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))
108 /* this is a chain of data about sub patterns we are processing that
109 need to be handled separately/specially in study_chunk. Its so
110 we can simulate recursion without losing state. */
112 typedef struct scan_frame {
113 regnode *last_regnode; /* last node to process in this frame */
114 regnode *next_regnode; /* next node to process when last is reached */
115 U32 prev_recursed_depth;
116 I32 stopparen; /* what stopparen do we use */
117 U32 is_top_frame; /* what flags do we use? */
119 struct scan_frame *this_prev_frame; /* this previous frame */
120 struct scan_frame *prev_frame; /* previous frame */
121 struct scan_frame *next_frame; /* next frame */
124 /* Certain characters are output as a sequence with the first being a
126 #define isBACKSLASHED_PUNCT(c) \
127 ((c) == '-' || (c) == ']' || (c) == '\\' || (c) == '^')
130 struct RExC_state_t {
131 U32 flags; /* RXf_* are we folding, multilining? */
132 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
133 char *precomp; /* uncompiled string. */
134 char *precomp_end; /* pointer to end of uncompiled string. */
135 REGEXP *rx_sv; /* The SV that is the regexp. */
136 regexp *rx; /* perl core regexp structure */
137 regexp_internal *rxi; /* internal data for regexp object
139 char *start; /* Start of input for compile */
140 char *end; /* End of input for compile */
141 char *parse; /* Input-scan pointer. */
142 char *adjusted_start; /* 'start', adjusted. See code use */
143 STRLEN precomp_adj; /* an offset beyond precomp. See code use */
144 SSize_t whilem_seen; /* number of WHILEM in this expr */
145 regnode *emit_start; /* Start of emitted-code area */
146 regnode *emit_bound; /* First regnode outside of the
148 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
149 implies compiling, so don't emit */
150 regnode_ssc emit_dummy; /* placeholder for emit to point to;
151 large enough for the largest
152 non-EXACTish node, so can use it as
154 I32 naughty; /* How bad is this pattern? */
155 I32 sawback; /* Did we see \1, ...? */
157 SSize_t size; /* Code size. */
158 I32 npar; /* Capture buffer count, (OPEN) plus
159 one. ("par" 0 is the whole
161 I32 nestroot; /* root parens we are in - used by
165 regnode **open_parens; /* pointers to open parens */
166 regnode **close_parens; /* pointers to close parens */
167 regnode *opend; /* END node in program */
168 I32 utf8; /* whether the pattern is utf8 or not */
169 I32 orig_utf8; /* whether the pattern was originally in utf8 */
170 /* XXX use this for future optimisation of case
171 * where pattern must be upgraded to utf8. */
172 I32 uni_semantics; /* If a d charset modifier should use unicode
173 rules, even if the pattern is not in
175 HV *paren_names; /* Paren names */
177 regnode **recurse; /* Recurse regops */
178 I32 recurse_count; /* Number of recurse regops */
179 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
181 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
185 I32 override_recoding;
187 I32 recode_x_to_native;
189 I32 in_multi_char_class;
190 struct reg_code_block *code_blocks; /* positions of literal (?{})
192 int num_code_blocks; /* size of code_blocks[] */
193 int code_index; /* next code_blocks[] slot */
194 SSize_t maxlen; /* mininum possible number of chars in string to match */
195 scan_frame *frame_head;
196 scan_frame *frame_last;
198 #ifdef ADD_TO_REGEXEC
199 char *starttry; /* -Dr: where regtry was called. */
200 #define RExC_starttry (pRExC_state->starttry)
202 SV *runtime_code_qr; /* qr with the runtime code blocks */
204 const char *lastparse;
206 AV *paren_name_list; /* idx -> name */
207 U32 study_chunk_recursed_count;
210 #define RExC_lastparse (pRExC_state->lastparse)
211 #define RExC_lastnum (pRExC_state->lastnum)
212 #define RExC_paren_name_list (pRExC_state->paren_name_list)
213 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
214 #define RExC_mysv (pRExC_state->mysv1)
215 #define RExC_mysv1 (pRExC_state->mysv1)
216 #define RExC_mysv2 (pRExC_state->mysv2)
219 bool seen_unfolded_sharp_s;
223 #define RExC_flags (pRExC_state->flags)
224 #define RExC_pm_flags (pRExC_state->pm_flags)
225 #define RExC_precomp (pRExC_state->precomp)
226 #define RExC_precomp_adj (pRExC_state->precomp_adj)
227 #define RExC_adjusted_start (pRExC_state->adjusted_start)
228 #define RExC_precomp_end (pRExC_state->precomp_end)
229 #define RExC_rx_sv (pRExC_state->rx_sv)
230 #define RExC_rx (pRExC_state->rx)
231 #define RExC_rxi (pRExC_state->rxi)
232 #define RExC_start (pRExC_state->start)
233 #define RExC_end (pRExC_state->end)
234 #define RExC_parse (pRExC_state->parse)
235 #define RExC_whilem_seen (pRExC_state->whilem_seen)
237 /* Set during the sizing pass when there is a LATIN SMALL LETTER SHARP S in any
238 * EXACTF node, hence was parsed under /di rules. If later in the parse,
239 * something forces the pattern into using /ui rules, the sharp s should be
240 * folded into the sequence 'ss', which takes up more space than previously
241 * calculated. This means that the sizing pass needs to be restarted. (The
242 * node also becomes an EXACTFU_SS.) For all other characters, an EXACTF node
243 * that gets converted to /ui (and EXACTFU) occupies the same amount of space,
244 * so there is no need to resize [perl #125990]. */
245 #define RExC_seen_unfolded_sharp_s (pRExC_state->seen_unfolded_sharp_s)
247 #ifdef RE_TRACK_PATTERN_OFFSETS
248 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
251 #define RExC_emit (pRExC_state->emit)
252 #define RExC_emit_dummy (pRExC_state->emit_dummy)
253 #define RExC_emit_start (pRExC_state->emit_start)
254 #define RExC_emit_bound (pRExC_state->emit_bound)
255 #define RExC_sawback (pRExC_state->sawback)
256 #define RExC_seen (pRExC_state->seen)
257 #define RExC_size (pRExC_state->size)
258 #define RExC_maxlen (pRExC_state->maxlen)
259 #define RExC_npar (pRExC_state->npar)
260 #define RExC_nestroot (pRExC_state->nestroot)
261 #define RExC_extralen (pRExC_state->extralen)
262 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
263 #define RExC_utf8 (pRExC_state->utf8)
264 #define RExC_uni_semantics (pRExC_state->uni_semantics)
265 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
266 #define RExC_open_parens (pRExC_state->open_parens)
267 #define RExC_close_parens (pRExC_state->close_parens)
268 #define RExC_opend (pRExC_state->opend)
269 #define RExC_paren_names (pRExC_state->paren_names)
270 #define RExC_recurse (pRExC_state->recurse)
271 #define RExC_recurse_count (pRExC_state->recurse_count)
272 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
273 #define RExC_study_chunk_recursed_bytes \
274 (pRExC_state->study_chunk_recursed_bytes)
275 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
276 #define RExC_contains_locale (pRExC_state->contains_locale)
277 #define RExC_contains_i (pRExC_state->contains_i)
278 #define RExC_override_recoding (pRExC_state->override_recoding)
280 # define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
282 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
283 #define RExC_frame_head (pRExC_state->frame_head)
284 #define RExC_frame_last (pRExC_state->frame_last)
285 #define RExC_frame_count (pRExC_state->frame_count)
286 #define RExC_strict (pRExC_state->strict)
288 /* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
289 * a flag to disable back-off on the fixed/floating substrings - if it's
290 * a high complexity pattern we assume the benefit of avoiding a full match
291 * is worth the cost of checking for the substrings even if they rarely help.
293 #define RExC_naughty (pRExC_state->naughty)
294 #define TOO_NAUGHTY (10)
295 #define MARK_NAUGHTY(add) \
296 if (RExC_naughty < TOO_NAUGHTY) \
297 RExC_naughty += (add)
298 #define MARK_NAUGHTY_EXP(exp, add) \
299 if (RExC_naughty < TOO_NAUGHTY) \
300 RExC_naughty += RExC_naughty / (exp) + (add)
302 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
303 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
304 ((*s) == '{' && regcurly(s)))
307 * Flags to be passed up and down.
309 #define WORST 0 /* Worst case. */
310 #define HASWIDTH 0x01 /* Known to match non-null strings. */
312 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
313 * character. (There needs to be a case: in the switch statement in regexec.c
314 * for any node marked SIMPLE.) Note that this is not the same thing as
317 #define SPSTART 0x04 /* Starts with * or + */
318 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
319 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
320 #define RESTART_PASS1 0x20 /* Need to restart sizing pass */
321 #define NEED_UTF8 0x40 /* In conjunction with RESTART_PASS1, need to
322 calcuate sizes as UTF-8 */
324 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
326 /* whether trie related optimizations are enabled */
327 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
328 #define TRIE_STUDY_OPT
329 #define FULL_TRIE_STUDY
335 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
336 #define PBITVAL(paren) (1 << ((paren) & 7))
337 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
338 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
339 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
341 #define REQUIRE_UTF8(flagp) STMT_START { \
344 *flagp = RESTART_PASS1|NEED_UTF8; \
349 /* Change from /d into /u rules, and restart the parse if we've already seen
350 * something whose size would increase as a result, by setting *flagp and
351 * returning 'restart_retval'. RExC_uni_semantics is a flag that indicates
352 * we've change to /u during the parse. */
353 #define REQUIRE_UNI_RULES(flagp, restart_retval) \
355 if (DEPENDS_SEMANTICS) { \
357 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
358 RExC_uni_semantics = 1; \
359 if (RExC_seen_unfolded_sharp_s) { \
360 *flagp |= RESTART_PASS1; \
361 return restart_retval; \
366 /* This converts the named class defined in regcomp.h to its equivalent class
367 * number defined in handy.h. */
368 #define namedclass_to_classnum(class) ((int) ((class) / 2))
369 #define classnum_to_namedclass(classnum) ((classnum) * 2)
371 #define _invlist_union_complement_2nd(a, b, output) \
372 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
373 #define _invlist_intersection_complement_2nd(a, b, output) \
374 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
376 /* About scan_data_t.
378 During optimisation we recurse through the regexp program performing
379 various inplace (keyhole style) optimisations. In addition study_chunk
380 and scan_commit populate this data structure with information about
381 what strings MUST appear in the pattern. We look for the longest
382 string that must appear at a fixed location, and we look for the
383 longest string that may appear at a floating location. So for instance
388 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
389 strings (because they follow a .* construct). study_chunk will identify
390 both FOO and BAR as being the longest fixed and floating strings respectively.
392 The strings can be composites, for instance
396 will result in a composite fixed substring 'foo'.
398 For each string some basic information is maintained:
400 - offset or min_offset
401 This is the position the string must appear at, or not before.
402 It also implicitly (when combined with minlenp) tells us how many
403 characters must match before the string we are searching for.
404 Likewise when combined with minlenp and the length of the string it
405 tells us how many characters must appear after the string we have
409 Only used for floating strings. This is the rightmost point that
410 the string can appear at. If set to SSize_t_MAX it indicates that the
411 string can occur infinitely far to the right.
414 A pointer to the minimum number of characters of the pattern that the
415 string was found inside. This is important as in the case of positive
416 lookahead or positive lookbehind we can have multiple patterns
421 The minimum length of the pattern overall is 3, the minimum length
422 of the lookahead part is 3, but the minimum length of the part that
423 will actually match is 1. So 'FOO's minimum length is 3, but the
424 minimum length for the F is 1. This is important as the minimum length
425 is used to determine offsets in front of and behind the string being
426 looked for. Since strings can be composites this is the length of the
427 pattern at the time it was committed with a scan_commit. Note that
428 the length is calculated by study_chunk, so that the minimum lengths
429 are not known until the full pattern has been compiled, thus the
430 pointer to the value.
434 In the case of lookbehind the string being searched for can be
435 offset past the start point of the final matching string.
436 If this value was just blithely removed from the min_offset it would
437 invalidate some of the calculations for how many chars must match
438 before or after (as they are derived from min_offset and minlen and
439 the length of the string being searched for).
440 When the final pattern is compiled and the data is moved from the
441 scan_data_t structure into the regexp structure the information
442 about lookbehind is factored in, with the information that would
443 have been lost precalculated in the end_shift field for the
446 The fields pos_min and pos_delta are used to store the minimum offset
447 and the delta to the maximum offset at the current point in the pattern.
