5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
10 /* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
12 * a regular expression.
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
20 /* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
24 /* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
29 /* The names of the functions have been changed from regcomp and
30 * regexec to pregcomp and pregexec in order to avoid conflicts
31 * with the POSIX routines of the same names.
34 #ifdef PERL_EXT_RE_BUILD
39 * pregcomp and pregexec -- regsub and regerror are not used in perl
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
59 **** Alterations to Henry's code are...
61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
74 #define PERL_IN_REGCOMP_C
78 #ifdef PERL_IN_XSUB_RE
80 EXTERN_C const struct regexp_engine my_reg_engine;
85 #include "dquote_inline.h"
86 #include "invlist_inline.h"
87 #include "unicode_constants.h"
89 #define HAS_NONLATIN1_FOLD_CLOSURE(i) \
90 _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
91 #define HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(i) \
92 _HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
93 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
94 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
100 /* this is a chain of data about sub patterns we are processing that
101 need to be handled separately/specially in study_chunk. Its so
102 we can simulate recursion without losing state. */
104 typedef struct scan_frame {
105 regnode *last_regnode; /* last node to process in this frame */
106 regnode *next_regnode; /* next node to process when last is reached */
107 U32 prev_recursed_depth;
108 I32 stopparen; /* what stopparen do we use */
110 struct scan_frame *this_prev_frame; /* this previous frame */
111 struct scan_frame *prev_frame; /* previous frame */
112 struct scan_frame *next_frame; /* next frame */
115 /* Certain characters are output as a sequence with the first being a
117 #define isBACKSLASHED_PUNCT(c) strchr("-[]\\^", c)
120 struct RExC_state_t {
121 U32 flags; /* RXf_* are we folding, multilining? */
122 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
123 char *precomp; /* uncompiled string. */
124 char *precomp_end; /* pointer to end of uncompiled string. */
125 REGEXP *rx_sv; /* The SV that is the regexp. */
126 regexp *rx; /* perl core regexp structure */
127 regexp_internal *rxi; /* internal data for regexp object
129 char *start; /* Start of input for compile */
130 char *end; /* End of input for compile */
131 char *parse; /* Input-scan pointer. */
132 char *copy_start; /* start of copy of input within
133 constructed parse string */
134 char *copy_start_in_input; /* Position in input string
135 corresponding to copy_start */
136 SSize_t whilem_seen; /* number of WHILEM in this expr */
137 regnode *emit_start; /* Start of emitted-code area */
138 regnode_offset emit; /* Code-emit pointer */
139 I32 naughty; /* How bad is this pattern? */
140 I32 sawback; /* Did we see \1, ...? */
142 SSize_t size; /* Number of regnode equivalents in
145 /* position beyond 'precomp' of the warning message furthest away from
146 * 'precomp'. During the parse, no warnings are raised for any problems
147 * earlier in the parse than this position. This works if warnings are
148 * raised the first time a given spot is parsed, and if only one
149 * independent warning is raised for any given spot */
150 Size_t latest_warn_offset;
152 I32 npar; /* Capture buffer count so far in the
153 parse, (OPEN) plus one. ("par" 0 is
155 I32 total_par; /* During initial parse, is either 0,
156 or -1; the latter indicating a
157 reparse is needed. After that pass,
158 it is what 'npar' became after the
159 pass. Hence, it being > 0 indicates
160 we are in a reparse situation */
161 I32 nestroot; /* root parens we are in - used by
164 regnode_offset *open_parens; /* offsets to open parens */
165 regnode_offset *close_parens; /* offsets to close parens */
166 regnode *end_op; /* END node in program */
167 I32 utf8; /* whether the pattern is utf8 or not */
168 I32 orig_utf8; /* whether the pattern was originally in utf8 */
169 /* XXX use this for future optimisation of case
170 * where pattern must be upgraded to utf8. */
171 I32 uni_semantics; /* If a d charset modifier should use unicode
172 rules, even if the pattern is not in
174 HV *paren_names; /* Paren names */
176 regnode **recurse; /* Recurse regops */
177 I32 recurse_count; /* Number of recurse regops we have generated */
178 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
180 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
183 I32 override_recoding;
185 I32 recode_x_to_native;
187 I32 in_multi_char_class;
188 struct reg_code_blocks *code_blocks;/* positions of literal (?{})
190 int code_index; /* next code_blocks[] slot */
191 SSize_t maxlen; /* mininum possible number of chars in string to match */
192 scan_frame *frame_head;
193 scan_frame *frame_last;
197 #ifdef ADD_TO_REGEXEC
198 char *starttry; /* -Dr: where regtry was called. */
199 #define RExC_starttry (pRExC_state->starttry)
201 SV *runtime_code_qr; /* qr with the runtime code blocks */
203 const char *lastparse;
205 AV *paren_name_list; /* idx -> name */
206 U32 study_chunk_recursed_count;
210 #define RExC_lastparse (pRExC_state->lastparse)
211 #define RExC_lastnum (pRExC_state->lastnum)
212 #define RExC_paren_name_list (pRExC_state->paren_name_list)
213 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
214 #define RExC_mysv (pRExC_state->mysv1)
215 #define RExC_mysv1 (pRExC_state->mysv1)
216 #define RExC_mysv2 (pRExC_state->mysv2)
226 #define RExC_flags (pRExC_state->flags)
227 #define RExC_pm_flags (pRExC_state->pm_flags)
228 #define RExC_precomp (pRExC_state->precomp)
229 #define RExC_copy_start_in_input (pRExC_state->copy_start_in_input)
230 #define RExC_copy_start_in_constructed (pRExC_state->copy_start)
231 #define RExC_precomp_end (pRExC_state->precomp_end)
232 #define RExC_rx_sv (pRExC_state->rx_sv)
233 #define RExC_rx (pRExC_state->rx)
234 #define RExC_rxi (pRExC_state->rxi)
235 #define RExC_start (pRExC_state->start)
236 #define RExC_end (pRExC_state->end)
237 #define RExC_parse (pRExC_state->parse)
238 #define RExC_latest_warn_offset (pRExC_state->latest_warn_offset )
239 #define RExC_whilem_seen (pRExC_state->whilem_seen)
240 #define RExC_seen_d_op (pRExC_state->seen_d_op) /* Seen something that differs
241 under /d from /u ? */
244 #ifdef RE_TRACK_PATTERN_OFFSETS
245 # define RExC_offsets (RExC_rxi->u.offsets) /* I am not like the
248 #define RExC_emit (pRExC_state->emit)
249 #define RExC_emit_start (pRExC_state->emit_start)
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_total_parens (pRExC_state->total_par)
256 #define RExC_nestroot (pRExC_state->nestroot)
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)
283 #define RExC_use_BRANCHJ (pRExC_state->use_BRANCHJ)
284 #define RExC_unlexed_names (pRExC_state->unlexed_names)
286 /* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
287 * a flag to disable back-off on the fixed/floating substrings - if it's
288 * a high complexity pattern we assume the benefit of avoiding a full match
289 * is worth the cost of checking for the substrings even if they rarely help.
291 #define RExC_naughty (pRExC_state->naughty)
292 #define TOO_NAUGHTY (10)
293 #define MARK_NAUGHTY(add) \
294 if (RExC_naughty < TOO_NAUGHTY) \
295 RExC_naughty += (add)
296 #define MARK_NAUGHTY_EXP(exp, add) \
297 if (RExC_naughty < TOO_NAUGHTY) \
298 RExC_naughty += RExC_naughty / (exp) + (add)
300 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
301 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
302 ((*s) == '{' && regcurly(s)))
305 * Flags to be passed up and down.
307 #define WORST 0 /* Worst case. */
308 #define HASWIDTH 0x01 /* Known to not match null strings, could match
311 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
312 * character. (There needs to be a case: in the switch statement in regexec.c
313 * for any node marked SIMPLE.) Note that this is not the same thing as
316 #define SPSTART 0x04 /* Starts with * or + */
317 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
318 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
319 #define RESTART_PARSE 0x20 /* Need to redo the parse */
320 #define NEED_UTF8 0x40 /* In conjunction with RESTART_PARSE, need to
321 calcuate sizes as UTF-8 */
323 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
325 /* whether trie related optimizations are enabled */
326 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
327 #define TRIE_STUDY_OPT
328 #define FULL_TRIE_STUDY
334 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
335 #define PBITVAL(paren) (1 << ((paren) & 7))
336 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
337 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
338 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
340 #define REQUIRE_UTF8(flagp) STMT_START { \
342 *flagp = RESTART_PARSE|NEED_UTF8; \
347 /* Change from /d into /u rules, and restart the parse. RExC_uni_semantics is
348 * a flag that indicates we need to override /d with /u as a result of
349 * something in the pattern. It should only be used in regards to calling
350 * set_regex_charset() or get_regex_charse() */
351 #define REQUIRE_UNI_RULES(flagp, restart_retval) \
353 if (DEPENDS_SEMANTICS) { \
354 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
355 RExC_uni_semantics = 1; \
356 if (RExC_seen_d_op && LIKELY(! IN_PARENS_PASS)) { \
357 /* No need to restart the parse if we haven't seen \
358 * anything that differs between /u and /d, and no need \
359 * to restart immediately if we're going to reparse \
360 * anyway to count parens */ \
361 *flagp |= RESTART_PARSE; \
362 return restart_retval; \
367 #define BRANCH_MAX_OFFSET U16_MAX
368 #define REQUIRE_BRANCHJ(flagp, restart_retval) \
370 RExC_use_BRANCHJ = 1; \
371 if (LIKELY(! IN_PARENS_PASS)) { \
372 /* No need to restart the parse immediately if we're \
373 * going to reparse anyway to count parens */ \
374 *flagp |= RESTART_PARSE; \
375 return restart_retval; \
379 /* Until we have completed the parse, we leave RExC_total_parens at 0 or
380 * less. After that, it must always be positive, because the whole re is
381 * considered to be surrounded by virtual parens. Setting it to negative
382 * indicates there is some construct that needs to know the actual number of
383 * parens to be properly handled. And that means an extra pass will be
384 * required after we've counted them all */
385 #define ALL_PARENS_COUNTED (RExC_total_parens > 0)
386 #define REQUIRE_PARENS_PASS \
387 STMT_START { /* No-op if have completed a pass */ \
388 if (! ALL_PARENS_COUNTED) RExC_total_parens = -1; \
390 #define IN_PARENS_PASS (RExC_total_parens < 0)
393 /* This is used to return failure (zero) early from the calling function if
394 * various flags in 'flags' are set. Two flags always cause a return:
395 * 'RESTART_PARSE' and 'NEED_UTF8'. 'extra' can be used to specify any
396 * additional flags that should cause a return; 0 if none. If the return will
397 * be done, '*flagp' is first set to be all of the flags that caused the
399 #define RETURN_FAIL_ON_RESTART_OR_FLAGS(flags,flagp,extra) \
401 if ((flags) & (RESTART_PARSE|NEED_UTF8|(extra))) { \
402 *(flagp) = (flags) & (RESTART_PARSE|NEED_UTF8|(extra)); \
407 #define MUST_RESTART(flags) ((flags) & (RESTART_PARSE))
409 #define RETURN_FAIL_ON_RESTART(flags,flagp) \
410 RETURN_FAIL_ON_RESTART_OR_FLAGS( flags, flagp, 0)
411 #define RETURN_FAIL_ON_RESTART_FLAGP(flagp) \
412 if (MUST_RESTART(*(flagp))) return 0
414 /* This converts the named class defined in regcomp.h to its equivalent class
415 * number defined in handy.h. */
416 #define namedclass_to_classnum(class) ((int) ((class) / 2))
417 #define classnum_to_namedclass(classnum) ((classnum) * 2)
419 #define _invlist_union_complement_2nd(a, b, output) \
420 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
421 #define _invlist_intersection_complement_2nd(a, b, output) \
422 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
424 /* About scan_data_t.
426 During optimisation we recurse through the regexp program performing
427 various inplace (keyhole style) optimisations. In addition study_chunk
428 and scan_commit populate this data structure with information about
429 what strings MUST appear in the pattern. We look for the longest
430 string that must appear at a fixed location, and we look for the
431 longest string that may appear at a floating location. So for instance
436 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
437 strings (because they follow a .* construct). study_chunk will identify
438 both FOO and BAR as being the longest fixed and floating strings respectively.
440 The strings can be composites, for instance
444 will result in a composite fixed substring 'foo'.
446 For each string some basic information is maintained:
449 This is the position the string must appear at, or not before.
450 It also implicitly (when combined with minlenp) tells us how many
451 characters must match before the string we are searching for.
452 Likewise when combined with minlenp and the length of the string it
453 tells us how many characters must appear after the string we have
457 Only used for floating strings. This is the rightmost point that
458 the string can appear at. If set to SSize_t_MAX it indicates that the
459 string can occur infinitely far to the right.
460 For fixed strings, it is equal to min_offset.
463 A pointer to the minimum number of characters of the pattern that the
464 string was found inside. This is important as in the case of positive
465 lookahead or positive lookbehind we can have multiple patterns
470 The minimum length of the pattern overall is 3, the minimum length
471 of the lookahead part is 3, but the minimum length of the part that
472 will actually match is 1. So 'FOO's minimum length is 3, but the
473 minimum length for the F is 1. This is important as the minimum length
474 is used to determine offsets in front of and behind the string being
475 looked for. Since strings can be composites this is the length of the
476 pattern at the time it was committed with a scan_commit. Note that
477 the length is calculated by study_chunk, so that the minimum lengths
478 are not known until the full pattern has been compiled, thus the
479 pointer to the value.
