlist_iterinit(ssc->invlist);
2000 while (invlist_iternext(ssc->invlist, &start, &end)) {
2001 if (start >= max_code_points) {
2004 end = MIN(end, max_code_points - 1);
2005 count += end - start + 1;
2006 if (count >= max_match) {
2007 invlist_iterfinish(ssc->invlist);
2017 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
2019 /* The inversion list in the SSC is marked mortal; now we need a more
2020 * permanent copy, which is stored the same way that is done in a regular
2021 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
2024 SV* invlist = invlist_clone(ssc->invlist, NULL);
2026 PERL_ARGS_ASSERT_SSC_FINALIZE;
2028 assert(is_ANYOF_SYNTHETIC(ssc));
2030 /* The code in this file assumes that all but these flags aren't relevant
2031 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
2032 * by the time we reach here */
2033 assert(! (ANYOF_FLAGS(ssc)
2034 & ~( ANYOF_COMMON_FLAGS
2035 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
2036 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
2038 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
2040 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
2041 NULL, NULL, NULL, FALSE);
2043 /* Make sure is clone-safe */
2044 ssc->invlist = NULL;
2046 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
2047 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
2048 OP(ssc) = ANYOFPOSIXL;
2050 else if (RExC_contains_locale) {
2054 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
2057 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
2058 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
2059 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
2060 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
2061 ? (TRIE_LIST_CUR( idx ) - 1) \
2067 dump_trie(trie,widecharmap,revcharmap)
2068 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
2069 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
2071 These routines dump out a trie in a somewhat readable format.
2072 The _interim_ variants are used for debugging the interim
2073 tables that are used to generate the final compressed
2074 representation which is what dump_trie expects.
2076 Part of the reason for their existence is to provide a form
2077 of documentation as to how the different representations function.
2082 Dumps the final compressed table form of the trie to Perl_debug_log.
2083 Used for debugging make_trie().
2087 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
2088 AV *revcharmap, U32 depth)
2091 SV *sv=sv_newmortal();
2092 int colwidth= widecharmap ? 6 : 4;
2094 GET_RE_DEBUG_FLAGS_DECL;
2096 PERL_ARGS_ASSERT_DUMP_TRIE;
2098 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
2099 depth+1, "Match","Base","Ofs" );
2101 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2102 SV ** const tmp = av_fetch( revcharmap, state, 0);
2104 Perl_re_printf( aTHX_ "%*s",
2106 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2107 PL_colors[0], PL_colors[1],
2108 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2109 PERL_PV_ESCAPE_FIRSTCHAR
2114 Perl_re_printf( aTHX_ "\n");
2115 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
2117 for( state = 0 ; state < trie->uniquecharcount ; state++ )
2118 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2119 Perl_re_printf( aTHX_ "\n");
2121 for( state = 1 ; state < trie->statecount ; state++ ) {
2122 const U32 base = trie->states[ state ].trans.base;
2124 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
2126 if ( trie->states[ state ].wordnum ) {
2127 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
2129 Perl_re_printf( aTHX_ "%6s", "" );
2132 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
2137 while( ( base + ofs < trie->uniquecharcount ) ||
2138 ( base + ofs - trie->uniquecharcount < trie->lasttrans
2139 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2143 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
2145 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2146 if ( ( base + ofs >= trie->uniquecharcount )
2147 && ( base + ofs - trie->uniquecharcount
2149 && trie->trans[ base + ofs
2150 - trie->uniquecharcount ].check == state )
2152 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
2153 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2156 Perl_re_printf( aTHX_ "%*s", colwidth," ." );
2160 Perl_re_printf( aTHX_ "]");
2163 Perl_re_printf( aTHX_ "\n" );
2165 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
2167 for (word=1; word <= trie->wordcount; word++) {
2168 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2169 (int)word, (int)(trie->wordinfo[word].prev),
2170 (int)(trie->wordinfo[word].len));
2172 Perl_re_printf( aTHX_ "\n" );
2175 Dumps a fully constructed but uncompressed trie in list form.
2176 List tries normally only are used for construction when the number of
2177 possible chars (trie->uniquecharcount) is very high.
2178 Used for debugging make_trie().
2181 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2182 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2186 SV *sv=sv_newmortal();
2187 int colwidth= widecharmap ? 6 : 4;
2188 GET_RE_DEBUG_FLAGS_DECL;
2190 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2192 /* print out the table precompression. */
2193 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
2195 Perl_re_indentf( aTHX_ "%s",
2196 depth+1, "------:-----+-----------------\n" );
2198 for( state=1 ; state < next_alloc ; state ++ ) {
2201 Perl_re_indentf( aTHX_ " %4" UVXf " :",
2202 depth+1, (UV)state );
2203 if ( ! trie->states[ state ].wordnum ) {
2204 Perl_re_printf( aTHX_ "%5s| ","");
2206 Perl_re_printf( aTHX_ "W%4x| ",
2207 trie->states[ state ].wordnum
2210 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2211 SV ** const tmp = av_fetch( revcharmap,
2212 TRIE_LIST_ITEM(state, charid).forid, 0);
2214 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
2216 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2218 PL_colors[0], PL_colors[1],
2219 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2220 | PERL_PV_ESCAPE_FIRSTCHAR
2222 TRIE_LIST_ITEM(state, charid).forid,
2223 (UV)TRIE_LIST_ITEM(state, charid).newstate
2226 Perl_re_printf( aTHX_ "\n%*s| ",
2227 (int)((depth * 2) + 14), "");
2230 Perl_re_printf( aTHX_ "\n");
2235 Dumps a fully constructed but uncompressed trie in table form.
2236 This is the normal DFA style state transition table, with a few
2237 twists to facilitate compression later.
2238 Used for debugging make_trie().
2241 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2242 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2247 SV *sv=sv_newmortal();
2248 int colwidth= widecharmap ? 6 : 4;
2249 GET_RE_DEBUG_FLAGS_DECL;
2251 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2254 print out the table precompression so that we can do a visual check
2255 that they are identical.
2258 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
2260 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2261 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2263 Perl_re_printf( aTHX_ "%*s",
2265 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2266 PL_colors[0], PL_colors[1],
2267 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2268 PERL_PV_ESCAPE_FIRSTCHAR
2274 Perl_re_printf( aTHX_ "\n");
2275 Perl_re_indentf( aTHX_ "State+-", depth+1 );
2277 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2278 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
2281 Perl_re_printf( aTHX_ "\n" );
2283 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2285 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
2287 (UV)TRIE_NODENUM( state ) );
2289 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2290 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2292 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
2294 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
2296 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2297 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
2298 (UV)trie->trans[ state ].check );
2300 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
2301 (UV)trie->trans[ state ].check,
2302 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2310 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2311 startbranch: the first branch in the whole branch sequence
2312 first : start branch of sequence of branch-exact nodes.
