5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
10 /* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
12 * a regular expression.
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
20 /* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
24 /* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
29 /* The names of the functions have been changed from regcomp and
30 * regexec to pregcomp and pregexec in order to avoid conflicts
31 * with the POSIX routines of the same names.
34 #ifdef PERL_EXT_RE_BUILD
39 * pregcomp and pregexec -- regsub and regerror are not used in perl
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
59 **** Alterations to Henry's code are...
61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
74 #define PERL_IN_REGCOMP_C
77 #ifndef PERL_IN_XSUB_RE
82 #ifdef PERL_IN_XSUB_RE
84 EXTERN_C const struct regexp_engine my_reg_engine;
89 #include "dquote_inline.h"
90 #include "invlist_inline.h"
91 #include "unicode_constants.h"
93 #define HAS_NONLATIN1_FOLD_CLOSURE(i) \
94 _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
95 #define HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(i) \
96 _HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
97 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
98 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
101 #define STATIC static
105 #define MIN(a,b) ((a) < (b) ? (a) : (b))
109 #define MAX(a,b) ((a) > (b) ? (a) : (b))
112 /* this is a chain of data about sub patterns we are processing that
113 need to be handled separately/specially in study_chunk. Its so
114 we can simulate recursion without losing state. */
116 typedef struct scan_frame {
117 regnode *last_regnode; /* last node to process in this frame */
118 regnode *next_regnode; /* next node to process when last is reached */
119 U32 prev_recursed_depth;
120 I32 stopparen; /* what stopparen do we use */
121 U32 is_top_frame; /* what flags do we use? */
123 struct scan_frame *this_prev_frame; /* this previous frame */
124 struct scan_frame *prev_frame; /* previous frame */
125 struct scan_frame *next_frame; /* next frame */
128 /* Certain characters are output as a sequence with the first being a
130 #define isBACKSLASHED_PUNCT(c) \
131 ((c) == '-' || (c) == ']' || (c) == '\\' || (c) == '^')
134 struct RExC_state_t {
135 U32 flags; /* RXf_* are we folding, multilining? */
136 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
137 char *precomp; /* uncompiled string. */
138 char *precomp_end; /* pointer to end of uncompiled string. */
139 REGEXP *rx_sv; /* The SV that is the regexp. */
140 regexp *rx; /* perl core regexp structure */
141 regexp_internal *rxi; /* internal data for regexp object
143 char *start; /* Start of input for compile */
144 char *end; /* End of input for compile */
145 char *parse; /* Input-scan pointer. */
146 char *adjusted_start; /* 'start', adjusted. See code use */
147 STRLEN precomp_adj; /* an offset beyond precomp. See code use */
148 SSize_t whilem_seen; /* number of WHILEM in this expr */
149 regnode *emit_start; /* Start of emitted-code area */
150 regnode *emit_bound; /* First regnode outside of the
152 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
153 implies compiling, so don't emit */
154 regnode_ssc emit_dummy; /* placeholder for emit to point to;
155 large enough for the largest
156 non-EXACTish node, so can use it as
158 I32 naughty; /* How bad is this pattern? */
159 I32 sawback; /* Did we see \1, ...? */
161 SSize_t size; /* Code size. */
162 I32 npar; /* Capture buffer count, (OPEN) plus
163 one. ("par" 0 is the whole
165 I32 nestroot; /* root parens we are in - used by
169 regnode **open_parens; /* pointers to open parens */
170 regnode **close_parens; /* pointers to close parens */
171 regnode *end_op; /* END node in program */
172 I32 utf8; /* whether the pattern is utf8 or not */
173 I32 orig_utf8; /* whether the pattern was originally in utf8 */
174 /* XXX use this for future optimisation of case
175 * where pattern must be upgraded to utf8. */
176 I32 uni_semantics; /* If a d charset modifier should use unicode
177 rules, even if the pattern is not in
179 HV *paren_names; /* Paren names */
181 regnode **recurse; /* Recurse regops */
182 I32 recurse_count; /* Number of recurse regops we have generated */
183 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
185 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
189 I32 override_recoding;
191 I32 recode_x_to_native;
193 I32 in_multi_char_class;
194 struct reg_code_block *code_blocks; /* positions of literal (?{})
196 int num_code_blocks; /* size of code_blocks[] */
197 int code_index; /* next code_blocks[] slot */
198 SSize_t maxlen; /* mininum possible number of chars in string to match */
199 scan_frame *frame_head;
200 scan_frame *frame_last;
202 #ifdef ADD_TO_REGEXEC
203 char *starttry; /* -Dr: where regtry was called. */
204 #define RExC_starttry (pRExC_state->starttry)
206 SV *runtime_code_qr; /* qr with the runtime code blocks */
208 const char *lastparse;
210 AV *paren_name_list; /* idx -> name */
211 U32 study_chunk_recursed_count;
214 #define RExC_lastparse (pRExC_state->lastparse)
215 #define RExC_lastnum (pRExC_state->lastnum)
216 #define RExC_paren_name_list (pRExC_state->paren_name_list)
217 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
218 #define RExC_mysv (pRExC_state->mysv1)
219 #define RExC_mysv1 (pRExC_state->mysv1)
220 #define RExC_mysv2 (pRExC_state->mysv2)
223 bool seen_unfolded_sharp_s;
227 #define RExC_flags (pRExC_state->flags)
228 #define RExC_pm_flags (pRExC_state->pm_flags)
229 #define RExC_precomp (pRExC_state->precomp)
230 #define RExC_precomp_adj (pRExC_state->precomp_adj)
231 #define RExC_adjusted_start (pRExC_state->adjusted_start)
232 #define RExC_precomp_end (pRExC_state->precomp_end)
233 #define RExC_rx_sv (pRExC_state->rx_sv)
234 #define RExC_rx (pRExC_state->rx)
235 #define RExC_rxi (pRExC_state->rxi)
236 #define RExC_start (pRExC_state->start)
237 #define RExC_end (pRExC_state->end)
238 #define RExC_parse (pRExC_state->parse)
239 #define RExC_whilem_seen (pRExC_state->whilem_seen)
241 /* Set during the sizing pass when there is a LATIN SMALL LETTER SHARP S in any
242 * EXACTF node, hence was parsed under /di rules. If later in the parse,
243 * something forces the pattern into using /ui rules, the sharp s should be
244 * folded into the sequence 'ss', which takes up more space than previously
245 * calculated. This means that the sizing pass needs to be restarted. (The
246 * node also becomes an EXACTFU_SS.) For all other characters, an EXACTF node
247 * that gets converted to /ui (and EXACTFU) occupies the same amount of space,
248 * so there is no need to resize [perl #125990]. */
249 #define RExC_seen_unfolded_sharp_s (pRExC_state->seen_unfolded_sharp_s)
251 #ifdef RE_TRACK_PATTERN_OFFSETS
252 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
255 #define RExC_emit (pRExC_state->emit)
256 #define RExC_emit_dummy (pRExC_state->emit_dummy)
257 #define RExC_emit_start (pRExC_state->emit_start)
258 #define RExC_emit_bound (pRExC_state->emit_bound)
259 #define RExC_sawback (pRExC_state->sawback)
260 #define RExC_seen (pRExC_state->seen)
261 #define RExC_size (pRExC_state->size)
262 #define RExC_maxlen (pRExC_state->maxlen)
263 #define RExC_npar (pRExC_state->npar)
264 #define RExC_nestroot (pRExC_state->nestroot)
265 #define RExC_extralen (pRExC_state->extralen)
266 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
267 #define RExC_utf8 (pRExC_state->utf8)
268 #define RExC_uni_semantics (pRExC_state->uni_semantics)
269 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
270 #define RExC_open_parens (pRExC_state->open_parens)
271 #define RExC_close_parens (pRExC_state->close_parens)
272 #define RExC_end_op (pRExC_state->end_op)
273 #define RExC_paren_names (pRExC_state->paren_names)
274 #define RExC_recurse (pRExC_state->recurse)
275 #define RExC_recurse_count (pRExC_state->recurse_count)
276 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
277 #define RExC_study_chunk_recursed_bytes \
278 (pRExC_state->study_chunk_recursed_bytes)
279 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
280 #define RExC_contains_locale (pRExC_state->contains_locale)
281 #define RExC_contains_i (pRExC_state->contains_i)
282 #define RExC_override_recoding (pRExC_state->override_recoding)
284 # define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
286 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
287 #define RExC_frame_head (pRExC_state->frame_head)
288 #define RExC_frame_last (pRExC_state->frame_last)
289 #define RExC_frame_count (pRExC_state->frame_count)
290 #define RExC_strict (pRExC_state->strict)
292 /* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
293 * a flag to disable back-off on the fixed/floating substrings - if it's
294 * a high complexity pattern we assume the benefit of avoiding a full match
295 * is worth the cost of checking for the substrings even if they rarely help.
297 #define RExC_naughty (pRExC_state->naughty)
298 #define TOO_NAUGHTY (10)
299 #define MARK_NAUGHTY(add) \
300 if (RExC_naughty < TOO_NAUGHTY) \
301 RExC_naughty += (add)
302 #define MARK_NAUGHTY_EXP(exp, add) \
303 if (RExC_naughty < TOO_NAUGHTY) \
304 RExC_naughty += RExC_naughty / (exp) + (add)
306 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
307 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
308 ((*s) == '{' && regcurly(s)))
311 * Flags to be passed up and down.
313 #define WORST 0 /* Worst case. */
314 #define HASWIDTH 0x01 /* Known to match non-null strings. */
316 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
317 * character. (There needs to be a case: in the switch statement in regexec.c
318 * for any node marked SIMPLE.) Note that this is not the same thing as
321 #define SPSTART 0x04 /* Starts with * or + */
322 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
323 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
324 #define RESTART_PASS1 0x20 /* Need to restart sizing pass */
325 #define NEED_UTF8 0x40 /* In conjunction with RESTART_PASS1, need to
326 calcuate sizes as UTF-8 */
328 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
330 /* whether trie related optimizations are enabled */
331 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
332 #define TRIE_STUDY_OPT
333 #define FULL_TRIE_STUDY
339 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
340 #define PBITVAL(paren) (1 << ((paren) & 7))
341 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
342 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
343 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
345 #define REQUIRE_UTF8(flagp) STMT_START { \
348 *flagp = RESTART_PASS1|NEED_UTF8; \
353 /* Change from /d into /u rules, and restart the parse if we've already seen
354 * something whose size would increase as a result, by setting *flagp and
355 * returning 'restart_retval'. RExC_uni_semantics is a flag that indicates
356 * we've change to /u during the parse. */
357 #define REQUIRE_UNI_RULES(flagp, restart_retval) \
359 if (DEPENDS_SEMANTICS) { \
361 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
362 RExC_uni_semantics = 1; \
363 if (RExC_seen_unfolded_sharp_s) { \
364 *flagp |= RESTART_PASS1; \
365 return restart_retval; \
370 /* This converts the named class defined in regcomp.h to its equivalent class
371 * number defined in handy.h. */
372 #define namedclass_to_classnum(class) ((int) ((class) / 2))
373 #define classnum_to_namedclass(classnum) ((classnum) * 2)
375 #define _invlist_union_complement_2nd(a, b, output) \
376 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
377 #define _invlist_intersection_complement_2nd(a, b, output) \
378 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
380 /* About scan_data_t.
382 During optimisation we recurse through the regexp program performing
383 various inplace (keyhole style) optimisations. In addition study_chunk
384 and scan_commit populate this data structure with information about
385 what strings MUST appear in the pattern. We look for the longest
386 string that must appear at a fixed location, and we look for the
387 longest string that may appear at a floating location. So for instance
392 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
393 strings (because they follow a .* construct). study_chunk will identify
394 both FOO and BAR as being the longest fixed and floating strings respectively.
396 The strings can be composites, for instance
400 will result in a composite fixed substring 'foo'.
402 For each string some basic information is maintained:
404 - offset or min_offset
405 This is the position the string must appear at, or not before.
406 It also implicitly (when combined with minlenp) tells us how many
407 characters must match before the string we are searching for.
408 Likewise when combined with minlenp and the length of the string it
409 tells us how many characters must appear after the string we have
413 Only used for floating strings. This is the rightmost point that
414 the string can appear at. If set to SSize_t_MAX it indicates that the
415 string can occur infinitely far to the right.
418 A pointer to the minimum number of characters of the pattern that the
419 string was found inside. This is important as in the case of positive
420 lookahead or positive lookbehind we can have multiple patterns
425 The minimum length of the pattern overall is 3, the minimum length
426 of the lookahead part is 3, but the minimum length of the part that
427 will actually match is 1. So 'FOO's minimum length is 3, but the
428 minimum length for the F is 1. This is important as the minimum length
429 is used to determine offsets in front of and behind the string being
430 looked for. Since strings can be composites this is the length of the
431 pattern at the time it was committed with a scan_commit. Note that
432 the length is calculated by study_chunk, so that the minimum lengths
433 are not known until the full pattern has been compiled, thus the
434 pointer to the value.
438 In the case of lookbehind the string being searched for can be
439 offset past the start point of the final matching string.