483 In the case of lookbehind the string being searched for can be
484 offset past the start point of the final matching string.
485 If this value was just blithely removed from the min_offset it would
486 invalidate some of the calculations for how many chars must match
487 before or after (as they are derived from min_offset and minlen and
488 the length of the string being searched for).
489 When the final pattern is compiled and the data is moved from the
490 scan_data_t structure into the regexp structure the information
491 about lookbehind is factored in, with the information that would
492 have been lost precalculated in the end_shift field for the
495 The fields pos_min and pos_delta are used to store the minimum offset
496 and the delta to the maximum offset at the current point in the pattern.
500 struct scan_data_substrs {
501 SV *str; /* longest substring found in pattern */
502 SSize_t min_offset; /* earliest point in string it can appear */
503 SSize_t max_offset; /* latest point in string it can appear */
504 SSize_t *minlenp; /* pointer to the minlen relevant to the string */
505 SSize_t lookbehind; /* is the pos of the string modified by LB */
506 I32 flags; /* per substring SF_* and SCF_* flags */
509 typedef struct scan_data_t {
510 /*I32 len_min; unused */
511 /*I32 len_delta; unused */
515 SSize_t last_end; /* min value, <0 unless valid. */
516 SSize_t last_start_min;
517 SSize_t last_start_max;
518 U8 cur_is_floating; /* whether the last_* values should be set as
519 * the next fixed (0) or floating (1)
522 /* [0] is longest fixed substring so far, [1] is longest float so far */
523 struct scan_data_substrs substrs[2];
525 I32 flags; /* common SF_* and SCF_* flags */
527 SSize_t *last_closep;
528 regnode_ssc *start_class;
532 * Forward declarations for pregcomp()'s friends.
535 static const scan_data_t zero_scan_data = {
536 0, 0, NULL, 0, 0, 0, 0,
538 { NULL, 0, 0, 0, 0, 0 },
539 { NULL, 0, 0, 0, 0, 0 },
546 #define SF_BEFORE_SEOL 0x0001
547 #define SF_BEFORE_MEOL 0x0002
548 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
550 #define SF_IS_INF 0x0040
551 #define SF_HAS_PAR 0x0080
552 #define SF_IN_PAR 0x0100
553 #define SF_HAS_EVAL 0x0200
556 /* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
557 * longest substring in the pattern. When it is not set the optimiser keeps
558 * track of position, but does not keep track of the actual strings seen,
560 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
563 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
564 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
565 * turned off because of the alternation (BRANCH). */
566 #define SCF_DO_SUBSTR 0x0400
568 #define SCF_DO_STCLASS_AND 0x0800
569 #define SCF_DO_STCLASS_OR 0x1000
570 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
571 #define SCF_WHILEM_VISITED_POS 0x2000
573 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
574 #define SCF_SEEN_ACCEPT 0x8000
575 #define SCF_TRIE_DOING_RESTUDY 0x10000
576 #define SCF_IN_DEFINE 0x20000
581 #define UTF cBOOL(RExC_utf8)
583 /* The enums for all these are ordered so things work out correctly */
584 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
585 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
586 == REGEX_DEPENDS_CHARSET)
587 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
588 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
589 >= REGEX_UNICODE_CHARSET)
590 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
591 == REGEX_ASCII_RESTRICTED_CHARSET)
592 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
593 >= REGEX_ASCII_RESTRICTED_CHARSET)
594 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
595 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
597 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
599 /* For programs that want to be strictly Unicode compatible by dying if any
600 * attempt is made to match a non-Unicode code point against a Unicode
602 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
604 #define OOB_NAMEDCLASS -1
606 /* There is no code point that is out-of-bounds, so this is problematic. But
607 * its only current use is to initialize a variable that is always set before
609 #define OOB_UNICODE 0xDEADBEEF
611 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
614 /* length of regex to show in messages that don't mark a position within */
615 #define RegexLengthToShowInErrorMessages 127
618 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
619 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
620 * op/pragma/warn/regcomp.
622 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
623 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
625 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
626 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
628 /* The code in this file in places uses one level of recursion with parsing
629 * rebased to an alternate string constructed by us in memory. This can take
630 * the form of something that is completely different from the input, or
631 * something that uses the input as part of the alternate. In the first case,
632 * there should be no possibility of an error, as we are in complete control of
633 * the alternate string. But in the second case we don't completely control
634 * the input portion, so there may be errors in that. Here's an example:
636 * is handled specially because \x{df} folds to a sequence of more than one
637 * character: 'ss'. What is done is to create and parse an alternate string,
638 * which looks like this:
639 * /(?:\x{DF}|[abc\x{DF}def])/ui
640 * where it uses the input unchanged in the middle of something it constructs,
641 * which is a branch for the DF outside the character class, and clustering
642 * parens around the whole thing. (It knows enough to skip the DF inside the
643 * class while in this substitute parse.) 'abc' and 'def' may have errors that
644 * need to be reported. The general situation looks like this:
646 * |<------- identical ------>|
648 * Input: ---------------------------------------------------------------
649 * Constructed: ---------------------------------------------------
651 * |<------- identical ------>|
653 * sI..eI is the portion of the input pattern we are concerned with here.
654 * sC..EC is the constructed substitute parse string.
655 * sC..tC is constructed by us
656 * tC..eC is an exact duplicate of the portion of the input pattern tI..eI.
657 * In the diagram, these are vertically aligned.
658 * eC..EC is also constructed by us.
659 * xC is the position in the substitute parse string where we found a
661 * xI is the position in the original pattern corresponding to xC.
663 * We want to display a message showing the real input string. Thus we need to
664 * translate from xC to xI. We know that xC >= tC, since the portion of the
665 * string sC..tC has been constructed by us, and so shouldn't have errors. We
667 * xI = tI + (xC - tC)
669 * When the substitute parse is constructed, the code needs to set:
672 * RExC_copy_start_in_input (tI)
673 * RExC_copy_start_in_constructed (tC)
674 * and restore them when done.
676 * During normal processing of the input pattern, both
677 * 'RExC_copy_start_in_input' and 'RExC_copy_start_in_constructed' are set to
678 * sI, so that xC equals xI.
681 #define sI RExC_precomp
682 #define eI RExC_precomp_end
683 #define sC RExC_start
685 #define tI RExC_copy_start_in_input
686 #define tC RExC_copy_start_in_constructed
687 #define xI(xC) (tI + (xC - tC))
688 #define xI_offset(xC) (xI(xC) - sI)
690 #define REPORT_LOCATION_ARGS(xC) \
692 (xI(xC) > eI) /* Don't run off end */ \
693 ? eI - sI /* Length before the <--HERE */ \
694 : ((xI_offset(xC) >= 0) \
696 : (Perl_croak(aTHX_ "panic: %s: %d: negative offset: %" \
697 IVdf " trying to output message for " \
699 __FILE__, __LINE__, (IV) xI_offset(xC), \
700 ((int) (eC - sC)), sC), 0)), \
701 sI), /* The input pattern printed up to the <--HERE */ \
703 (xI(xC) > eI) ? 0 : eI - xI(xC), /* Length after <--HERE */ \
704 (xI(xC) > eI) ? eI : xI(xC)) /* pattern after <--HERE */
706 /* Used to point after bad bytes for an error message, but avoid skipping
707 * past a nul byte. */
708 #define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
710 /* Set up to clean up after our imminent demise */
711 #define PREPARE_TO_DIE \
714 SAVEFREESV(RExC_rx_sv); \
715 if (RExC_open_parens) \
716 SAVEFREEPV(RExC_open_parens); \
717 if (RExC_close_parens) \
718 SAVEFREEPV(RExC_close_parens); \
722 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
723 * arg. Show regex, up to a maximum length. If it's too long, chop and add
726 #define _FAIL(code) STMT_START { \
727 const char *ellipses = ""; \
728 IV len = RExC_precomp_end - RExC_precomp; \
731 if (len > RegexLengthToShowInErrorMessages) { \
732 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
733 len = RegexLengthToShowInErrorMessages - 10; \
739 #define FAIL(msg) _FAIL( \
740 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
741 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
743 #define FAIL2(msg,arg) _FAIL( \
744 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
745 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
748 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
750 #define Simple_vFAIL(m) STMT_START { \
751 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
752 m, REPORT_LOCATION_ARGS(RExC_parse)); \
756 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
758 #define vFAIL(m) STMT_START { \
764 * Like Simple_vFAIL(), but accepts two arguments.
766 #define Simple_vFAIL2(m,a1) STMT_START { \
767 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
768 REPORT_LOCATION_ARGS(RExC_parse)); \
772 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
774 #define vFAIL2(m,a1) STMT_START { \
776 Simple_vFAIL2(m, a1); \
781 * Like Simple_vFAIL(), but accepts three arguments.
783 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
784 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
785 REPORT_LOCATION_ARGS(RExC_parse)); \
789 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
791 #define vFAIL3(m,a1,a2) STMT_START { \
793 Simple_vFAIL3(m, a1, a2); \
797 * Like Simple_vFAIL(), but accepts four arguments.
799 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
800 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
801 REPORT_LOCATION_ARGS(RExC_parse)); \
804 #define vFAIL4(m,a1,a2,a3) STMT_START { \
806 Simple_vFAIL4(m, a1, a2, a3); \
809 /* A specialized version of vFAIL2 that works with UTF8f */
810 #define vFAIL2utf8f(m, a1) STMT_START { \
812 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
813 REPORT_LOCATION_ARGS(RExC_parse)); \
816 #define vFAIL3utf8f(m, a1, a2) STMT_START { \
818 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
819 REPORT_LOCATION_ARGS(RExC_parse)); \
822 /* Setting this to NULL is a signal to not output warnings */
823 #define TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE RExC_copy_start_in_constructed = NULL
824 #define RESTORE_WARNINGS RExC_copy_start_in_constructed = RExC_precomp
826 /* Since a warning can be generated multiple times as the input is reparsed, we
827 * output it the first time we come to that point in the parse, but suppress it
828 * otherwise. 'RExC_copy_start_in_constructed' being NULL is a flag to not
829 * generate any warnings */
830 #define TO_OUTPUT_WARNINGS(loc) \
831 ( RExC_copy_start_in_constructed \
832 && ((xI(loc)) - RExC_precomp) > (Ptrdiff_t) RExC_latest_warn_offset)
834 /* After we've emitted a warning, we save the position in the input so we don't
836 #define UPDATE_WARNINGS_LOC(loc) \
838 if (TO_OUTPUT_WARNINGS(loc)) { \
839 RExC_latest_warn_offset = (xI(loc)) - RExC_precomp; \
843 /* 'warns' is the output of the packWARNx macro used in 'code' */
844 #define _WARN_HELPER(loc, warns, code) \
846 if (! RExC_copy_start_in_constructed) { \
847 Perl_croak( aTHX_ "panic! %s: %d: Tried to warn when none" \
848 " expected at '%s'", \
849 __FILE__, __LINE__, loc); \
851 if (TO_OUTPUT_WARNINGS(loc)) { \
855 UPDATE_WARNINGS_LOC(loc); \
859 /* m is not necessarily a "literal string", in this macro */
860 #define reg_warn_non_literal_string(loc, m) \
861 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
862 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
863 "%s" REPORT_LOCATION, \
864 m, REPORT_LOCATION_ARGS(loc)))
866 #define ckWARNreg(loc,m) \
867 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
868 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
870 REPORT_LOCATION_ARGS(loc)))
872 #define vWARN(loc, m) \
873 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
874 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
876 REPORT_LOCATION_ARGS(loc))) \
878 #define vWARN_dep(loc, m) \
879 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
880 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
882 REPORT_LOCATION_ARGS(loc)))
884 #define ckWARNdep(loc,m) \
885 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
886 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
888 REPORT_LOCATION_ARGS(loc)))
890 #define ckWARNregdep(loc,m) \
891 _WARN_HELPER(loc, packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
892 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
895 REPORT_LOCATION_ARGS(loc)))
897 #define ckWARN2reg_d(loc,m, a1) \
898 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
899 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
901 a1, REPORT_LOCATION_ARGS(loc)))
903 #define ckWARN2reg(loc, m, a1) \
904 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
905 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
907 a1, REPORT_LOCATION_ARGS(loc)))
909 #define vWARN3(loc, m, a1, a2) \
910 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
911 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
913 a1, a2, REPORT_LOCATION_ARGS(loc)))
915 #define ckWARN3reg(loc, m, a1, a2) \
916 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
917 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
920 REPORT_LOCATION_ARGS(loc)))
922 #define vWARN4(loc, m, a1, a2, a3) \
923 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
924 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
927 REPORT_LOCATION_ARGS(loc)))
929 #define ckWARN4reg(loc, m, a1, a2, a3) \
930 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
931 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
934 REPORT_LOCATION_ARGS(loc)))
936 #define vWARN5(loc, m, a1, a2, a3, a4) \
937 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
938 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
941 REPORT_LOCATION_ARGS(loc)))
943 #define ckWARNexperimental(loc, class, m) \
944 _WARN_HELPER(loc, packWARN(class), \
945 Perl_ck_warner_d(aTHX_ packWARN(class), \
947 REPORT_LOCATION_ARGS(loc)))
949 /* Convert between a pointer to a node and its offset from the beginning of the
951 #define REGNODE_p(offset) (RExC_emit_start + (offset))
952 #define REGNODE_OFFSET(node) ((node) - RExC_emit_start)
954 /* Macros for recording node offsets. 20001227 mjd@plover.com
955 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
956 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
957 * Element 0 holds the number n.