2313 May be the same as startbranch
2314 last : Thing following the last branch.
2315 May be the same as tail.
2316 tail : item following the branch sequence
2317 count : words in the sequence
2318 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2319 depth : indent depth
2321 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2323 A trie is an N'ary tree where the branches are determined by digital
2324 decomposition of the key. IE, at the root node you look up the 1st character and
2325 follow that branch repeat until you find the end of the branches. Nodes can be
2326 marked as "accepting" meaning they represent a complete word. Eg:
2330 would convert into the following structure. Numbers represent states, letters
2331 following numbers represent valid transitions on the letter from that state, if
2332 the number is in square brackets it represents an accepting state, otherwise it
2333 will be in parenthesis.
2335 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2339 (1) +-i->(6)-+-s->[7]
2341 +-s->(3)-+-h->(4)-+-e->[5]
2343 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2345 This shows that when matching against the string 'hers' we will begin at state 1
2346 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2347 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2348 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2349 single traverse. We store a mapping from accepting to state to which word was
2350 matched, and then when we have multiple possibilities we try to complete the
2351 rest of the regex in the order in which they occurred in the alternation.
2353 The only prior NFA like behaviour that would be changed by the TRIE support is
2354 the silent ignoring of duplicate alternations which are of the form:
2356 / (DUPE|DUPE) X? (?{ ... }) Y /x
2358 Thus EVAL blocks following a trie may be called a different number of times with
2359 and without the optimisation. With the optimisations dupes will be silently
2360 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2361 the following demonstrates:
2363 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2365 which prints out 'word' three times, but
2367 'words'=~/(word|word|word)(?{ print $1 })S/
2369 which doesnt print it out at all. This is due to other optimisations kicking in.
2371 Example of what happens on a structural level:
2373 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2375 1: CURLYM[1] {1,32767}(18)
2386 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2387 and should turn into:
2389 1: CURLYM[1] {1,32767}(18)
2391 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2399 Cases where tail != last would be like /(?foo|bar)baz/:
2409 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2410 and would end up looking like:
2413 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2420 d = uvchr_to_utf8_flags(d, uv, 0);
2422 is the recommended Unicode-aware way of saying
2427 #define TRIE_STORE_REVCHAR(val) \
2430 SV *zlopp = newSV(UTF8_MAXBYTES); \
2431 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2432 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2433 SvCUR_set(zlopp, kapow - flrbbbbb); \
2436 av_push(revcharmap, zlopp); \
2438 char ooooff = (char)val; \
2439 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2443 /* This gets the next character from the input, folding it if not already
2445 #define TRIE_READ_CHAR STMT_START { \
2448 /* if it is UTF then it is either already folded, or does not need \
2450 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2452 else if (folder == PL_fold_latin1) { \
2453 /* This folder implies Unicode rules, which in the range expressible \
2454 * by not UTF is the lower case, with the two exceptions, one of \
2455 * which should have been taken care of before calling this */ \
2456 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2457 uvc = toLOWER_L1(*uc); \
2458 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2461 /* raw data, will be folded later if needed */ \
2469 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2470 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2471 U32 ging = TRIE_LIST_LEN( state ) * 2; \
2472 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2473 TRIE_LIST_LEN( state ) = ging; \
2475 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2476 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2477 TRIE_LIST_CUR( state )++; \
2480 #define TRIE_LIST_NEW(state) STMT_START { \
2481 Newx( trie->states[ state ].trans.list, \
2482 4, reg_trie_trans_le ); \
2483 TRIE_LIST_CUR( state ) = 1; \
2484 TRIE_LIST_LEN( state ) = 4; \
2487 #define TRIE_HANDLE_WORD(state) STMT_START { \
2488 U16 dupe= trie->states[ state ].wordnum; \
2489 regnode * const noper_next = regnext( noper ); \
2492 /* store the word for dumping */ \
2494 if (OP(noper) != NOTHING) \
2495 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2497 tmp = newSVpvn_utf8( "", 0, UTF ); \
2498 av_push( trie_words, tmp ); \
2502 trie->wordinfo[curword].prev = 0; \
2503 trie->wordinfo[curword].len = wordlen; \
2504 trie->wordinfo[curword].accept = state; \
2506 if ( noper_next < tail ) { \
2508 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2510 trie->jump[curword] = (U16)(noper_next - convert); \
2512 jumper = noper_next; \
2514 nextbranch= regnext(cur); \
2518 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2519 /* chain, so that when the bits of chain are later */\
2520 /* linked together, the dups appear in the chain */\
2521 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2522 trie->wordinfo[dupe].prev = curword; \
2524 /* we haven't inserted this word yet. */ \
2525 trie->states[ state ].wordnum = curword; \
2530 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2531 ( ( base + charid >= ucharcount \
2532 && base + charid < ubound \
2533 && state == trie->trans[ base - ucharcount + charid ].check \
2534 && trie->trans[ base - ucharcount + charid ].next ) \
2535 ? trie->trans[ base - ucharcount + charid ].next \
2536 : ( state==1 ? special : 0 ) \
2539 #define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2541 TRIE_BITMAP_SET(trie, uvc); \
2542 /* store the folded codepoint */ \
2544 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2547 /* store first byte of utf8 representation of */ \
2548 /* variant codepoints */ \
2549 if (! UVCHR_IS_INVARIANT(uvc)) { \
2550 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2555 #define MADE_JUMP_TRIE 2
2556 #define MADE_EXACT_TRIE 4
2559 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2560 regnode *first, regnode *last, regnode *tail,
2561 U32 word_count, U32 flags, U32 depth)
2563 /* first pass, loop through and scan words */
2564 reg_trie_data *trie;
2565 HV *widecharmap = NULL;
2566 AV *revcharmap = newAV();
2572 regnode *jumper = NULL;
2573 regnode *nextbranch = NULL;
2574 regnode *convert = NULL;
2575 U32 *prev_states; /* temp array mapping each state to previous one */
2576 /* we just use folder as a flag in utf8 */
2577 const U8 * folder = NULL;
2579 /* in the below add_data call we are storing either 'tu' or 'tuaa'
2580 * which stands for one trie structure, one hash, optionally followed
2583 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuaa"));
2584 AV *trie_words = NULL;
2585 /* along with revcharmap, this only used during construction but both are
2586 * useful during debugging so we store them in the struct when debugging.