440 If this value was just blithely removed from the min_offset it would
441 invalidate some of the calculations for how many chars must match
442 before or after (as they are derived from min_offset and minlen and
443 the length of the string being searched for).
444 When the final pattern is compiled and the data is moved from the
445 scan_data_t structure into the regexp structure the information
446 about lookbehind is factored in, with the information that would
447 have been lost precalculated in the end_shift field for the
450 The fields pos_min and pos_delta are used to store the minimum offset
451 and the delta to the maximum offset at the current point in the pattern.
455 typedef struct scan_data_t {
456 /*I32 len_min; unused */
457 /*I32 len_delta; unused */
461 SSize_t last_end; /* min value, <0 unless valid. */
462 SSize_t last_start_min;
463 SSize_t last_start_max;
464 SV **longest; /* Either &l_fixed, or &l_float. */
465 SV *longest_fixed; /* longest fixed string found in pattern */
466 SSize_t offset_fixed; /* offset where it starts */
467 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
468 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
469 SV *longest_float; /* longest floating string found in pattern */
470 SSize_t offset_float_min; /* earliest point in string it can appear */
471 SSize_t offset_float_max; /* latest point in string it can appear */
472 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
473 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
476 SSize_t *last_closep;
477 regnode_ssc *start_class;
481 * Forward declarations for pregcomp()'s friends.
484 static const scan_data_t zero_scan_data =
485 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
487 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
488 #define SF_BEFORE_SEOL 0x0001
489 #define SF_BEFORE_MEOL 0x0002
490 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
491 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
493 #define SF_FIX_SHIFT_EOL (+2)
494 #define SF_FL_SHIFT_EOL (+4)
496 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
497 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
499 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
500 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
501 #define SF_IS_INF 0x0040
502 #define SF_HAS_PAR 0x0080
503 #define SF_IN_PAR 0x0100
504 #define SF_HAS_EVAL 0x0200
505 #define SCF_DO_SUBSTR 0x0400
506 #define SCF_DO_STCLASS_AND 0x0800
507 #define SCF_DO_STCLASS_OR 0x1000
508 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
509 #define SCF_WHILEM_VISITED_POS 0x2000
511 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
512 #define SCF_SEEN_ACCEPT 0x8000
513 #define SCF_TRIE_DOING_RESTUDY 0x10000
514 #define SCF_IN_DEFINE 0x20000
519 #define UTF cBOOL(RExC_utf8)
521 /* The enums for all these are ordered so things work out correctly */
522 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
523 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
524 == REGEX_DEPENDS_CHARSET)
525 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
526 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
527 >= REGEX_UNICODE_CHARSET)
528 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
529 == REGEX_ASCII_RESTRICTED_CHARSET)
530 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
531 >= REGEX_ASCII_RESTRICTED_CHARSET)
532 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
533 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
535 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
537 /* For programs that want to be strictly Unicode compatible by dying if any
538 * attempt is made to match a non-Unicode code point against a Unicode
540 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
542 #define OOB_NAMEDCLASS -1
544 /* There is no code point that is out-of-bounds, so this is problematic. But
545 * its only current use is to initialize a variable that is always set before
547 #define OOB_UNICODE 0xDEADBEEF
549 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
550 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
553 /* length of regex to show in messages that don't mark a position within */
554 #define RegexLengthToShowInErrorMessages 127
557 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
558 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
559 * op/pragma/warn/regcomp.
561 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
562 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
564 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
565 " in m/%"UTF8f MARKER2 "%"UTF8f"/"
567 /* The code in this file in places uses one level of recursion with parsing
568 * rebased to an alternate string constructed by us in memory. This can take
569 * the form of something that is completely different from the input, or
570 * something that uses the input as part of the alternate. In the first case,
571 * there should be no possibility of an error, as we are in complete control of
572 * the alternate string. But in the second case we don't control the input
573 * portion, so there may be errors in that. Here's an example:
575 * is handled specially because \x{df} folds to a sequence of more than one
576 * character, 'ss'. What is done is to create and parse an alternate string,
577 * which looks like this:
578 * /(?:\x{DF}|[abc\x{DF}def])/ui
579 * where it uses the input unchanged in the middle of something it constructs,
580 * which is a branch for the DF outside the character class, and clustering
581 * parens around the whole thing. (It knows enough to skip the DF inside the
582 * class while in this substitute parse.) 'abc' and 'def' may have errors that
583 * need to be reported. The general situation looks like this:
586 * Input: ----------------------------------------------------
587 * Constructed: ---------------------------------------------------
590 * The input string sI..eI is the input pattern. The string sC..EC is the
591 * constructed substitute parse string. The portions sC..tC and eC..EC are
592 * constructed by us. The portion tC..eC is an exact duplicate of the input
593 * pattern tI..eI. In the diagram, these are vertically aligned. Suppose that
594 * while parsing, we find an error at xC. We want to display a message showing
595 * the real input string. Thus we need to find the point xI in it which
596 * corresponds to xC. xC >= tC, since the portion of the string sC..tC has
597 * been constructed by us, and so shouldn't have errors. We get:
599 * xI = sI + (tI - sI) + (xC - tC)
601 * and, the offset into sI is:
603 * (xI - sI) = (tI - sI) + (xC - tC)
605 * When the substitute is constructed, we save (tI -sI) as RExC_precomp_adj,
606 * and we save tC as RExC_adjusted_start.
608 * During normal processing of the input pattern, everything points to that,
609 * with RExC_precomp_adj set to 0, and RExC_adjusted_start set to sI.
612 #define tI_sI RExC_precomp_adj
613 #define tC RExC_adjusted_start
614 #define sC RExC_precomp
615 #define xI_offset(xC) ((IV) (tI_sI + (xC - tC)))
616 #define xI(xC) (sC + xI_offset(xC))
617 #define eC RExC_precomp_end
619 #define REPORT_LOCATION_ARGS(xC) \
621 (xI(xC) > eC) /* Don't run off end */ \
622 ? eC - sC /* Length before the <--HERE */ \
624 sC), /* The input pattern printed up to the <--HERE */ \
626 (xI(xC) > eC) ? 0 : eC - xI(xC), /* Length after <--HERE */ \
627 (xI(xC) > eC) ? eC : xI(xC)) /* pattern after <--HERE */
629 /* Used to point after bad bytes for an error message, but avoid skipping
630 * past a nul byte. */
631 #define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
634 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
635 * arg. Show regex, up to a maximum length. If it's too long, chop and add
638 #define _FAIL(code) STMT_START { \
639 const char *ellipses = ""; \
640 IV len = RExC_precomp_end - RExC_precomp; \
643 SAVEFREESV(RExC_rx_sv); \
644 if (len > RegexLengthToShowInErrorMessages) { \
645 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
646 len = RegexLengthToShowInErrorMessages - 10; \
652 #define FAIL(msg) _FAIL( \
653 Perl_croak(aTHX_ "%s in regex m/%"UTF8f"%s/", \
654 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
656 #define FAIL2(msg,arg) _FAIL( \
657 Perl_croak(aTHX_ msg " in regex m/%"UTF8f"%s/", \
658 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
661 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
663 #define Simple_vFAIL(m) STMT_START { \
664 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
665 m, REPORT_LOCATION_ARGS(RExC_parse)); \
669 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
671 #define vFAIL(m) STMT_START { \
673 SAVEFREESV(RExC_rx_sv); \
678 * Like Simple_vFAIL(), but accepts two arguments.
680 #define Simple_vFAIL2(m,a1) STMT_START { \
681 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
682 REPORT_LOCATION_ARGS(RExC_parse)); \
686 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
688 #define vFAIL2(m,a1) STMT_START { \
690 SAVEFREESV(RExC_rx_sv); \
691 Simple_vFAIL2(m, a1); \
696 * Like Simple_vFAIL(), but accepts three arguments.
698 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
699 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
700 REPORT_LOCATION_ARGS(RExC_parse)); \
704 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
706 #define vFAIL3(m,a1,a2) STMT_START { \
708 SAVEFREESV(RExC_rx_sv); \
709 Simple_vFAIL3(m, a1, a2); \
713 * Like Simple_vFAIL(), but accepts four arguments.
715 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
716 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
717 REPORT_LOCATION_ARGS(RExC_parse)); \
720 #define vFAIL4(m,a1,a2,a3) STMT_START { \
722 SAVEFREESV(RExC_rx_sv); \
723 Simple_vFAIL4(m, a1, a2, a3); \
726 /* A specialized version of vFAIL2 that works with UTF8f */
727 #define vFAIL2utf8f(m, a1) STMT_START { \
729 SAVEFREESV(RExC_rx_sv); \
730 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
731 REPORT_LOCATION_ARGS(RExC_parse)); \
734 #define vFAIL3utf8f(m, a1, a2) STMT_START { \
736 SAVEFREESV(RExC_rx_sv); \
737 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
738 REPORT_LOCATION_ARGS(RExC_parse)); \
741 /* These have asserts in them because of [perl #122671] Many warnings in
742 * regcomp.c can occur twice. If they get output in pass1 and later in that
743 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
744 * would get output again. So they should be output in pass2, and these
745 * asserts make sure new warnings follow that paradigm. */
747 /* m is not necessarily a "literal string", in this macro */
748 #define reg_warn_non_literal_string(loc, m) STMT_START { \
749 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
750 "%s" REPORT_LOCATION, \
751 m, REPORT_LOCATION_ARGS(loc)); \
754 #define ckWARNreg(loc,m) STMT_START { \
755 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
757 REPORT_LOCATION_ARGS(loc)); \
760 #define vWARN(loc, m) STMT_START { \
761 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
763 REPORT_LOCATION_ARGS(loc)); \
766 #define vWARN_dep(loc, m) STMT_START { \
767 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
769 REPORT_LOCATION_ARGS(loc)); \
772 #define ckWARNdep(loc,m) STMT_START { \
773 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
775 REPORT_LOCATION_ARGS(loc)); \
778 #define ckWARNregdep(loc,m) STMT_START { \
779 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
782 REPORT_LOCATION_ARGS(loc)); \
785 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
786 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
788 a1, REPORT_LOCATION_ARGS(loc)); \
791 #define ckWARN2reg(loc, m, a1) STMT_START { \
792 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
794 a1, REPORT_LOCATION_ARGS(loc)); \
797 #define vWARN3(loc, m, a1, a2) STMT_START { \
798 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
800 a1, a2, REPORT_LOCATION_ARGS(loc)); \
803 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
804 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
807 REPORT_LOCATION_ARGS(loc)); \
810 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
811 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
814 REPORT_LOCATION_ARGS(loc)); \
817 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
818 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
821 REPORT_LOCATION_ARGS(loc)); \
824 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
825 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
828 REPORT_LOCATION_ARGS(loc)); \
831 /* Macros for recording node offsets. 20001227 mjd@plover.com
832 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
833 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
834 * Element 0 holds the number n.
835 * Position is 1 indexed.