451 typedef struct scan_data_t {
452 /*I32 len_min; unused */
453 /*I32 len_delta; unused */
457 SSize_t last_end; /* min value, <0 unless valid. */
458 SSize_t last_start_min;
459 SSize_t last_start_max;
460 SV **longest; /* Either &l_fixed, or &l_float. */
461 SV *longest_fixed; /* longest fixed string found in pattern */
462 SSize_t offset_fixed; /* offset where it starts */
463 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
464 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
465 SV *longest_float; /* longest floating string found in pattern */
466 SSize_t offset_float_min; /* earliest point in string it can appear */
467 SSize_t offset_float_max; /* latest point in string it can appear */
468 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
469 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
472 SSize_t *last_closep;
473 regnode_ssc *start_class;
477 * Forward declarations for pregcomp()'s friends.
480 static const scan_data_t zero_scan_data =
481 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
483 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
484 #define SF_BEFORE_SEOL 0x0001
485 #define SF_BEFORE_MEOL 0x0002
486 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
487 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
489 #define SF_FIX_SHIFT_EOL (+2)
490 #define SF_FL_SHIFT_EOL (+4)
492 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
493 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
495 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
496 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
497 #define SF_IS_INF 0x0040
498 #define SF_HAS_PAR 0x0080
499 #define SF_IN_PAR 0x0100
500 #define SF_HAS_EVAL 0x0200
501 #define SCF_DO_SUBSTR 0x0400
502 #define SCF_DO_STCLASS_AND 0x0800
503 #define SCF_DO_STCLASS_OR 0x1000
504 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
505 #define SCF_WHILEM_VISITED_POS 0x2000
507 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
508 #define SCF_SEEN_ACCEPT 0x8000
509 #define SCF_TRIE_DOING_RESTUDY 0x10000
510 #define SCF_IN_DEFINE 0x20000
515 #define UTF cBOOL(RExC_utf8)
517 /* The enums for all these are ordered so things work out correctly */
518 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
519 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
520 == REGEX_DEPENDS_CHARSET)
521 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
522 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
523 >= REGEX_UNICODE_CHARSET)
524 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
525 == REGEX_ASCII_RESTRICTED_CHARSET)
526 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
527 >= REGEX_ASCII_RESTRICTED_CHARSET)
528 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
529 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
531 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
533 /* For programs that want to be strictly Unicode compatible by dying if any
534 * attempt is made to match a non-Unicode code point against a Unicode
536 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
538 #define OOB_NAMEDCLASS -1
540 /* There is no code point that is out-of-bounds, so this is problematic. But
541 * its only current use is to initialize a variable that is always set before
543 #define OOB_UNICODE 0xDEADBEEF
545 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
546 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
549 /* length of regex to show in messages that don't mark a position within */
550 #define RegexLengthToShowInErrorMessages 127
553 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
554 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
555 * op/pragma/warn/regcomp.
557 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
558 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
560 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
561 " in m/%"UTF8f MARKER2 "%"UTF8f"/"
563 /* The code in this file in places uses one level of recursion with parsing
564 * rebased to an alternate string constructed by us in memory. This can take
565 * the form of something that is completely different from the input, or
566 * something that uses the input as part of the alternate. In the first case,
567 * there should be no possibility of an error, as we are in complete control of
568 * the alternate string. But in the second case we don't control the input
569 * portion, so there may be errors in that. Here's an example:
571 * is handled specially because \x{df} folds to a sequence of more than one
572 * character, 'ss'. What is done is to create and parse an alternate string,
573 * which looks like this:
574 * /(?:\x{DF}|[abc\x{DF}def])/ui
575 * where it uses the input unchanged in the middle of something it constructs,
576 * which is a branch for the DF outside the character class, and clustering
577 * parens around the whole thing. (It knows enough to skip the DF inside the
578 * class while in this substitute parse.) 'abc' and 'def' may have errors that
579 * need to be reported. The general situation looks like this:
582 * Input: ----------------------------------------------------
583 * Constructed: ---------------------------------------------------
586 * The input string sI..eI is the input pattern. The string sC..EC is the
587 * constructed substitute parse string. The portions sC..tC and eC..EC are
588 * constructed by us. The portion tC..eC is an exact duplicate of the input
589 * pattern tI..eI. In the diagram, these are vertically aligned. Suppose that
590 * while parsing, we find an error at xC. We want to display a message showing
591 * the real input string. Thus we need to find the point xI in it which
592 * corresponds to xC. xC >= tC, since the portion of the string sC..tC has
593 * been constructed by us, and so shouldn't have errors. We get:
595 * xI = sI + (tI - sI) + (xC - tC)
597 * and, the offset into sI is:
599 * (xI - sI) = (tI - sI) + (xC - tC)
601 * When the substitute is constructed, we save (tI -sI) as RExC_precomp_adj,
602 * and we save tC as RExC_adjusted_start.
604 * During normal processing of the input pattern, everything points to that,
605 * with RExC_precomp_adj set to 0, and RExC_adjusted_start set to sI.
608 #define tI_sI RExC_precomp_adj
609 #define tC RExC_adjusted_start
610 #define sC RExC_precomp
611 #define xI_offset(xC) ((IV) (tI_sI + (xC - tC)))
612 #define xI(xC) (sC + xI_offset(xC))
613 #define eC RExC_precomp_end
615 #define REPORT_LOCATION_ARGS(xC) \
617 (xI(xC) > eC) /* Don't run off end */ \
618 ? eC - sC /* Length before the <--HERE */ \
620 sC), /* The input pattern printed up to the <--HERE */ \
622 (xI(xC) > eC) ? 0 : eC - xI(xC), /* Length after <--HERE */ \
623 (xI(xC) > eC) ? eC : xI(xC)) /* pattern after <--HERE */
625 /* Used to point after bad bytes for an error message, but avoid skipping
626 * past a nul byte. */
627 #define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
630 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
631 * arg. Show regex, up to a maximum length. If it's too long, chop and add
634 #define _FAIL(code) STMT_START { \
635 const char *ellipses = ""; \
636 IV len = RExC_precomp_end - RExC_precomp; \
639 SAVEFREESV(RExC_rx_sv); \
640 if (len > RegexLengthToShowInErrorMessages) { \
641 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
642 len = RegexLengthToShowInErrorMessages - 10; \
648 #define FAIL(msg) _FAIL( \
649 Perl_croak(aTHX_ "%s in regex m/%"UTF8f"%s/", \
650 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
652 #define FAIL2(msg,arg) _FAIL( \
653 Perl_croak(aTHX_ msg " in regex m/%"UTF8f"%s/", \
654 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
657 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
659 #define Simple_vFAIL(m) STMT_START { \
660 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
661 m, REPORT_LOCATION_ARGS(RExC_parse)); \
665 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
667 #define vFAIL(m) STMT_START { \
669 SAVEFREESV(RExC_rx_sv); \
674 * Like Simple_vFAIL(), but accepts two arguments.
676 #define Simple_vFAIL2(m,a1) STMT_START { \
677 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
678 REPORT_LOCATION_ARGS(RExC_parse)); \
682 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
684 #define vFAIL2(m,a1) STMT_START { \
686 SAVEFREESV(RExC_rx_sv); \
687 Simple_vFAIL2(m, a1); \
692 * Like Simple_vFAIL(), but accepts three arguments.
694 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
695 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
696 REPORT_LOCATION_ARGS(RExC_parse)); \
700 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
702 #define vFAIL3(m,a1,a2) STMT_START { \
704 SAVEFREESV(RExC_rx_sv); \
705 Simple_vFAIL3(m, a1, a2); \
709 * Like Simple_vFAIL(), but accepts four arguments.
711 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
712 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
713 REPORT_LOCATION_ARGS(RExC_parse)); \
716 #define vFAIL4(m,a1,a2,a3) STMT_START { \
718 SAVEFREESV(RExC_rx_sv); \
719 Simple_vFAIL4(m, a1, a2, a3); \
722 /* A specialized version of vFAIL2 that works with UTF8f */
723 #define vFAIL2utf8f(m, a1) STMT_START { \
725 SAVEFREESV(RExC_rx_sv); \
726 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
727 REPORT_LOCATION_ARGS(RExC_parse)); \
730 #define vFAIL3utf8f(m, a1, a2) STMT_START { \
732 SAVEFREESV(RExC_rx_sv); \
733 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
734 REPORT_LOCATION_ARGS(RExC_parse)); \
737 /* These have asserts in them because of [perl #122671] Many warnings in
738 * regcomp.c can occur twice. If they get output in pass1 and later in that
739 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
740 * would get output again. So they should be output in pass2, and these
741 * asserts make sure new warnings follow that paradigm. */
743 /* m is not necessarily a "literal string", in this macro */
744 #define reg_warn_non_literal_string(loc, m) STMT_START { \
745 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
746 "%s" REPORT_LOCATION, \
747 m, REPORT_LOCATION_ARGS(loc)); \
750 #define ckWARNreg(loc,m) STMT_START { \
751 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
753 REPORT_LOCATION_ARGS(loc)); \
756 #define vWARN(loc, m) STMT_START { \
757 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
759 REPORT_LOCATION_ARGS(loc)); \
762 #define vWARN_dep(loc, m) STMT_START { \
763 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
765 REPORT_LOCATION_ARGS(loc)); \
768 #define ckWARNdep(loc,m) STMT_START { \
769 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
771 REPORT_LOCATION_ARGS(loc)); \
774 #define ckWARNregdep(loc,m) STMT_START { \
775 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
778 REPORT_LOCATION_ARGS(loc)); \
781 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
782 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
784 a1, REPORT_LOCATION_ARGS(loc)); \
787 #define ckWARN2reg(loc, m, a1) STMT_START { \
788 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
790 a1, REPORT_LOCATION_ARGS(loc)); \
793 #define vWARN3(loc, m, a1, a2) STMT_START { \
794 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
796 a1, a2, REPORT_LOCATION_ARGS(loc)); \
799 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
800 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
803 REPORT_LOCATION_ARGS(loc)); \
806 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
807 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
810 REPORT_LOCATION_ARGS(loc)); \
813 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
814 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
817 REPORT_LOCATION_ARGS(loc)); \
820 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
821 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
824 REPORT_LOCATION_ARGS(loc)); \
827 /* Macros for recording node offsets. 20001227 mjd@plover.com
828 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
829 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
830 * Element 0 holds the number n.
831 * Position is 1 indexed.