2589 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2590 STRLEN trie_charcount=0;
2592 SV *re_trie_maxbuff;
2593 GET_RE_DEBUG_FLAGS_DECL;
2595 PERL_ARGS_ASSERT_MAKE_TRIE;
2597 PERL_UNUSED_ARG(depth);
2601 case EXACT: case EXACTL: break;
2605 case EXACTFLU8: folder = PL_fold_latin1; break;
2606 case EXACTF: folder = PL_fold; break;
2607 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2610 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2612 trie->startstate = 1;
2613 trie->wordcount = word_count;
2614 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2615 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2616 if (flags == EXACT || flags == EXACTL)
2617 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2618 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2619 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2622 trie_words = newAV();
2625 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2626 assert(re_trie_maxbuff);
2627 if (!SvIOK(re_trie_maxbuff)) {
2628 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2630 DEBUG_TRIE_COMPILE_r({
2631 Perl_re_indentf( aTHX_
2632 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2634 REG_NODE_NUM(startbranch), REG_NODE_NUM(first),
2635 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2638 /* Find the node we are going to overwrite */
2639 if ( first == startbranch && OP( last ) != BRANCH ) {
2640 /* whole branch chain */
2643 /* branch sub-chain */
2644 convert = NEXTOPER( first );
2647 /* -- First loop and Setup --
2649 We first traverse the branches and scan each word to determine if it
2650 contains widechars, and how many unique chars there are, this is
2651 important as we have to build a table with at least as many columns as we
2654 We use an array of integers to represent the character codes 0..255
2655 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2656 the native representation of the character value as the key and IV's for
2659 *TODO* If we keep track of how many times each character is used we can
2660 remap the columns so that the table compression later on is more
2661 efficient in terms of memory by ensuring the most common value is in the
2662 middle and the least common are on the outside. IMO this would be better
2663 than a most to least common mapping as theres a decent chance the most
2664 common letter will share a node with the least common, meaning the node
2665 will not be compressible. With a middle is most common approach the worst
2666 case is when we have the least common nodes twice.
2670 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2671 regnode *noper = NEXTOPER( cur );
2675 U32 wordlen = 0; /* required init */
2676 STRLEN minchars = 0;
2677 STRLEN maxchars = 0;
2678 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2681 if (OP(noper) == NOTHING) {
2682 /* skip past a NOTHING at the start of an alternation
2683 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2685 regnode *noper_next= regnext(noper);
2686 if (noper_next < tail)
2690 if ( noper < tail &&
2692 OP(noper) == flags ||
2695 OP(noper) == EXACTFU_SS
2699 uc= (U8*)STRING(noper);
2700 e= uc + STR_LEN(noper);
2707 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2708 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2709 regardless of encoding */
2710 if (OP( noper ) == EXACTFU_SS) {
2711 /* false positives are ok, so just set this */
2712 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2716 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2718 TRIE_CHARCOUNT(trie)++;
2721 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2722 * is in effect. Under /i, this character can match itself, or
2723 * anything that folds to it. If not under /i, it can match just
2724 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2725 * all fold to k, and all are single characters. But some folds
2726 * expand to more than one character, so for example LATIN SMALL
2727 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2728 * the string beginning at 'uc' is 'ffi', it could be matched by
2729 * three characters, or just by the one ligature character. (It
2730 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2731 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2732 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2733 * match.) The trie needs to know the minimum and maximum number
2734 * of characters that could match so that it can use size alone to
2735 * quickly reject many match attempts. The max is simple: it is
2736 * the number of folded characters in this branch (since a fold is
2737 * never shorter than what folds to it. */
2741 /* And the min is equal to the max if not under /i (indicated by
2742 * 'folder' being NULL), or there are no multi-character folds. If
2743 * there is a multi-character fold, the min is incremented just
2744 * once, for the character that folds to the sequence. Each
2745 * character in the sequence needs to be added to the list below of
2746 * characters in the trie, but we count only the first towards the
2747 * min number of characters needed. This is done through the
2748 * variable 'foldlen', which is returned by the macros that look
2749 * for these sequences as the number of bytes the sequence
2750 * occupies. Each time through the loop, we decrement 'foldlen' by
2751 * how many bytes the current char occupies. Only when it reaches
2752 * 0 do we increment 'minchars' or look for another multi-character
2754 if (folder == NULL) {
2757 else if (foldlen > 0) {
2758 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2763 /* See if *uc is the beginning of a multi-character fold. If
2764 * so, we decrement the length remaining to look at, to account
2765 * for the current character this iteration. (We can use 'uc'
2766 * instead of the fold returned by TRIE_READ_CHAR because for
2767 * non-UTF, the latin1_safe macro is smart enough to account
2768 * for all the unfolded characters, and because for UTF, the
2769 * string will already have been folded earlier in the
2770 * compilation process */
2772 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2773 foldlen -= UTF8SKIP(uc);
2776 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2781 /* The current character (and any potential folds) should be added
2782 * to the possible matching characters for this position in this
2786 U8 folded= folder[ (U8) uvc ];
2787 if ( !trie->charmap[ folded ] ) {
2788 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2789 TRIE_STORE_REVCHAR( folded );
2792 if ( !trie->charmap[ uvc ] ) {
2793 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2794 TRIE_STORE_REVCHAR( uvc );
2797 /* store the codepoint in the bitmap, and its folded
2799 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
2800 set_bit = 0; /* We've done our bit :-) */
2804 /* XXX We could come up with the list of code points that fold
2805 * to this using PL_utf8_foldclosures, except not for
2806 * multi-char folds, as there may be multiple combinations
2807 * there that could work, which needs to wait until runtime to
2808 * resolve (The comment about LIGATURE FFI above is such an
2813 widecharmap = newHV();
2815 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2818 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
2820 if ( !SvTRUE( *svpp ) ) {
2821 sv_setiv( *svpp, ++trie->uniquecharcount );
2822 TRIE_STORE_REVCHAR(uvc);
2825 } /* end loop through characters in this branch of the trie */
2827 /* We take the min and max for this branch and combine to find the min
2828 * and max for all branches processed so far */
2829 if( cur == first ) {
2830 trie->minlen = minchars;
2831 trie->maxlen = maxchars;
2832 } else if (minchars < trie->minlen) {
2833 trie->minlen = minchars;
2834 } else if (maxchars > trie->maxlen) {
2835 trie->maxlen = maxchars;
2837 } /* end first pass */
2838 DEBUG_TRIE_COMPILE_r(
2839 Perl_re_indentf( aTHX_
2840 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2842 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2843 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2844 (int)trie->minlen, (int)trie->maxlen )
2848 We now know what we are dealing with in terms of unique chars and
2849 string sizes so we can calculate how much memory a naive
2850 representation using a flat table will take. If it's over a reasonable
2851 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2852 conservative but potentially much slower representation using an array
2855 At the end we convert both representations into the same compressed
2856 form that will be used in regexec.c for matching with. The latter
2857 is a form that cannot be used to construct with but has memory
2858 properties similar to the list form and access properties similar
2859 to the table form making it both suitable for fast searches and
2860 small enough that its feasable to store for the duration of a program.