837 #ifndef RE_TRACK_PATTERN_OFFSETS
838 #define Set_Node_Offset_To_R(node,byte)
839 #define Set_Node_Offset(node,byte)
840 #define Set_Cur_Node_Offset
841 #define Set_Node_Length_To_R(node,len)
842 #define Set_Node_Length(node,len)
843 #define Set_Node_Cur_Length(node,start)
844 #define Node_Offset(n)
845 #define Node_Length(n)
846 #define Set_Node_Offset_Length(node,offset,len)
847 #define ProgLen(ri) ri->u.proglen
848 #define SetProgLen(ri,x) ri->u.proglen = x
850 #define ProgLen(ri) ri->u.offsets[0]
851 #define SetProgLen(ri,x) ri->u.offsets[0] = x
852 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
854 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
855 __LINE__, (int)(node), (int)(byte))); \
857 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
860 RExC_offsets[2*(node)-1] = (byte); \
865 #define Set_Node_Offset(node,byte) \
866 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
867 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
869 #define Set_Node_Length_To_R(node,len) STMT_START { \
871 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
872 __LINE__, (int)(node), (int)(len))); \
874 Perl_croak(aTHX_ "value of node is %d in Length macro", \
877 RExC_offsets[2*(node)] = (len); \
882 #define Set_Node_Length(node,len) \
883 Set_Node_Length_To_R((node)-RExC_emit_start, len)
884 #define Set_Node_Cur_Length(node, start) \
885 Set_Node_Length(node, RExC_parse - start)
887 /* Get offsets and lengths */
888 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
889 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
891 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
892 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
893 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
897 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
898 #define EXPERIMENTAL_INPLACESCAN
899 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
901 #define DEBUG_RExC_seen() \
902 DEBUG_OPTIMISE_MORE_r({ \
903 PerlIO_printf(Perl_debug_log,"RExC_seen: "); \
905 if (RExC_seen & REG_ZERO_LEN_SEEN) \
906 PerlIO_printf(Perl_debug_log,"REG_ZERO_LEN_SEEN "); \
908 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
909 PerlIO_printf(Perl_debug_log,"REG_LOOKBEHIND_SEEN "); \
911 if (RExC_seen & REG_GPOS_SEEN) \
912 PerlIO_printf(Perl_debug_log,"REG_GPOS_SEEN "); \
914 if (RExC_seen & REG_RECURSE_SEEN) \
915 PerlIO_printf(Perl_debug_log,"REG_RECURSE_SEEN "); \
917 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
918 PerlIO_printf(Perl_debug_log,"REG_TOP_LEVEL_BRANCHES_SEEN "); \
920 if (RExC_seen & REG_VERBARG_SEEN) \
921 PerlIO_printf(Perl_debug_log,"REG_VERBARG_SEEN "); \
923 if (RExC_seen & REG_CUTGROUP_SEEN) \
924 PerlIO_printf(Perl_debug_log,"REG_CUTGROUP_SEEN "); \
926 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
927 PerlIO_printf(Perl_debug_log,"REG_RUN_ON_COMMENT_SEEN "); \
929 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
930 PerlIO_printf(Perl_debug_log,"REG_UNFOLDED_MULTI_SEEN "); \
932 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
933 PerlIO_printf(Perl_debug_log,"REG_UNBOUNDED_QUANTIFIER_SEEN "); \
935 PerlIO_printf(Perl_debug_log,"\n"); \
938 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
939 if ((flags) & flag) PerlIO_printf(Perl_debug_log, "%s ", #flag)
941 #define DEBUG_SHOW_STUDY_FLAGS(flags,open_str,close_str) \
943 PerlIO_printf(Perl_debug_log, "%s", open_str); \
944 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_SEOL); \
945 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_MEOL); \
946 DEBUG_SHOW_STUDY_FLAG(flags,SF_IS_INF); \
947 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_PAR); \
948 DEBUG_SHOW_STUDY_FLAG(flags,SF_IN_PAR); \
949 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_EVAL); \
950 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_SUBSTR); \
951 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_AND); \
952 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_OR); \
953 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS); \
954 DEBUG_SHOW_STUDY_FLAG(flags,SCF_WHILEM_VISITED_POS); \
955 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_RESTUDY); \
956 DEBUG_SHOW_STUDY_FLAG(flags,SCF_SEEN_ACCEPT); \
957 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_DOING_RESTUDY); \
958 DEBUG_SHOW_STUDY_FLAG(flags,SCF_IN_DEFINE); \
959 PerlIO_printf(Perl_debug_log, "%s", close_str); \
963 #define DEBUG_STUDYDATA(str,data,depth) \
964 DEBUG_OPTIMISE_MORE_r(if(data){ \
965 PerlIO_printf(Perl_debug_log, \
966 "%*s" str "Pos:%"IVdf"/%"IVdf \
968 (int)(depth)*2, "", \
969 (IV)((data)->pos_min), \
970 (IV)((data)->pos_delta), \
971 (UV)((data)->flags) \
973 DEBUG_SHOW_STUDY_FLAGS((data)->flags," [ ","]"); \
974 PerlIO_printf(Perl_debug_log, \
975 " Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
976 (IV)((data)->whilem_c), \
977 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
978 is_inf ? "INF " : "" \
980 if ((data)->last_found) \
981 PerlIO_printf(Perl_debug_log, \
982 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
983 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
984 SvPVX_const((data)->last_found), \
985 (IV)((data)->last_end), \
986 (IV)((data)->last_start_min), \
987 (IV)((data)->last_start_max), \
988 ((data)->longest && \
989 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
990 SvPVX_const((data)->longest_fixed), \
991 (IV)((data)->offset_fixed), \
992 ((data)->longest && \
993 (data)->longest==&((data)->longest_float)) ? "*" : "", \
994 SvPVX_const((data)->longest_float), \
995 (IV)((data)->offset_float_min), \
996 (IV)((data)->offset_float_max) \
998 PerlIO_printf(Perl_debug_log,"\n"); \
1001 /* =========================================================
1002 * BEGIN edit_distance stuff.
1004 * This calculates how many single character changes of any type are needed to
1005 * transform a string into another one. It is taken from version 3.1 of
1007 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1010 /* Our unsorted dictionary linked list. */
1011 /* Note we use UVs, not chars. */
1016 struct dictionary* next;
1018 typedef struct dictionary item;
1021 PERL_STATIC_INLINE item*
1022 push(UV key,item* curr)
1025 Newxz(head, 1, item);
1033 PERL_STATIC_INLINE item*
1034 find(item* head, UV key)
1036 item* iterator = head;
1038 if (iterator->key == key){
1041 iterator = iterator->next;
1047 PERL_STATIC_INLINE item*
1048 uniquePush(item* head,UV key)
1050 item* iterator = head;
1053 if (iterator->key == key) {
1056 iterator = iterator->next;
1059 return push(key,head);
1062 PERL_STATIC_INLINE void
1063 dict_free(item* head)
1065 item* iterator = head;
1068 item* temp = iterator;
1069 iterator = iterator->next;
1076 /* End of Dictionary Stuff */
1078 /* All calculations/work are done here */
1080 S_edit_distance(const UV* src,
1082 const STRLEN x, /* length of src[] */
1083 const STRLEN y, /* length of tgt[] */
1084 const SSize_t maxDistance
1088 UV swapCount,swapScore,targetCharCount,i,j;
1090 UV score_ceil = x + y;
1092 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1094 /* intialize matrix start values */
1095 Newxz(scores, ( (x + 2) * (y + 2)), UV);
1096 scores[0] = score_ceil;
1097 scores[1 * (y + 2) + 0] = score_ceil;
1098 scores[0 * (y + 2) + 1] = score_ceil;
1099 scores[1 * (y + 2) + 1] = 0;
1100 head = uniquePush(uniquePush(head,src[0]),tgt[0]);
1105 for (i=1;i<=x;i++) {
1107 head = uniquePush(head,src[i]);
1108 scores[(i+1) * (y + 2) + 1] = i;
1109 scores[(i+1) * (y + 2) + 0] = score_ceil;
1112 for (j=1;j<=y;j++) {
1115 head = uniquePush(head,tgt[j]);
1116 scores[1 * (y + 2) + (j + 1)] = j;
1117 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1120 targetCharCount = find(head,tgt[j-1])->value;
1121 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1123 if (src[i-1] != tgt[j-1]){
1124 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));
1128 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1132 find(head,src[i-1])->value = i;
1136 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1139 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1143 /* END of edit_distance() stuff
1144 * ========================================================= */
1146 /* is c a control character for which we have a mnemonic? */
1147 #define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
1150 S_cntrl_to_mnemonic(const U8 c)
1152 /* Returns the mnemonic string that represents character 'c', if one
1153 * exists; NULL otherwise. The only ones that exist for the purposes of
1154 * this routine are a few control characters */
1157 case '\a': return "\\a";
1158 case '\b': return "\\b";
1159 case ESC_NATIVE: return "\\e";
1160 case '\f': return "\\f";
1161 case '\n': return "\\n";
1162 case '\r': return "\\r";
1163 case '\t': return "\\t";
1169 /* Mark that we cannot extend a found fixed substring at this point.
1170 Update the longest found anchored substring and the longest found
1171 floating substrings if needed. */
1174 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1175 SSize_t *minlenp, int is_inf)
1177 const STRLEN l = CHR_SVLEN(data->last_found);
1178 const STRLEN old_l = CHR_SVLEN(*data->longest);
1179 GET_RE_DEBUG_FLAGS_DECL;
1181 PERL_ARGS_ASSERT_SCAN_COMMIT;
1183 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
1184 SvSetMagicSV(*data->longest, data->last_found);
1185 if (*data->longest == data->longest_fixed) {
1186 data->offset_fixed = l ? data->last_start_min : data->pos_min;
1187 if (data->flags & SF_BEFORE_EOL)
1189 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
1191 data->flags &= ~SF_FIX_BEFORE_EOL;
1192 data->minlen_fixed=minlenp;
1193 data->lookbehind_fixed=0;
1195 else { /* *data->longest == data->longest_float */
1196 data->offset_float_min = l ? data->last_start_min : data->pos_min;
1197 data->offset_float_max = (l
1198 ? data->last_start_max
1199 : (data->pos_delta > SSize_t_MAX - data->pos_min
1201 : data->pos_min + data->pos_delta));
1203 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
1204 data->offset_float_max = SSize_t_MAX;
1205 if (data->flags & SF_BEFORE_EOL)
1207 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
1209 data->flags &= ~SF_FL_BEFORE_EOL;
1210 data->minlen_float=minlenp;
1211 data->lookbehind_float=0;
1214 SvCUR_set(data->last_found, 0);
1216 SV * const sv = data->last_found;
1217 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1218 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1223 data->last_end = -1;
1224 data->flags &= ~SF_BEFORE_EOL;
1225 DEBUG_STUDYDATA("commit: ",data,0);
1228 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1229 * list that describes which code points it matches */
1232 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1234 /* Set the SSC 'ssc' to match an empty string or any code point */
1236 PERL_ARGS_ASSERT_SSC_ANYTHING;
1238 assert(is_ANYOF_SYNTHETIC(ssc));
1240 ssc->invlist = sv_2mortal(_new_invlist(2)); /* mortalize so won't leak */
1241 _append_range_to_invlist(ssc->invlist, 0, UV_MAX);
1242 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1246 S_ssc_is_anything(const regnode_ssc *ssc)
1248 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1249 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1250 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1251 * in any way, so there's no point in using it */
1256 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1258 assert(is_ANYOF_SYNTHETIC(ssc));
1260 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1264 /* See if the list consists solely of the range 0 - Infinity */
1265 invlist_iterinit(ssc->invlist);
1266 ret = invlist_iternext(ssc->invlist, &start, &end)
1270 invlist_iterfinish(ssc->invlist);
1276 /* If e.g., both \w and \W are set, matches everything */
1277 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1279 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1280 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1290 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1292 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1293 * string, any code point, or any posix class under locale */
1295 PERL_ARGS_ASSERT_SSC_INIT;
1297 Zero(ssc, 1, regnode_ssc);
1298 set_ANYOF_SYNTHETIC(ssc);
1299 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1302 /* If any portion of the regex is to operate under locale rules that aren't
1303 * fully known at compile time, initialization includes it. The reason
1304 * this isn't done for all regexes is that the optimizer was written under
1305 * the assumption that locale was all-or-nothing. Given the complexity and
1306 * lack of documentation in the optimizer, and that there are inadequate
1307 * test cases for locale, many parts of it may not work properly, it is
1308 * safest to avoid locale unless necessary. */
1309 if (RExC_contains_locale) {
1310 ANYOF_POSIXL_SETALL(ssc);
1313 ANYOF_POSIXL_ZERO(ssc);
1318 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1319 const regnode_ssc *ssc)
1321 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1322 * to the list of code points matched, and locale posix classes; hence does
1323 * not check its flags) */
1328 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1330 assert(is_ANYOF_SYNTHETIC(ssc));
1332 invlist_iterinit(ssc->invlist);
1333 ret = invlist_iternext(ssc->invlist, &start, &end)
1337 invlist_iterfinish(ssc->invlist);
1343 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1351 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1352 const regnode_charclass* const node)
1354 /* Returns a mortal inversion list defining which code points are matched
1355 * by 'node', which is of type ANYOF. Handles complementing the result if
1356 * appropriate. If some code points aren't knowable at this time, the
1357 * returned list must, and will, contain every code point that is a
1361 SV* only_utf8_locale_invlist = NULL;
1363 const U32 n = ARG(node);
1364 bool new_node_has_latin1 = FALSE;
1366 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1368 /* Look at the data structure created by S_set_ANYOF_arg() */
1369 if (n != ANYOF_ONLY_HAS_BITMAP) {
1370 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1371 AV * const av = MUTABLE_AV(SvRV(rv));
1372 SV **const ary = AvARRAY(av);
1373 assert(RExC_rxi->data->what[n] == 's');
1375 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1376 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1378 else if (ary[0] && ary[0] != &PL_sv_undef) {
1380 /* Here, no compile-time swash, and there are things that won't be
1381 * known until runtime -- we have to assume it could be anything */
1382 invlist = sv_2mortal(_new_invlist(1));
1383 return _add_range_to_invlist(invlist, 0, UV_MAX);
1385 else if (ary[3] && ary[3] != &PL_sv_undef) {
1387 /* Here no compile-time swash, and no run-time only data. Use the
1388 * node's inversion list */
1389 invlist = sv_2mortal(invlist_clone(ary[3]));
1392 /* Get the code points valid only under UTF-8 locales */
1393 if ((ANYOF_FLAGS(node) & ANYOFL_FOLD)
1394 && ary[2] && ary[2] != &PL_sv_undef)
1396 only_utf8_locale_invlist = ary[2];
1401 invlist = sv_2mortal(_new_invlist(0));
1404 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1405 * code points, and an inversion list for the others, but if there are code
1406 * points that should match only conditionally on the target string being
1407 * UTF-8, those are placed in the inversion list, and not the bitmap.