833 #ifndef RE_TRACK_PATTERN_OFFSETS
834 #define Set_Node_Offset_To_R(node,byte)
835 #define Set_Node_Offset(node,byte)
836 #define Set_Cur_Node_Offset
837 #define Set_Node_Length_To_R(node,len)
838 #define Set_Node_Length(node,len)
839 #define Set_Node_Cur_Length(node,start)
840 #define Node_Offset(n)
841 #define Node_Length(n)
842 #define Set_Node_Offset_Length(node,offset,len)
843 #define ProgLen(ri) ri->u.proglen
844 #define SetProgLen(ri,x) ri->u.proglen = x
846 #define ProgLen(ri) ri->u.offsets[0]
847 #define SetProgLen(ri,x) ri->u.offsets[0] = x
848 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
850 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
851 __LINE__, (int)(node), (int)(byte))); \
853 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
856 RExC_offsets[2*(node)-1] = (byte); \
861 #define Set_Node_Offset(node,byte) \
862 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
863 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
865 #define Set_Node_Length_To_R(node,len) STMT_START { \
867 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
868 __LINE__, (int)(node), (int)(len))); \
870 Perl_croak(aTHX_ "value of node is %d in Length macro", \
873 RExC_offsets[2*(node)] = (len); \
878 #define Set_Node_Length(node,len) \
879 Set_Node_Length_To_R((node)-RExC_emit_start, len)
880 #define Set_Node_Cur_Length(node, start) \
881 Set_Node_Length(node, RExC_parse - start)
883 /* Get offsets and lengths */
884 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
885 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
887 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
888 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
889 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
893 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
894 #define EXPERIMENTAL_INPLACESCAN
895 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
897 #define DEBUG_RExC_seen() \
898 DEBUG_OPTIMISE_MORE_r({ \
899 PerlIO_printf(Perl_debug_log,"RExC_seen: "); \
901 if (RExC_seen & REG_ZERO_LEN_SEEN) \
902 PerlIO_printf(Perl_debug_log,"REG_ZERO_LEN_SEEN "); \
904 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
905 PerlIO_printf(Perl_debug_log,"REG_LOOKBEHIND_SEEN "); \
907 if (RExC_seen & REG_GPOS_SEEN) \
908 PerlIO_printf(Perl_debug_log,"REG_GPOS_SEEN "); \
910 if (RExC_seen & REG_RECURSE_SEEN) \
911 PerlIO_printf(Perl_debug_log,"REG_RECURSE_SEEN "); \
913 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
914 PerlIO_printf(Perl_debug_log,"REG_TOP_LEVEL_BRANCHES_SEEN "); \
916 if (RExC_seen & REG_VERBARG_SEEN) \
917 PerlIO_printf(Perl_debug_log,"REG_VERBARG_SEEN "); \
919 if (RExC_seen & REG_CUTGROUP_SEEN) \
920 PerlIO_printf(Perl_debug_log,"REG_CUTGROUP_SEEN "); \
922 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
923 PerlIO_printf(Perl_debug_log,"REG_RUN_ON_COMMENT_SEEN "); \
925 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
926 PerlIO_printf(Perl_debug_log,"REG_UNFOLDED_MULTI_SEEN "); \
928 if (RExC_seen & REG_GOSTART_SEEN) \
929 PerlIO_printf(Perl_debug_log,"REG_GOSTART_SEEN "); \
931 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
932 PerlIO_printf(Perl_debug_log,"REG_UNBOUNDED_QUANTIFIER_SEEN "); \
934 PerlIO_printf(Perl_debug_log,"\n"); \
937 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
938 if ((flags) & flag) PerlIO_printf(Perl_debug_log, "%s ", #flag)
940 #define DEBUG_SHOW_STUDY_FLAGS(flags,open_str,close_str) \
942 PerlIO_printf(Perl_debug_log, "%s", open_str); \
943 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_SEOL); \
944 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_MEOL); \
945 DEBUG_SHOW_STUDY_FLAG(flags,SF_IS_INF); \
946 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_PAR); \
947 DEBUG_SHOW_STUDY_FLAG(flags,SF_IN_PAR); \
948 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_EVAL); \
949 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_SUBSTR); \
950 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_AND); \
951 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_OR); \
952 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS); \
953 DEBUG_SHOW_STUDY_FLAG(flags,SCF_WHILEM_VISITED_POS); \
954 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_RESTUDY); \
955 DEBUG_SHOW_STUDY_FLAG(flags,SCF_SEEN_ACCEPT); \
956 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_DOING_RESTUDY); \
957 DEBUG_SHOW_STUDY_FLAG(flags,SCF_IN_DEFINE); \
958 PerlIO_printf(Perl_debug_log, "%s", close_str); \
962 #define DEBUG_STUDYDATA(str,data,depth) \
963 DEBUG_OPTIMISE_MORE_r(if(data){ \
964 PerlIO_printf(Perl_debug_log, \
965 "%*s" str "Pos:%"IVdf"/%"IVdf \
967 (int)(depth)*2, "", \
968 (IV)((data)->pos_min), \
969 (IV)((data)->pos_delta), \
970 (UV)((data)->flags) \
972 DEBUG_SHOW_STUDY_FLAGS((data)->flags," [ ","]"); \
973 PerlIO_printf(Perl_debug_log, \
974 " Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
975 (IV)((data)->whilem_c), \
976 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
977 is_inf ? "INF " : "" \
979 if ((data)->last_found) \
980 PerlIO_printf(Perl_debug_log, \
981 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
982 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
983 SvPVX_const((data)->last_found), \
984 (IV)((data)->last_end), \
985 (IV)((data)->last_start_min), \
986 (IV)((data)->last_start_max), \
987 ((data)->longest && \
988 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
989 SvPVX_const((data)->longest_fixed), \
990 (IV)((data)->offset_fixed), \
991 ((data)->longest && \
992 (data)->longest==&((data)->longest_float)) ? "*" : "", \
993 SvPVX_const((data)->longest_float), \
994 (IV)((data)->offset_float_min), \
995 (IV)((data)->offset_float_max) \
997 PerlIO_printf(Perl_debug_log,"\n"); \
1000 /* =========================================================
1001 * BEGIN edit_distance stuff.
1003 * This calculates how many single character changes of any type are needed to
1004 * transform a string into another one. It is taken from version 3.1 of
1006 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1009 /* Our unsorted dictionary linked list. */
1010 /* Note we use UVs, not chars. */
1015 struct dictionary* next;
1017 typedef struct dictionary item;
1020 PERL_STATIC_INLINE item*
1021 push(UV key,item* curr)
1024 Newxz(head, 1, item);
1032 PERL_STATIC_INLINE item*
1033 find(item* head, UV key)
1035 item* iterator = head;
1037 if (iterator->key == key){
1040 iterator = iterator->next;
1046 PERL_STATIC_INLINE item*
1047 uniquePush(item* head,UV key)
1049 item* iterator = head;
1052 if (iterator->key == key) {
1055 iterator = iterator->next;
1058 return push(key,head);
1061 PERL_STATIC_INLINE void
1062 dict_free(item* head)
1064 item* iterator = head;
1067 item* temp = iterator;
1068 iterator = iterator->next;
1075 /* End of Dictionary Stuff */
1077 /* All calculations/work are done here */
1079 S_edit_distance(const UV* src,
1081 const STRLEN x, /* length of src[] */
1082 const STRLEN y, /* length of tgt[] */
1083 const SSize_t maxDistance
1087 UV swapCount,swapScore,targetCharCount,i,j;
1089 UV score_ceil = x + y;
1091 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1093 /* intialize matrix start values */
1094 Newxz(scores, ( (x + 2) * (y + 2)), UV);
1095 scores[0] = score_ceil;
1096 scores[1 * (y + 2) + 0] = score_ceil;
1097 scores[0 * (y + 2) + 1] = score_ceil;
1098 scores[1 * (y + 2) + 1] = 0;
1099 head = uniquePush(uniquePush(head,src[0]),tgt[0]);
1104 for (i=1;i<=x;i++) {
1106 head = uniquePush(head,src[i]);
1107 scores[(i+1) * (y + 2) + 1] = i;
1108 scores[(i+1) * (y + 2) + 0] = score_ceil;
1111 for (j=1;j<=y;j++) {
1114 head = uniquePush(head,tgt[j]);
1115 scores[1 * (y + 2) + (j + 1)] = j;
1116 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1119 targetCharCount = find(head,tgt[j-1])->value;
1120 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1122 if (src[i-1] != tgt[j-1]){
1123 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));
1127 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1131 find(head,src[i-1])->value = i;
1135 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1138 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1142 /* END of edit_distance() stuff
1143 * ========================================================= */
1145 /* is c a control character for which we have a mnemonic? */
1146 #define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
1149 S_cntrl_to_mnemonic(const U8 c)
1151 /* Returns the mnemonic string that represents character 'c', if one
1152 * exists; NULL otherwise. The only ones that exist for the purposes of
1153 * this routine are a few control characters */
1156 case '\a': return "\\a";
1157 case '\b': return "\\b";
1158 case ESC_NATIVE: return "\\e";
1159 case '\f': return "\\f";
1160 case '\n': return "\\n";
1161 case '\r': return "\\r";
1162 case '\t': return "\\t";
1168 /* Mark that we cannot extend a found fixed substring at this point.
1169 Update the longest found anchored substring and the longest found
1170 floating substrings if needed. */
1173 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1174 SSize_t *minlenp, int is_inf)
1176 const STRLEN l = CHR_SVLEN(data->last_found);
1177 const STRLEN old_l = CHR_SVLEN(*data->longest);
1178 GET_RE_DEBUG_FLAGS_DECL;
1180 PERL_ARGS_ASSERT_SCAN_COMMIT;
1182 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
1183 SvSetMagicSV(*data->longest, data->last_found);
1184 if (*data->longest == data->longest_fixed) {
1185 data->offset_fixed = l ? data->last_start_min : data->pos_min;
1186 if (data->flags & SF_BEFORE_EOL)
1188 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
1190 data->flags &= ~SF_FIX_BEFORE_EOL;
1191 data->minlen_fixed=minlenp;
1192 data->lookbehind_fixed=0;
1194 else { /* *data->longest == data->longest_float */
1195 data->offset_float_min = l ? data->last_start_min : data->pos_min;
1196 data->offset_float_max = (l
1197 ? data->last_start_max
1198 : (data->pos_delta > SSize_t_MAX - data->pos_min
1200 : data->pos_min + data->pos_delta));
1202 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
1203 data->offset_float_max = SSize_t_MAX;
1204 if (data->flags & SF_BEFORE_EOL)
1206 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
1208 data->flags &= ~SF_FL_BEFORE_EOL;
1209 data->minlen_float=minlenp;
1210 data->lookbehind_float=0;
1213 SvCUR_set(data->last_found, 0);
1215 SV * const sv = data->last_found;
1216 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1217 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1222 data->last_end = -1;
1223 data->flags &= ~SF_BEFORE_EOL;
1224 DEBUG_STUDYDATA("commit: ",data,0);
1227 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1228 * list that describes which code points it matches */
1231 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1233 /* Set the SSC 'ssc' to match an empty string or any code point */
1235 PERL_ARGS_ASSERT_SSC_ANYTHING;
1237 assert(is_ANYOF_SYNTHETIC(ssc));
1239 ssc->invlist = sv_2mortal(_new_invlist(2)); /* mortalize so won't leak */
1240 _append_range_to_invlist(ssc->invlist, 0, UV_MAX);
1241 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1245 S_ssc_is_anything(const regnode_ssc *ssc)
1247 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1248 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1249 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1250 * in any way, so there's no point in using it */
1255 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1257 assert(is_ANYOF_SYNTHETIC(ssc));
1259 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1263 /* See if the list consists solely of the range 0 - Infinity */
1264 invlist_iterinit(ssc->invlist);
1265 ret = invlist_iternext(ssc->invlist, &start, &end)
1269 invlist_iterfinish(ssc->invlist);
1275 /* If e.g., both \w and \W are set, matches everything */
1276 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1278 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1279 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1289 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1291 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1292 * string, any code point, or any posix class under locale */
1294 PERL_ARGS_ASSERT_SSC_INIT;
1296 Zero(ssc, 1, regnode_ssc);
1297 set_ANYOF_SYNTHETIC(ssc);
1298 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1301 /* If any portion of the regex is to operate under locale rules that aren't
1302 * fully known at compile time, initialization includes it. The reason
1303 * this isn't done for all regexes is that the optimizer was written under
1304 * the assumption that locale was all-or-nothing. Given the complexity and
1305 * lack of documentation in the optimizer, and that there are inadequate
1306 * test cases for locale, many parts of it may not work properly, it is
1307 * safest to avoid locale unless necessary. */
1308 if (RExC_contains_locale) {
1309 ANYOF_POSIXL_SETALL(ssc);
1312 ANYOF_POSIXL_ZERO(ssc);
1317 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1318 const regnode_ssc *ssc)
1320 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1321 * to the list of code points matched, and locale posix classes; hence does
1322 * not check its flags) */
1327 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1329 assert(is_ANYOF_SYNTHETIC(ssc));
1331 invlist_iterinit(ssc->invlist);
1332 ret = invlist_iternext(ssc->invlist, &start, &end)
1336 invlist_iterfinish(ssc->invlist);
1342 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1350 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1351 const regnode_charclass* const node)
1353 /* Returns a mortal inversion list defining which code points are matched
1354 * by 'node', which is of type ANYOF. Handles complementing the result if
1355 * appropriate. If some code points aren't knowable at this time, the
1356 * returned list must, and will, contain every code point that is a
1359 SV* invlist = sv_2mortal(_new_invlist(0));
1360 SV* only_utf8_locale_invlist = NULL;
1362 const U32 n = ARG(node);
1363 bool new_node_has_latin1 = FALSE;
1365 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1367 /* Look at the data structure created by S_set_ANYOF_arg() */
1368 if (n != ANYOF_ONLY_HAS_BITMAP) {
1369 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1370 AV * const av = MUTABLE_AV(SvRV(rv));
1371 SV **const ary = AvARRAY(av);
1372 assert(RExC_rxi->data->what[n] == 's');
1374 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1375 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1377 else if (ary[0] && ary[0] != &PL_sv_undef) {
1379 /* Here, no compile-time swash, and there are things that won't be
1380 * known until runtime -- we have to assume it could be anything */
1381 return _add_range_to_invlist(invlist, 0, UV_MAX);
1383 else if (ary[3] && ary[3] != &PL_sv_undef) {
1385 /* Here no compile-time swash, and no run-time only data. Use the
1386 * node's inversion list */
1387 invlist = sv_2mortal(invlist_clone(ary[3]));
1390 /* Get the code points valid only under UTF-8 locales */
1391 if ((ANYOF_FLAGS(node) & ANYOFL_FOLD)
1392 && ary[2] && ary[2] != &PL_sv_undef)
1394 only_utf8_locale_invlist = ary[2];
1398 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1399 * code points, and an inversion list for the others, but if there are code
1400 * points that should match only conditionally on the target string being
1401 * UTF-8, those are placed in the inversion list, and not the bitmap.