2862 See the comment in the code where the compressed table is produced
2863 inplace from the flat tabe representation for an explanation of how
2864 the compression works.
2869 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2872 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2873 > SvIV(re_trie_maxbuff) )
2876 Second Pass -- Array Of Lists Representation
2878 Each state will be represented by a list of charid:state records
2879 (reg_trie_trans_le) the first such element holds the CUR and LEN
2880 points of the allocated array. (See defines above).
2882 We build the initial structure using the lists, and then convert
2883 it into the compressed table form which allows faster lookups
2884 (but cant be modified once converted).
2887 STRLEN transcount = 1;
2889 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
2892 trie->states = (reg_trie_state *)
2893 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2894 sizeof(reg_trie_state) );
2898 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2900 regnode *noper = NEXTOPER( cur );
2901 U32 state = 1; /* required init */
2902 U16 charid = 0; /* sanity init */
2903 U32 wordlen = 0; /* required init */
2905 if (OP(noper) == NOTHING) {
2906 regnode *noper_next= regnext(noper);
2907 if (noper_next < tail)
2911 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
2912 const U8 *uc= (U8*)STRING(noper);
2913 const U8 *e= uc + STR_LEN(noper);
2915 for ( ; uc < e ; uc += len ) {
2920 charid = trie->charmap[ uvc ];
2922 SV** const svpp = hv_fetch( widecharmap,
2929 charid=(U16)SvIV( *svpp );
2932 /* charid is now 0 if we dont know the char read, or
2933 * nonzero if we do */
2940 if ( !trie->states[ state ].trans.list ) {
2941 TRIE_LIST_NEW( state );
2944 check <= TRIE_LIST_USED( state );
2947 if ( TRIE_LIST_ITEM( state, check ).forid
2950 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2955 newstate = next_alloc++;
2956 prev_states[newstate] = state;
2957 TRIE_LIST_PUSH( state, charid, newstate );
2962 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
2966 TRIE_HANDLE_WORD(state);
2968 } /* end second pass */
2970 /* next alloc is the NEXT state to be allocated */
2971 trie->statecount = next_alloc;
2972 trie->states = (reg_trie_state *)
2973 PerlMemShared_realloc( trie->states,
2975 * sizeof(reg_trie_state) );
2977 /* and now dump it out before we compress it */
2978 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2979 revcharmap, next_alloc,
2983 trie->trans = (reg_trie_trans *)
2984 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2991 for( state=1 ; state < next_alloc ; state ++ ) {
2995 DEBUG_TRIE_COMPILE_MORE_r(
2996 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
3000 if (trie->states[state].trans.list) {
3001 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
3005 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3006 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
3007 if ( forid < minid ) {
3009 } else if ( forid > maxid ) {
3013 if ( transcount < tp + maxid - minid + 1) {
3015 trie->trans = (reg_trie_trans *)
3016 PerlMemShared_realloc( trie->trans,
3018 * sizeof(reg_trie_trans) );
3019 Zero( trie->trans + (transcount / 2),
3023 base = trie->uniquecharcount + tp - minid;
3024 if ( maxid == minid ) {
3026 for ( ; zp < tp ; zp++ ) {
3027 if ( ! trie->trans[ zp ].next ) {
3028 base = trie->uniquecharcount + zp - minid;
3029 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
3031 trie->trans[ zp ].check = state;
3037 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
3039 trie->trans[ tp ].check = state;
3044 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3045 const U32 tid = base
3046 - trie->uniquecharcount
3047 + TRIE_LIST_ITEM( state, idx ).forid;
3048 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
3050 trie->trans[ tid ].check = state;
3052 tp += ( maxid - minid + 1 );
3054 Safefree(trie->states[ state ].trans.list);
3057 DEBUG_TRIE_COMPILE_MORE_r(
3058 Perl_re_printf( aTHX_ " base: %d\n",base);
3061 trie->states[ state ].trans.base=base;
3063 trie->lasttrans = tp + 1;
3067 Second Pass -- Flat Table Representation.
3069 we dont use the 0 slot of either trans[] or states[] so we add 1 to
3070 each. We know that we will need Charcount+1 trans at most to store
3071 the data (one row per char at worst case) So we preallocate both
3072 structures assuming worst case.
3074 We then construct the trie using only the .next slots of the entry
3077 We use the .check field of the first entry of the node temporarily
3078 to make compression both faster and easier by keeping track of how
3079 many non zero fields are in the node.
3081 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
3084 There are two terms at use here: state as a TRIE_NODEIDX() which is
3085 a number representing the first entry of the node, and state as a
3086 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
3087 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
3088 if there are 2 entrys per node. eg:
3096 The table is internally in the right hand, idx form. However as we
3097 also have to deal with the states array which is indexed by nodenum
3098 we have to use TRIE_NODENUM() to convert.