1408 * Since there are circumstances under which they could match, they are
1409 * included in the SSC. But if the ANYOF node is to be inverted, we have
1410 * to exclude them here, so that when we invert below, the end result
1411 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1412 * have to do this here before we add the unconditionally matched code
1414 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1415 _invlist_intersection_complement_2nd(invlist,
1420 /* Add in the points from the bit map */
1421 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1422 if (ANYOF_BITMAP_TEST(node, i)) {
1423 unsigned int start = i++;
1425 for (; i < NUM_ANYOF_CODE_POINTS && ANYOF_BITMAP_TEST(node, i); ++i) {
1428 invlist = _add_range_to_invlist(invlist, start, i-1);
1429 new_node_has_latin1 = TRUE;
1433 /* If this can match all upper Latin1 code points, have to add them
1434 * as well. But don't add them if inverting, as when that gets done below,
1435 * it would exclude all these characters, including the ones it shouldn't
1436 * that were added just above */
1437 if (! (ANYOF_FLAGS(node) & ANYOF_INVERT) && OP(node) == ANYOFD
1438 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1440 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1443 /* Similarly for these */
1444 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1445 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1448 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1449 _invlist_invert(invlist);
1451 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOFL_FOLD) {
1453 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1454 * locale. We can skip this if there are no 0-255 at all. */
1455 _invlist_union(invlist, PL_Latin1, &invlist);
1458 /* Similarly add the UTF-8 locale possible matches. These have to be
1459 * deferred until after the non-UTF-8 locale ones are taken care of just
1460 * above, or it leads to wrong results under ANYOF_INVERT */
1461 if (only_utf8_locale_invlist) {
1462 _invlist_union_maybe_complement_2nd(invlist,
1463 only_utf8_locale_invlist,
1464 ANYOF_FLAGS(node) & ANYOF_INVERT,
1471 /* These two functions currently do the exact same thing */
1472 #define ssc_init_zero ssc_init
1474 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1475 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1477 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1478 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1479 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1482 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1483 const regnode_charclass *and_with)
1485 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1486 * another SSC or a regular ANYOF class. Can create false positives. */
1491 PERL_ARGS_ASSERT_SSC_AND;
1493 assert(is_ANYOF_SYNTHETIC(ssc));
1495 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1496 * the code point inversion list and just the relevant flags */
1497 if (is_ANYOF_SYNTHETIC(and_with)) {
1498 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1499 anded_flags = ANYOF_FLAGS(and_with);
1501 /* XXX This is a kludge around what appears to be deficiencies in the
1502 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1503 * there are paths through the optimizer where it doesn't get weeded
1504 * out when it should. And if we don't make some extra provision for
1505 * it like the code just below, it doesn't get added when it should.
1506 * This solution is to add it only when AND'ing, which is here, and
1507 * only when what is being AND'ed is the pristine, original node
1508 * matching anything. Thus it is like adding it to ssc_anything() but
1509 * only when the result is to be AND'ed. Probably the same solution
1510 * could be adopted for the same problem we have with /l matching,
1511 * which is solved differently in S_ssc_init(), and that would lead to
1512 * fewer false positives than that solution has. But if this solution
1513 * creates bugs, the consequences are only that a warning isn't raised
1514 * that should be; while the consequences for having /l bugs is
1515 * incorrect matches */
1516 if (ssc_is_anything((regnode_ssc *)and_with)) {
1517 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1521 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1522 if (OP(and_with) == ANYOFD) {
1523 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1526 anded_flags = ANYOF_FLAGS(and_with)
1527 &( ANYOF_COMMON_FLAGS
1528 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1529 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1530 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(and_with))) {
1532 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1537 ANYOF_FLAGS(ssc) &= anded_flags;
1539 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1540 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1541 * 'and_with' may be inverted. When not inverted, we have the situation of
1543 * (C1 | P1) & (C2 | P2)
1544 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1545 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1546 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1547 * <= ((C1 & C2) | P1 | P2)
1548 * Alternatively, the last few steps could be:
1549 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1550 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1551 * <= (C1 | C2 | (P1 & P2))
1552 * We favor the second approach if either P1 or P2 is non-empty. This is
1553 * because these components are a barrier to doing optimizations, as what
1554 * they match cannot be known until the moment of matching as they are
1555 * dependent on the current locale, 'AND"ing them likely will reduce or
1557 * But we can do better if we know that C1,P1 are in their initial state (a
1558 * frequent occurrence), each matching everything:
1559 * (<everything>) & (C2 | P2) = C2 | P2
1560 * Similarly, if C2,P2 are in their initial state (again a frequent
1561 * occurrence), the result is a no-op
1562 * (C1 | P1) & (<everything>) = C1 | P1
1565 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1566 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1567 * <= (C1 & ~C2) | (P1 & ~P2)
1570 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1571 && ! is_ANYOF_SYNTHETIC(and_with))
1575 ssc_intersection(ssc,
1577 FALSE /* Has already been inverted */
1580 /* If either P1 or P2 is empty, the intersection will be also; can skip
1582 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
1583 ANYOF_POSIXL_ZERO(ssc);
1585 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1587 /* Note that the Posix class component P from 'and_with' actually
1589 * P = Pa | Pb | ... | Pn
1590 * where each component is one posix class, such as in [\w\s].
1592 * ~P = ~(Pa | Pb | ... | Pn)
1593 * = ~Pa & ~Pb & ... & ~Pn
1594 * <= ~Pa | ~Pb | ... | ~Pn
1595 * The last is something we can easily calculate, but unfortunately
1596 * is likely to have many false positives. We could do better
1597 * in some (but certainly not all) instances if two classes in
1598 * P have known relationships. For example
1599 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1601 * :lower: & :print: = :lower:
1602 * And similarly for classes that must be disjoint. For example,
1603 * since \s and \w can have no elements in common based on rules in
1604 * the POSIX standard,
1605 * \w & ^\S = nothing
1606 * Unfortunately, some vendor locales do not meet the Posix
1607 * standard, in particular almost everything by Microsoft.
1608 * The loop below just changes e.g., \w into \W and vice versa */
1610 regnode_charclass_posixl temp;
1611 int add = 1; /* To calculate the index of the complement */
1613 ANYOF_POSIXL_ZERO(&temp);
1614 for (i = 0; i < ANYOF_MAX; i++) {
1616 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1617 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1619 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1620 ANYOF_POSIXL_SET(&temp, i + add);
1622 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1624 ANYOF_POSIXL_AND(&temp, ssc);
1626 } /* else ssc already has no posixes */
1627 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1628 in its initial state */
1629 else if (! is_ANYOF_SYNTHETIC(and_with)
1630 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1632 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1633 * copy it over 'ssc' */
1634 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1635 if (is_ANYOF_SYNTHETIC(and_with)) {
1636 StructCopy(and_with, ssc, regnode_ssc);
1639 ssc->invlist = anded_cp_list;
1640 ANYOF_POSIXL_ZERO(ssc);
1641 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1642 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1646 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1647 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
1649 /* One or the other of P1, P2 is non-empty. */
1650 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1651 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1653 ssc_union(ssc, anded_cp_list, FALSE);
1655 else { /* P1 = P2 = empty */
1656 ssc_intersection(ssc, anded_cp_list, FALSE);
1662 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1663 const regnode_charclass *or_with)
1665 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1666 * another SSC or a regular ANYOF class. Can create false positives if
1667 * 'or_with' is to be inverted. */
1672 PERL_ARGS_ASSERT_SSC_OR;
1674 assert(is_ANYOF_SYNTHETIC(ssc));
1676 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1677 * the code point inversion list and just the relevant flags */
1678 if (is_ANYOF_SYNTHETIC(or_with)) {
1679 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1680 ored_flags = ANYOF_FLAGS(or_with);
1683 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1684 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1685 if (OP(or_with) != ANYOFD) {
1687 |= ANYOF_FLAGS(or_with)
1688 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1689 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1690 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(or_with))) {
1692 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1697 ANYOF_FLAGS(ssc) |= ored_flags;
1699 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1700 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1701 * 'or_with' may be inverted. When not inverted, we have the simple
1702 * situation of computing:
1703 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1704 * If P1|P2 yields a situation with both a class and its complement are
1705 * set, like having both \w and \W, this matches all code points, and we
1706 * can delete these from the P component of the ssc going forward. XXX We
1707 * might be able to delete all the P components, but I (khw) am not certain
1708 * about this, and it is better to be safe.
1711 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1712 * <= (C1 | P1) | ~C2
1713 * <= (C1 | ~C2) | P1
1714 * (which results in actually simpler code than the non-inverted case)
1717 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1718 && ! is_ANYOF_SYNTHETIC(or_with))
1720 /* We ignore P2, leaving P1 going forward */
1721 } /* else Not inverted */
1722 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
1723 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1724 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1726 for (i = 0; i < ANYOF_MAX; i += 2) {
1727 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1729 ssc_match_all_cp(ssc);
1730 ANYOF_POSIXL_CLEAR(ssc, i);
1731 ANYOF_POSIXL_CLEAR(ssc, i+1);
1739 FALSE /* Already has been inverted */
1743 PERL_STATIC_INLINE void
1744 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1746 PERL_ARGS_ASSERT_SSC_UNION;
1748 assert(is_ANYOF_SYNTHETIC(ssc));
1750 _invlist_union_maybe_complement_2nd(ssc->invlist,
1756 PERL_STATIC_INLINE void
1757 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1759 const bool invert2nd)
1761 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1763 assert(is_ANYOF_SYNTHETIC(ssc));
1765 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1771 PERL_STATIC_INLINE void
1772 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1774 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1776 assert(is_ANYOF_SYNTHETIC(ssc));
1778 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1781 PERL_STATIC_INLINE void
1782 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1784 /* AND just the single code point 'cp' into the SSC 'ssc' */
1786 SV* cp_list = _new_invlist(2);
1788 PERL_ARGS_ASSERT_SSC_CP_AND;
1790 assert(is_ANYOF_SYNTHETIC(ssc));
1792 cp_list = add_cp_to_invlist(cp_list, cp);
1793 ssc_intersection(ssc, cp_list,
1794 FALSE /* Not inverted */
1796 SvREFCNT_dec_NN(cp_list);
1799 PERL_STATIC_INLINE void
1800 S_ssc_clear_locale(regnode_ssc *ssc)
1802 /* Set the SSC 'ssc' to not match any locale things */
1803 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1805 assert(is_ANYOF_SYNTHETIC(ssc));
1807 ANYOF_POSIXL_ZERO(ssc);
1808 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1811 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1814 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1816 /* The synthetic start class is used to hopefully quickly winnow down
1817 * places where a pattern could start a match in the target string. If it
1818 * doesn't really narrow things down that much, there isn't much point to
1819 * having the overhead of using it. This function uses some very crude
1820 * heuristics to decide if to use the ssc or not.
1822 * It returns TRUE if 'ssc' rules out more than half what it considers to
1823 * be the "likely" possible matches, but of course it doesn't know what the
1824 * actual things being matched are going to be; these are only guesses
1826 * For /l matches, it assumes that the only likely matches are going to be
1827 * in the 0-255 range, uniformly distributed, so half of that is 127
1828 * For /a and /d matches, it assumes that the likely matches will be just
1829 * the ASCII range, so half of that is 63
1830 * For /u and there isn't anything matching above the Latin1 range, it
1831 * assumes that that is the only range likely to be matched, and uses
1832 * half that as the cut-off: 127. If anything matches above Latin1,
1833 * it assumes that all of Unicode could match (uniformly), except for
1834 * non-Unicode code points and things in the General Category "Other"
1835 * (unassigned, private use, surrogates, controls and formats). This
1836 * is a much large number. */
1838 U32 count = 0; /* Running total of number of code points matched by
1840 UV start, end; /* Start and end points of current range in inversion
1842 const U32 max_code_points = (LOC)
1844 : (( ! UNI_SEMANTICS
1845 || invlist_highest(ssc->invlist) < 256)
1848 const U32 max_match = max_code_points / 2;
1850 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1852 invlist_iterinit(ssc->invlist);
1853 while (invlist_iternext(ssc->invlist, &start, &end)) {
1854 if (start >= max_code_points) {
1857 end = MIN(end, max_code_points - 1);
1858 count += end - start + 1;
1859 if (count >= max_match) {
1860 invlist_iterfinish(ssc->invlist);
1870 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1872 /* The inversion list in the SSC is marked mortal; now we need a more
1873 * permanent copy, which is stored the same way that is done in a regular
1874 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
1877 SV* invlist = invlist_clone(ssc->invlist);
1879 PERL_ARGS_ASSERT_SSC_FINALIZE;
1881 assert(is_ANYOF_SYNTHETIC(ssc));
1883 /* The code in this file assumes that all but these flags aren't relevant
1884 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
1885 * by the time we reach here */
1886 assert(! (ANYOF_FLAGS(ssc)
1887 & ~( ANYOF_COMMON_FLAGS
1888 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1889 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
1891 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1893 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1894 NULL, NULL, NULL, FALSE);
1896 /* Make sure is clone-safe */
1897 ssc->invlist = NULL;
1899 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1900 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
1903 if (RExC_contains_locale) {
1907 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
1910 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1911 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1912 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1913 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1914 ? (TRIE_LIST_CUR( idx ) - 1) \
1920 dump_trie(trie,widecharmap,revcharmap)
1921 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1922 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1924 These routines dump out a trie in a somewhat readable format.