1402 * Since there are circumstances under which they could match, they are
1403 * included in the SSC. But if the ANYOF node is to be inverted, we have
1404 * to exclude them here, so that when we invert below, the end result
1405 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1406 * have to do this here before we add the unconditionally matched code
1408 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1409 _invlist_intersection_complement_2nd(invlist,
1414 /* Add in the points from the bit map */
1415 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1416 if (ANYOF_BITMAP_TEST(node, i)) {
1417 invlist = add_cp_to_invlist(invlist, i);
1418 new_node_has_latin1 = TRUE;
1422 /* If this can match all upper Latin1 code points, have to add them
1423 * as well. But don't add them if inverting, as when that gets done below,
1424 * it would exclude all these characters, including the ones it shouldn't
1425 * that were added just above */
1426 if (! (ANYOF_FLAGS(node) & ANYOF_INVERT) && OP(node) == ANYOFD
1427 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1429 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1432 /* Similarly for these */
1433 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1434 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1437 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1438 _invlist_invert(invlist);
1440 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOFL_FOLD) {
1442 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1443 * locale. We can skip this if there are no 0-255 at all. */
1444 _invlist_union(invlist, PL_Latin1, &invlist);
1447 /* Similarly add the UTF-8 locale possible matches. These have to be
1448 * deferred until after the non-UTF-8 locale ones are taken care of just
1449 * above, or it leads to wrong results under ANYOF_INVERT */
1450 if (only_utf8_locale_invlist) {
1451 _invlist_union_maybe_complement_2nd(invlist,
1452 only_utf8_locale_invlist,
1453 ANYOF_FLAGS(node) & ANYOF_INVERT,
1460 /* These two functions currently do the exact same thing */
1461 #define ssc_init_zero ssc_init
1463 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1464 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1466 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1467 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1468 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1471 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1472 const regnode_charclass *and_with)
1474 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1475 * another SSC or a regular ANYOF class. Can create false positives. */
1480 PERL_ARGS_ASSERT_SSC_AND;
1482 assert(is_ANYOF_SYNTHETIC(ssc));
1484 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1485 * the code point inversion list and just the relevant flags */
1486 if (is_ANYOF_SYNTHETIC(and_with)) {
1487 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1488 anded_flags = ANYOF_FLAGS(and_with);
1490 /* XXX This is a kludge around what appears to be deficiencies in the
1491 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1492 * there are paths through the optimizer where it doesn't get weeded
1493 * out when it should. And if we don't make some extra provision for
1494 * it like the code just below, it doesn't get added when it should.
1495 * This solution is to add it only when AND'ing, which is here, and
1496 * only when what is being AND'ed is the pristine, original node
1497 * matching anything. Thus it is like adding it to ssc_anything() but
1498 * only when the result is to be AND'ed. Probably the same solution
1499 * could be adopted for the same problem we have with /l matching,
1500 * which is solved differently in S_ssc_init(), and that would lead to
1501 * fewer false positives than that solution has. But if this solution
1502 * creates bugs, the consequences are only that a warning isn't raised
1503 * that should be; while the consequences for having /l bugs is
1504 * incorrect matches */
1505 if (ssc_is_anything((regnode_ssc *)and_with)) {
1506 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1510 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1511 if (OP(and_with) == ANYOFD) {
1512 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1515 anded_flags = ANYOF_FLAGS(and_with)
1516 &( ANYOF_COMMON_FLAGS
1517 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1518 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1519 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(and_with))) {
1521 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1526 ANYOF_FLAGS(ssc) &= anded_flags;
1528 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1529 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1530 * 'and_with' may be inverted. When not inverted, we have the situation of
1532 * (C1 | P1) & (C2 | P2)
1533 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1534 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1535 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1536 * <= ((C1 & C2) | P1 | P2)
1537 * Alternatively, the last few steps could be:
1538 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1539 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1540 * <= (C1 | C2 | (P1 & P2))
1541 * We favor the second approach if either P1 or P2 is non-empty. This is
1542 * because these components are a barrier to doing optimizations, as what
1543 * they match cannot be known until the moment of matching as they are
1544 * dependent on the current locale, 'AND"ing them likely will reduce or
1546 * But we can do better if we know that C1,P1 are in their initial state (a
1547 * frequent occurrence), each matching everything:
1548 * (<everything>) & (C2 | P2) = C2 | P2
1549 * Similarly, if C2,P2 are in their initial state (again a frequent
1550 * occurrence), the result is a no-op
1551 * (C1 | P1) & (<everything>) = C1 | P1
1554 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1555 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1556 * <= (C1 & ~C2) | (P1 & ~P2)
1559 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1560 && ! is_ANYOF_SYNTHETIC(and_with))
1564 ssc_intersection(ssc,
1566 FALSE /* Has already been inverted */
1569 /* If either P1 or P2 is empty, the intersection will be also; can skip
1571 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
1572 ANYOF_POSIXL_ZERO(ssc);
1574 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1576 /* Note that the Posix class component P from 'and_with' actually
1578 * P = Pa | Pb | ... | Pn
1579 * where each component is one posix class, such as in [\w\s].
1581 * ~P = ~(Pa | Pb | ... | Pn)
1582 * = ~Pa & ~Pb & ... & ~Pn
1583 * <= ~Pa | ~Pb | ... | ~Pn
1584 * The last is something we can easily calculate, but unfortunately
1585 * is likely to have many false positives. We could do better
1586 * in some (but certainly not all) instances if two classes in
1587 * P have known relationships. For example
1588 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1590 * :lower: & :print: = :lower:
1591 * And similarly for classes that must be disjoint. For example,
1592 * since \s and \w can have no elements in common based on rules in
1593 * the POSIX standard,
1594 * \w & ^\S = nothing
1595 * Unfortunately, some vendor locales do not meet the Posix
1596 * standard, in particular almost everything by Microsoft.
1597 * The loop below just changes e.g., \w into \W and vice versa */
1599 regnode_charclass_posixl temp;
1600 int add = 1; /* To calculate the index of the complement */
1602 ANYOF_POSIXL_ZERO(&temp);
1603 for (i = 0; i < ANYOF_MAX; i++) {
1605 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1606 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1608 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1609 ANYOF_POSIXL_SET(&temp, i + add);
1611 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1613 ANYOF_POSIXL_AND(&temp, ssc);
1615 } /* else ssc already has no posixes */
1616 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1617 in its initial state */
1618 else if (! is_ANYOF_SYNTHETIC(and_with)
1619 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1621 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1622 * copy it over 'ssc' */
1623 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1624 if (is_ANYOF_SYNTHETIC(and_with)) {
1625 StructCopy(and_with, ssc, regnode_ssc);
1628 ssc->invlist = anded_cp_list;
1629 ANYOF_POSIXL_ZERO(ssc);
1630 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1631 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1635 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1636 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
1638 /* One or the other of P1, P2 is non-empty. */
1639 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1640 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1642 ssc_union(ssc, anded_cp_list, FALSE);
1644 else { /* P1 = P2 = empty */
1645 ssc_intersection(ssc, anded_cp_list, FALSE);
1651 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1652 const regnode_charclass *or_with)
1654 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1655 * another SSC or a regular ANYOF class. Can create false positives if
1656 * 'or_with' is to be inverted. */
1661 PERL_ARGS_ASSERT_SSC_OR;
1663 assert(is_ANYOF_SYNTHETIC(ssc));
1665 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1666 * the code point inversion list and just the relevant flags */
1667 if (is_ANYOF_SYNTHETIC(or_with)) {
1668 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1669 ored_flags = ANYOF_FLAGS(or_with);
1672 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1673 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1674 if (OP(or_with) != ANYOFD) {
1676 |= ANYOF_FLAGS(or_with)
1677 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1678 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1679 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(or_with))) {
1681 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1686 ANYOF_FLAGS(ssc) |= ored_flags;
1688 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1689 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1690 * 'or_with' may be inverted. When not inverted, we have the simple
1691 * situation of computing:
1692 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1693 * If P1|P2 yields a situation with both a class and its complement are
1694 * set, like having both \w and \W, this matches all code points, and we
1695 * can delete these from the P component of the ssc going forward. XXX We
1696 * might be able to delete all the P components, but I (khw) am not certain
1697 * about this, and it is better to be safe.
1700 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1701 * <= (C1 | P1) | ~C2
1702 * <= (C1 | ~C2) | P1
1703 * (which results in actually simpler code than the non-inverted case)
1706 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1707 && ! is_ANYOF_SYNTHETIC(or_with))
1709 /* We ignore P2, leaving P1 going forward */
1710 } /* else Not inverted */
1711 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
1712 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1713 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1715 for (i = 0; i < ANYOF_MAX; i += 2) {
1716 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1718 ssc_match_all_cp(ssc);
1719 ANYOF_POSIXL_CLEAR(ssc, i);
1720 ANYOF_POSIXL_CLEAR(ssc, i+1);
1728 FALSE /* Already has been inverted */
1732 PERL_STATIC_INLINE void
1733 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1735 PERL_ARGS_ASSERT_SSC_UNION;
1737 assert(is_ANYOF_SYNTHETIC(ssc));
1739 _invlist_union_maybe_complement_2nd(ssc->invlist,
1745 PERL_STATIC_INLINE void
1746 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1748 const bool invert2nd)
1750 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1752 assert(is_ANYOF_SYNTHETIC(ssc));
1754 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1760 PERL_STATIC_INLINE void
1761 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1763 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1765 assert(is_ANYOF_SYNTHETIC(ssc));
1767 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1770 PERL_STATIC_INLINE void
1771 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1773 /* AND just the single code point 'cp' into the SSC 'ssc' */
1775 SV* cp_list = _new_invlist(2);
1777 PERL_ARGS_ASSERT_SSC_CP_AND;
1779 assert(is_ANYOF_SYNTHETIC(ssc));
1781 cp_list = add_cp_to_invlist(cp_list, cp);
1782 ssc_intersection(ssc, cp_list,
1783 FALSE /* Not inverted */
1785 SvREFCNT_dec_NN(cp_list);
1788 PERL_STATIC_INLINE void
1789 S_ssc_clear_locale(regnode_ssc *ssc)
1791 /* Set the SSC 'ssc' to not match any locale things */
1792 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1794 assert(is_ANYOF_SYNTHETIC(ssc));
1796 ANYOF_POSIXL_ZERO(ssc);
1797 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1800 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1803 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1805 /* The synthetic start class is used to hopefully quickly winnow down
1806 * places where a pattern could start a match in the target string. If it
1807 * doesn't really narrow things down that much, there isn't much point to
1808 * having the overhead of using it. This function uses some very crude
1809 * heuristics to decide if to use the ssc or not.
1811 * It returns TRUE if 'ssc' rules out more than half what it considers to
1812 * be the "likely" possible matches, but of course it doesn't know what the
1813 * actual things being matched are going to be; these are only guesses
1815 * For /l matches, it assumes that the only likely matches are going to be
1816 * in the 0-255 range, uniformly distributed, so half of that is 127
1817 * For /a and /d matches, it assumes that the likely matches will be just
1818 * the ASCII range, so half of that is 63
1819 * For /u and there isn't anything matching above the Latin1 range, it
1820 * assumes that that is the only range likely to be matched, and uses
1821 * half that as the cut-off: 127. If anything matches above Latin1,
1822 * it assumes that all of Unicode could match (uniformly), except for
1823 * non-Unicode code points and things in the General Category "Other"
1824 * (unassigned, private use, surrogates, controls and formats). This
1825 * is a much large number. */
1827 U32 count = 0; /* Running total of number of code points matched by
1829 UV start, end; /* Start and end points of current range in inversion
1831 const U32 max_code_points = (LOC)
1833 : (( ! UNI_SEMANTICS
1834 || invlist_highest(ssc->invlist) < 256)
1837 const U32 max_match = max_code_points / 2;
1839 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1841 invlist_iterinit(ssc->invlist);
1842 while (invlist_iternext(ssc->invlist, &start, &end)) {
1843 if (start >= max_code_points) {
1846 end = MIN(end, max_code_points - 1);
1847 count += end - start + 1;
1848 if (count >= max_match) {
1849 invlist_iterfinish(ssc->invlist);
1859 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1861 /* The inversion list in the SSC is marked mortal; now we need a more
1862 * permanent copy, which is stored the same way that is done in a regular
1863 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
1866 SV* invlist = invlist_clone(ssc->invlist);
1868 PERL_ARGS_ASSERT_SSC_FINALIZE;
1870 assert(is_ANYOF_SYNTHETIC(ssc));
1872 /* The code in this file assumes that all but these flags aren't relevant
1873 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
1874 * by the time we reach here */
1875 assert(! (ANYOF_FLAGS(ssc)
1876 & ~( ANYOF_COMMON_FLAGS
1877 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1878 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
1880 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1882 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1883 NULL, NULL, NULL, FALSE);
1885 /* Make sure is clone-safe */
1886 ssc->invlist = NULL;
1888 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1889 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
1892 if (RExC_contains_locale) {
1896 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
1899 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1900 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1901 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1902 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1903 ? (TRIE_LIST_CUR( idx ) - 1) \
1909 dump_trie(trie,widecharmap,revcharmap)
1910 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1911 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1913 These routines dump out a trie in a somewhat readable format.