958 * Position is 1 indexed.
960 #ifndef RE_TRACK_PATTERN_OFFSETS
961 #define Set_Node_Offset_To_R(offset,byte)
962 #define Set_Node_Offset(node,byte)
963 #define Set_Cur_Node_Offset
964 #define Set_Node_Length_To_R(node,len)
965 #define Set_Node_Length(node,len)
966 #define Set_Node_Cur_Length(node,start)
967 #define Node_Offset(n)
968 #define Node_Length(n)
969 #define Set_Node_Offset_Length(node,offset,len)
970 #define ProgLen(ri) ri->u.proglen
971 #define SetProgLen(ri,x) ri->u.proglen = x
972 #define Track_Code(code)
974 #define ProgLen(ri) ri->u.offsets[0]
975 #define SetProgLen(ri,x) ri->u.offsets[0] = x
976 #define Set_Node_Offset_To_R(offset,byte) STMT_START { \
977 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
978 __LINE__, (int)(offset), (int)(byte))); \
980 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
983 RExC_offsets[2*(offset)-1] = (byte); \
987 #define Set_Node_Offset(node,byte) \
988 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (byte)-RExC_start)
989 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
991 #define Set_Node_Length_To_R(node,len) STMT_START { \
992 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
993 __LINE__, (int)(node), (int)(len))); \
995 Perl_croak(aTHX_ "value of node is %d in Length macro", \
998 RExC_offsets[2*(node)] = (len); \
1002 #define Set_Node_Length(node,len) \
1003 Set_Node_Length_To_R(REGNODE_OFFSET(node), len)
1004 #define Set_Node_Cur_Length(node, start) \
1005 Set_Node_Length(node, RExC_parse - start)
1007 /* Get offsets and lengths */
1008 #define Node_Offset(n) (RExC_offsets[2*(REGNODE_OFFSET(n))-1])
1009 #define Node_Length(n) (RExC_offsets[2*(REGNODE_OFFSET(n))])
1011 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
1012 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (offset)); \
1013 Set_Node_Length_To_R(REGNODE_OFFSET(node), (len)); \
1016 #define Track_Code(code) STMT_START { code } STMT_END
1019 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
1020 #define EXPERIMENTAL_INPLACESCAN
1021 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
1025 Perl_re_printf(pTHX_ const char *fmt, ...)
1029 PerlIO *f= Perl_debug_log;
1030 PERL_ARGS_ASSERT_RE_PRINTF;
1032 result = PerlIO_vprintf(f, fmt, ap);
1038 Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
1042 PerlIO *f= Perl_debug_log;
1043 PERL_ARGS_ASSERT_RE_INDENTF;
1044 va_start(ap, depth);
1045 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
1046 result = PerlIO_vprintf(f, fmt, ap);
1050 #endif /* DEBUGGING */
1052 #define DEBUG_RExC_seen() \
1053 DEBUG_OPTIMISE_MORE_r({ \
1054 Perl_re_printf( aTHX_ "RExC_seen: "); \
1056 if (RExC_seen & REG_ZERO_LEN_SEEN) \
1057 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
1059 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
1060 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
1062 if (RExC_seen & REG_GPOS_SEEN) \
1063 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
1065 if (RExC_seen & REG_RECURSE_SEEN) \
1066 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
1068 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
1069 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
1071 if (RExC_seen & REG_VERBARG_SEEN) \
1072 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
1074 if (RExC_seen & REG_CUTGROUP_SEEN) \
1075 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
1077 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
1078 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
1080 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
1081 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
1083 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
1084 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
1086 Perl_re_printf( aTHX_ "\n"); \
1089 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
1090 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
1095 S_debug_show_study_flags(pTHX_ U32 flags, const char *open_str,
1096 const char *close_str)
1101 Perl_re_printf( aTHX_ "%s", open_str);
1102 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_SEOL);
1103 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_MEOL);
1104 DEBUG_SHOW_STUDY_FLAG(flags, SF_IS_INF);
1105 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_PAR);
1106 DEBUG_SHOW_STUDY_FLAG(flags, SF_IN_PAR);
1107 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_EVAL);
1108 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_SUBSTR);
1109 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_AND);
1110 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_OR);
1111 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS);
1112 DEBUG_SHOW_STUDY_FLAG(flags, SCF_WHILEM_VISITED_POS);
1113 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_RESTUDY);
1114 DEBUG_SHOW_STUDY_FLAG(flags, SCF_SEEN_ACCEPT);
1115 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_DOING_RESTUDY);
1116 DEBUG_SHOW_STUDY_FLAG(flags, SCF_IN_DEFINE);
1117 Perl_re_printf( aTHX_ "%s", close_str);
1122 S_debug_studydata(pTHX_ const char *where, scan_data_t *data,
1123 U32 depth, int is_inf)
1125 GET_RE_DEBUG_FLAGS_DECL;
1127 DEBUG_OPTIMISE_MORE_r({
1130 Perl_re_indentf(aTHX_ "%s: Pos:%" IVdf "/%" IVdf " Flags: 0x%" UVXf,
1134 (IV)data->pos_delta,
1138 S_debug_show_study_flags(aTHX_ data->flags," [","]");
1140 Perl_re_printf( aTHX_
1141 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s",
1143 (IV)(data->last_closep ? *((data)->last_closep) : -1),
1144 is_inf ? "INF " : ""
1147 if (data->last_found) {
1149 Perl_re_printf(aTHX_
1150 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf,
1151 SvPVX_const(data->last_found),
1153 (IV)data->last_start_min,
1154 (IV)data->last_start_max
1157 for (i = 0; i < 2; i++) {
1158 Perl_re_printf(aTHX_
1159 " %s%s: '%s' @ %" IVdf "/%" IVdf,
1160 data->cur_is_floating == i ? "*" : "",
1161 i ? "Float" : "Fixed",
1162 SvPVX_const(data->substrs[i].str),
1163 (IV)data->substrs[i].min_offset,
1164 (IV)data->substrs[i].max_offset
1166 S_debug_show_study_flags(aTHX_ data->substrs[i].flags," [","]");
1170 Perl_re_printf( aTHX_ "\n");
1176 S_debug_peep(pTHX_ const char *str, const RExC_state_t *pRExC_state,
1177 regnode *scan, U32 depth, U32 flags)
1179 GET_RE_DEBUG_FLAGS_DECL;
1186 Next = regnext(scan);
1187 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
1188 Perl_re_indentf( aTHX_ "%s>%3d: %s (%d)",
1191 REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),
1192 Next ? (REG_NODE_NUM(Next)) : 0 );
1193 S_debug_show_study_flags(aTHX_ flags," [ ","]");
1194 Perl_re_printf( aTHX_ "\n");
1199 # define DEBUG_STUDYDATA(where, data, depth, is_inf) \
1200 S_debug_studydata(aTHX_ where, data, depth, is_inf)
1202 # define DEBUG_PEEP(str, scan, depth, flags) \
1203 S_debug_peep(aTHX_ str, pRExC_state, scan, depth, flags)
1206 # define DEBUG_STUDYDATA(where, data, depth, is_inf) NOOP
1207 # define DEBUG_PEEP(str, scan, depth, flags) NOOP
1211 /* =========================================================
1212 * BEGIN edit_distance stuff.
1214 * This calculates how many single character changes of any type are needed to
1215 * transform a string into another one. It is taken from version 3.1 of
1217 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1220 /* Our unsorted dictionary linked list. */
1221 /* Note we use UVs, not chars. */
1226 struct dictionary* next;
1228 typedef struct dictionary item;
1231 PERL_STATIC_INLINE item*
1232 push(UV key, item* curr)
1235 Newx(head, 1, item);
1243 PERL_STATIC_INLINE item*
1244 find(item* head, UV key)
1246 item* iterator = head;
1248 if (iterator->key == key){
1251 iterator = iterator->next;
1257 PERL_STATIC_INLINE item*
1258 uniquePush(item* head, UV key)
1260 item* iterator = head;
1263 if (iterator->key == key) {
1266 iterator = iterator->next;
1269 return push(key, head);
1272 PERL_STATIC_INLINE void
1273 dict_free(item* head)
1275 item* iterator = head;
1278 item* temp = iterator;
1279 iterator = iterator->next;
1286 /* End of Dictionary Stuff */
1288 /* All calculations/work are done here */
1290 S_edit_distance(const UV* src,
1292 const STRLEN x, /* length of src[] */
1293 const STRLEN y, /* length of tgt[] */
1294 const SSize_t maxDistance
1298 UV swapCount, swapScore, targetCharCount, i, j;
1300 UV score_ceil = x + y;
1302 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1304 /* intialize matrix start values */
1305 Newx(scores, ( (x + 2) * (y + 2)), UV);
1306 scores[0] = score_ceil;
1307 scores[1 * (y + 2) + 0] = score_ceil;
1308 scores[0 * (y + 2) + 1] = score_ceil;
1309 scores[1 * (y + 2) + 1] = 0;
1310 head = uniquePush(uniquePush(head, src[0]), tgt[0]);
1315 for (i=1;i<=x;i++) {
1317 head = uniquePush(head, src[i]);
1318 scores[(i+1) * (y + 2) + 1] = i;
1319 scores[(i+1) * (y + 2) + 0] = score_ceil;
1322 for (j=1;j<=y;j++) {
1325 head = uniquePush(head, tgt[j]);
1326 scores[1 * (y + 2) + (j + 1)] = j;
1327 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1330 targetCharCount = find(head, tgt[j-1])->value;
1331 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1333 if (src[i-1] != tgt[j-1]){
1334 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));
1338 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1342 find(head, src[i-1])->value = i;
1346 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1349 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1353 /* END of edit_distance() stuff
1354 * ========================================================= */
1356 /* is c a control character for which we have a mnemonic? */
1357 #define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
1360 S_cntrl_to_mnemonic(const U8 c)
1362 /* Returns the mnemonic string that represents character 'c', if one
1363 * exists; NULL otherwise. The only ones that exist for the purposes of
1364 * this routine are a few control characters */
1367 case '\a': return "\\a";
1368 case '\b': return "\\b";
1369 case ESC_NATIVE: return "\\e";
1370 case '\f': return "\\f";
1371 case '\n': return "\\n";
1372 case '\r': return "\\r";
1373 case '\t': return "\\t";
1379 /* Mark that we cannot extend a found fixed substring at this point.