3101 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
3104 trie->trans = (reg_trie_trans *)
3105 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
3106 * trie->uniquecharcount + 1,
3107 sizeof(reg_trie_trans) );
3108 trie->states = (reg_trie_state *)
3109 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3110 sizeof(reg_trie_state) );
3111 next_alloc = trie->uniquecharcount + 1;
3114 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3116 regnode *noper = NEXTOPER( cur );
3118 U32 state = 1; /* required init */
3120 U16 charid = 0; /* sanity init */
3121 U32 accept_state = 0; /* sanity init */
3123 U32 wordlen = 0; /* required init */
3125 if (OP(noper) == NOTHING) {
3126 regnode *noper_next= regnext(noper);
3127 if (noper_next < tail)
3131 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
3132 const U8 *uc= (U8*)STRING(noper);
3133 const U8 *e= uc + STR_LEN(noper);
3135 for ( ; uc < e ; uc += len ) {
3140 charid = trie->charmap[ uvc ];
3142 SV* const * const svpp = hv_fetch( widecharmap,
3146 charid = svpp ? (U16)SvIV(*svpp) : 0;
3150 if ( !trie->trans[ state + charid ].next ) {
3151 trie->trans[ state + charid ].next = next_alloc;
3152 trie->trans[ state ].check++;
3153 prev_states[TRIE_NODENUM(next_alloc)]
3154 = TRIE_NODENUM(state);
3155 next_alloc += trie->uniquecharcount;
3157 state = trie->trans[ state + charid ].next;
3159 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3161 /* charid is now 0 if we dont know the char read, or
3162 * nonzero if we do */
3165 accept_state = TRIE_NODENUM( state );
3166 TRIE_HANDLE_WORD(accept_state);
3168 } /* end second pass */
3170 /* and now dump it out before we compress it */
3171 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3173 next_alloc, depth+1));
3177 * Inplace compress the table.*
3179 For sparse data sets the table constructed by the trie algorithm will
3180 be mostly 0/FAIL transitions or to put it another way mostly empty.
3181 (Note that leaf nodes will not contain any transitions.)
3183 This algorithm compresses the tables by eliminating most such
3184 transitions, at the cost of a modest bit of extra work during lookup:
3186 - Each states[] entry contains a .base field which indicates the
3187 index in the state[] array wheres its transition data is stored.
3189 - If .base is 0 there are no valid transitions from that node.
3191 - If .base is nonzero then charid is added to it to find an entry in
3194 -If trans[states[state].base+charid].check!=state then the
3195 transition is taken to be a 0/Fail transition. Thus if there are fail
3196 transitions at the front of the node then the .base offset will point
3197 somewhere inside the previous nodes data (or maybe even into a node
3198 even earlier), but the .check field determines if the transition is
3202 The following process inplace converts the table to the compressed
3203 table: We first do not compress the root node 1,and mark all its
3204 .check pointers as 1 and set its .base pointer as 1 as well. This
3205 allows us to do a DFA construction from the compressed table later,
3206 and ensures that any .base pointers we calculate later are greater
3209 - We set 'pos' to indicate the first entry of the second node.
3211 - We then iterate over the columns of the node, finding the first and
3212 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3213 and set the .check pointers accordingly, and advance pos
3214 appropriately and repreat for the next node. Note that when we copy
3215 the next pointers we have to convert them from the original
3216 NODEIDX form to NODENUM form as the former is not valid post
3219 - If a node has no transitions used we mark its base as 0 and do not
3220 advance the pos pointer.
3222 - If a node only has one transition we use a second pointer into the
3223 structure to fill in allocated fail transitions from other states.
3224 This pointer is independent of the main pointer and scans forward
3225 looking for null transitions that are allocated to a state. When it
3226 finds one it writes the single transition into the "hole". If the
3227 pointer doesnt find one the single transition is appended as normal.
3229 - Once compressed we can Renew/realloc the structures to release the
3232 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3233 specifically Fig 3.47 and the associated pseudocode.
3237 const U32 laststate = TRIE_NODENUM( next_alloc );
3240 trie->statecount = laststate;
3242 for ( state = 1 ; state < laststate ; state++ ) {
3244 const U32 stateidx = TRIE_NODEIDX( state );
3245 const U32 o_used = trie->trans[ stateidx ].check;
3246 U32 used = trie->trans[ stateidx ].check;
3247 trie->trans[ stateidx ].check = 0;
3250 used && charid < trie->uniquecharcount;
3253 if ( flag || trie->trans[ stateidx + charid ].next ) {
3254 if ( trie->trans[ stateidx + charid ].next ) {
3256 for ( ; zp < pos ; zp++ ) {
3257 if ( ! trie->trans[ zp ].next ) {
3261 trie->states[ state ].trans.base
3263 + trie->uniquecharcount
3265 trie->trans[ zp ].next
3266 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3268 trie->trans[ zp ].check = state;
3269 if ( ++zp > pos ) pos = zp;
3276 trie->states[ state ].trans.base
3277 = pos + trie->uniquecharcount - charid ;
3279 trie->trans[ pos ].next
3280 = SAFE_TRIE_NODENUM(
3281 trie->trans[ stateidx + charid ].next );
3282 trie->trans[ pos ].check = state;
3287 trie->lasttrans = pos + 1;
3288 trie->states = (reg_trie_state *)
3289 PerlMemShared_realloc( trie->states, laststate
3290 * sizeof(reg_trie_state) );
3291 DEBUG_TRIE_COMPILE_MORE_r(
3292 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
3294 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3298 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3301 } /* end table compress */
3303 DEBUG_TRIE_COMPILE_MORE_r(
3304 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
3306 (UV)trie->statecount,
3307 (UV)trie->lasttrans)
3309 /* resize the trans array to remove unused space */
3310 trie->trans = (reg_trie_trans *)
3311 PerlMemShared_realloc( trie->trans, trie->lasttrans
3312 * sizeof(reg_trie_trans) );
3314 { /* Modify the program and insert the new TRIE node */
3315 U8 nodetype =(U8)(flags & 0xFF);
3319 regnode *optimize = NULL;
3320 #ifdef RE_TRACK_PATTERN_OFFSETS
3323 U32 mjd_nodelen = 0;
3324 #endif /* RE_TRACK_PATTERN_OFFSETS */
3325 #endif /* DEBUGGING */
3327 This means we convert either the first branch or the first Exact,
3328 depending on whether the thing following (in 'last') is a branch
3329 or not and whther first is the startbranch (ie is it a sub part of
3330 the alternation or is it the whole thing.)
3331 Assuming its a sub part we convert the EXACT otherwise we convert
3332 the whole branch sequence, including the first.
3334 /* Find the node we are going to overwrite */
3335 if ( first != startbranch || OP( last ) == BRANCH ) {
3336 /* branch sub-chain */
3337 NEXT_OFF( first ) = (U16)(last - first);
3338 #ifdef RE_TRACK_PATTERN_OFFSETS
3340 mjd_offset= Node_Offset((convert));
3341 mjd_nodelen= Node_Length((convert));
3344 /* whole branch chain */
3346 #ifdef RE_TRACK_PATTERN_OFFSETS
3349 const regnode *nop = NEXTOPER( convert );
3350 mjd_offset= Node_Offset((nop));
3351 mjd_nodelen= Node_Length((nop));
3355 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
3357 (UV)mjd_offset, (UV)mjd_nodelen)
3360 /* But first we check to see if there is a common prefix we can
3361 split out as an EXACT and put in front of the TRIE node. */
3362 trie->startstate= 1;
3363 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3364 /* we want to find the first state that has more than
3365 * one transition, if that state is not the first state
3366 * then we have a common prefix which we can remove.