1925 The _interim_ variants are used for debugging the interim
1926 tables that are used to generate the final compressed
1927 representation which is what dump_trie expects.
1929 Part of the reason for their existence is to provide a form
1930 of documentation as to how the different representations function.
1935 Dumps the final compressed table form of the trie to Perl_debug_log.
1936 Used for debugging make_trie().
1940 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1941 AV *revcharmap, U32 depth)
1944 SV *sv=sv_newmortal();
1945 int colwidth= widecharmap ? 6 : 4;
1947 GET_RE_DEBUG_FLAGS_DECL;
1949 PERL_ARGS_ASSERT_DUMP_TRIE;
1951 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1952 (int)depth * 2 + 2,"",
1953 "Match","Base","Ofs" );
1955 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1956 SV ** const tmp = av_fetch( revcharmap, state, 0);
1958 PerlIO_printf( Perl_debug_log, "%*s",
1960 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1961 PL_colors[0], PL_colors[1],
1962 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1963 PERL_PV_ESCAPE_FIRSTCHAR
1968 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1969 (int)depth * 2 + 2,"");
1971 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1972 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1973 PerlIO_printf( Perl_debug_log, "\n");
1975 for( state = 1 ; state < trie->statecount ; state++ ) {
1976 const U32 base = trie->states[ state ].trans.base;
1978 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|",
1979 (int)depth * 2 + 2,"", (UV)state);
1981 if ( trie->states[ state ].wordnum ) {
1982 PerlIO_printf( Perl_debug_log, " W%4X",
1983 trie->states[ state ].wordnum );
1985 PerlIO_printf( Perl_debug_log, "%6s", "" );
1988 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1993 while( ( base + ofs < trie->uniquecharcount ) ||
1994 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1995 && trie->trans[ base + ofs - trie->uniquecharcount ].check
1999 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
2001 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2002 if ( ( base + ofs >= trie->uniquecharcount )
2003 && ( base + ofs - trie->uniquecharcount
2005 && trie->trans[ base + ofs
2006 - trie->uniquecharcount ].check == state )
2008 PerlIO_printf( Perl_debug_log, "%*"UVXf,
2010 (UV)trie->trans[ base + ofs
2011 - trie->uniquecharcount ].next );
2013 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
2017 PerlIO_printf( Perl_debug_log, "]");
2020 PerlIO_printf( Perl_debug_log, "\n" );
2022 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=",
2024 for (word=1; word <= trie->wordcount; word++) {
2025 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
2026 (int)word, (int)(trie->wordinfo[word].prev),
2027 (int)(trie->wordinfo[word].len));
2029 PerlIO_printf(Perl_debug_log, "\n" );
2032 Dumps a fully constructed but uncompressed trie in list form.
2033 List tries normally only are used for construction when the number of
2034 possible chars (trie->uniquecharcount) is very high.
2035 Used for debugging make_trie().
2038 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2039 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2043 SV *sv=sv_newmortal();
2044 int colwidth= widecharmap ? 6 : 4;
2045 GET_RE_DEBUG_FLAGS_DECL;
2047 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2049 /* print out the table precompression. */
2050 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
2051 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
2052 "------:-----+-----------------\n" );
2054 for( state=1 ; state < next_alloc ; state ++ ) {
2057 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
2058 (int)depth * 2 + 2,"", (UV)state );
2059 if ( ! trie->states[ state ].wordnum ) {
2060 PerlIO_printf( Perl_debug_log, "%5s| ","");
2062 PerlIO_printf( Perl_debug_log, "W%4x| ",
2063 trie->states[ state ].wordnum
2066 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2067 SV ** const tmp = av_fetch( revcharmap,
2068 TRIE_LIST_ITEM(state,charid).forid, 0);
2070 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
2072 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2074 PL_colors[0], PL_colors[1],
2075 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2076 | PERL_PV_ESCAPE_FIRSTCHAR
2078 TRIE_LIST_ITEM(state,charid).forid,
2079 (UV)TRIE_LIST_ITEM(state,charid).newstate
2082 PerlIO_printf(Perl_debug_log, "\n%*s| ",
2083 (int)((depth * 2) + 14), "");
2086 PerlIO_printf( Perl_debug_log, "\n");
2091 Dumps a fully constructed but uncompressed trie in table form.
2092 This is the normal DFA style state transition table, with a few
2093 twists to facilitate compression later.
2094 Used for debugging make_trie().
2097 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2098 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2103 SV *sv=sv_newmortal();
2104 int colwidth= widecharmap ? 6 : 4;
2105 GET_RE_DEBUG_FLAGS_DECL;
2107 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2110 print out the table precompression so that we can do a visual check
2111 that they are identical.
2114 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
2116 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2117 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2119 PerlIO_printf( Perl_debug_log, "%*s",
2121 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2122 PL_colors[0], PL_colors[1],
2123 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2124 PERL_PV_ESCAPE_FIRSTCHAR
2130 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
2132 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2133 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
2136 PerlIO_printf( Perl_debug_log, "\n" );
2138 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2140 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
2141 (int)depth * 2 + 2,"",
2142 (UV)TRIE_NODENUM( state ) );
2144 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2145 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2147 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
2149 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
2151 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2152 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n",
2153 (UV)trie->trans[ state ].check );
2155 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n",
2156 (UV)trie->trans[ state ].check,
2157 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2165 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2166 startbranch: the first branch in the whole branch sequence
2167 first : start branch of sequence of branch-exact nodes.
2168 May be the same as startbranch
2169 last : Thing following the last branch.
2170 May be the same as tail.
2171 tail : item following the branch sequence
2172 count : words in the sequence
2173 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2174 depth : indent depth
2176 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2178 A trie is an N'ary tree where the branches are determined by digital
2179 decomposition of the key. IE, at the root node you look up the 1st character and
2180 follow that branch repeat until you find the end of the branches. Nodes can be
2181 marked as "accepting" meaning they represent a complete word. Eg:
2185 would convert into the following structure. Numbers represent states, letters
2186 following numbers represent valid transitions on the letter from that state, if
2187 the number is in square brackets it represents an accepting state, otherwise it
2188 will be in parenthesis.
2190 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2194 (1) +-i->(6)-+-s->[7]
2196 +-s->(3)-+-h->(4)-+-e->[5]
2198 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2200 This shows that when matching against the string 'hers' we will begin at state 1
2201 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2202 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2203 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2204 single traverse. We store a mapping from accepting to state to which word was
2205 matched, and then when we have multiple possibilities we try to complete the
2206 rest of the regex in the order in which they occurred in the alternation.
2208 The only prior NFA like behaviour that would be changed by the TRIE support is
2209 the silent ignoring of duplicate alternations which are of the form:
2211 / (DUPE|DUPE) X? (?{ ... }) Y /x
2213 Thus EVAL blocks following a trie may be called a different number of times with
2214 and without the optimisation. With the optimisations dupes will be silently
2215 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2216 the following demonstrates:
2218 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2220 which prints out 'word' three times, but
2222 'words'=~/(word|word|word)(?{ print $1 })S/
2224 which doesnt print it out at all. This is due to other optimisations kicking in.
2226 Example of what happens on a structural level:
2228 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2230 1: CURLYM[1] {1,32767}(18)
2241 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2242 and should turn into:
2244 1: CURLYM[1] {1,32767}(18)
2246 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2254 Cases where tail != last would be like /(?foo|bar)baz/:
2264 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2265 and would end up looking like:
2268 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2275 d = uvchr_to_utf8_flags(d, uv, 0);
2277 is the recommended Unicode-aware way of saying
2282 #define TRIE_STORE_REVCHAR(val) \
2285 SV *zlopp = newSV(UTF8_MAXBYTES); \
2286 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2287 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2288 SvCUR_set(zlopp, kapow - flrbbbbb); \
2291 av_push(revcharmap, zlopp); \
2293 char ooooff = (char)val; \
2294 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2298 /* This gets the next character from the input, folding it if not already
2300 #define TRIE_READ_CHAR STMT_START { \
2303 /* if it is UTF then it is either already folded, or does not need \
2305 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2307 else if (folder == PL_fold_latin1) { \
2308 /* This folder implies Unicode rules, which in the range expressible \
2309 * by not UTF is the lower case, with the two exceptions, one of \
2310 * which should have been taken care of before calling this */ \
2311 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2312 uvc = toLOWER_L1(*uc); \
2313 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2316 /* raw data, will be folded later if needed */ \
2324 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2325 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2326 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
2327 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2329 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2330 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2331 TRIE_LIST_CUR( state )++; \
2334 #define TRIE_LIST_NEW(state) STMT_START { \
2335 Newxz( trie->states[ state ].trans.list, \
2336 4, reg_trie_trans_le ); \
2337 TRIE_LIST_CUR( state ) = 1; \
2338 TRIE_LIST_LEN( state ) = 4; \
2341 #define TRIE_HANDLE_WORD(state) STMT_START { \
2342 U16 dupe= trie->states[ state ].wordnum; \
2343 regnode * const noper_next = regnext( noper ); \
2346 /* store the word for dumping */ \
2348 if (OP(noper) != NOTHING) \
2349 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2351 tmp = newSVpvn_utf8( "", 0, UTF ); \
2352 av_push( trie_words, tmp ); \
2356 trie->wordinfo[curword].prev = 0; \
2357 trie->wordinfo[curword].len = wordlen; \
2358 trie->wordinfo[curword].accept = state; \
2360 if ( noper_next < tail ) { \
2362 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2364 trie->jump[curword] = (U16)(noper_next - convert); \
2366 jumper = noper_next; \
2368 nextbranch= regnext(cur); \
2372 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2373 /* chain, so that when the bits of chain are later */\
2374 /* linked together, the dups appear in the chain */\
2375 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2376 trie->wordinfo[dupe].prev = curword; \
2378 /* we haven't inserted this word yet. */ \
2379 trie->states[ state ].wordnum = curword; \
2384 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2385 ( ( base + charid >= ucharcount \
2386 && base + charid < ubound \
2387 && state == trie->trans[ base - ucharcount + charid ].check \
2388 && trie->trans[ base - ucharcount + charid ].next ) \
2389 ? trie->trans[ base - ucharcount + charid ].next \
2390 : ( state==1 ? special : 0 ) \
2394 #define MADE_JUMP_TRIE 2
2395 #define MADE_EXACT_TRIE 4
2398 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2399 regnode *first, regnode *last, regnode *tail,
2400 U32 word_count, U32 flags, U32 depth)
2402 /* first pass, loop through and scan words */
2403 reg_trie_data *trie;
2404 HV *widecharmap = NULL;
2405 AV *revcharmap = newAV();
2411 regnode *jumper = NULL;
2412 regnode *nextbranch = NULL;
2413 regnode *convert = NULL;
2414 U32 *prev_states; /* temp array mapping each state to previous one */
2415 /* we just use folder as a flag in utf8 */
2416 const U8 * folder = NULL;
2419 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
2420 AV *trie_words = NULL;
2421 /* along with revcharmap, this only used during construction but both are
2422 * useful during debugging so we store them in the struct when debugging.
2425 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2426 STRLEN trie_charcount=0;
2428 SV *re_trie_maxbuff;
2429 GET_RE_DEBUG_FLAGS_DECL;
2431 PERL_ARGS_ASSERT_MAKE_TRIE;
2433 PERL_UNUSED_ARG(depth);
2437 case EXACT: case EXACTL: break;
2441 case EXACTFLU8: folder = PL_fold_latin1; break;
2442 case EXACTF: folder = PL_fold; break;
2443 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2446 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2448 trie->startstate = 1;
2449 trie->wordcount = word_count;
2450 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2451 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2452 if (flags == EXACT || flags == EXACTL)
2453 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2454 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2455 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2458 trie_words = newAV();
2461 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2462 assert(re_trie_maxbuff);
2463 if (!SvIOK(re_trie_maxbuff)) {
2464 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2466 DEBUG_TRIE_COMPILE_r({
2467 PerlIO_printf( Perl_debug_log,
2468 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2469 (int)depth * 2 + 2, "",
2470 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2471 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2474 /* Find the node we are going to overwrite */
2475 if ( first == startbranch && OP( last ) != BRANCH ) {
2476 /* whole branch chain */
2479 /* branch sub-chain */
2480 convert = NEXTOPER( first );
2483 /* -- First loop and Setup --
2485 We first traverse the branches and scan each word to determine if it
2486 contains widechars, and how many unique chars there are, this is
2487 important as we have to build a table with at least as many columns as we
2490 We use an array of integers to represent the character codes 0..255
2491 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2492 the native representation of the character value as the key and IV's for
2495 *TODO* If we keep track of how many times each character is used we can
2496 remap the columns so that the table compression later on is more
2497 efficient in terms of memory by ensuring the most common value is in the
2498 middle and the least common are on the outside. IMO this would be better
2499 than a most to least common mapping as theres a decent chance the most
2500 common letter will share a node with the least common, meaning the node
2501 will not be compressible. With a middle is most common approach the worst
2502 case is when we have the least common nodes twice.