1380 Update the longest found anchored substring or the longest found
1381 floating substrings if needed. */
1384 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1385 SSize_t *minlenp, int is_inf)
1387 const STRLEN l = CHR_SVLEN(data->last_found);
1388 SV * const longest_sv = data->substrs[data->cur_is_floating].str;
1389 const STRLEN old_l = CHR_SVLEN(longest_sv);
1390 GET_RE_DEBUG_FLAGS_DECL;
1392 PERL_ARGS_ASSERT_SCAN_COMMIT;
1394 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
1395 const U8 i = data->cur_is_floating;
1396 SvSetMagicSV(longest_sv, data->last_found);
1397 data->substrs[i].min_offset = l ? data->last_start_min : data->pos_min;
1400 data->substrs[0].max_offset = data->substrs[0].min_offset;
1402 data->substrs[1].max_offset = (l
1403 ? data->last_start_max
1404 : (data->pos_delta > SSize_t_MAX - data->pos_min
1406 : data->pos_min + data->pos_delta));
1408 || (STRLEN)data->substrs[1].max_offset > (STRLEN)SSize_t_MAX)
1409 data->substrs[1].max_offset = SSize_t_MAX;
1412 if (data->flags & SF_BEFORE_EOL)
1413 data->substrs[i].flags |= (data->flags & SF_BEFORE_EOL);
1415 data->substrs[i].flags &= ~SF_BEFORE_EOL;
1416 data->substrs[i].minlenp = minlenp;
1417 data->substrs[i].lookbehind = 0;
1420 SvCUR_set(data->last_found, 0);
1422 SV * const sv = data->last_found;
1423 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1424 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1429 data->last_end = -1;
1430 data->flags &= ~SF_BEFORE_EOL;
1431 DEBUG_STUDYDATA("commit", data, 0, is_inf);
1434 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1435 * list that describes which code points it matches */
1438 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1440 /* Set the SSC 'ssc' to match an empty string or any code point */
1442 PERL_ARGS_ASSERT_SSC_ANYTHING;
1444 assert(is_ANYOF_SYNTHETIC(ssc));
1446 /* mortalize so won't leak */
1447 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
1448 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1452 S_ssc_is_anything(const regnode_ssc *ssc)
1454 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1455 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1456 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1457 * in any way, so there's no point in using it */
1462 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1464 assert(is_ANYOF_SYNTHETIC(ssc));
1466 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1470 /* See if the list consists solely of the range 0 - Infinity */
1471 invlist_iterinit(ssc->invlist);
1472 ret = invlist_iternext(ssc->invlist, &start, &end)
1476 invlist_iterfinish(ssc->invlist);
1482 /* If e.g., both \w and \W are set, matches everything */
1483 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1485 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1486 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1496 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1498 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1499 * string, any code point, or any posix class under locale */
1501 PERL_ARGS_ASSERT_SSC_INIT;
1503 Zero(ssc, 1, regnode_ssc);
1504 set_ANYOF_SYNTHETIC(ssc);
1505 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1508 /* If any portion of the regex is to operate under locale rules that aren't
1509 * fully known at compile time, initialization includes it. The reason
1510 * this isn't done for all regexes is that the optimizer was written under
1511 * the assumption that locale was all-or-nothing. Given the complexity and
1512 * lack of documentation in the optimizer, and that there are inadequate
1513 * test cases for locale, many parts of it may not work properly, it is
1514 * safest to avoid locale unless necessary. */
1515 if (RExC_contains_locale) {
1516 ANYOF_POSIXL_SETALL(ssc);
1519 ANYOF_POSIXL_ZERO(ssc);
1524 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1525 const regnode_ssc *ssc)
1527 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1528 * to the list of code points matched, and locale posix classes; hence does
1529 * not check its flags) */
1534 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1536 assert(is_ANYOF_SYNTHETIC(ssc));
1538 invlist_iterinit(ssc->invlist);
1539 ret = invlist_iternext(ssc->invlist, &start, &end)
1543 invlist_iterfinish(ssc->invlist);
1549 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1556 #define INVLIST_INDEX 0
1557 #define ONLY_LOCALE_MATCHES_INDEX 1
1558 #define DEFERRED_USER_DEFINED_INDEX 2
1561 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1562 const regnode_charclass* const node)
1564 /* Returns a mortal inversion list defining which code points are matched
1565 * by 'node', which is of type ANYOF. Handles complementing the result if
1566 * appropriate. If some code points aren't knowable at this time, the
1567 * returned list must, and will, contain every code point that is a
1572 SV* only_utf8_locale_invlist = NULL;
1574 const U32 n = ARG(node);
1575 bool new_node_has_latin1 = FALSE;
1577 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1579 /* Look at the data structure created by S_set_ANYOF_arg() */
1580 if (n != ANYOF_ONLY_HAS_BITMAP) {
1581 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1582 AV * const av = MUTABLE_AV(SvRV(rv));
1583 SV **const ary = AvARRAY(av);
1584 assert(RExC_rxi->data->what[n] == 's');
1586 if (av_tindex_skip_len_mg(av) >= DEFERRED_USER_DEFINED_INDEX) {
1588 /* Here there are things that won't be known until runtime -- we
1589 * have to assume it could be anything */
1590 invlist = sv_2mortal(_new_invlist(1));
1591 return _add_range_to_invlist(invlist, 0, UV_MAX);
1593 else if (ary[INVLIST_INDEX]) {
1595 /* Use the node's inversion list */
1596 invlist = sv_2mortal(invlist_clone(ary[INVLIST_INDEX], NULL));
1599 /* Get the code points valid only under UTF-8 locales */
1600 if ( (ANYOF_FLAGS(node) & ANYOFL_FOLD)
1601 && av_tindex_skip_len_mg(av) >= ONLY_LOCALE_MATCHES_INDEX)
1603 only_utf8_locale_invlist = ary[ONLY_LOCALE_MATCHES_INDEX];
1608 invlist = sv_2mortal(_new_invlist(0));
1611 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1612 * code points, and an inversion list for the others, but if there are code
1613 * points that should match only conditionally on the target string being
1614 * UTF-8, those are placed in the inversion list, and not the bitmap.
1615 * Since there are circumstances under which they could match, they are
1616 * included in the SSC. But if the ANYOF node is to be inverted, we have
1617 * to exclude them here, so that when we invert below, the end result
1618 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1619 * have to do this here before we add the unconditionally matched code
1621 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1622 _invlist_intersection_complement_2nd(invlist,
1627 /* Add in the points from the bit map */
1628 if (OP(node) != ANYOFH) {
1629 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1630 if (ANYOF_BITMAP_TEST(node, i)) {
1631 unsigned int start = i++;
1633 for (; i < NUM_ANYOF_CODE_POINTS
1634 && ANYOF_BITMAP_TEST(node, i); ++i)
1638 invlist = _add_range_to_invlist(invlist, start, i-1);
1639 new_node_has_latin1 = TRUE;
1644 /* If this can match all upper Latin1 code points, have to add them
1645 * as well. But don't add them if inverting, as when that gets done below,
1646 * it would exclude all these characters, including the ones it shouldn't
1647 * that were added just above */
1648 if (! (ANYOF_FLAGS(node) & ANYOF_INVERT) && OP(node) == ANYOFD
1649 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1651 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1654 /* Similarly for these */
1655 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1656 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1659 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1660 _invlist_invert(invlist);
1662 else if (ANYOF_FLAGS(node) & ANYOFL_FOLD) {
1663 if (new_node_has_latin1) {
1665 /* Under /li, any 0-255 could fold to any other 0-255, depending on
1666 * the locale. We can skip this if there are no 0-255 at all. */
1667 _invlist_union(invlist, PL_Latin1, &invlist);
1669 invlist = add_cp_to_invlist(invlist, LATIN_SMALL_LETTER_DOTLESS_I);
1670 invlist = add_cp_to_invlist(invlist, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
1673 if (_invlist_contains_cp(invlist, LATIN_SMALL_LETTER_DOTLESS_I)) {
1674 invlist = add_cp_to_invlist(invlist, 'I');
1676 if (_invlist_contains_cp(invlist,
1677 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE))
1679 invlist = add_cp_to_invlist(invlist, 'i');
1684 /* Similarly add the UTF-8 locale possible matches. These have to be
1685 * deferred until after the non-UTF-8 locale ones are taken care of just
1686 * above, or it leads to wrong results under ANYOF_INVERT */
1687 if (only_utf8_locale_invlist) {
1688 _invlist_union_maybe_complement_2nd(invlist,
1689 only_utf8_locale_invlist,
1690 ANYOF_FLAGS(node) & ANYOF_INVERT,
1697 /* These two functions currently do the exact same thing */
1698 #define ssc_init_zero ssc_init
1700 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1701 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1703 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1704 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1705 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1708 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1709 const regnode_charclass *and_with)
1711 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1712 * another SSC or a regular ANYOF class. Can create false positives. */
1717 PERL_ARGS_ASSERT_SSC_AND;
1719 assert(is_ANYOF_SYNTHETIC(ssc));
1721 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1722 * the code point inversion list and just the relevant flags */
1723 if (is_ANYOF_SYNTHETIC(and_with)) {
1724 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1725 anded_flags = ANYOF_FLAGS(and_with);
1727 /* XXX This is a kludge around what appears to be deficiencies in the
1728 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1729 * there are paths through the optimizer where it doesn't get weeded
1730 * out when it should. And if we don't make some extra provision for
1731 * it like the code just below, it doesn't get added when it should.
1732 * This solution is to add it only when AND'ing, which is here, and
1733 * only when what is being AND'ed is the pristine, original node
1734 * matching anything. Thus it is like adding it to ssc_anything() but
1735 * only when the result is to be AND'ed. Probably the same solution
1736 * could be adopted for the same problem we have with /l matching,
1737 * which is solved differently in S_ssc_init(), and that would lead to
1738 * fewer false positives than that solution has. But if this solution
1739 * creates bugs, the consequences are only that a warning isn't raised
1740 * that should be; while the consequences for having /l bugs is
1741 * incorrect matches */
1742 if (ssc_is_anything((regnode_ssc *)and_with)) {
1743 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1747 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1748 if (OP(and_with) == ANYOFD) {
1749 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1752 anded_flags = ANYOF_FLAGS(and_with)
1753 &( ANYOF_COMMON_FLAGS
1754 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1755 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1756 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(and_with))) {
1758 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1763 ANYOF_FLAGS(ssc) &= anded_flags;
1765 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1766 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1767 * 'and_with' may be inverted. When not inverted, we have the situation of
1769 * (C1 | P1) & (C2 | P2)
1770 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1771 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1772 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1773 * <= ((C1 & C2) | P1 | P2)
1774 * Alternatively, the last few steps could be:
1775 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1776 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1777 * <= (C1 | C2 | (P1 & P2))
1778 * We favor the second approach if either P1 or P2 is non-empty. This is
1779 * because these components are a barrier to doing optimizations, as what
1780 * they match cannot be known until the moment of matching as they are
1781 * dependent on the current locale, 'AND"ing them likely will reduce or
1783 * But we can do better if we know that C1,P1 are in their initial state (a
1784 * frequent occurrence), each matching everything:
1785 * (<everything>) & (C2 | P2) = C2 | P2
1786 * Similarly, if C2,P2 are in their initial state (again a frequent
1787 * occurrence), the result is a no-op
1788 * (C1 | P1) & (<everything>) = C1 | P1
1791 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1792 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1793 * <= (C1 & ~C2) | (P1 & ~P2)
1796 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1797 && ! is_ANYOF_SYNTHETIC(and_with))
1801 ssc_intersection(ssc,
1803 FALSE /* Has already been inverted */
1806 /* If either P1 or P2 is empty, the intersection will be also; can skip
1808 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
1809 ANYOF_POSIXL_ZERO(ssc);
1811 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1813 /* Note that the Posix class component P from 'and_with' actually
1815 * P = Pa | Pb | ... | Pn
1816 * where each component is one posix class, such as in [\w\s].
1818 * ~P = ~(Pa | Pb | ... | Pn)
1819 * = ~Pa & ~Pb & ... & ~Pn
1820 * <= ~Pa | ~Pb | ... | ~Pn
1821 * The last is something we can easily calculate, but unfortunately
1822 * is likely to have many false positives. We could do better
1823 * in some (but certainly not all) instances if two classes in
1824 * P have known relationships. For example
1825 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1827 * :lower: & :print: = :lower:
1828 * And similarly for classes that must be disjoint. For example,
1829 * since \s and \w can have no elements in common based on rules in
1830 * the POSIX standard,
1831 * \w & ^\S = nothing
1832 * Unfortunately, some vendor locales do not meet the Posix
1833 * standard, in particular almost everything by Microsoft.
1834 * The loop below just changes e.g., \w into \W and vice versa */
1836 regnode_charclass_posixl temp;
1837 int add = 1; /* To calculate the index of the complement */
1839 Zero(&temp, 1, regnode_charclass_posixl);
1840 ANYOF_POSIXL_ZERO(&temp);
1841 for (i = 0; i < ANYOF_MAX; i++) {
1843 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1844 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1846 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1847 ANYOF_POSIXL_SET(&temp, i + add);
1849 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1851 ANYOF_POSIXL_AND(&temp, ssc);
1853 } /* else ssc already has no posixes */
1854 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1855 in its initial state */
1856 else if (! is_ANYOF_SYNTHETIC(and_with)
1857 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1859 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1860 * copy it over 'ssc' */
1861 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1862 if (is_ANYOF_SYNTHETIC(and_with)) {
1863 StructCopy(and_with, ssc, regnode_ssc);
1866 ssc->invlist = anded_cp_list;
1867 ANYOF_POSIXL_ZERO(ssc);
1868 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1869 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1873 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1874 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
1876 /* One or the other of P1, P2 is non-empty. */
1877 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1878 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1880 ssc_union(ssc, anded_cp_list, FALSE);
1882 else { /* P1 = P2 = empty */
1883 ssc_intersection(ssc, anded_cp_list, FALSE);
1889 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1890 const regnode_charclass *or_with)
1892 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1893 * another SSC or a regular ANYOF class. Can create false positives if
1894 * 'or_with' is to be inverted. */
1899 PERL_ARGS_ASSERT_SSC_OR;
1901 assert(is_ANYOF_SYNTHETIC(ssc));
1903 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1904 * the code point inversion list and just the relevant flags */
1905 if (is_ANYOF_SYNTHETIC(or_with)) {
1906 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1907 ored_flags = ANYOF_FLAGS(or_with);
1910 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1911 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1912 if (OP(or_with) != ANYOFD) {
1914 |= ANYOF_FLAGS(or_with)
1915 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1916 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1917 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(or_with))) {
1919 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1924 ANYOF_FLAGS(ssc) |= ored_flags;
1926 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1927 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1928 * 'or_with' may be inverted. When not inverted, we have the simple
1929 * situation of computing:
1930 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1931 * If P1|P2 yields a situation with both a class and its complement are
1932 * set, like having both \w and \W, this matches all code points, and we
1933 * can delete these from the P component of the ssc going forward. XXX We
1934 * might be able to delete all the P components, but I (khw) am not certain
1935 * about this, and it is better to be safe.