3369 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3371 I32 first_ofs = -1; /* keeps track of the ofs of the first
3372 transition, -1 means none */
3374 const U32 base = trie->states[ state ].trans.base;
3376 /* does this state terminate an alternation? */
3377 if ( trie->states[state].wordnum )
3380 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3381 if ( ( base + ofs >= trie->uniquecharcount ) &&
3382 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3383 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3385 if ( ++count > 1 ) {
3386 /* we have more than one transition */
3389 /* if this is the first state there is no common prefix
3390 * to extract, so we can exit */
3391 if ( state == 1 ) break;
3392 tmp = av_fetch( revcharmap, ofs, 0);
3393 ch = (U8*)SvPV_nolen_const( *tmp );
3395 /* if we are on count 2 then we need to initialize the
3396 * bitmap, and store the previous char if there was one
3399 /* clear the bitmap */
3400 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3402 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
3405 if (first_ofs >= 0) {
3406 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
3407 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3409 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3411 Perl_re_printf( aTHX_ "%s", (char*)ch)
3415 /* store the current firstchar in the bitmap */
3416 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3417 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
3423 /* This state has only one transition, its transition is part
3424 * of a common prefix - we need to concatenate the char it
3425 * represents to what we have so far. */
3426 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
3428 char *ch = SvPV( *tmp, len );
3430 SV *sv=sv_newmortal();
3431 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
3433 (UV)state, (UV)first_ofs,
3434 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3435 PL_colors[0], PL_colors[1],
3436 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3437 PERL_PV_ESCAPE_FIRSTCHAR
3442 OP( convert ) = nodetype;
3443 str=STRING(convert);
3446 STR_LEN(convert) += len;
3452 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
3457 trie->prefixlen = (state-1);
3459 regnode *n = convert+NODE_SZ_STR(convert);
3460 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3461 trie->startstate = state;
3462 trie->minlen -= (state - 1);
3463 trie->maxlen -= (state - 1);
3465 /* At least the UNICOS C compiler choked on this
3466 * being argument to DEBUG_r(), so let's just have
3469 #ifdef PERL_EXT_RE_BUILD
3475 regnode *fix = convert;
3476 U32 word = trie->wordcount;
3477 #ifdef RE_TRACK_PATTERN_OFFSETS
3480 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3481 while( ++fix < n ) {
3482 Set_Node_Offset_Length(fix, 0, 0);
3485 SV ** const tmp = av_fetch( trie_words, word, 0 );
3487 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3488 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3490 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3498 NEXT_OFF(convert) = (U16)(tail - convert);
3499 DEBUG_r(optimize= n);
3505 if ( trie->maxlen ) {
3506 NEXT_OFF( convert ) = (U16)(tail - convert);
3507 ARG_SET( convert, data_slot );
3508 /* Store the offset to the first unabsorbed branch in
3509 jump[0], which is otherwise unused by the jump logic.
3510 We use this when dumping a trie and during optimisation. */
3512 trie->jump[0] = (U16)(nextbranch - convert);
3514 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3515 * and there is a bitmap
3516 * and the first "jump target" node we found leaves enough room
3517 * then convert the TRIE node into a TRIEC node, with the bitmap
3518 * embedded inline in the opcode - this is hypothetically faster.
3520 if ( !trie->states[trie->startstate].wordnum
3522 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3524 OP( convert ) = TRIEC;
3525 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3526 PerlMemShared_free(trie->bitmap);
3529 OP( convert ) = TRIE;
3531 /* store the type in the flags */
3532 convert->flags = nodetype;
3536 + regarglen[ OP( convert ) ];
3538 /* XXX We really should free up the resource in trie now,
3539 as we won't use them - (which resources?) dmq */
3541 /* needed for dumping*/
3542 DEBUG_r(if (optimize) {
3543 regnode *opt = convert;
3545 while ( ++opt < optimize) {
3546 Set_Node_Offset_Length(opt, 0, 0);
3549 Try to clean up some of the debris left after the
3552 while( optimize < jumper ) {
3553 #ifdef RE_TRACK_PATTERN_OFFSETS
3554 mjd_nodelen += Node_Length((optimize));
3556 OP( optimize ) = OPTIMIZED;
3557 Set_Node_Offset_Length(optimize, 0, 0);
3560 Set_Node_Offset_Length(convert, mjd_offset, mjd_nodelen);
3562 } /* end node insert */
3564 /* Finish populating the prev field of the wordinfo array. Walk back
3565 * from each accept state until we find another accept state, and if
3566 * so, point the first word's .prev field at the second word. If the
3567 * second already has a .prev field set, stop now. This will be the
3568 * case either if we've already processed that word's accept state,
3569 * or that state had multiple words, and the overspill words were
3570 * already linked up earlier.
3577 for (word=1; word <= trie->wordcount; word++) {
3579 if (trie->wordinfo[word].prev)
3581 state = trie->wordinfo[word].accept;
3583 state = prev_states[state];
3586 prev = trie->states[state].wordnum;
3590 trie->wordinfo[word].prev = prev;
3592 Safefree(prev_states);
3596 /* and now dump out the compressed format */
3597 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3599 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3601 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3602 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3604 SvREFCNT_dec_NN(revcharmap);
3608 : trie->startstate>1
3614 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3616 /* The Trie is constructed and compressed now so we can build a fail array if
3619 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3621 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3625 We find the fail state for each state in the trie, this state is the longest
3626 proper suffix of the current state's 'word' that is also a proper prefix of
3627 another word in our trie. State 1 represents the word '' and is thus the
3628 default fail state. This allows the DFA not to have to restart after its
3629 tried and failed a word at a given point, it simply continues as though it
3630 had been matching the other word in the first place.