2506 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2507 regnode *noper = NEXTOPER( cur );
2508 const U8 *uc = (U8*)STRING( noper );
2509 const U8 *e = uc + STR_LEN( noper );
2511 U32 wordlen = 0; /* required init */
2512 STRLEN minchars = 0;
2513 STRLEN maxchars = 0;
2514 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2517 if (OP(noper) == NOTHING) {
2518 regnode *noper_next= regnext(noper);
2519 if (noper_next != tail && OP(noper_next) == flags) {
2521 uc= (U8*)STRING(noper);
2522 e= uc + STR_LEN(noper);
2523 trie->minlen= STR_LEN(noper);
2530 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2531 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2532 regardless of encoding */
2533 if (OP( noper ) == EXACTFU_SS) {
2534 /* false positives are ok, so just set this */
2535 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2538 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2540 TRIE_CHARCOUNT(trie)++;
2543 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2544 * is in effect. Under /i, this character can match itself, or
2545 * anything that folds to it. If not under /i, it can match just
2546 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2547 * all fold to k, and all are single characters. But some folds
2548 * expand to more than one character, so for example LATIN SMALL
2549 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2550 * the string beginning at 'uc' is 'ffi', it could be matched by
2551 * three characters, or just by the one ligature character. (It
2552 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2553 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2554 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2555 * match.) The trie needs to know the minimum and maximum number
2556 * of characters that could match so that it can use size alone to
2557 * quickly reject many match attempts. The max is simple: it is
2558 * the number of folded characters in this branch (since a fold is
2559 * never shorter than what folds to it. */
2563 /* And the min is equal to the max if not under /i (indicated by
2564 * 'folder' being NULL), or there are no multi-character folds. If
2565 * there is a multi-character fold, the min is incremented just
2566 * once, for the character that folds to the sequence. Each
2567 * character in the sequence needs to be added to the list below of
2568 * characters in the trie, but we count only the first towards the
2569 * min number of characters needed. This is done through the
2570 * variable 'foldlen', which is returned by the macros that look
2571 * for these sequences as the number of bytes the sequence
2572 * occupies. Each time through the loop, we decrement 'foldlen' by
2573 * how many bytes the current char occupies. Only when it reaches
2574 * 0 do we increment 'minchars' or look for another multi-character
2576 if (folder == NULL) {
2579 else if (foldlen > 0) {
2580 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2585 /* See if *uc is the beginning of a multi-character fold. If
2586 * so, we decrement the length remaining to look at, to account
2587 * for the current character this iteration. (We can use 'uc'
2588 * instead of the fold returned by TRIE_READ_CHAR because for
2589 * non-UTF, the latin1_safe macro is smart enough to account
2590 * for all the unfolded characters, and because for UTF, the
2591 * string will already have been folded earlier in the
2592 * compilation process */
2594 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2595 foldlen -= UTF8SKIP(uc);
2598 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2603 /* The current character (and any potential folds) should be added
2604 * to the possible matching characters for this position in this
2608 U8 folded= folder[ (U8) uvc ];
2609 if ( !trie->charmap[ folded ] ) {
2610 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2611 TRIE_STORE_REVCHAR( folded );
2614 if ( !trie->charmap[ uvc ] ) {
2615 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2616 TRIE_STORE_REVCHAR( uvc );
2619 /* store the codepoint in the bitmap, and its folded
2621 TRIE_BITMAP_SET(trie, uvc);
2623 /* store the folded codepoint */
2624 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
2627 /* store first byte of utf8 representation of
2628 variant codepoints */
2629 if (! UVCHR_IS_INVARIANT(uvc)) {
2630 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
2633 set_bit = 0; /* We've done our bit :-) */
2637 /* XXX We could come up with the list of code points that fold
2638 * to this using PL_utf8_foldclosures, except not for
2639 * multi-char folds, as there may be multiple combinations
2640 * there that could work, which needs to wait until runtime to
2641 * resolve (The comment about LIGATURE FFI above is such an
2646 widecharmap = newHV();
2648 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2651 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
2653 if ( !SvTRUE( *svpp ) ) {
2654 sv_setiv( *svpp, ++trie->uniquecharcount );
2655 TRIE_STORE_REVCHAR(uvc);
2658 } /* end loop through characters in this branch of the trie */
2660 /* We take the min and max for this branch and combine to find the min
2661 * and max for all branches processed so far */
2662 if( cur == first ) {
2663 trie->minlen = minchars;
2664 trie->maxlen = maxchars;
2665 } else if (minchars < trie->minlen) {
2666 trie->minlen = minchars;
2667 } else if (maxchars > trie->maxlen) {
2668 trie->maxlen = maxchars;
2670 } /* end first pass */
2671 DEBUG_TRIE_COMPILE_r(
2672 PerlIO_printf( Perl_debug_log,
2673 "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2674 (int)depth * 2 + 2,"",
2675 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2676 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2677 (int)trie->minlen, (int)trie->maxlen )
2681 We now know what we are dealing with in terms of unique chars and
2682 string sizes so we can calculate how much memory a naive
2683 representation using a flat table will take. If it's over a reasonable
2684 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2685 conservative but potentially much slower representation using an array
2688 At the end we convert both representations into the same compressed
2689 form that will be used in regexec.c for matching with. The latter
2690 is a form that cannot be used to construct with but has memory
2691 properties similar to the list form and access properties similar
2692 to the table form making it both suitable for fast searches and
2693 small enough that its feasable to store for the duration of a program.
2695 See the comment in the code where the compressed table is produced
2696 inplace from the flat tabe representation for an explanation of how
2697 the compression works.
2702 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2705 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2706 > SvIV(re_trie_maxbuff) )
2709 Second Pass -- Array Of Lists Representation
2711 Each state will be represented by a list of charid:state records
2712 (reg_trie_trans_le) the first such element holds the CUR and LEN
2713 points of the allocated array. (See defines above).
2715 We build the initial structure using the lists, and then convert
2716 it into the compressed table form which allows faster lookups
2717 (but cant be modified once converted).
2720 STRLEN transcount = 1;
2722 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2723 "%*sCompiling trie using list compiler\n",
2724 (int)depth * 2 + 2, ""));
2726 trie->states = (reg_trie_state *)
2727 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2728 sizeof(reg_trie_state) );
2732 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2734 regnode *noper = NEXTOPER( cur );
2735 U8 *uc = (U8*)STRING( noper );
2736 const U8 *e = uc + STR_LEN( noper );
2737 U32 state = 1; /* required init */
2738 U16 charid = 0; /* sanity init */
2739 U32 wordlen = 0; /* required init */
2741 if (OP(noper) == NOTHING) {
2742 regnode *noper_next= regnext(noper);
2743 if (noper_next != tail && OP(noper_next) == flags) {
2745 uc= (U8*)STRING(noper);
2746 e= uc + STR_LEN(noper);
2750 if (OP(noper) != NOTHING) {
2751 for ( ; uc < e ; uc += len ) {
2756 charid = trie->charmap[ uvc ];
2758 SV** const svpp = hv_fetch( widecharmap,
2765 charid=(U16)SvIV( *svpp );
2768 /* charid is now 0 if we dont know the char read, or
2769 * nonzero if we do */
2776 if ( !trie->states[ state ].trans.list ) {
2777 TRIE_LIST_NEW( state );
2780 check <= TRIE_LIST_USED( state );
2783 if ( TRIE_LIST_ITEM( state, check ).forid
2786 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2791 newstate = next_alloc++;
2792 prev_states[newstate] = state;
2793 TRIE_LIST_PUSH( state, charid, newstate );
2798 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2802 TRIE_HANDLE_WORD(state);
2804 } /* end second pass */
2806 /* next alloc is the NEXT state to be allocated */
2807 trie->statecount = next_alloc;
2808 trie->states = (reg_trie_state *)
2809 PerlMemShared_realloc( trie->states,
2811 * sizeof(reg_trie_state) );
2813 /* and now dump it out before we compress it */
2814 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2815 revcharmap, next_alloc,
2819 trie->trans = (reg_trie_trans *)
2820 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2827 for( state=1 ; state < next_alloc ; state ++ ) {
2831 DEBUG_TRIE_COMPILE_MORE_r(
2832 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
2836 if (trie->states[state].trans.list) {
2837 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2841 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2842 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2843 if ( forid < minid ) {
2845 } else if ( forid > maxid ) {
2849 if ( transcount < tp + maxid - minid + 1) {
2851 trie->trans = (reg_trie_trans *)
2852 PerlMemShared_realloc( trie->trans,
2854 * sizeof(reg_trie_trans) );
2855 Zero( trie->trans + (transcount / 2),
2859 base = trie->uniquecharcount + tp - minid;
2860 if ( maxid == minid ) {
2862 for ( ; zp < tp ; zp++ ) {
2863 if ( ! trie->trans[ zp ].next ) {
2864 base = trie->uniquecharcount + zp - minid;
2865 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2867 trie->trans[ zp ].check = state;
2873 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2875 trie->trans[ tp ].check = state;
2880 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2881 const U32 tid = base
2882 - trie->uniquecharcount
2883 + TRIE_LIST_ITEM( state, idx ).forid;
2884 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2886 trie->trans[ tid ].check = state;
2888 tp += ( maxid - minid + 1 );
2890 Safefree(trie->states[ state ].trans.list);
2893 DEBUG_TRIE_COMPILE_MORE_r(
2894 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
2897 trie->states[ state ].trans.base=base;
2899 trie->lasttrans = tp + 1;
2903 Second Pass -- Flat Table Representation.
2905 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2906 each. We know that we will need Charcount+1 trans at most to store
2907 the data (one row per char at worst case) So we preallocate both
2908 structures assuming worst case.
2910 We then construct the trie using only the .next slots of the entry
2913 We use the .check field of the first entry of the node temporarily
2914 to make compression both faster and easier by keeping track of how
2915 many non zero fields are in the node.
2917 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2920 There are two terms at use here: state as a TRIE_NODEIDX() which is
2921 a number representing the first entry of the node, and state as a
2922 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2923 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2924 if there are 2 entrys per node. eg:
2932 The table is internally in the right hand, idx form. However as we
2933 also have to deal with the states array which is indexed by nodenum
2934 we have to use TRIE_NODENUM() to convert.
2937 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2938 "%*sCompiling trie using table compiler\n",
2939 (int)depth * 2 + 2, ""));
2941 trie->trans = (reg_trie_trans *)
2942 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2943 * trie->uniquecharcount + 1,
2944 sizeof(reg_trie_trans) );
2945 trie->states = (reg_trie_state *)
2946 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2947 sizeof(reg_trie_state) );
2948 next_alloc = trie->uniquecharcount + 1;
2951 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2953 regnode *noper = NEXTOPER( cur );
2954 const U8 *uc = (U8*)STRING( noper );
2955 const U8 *e = uc + STR_LEN( noper );
2957 U32 state = 1; /* required init */
2959 U16 charid = 0; /* sanity init */
2960 U32 accept_state = 0; /* sanity init */
2962 U32 wordlen = 0; /* required init */
2964 if (OP(noper) == NOTHING) {
2965 regnode *noper_next= regnext(noper);
2966 if (noper_next != tail && OP(noper_next) == flags) {
2968 uc= (U8*)STRING(noper);
2969 e= uc + STR_LEN(noper);
2973 if ( OP(noper) != NOTHING ) {
2974 for ( ; uc < e ; uc += len ) {
2979 charid = trie->charmap[ uvc ];
2981 SV* const * const svpp = hv_fetch( widecharmap,
2985 charid = svpp ? (U16)SvIV(*svpp) : 0;
2989 if ( !trie->trans[ state + charid ].next ) {
2990 trie->trans[ state + charid ].next = next_alloc;
2991 trie->trans[ state ].check++;
2992 prev_states[TRIE_NODENUM(next_alloc)]
2993 = TRIE_NODENUM(state);
2994 next_alloc += trie->uniquecharcount;
2996 state = trie->trans[ state + charid ].next;
2998 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
3000 /* charid is now 0 if we dont know the char read, or
3001 * nonzero if we do */
3004 accept_state = TRIE_NODENUM( state );
3005 TRIE_HANDLE_WORD(accept_state);
3007 } /* end second pass */
3009 /* and now dump it out before we compress it */
3010 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3012 next_alloc, depth+1));
3016 * Inplace compress the table.*
3018 For sparse data sets the table constructed by the trie algorithm will
3019 be mostly 0/FAIL transitions or to put it another way mostly empty.
3020 (Note that leaf nodes will not contain any transitions.)
3022 This algorithm compresses the tables by eliminating most such
3023 transitions, at the cost of a modest bit of extra work during lookup:
3025 - Each states[] entry contains a .base field which indicates the
3026 index in the state[] array wheres its transition data is stored.
3028 - If .base is 0 there are no valid transitions from that node.