1914 The _interim_ variants are used for debugging the interim
1915 tables that are used to generate the final compressed
1916 representation which is what dump_trie expects.
1918 Part of the reason for their existence is to provide a form
1919 of documentation as to how the different representations function.
1924 Dumps the final compressed table form of the trie to Perl_debug_log.
1925 Used for debugging make_trie().
1929 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1930 AV *revcharmap, U32 depth)
1933 SV *sv=sv_newmortal();
1934 int colwidth= widecharmap ? 6 : 4;
1936 GET_RE_DEBUG_FLAGS_DECL;
1938 PERL_ARGS_ASSERT_DUMP_TRIE;
1940 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1941 (int)depth * 2 + 2,"",
1942 "Match","Base","Ofs" );
1944 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1945 SV ** const tmp = av_fetch( revcharmap, state, 0);
1947 PerlIO_printf( Perl_debug_log, "%*s",
1949 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1950 PL_colors[0], PL_colors[1],
1951 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1952 PERL_PV_ESCAPE_FIRSTCHAR
1957 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1958 (int)depth * 2 + 2,"");
1960 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1961 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1962 PerlIO_printf( Perl_debug_log, "\n");
1964 for( state = 1 ; state < trie->statecount ; state++ ) {
1965 const U32 base = trie->states[ state ].trans.base;
1967 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|",
1968 (int)depth * 2 + 2,"", (UV)state);
1970 if ( trie->states[ state ].wordnum ) {
1971 PerlIO_printf( Perl_debug_log, " W%4X",
1972 trie->states[ state ].wordnum );
1974 PerlIO_printf( Perl_debug_log, "%6s", "" );
1977 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1982 while( ( base + ofs < trie->uniquecharcount ) ||
1983 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1984 && trie->trans[ base + ofs - trie->uniquecharcount ].check
1988 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1990 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1991 if ( ( base + ofs >= trie->uniquecharcount )
1992 && ( base + ofs - trie->uniquecharcount
1994 && trie->trans[ base + ofs
1995 - trie->uniquecharcount ].check == state )
1997 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1999 (UV)trie->trans[ base + ofs
2000 - trie->uniquecharcount ].next );
2002 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
2006 PerlIO_printf( Perl_debug_log, "]");
2009 PerlIO_printf( Perl_debug_log, "\n" );
2011 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=",
2013 for (word=1; word <= trie->wordcount; word++) {
2014 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
2015 (int)word, (int)(trie->wordinfo[word].prev),
2016 (int)(trie->wordinfo[word].len));
2018 PerlIO_printf(Perl_debug_log, "\n" );
2021 Dumps a fully constructed but uncompressed trie in list form.
2022 List tries normally only are used for construction when the number of
2023 possible chars (trie->uniquecharcount) is very high.
2024 Used for debugging make_trie().
2027 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2028 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2032 SV *sv=sv_newmortal();
2033 int colwidth= widecharmap ? 6 : 4;
2034 GET_RE_DEBUG_FLAGS_DECL;
2036 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2038 /* print out the table precompression. */
2039 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
2040 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
2041 "------:-----+-----------------\n" );
2043 for( state=1 ; state < next_alloc ; state ++ ) {
2046 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
2047 (int)depth * 2 + 2,"", (UV)state );
2048 if ( ! trie->states[ state ].wordnum ) {
2049 PerlIO_printf( Perl_debug_log, "%5s| ","");
2051 PerlIO_printf( Perl_debug_log, "W%4x| ",
2052 trie->states[ state ].wordnum
2055 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2056 SV ** const tmp = av_fetch( revcharmap,
2057 TRIE_LIST_ITEM(state,charid).forid, 0);
2059 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
2061 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2063 PL_colors[0], PL_colors[1],
2064 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2065 | PERL_PV_ESCAPE_FIRSTCHAR
2067 TRIE_LIST_ITEM(state,charid).forid,
2068 (UV)TRIE_LIST_ITEM(state,charid).newstate
2071 PerlIO_printf(Perl_debug_log, "\n%*s| ",
2072 (int)((depth * 2) + 14), "");
2075 PerlIO_printf( Perl_debug_log, "\n");
2080 Dumps a fully constructed but uncompressed trie in table form.
2081 This is the normal DFA style state transition table, with a few
2082 twists to facilitate compression later.
2083 Used for debugging make_trie().
2086 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2087 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2092 SV *sv=sv_newmortal();
2093 int colwidth= widecharmap ? 6 : 4;
2094 GET_RE_DEBUG_FLAGS_DECL;
2096 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2099 print out the table precompression so that we can do a visual check
2100 that they are identical.
2103 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
2105 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2106 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2108 PerlIO_printf( Perl_debug_log, "%*s",
2110 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2111 PL_colors[0], PL_colors[1],
2112 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2113 PERL_PV_ESCAPE_FIRSTCHAR
2119 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
2121 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2122 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
2125 PerlIO_printf( Perl_debug_log, "\n" );
2127 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2129 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
2130 (int)depth * 2 + 2,"",
2131 (UV)TRIE_NODENUM( state ) );
2133 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2134 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2136 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
2138 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
2140 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2141 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n",
2142 (UV)trie->trans[ state ].check );
2144 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n",
2145 (UV)trie->trans[ state ].check,
2146 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2154 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2155 startbranch: the first branch in the whole branch sequence
2156 first : start branch of sequence of branch-exact nodes.
2157 May be the same as startbranch
2158 last : Thing following the last branch.
2159 May be the same as tail.
2160 tail : item following the branch sequence
2161 count : words in the sequence
2162 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2163 depth : indent depth
2165 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2167 A trie is an N'ary tree where the branches are determined by digital
2168 decomposition of the key. IE, at the root node you look up the 1st character and
2169 follow that branch repeat until you find the end of the branches. Nodes can be
2170 marked as "accepting" meaning they represent a complete word. Eg:
2174 would convert into the following structure. Numbers represent states, letters
2175 following numbers represent valid transitions on the letter from that state, if
2176 the number is in square brackets it represents an accepting state, otherwise it
2177 will be in parenthesis.
2179 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2183 (1) +-i->(6)-+-s->[7]
2185 +-s->(3)-+-h->(4)-+-e->[5]
2187 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2189 This shows that when matching against the string 'hers' we will begin at state 1
2190 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2191 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2192 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2193 single traverse. We store a mapping from accepting to state to which word was
2194 matched, and then when we have multiple possibilities we try to complete the
2195 rest of the regex in the order in which they occurred in the alternation.
2197 The only prior NFA like behaviour that would be changed by the TRIE support is
2198 the silent ignoring of duplicate alternations which are of the form:
2200 / (DUPE|DUPE) X? (?{ ... }) Y /x
2202 Thus EVAL blocks following a trie may be called a different number of times with
2203 and without the optimisation. With the optimisations dupes will be silently
2204 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2205 the following demonstrates:
2207 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2209 which prints out 'word' three times, but
2211 'words'=~/(word|word|word)(?{ print $1 })S/
2213 which doesnt print it out at all. This is due to other optimisations kicking in.
2215 Example of what happens on a structural level:
2217 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2219 1: CURLYM[1] {1,32767}(18)
2230 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2231 and should turn into:
2233 1: CURLYM[1] {1,32767}(18)
2235 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2243 Cases where tail != last would be like /(?foo|bar)baz/:
2253 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2254 and would end up looking like:
2257 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2264 d = uvchr_to_utf8_flags(d, uv, 0);
2266 is the recommended Unicode-aware way of saying
2271 #define TRIE_STORE_REVCHAR(val) \
2274 SV *zlopp = newSV(UTF8_MAXBYTES); \
2275 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2276 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2277 SvCUR_set(zlopp, kapow - flrbbbbb); \
2280 av_push(revcharmap, zlopp); \
2282 char ooooff = (char)val; \
2283 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2287 /* This gets the next character from the input, folding it if not already
2289 #define TRIE_READ_CHAR STMT_START { \
2292 /* if it is UTF then it is either already folded, or does not need \
2294 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2296 else if (folder == PL_fold_latin1) { \
2297 /* This folder implies Unicode rules, which in the range expressible \
2298 * by not UTF is the lower case, with the two exceptions, one of \
2299 * which should have been taken care of before calling this */ \
2300 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2301 uvc = toLOWER_L1(*uc); \
2302 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2305 /* raw data, will be folded later if needed */ \
2313 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2314 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2315 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
2316 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2318 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2319 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2320 TRIE_LIST_CUR( state )++; \
2323 #define TRIE_LIST_NEW(state) STMT_START { \
2324 Newxz( trie->states[ state ].trans.list, \
2325 4, reg_trie_trans_le ); \
2326 TRIE_LIST_CUR( state ) = 1; \
2327 TRIE_LIST_LEN( state ) = 4; \
2330 #define TRIE_HANDLE_WORD(state) STMT_START { \
2331 U16 dupe= trie->states[ state ].wordnum; \
2332 regnode * const noper_next = regnext( noper ); \
2335 /* store the word for dumping */ \
2337 if (OP(noper) != NOTHING) \
2338 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2340 tmp = newSVpvn_utf8( "", 0, UTF ); \
2341 av_push( trie_words, tmp ); \
2345 trie->wordinfo[curword].prev = 0; \
2346 trie->wordinfo[curword].len = wordlen; \
2347 trie->wordinfo[curword].accept = state; \
2349 if ( noper_next < tail ) { \
2351 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2353 trie->jump[curword] = (U16)(noper_next - convert); \
2355 jumper = noper_next; \
2357 nextbranch= regnext(cur); \
2361 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2362 /* chain, so that when the bits of chain are later */\
2363 /* linked together, the dups appear in the chain */\
2364 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2365 trie->wordinfo[dupe].prev = curword; \
2367 /* we haven't inserted this word yet. */ \
2368 trie->states[ state ].wordnum = curword; \
2373 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2374 ( ( base + charid >= ucharcount \
2375 && base + charid < ubound \
2376 && state == trie->trans[ base - ucharcount + charid ].check \
2377 && trie->trans[ base - ucharcount + charid ].next ) \
2378 ? trie->trans[ base - ucharcount + charid ].next \
2379 : ( state==1 ? special : 0 ) \
2383 #define MADE_JUMP_TRIE 2
2384 #define MADE_EXACT_TRIE 4
2387 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2388 regnode *first, regnode *last, regnode *tail,
2389 U32 word_count, U32 flags, U32 depth)
2391 /* first pass, loop through and scan words */
2392 reg_trie_data *trie;
2393 HV *widecharmap = NULL;
2394 AV *revcharmap = newAV();
2400 regnode *jumper = NULL;
2401 regnode *nextbranch = NULL;
2402 regnode *convert = NULL;
2403 U32 *prev_states; /* temp array mapping each state to previous one */
2404 /* we just use folder as a flag in utf8 */
2405 const U8 * folder = NULL;
2408 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
2409 AV *trie_words = NULL;
2410 /* along with revcharmap, this only used during construction but both are
2411 * useful during debugging so we store them in the struct when debugging.