3632 'abcdgu'=~/abcdefg|cdgu/
3633 When we get to 'd' we are still matching the first word, we would encounter
3634 'g' which would fail, which would bring us to the state representing 'd' in
3635 the second word where we would try 'g' and succeed, proceeding to match
3638 /* add a fail transition */
3639 const U32 trie_offset = ARG(source);
3640 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3642 const U32 ucharcount = trie->uniquecharcount;
3643 const U32 numstates = trie->statecount;
3644 const U32 ubound = trie->lasttrans + ucharcount;
3648 U32 base = trie->states[ 1 ].trans.base;
3651 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3653 GET_RE_DEBUG_FLAGS_DECL;
3655 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3656 PERL_UNUSED_CONTEXT;
3658 PERL_UNUSED_ARG(depth);
3661 if ( OP(source) == TRIE ) {
3662 struct regnode_1 *op = (struct regnode_1 *)
3663 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3664 StructCopy(source, op, struct regnode_1);
3665 stclass = (regnode *)op;
3667 struct regnode_charclass *op = (struct regnode_charclass *)
3668 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3669 StructCopy(source, op, struct regnode_charclass);
3670 stclass = (regnode *)op;
3672 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3674 ARG_SET( stclass, data_slot );
3675 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3676 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3677 aho->trie=trie_offset;
3678 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3679 Copy( trie->states, aho->states, numstates, reg_trie_state );
3680 Newx( q, numstates, U32);
3681 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3684 /* initialize fail[0..1] to be 1 so that we always have
3685 a valid final fail state */
3686 fail[ 0 ] = fail[ 1 ] = 1;
3688 for ( charid = 0; charid < ucharcount ; charid++ ) {
3689 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3691 q[ q_write ] = newstate;
3692 /* set to point at the root */
3693 fail[ q[ q_write++ ] ]=1;
3696 while ( q_read < q_write) {
3697 const U32 cur = q[ q_read++ % numstates ];
3698 base = trie->states[ cur ].trans.base;
3700 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3701 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3703 U32 fail_state = cur;
3706 fail_state = fail[ fail_state ];
3707 fail_base = aho->states[ fail_state ].trans.base;
3708 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3710 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3711 fail[ ch_state ] = fail_state;
3712 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3714 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3716 q[ q_write++ % numstates] = ch_state;
3720 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3721 when we fail in state 1, this allows us to use the
3722 charclass scan to find a valid start char. This is based on the principle
3723 that theres a good chance the string being searched contains lots of stuff
3724 that cant be a start char.
3726 fail[ 0 ] = fail[ 1 ] = 0;
3727 DEBUG_TRIE_COMPILE_r({
3728 Perl_re_indentf( aTHX_ "Stclass Failtable (%" UVuf " states): 0",
3729 depth, (UV)numstates
3731 for( q_read=1; q_read<numstates; q_read++ ) {
3732 Perl_re_printf( aTHX_ ", %" UVuf, (UV)fail[q_read]);
3734 Perl_re_printf( aTHX_ "\n");
3737 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3742 /* The below joins as many adjacent EXACTish nodes as possible into a single
3743 * one. The regop may be changed if the node(s) contain certain sequences that
3744 * require special handling. The joining is only done if:
3745 * 1) there is room in the current conglomerated node to entirely contain the
3747 * 2) they are the exact same node type
3749 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3750 * these get optimized out
3752 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3753 * as possible, even if that means splitting an existing node so that its first
3754 * part is moved to the preceeding node. This would maximise the efficiency of
3755 * memEQ during matching.
3757 * If a node is to match under /i (folded), the number of characters it matches
3758 * can be different than its character length if it contains a multi-character
3759 * fold. *min_subtract is set to the total delta number of characters of the
3762 * And *unfolded_multi_char is set to indicate whether or not the node contains
3763 * an unfolded multi-char fold. This happens when it won't be known until
3764 * runtime whether the fold is valid or not; namely
3765 * 1) for EXACTF nodes that contain LATIN SMALL LETTER SHARP S, as only if the
3766 * target string being matched against turns out to be UTF-8 is that fold
3768 * 2) for EXACTFL nodes whose folding rules depend on the locale in force at
3770 * (Multi-char folds whose components are all above the Latin1 range are not
3771 * run-time locale dependent, and have already been folded by the time this
3772 * function is called.)
3774 * This is as good a place as any to discuss the design of handling these
3775 * multi-character fold sequences. It's been wrong in Perl for a very long
3776 * time. There are three code points in Unicode whose multi-character folds
3777 * were long ago discovered to mess things up. The previous designs for
3778 * dealing with these involved assigning a special node for them. This
3779 * approach doesn't always work, as evidenced by this example:
3780 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3781 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3782 * would match just the \xDF, it won't be able to handle the case where a
3783 * successful match would have to cross the node's boundary. The new approach
3784 * that hopefully generally solves the problem generates an EXACTFU_SS node
3785 * that is "sss" in this case.
3787 * It turns out that there are problems with all multi-character folds, and not
3788 * just these three. Now the code is general, for all such cases. The
3789 * approach taken is:
3790 * 1) This routine examines each EXACTFish node that could contain multi-
3791 * character folded sequences. Since a single character can fold into
3792 * such a sequence, the minimum match length for this node is less than
3793 * the number of characters in the node. This routine returns in
3794 * *min_subtract how many characters to subtract from the the actual
3795 * length of the string to get a real minimum match length; it is 0 if
3796 * there are no multi-char foldeds. This delta is used by the caller to
3797 * adjust the min length of the match, and the delta between min and max,
3798 * so that the optimizer doesn't reject these possibilities based on size
3800 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3801 * is used for an EXACTFU node that contains at least one "ss" sequence in
3802 * it. For non-UTF-8 patterns and strings, this is the only case where
3803 * there is a possible fold length change. That means that a regular
3804 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3805 * with length changes, and so can be processed faster. regexec.c takes
3806 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3807 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3808 * known until runtime). This saves effort in regex matching. However,
3809 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3810 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3811 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3812 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3813 * possibilities for the non-UTF8 patterns are quite simple, except for
3814 * the sharp s. All the ones that don't involve a UTF-8 target string are
3815 * members of a fold-pair, and arrays are set up for all of them so that
3816 * the other member of the pair can be found quickly. Code elsewhere in
3817 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3818 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3819 * described in the next item.