1938 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1939 * <= (C1 | P1) | ~C2
1940 * <= (C1 | ~C2) | P1
1941 * (which results in actually simpler code than the non-inverted case)
1944 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1945 && ! is_ANYOF_SYNTHETIC(or_with))
1947 /* We ignore P2, leaving P1 going forward */
1948 } /* else Not inverted */
1949 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
1950 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1951 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1953 for (i = 0; i < ANYOF_MAX; i += 2) {
1954 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1956 ssc_match_all_cp(ssc);
1957 ANYOF_POSIXL_CLEAR(ssc, i);
1958 ANYOF_POSIXL_CLEAR(ssc, i+1);
1966 FALSE /* Already has been inverted */
1970 PERL_STATIC_INLINE void
1971 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1973 PERL_ARGS_ASSERT_SSC_UNION;
1975 assert(is_ANYOF_SYNTHETIC(ssc));
1977 _invlist_union_maybe_complement_2nd(ssc->invlist,
1983 PERL_STATIC_INLINE void
1984 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1986 const bool invert2nd)
1988 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1990 assert(is_ANYOF_SYNTHETIC(ssc));
1992 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1998 PERL_STATIC_INLINE void
1999 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
2001 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
2003 assert(is_ANYOF_SYNTHETIC(ssc));
2005 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
2008 PERL_STATIC_INLINE void
2009 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
2011 /* AND just the single code point 'cp' into the SSC 'ssc' */
2013 SV* cp_list = _new_invlist(2);
2015 PERL_ARGS_ASSERT_SSC_CP_AND;
2017 assert(is_ANYOF_SYNTHETIC(ssc));
2019 cp_list = add_cp_to_invlist(cp_list, cp);
2020 ssc_intersection(ssc, cp_list,
2021 FALSE /* Not inverted */
2023 SvREFCNT_dec_NN(cp_list);
2026 PERL_STATIC_INLINE void
2027 S_ssc_clear_locale(regnode_ssc *ssc)
2029 /* Set the SSC 'ssc' to not match any locale things */
2030 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
2032 assert(is_ANYOF_SYNTHETIC(ssc));
2034 ANYOF_POSIXL_ZERO(ssc);
2035 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
2038 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
2041 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
2043 /* The synthetic start class is used to hopefully quickly winnow down
2044 * places where a pattern could start a match in the target string. If it
2045 * doesn't really narrow things down that much, there isn't much point to
2046 * having the overhead of using it. This function uses some very crude
2047 * heuristics to decide if to use the ssc or not.
2049 * It returns TRUE if 'ssc' rules out more than half what it considers to
2050 * be the "likely" possible matches, but of course it doesn't know what the
2051 * actual things being matched are going to be; these are only guesses
2053 * For /l matches, it assumes that the only likely matches are going to be
2054 * in the 0-255 range, uniformly distributed, so half of that is 127
2055 * For /a and /d matches, it assumes that the likely matches will be just
2056 * the ASCII range, so half of that is 63
2057 * For /u and there isn't anything matching above the Latin1 range, it
2058 * assumes that that is the only range likely to be matched, and uses
2059 * half that as the cut-off: 127. If anything matches above Latin1,
2060 * it assumes that all of Unicode could match (uniformly), except for
2061 * non-Unicode code points and things in the General Category "Other"
2062 * (unassigned, private use, surrogates, controls and formats). This
2063 * is a much large number. */
2065 U32 count = 0; /* Running total of number of code points matched by
2067 UV start, end; /* Start and end points of current range in inversion
2068 XXX outdated. UTF-8 locales are common, what about invert? list */
2069 const U32 max_code_points = (LOC)
2071 : (( ! UNI_SEMANTICS
2072 || invlist_highest(ssc->invlist) < 256)
2075 const U32 max_match = max_code_points / 2;
2077 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
2079 invlist_iterinit(ssc->invlist);
2080 while (invlist_iternext(ssc->invlist, &start, &end)) {
2081 if (start >= max_code_points) {
2084 end = MIN(end, max_code_points - 1);
2085 count += end - start + 1;
2086 if (count >= max_match) {
2087 invlist_iterfinish(ssc->invlist);
2097 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
2099 /* The inversion list in the SSC is marked mortal; now we need a more
2100 * permanent copy, which is stored the same way that is done in a regular
2101 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
2104 SV* invlist = invlist_clone(ssc->invlist, NULL);
2106 PERL_ARGS_ASSERT_SSC_FINALIZE;
2108 assert(is_ANYOF_SYNTHETIC(ssc));
2110 /* The code in this file assumes that all but these flags aren't relevant
2111 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
2112 * by the time we reach here */
2113 assert(! (ANYOF_FLAGS(ssc)
2114 & ~( ANYOF_COMMON_FLAGS
2115 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
2116 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
2118 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
2120 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist, NULL, NULL);
2122 /* Make sure is clone-safe */
2123 ssc->invlist = NULL;
2125 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
2126 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
2127 OP(ssc) = ANYOFPOSIXL;
2129 else if (RExC_contains_locale) {
2133 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
2136 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
2137 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
2138 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
2139 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
2140 ? (TRIE_LIST_CUR( idx ) - 1) \
2146 dump_trie(trie,widecharmap,revcharmap)
2147 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
2148 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
2150 These routines dump out a trie in a somewhat readable format.
2151 The _interim_ variants are used for debugging the interim
2152 tables that are used to generate the final compressed
2153 representation which is what dump_trie expects.
2155 Part of the reason for their existence is to provide a form
2156 of documentation as to how the different representations function.
2161 Dumps the final compressed table form of the trie to Perl_debug_log.
2162 Used for debugging make_trie().
2166 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
2167 AV *revcharmap, U32 depth)
2170 SV *sv=sv_newmortal();
2171 int colwidth= widecharmap ? 6 : 4;
2173 GET_RE_DEBUG_FLAGS_DECL;
2175 PERL_ARGS_ASSERT_DUMP_TRIE;
2177 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
2178 depth+1, "Match","Base","Ofs" );
2180 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2181 SV ** const tmp = av_fetch( revcharmap, state, 0);
2183 Perl_re_printf( aTHX_ "%*s",
2185 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2186 PL_colors[0], PL_colors[1],
2187 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2188 PERL_PV_ESCAPE_FIRSTCHAR
2193 Perl_re_printf( aTHX_ "\n");
2194 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
2196 for( state = 0 ; state < trie->uniquecharcount ; state++ )
2197 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2198 Perl_re_printf( aTHX_ "\n");
2200 for( state = 1 ; state < trie->statecount ; state++ ) {
2201 const U32 base = trie->states[ state ].trans.base;
2203 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
2205 if ( trie->states[ state ].wordnum ) {
2206 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
2208 Perl_re_printf( aTHX_ "%6s", "" );
2211 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
2216 while( ( base + ofs < trie->uniquecharcount ) ||
2217 ( base + ofs - trie->uniquecharcount < trie->lasttrans
2218 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2222 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
2224 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2225 if ( ( base + ofs >= trie->uniquecharcount )
2226 && ( base + ofs - trie->uniquecharcount
2228 && trie->trans[ base + ofs
2229 - trie->uniquecharcount ].check == state )
2231 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
2232 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2235 Perl_re_printf( aTHX_ "%*s", colwidth," ." );
2239 Perl_re_printf( aTHX_ "]");
2242 Perl_re_printf( aTHX_ "\n" );
2244 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
2246 for (word=1; word <= trie->wordcount; word++) {
2247 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2248 (int)word, (int)(trie->wordinfo[word].prev),
2249 (int)(trie->wordinfo[word].len));
2251 Perl_re_printf( aTHX_ "\n" );
2254 Dumps a fully constructed but uncompressed trie in list form.
2255 List tries normally only are used for construction when the number of
2256 possible chars (trie->uniquecharcount) is very high.
2257 Used for debugging make_trie().
2260 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2261 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2265 SV *sv=sv_newmortal();
2266 int colwidth= widecharmap ? 6 : 4;
2267 GET_RE_DEBUG_FLAGS_DECL;
2269 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2271 /* print out the table precompression. */
2272 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
2274 Perl_re_indentf( aTHX_ "%s",
2275 depth+1, "------:-----+-----------------\n" );
2277 for( state=1 ; state < next_alloc ; state ++ ) {
2280 Perl_re_indentf( aTHX_ " %4" UVXf " :",
2281 depth+1, (UV)state );
2282 if ( ! trie->states[ state ].wordnum ) {
2283 Perl_re_printf( aTHX_ "%5s| ","");
2285 Perl_re_printf( aTHX_ "W%4x| ",
2286 trie->states[ state ].wordnum
2289 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2290 SV ** const tmp = av_fetch( revcharmap,
2291 TRIE_LIST_ITEM(state, charid).forid, 0);
2293 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
2295 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2297 PL_colors[0], PL_colors[1],
2298 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2299 | PERL_PV_ESCAPE_FIRSTCHAR
2301 TRIE_LIST_ITEM(state, charid).forid,
2302 (UV)TRIE_LIST_ITEM(state, charid).newstate
2305 Perl_re_printf( aTHX_ "\n%*s| ",
2306 (int)((depth * 2) + 14), "");
2309 Perl_re_printf( aTHX_ "\n");
2314 Dumps a fully constructed but uncompressed trie in table form.
2315 This is the normal DFA style state transition table, with a few
2316 twists to facilitate compression later.
2317 Used for debugging make_trie().
2320 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2321 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2326 SV *sv=sv_newmortal();
2327 int colwidth= widecharmap ? 6 : 4;
2328 GET_RE_DEBUG_FLAGS_DECL;
2330 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2333 print out the table precompression so that we can do a visual check
2334 that they are identical.
2337 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
2339 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2340 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2342 Perl_re_printf( aTHX_ "%*s",
2344 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2345 PL_colors[0], PL_colors[1],
2346 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2347 PERL_PV_ESCAPE_FIRSTCHAR
2353 Perl_re_printf( aTHX_ "\n");
2354 Perl_re_indentf( aTHX_ "State+-", depth+1 );
2356 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2357 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
2360 Perl_re_printf( aTHX_ "\n" );
2362 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2364 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
2366 (UV)TRIE_NODENUM( state ) );
2368 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2369 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2371 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
2373 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
2375 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2376 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
2377 (UV)trie->trans[ state ].check );
2379 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
2380 (UV)trie->trans[ state ].check,
2381 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2389 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2390 startbranch: the first branch in the whole branch sequence
2391 first : start branch of sequence of branch-exact nodes.
2392 May be the same as startbranch
2393 last : Thing following the last branch.
2394 May be the same as tail.
2395 tail : item following the branch sequence
2396 count : words in the sequence
2397 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2398 depth : indent depth
2400 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2402 A trie is an N'ary tree where the branches are determined by digital
2403 decomposition of the key. IE, at the root node you look up the 1st character and
2404 follow that branch repeat until you find the end of the branches. Nodes can be
2405 marked as "accepting" meaning they represent a complete word. Eg:
2409 would convert into the following structure. Numbers represent states, letters
2410 following numbers represent valid transitions on the letter from that state, if
2411 the number is in square brackets it represents an accepting state, otherwise it
2412 will be in parenthesis.
2414 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2418 (1) +-i->(6)-+-s->[7]
2420 +-s->(3)-+-h->(4)-+-e->[5]
2422 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2424 This shows that when matching against the string 'hers' we will begin at state 1
2425 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2426 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2427 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2428 single traverse. We store a mapping from accepting to state to which word was
2429 matched, and then when we have multiple possibilities we try to complete the
2430 rest of the regex in the order in which they occurred in the alternation.
2432 The only prior NFA like behaviour that would be changed by the TRIE support is
2433 the silent ignoring of duplicate alternations which are of the form:
2435 / (DUPE|DUPE) X? (?{ ... }) Y /x
2437 Thus EVAL blocks following a trie may be called a different number of times with
2438 and without the optimisation. With the optimisations dupes will be silently
2439 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2440 the following demonstrates:
2442 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2444 which prints out 'word' three times, but
2446 'words'=~/(word|word|word)(?{ print $1 })S/
2448 which doesnt print it out at all. This is due to other optimisations kicking in.