2414 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2415 STRLEN trie_charcount=0;
2417 SV *re_trie_maxbuff;
2418 GET_RE_DEBUG_FLAGS_DECL;
2420 PERL_ARGS_ASSERT_MAKE_TRIE;
2422 PERL_UNUSED_ARG(depth);
2426 case EXACT: case EXACTL: break;
2430 case EXACTFLU8: folder = PL_fold_latin1; break;
2431 case EXACTF: folder = PL_fold; break;
2432 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2435 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2437 trie->startstate = 1;
2438 trie->wordcount = word_count;
2439 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2440 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2441 if (flags == EXACT || flags == EXACTL)
2442 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2443 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2444 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2447 trie_words = newAV();
2450 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2451 assert(re_trie_maxbuff);
2452 if (!SvIOK(re_trie_maxbuff)) {
2453 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2455 DEBUG_TRIE_COMPILE_r({
2456 PerlIO_printf( Perl_debug_log,
2457 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2458 (int)depth * 2 + 2, "",
2459 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2460 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2463 /* Find the node we are going to overwrite */
2464 if ( first == startbranch && OP( last ) != BRANCH ) {
2465 /* whole branch chain */
2468 /* branch sub-chain */
2469 convert = NEXTOPER( first );
2472 /* -- First loop and Setup --
2474 We first traverse the branches and scan each word to determine if it
2475 contains widechars, and how many unique chars there are, this is
2476 important as we have to build a table with at least as many columns as we
2479 We use an array of integers to represent the character codes 0..255
2480 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2481 the native representation of the character value as the key and IV's for
2484 *TODO* If we keep track of how many times each character is used we can
2485 remap the columns so that the table compression later on is more
2486 efficient in terms of memory by ensuring the most common value is in the
2487 middle and the least common are on the outside. IMO this would be better
2488 than a most to least common mapping as theres a decent chance the most
2489 common letter will share a node with the least common, meaning the node
2490 will not be compressible. With a middle is most common approach the worst
2491 case is when we have the least common nodes twice.
2495 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2496 regnode *noper = NEXTOPER( cur );
2497 const U8 *uc = (U8*)STRING( noper );
2498 const U8 *e = uc + STR_LEN( noper );
2500 U32 wordlen = 0; /* required init */
2501 STRLEN minchars = 0;
2502 STRLEN maxchars = 0;
2503 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2506 if (OP(noper) == NOTHING) {
2507 regnode *noper_next= regnext(noper);
2508 if (noper_next != tail && OP(noper_next) == flags) {
2510 uc= (U8*)STRING(noper);
2511 e= uc + STR_LEN(noper);
2512 trie->minlen= STR_LEN(noper);
2519 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2520 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2521 regardless of encoding */
2522 if (OP( noper ) == EXACTFU_SS) {
2523 /* false positives are ok, so just set this */
2524 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2527 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2529 TRIE_CHARCOUNT(trie)++;
2532 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2533 * is in effect. Under /i, this character can match itself, or
2534 * anything that folds to it. If not under /i, it can match just
2535 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2536 * all fold to k, and all are single characters. But some folds
2537 * expand to more than one character, so for example LATIN SMALL
2538 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2539 * the string beginning at 'uc' is 'ffi', it could be matched by
2540 * three characters, or just by the one ligature character. (It
2541 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2542 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2543 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2544 * match.) The trie needs to know the minimum and maximum number
2545 * of characters that could match so that it can use size alone to
2546 * quickly reject many match attempts. The max is simple: it is
2547 * the number of folded characters in this branch (since a fold is
2548 * never shorter than what folds to it. */
2552 /* And the min is equal to the max if not under /i (indicated by
2553 * 'folder' being NULL), or there are no multi-character folds. If
2554 * there is a multi-character fold, the min is incremented just
2555 * once, for the character that folds to the sequence. Each
2556 * character in the sequence needs to be added to the list below of
2557 * characters in the trie, but we count only the first towards the
2558 * min number of characters needed. This is done through the
2559 * variable 'foldlen', which is returned by the macros that look
2560 * for these sequences as the number of bytes the sequence
2561 * occupies. Each time through the loop, we decrement 'foldlen' by
2562 * how many bytes the current char occupies. Only when it reaches
2563 * 0 do we increment 'minchars' or look for another multi-character
2565 if (folder == NULL) {
2568 else if (foldlen > 0) {
2569 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2574 /* See if *uc is the beginning of a multi-character fold. If
2575 * so, we decrement the length remaining to look at, to account
2576 * for the current character this iteration. (We can use 'uc'
2577 * instead of the fold returned by TRIE_READ_CHAR because for
2578 * non-UTF, the latin1_safe macro is smart enough to account
2579 * for all the unfolded characters, and because for UTF, the
2580 * string will already have been folded earlier in the
2581 * compilation process */
2583 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2584 foldlen -= UTF8SKIP(uc);
2587 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2592 /* The current character (and any potential folds) should be added
2593 * to the possible matching characters for this position in this
2597 U8 folded= folder[ (U8) uvc ];
2598 if ( !trie->charmap[ folded ] ) {
2599 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2600 TRIE_STORE_REVCHAR( folded );
2603 if ( !trie->charmap[ uvc ] ) {
2604 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2605 TRIE_STORE_REVCHAR( uvc );
2608 /* store the codepoint in the bitmap, and its folded
2610 TRIE_BITMAP_SET(trie, uvc);
2612 /* store the folded codepoint */
2613 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
2616 /* store first byte of utf8 representation of
2617 variant codepoints */
2618 if (! UVCHR_IS_INVARIANT(uvc)) {
2619 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
2622 set_bit = 0; /* We've done our bit :-) */
2626 /* XXX We could come up with the list of code points that fold
2627 * to this using PL_utf8_foldclosures, except not for
2628 * multi-char folds, as there may be multiple combinations
2629 * there that could work, which needs to wait until runtime to
2630 * resolve (The comment about LIGATURE FFI above is such an
2635 widecharmap = newHV();
2637 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2640 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
2642 if ( !SvTRUE( *svpp ) ) {
2643 sv_setiv( *svpp, ++trie->uniquecharcount );
2644 TRIE_STORE_REVCHAR(uvc);
2647 } /* end loop through characters in this branch of the trie */
2649 /* We take the min and max for this branch and combine to find the min
2650 * and max for all branches processed so far */
2651 if( cur == first ) {
2652 trie->minlen = minchars;
2653 trie->maxlen = maxchars;
2654 } else if (minchars < trie->minlen) {
2655 trie->minlen = minchars;
2656 } else if (maxchars > trie->maxlen) {
2657 trie->maxlen = maxchars;
2659 } /* end first pass */
2660 DEBUG_TRIE_COMPILE_r(
2661 PerlIO_printf( Perl_debug_log,
2662 "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2663 (int)depth * 2 + 2,"",
2664 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2665 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2666 (int)trie->minlen, (int)trie->maxlen )
2670 We now know what we are dealing with in terms of unique chars and
2671 string sizes so we can calculate how much memory a naive
2672 representation using a flat table will take. If it's over a reasonable
2673 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2674 conservative but potentially much slower representation using an array
2677 At the end we convert both representations into the same compressed
2678 form that will be used in regexec.c for matching with. The latter
2679 is a form that cannot be used to construct with but has memory
2680 properties similar to the list form and access properties similar
2681 to the table form making it both suitable for fast searches and
2682 small enough that its feasable to store for the duration of a program.
2684 See the comment in the code where the compressed table is produced
2685 inplace from the flat tabe representation for an explanation of how
2686 the compression works.
2691 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2694 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2695 > SvIV(re_trie_maxbuff) )
2698 Second Pass -- Array Of Lists Representation
2700 Each state will be represented by a list of charid:state records
2701 (reg_trie_trans_le) the first such element holds the CUR and LEN
2702 points of the allocated array. (See defines above).
2704 We build the initial structure using the lists, and then convert
2705 it into the compressed table form which allows faster lookups
2706 (but cant be modified once converted).
2709 STRLEN transcount = 1;
2711 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2712 "%*sCompiling trie using list compiler\n",
2713 (int)depth * 2 + 2, ""));
2715 trie->states = (reg_trie_state *)
2716 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2717 sizeof(reg_trie_state) );
2721 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2723 regnode *noper = NEXTOPER( cur );
2724 U8 *uc = (U8*)STRING( noper );
2725 const U8 *e = uc + STR_LEN( noper );
2726 U32 state = 1; /* required init */
2727 U16 charid = 0; /* sanity init */
2728 U32 wordlen = 0; /* required init */
2730 if (OP(noper) == NOTHING) {
2731 regnode *noper_next= regnext(noper);
2732 if (noper_next != tail && OP(noper_next) == flags) {
2734 uc= (U8*)STRING(noper);
2735 e= uc + STR_LEN(noper);
2739 if (OP(noper) != NOTHING) {
2740 for ( ; uc < e ; uc += len ) {
2745 charid = trie->charmap[ uvc ];
2747 SV** const svpp = hv_fetch( widecharmap,
2754 charid=(U16)SvIV( *svpp );
2757 /* charid is now 0 if we dont know the char read, or
2758 * nonzero if we do */
2765 if ( !trie->states[ state ].trans.list ) {
2766 TRIE_LIST_NEW( state );
2769 check <= TRIE_LIST_USED( state );
2772 if ( TRIE_LIST_ITEM( state, check ).forid
2775 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2780 newstate = next_alloc++;
2781 prev_states[newstate] = state;
2782 TRIE_LIST_PUSH( state, charid, newstate );
2787 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2791 TRIE_HANDLE_WORD(state);
2793 } /* end second pass */
2795 /* next alloc is the NEXT state to be allocated */
2796 trie->statecount = next_alloc;
2797 trie->states = (reg_trie_state *)
2798 PerlMemShared_realloc( trie->states,
2800 * sizeof(reg_trie_state) );
2802 /* and now dump it out before we compress it */
2803 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2804 revcharmap, next_alloc,
2808 trie->trans = (reg_trie_trans *)
2809 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2816 for( state=1 ; state < next_alloc ; state ++ ) {
2820 DEBUG_TRIE_COMPILE_MORE_r(
2821 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
2825 if (trie->states[state].trans.list) {
2826 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2830 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2831 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2832 if ( forid < minid ) {
2834 } else if ( forid > maxid ) {
2838 if ( transcount < tp + maxid - minid + 1) {
2840 trie->trans = (reg_trie_trans *)
2841 PerlMemShared_realloc( trie->trans,
2843 * sizeof(reg_trie_trans) );
2844 Zero( trie->trans + (transcount / 2),
2848 base = trie->uniquecharcount + tp - minid;
2849 if ( maxid == minid ) {
2851 for ( ; zp < tp ; zp++ ) {
2852 if ( ! trie->trans[ zp ].next ) {
2853 base = trie->uniquecharcount + zp - minid;
2854 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2856 trie->trans[ zp ].check = state;
2862 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2864 trie->trans[ tp ].check = state;
2869 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2870 const U32 tid = base
2871 - trie->uniquecharcount
2872 + TRIE_LIST_ITEM( state, idx ).forid;
2873 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2875 trie->trans[ tid ].check = state;
2877 tp += ( maxid - minid + 1 );
2879 Safefree(trie->states[ state ].trans.list);
2882 DEBUG_TRIE_COMPILE_MORE_r(
2883 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
2886 trie->states[ state ].trans.base=base;
2888 trie->lasttrans = tp + 1;
2892 Second Pass -- Flat Table Representation.
2894 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2895 each. We know that we will need Charcount+1 trans at most to store
2896 the data (one row per char at worst case) So we preallocate both
2897 structures assuming worst case.
2899 We then construct the trie using only the .next slots of the entry
2902 We use the .check field of the first entry of the node temporarily
2903 to make compression both faster and easier by keeping track of how
2904 many non zero fields are in the node.
2906 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2909 There are two terms at use here: state as a TRIE_NODEIDX() which is
2910 a number representing the first entry of the node, and state as a
2911 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2912 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2913 if there are 2 entrys per node. eg:
2921 The table is internally in the right hand, idx form. However as we
2922 also have to deal with the states array which is indexed by nodenum
2923 we have to use TRIE_NODENUM() to convert.