3030 - If .base is nonzero then charid is added to it to find an entry in
3033 -If trans[states[state].base+charid].check!=state then the
3034 transition is taken to be a 0/Fail transition. Thus if there are fail
3035 transitions at the front of the node then the .base offset will point
3036 somewhere inside the previous nodes data (or maybe even into a node
3037 even earlier), but the .check field determines if the transition is
3041 The following process inplace converts the table to the compressed
3042 table: We first do not compress the root node 1,and mark all its
3043 .check pointers as 1 and set its .base pointer as 1 as well. This
3044 allows us to do a DFA construction from the compressed table later,
3045 and ensures that any .base pointers we calculate later are greater
3048 - We set 'pos' to indicate the first entry of the second node.
3050 - We then iterate over the columns of the node, finding the first and
3051 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3052 and set the .check pointers accordingly, and advance pos
3053 appropriately and repreat for the next node. Note that when we copy
3054 the next pointers we have to convert them from the original
3055 NODEIDX form to NODENUM form as the former is not valid post
3058 - If a node has no transitions used we mark its base as 0 and do not
3059 advance the pos pointer.
3061 - If a node only has one transition we use a second pointer into the
3062 structure to fill in allocated fail transitions from other states.
3063 This pointer is independent of the main pointer and scans forward
3064 looking for null transitions that are allocated to a state. When it
3065 finds one it writes the single transition into the "hole". If the
3066 pointer doesnt find one the single transition is appended as normal.
3068 - Once compressed we can Renew/realloc the structures to release the
3071 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3072 specifically Fig 3.47 and the associated pseudocode.
3076 const U32 laststate = TRIE_NODENUM( next_alloc );
3079 trie->statecount = laststate;
3081 for ( state = 1 ; state < laststate ; state++ ) {
3083 const U32 stateidx = TRIE_NODEIDX( state );
3084 const U32 o_used = trie->trans[ stateidx ].check;
3085 U32 used = trie->trans[ stateidx ].check;
3086 trie->trans[ stateidx ].check = 0;
3089 used && charid < trie->uniquecharcount;
3092 if ( flag || trie->trans[ stateidx + charid ].next ) {
3093 if ( trie->trans[ stateidx + charid ].next ) {
3095 for ( ; zp < pos ; zp++ ) {
3096 if ( ! trie->trans[ zp ].next ) {
3100 trie->states[ state ].trans.base
3102 + trie->uniquecharcount
3104 trie->trans[ zp ].next
3105 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3107 trie->trans[ zp ].check = state;
3108 if ( ++zp > pos ) pos = zp;
3115 trie->states[ state ].trans.base
3116 = pos + trie->uniquecharcount - charid ;
3118 trie->trans[ pos ].next
3119 = SAFE_TRIE_NODENUM(
3120 trie->trans[ stateidx + charid ].next );
3121 trie->trans[ pos ].check = state;
3126 trie->lasttrans = pos + 1;
3127 trie->states = (reg_trie_state *)
3128 PerlMemShared_realloc( trie->states, laststate
3129 * sizeof(reg_trie_state) );
3130 DEBUG_TRIE_COMPILE_MORE_r(
3131 PerlIO_printf( Perl_debug_log,
3132 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
3133 (int)depth * 2 + 2,"",
3134 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3138 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3141 } /* end table compress */
3143 DEBUG_TRIE_COMPILE_MORE_r(
3144 PerlIO_printf(Perl_debug_log,
3145 "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
3146 (int)depth * 2 + 2, "",
3147 (UV)trie->statecount,
3148 (UV)trie->lasttrans)
3150 /* resize the trans array to remove unused space */
3151 trie->trans = (reg_trie_trans *)
3152 PerlMemShared_realloc( trie->trans, trie->lasttrans
3153 * sizeof(reg_trie_trans) );
3155 { /* Modify the program and insert the new TRIE node */
3156 U8 nodetype =(U8)(flags & 0xFF);
3160 regnode *optimize = NULL;
3161 #ifdef RE_TRACK_PATTERN_OFFSETS
3164 U32 mjd_nodelen = 0;
3165 #endif /* RE_TRACK_PATTERN_OFFSETS */
3166 #endif /* DEBUGGING */
3168 This means we convert either the first branch or the first Exact,
3169 depending on whether the thing following (in 'last') is a branch
3170 or not and whther first is the startbranch (ie is it a sub part of
3171 the alternation or is it the whole thing.)
3172 Assuming its a sub part we convert the EXACT otherwise we convert
3173 the whole branch sequence, including the first.
3175 /* Find the node we are going to overwrite */
3176 if ( first != startbranch || OP( last ) == BRANCH ) {
3177 /* branch sub-chain */
3178 NEXT_OFF( first ) = (U16)(last - first);
3179 #ifdef RE_TRACK_PATTERN_OFFSETS
3181 mjd_offset= Node_Offset((convert));
3182 mjd_nodelen= Node_Length((convert));
3185 /* whole branch chain */
3187 #ifdef RE_TRACK_PATTERN_OFFSETS
3190 const regnode *nop = NEXTOPER( convert );
3191 mjd_offset= Node_Offset((nop));
3192 mjd_nodelen= Node_Length((nop));
3196 PerlIO_printf(Perl_debug_log,
3197 "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
3198 (int)depth * 2 + 2, "",
3199 (UV)mjd_offset, (UV)mjd_nodelen)
3202 /* But first we check to see if there is a common prefix we can
3203 split out as an EXACT and put in front of the TRIE node. */
3204 trie->startstate= 1;
3205 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3207 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3211 const U32 base = trie->states[ state ].trans.base;
3213 if ( trie->states[state].wordnum )
3216 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3217 if ( ( base + ofs >= trie->uniquecharcount ) &&
3218 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3219 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3221 if ( ++count > 1 ) {
3222 SV **tmp = av_fetch( revcharmap, ofs, 0);
3223 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
3224 if ( state == 1 ) break;
3226 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3228 PerlIO_printf(Perl_debug_log,
3229 "%*sNew Start State=%"UVuf" Class: [",
3230 (int)depth * 2 + 2, "",
3233 SV ** const tmp = av_fetch( revcharmap, idx, 0);
3234 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3236 TRIE_BITMAP_SET(trie,*ch);
3238 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3240 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3244 TRIE_BITMAP_SET(trie,*ch);
3246 TRIE_BITMAP_SET(trie,folder[ *ch ]);
3247 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
3253 SV **tmp = av_fetch( revcharmap, idx, 0);
3255 char *ch = SvPV( *tmp, len );
3257 SV *sv=sv_newmortal();
3258 PerlIO_printf( Perl_debug_log,
3259 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
3260 (int)depth * 2 + 2, "",
3262 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3263 PL_colors[0], PL_colors[1],
3264 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3265 PERL_PV_ESCAPE_FIRSTCHAR
3270 OP( convert ) = nodetype;
3271 str=STRING(convert);
3274 STR_LEN(convert) += len;
3280 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
3285 trie->prefixlen = (state-1);
3287 regnode *n = convert+NODE_SZ_STR(convert);
3288 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3289 trie->startstate = state;
3290 trie->minlen -= (state - 1);
3291 trie->maxlen -= (state - 1);
3293 /* At least the UNICOS C compiler choked on this
3294 * being argument to DEBUG_r(), so let's just have
3297 #ifdef PERL_EXT_RE_BUILD
3303 regnode *fix = convert;
3304 U32 word = trie->wordcount;
3306 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3307 while( ++fix < n ) {
3308 Set_Node_Offset_Length(fix, 0, 0);
3311 SV ** const tmp = av_fetch( trie_words, word, 0 );
3313 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3314 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3316 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3324 NEXT_OFF(convert) = (U16)(tail - convert);
3325 DEBUG_r(optimize= n);
3331 if ( trie->maxlen ) {
3332 NEXT_OFF( convert ) = (U16)(tail - convert);
3333 ARG_SET( convert, data_slot );
3334 /* Store the offset to the first unabsorbed branch in
3335 jump[0], which is otherwise unused by the jump logic.
3336 We use this when dumping a trie and during optimisation. */
3338 trie->jump[0] = (U16)(nextbranch - convert);
3340 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3341 * and there is a bitmap
3342 * and the first "jump target" node we found leaves enough room
3343 * then convert the TRIE node into a TRIEC node, with the bitmap
3344 * embedded inline in the opcode - this is hypothetically faster.
3346 if ( !trie->states[trie->startstate].wordnum
3348 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3350 OP( convert ) = TRIEC;
3351 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3352 PerlMemShared_free(trie->bitmap);
3355 OP( convert ) = TRIE;
3357 /* store the type in the flags */
3358 convert->flags = nodetype;
3362 + regarglen[ OP( convert ) ];
3364 /* XXX We really should free up the resource in trie now,
3365 as we won't use them - (which resources?) dmq */
3367 /* needed for dumping*/
3368 DEBUG_r(if (optimize) {
3369 regnode *opt = convert;
3371 while ( ++opt < optimize) {
3372 Set_Node_Offset_Length(opt,0,0);
3375 Try to clean up some of the debris left after the
3378 while( optimize < jumper ) {
3379 mjd_nodelen += Node_Length((optimize));
3380 OP( optimize ) = OPTIMIZED;
3381 Set_Node_Offset_Length(optimize,0,0);
3384 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
3386 } /* end node insert */
3388 /* Finish populating the prev field of the wordinfo array. Walk back
3389 * from each accept state until we find another accept state, and if
3390 * so, point the first word's .prev field at the second word. If the
3391 * second already has a .prev field set, stop now. This will be the
3392 * case either if we've already processed that word's accept state,
3393 * or that state had multiple words, and the overspill words were
3394 * already linked up earlier.
3401 for (word=1; word <= trie->wordcount; word++) {
3403 if (trie->wordinfo[word].prev)
3405 state = trie->wordinfo[word].accept;
3407 state = prev_states[state];
3410 prev = trie->states[state].wordnum;
3414 trie->wordinfo[word].prev = prev;
3416 Safefree(prev_states);
3420 /* and now dump out the compressed format */
3421 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3423 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3425 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3426 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3428 SvREFCNT_dec_NN(revcharmap);
3432 : trie->startstate>1
3438 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3440 /* The Trie is constructed and compressed now so we can build a fail array if
3443 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3445 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3449 We find the fail state for each state in the trie, this state is the longest
3450 proper suffix of the current state's 'word' that is also a proper prefix of
3451 another word in our trie. State 1 represents the word '' and is thus the
3452 default fail state. This allows the DFA not to have to restart after its
3453 tried and failed a word at a given point, it simply continues as though it
3454 had been matching the other word in the first place.
3456 'abcdgu'=~/abcdefg|cdgu/
3457 When we get to 'd' we are still matching the first word, we would encounter
3458 'g' which would fail, which would bring us to the state representing 'd' in
3459 the second word where we would try 'g' and succeed, proceeding to match
3462 /* add a fail transition */
3463 const U32 trie_offset = ARG(source);
3464 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3466 const U32 ucharcount = trie->uniquecharcount;
3467 const U32 numstates = trie->statecount;
3468 const U32 ubound = trie->lasttrans + ucharcount;
3472 U32 base = trie->states[ 1 ].trans.base;
3475 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3477 GET_RE_DEBUG_FLAGS_DECL;
3479 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3480 PERL_UNUSED_CONTEXT;
3482 PERL_UNUSED_ARG(depth);
3485 if ( OP(source) == TRIE ) {
3486 struct regnode_1 *op = (struct regnode_1 *)
3487 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3488 StructCopy(source,op,struct regnode_1);
3489 stclass = (regnode *)op;
3491 struct regnode_charclass *op = (struct regnode_charclass *)
3492 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3493 StructCopy(source,op,struct regnode_charclass);
3494 stclass = (regnode *)op;
3496 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3498 ARG_SET( stclass, data_slot );
3499 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3500 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3501 aho->trie=trie_offset;
3502 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3503 Copy( trie->states, aho->states, numstates, reg_trie_state );
3504 Newxz( q, numstates, U32);
3505 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3508 /* initialize fail[0..1] to be 1 so that we always have
3509 a valid final fail state */
3510 fail[ 0 ] = fail[ 1 ] = 1;
3512 for ( charid = 0; charid < ucharcount ; charid++ ) {
3513 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3515 q[ q_write ] = newstate;
3516 /* set to point at the root */
3517 fail[ q[ q_write++ ] ]=1;
3520 while ( q_read < q_write) {
3521 const U32 cur = q[ q_read++ % numstates ];
3522 base = trie->states[ cur ].trans.base;
3524 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3525 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3527 U32 fail_state = cur;
3530 fail_state = fail[ fail_state ];
3531 fail_base = aho->states[ fail_state ].trans.base;
3532 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3534 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3535 fail[ ch_state ] = fail_state;
3536 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3538 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3540 q[ q_write++ % numstates] = ch_state;
3544 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3545 when we fail in state 1, this allows us to use the
3546 charclass scan to find a valid start char. This is based on the principle
3547 that theres a good chance the string being searched contains lots of stuff
3548 that cant be a start char.