2450 Example of what happens on a structural level:
2452 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2454 1: CURLYM[1] {1,32767}(18)
2465 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2466 and should turn into:
2468 1: CURLYM[1] {1,32767}(18)
2470 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2478 Cases where tail != last would be like /(?foo|bar)baz/:
2488 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2489 and would end up looking like:
2492 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2499 d = uvchr_to_utf8_flags(d, uv, 0);
2501 is the recommended Unicode-aware way of saying
2506 #define TRIE_STORE_REVCHAR(val) \
2509 SV *zlopp = newSV(UTF8_MAXBYTES); \
2510 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2511 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2512 SvCUR_set(zlopp, kapow - flrbbbbb); \
2515 av_push(revcharmap, zlopp); \
2517 char ooooff = (char)val; \
2518 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2522 /* This gets the next character from the input, folding it if not already
2524 #define TRIE_READ_CHAR STMT_START { \
2527 /* if it is UTF then it is either already folded, or does not need \
2529 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2531 else if (folder == PL_fold_latin1) { \
2532 /* This folder implies Unicode rules, which in the range expressible \
2533 * by not UTF is the lower case, with the two exceptions, one of \
2534 * which should have been taken care of before calling this */ \
2535 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2536 uvc = toLOWER_L1(*uc); \
2537 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2540 /* raw data, will be folded later if needed */ \
2548 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2549 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2550 U32 ging = TRIE_LIST_LEN( state ) * 2; \
2551 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2552 TRIE_LIST_LEN( state ) = ging; \
2554 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2555 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2556 TRIE_LIST_CUR( state )++; \
2559 #define TRIE_LIST_NEW(state) STMT_START { \
2560 Newx( trie->states[ state ].trans.list, \
2561 4, reg_trie_trans_le ); \
2562 TRIE_LIST_CUR( state ) = 1; \
2563 TRIE_LIST_LEN( state ) = 4; \
2566 #define TRIE_HANDLE_WORD(state) STMT_START { \
2567 U16 dupe= trie->states[ state ].wordnum; \
2568 regnode * const noper_next = regnext( noper ); \
2571 /* store the word for dumping */ \
2573 if (OP(noper) != NOTHING) \
2574 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2576 tmp = newSVpvn_utf8( "", 0, UTF ); \
2577 av_push( trie_words, tmp ); \
2581 trie->wordinfo[curword].prev = 0; \
2582 trie->wordinfo[curword].len = wordlen; \
2583 trie->wordinfo[curword].accept = state; \
2585 if ( noper_next < tail ) { \
2587 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2589 trie->jump[curword] = (U16)(noper_next - convert); \
2591 jumper = noper_next; \
2593 nextbranch= regnext(cur); \
2597 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2598 /* chain, so that when the bits of chain are later */\
2599 /* linked together, the dups appear in the chain */\
2600 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2601 trie->wordinfo[dupe].prev = curword; \
2603 /* we haven't inserted this word yet. */ \
2604 trie->states[ state ].wordnum = curword; \
2609 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2610 ( ( base + charid >= ucharcount \
2611 && base + charid < ubound \
2612 && state == trie->trans[ base - ucharcount + charid ].check \
2613 && trie->trans[ base - ucharcount + charid ].next ) \
2614 ? trie->trans[ base - ucharcount + charid ].next \
2615 : ( state==1 ? special : 0 ) \
2618 #define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2620 TRIE_BITMAP_SET(trie, uvc); \
2621 /* store the folded codepoint */ \
2623 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2626 /* store first byte of utf8 representation of */ \
2627 /* variant codepoints */ \
2628 if (! UVCHR_IS_INVARIANT(uvc)) { \
2629 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2634 #define MADE_JUMP_TRIE 2
2635 #define MADE_EXACT_TRIE 4
2638 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2639 regnode *first, regnode *last, regnode *tail,
2640 U32 word_count, U32 flags, U32 depth)
2642 /* first pass, loop through and scan words */
2643 reg_trie_data *trie;
2644 HV *widecharmap = NULL;
2645 AV *revcharmap = newAV();
2651 regnode *jumper = NULL;
2652 regnode *nextbranch = NULL;
2653 regnode *convert = NULL;
2654 U32 *prev_states; /* temp array mapping each state to previous one */
2655 /* we just use folder as a flag in utf8 */
2656 const U8 * folder = NULL;
2658 /* in the below add_data call we are storing either 'tu' or 'tuaa'
2659 * which stands for one trie structure, one hash, optionally followed
2662 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuaa"));
2663 AV *trie_words = NULL;
2664 /* along with revcharmap, this only used during construction but both are
2665 * useful during debugging so we store them in the struct when debugging.
2668 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2669 STRLEN trie_charcount=0;
2671 SV *re_trie_maxbuff;
2672 GET_RE_DEBUG_FLAGS_DECL;
2674 PERL_ARGS_ASSERT_MAKE_TRIE;
2676 PERL_UNUSED_ARG(depth);
2680 case EXACT: case EXACT_ONLY8: case EXACTL: break;
2684 case EXACTFLU8: folder = PL_fold_latin1; break;
2685 case EXACTF: folder = PL_fold; break;
2686 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2689 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2691 trie->startstate = 1;
2692 trie->wordcount = word_count;
2693 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2694 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2695 if (flags == EXACT || flags == EXACT_ONLY8 || flags == EXACTL)
2696 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2697 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2698 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2701 trie_words = newAV();
2704 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2705 assert(re_trie_maxbuff);
2706 if (!SvIOK(re_trie_maxbuff)) {
2707 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2709 DEBUG_TRIE_COMPILE_r({
2710 Perl_re_indentf( aTHX_
2711 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2713 REG_NODE_NUM(startbranch), REG_NODE_NUM(first),
2714 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2717 /* Find the node we are going to overwrite */
2718 if ( first == startbranch && OP( last ) != BRANCH ) {
2719 /* whole branch chain */
2722 /* branch sub-chain */
2723 convert = NEXTOPER( first );
2726 /* -- First loop and Setup --
2728 We first traverse the branches and scan each word to determine if it
2729 contains widechars, and how many unique chars there are, this is
2730 important as we have to build a table with at least as many columns as we
2733 We use an array of integers to represent the character codes 0..255
2734 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2735 the native representation of the character value as the key and IV's for
2738 *TODO* If we keep track of how many times each character is used we can
2739 remap the columns so that the table compression later on is more
2740 efficient in terms of memory by ensuring the most common value is in the
2741 middle and the least common are on the outside. IMO this would be better
2742 than a most to least common mapping as theres a decent chance the most
2743 common letter will share a node with the least common, meaning the node
2744 will not be compressible. With a middle is most common approach the worst
2745 case is when we have the least common nodes twice.
2749 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2750 regnode *noper = NEXTOPER( cur );
2754 U32 wordlen = 0; /* required init */
2755 STRLEN minchars = 0;
2756 STRLEN maxchars = 0;
2757 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2760 if (OP(noper) == NOTHING) {
2761 /* skip past a NOTHING at the start of an alternation
2762 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2764 regnode *noper_next= regnext(noper);
2765 if (noper_next < tail)
2770 && ( OP(noper) == flags
2771 || (flags == EXACT && OP(noper) == EXACT_ONLY8)
2772 || (flags == EXACTFU && ( OP(noper) == EXACTFU_ONLY8
2773 || OP(noper) == EXACTFUP))))
2775 uc= (U8*)STRING(noper);
2776 e= uc + STR_LEN(noper);
2783 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2784 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2785 regardless of encoding */
2786 if (OP( noper ) == EXACTFUP) {
2787 /* false positives are ok, so just set this */
2788 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2792 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2794 TRIE_CHARCOUNT(trie)++;
2797 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2798 * is in effect. Under /i, this character can match itself, or
2799 * anything that folds to it. If not under /i, it can match just
2800 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2801 * all fold to k, and all are single characters. But some folds
2802 * expand to more than one character, so for example LATIN SMALL
2803 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2804 * the string beginning at 'uc' is 'ffi', it could be matched by
2805 * three characters, or just by the one ligature character. (It
2806 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2807 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2808 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2809 * match.) The trie needs to know the minimum and maximum number
2810 * of characters that could match so that it can use size alone to
2811 * quickly reject many match attempts. The max is simple: it is
2812 * the number of folded characters in this branch (since a fold is
2813 * never shorter than what folds to it. */
2817 /* And the min is equal to the max if not under /i (indicated by
2818 * 'folder' being NULL), or there are no multi-character folds. If
2819 * there is a multi-character fold, the min is incremented just
2820 * once, for the character that folds to the sequence. Each
2821 * character in the sequence needs to be added to the list below of
2822 * characters in the trie, but we count only the first towards the
2823 * min number of characters needed. This is done through the
2824 * variable 'foldlen', which is returned by the macros that look
2825 * for these sequences as the number of bytes the sequence
2826 * occupies. Each time through the loop, we decrement 'foldlen' by
2827 * how many bytes the current char occupies. Only when it reaches
2828 * 0 do we increment 'minchars' or look for another multi-character
2830 if (folder == NULL) {
2833 else if (foldlen > 0) {
2834 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2839 /* See if *uc is the beginning of a multi-character fold. If
2840 * so, we decrement the length remaining to look at, to account
2841 * for the current character this iteration. (We can use 'uc'
2842 * instead of the fold returned by TRIE_READ_CHAR because for
2843 * non-UTF, the latin1_safe macro is smart enough to account
2844 * for all the unfolded characters, and because for UTF, the
2845 * string will already have been folded earlier in the
2846 * compilation process */
2848 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2849 foldlen -= UTF8SKIP(uc);
2852 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2857 /* The current character (and any potential folds) should be added
2858 * to the possible matching characters for this position in this
2862 U8 folded= folder[ (U8) uvc ];
2863 if ( !trie->charmap[ folded ] ) {
2864 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2865 TRIE_STORE_REVCHAR( folded );
2868 if ( !trie->charmap[ uvc ] ) {
2869 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2870 TRIE_STORE_REVCHAR( uvc );
2873 /* store the codepoint in the bitmap, and its folded
2875 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
2876 set_bit = 0; /* We've done our bit :-) */
2880 /* XXX We could come up with the list of code points that fold
2881 * to this using PL_utf8_foldclosures, except not for
2882 * multi-char folds, as there may be multiple combinations
2883 * there that could work, which needs to wait until runtime to
2884 * resolve (The comment about LIGATURE FFI above is such an
2889 widecharmap = newHV();
2891 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2894 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
2896 if ( !SvTRUE( *svpp ) ) {
2897 sv_setiv( *svpp, ++trie->uniquecharcount );
2898 TRIE_STORE_REVCHAR(uvc);
2901 } /* end loop through characters in this branch of the trie */
2903 /* We take the min and max for this branch and combine to find the min
2904 * and max for all branches processed so far */
2905 if( cur == first ) {
2906 trie->minlen = minchars;
2907 trie->maxlen = maxchars;
2908 } else if (minchars < trie->minlen) {
2909 trie->minlen = minchars;
2910 } else if (maxchars > trie->maxlen) {
2911 trie->maxlen = maxchars;
2913 } /* end first pass */
2914 DEBUG_TRIE_COMPILE_r(
2915 Perl_re_indentf( aTHX_
2916 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2918 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2919 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2920 (int)trie->minlen, (int)trie->maxlen )
2924 We now know what we are dealing with in terms of unique chars and
2925 string sizes so we can calculate how much memory a naive
2926 representation using a flat table will take. If it's over a reasonable
2927 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2928 conservative but potentially much slower representation using an array
2931 At the end we convert both representations into the same compressed
2932 form that will be used in regexec.c for matching with. The latter
2933 is a form that cannot be used to construct with but has memory
2934 properties similar to the list form and access properties similar
2935 to the table form making it both suitable for fast searches and
2936 small enough that its feasable to store for the duration of a program.
2938 See the comment in the code where the compressed table is produced
2939 inplace from the flat tabe representation for an explanation of how
2940 the compression works.
2945 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2948 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2949 > SvIV(re_trie_maxbuff) )
2952 Second Pass -- Array Of Lists Representation
2954 Each state will be represented by a list of charid:state records
2955 (reg_trie_trans_le) the first such element holds the CUR and LEN
2956 points of the allocated array. (See defines above).
2958 We build the initial structure using the lists, and then convert
2959 it into the compressed table form which allows faster lookups
2960 (but cant be modified once converted).
2963 STRLEN transcount = 1;
2965 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
2968 trie->states = (reg_trie_state *)
2969 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2970 sizeof(reg_trie_state) );
2974 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2976 regnode *noper = NEXTOPER( cur );
2977 U32 state = 1; /* required init */
2978 U16 charid = 0; /* sanity init */
2979 U32 wordlen = 0; /* required init */
2981 if (OP(noper) == NOTHING) {
2982 regnode *noper_next= regnext(noper);
2983 if (noper_next < tail)
2988 && ( OP(noper) == flags
2989 || (flags == EXACT && OP(noper) == EXACT_ONLY8)
2990 || (flags == EXACTFU && ( OP(noper) == EXACTFU_ONLY8
2991 || OP(noper) == EXACTFUP))))
2993 const U8 *uc= (U8*)STRING(noper);
2994 const U8 *e= uc + STR_LEN(noper);
2996 for ( ; uc < e ; uc += len ) {
3001 charid = trie->charmap[ uvc ];
3003 SV** const svpp = hv_fetch( widecharmap,
3010 charid=(U16)SvIV( *svpp );
3013 /* charid is now 0 if we dont know the char read, or
3014 * nonzero if we do */
3021 if ( !trie->states[ state ].trans.list ) {
3022 TRIE_LIST_NEW( state );
3025 check <= TRIE_LIST_USED( state );
3028 if ( TRIE_LIST_ITEM( state, check ).forid
3031 newstate = TRIE_LIST_ITEM( state, check ).newstate;
3036 newstate = next_alloc++;
3037 prev_states[newstate] = state;
3038 TRIE_LIST_PUSH( state, charid, newstate );
3043 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3047 TRIE_HANDLE_WORD(state);
3049 } /* end second pass */
3051 /* next alloc is the NEXT state to be allocated */
3052 trie->statecount = next_alloc;
3053 trie->states = (reg_trie_state *)
3054 PerlMemShared_realloc( trie->states,
3056 * sizeof(reg_trie_state) );
3058 /* and now dump it out before we compress it */
3059 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
3060 revcharmap, next_alloc,
3064 trie->trans = (reg_trie_trans *)
3065 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
3072 for( state=1 ; state < next_alloc ; state ++ ) {
3076 DEBUG_TRIE_COMPILE_MORE_r(
3077 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
3081 if (trie->states[state].trans.list) {
3082 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
3086 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3087 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
3088 if ( forid < minid ) {
3090 } else if ( forid > maxid ) {
3094 if ( transcount < tp + maxid - minid + 1) {
3096 trie->trans = (reg_trie_trans *)
3097 PerlMemShared_realloc( trie->trans,
3099 * sizeof(reg_trie_trans) );
3100 Zero( trie->trans + (transcount / 2),
3104 base = trie->uniquecharcount + tp - minid;
3105 if ( maxid == minid ) {
3107 for ( ; zp < tp ; zp++ ) {
3108 if ( ! trie->trans[ zp ].next ) {
3109 base = trie->uniquecharcount + zp - minid;
3110 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
3112 trie->trans[ zp ].check = state;
3118 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
3120 trie->trans[ tp ].check = state;
3125 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3126 const U32 tid = base
3127 - trie->uniquecharcount
3128 + TRIE_LIST_ITEM( state, idx ).forid;
3129 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
3131 trie->trans[ tid ].check = state;
3133 tp += ( maxid - minid + 1 );
3135 Safefree(trie->states[ state ].trans.list);
3138 DEBUG_TRIE_COMPILE_MORE_r(
3139 Perl_re_printf( aTHX_ " base: %d\n",base);
3142 trie->states[ state ].trans.base=base;
3144 trie->lasttrans = tp + 1;
3148 Second Pass -- Flat Table Representation.
3150 we dont use the 0 slot of either trans[] or states[] so we add 1 to
3151 each. We know that we will need Charcount+1 trans at most to store
3152 the data (one row per char at worst case) So we preallocate both
3153 structures assuming worst case.