2926 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2927 "%*sCompiling trie using table compiler\n",
2928 (int)depth * 2 + 2, ""));
2930 trie->trans = (reg_trie_trans *)
2931 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2932 * trie->uniquecharcount + 1,
2933 sizeof(reg_trie_trans) );
2934 trie->states = (reg_trie_state *)
2935 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2936 sizeof(reg_trie_state) );
2937 next_alloc = trie->uniquecharcount + 1;
2940 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2942 regnode *noper = NEXTOPER( cur );
2943 const U8 *uc = (U8*)STRING( noper );
2944 const U8 *e = uc + STR_LEN( noper );
2946 U32 state = 1; /* required init */
2948 U16 charid = 0; /* sanity init */
2949 U32 accept_state = 0; /* sanity init */
2951 U32 wordlen = 0; /* required init */
2953 if (OP(noper) == NOTHING) {
2954 regnode *noper_next= regnext(noper);
2955 if (noper_next != tail && OP(noper_next) == flags) {
2957 uc= (U8*)STRING(noper);
2958 e= uc + STR_LEN(noper);
2962 if ( OP(noper) != NOTHING ) {
2963 for ( ; uc < e ; uc += len ) {
2968 charid = trie->charmap[ uvc ];
2970 SV* const * const svpp = hv_fetch( widecharmap,
2974 charid = svpp ? (U16)SvIV(*svpp) : 0;
2978 if ( !trie->trans[ state + charid ].next ) {
2979 trie->trans[ state + charid ].next = next_alloc;
2980 trie->trans[ state ].check++;
2981 prev_states[TRIE_NODENUM(next_alloc)]
2982 = TRIE_NODENUM(state);
2983 next_alloc += trie->uniquecharcount;
2985 state = trie->trans[ state + charid ].next;
2987 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2989 /* charid is now 0 if we dont know the char read, or
2990 * nonzero if we do */
2993 accept_state = TRIE_NODENUM( state );
2994 TRIE_HANDLE_WORD(accept_state);
2996 } /* end second pass */
2998 /* and now dump it out before we compress it */
2999 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3001 next_alloc, depth+1));
3005 * Inplace compress the table.*
3007 For sparse data sets the table constructed by the trie algorithm will
3008 be mostly 0/FAIL transitions or to put it another way mostly empty.
3009 (Note that leaf nodes will not contain any transitions.)
3011 This algorithm compresses the tables by eliminating most such
3012 transitions, at the cost of a modest bit of extra work during lookup:
3014 - Each states[] entry contains a .base field which indicates the
3015 index in the state[] array wheres its transition data is stored.
3017 - If .base is 0 there are no valid transitions from that node.
3019 - If .base is nonzero then charid is added to it to find an entry in
3022 -If trans[states[state].base+charid].check!=state then the
3023 transition is taken to be a 0/Fail transition. Thus if there are fail
3024 transitions at the front of the node then the .base offset will point
3025 somewhere inside the previous nodes data (or maybe even into a node
3026 even earlier), but the .check field determines if the transition is
3030 The following process inplace converts the table to the compressed
3031 table: We first do not compress the root node 1,and mark all its
3032 .check pointers as 1 and set its .base pointer as 1 as well. This
3033 allows us to do a DFA construction from the compressed table later,
3034 and ensures that any .base pointers we calculate later are greater
3037 - We set 'pos' to indicate the first entry of the second node.
3039 - We then iterate over the columns of the node, finding the first and
3040 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3041 and set the .check pointers accordingly, and advance pos
3042 appropriately and repreat for the next node. Note that when we copy
3043 the next pointers we have to convert them from the original
3044 NODEIDX form to NODENUM form as the former is not valid post
3047 - If a node has no transitions used we mark its base as 0 and do not
3048 advance the pos pointer.
3050 - If a node only has one transition we use a second pointer into the
3051 structure to fill in allocated fail transitions from other states.
3052 This pointer is independent of the main pointer and scans forward
3053 looking for null transitions that are allocated to a state. When it
3054 finds one it writes the single transition into the "hole". If the
3055 pointer doesnt find one the single transition is appended as normal.
3057 - Once compressed we can Renew/realloc the structures to release the
3060 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3061 specifically Fig 3.47 and the associated pseudocode.
3065 const U32 laststate = TRIE_NODENUM( next_alloc );
3068 trie->statecount = laststate;
3070 for ( state = 1 ; state < laststate ; state++ ) {
3072 const U32 stateidx = TRIE_NODEIDX( state );
3073 const U32 o_used = trie->trans[ stateidx ].check;
3074 U32 used = trie->trans[ stateidx ].check;
3075 trie->trans[ stateidx ].check = 0;
3078 used && charid < trie->uniquecharcount;
3081 if ( flag || trie->trans[ stateidx + charid ].next ) {
3082 if ( trie->trans[ stateidx + charid ].next ) {
3084 for ( ; zp < pos ; zp++ ) {
3085 if ( ! trie->trans[ zp ].next ) {
3089 trie->states[ state ].trans.base
3091 + trie->uniquecharcount
3093 trie->trans[ zp ].next
3094 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3096 trie->trans[ zp ].check = state;
3097 if ( ++zp > pos ) pos = zp;
3104 trie->states[ state ].trans.base
3105 = pos + trie->uniquecharcount - charid ;
3107 trie->trans[ pos ].next
3108 = SAFE_TRIE_NODENUM(
3109 trie->trans[ stateidx + charid ].next );
3110 trie->trans[ pos ].check = state;
3115 trie->lasttrans = pos + 1;
3116 trie->states = (reg_trie_state *)
3117 PerlMemShared_realloc( trie->states, laststate
3118 * sizeof(reg_trie_state) );
3119 DEBUG_TRIE_COMPILE_MORE_r(
3120 PerlIO_printf( Perl_debug_log,
3121 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
3122 (int)depth * 2 + 2,"",
3123 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3127 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3130 } /* end table compress */
3132 DEBUG_TRIE_COMPILE_MORE_r(
3133 PerlIO_printf(Perl_debug_log,
3134 "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
3135 (int)depth * 2 + 2, "",
3136 (UV)trie->statecount,
3137 (UV)trie->lasttrans)
3139 /* resize the trans array to remove unused space */
3140 trie->trans = (reg_trie_trans *)
3141 PerlMemShared_realloc( trie->trans, trie->lasttrans
3142 * sizeof(reg_trie_trans) );
3144 { /* Modify the program and insert the new TRIE node */
3145 U8 nodetype =(U8)(flags & 0xFF);
3149 regnode *optimize = NULL;
3150 #ifdef RE_TRACK_PATTERN_OFFSETS
3153 U32 mjd_nodelen = 0;
3154 #endif /* RE_TRACK_PATTERN_OFFSETS */
3155 #endif /* DEBUGGING */
3157 This means we convert either the first branch or the first Exact,
3158 depending on whether the thing following (in 'last') is a branch
3159 or not and whther first is the startbranch (ie is it a sub part of
3160 the alternation or is it the whole thing.)
3161 Assuming its a sub part we convert the EXACT otherwise we convert
3162 the whole branch sequence, including the first.
3164 /* Find the node we are going to overwrite */
3165 if ( first != startbranch || OP( last ) == BRANCH ) {
3166 /* branch sub-chain */
3167 NEXT_OFF( first ) = (U16)(last - first);
3168 #ifdef RE_TRACK_PATTERN_OFFSETS
3170 mjd_offset= Node_Offset((convert));
3171 mjd_nodelen= Node_Length((convert));
3174 /* whole branch chain */
3176 #ifdef RE_TRACK_PATTERN_OFFSETS
3179 const regnode *nop = NEXTOPER( convert );
3180 mjd_offset= Node_Offset((nop));
3181 mjd_nodelen= Node_Length((nop));
3185 PerlIO_printf(Perl_debug_log,
3186 "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
3187 (int)depth * 2 + 2, "",
3188 (UV)mjd_offset, (UV)mjd_nodelen)
3191 /* But first we check to see if there is a common prefix we can
3192 split out as an EXACT and put in front of the TRIE node. */
3193 trie->startstate= 1;
3194 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3196 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3200 const U32 base = trie->states[ state ].trans.base;
3202 if ( trie->states[state].wordnum )
3205 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3206 if ( ( base + ofs >= trie->uniquecharcount ) &&
3207 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3208 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3210 if ( ++count > 1 ) {
3211 SV **tmp = av_fetch( revcharmap, ofs, 0);
3212 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
3213 if ( state == 1 ) break;
3215 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3217 PerlIO_printf(Perl_debug_log,
3218 "%*sNew Start State=%"UVuf" Class: [",
3219 (int)depth * 2 + 2, "",
3222 SV ** const tmp = av_fetch( revcharmap, idx, 0);
3223 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3225 TRIE_BITMAP_SET(trie,*ch);
3227 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3229 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3233 TRIE_BITMAP_SET(trie,*ch);
3235 TRIE_BITMAP_SET(trie,folder[ *ch ]);
3236 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
3242 SV **tmp = av_fetch( revcharmap, idx, 0);
3244 char *ch = SvPV( *tmp, len );
3246 SV *sv=sv_newmortal();
3247 PerlIO_printf( Perl_debug_log,
3248 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
3249 (int)depth * 2 + 2, "",
3251 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3252 PL_colors[0], PL_colors[1],
3253 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3254 PERL_PV_ESCAPE_FIRSTCHAR
3259 OP( convert ) = nodetype;
3260 str=STRING(convert);
3263 STR_LEN(convert) += len;
3269 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
3274 trie->prefixlen = (state-1);
3276 regnode *n = convert+NODE_SZ_STR(convert);
3277 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3278 trie->startstate = state;
3279 trie->minlen -= (state - 1);
3280 trie->maxlen -= (state - 1);
3282 /* At least the UNICOS C compiler choked on this
3283 * being argument to DEBUG_r(), so let's just have
3286 #ifdef PERL_EXT_RE_BUILD
3292 regnode *fix = convert;
3293 U32 word = trie->wordcount;
3295 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3296 while( ++fix < n ) {
3297 Set_Node_Offset_Length(fix, 0, 0);
3300 SV ** const tmp = av_fetch( trie_words, word, 0 );
3302 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3303 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3305 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3313 NEXT_OFF(convert) = (U16)(tail - convert);
3314 DEBUG_r(optimize= n);
3320 if ( trie->maxlen ) {
3321 NEXT_OFF( convert ) = (U16)(tail - convert);
3322 ARG_SET( convert, data_slot );
3323 /* Store the offset to the first unabsorbed branch in
3324 jump[0], which is otherwise unused by the jump logic.
3325 We use this when dumping a trie and during optimisation. */
3327 trie->jump[0] = (U16)(nextbranch - convert);
3329 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3330 * and there is a bitmap
3331 * and the first "jump target" node we found leaves enough room
3332 * then convert the TRIE node into a TRIEC node, with the bitmap
3333 * embedded inline in the opcode - this is hypothetically faster.
3335 if ( !trie->states[trie->startstate].wordnum
3337 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3339 OP( convert ) = TRIEC;
3340 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3341 PerlMemShared_free(trie->bitmap);
3344 OP( convert ) = TRIE;
3346 /* store the type in the flags */
3347 convert->flags = nodetype;
3351 + regarglen[ OP( convert ) ];
3353 /* XXX We really should free up the resource in trie now,
3354 as we won't use them - (which resources?) dmq */
3356 /* needed for dumping*/
3357 DEBUG_r(if (optimize) {
3358 regnode *opt = convert;
3360 while ( ++opt < optimize) {
3361 Set_Node_Offset_Length(opt,0,0);
3364 Try to clean up some of the debris left after the
3367 while( optimize < jumper ) {
3368 mjd_nodelen += Node_Length((optimize));
3369 OP( optimize ) = OPTIMIZED;
3370 Set_Node_Offset_Length(optimize,0,0);
3373 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
3375 } /* end node insert */
3377 /* Finish populating the prev field of the wordinfo array. Walk back
3378 * from each accept state until we find another accept state, and if
3379 * so, point the first word's .prev field at the second word. If the
3380 * second already has a .prev field set, stop now. This will be the
3381 * case either if we've already processed that word's accept state,
3382 * or that state had multiple words, and the overspill words were
3383 * already linked up earlier.
3390 for (word=1; word <= trie->wordcount; word++) {
3392 if (trie->wordinfo[word].prev)
3394 state = trie->wordinfo[word].accept;
3396 state = prev_states[state];
3399 prev = trie->states[state].wordnum;
3403 trie->wordinfo[word].prev = prev;
3405 Safefree(prev_states);
3409 /* and now dump out the compressed format */
3410 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3412 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3414 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3415 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3417 SvREFCNT_dec_NN(revcharmap);
3421 : trie->startstate>1
3427 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3429 /* The Trie is constructed and compressed now so we can build a fail array if
3432 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3434 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3438 We find the fail state for each state in the trie, this state is the longest
3439 proper suffix of the current state's 'word' that is also a proper prefix of
3440 another word in our trie. State 1 represents the word '' and is thus the
3441 default fail state. This allows the DFA not to have to restart after its
3442 tried and failed a word at a given point, it simply continues as though it
3443 had been matching the other word in the first place.