3550 fail[ 0 ] = fail[ 1 ] = 0;
3551 DEBUG_TRIE_COMPILE_r({
3552 PerlIO_printf(Perl_debug_log,
3553 "%*sStclass Failtable (%"UVuf" states): 0",
3554 (int)(depth * 2), "", (UV)numstates
3556 for( q_read=1; q_read<numstates; q_read++ ) {
3557 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
3559 PerlIO_printf(Perl_debug_log, "\n");
3562 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3567 #define DEBUG_PEEP(str,scan,depth) \
3568 DEBUG_OPTIMISE_r({if (scan){ \
3569 regnode *Next = regnext(scan); \
3570 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state); \
3571 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)", \
3572 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),\
3573 Next ? (REG_NODE_NUM(Next)) : 0 ); \
3574 DEBUG_SHOW_STUDY_FLAGS(flags," [ ","]");\
3575 PerlIO_printf(Perl_debug_log, "\n"); \
3578 /* The below joins as many adjacent EXACTish nodes as possible into a single
3579 * one. The regop may be changed if the node(s) contain certain sequences that
3580 * require special handling. The joining is only done if:
3581 * 1) there is room in the current conglomerated node to entirely contain the
3583 * 2) they are the exact same node type
3585 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3586 * these get optimized out
3588 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3589 * as possible, even if that means splitting an existing node so that its first
3590 * part is moved to the preceeding node. This would maximise the efficiency of
3591 * memEQ during matching. Elsewhere in this file, khw proposes splitting
3592 * EXACTFish nodes into portions that don't change under folding vs those that
3593 * do. Those portions that don't change may be the only things in the pattern that
3594 * could be used to find fixed and floating strings.
3596 * If a node is to match under /i (folded), the number of characters it matches
3597 * can be different than its character length if it contains a multi-character
3598 * fold. *min_subtract is set to the total delta number of characters of the
3601 * And *unfolded_multi_char is set to indicate whether or not the node contains
3602 * an unfolded multi-char fold. This happens when whether the fold is valid or
3603 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3604 * SMALL LETTER SHARP S, as only if the target string being matched against
3605 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3606 * folding rules depend on the locale in force at runtime. (Multi-char folds
3607 * whose components are all above the Latin1 range are not run-time locale
3608 * dependent, and have already been folded by the time this function is
3611 * This is as good a place as any to discuss the design of handling these
3612 * multi-character fold sequences. It's been wrong in Perl for a very long
3613 * time. There are three code points in Unicode whose multi-character folds
3614 * were long ago discovered to mess things up. The previous designs for
3615 * dealing with these involved assigning a special node for them. This
3616 * approach doesn't always work, as evidenced by this example:
3617 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3618 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3619 * would match just the \xDF, it won't be able to handle the case where a
3620 * successful match would have to cross the node's boundary. The new approach
3621 * that hopefully generally solves the problem generates an EXACTFU_SS node
3622 * that is "sss" in this case.
3624 * It turns out that there are problems with all multi-character folds, and not
3625 * just these three. Now the code is general, for all such cases. The
3626 * approach taken is:
3627 * 1) This routine examines each EXACTFish node that could contain multi-
3628 * character folded sequences. Since a single character can fold into
3629 * such a sequence, the minimum match length for this node is less than
3630 * the number of characters in the node. This routine returns in
3631 * *min_subtract how many characters to subtract from the the actual
3632 * length of the string to get a real minimum match length; it is 0 if
3633 * there are no multi-char foldeds. This delta is used by the caller to
3634 * adjust the min length of the match, and the delta between min and max,
3635 * so that the optimizer doesn't reject these possibilities based on size
3637 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3638 * is used for an EXACTFU node that contains at least one "ss" sequence in
3639 * it. For non-UTF-8 patterns and strings, this is the only case where
3640 * there is a possible fold length change. That means that a regular
3641 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3642 * with length changes, and so can be processed faster. regexec.c takes
3643 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3644 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3645 * known until runtime). This saves effort in regex matching. However,
3646 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3647 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3648 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3649 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3650 * possibilities for the non-UTF8 patterns are quite simple, except for
3651 * the sharp s. All the ones that don't involve a UTF-8 target string are
3652 * members of a fold-pair, and arrays are set up for all of them so that
3653 * the other member of the pair can be found quickly. Code elsewhere in
3654 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3655 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3656 * described in the next item.
3657 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3658 * validity of the fold won't be known until runtime, and so must remain
3659 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3660 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3661 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3662 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3663 * The reason this is a problem is that the optimizer part of regexec.c
3664 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3665 * that a character in the pattern corresponds to at most a single
3666 * character in the target string. (And I do mean character, and not byte
3667 * here, unlike other parts of the documentation that have never been
3668 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3669 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3670 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3671 * nodes, violate the assumption, and they are the only instances where it
3672 * is violated. I'm reluctant to try to change the assumption, as the
3673 * code involved is impenetrable to me (khw), so instead the code here
3674 * punts. This routine examines EXACTFL nodes, and (when the pattern
3675 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3676 * boolean indicating whether or not the node contains such a fold. When
3677 * it is true, the caller sets a flag that later causes the optimizer in
3678 * this file to not set values for the floating and fixed string lengths,
3679 * and thus avoids the optimizer code in regexec.c that makes the invalid
3680 * assumption. Thus, there is no optimization based on string lengths for
3681 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3682 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3683 * assumption is wrong only in these cases is that all other non-UTF-8
3684 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3685 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3686 * EXACTF nodes because we don't know at compile time if it actually
3687 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3688 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3689 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3690 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3691 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3692 * string would require the pattern to be forced into UTF-8, the overhead
3693 * of which we want to avoid. Similarly the unfolded multi-char folds in
3694 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3697 * Similarly, the code that generates tries doesn't currently handle
3698 * not-already-folded multi-char folds, and it looks like a pain to change
3699 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3700 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3701 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3702 * using /iaa matching will be doing so almost entirely with ASCII
3703 * strings, so this should rarely be encountered in practice */
3705 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3706 if (PL_regkind[OP(scan)] == EXACT) \
3707 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3710 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3711 UV *min_subtract, bool *unfolded_multi_char,
3712 U32 flags,regnode *val, U32 depth)
3714 /* Merge several consecutive EXACTish nodes into one. */
3715 regnode *n = regnext(scan);
3717 regnode *next = scan + NODE_SZ_STR(scan);
3721 regnode *stop = scan;
3722 GET_RE_DEBUG_FLAGS_DECL;
3724 PERL_UNUSED_ARG(depth);
3727 PERL_ARGS_ASSERT_JOIN_EXACT;
3728 #ifndef EXPERIMENTAL_INPLACESCAN
3729 PERL_UNUSED_ARG(flags);
3730 PERL_UNUSED_ARG(val);
3732 DEBUG_PEEP("join",scan,depth);
3734 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3735 * EXACT ones that are mergeable to the current one. */
3737 && (PL_regkind[OP(n)] == NOTHING
3738 || (stringok && OP(n) == OP(scan)))
3740 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3743 if (OP(n) == TAIL || n > next)
3745 if (PL_regkind[OP(n)] == NOTHING) {
3746 DEBUG_PEEP("skip:",n,depth);
3747 NEXT_OFF(scan) += NEXT_OFF(n);
3748 next = n + NODE_STEP_REGNODE;
3755 else if (stringok) {
3756 const unsigned int oldl = STR_LEN(scan);
3757 regnode * const nnext = regnext(n);
3759 /* XXX I (khw) kind of doubt that this works on platforms (should
3760 * Perl ever run on one) where U8_MAX is above 255 because of lots
3761 * of other assumptions */
3762 /* Don't join if the sum can't fit into a single node */
3763 if (oldl + STR_LEN(n) > U8_MAX)
3766 DEBUG_PEEP("merg",n,depth);
3769 NEXT_OFF(scan) += NEXT_OFF(n);
3770 STR_LEN(scan) += STR_LEN(n);
3771 next = n + NODE_SZ_STR(n);
3772 /* Now we can overwrite *n : */
3773 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3781 #ifdef EXPERIMENTAL_INPLACESCAN
3782 if (flags && !NEXT_OFF(n)) {
3783 DEBUG_PEEP("atch", val, depth);
3784 if (reg_off_by_arg[OP(n)]) {
3785 ARG_SET(n, val - n);
3788 NEXT_OFF(n) = val - n;
3796 *unfolded_multi_char = FALSE;
3798 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3799 * can now analyze for sequences of problematic code points. (Prior to
3800 * this final joining, sequences could have been split over boundaries, and
3801 * hence missed). The sequences only happen in folding, hence for any
3802 * non-EXACT EXACTish node */
3803 if (OP(scan) != EXACT && OP(scan) != EXACTL) {
3804 U8* s0 = (U8*) STRING(scan);
3806 U8* s_end = s0 + STR_LEN(scan);
3808 int total_count_delta = 0; /* Total delta number of characters that
3809 multi-char folds expand to */
3811 /* One pass is made over the node's string looking for all the
3812 * possibilities. To avoid some tests in the loop, there are two main
3813 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3818 if (OP(scan) == EXACTFL) {
3821 /* An EXACTFL node would already have been changed to another
3822 * node type unless there is at least one character in it that
3823 * is problematic; likely a character whose fold definition
3824 * won't be known until runtime, and so has yet to be folded.
3825 * For all but the UTF-8 locale, folds are 1-1 in length, but
3826 * to handle the UTF-8 case, we need to create a temporary
3827 * folded copy using UTF-8 locale rules in order to analyze it.
3828 * This is because our macros that look to see if a sequence is
3829 * a multi-char fold assume everything is folded (otherwise the
3830 * tests in those macros would be too complicated and slow).
3831 * Note that here, the non-problematic folds will have already
3832 * been done, so we can just copy such characters. We actually
3833 * don't completely fold the EXACTFL string. We skip the
3834 * unfolded multi-char folds, as that would just create work
3835 * below to figure out the size they already are */
3837 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3840 STRLEN s_len = UTF8SKIP(s);
3841 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3842 Copy(s, d, s_len, U8);
3845 else if (is_FOLDS_TO_MULTI_utf8(s)) {
3846 *unfolded_multi_char = TRUE;
3847 Copy(s, d, s_len, U8);
3850 else if (isASCII(*s)) {
3851 *(d++) = toFOLD(*s);
3855 _to_utf8_fold_flags(s, d, &len, FOLD_FLAGS_FULL);
3861 /* Point the remainder of the routine to look at our temporary
3865 } /* End of creating folded copy of EXACTFL string */
3867 /* Examine the string for a multi-character fold sequence. UTF-8
3868 * patterns have all characters pre-folded by the time this code is
3870 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
3871 length sequence we are looking for is 2 */
3873 int count = 0; /* How many characters in a multi-char fold */
3874 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
3875 if (! len) { /* Not a multi-char fold: get next char */
3880 /* Nodes with 'ss' require special handling, except for
3881 * EXACTFA-ish for which there is no multi-char fold to this */
3882 if (len == 2 && *s == 's' && *(s+1) == 's'
3883 && OP(scan) != EXACTFA
3884 && OP(scan) != EXACTFA_NO_TRIE)
3887 if (OP(scan) != EXACTFL) {
3888 OP(scan) = EXACTFU_SS;
3892 else { /* Here is a generic multi-char fold. */
3893 U8* multi_end = s + len;
3895 /* Count how many characters are in it. In the case of
3896 * /aa, no folds which contain ASCII code points are
3897 * allowed, so check for those, and skip if found. */
3898 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE) {
3899 count = utf8_length(s, multi_end);
3903 while (s < multi_end) {
3906 goto next_iteration;
3916 /* The delta is how long the sequence is minus 1 (1 is how long
3917 * the character that folds to the sequence is) */
3918 total_count_delta += count - 1;
3922 /* We created a temporary folded copy of the string in EXACTFL
3923 * nodes. Therefore we need to be sure it doesn't go below zero,
3924 * as the real string could be shorter */
3925 if (OP(scan) == EXACTFL) {
3926 int total_chars = utf8_length((U8*) STRING(scan),
3927 (U8*) STRING(scan) + STR_LEN(scan));
3928 if (total_count_delta > total_chars) {
3929 total_count_delta = total_chars;
3933 *min_subtract += total_count_delta;
3936 else if (OP(scan) == EXACTFA) {
3938 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
3939 * fold to the ASCII range (and there are no existing ones in the
3940 * upper latin1 range). But, as outlined in the comments preceding
3941 * this function, we need to flag any occurrences of the sharp s.
3942 * This character forbids trie formation (because of added
3944 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
3945 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
3946 || UNICODE_DOT_DOT_VERSION > 0)
3948 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
3949 OP(scan) = EXACTFA_NO_TRIE;
3950 *unfolded_multi_char = TRUE;
3958 /* Non-UTF-8 pattern, not EXACTFA node. Look for the multi-char
3959 * folds that are all Latin1. As explained in the comments
3960 * preceding this function, we look also for the sharp s in EXACTF
3961 * and EXACTFL nodes; it can be in the final position. Otherwise
3962 * we can stop looking 1 byte earlier because have to find at least
3963 * two characters for a multi-fold */
3964 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
3969 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
3970 if (! len) { /* Not a multi-char fold. */
3971 if (*s == LATIN_SMALL_LETTER_SHARP_S
3972 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
3974 *unfolded_multi_char = TRUE;
https://perl5.git.perl.org / perl5.git / blob