3820 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3821 * validity of the fold won't be known until runtime, and so must remain
3822 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFAA
3823 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3824 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3825 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3826 * The reason this is a problem is that the optimizer part of regexec.c
3827 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3828 * that a character in the pattern corresponds to at most a single
3829 * character in the target string. (And I do mean character, and not byte
3830 * here, unlike other parts of the documentation that have never been
3831 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3832 * EXACTFL nodes can match the two character string 'ss'; in EXACTFAA
3833 * nodes it can match "\x{17F}\x{17F}". These, along with other ones in
3834 * EXACTFL nodes, violate the assumption, and they are the only instances
3835 * where it is violated. I'm reluctant to try to change the assumption,
3836 * as the code involved is impenetrable to me (khw), so instead the code
3837 * here punts. This routine examines EXACTFL nodes, and (when the pattern
3838 * isn't UTF-8) EXACTF and EXACTFAA for such unfolded folds, and returns a
3839 * boolean indicating whether or not the node contains such a fold. When
3840 * it is true, the caller sets a flag that later causes the optimizer in
3841 * this file to not set values for the floating and fixed string lengths,
3842 * and thus avoids the optimizer code in regexec.c that makes the invalid
3843 * assumption. Thus, there is no optimization based on string lengths for
3844 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3845 * EXACTF and EXACTFAA nodes that contain the sharp s. (The reason the
3846 * assumption is wrong only in these cases is that all other non-UTF-8
3847 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3848 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3849 * EXACTF nodes because we don't know at compile time if it actually
3850 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3851 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3852 * always matches; and EXACTFAA where it never does. In an EXACTFAA node
3853 * in a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3854 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3855 * string would require the pattern to be forced into UTF-8, the overhead
3856 * of which we want to avoid. Similarly the unfolded multi-char folds in
3857 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3860 * Similarly, the code that generates tries doesn't currently handle
3861 * not-already-folded multi-char folds, and it looks like a pain to change
3862 * that. Therefore, trie generation of EXACTFAA nodes with the sharp s
3863 * doesn't work. Instead, such an EXACTFAA is turned into a new regnode,
3864 * EXACTFAA_NO_TRIE, which the trie code knows not to handle. Most people
3865 * using /iaa matching will be doing so almost entirely with ASCII
3866 * strings, so this should rarely be encountered in practice */
3868 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3869 if (PL_regkind[OP(scan)] == EXACT) \
3870 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags), NULL, depth+1)
3873 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3874 UV *min_subtract, bool *unfolded_multi_char,
3875 U32 flags, regnode *val, U32 depth)
3877 /* Merge several consecutive EXACTish nodes into one. */
3878 regnode *n = regnext(scan);
3880 regnode *next = scan + NODE_SZ_STR(scan);
3884 regnode *stop = scan;
3885 GET_RE_DEBUG_FLAGS_DECL;
3887 PERL_UNUSED_ARG(depth);
3890 PERL_ARGS_ASSERT_JOIN_EXACT;
3891 #ifndef EXPERIMENTAL_INPLACESCAN
3892 PERL_UNUSED_ARG(flags);
3893 PERL_UNUSED_ARG(val);
3895 DEBUG_PEEP("join", scan, depth, 0);
3897 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3898 * EXACT ones that are mergeable to the current one. */
3900 && (PL_regkind[OP(n)] == NOTHING
3901 || (stringok && OP(n) == OP(scan)))
3903 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3906 if (OP(n) == TAIL || n > next)
3908 if (PL_regkind[OP(n)] == NOTHING) {
3909 DEBUG_PEEP("skip:", n, depth, 0);
3910 NEXT_OFF(scan) += NEXT_OFF(n);
3911 next = n + NODE_STEP_REGNODE;
3918 else if (stringok) {
3919 const unsigned int oldl = STR_LEN(scan);
3920 regnode * const nnext = regnext(n);
3922 /* XXX I (khw) kind of doubt that this works on platforms (should
3923 * Perl ever run on one) where U8_MAX is above 255 because of lots
3924 * of other assumptions */
3925 /* Don't join if the sum can't fit into a single node */
3926 if (oldl + STR_LEN(n) > U8_MAX)
3929 DEBUG_PEEP("merg", n, depth, 0);
3932 NEXT_OFF(scan) += NEXT_OFF(n);
3933 STR_LEN(scan) += STR_LEN(n);
3934 next = n + NODE_SZ_STR(n);
3935 /* Now we can overwrite *n : */
3936 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3944 #ifdef EXPERIMENTAL_INPLACESCAN
3945 if (flags && !NEXT_OFF(n)) {
3946 DEBUG_PEEP("atch", val, depth, 0);
3947 if (reg_off_by_arg[OP(n)]) {
3948 ARG_SET(n, val - n);
3951 NEXT_OFF(n) = val - n;
3959 *unfolded_multi_char = FALSE;
3961 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3962 * can now analyze for sequences of problematic code points. (Prior to
3963 * this final joining, sequences could have been split over boundaries, and
3964 * hence missed). The sequences only happen in folding, hence for any
3965 * non-EXACT EXACTish node */
3966 if (OP(scan) != EXACT && OP(scan) != EXACTL) {
3967 U8* s0 = (U8*) STRING(scan);
3969 U8* s_end = s0 + STR_LEN(scan);
3971 int total_count_delta = 0; /* Total delta number of characters that
3972 multi-char folds expand to */
3974 /* One pass is made over the node's string looking for all the
3975 * possibilities. To avoid some tests in the loop, there are two main
3976 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3981 if (OP(scan) == EXACTFL) {
3984 /* An EXACTFL node would already have been changed to another
3985 * node type unless there is at least one character in it that
3986 * is problematic; likely a character whose fold definition
3987 * won't be known until runtime, and so has yet to be folded.
3988 * For all but the UTF-8 locale, folds are 1-1 in length, but
3989 * to handle the UTF-8 case, we need to create a temporary
3990 * folded copy using UTF-8 locale rules in order to analyze it.
3991 * This is because our macros that look to see if a sequence is
3992 * a multi-char fold assume everything is folded (otherwise the
3993 * tests in those macros would be too complicated and slow).
3994 * Note that here, the non-problematic folds will have already
3995 * been done, so we can just copy such characters. We actually
3996 * don't completely fold the EXACTFL string. We skip the
3997 * unfolded multi-char folds, as that would just create work
3998 * below to figure out the size they already are */
4000 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
4003 STRLEN s_len = UTF8SKIP(s);
4004 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
4005 Copy(s, d, s_len, U8);
4008 else if (is_FOLDS_TO_MULTI_utf8(s)) {
4009 *unfolded_multi_char = TRUE;
4010 Copy(s, d, s_len, U8);
4013 else if (isASCII(*s)) {
4014 *(d++) = toFOLD(*s);
4018 _toFOLD_utf8_flags(s, s_end, d, &len, FOLD_FLAGS_FULL);
https://perl5.git.perl.org / perl5.git / blob