3155 We then construct the trie using only the .next slots of the entry
3158 We use the .check field of the first entry of the node temporarily
3159 to make compression both faster and easier by keeping track of how
3160 many non zero fields are in the node.
3162 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
3165 There are two terms at use here: state as a TRIE_NODEIDX() which is
3166 a number representing the first entry of the node, and state as a
3167 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
3168 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
3169 if there are 2 entrys per node. eg:
3177 The table is internally in the right hand, idx form. However as we
3178 also have to deal with the states array which is indexed by nodenum
3179 we have to use TRIE_NODENUM() to convert.
3182 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
3185 trie->trans = (reg_trie_trans *)
3186 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
3187 * trie->uniquecharcount + 1,
3188 sizeof(reg_trie_trans) );
3189 trie->states = (reg_trie_state *)
3190 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3191 sizeof(reg_trie_state) );
3192 next_alloc = trie->uniquecharcount + 1;
3195 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3197 regnode *noper = NEXTOPER( cur );
3199 U32 state = 1; /* required init */
3201 U16 charid = 0; /* sanity init */
3202 U32 accept_state = 0; /* sanity init */
3204 U32 wordlen = 0; /* required init */
3206 if (OP(noper) == NOTHING) {
3207 regnode *noper_next= regnext(noper);
3208 if (noper_next < tail)
3213 && ( OP(noper) == flags
3214 || (flags == EXACT && OP(noper) == EXACT_ONLY8)
3215 || (flags == EXACTFU && ( OP(noper) == EXACTFU_ONLY8
3216 || OP(noper) == EXACTFUP))))
3218 const U8 *uc= (U8*)STRING(noper);
3219 const U8 *e= uc + STR_LEN(noper);
3221 for ( ; uc < e ; uc += len ) {
3226 charid = trie->charmap[ uvc ];
3228 SV* const * const svpp = hv_fetch( widecharmap,
3232 charid = svpp ? (U16)SvIV(*svpp) : 0;
3236 if ( !trie->trans[ state + charid ].next ) {
3237 trie->trans[ state + charid ].next = next_alloc;
3238 trie->trans[ state ].check++;
3239 prev_states[TRIE_NODENUM(next_alloc)]
3240 = TRIE_NODENUM(state);
3241 next_alloc += trie->uniquecharcount;
3243 state = trie->trans[ state + charid ].next;
3245 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3247 /* charid is now 0 if we dont know the char read, or
3248 * nonzero if we do */
3251 accept_state = TRIE_NODENUM( state );
3252 TRIE_HANDLE_WORD(accept_state);
3254 } /* end second pass */
3256 /* and now dump it out before we compress it */
3257 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3259 next_alloc, depth+1));
3263 * Inplace compress the table.*
3265 For sparse data sets the table constructed by the trie algorithm will
3266 be mostly 0/FAIL transitions or to put it another way mostly empty.
3267 (Note that leaf nodes will not contain any transitions.)
3269 This algorithm compresses the tables by eliminating most such
3270 transitions, at the cost of a modest bit of extra work during lookup:
3272 - Each states[] entry contains a .base field which indicates the
3273 index in the state[] array wheres its transition data is stored.
3275 - If .base is 0 there are no valid transitions from that node.
3277 - If .base is nonzero then charid is added to it to find an entry in
3280 -If trans[states[state].base+charid].check!=state then the
3281 transition is taken to be a 0/Fail transition. Thus if there are fail
3282 transitions at the front of the node then the .base offset will point
3283 somewhere inside the previous nodes data (or maybe even into a node
3284 even earlier), but the .check field determines if the transition is
3288 The following process inplace converts the table to the compressed
3289 table: We first do not compress the root node 1,and mark all its
3290 .check pointers as 1 and set its .base pointer as 1 as well. This
3291 allows us to do a DFA construction from the compressed table later,
3292 and ensures that any .base pointers we calculate later are greater
3295 - We set 'pos' to indicate the first entry of the second node.
3297 - We then iterate over the columns of the node, finding the first and
3298 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3299 and set the .check pointers accordingly, and advance pos
3300 appropriately and repreat for the next node. Note that when we copy
3301 the next pointers we have to convert them from the original
3302 NODEIDX form to NODENUM form as the former is not valid post
3305 - If a node has no transitions used we mark its base as 0 and do not
3306 advance the pos pointer.
3308 - If a node only has one transition we use a second pointer into the
3309 structure to fill in allocated fail transitions from other states.
3310 This pointer is independent of the main pointer and scans forward
3311 looking for null transitions that are allocated to a state. When it
3312 finds one it writes the single transition into the "hole". If the
3313 pointer doesnt find one the single transition is appended as normal.
3315 - Once compressed we can Renew/realloc the structures to release the
3318 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3319 specifically Fig 3.47 and the associated pseudocode.
3323 const U32 laststate = TRIE_NODENUM( next_alloc );
3326 trie->statecount = laststate;
3328 for ( state = 1 ; state < laststate ; state++ ) {
3330 const U32 stateidx = TRIE_NODEIDX( state );
3331 const U32 o_used = trie->trans[ stateidx ].check;
3332 U32 used = trie->trans[ stateidx ].check;
3333 trie->trans[ stateidx ].check = 0;
3336 used && charid < trie->uniquecharcount;
3339 if ( flag || trie->trans[ stateidx + charid ].next ) {
3340 if ( trie->trans[ stateidx + charid ].next ) {
3342 for ( ; zp < pos ; zp++ ) {
3343 if ( ! trie->trans[ zp ].next ) {
3347 trie->states[ state ].trans.base
3349 + trie->uniquecharcount
3351 trie->trans[ zp ].next
3352 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3354 trie->trans[ zp ].check = state;
3355 if ( ++zp > pos ) pos = zp;
3362 trie->states[ state ].trans.base
3363 = pos + trie->uniquecharcount - charid ;
3365 trie->trans[ pos ].next
3366 = SAFE_TRIE_NODENUM(
3367 trie->trans[ stateidx + charid ].next );
3368 trie->trans[ pos ].check = state;
3373 trie->lasttrans = pos + 1;
3374 trie->states = (reg_trie_state *)
3375 PerlMemShared_realloc( trie->states, laststate
3376 * sizeof(reg_trie_state) );
3377 DEBUG_TRIE_COMPILE_MORE_r(
3378 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
3380 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3384 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3387 } /* end table compress */
3389 DEBUG_TRIE_COMPILE_MORE_r(
3390 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
3392 (UV)trie->statecount,
3393 (UV)trie->lasttrans)
3395 /* resize the trans array to remove unused space */
3396 trie->trans = (reg_trie_trans *)
3397 PerlMemShared_realloc( trie->trans, trie->lasttrans
3398 * sizeof(reg_trie_trans) );
3400 { /* Modify the program and insert the new TRIE node */
3401 U8 nodetype =(U8)(flags & 0xFF);
3405 regnode *optimize = NULL;
3406 #ifdef RE_TRACK_PATTERN_OFFSETS
3409 U32 mjd_nodelen = 0;
3410 #endif /* RE_TRACK_PATTERN_OFFSETS */
3411 #endif /* DEBUGGING */
3413 This means we convert either the first branch or the first Exact,
3414 depending on whether the thing following (in 'last') is a branch
3415 or not and whther first is the startbranch (ie is it a sub part of
3416 the alternation or is it the whole thing.)
3417 Assuming its a sub part we convert the EXACT otherwise we convert
3418 the whole branch sequence, including the first.
3420 /* Find the node we are going to overwrite */
3421 if ( first != startbranch || OP( last ) == BRANCH ) {
3422 /* branch sub-chain */
3423 NEXT_OFF( first ) = (U16)(last - first);
3424 #ifdef RE_TRACK_PATTERN_OFFSETS
3426 mjd_offset= Node_Offset((convert));
3427 mjd_nodelen= Node_Length((convert));
3430 /* whole branch chain */
3432 #ifdef RE_TRACK_PATTERN_OFFSETS
3435 const regnode *nop = NEXTOPER( convert );
3436 mjd_offset= Node_Offset((nop));
3437 mjd_nodelen= Node_Length((nop));
3441 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
3443 (UV)mjd_offset, (UV)mjd_nodelen)
3446 /* But first we check to see if there is a common prefix we can
3447 split out as an EXACT and put in front of the TRIE node. */
3448 trie->startstate= 1;
3449 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3450 /* we want to find the first state that has more than
3451 * one transition, if that state is not the first state
3452 * then we have a common prefix which we can remove.
3455 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3457 I32 first_ofs = -1; /* keeps track of the ofs of the first
3458 transition, -1 means none */
3460 const U32 base = trie->states[ state ].trans.base;
3462 /* does this state terminate an alternation? */
3463 if ( trie->states[state].wordnum )
3466 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3467 if ( ( base + ofs >= trie->uniquecharcount ) &&
3468 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3469 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3471 if ( ++count > 1 ) {
3472 /* we have more than one transition */
3475 /* if this is the first state there is no common prefix
3476 * to extract, so we can exit */
3477 if ( state == 1 ) break;
3478 tmp = av_fetch( revcharmap, ofs, 0);
3479 ch = (U8*)SvPV_nolen_const( *tmp );
3481 /* if we are on count 2 then we need to initialize the
3482 * bitmap, and store the previous char if there was one
3485 /* clear the bitmap */
3486 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3488 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
3491 if (first_ofs >= 0) {
3492 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
3493 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3495 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3497 Perl_re_printf( aTHX_ "%s", (char*)ch)
3501 /* store the current firstchar in the bitmap */
3502 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3503 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
3509 /* This state has only one transition, its transition is part
3510 * of a common prefix - we need to concatenate the char it
3511 * represents to what we have so far. */
3512 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
3514 char *ch = SvPV( *tmp, len );
3516 SV *sv=sv_newmortal();
3517 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
3519 (UV)state, (UV)first_ofs,
3520 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3521 PL_colors[0], PL_colors[1],
3522 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3523 PERL_PV_ESCAPE_FIRSTCHAR
3528 OP( convert ) = nodetype;
3529 str=STRING(convert);
3532 STR_LEN(convert) += len;
3538 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
3543 trie->prefixlen = (state-1);
3545 regnode *n = convert+NODE_SZ_STR(convert);
3546 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3547 trie->startstate = state;
3548 trie->minlen -= (state - 1);
3549 trie->maxlen -= (state - 1);
3551 /* At least the UNICOS C compiler choked on this
3552 * being argument to DEBUG_r(), so let's just have
3555 #ifdef PERL_EXT_RE_BUILD
3561 regnode *fix = convert;
3562 U32 word = trie->wordcount;
3563 #ifdef RE_TRACK_PATTERN_OFFSETS
3566 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3567 while( ++fix < n ) {
3568 Set_Node_Offset_Length(fix, 0, 0);
3571 SV ** const tmp = av_fetch( trie_words, word, 0 );
3573 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3574 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3576 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3584 NEXT_OFF(convert) = (U16)(tail - convert);
3585 DEBUG_r(optimize= n);
3591 if ( trie->maxlen ) {
3592 NEXT_OFF( convert ) = (U16)(tail - convert);
3593 ARG_SET( convert, data_slot );
3594 /* Store the offset to the first unabsorbed branch in
3595 jump[0], which is otherwise unused by the jump logic.
3596 We use this when dumping a trie and during optimisation. */
3598 trie->jump[0] = (U16)(nextbranch - convert);
3600 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3601 * and there is a bitmap
3602 * and the first "jump target" node we found leaves enough room
3603 * then convert the TRIE node into a TRIEC node, with the bitmap
3604 * embedded inline in the opcode - this is hypothetically faster.
3606 if ( !trie->states[trie->startstate].wordnum
3608 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3610 OP( convert ) = TRIEC;
3611 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3612 PerlMemShared_free(trie->bitmap);
3615 OP( convert ) = TRIE;
3617 /* store the type in the flags */
3618 convert->flags = nodetype;
3622 + regarglen[ OP( convert ) ];
3624 /* XXX We really should free up the resource in trie now,
3625 as we won't use them - (which resources?) dmq */
3627 /* needed for dumping*/
3628 DEBUG_r(if (optimize) {
3629 regnode *opt = convert;
3631 while ( ++opt < optimize) {
3632 Set_Node_Offset_Length(opt, 0, 0);
3635 Try to clean up some of the debris left after the
3638 while( optimize < jumper ) {
3639 Track_Code( mjd_nodelen += Node_Length((optimize)); );
3640 OP( optimize ) = OPTIMIZED;
3641 Set_Node_Offset_Length(optimize, 0, 0);
3644 Set_Node_Offset_Length(convert, mjd_offset, mjd_nodelen);
3646 } /* end node insert */
3648 /* Finish populating the prev field of the wordinfo array. Walk back
3649 * from each accept state until we find another accept state, and if
3650 * so, point the first word's .prev field at the second word. If the
3651 * second already has a .prev field set, stop now. This will be the
3652 * case either if we've already processed that word's accept state,
3653 * or that state had multiple words, and the overspill words were
3654 * already linked up earlier.
3661 for (word=1; word <= trie->wordcount; word++) {
3663 if (trie->wordinfo[word].prev)
3665 state = trie->wordinfo[word].accept;
3667 state = prev_states[state];
3670 prev = trie->states[state].wordnum;
3674 trie->wordinfo[word].prev = prev;
3676 Safefree(prev_states);
3680 /* and now dump out the compressed format */
3681 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3683 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3685 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3686 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3688 SvREFCNT_dec_NN(revcharmap);
3692 : trie->startstate>1
3698 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3700 /* The Trie is constructed and compressed now so we can build a fail array if
3703 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3705 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3709 We find the fail state for each state in the trie, this state is the longest
3710 proper suffix of the current state's 'word' that is also a proper prefix of
3711 another word in our trie. State 1 represents the word '' and is thus the
3712 default fail state. This allows the DFA not to have to restart after its
3713 tried and failed a word at a given point, it simply continues as though it
3714 had been matching the other word in the first place.
3716 'abcdgu'=~/abcdefg|cdgu/
3717 When we get to 'd' we are still matching the first word, we would encounter
3718 'g' which would fail, which would bring us to the state representing 'd' in
3719 the second word where we would try 'g' and succeed, proceeding to match
3722 /* add a fail transition */
3723 const U32 trie_offset = ARG(source);
3724 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3726 const U32 ucharcount = trie->uniquecharcount;
3727 const U32 numstates = trie->statecount;
3728 const U32 ubound = trie->lasttrans + ucharcount;
3732 U32 base = trie->states[ 1 ].trans.base;
3735 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3737 GET_RE_DEBUG_FLAGS_DECL;
3739 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3740 PERL_UNUSED_CONTEXT;
3742 PERL_UNUSED_ARG(depth);
3745 if ( OP(source) == TRIE ) {
3746 struct regnode_1 *op = (struct regnode_1 *)
3747 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3748 StructCopy(source, op, struct regnode_1);
3749 stclass = (regnode *)op;
3751 struct regnode_charclass *op = (struct regnode_charclass *)
3752 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3753 StructCopy(source, op, struct regnode_charclass);
3754 stclass = (regnode *)op;
3756 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3758 ARG_SET( stclass, data_slot );
3759 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3760 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3761 aho->trie=trie_offset;
3762 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3763 Copy( trie->states, aho->states, numstates, reg_trie_state );
3764 Newx( q, numstates, U32);
3765 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3768 /* initialize fail[0..1] to be 1 so that we always have
3769 a valid final fail state */
3770 fail[ 0 ] = fail[ 1 ] = 1;
3772 for ( charid = 0; charid < ucharcount ; charid++ ) {
3773 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3775 q[ q_write ] = newstate;
3776 /* set to point at the root */
3777 fail[ q[ q_write++ ] ]=1;
3780 while ( q_read < q_write) {
3781 const U32 cur = q[ q_read++ % numstates ];
3782 base = trie->states[ cur ].trans.base;
3784 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3785 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3787 U32 fail_state = cur;
3790 fail_state = fail[ fail_state ];
3791 fail_base = aho->states[ fail_state ].trans.base;
3792 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3794 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3795 fail[ ch_state ] = fail_state;
3796 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3798 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3800 q[ q_write++ % numstates] = ch_state;
3804 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3805 when we fail in state 1, this allows us to use the
3806 charclass scan to find a valid start char. This is based on the principle
3807 that theres a good chance the string being searched contains lots of stuff
3808 that cant be a start char.
3810 fail[ 0 ] = fail[ 1 ] = 0;
3811 DEBUG_TRIE_COMPILE_r({
3812 Perl_re_indentf( aTHX_ "Stclass Failtable (%" UVuf " states): 0",
3813 depth, (UV)numstates
3815 for( q_read=1; q_read<numstates; q_read++ ) {
3816 Perl_re_printf( aTHX_ ", %" UVuf, (UV)fail[q_read]);
3818 Perl_re_printf( aTHX_ "\n");
3821 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3826 /* The below joins as many adjacent EXACTish nodes as possible into a single
3827 * one. The regop may be changed if the node(s) contain certain sequences that
3828 * require special handling. The joining is only done if:
3829 * 1) there is room in the current conglomerated node to entirely contain the
3831 * 2) they are compatible node types
3833 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3834 * these get optimized out
3836 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3837 * as possible, even if that means splitting an existing node so that its first
3838 * part is moved to the preceeding node. This would maximise the efficiency of
3839 * memEQ during matching.
3841 * If a node is to match under /i (folded), the number of characters it matches
3842 * can be different than its character length if it contains a multi-character
3843 * fold. *min_subtract is set to the total delta number of characters of the
3846 * And *unfolded_multi_char is set to indicate whether or not the node contains
3847 * an unfolded multi-char fold. This happens when it won't be known until
3848 * runtime whether the fold is valid or not; namely
3849 * 1) for EXACTF nodes that contain LATIN SMALL LETTER SHARP S, as only if the
3850 * target string being matched against turns out to be UTF-8 is that fold
3852 * 2) for EXACTFL nodes whose folding rules depend on the locale in force at
3854 * (Multi-char folds whose components are all above the Latin1 range are not
3855 * run-time locale dependent, and have already been folded by the time this
3856 * function is called.)
3858 * This is as good a place as any to discuss the design of handling these
3859 * multi-character fold sequences. It's been wrong in Perl for a very long
3860 * time. There are three code points in Unicode whose multi-character folds
3861 * were long ago discovered to mess things up. The previous designs for
3862 * dealing with these involved assigning a special node for them. This
3863 * approach doesn't always work, as evidenced by this example:
3864 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3865 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3866 * would match just the \xDF, it won't be able to handle the case where a
3867 * successful match would have to cross the node's boundary. The new approach
3868 * that hopefully generally solves the problem generates an EXACTFUP node
3869 * that is "sss" in this case.
3871 * It turns out that there are problems with all multi-character folds, and not
3872 * just these three. Now the code is general, for all such cases. The
3873 * approach taken is:
3874 * 1) This routine examines each EXACTFish node that could contain multi-
3875 * character folded sequences. Since a single character can fold into
3876 * such a sequence, the minimum match length for this node is less than
3877 * the number of characters in the node. This routine returns in
3878 * *min_subtract how many characters to subtract from the the actual
3879 * length of the string to get a real minimum match length; it is 0 if
3880 * there are no multi-char foldeds. This delta is used by the caller to
3881 * adjust the min length of the match, and the delta between min and max,
3882 * so that the optimizer doesn't reject these possibilities based on size
3885 * 2) For the sequence involving the LATIN SMALL LETTER SHARP S (U+00DF)
3886 * under /u, we fold it to 'ss' in regatom(), and in this routine, after
3887 * joining, we scan for occurrences of the sequence 'ss' in non-UTF-8
3888 * EXACTFU nodes. The node type of such nodes is then changed to
3889 * EXACTFUP, indicating it is problematic, and needs careful handling.
3890 * (The procedures in step 1) above are sufficient to handle this case in
3891 * UTF-8 encoded nodes.) The reason this is problematic is that this is
3892 * the only case where there is a possible fold length change in non-UTF-8
3893 * patterns. By reserving a special node type for problematic cases, the
3894 * far more common regular EXACTFU nodes can be processed faster.
3895 * regexec.c takes advantage of this.
3897 * EXACTFUP has been created as a grab-bag for (hopefully uncommon)
3898 * problematic cases. These all only occur when the pattern is not
3899 * UTF-8. In addition to the 'ss' sequence where there is a possible fold
3900 * length change, it handles the situation where the string cannot be
3901 * entirely folded. The strings in an EXACTFish node are folded as much
3902 * as possible during compilation in regcomp.c. This saves effort in
3903 * regex matching. By using an EXACTFUP node when it is not possible to
3904 * fully fold at compile time, regexec.c can know that everything in an
3905 * EXACTFU node is folded, so folding can be skipped at runtime. The only
3906 * case where folding in EXACTFU nodes can't be done at compile time is
3907 * the presumably uncommon MICRO SIGN, when the pattern isn't UTF-8. This
3908 * is because its fold requires UTF-8 to represent. Thus EXACTFUP nodes
3909 * handle two very different cases. Alternatively, there could have been
3910 * a node type where there are length changes, one for unfolded, and one
3911 * for both. If yet another special case needed to be created, the number
3912 * of required node types would have to go to 7. khw figures that even
3913 * though there are plenty of node types to spare, that the maintenance
3914 * cost wasn't worth the small speedup of doing it that way, especially
3915 * since he thinks the MICRO SIGN is rarely encountered in practice.
3917 * There are other cases where folding isn't done at compile time, but
3918 * none of them are under /u, and hence not for EXACTFU nodes. The folds
3919 * in EXACTFL nodes aren't known until runtime, and vary as the locale
3920 * changes. Some folds in EXACTF depend on if the runtime target string
3921 * is UTF-8 or not. (regatom() will create an EXACTFU node even under /di
3922 * when no fold in it depends on the UTF-8ness of the target string.)
3924 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3925 * validity of the fold won't be known until runtime, and so must remain
3926 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFAA
3927 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3928 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3929 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3930 * The reason this is a problem is that the optimizer part of regexec.c
3931 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3932 * that a character in the pattern corresponds to at most a single
3933 * character in the target string. (And I do mean character, and not byte
3934 * here, unlike other parts of the documentation that have never been
3935 * updated to account for multibyte Unicode.) Sharp s in EXACTF and
3936 * EXACTFL nodes can match the two character string 'ss'; in EXACTFAA
3937 * nodes it can match "\x{17F}\x{17F}". These, along with other ones in
3938 * EXACTFL nodes, violate the assumption, and they are the only instances
3939 * where it is violated. I'm reluctant to try to change the assumption,
3940 * as the code involved is impenetrable to me (khw), so instead the code
3941 * here punts. This routine examines EXACTFL nodes, and (when the pattern
3942 * isn't UTF-8) EXACTF and EXACTFAA for such unfolded folds, and returns a
3943 * boolean indicating whether or not the node contains such a fold. When
3944 * it is true, the caller sets a flag that later causes the optimizer in
3945 * this file to not set values for the floating and fixed string lengths,
3946 * and thus avoids the optimizer code in regexec.c that makes the invalid
3947 * assumption. Thus, there is no optimization based on string lengths for
3948 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3949 * EXACTF and EXACTFAA nodes that contain the sharp s. (The reason the
3950 * assumption is wrong only in these cases is that all other non-UTF-8
3951 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3952 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3953 * EXACTF nodes because we don't know at compile time if it actually
3954 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3955 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3956 * always matches; and EXACTFAA where it never does. In an EXACTFAA node
3957 * in a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3958 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3959 * string would require the pattern to be forced into UTF-8, the overhead
3960 * of which we want to avoid. Similarly the unfolded multi-char folds in
3961 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3964 * Similarly, the code that generates tries doesn't currently handle
3965 * not-already-folded multi-char folds, and it looks like a pain to change
3966 * that. Therefore, trie generation of EXACTFAA nodes with the sharp s
3967 * doesn't work. Instead, such an EXACTFAA is turned into a new regnode,
3968 * EXACTFAA_NO_TRIE, which the trie code knows not to handle. Most people
3969 * using /iaa matching will be doing so almost entirely with ASCII
3970 * strings, so this should rarely be encountered in practice */
3972 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3973 if (PL_regkind[OP(scan)] == EXACT) \
3974 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags), NULL, depth+1)
3977 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3978 UV *min_subtract, bool *unfolded_multi_char,
3979 U32 flags, regnode *val, U32 depth)
3981 /* Merge several consecutive EXACTish nodes into one. */
3983 regnode *n = regnext(scan);
3985 regnode *next = scan + NODE_SZ_STR(scan);
3989 regnode *stop = scan;
3990 GET_RE_DEBUG_FLAGS_DECL;
3992 PERL_UNUSED_ARG(depth);