3445 'abcdgu'=~/abcdefg|cdgu/
3446 When we get to 'd' we are still matching the first word, we would encounter
3447 'g' which would fail, which would bring us to the state representing 'd' in
3448 the second word where we would try 'g' and succeed, proceeding to match
3451 /* add a fail transition */
3452 const U32 trie_offset = ARG(source);
3453 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3455 const U32 ucharcount = trie->uniquecharcount;
3456 const U32 numstates = trie->statecount;
3457 const U32 ubound = trie->lasttrans + ucharcount;
3461 U32 base = trie->states[ 1 ].trans.base;
3464 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3466 GET_RE_DEBUG_FLAGS_DECL;
3468 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3469 PERL_UNUSED_CONTEXT;
3471 PERL_UNUSED_ARG(depth);
3474 if ( OP(source) == TRIE ) {
3475 struct regnode_1 *op = (struct regnode_1 *)
3476 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3477 StructCopy(source,op,struct regnode_1);
3478 stclass = (regnode *)op;
3480 struct regnode_charclass *op = (struct regnode_charclass *)
3481 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3482 StructCopy(source,op,struct regnode_charclass);
3483 stclass = (regnode *)op;
3485 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3487 ARG_SET( stclass, data_slot );
3488 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3489 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3490 aho->trie=trie_offset;
3491 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3492 Copy( trie->states, aho->states, numstates, reg_trie_state );
3493 Newxz( q, numstates, U32);
3494 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3497 /* initialize fail[0..1] to be 1 so that we always have
3498 a valid final fail state */
3499 fail[ 0 ] = fail[ 1 ] = 1;
3501 for ( charid = 0; charid < ucharcount ; charid++ ) {
3502 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3504 q[ q_write ] = newstate;
3505 /* set to point at the root */
3506 fail[ q[ q_write++ ] ]=1;
3509 while ( q_read < q_write) {
3510 const U32 cur = q[ q_read++ % numstates ];
3511 base = trie->states[ cur ].trans.base;
3513 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3514 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3516 U32 fail_state = cur;
3519 fail_state = fail[ fail_state ];
3520 fail_base = aho->states[ fail_state ].trans.base;
3521 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3523 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3524 fail[ ch_state ] = fail_state;
3525 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3527 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3529 q[ q_write++ % numstates] = ch_state;
3533 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3534 when we fail in state 1, this allows us to use the
3535 charclass scan to find a valid start char. This is based on the principle
3536 that theres a good chance the string being searched contains lots of stuff
3537 that cant be a start char.
3539 fail[ 0 ] = fail[ 1 ] = 0;
3540 DEBUG_TRIE_COMPILE_r({
3541 PerlIO_printf(Perl_debug_log,
3542 "%*sStclass Failtable (%"UVuf" states): 0",
3543 (int)(depth * 2), "", (UV)numstates
3545 for( q_read=1; q_read<numstates; q_read++ ) {
3546 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
3548 PerlIO_printf(Perl_debug_log, "\n");
3551 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3556 #define DEBUG_PEEP(str,scan,depth) \
3557 DEBUG_OPTIMISE_r({if (scan){ \
3558 regnode *Next = regnext(scan); \
3559 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state); \
3560 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)", \
3561 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),\
3562 Next ? (REG_NODE_NUM(Next)) : 0 ); \
3563 DEBUG_SHOW_STUDY_FLAGS(flags," [ ","]");\
3564 PerlIO_printf(Perl_debug_log, "\n"); \
3567 /* The below joins as many adjacent EXACTish nodes as possible into a single
3568 * one. The regop may be changed if the node(s) contain certain sequences that
3569 * require special handling. The joining is only done if:
3570 * 1) there is room in the current conglomerated node to entirely contain the
3572 * 2) they are the exact same node type
3574 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3575 * these get optimized out
3577 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3578 * as possible, even if that means splitting an existing node so that its first
3579 * part is moved to the preceeding node. This would maximise the efficiency of
3580 * memEQ during matching. Elsewhere in this file, khw proposes splitting
3581 * EXACTFish nodes into portions that don't change under folding vs those that
3582 * do. Those portions that don't change may be the only things in the pattern that
3583 * could be used to find fixed and floating strings.
3585 * If a node is to match under /i (folded), the number of characters it matches
3586 * can be different than its character length if it contains a multi-character
3587 * fold. *min_subtract is set to the total delta number of characters of the
3590 * And *unfolded_multi_char is set to indicate whether or not the node contains
3591 * an unfolded multi-char fold. This happens when whether the fold is valid or
3592 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3593 * SMALL LETTER SHARP S, as only if the target string being matched against
3594 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3595 * folding rules depend on the locale in force at runtime. (Multi-char folds
3596 * whose components are all above the Latin1 range are not run-time locale
3597 * dependent, and have already been folded by the time this function is
3600 * This is as good a place as any to discuss the design of handling these
3601 * multi-character fold sequences. It's been wrong in Perl for a very long
3602 * time. There are three code points in Unicode whose multi-character folds
3603 * were long ago discovered to mess things up. The previous designs for
3604 * dealing with these involved assigning a special node for them. This
3605 * approach doesn't always work, as evidenced by this example:
3606 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3607 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3608 * would match just the \xDF, it won't be able to handle the case where a
3609 * successful match would have to cross the node's boundary. The new approach
3610 * that hopefully generally solves the problem generates an EXACTFU_SS node
3611 * that is "sss" in this case.
3613 * It turns out that there are problems with all multi-character folds, and not
3614 * just these three. Now the code is general, for all such cases. The
3615 * approach taken is:
3616 * 1) This routine examines each EXACTFish node that could contain multi-
3617 * character folded sequences. Since a single character can fold into
3618 * such a sequence, the minimum match length for this node is less than
3619 * the number of characters in the node. This routine returns in
3620 * *min_subtract how many characters to subtract from the the actual
3621 * length of the string to get a real minimum match length; it is 0 if
3622 * there are no multi-char foldeds. This delta is used by the caller to
3623 * adjust the min length of the match, and the delta between min and max,
3624 * so that the optimizer doesn't reject these possibilities based on size
3626 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3627 * is used for an EXACTFU node that contains at least one "ss" sequence in
3628 * it. For non-UTF-8 patterns and strings, this is the only case where
3629 * there is a possible fold length change. That means that a regular
3630 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3631 * with length changes, and so can be processed faster. regexec.c takes
3632 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3633 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3634 * known until runtime). This saves effort in regex matching. However,
3635 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3636 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3637 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3638 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3639 * possibilities for the non-UTF8 patterns are quite simple, except for
3640 * the sharp s. All the ones that don't involve a UTF-8 target string are
3641 * members of a fold-pair, and arrays are set up for all of them so that
3642 * the other member of the pair can be found quickly. Code elsewhere in
3643 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3644 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3645 * described in the next item.
3646 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3647 * validity of the fold won't be known until runtime, and so must remain
3648 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3649 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3650 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3651 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3652 * The reason this is a problem is that the optimizer part of regexec.c
3653 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3654 * that a character in the pattern corresponds to at most a single
3655 * character in the target string. (And I do mean character, and not byte
3656 * here, unlike other parts of the documentation that have never been
3657 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3658 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3659 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3660 * nodes, violate the assumption, and they are the only instances where it
3661 * is violated. I'm reluctant to try to change the assumption, as the
3662 * code involved is impenetrable to me (khw), so instead the code here
3663 * punts. This routine examines EXACTFL nodes, and (when the pattern
3664 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3665 * boolean indicating whether or not the node contains such a fold. When
3666 * it is true, the caller sets a flag that later causes the optimizer in
3667 * this file to not set values for the floating and fixed string lengths,
3668 * and thus avoids the optimizer code in regexec.c that makes the invalid
3669 * assumption. Thus, there is no optimization based on string lengths for
3670 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3671 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3672 * assumption is wrong only in these cases is that all other non-UTF-8
3673 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3674 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3675 * EXACTF nodes because we don't know at compile time if it actually
3676 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3677 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3678 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3679 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3680 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3681 * string would require the pattern to be forced into UTF-8, the overhead
3682 * of which we want to avoid. Similarly the unfolded multi-char folds in
3683 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3686 * Similarly, the code that generates tries doesn't currently handle
3687 * not-already-folded multi-char folds, and it looks like a pain to change
3688 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3689 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3690 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3691 * using /iaa matching will be doing so almost entirely with ASCII
3692 * strings, so this should rarely be encountered in practice */
3694 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3695 if (PL_regkind[OP(scan)] == EXACT) \
3696 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3699 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3700 UV *min_subtract, bool *unfolded_multi_char,
3701 U32 flags,regnode *val, U32 depth)
3703 /* Merge several consecutive EXACTish nodes into one. */
3704 regnode *n = regnext(scan);
3706 regnode *next = scan + NODE_SZ_STR(scan);
3710 regnode *stop = scan;
3711 GET_RE_DEBUG_FLAGS_DECL;
3713 PERL_UNUSED_ARG(depth);
3716 PERL_ARGS_ASSERT_JOIN_EXACT;
3717 #ifndef EXPERIMENTAL_INPLACESCAN
3718 PERL_UNUSED_ARG(flags);
3719 PERL_UNUSED_ARG(val);
3721 DEBUG_PEEP("join",scan,depth);
3723 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3724 * EXACT ones that are mergeable to the current one. */
3726 && (PL_regkind[OP(n)] == NOTHING
3727 || (stringok && OP(n) == OP(scan)))
3729 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3732 if (OP(n) == TAIL || n > next)
3734 if (PL_regkind[OP(n)] == NOTHING) {
3735 DEBUG_PEEP("skip:",n,depth);
3736 NEXT_OFF(scan) += NEXT_OFF(n);
3737 next = n + NODE_STEP_REGNODE;
3744 else if (stringok) {
3745 const unsigned int oldl = STR_LEN(scan);
3746 regnode * const nnext = regnext(n);
3748 /* XXX I (khw) kind of doubt that this works on platforms (should
3749 * Perl ever run on one) where U8_MAX is above 255 because of lots
3750 * of other assumptions */
3751 /* Don't join if the sum can't fit into a single node */
3752 if (oldl + STR_LEN(n) > U8_MAX)
3755 DEBUG_PEEP("merg",n,depth);
3758 NEXT_OFF(scan) += NEXT_OFF(n);
3759 STR_LEN(scan) += STR_LEN(n);
3760 next = n + NODE_SZ_STR(n);
3761 /* Now we can overwrite *n : */
3762 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3770 #ifdef EXPERIMENTAL_INPLACESCAN
3771 if (flags && !NEXT_OFF(n)) {
3772 DEBUG_PEEP("atch", val, depth);
3773 if (reg_off_by_arg[OP(n)]) {
3774 ARG_SET(n, val - n);
3777 NEXT_OFF(n) = val - n;
3785 *unfolded_multi_char = FALSE;
3787 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3788 * can now analyze for sequences of problematic code points. (Prior to
3789 * this final joining, sequences could have been split over boundaries, and
3790 * hence missed). The sequences only happen in folding, hence for any
3791 * non-EXACT EXACTish node */
3792 if (OP(scan) != EXACT && OP(scan) != EXACTL) {
3793 U8* s0 = (U8*) STRING(scan);
3795 U8* s_end = s0 + STR_LEN(scan);
3797 int total_count_delta = 0; /* Total delta number of characters that
3798 multi-char folds expand to */
3800 /* One pass is made over the node's string looking for all the
3801 * possibilities. To avoid some tests in the loop, there are two main
3802 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3807 if (OP(scan) == EXACTFL) {
3810 /* An EXACTFL node would already have been changed to another
3811 * node type unless there is at least one character in it that
3812 * is problematic; likely a character whose fold definition
3813 * won't be known until runtime, and so has yet to be folded.
3814 * For all but the UTF-8 locale, folds are 1-1 in length, but
3815 * to handle the UTF-8 case, we need to create a temporary
3816 * folded copy using UTF-8 locale rules in order to analyze it.
3817 * This is because our macros that look to see if a sequence is
3818 * a multi-char fold assume everything is folded (otherwise the
3819 * tests in those macros would be too complicated and slow).
3820 * Note that here, the non-problematic folds will have already
3821 * been done, so we can just copy such characters. We actually
3822 * don't completely fold the EXACTFL string. We skip the
3823 * unfolded multi-char folds, as that would just create work
3824 * below to figure out the size they already are */
3826 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3829 STRLEN s_len = UTF8SKIP(s);
3830 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3831 Copy(s, d, s_len, U8);
3834 else if (is_FOLDS_TO_MULTI_utf8(s)) {
3835 *unfolded_multi_char = TRUE;
3836 Copy(s, d, s_len, U8);
3839 else if (isASCII(*s)) {
3840 *(d++) = toFOLD(*s);
3844 _to_utf8_fold_flags(s, d, &len, FOLD_FLAGS_FULL);
3850 /* Point the remainder of the routine to look at our temporary
3854 } /* End of creating folded copy of EXACTFL string */
3856 /* Examine the string for a multi-character fold sequence. UTF-8
3857 * patterns have all characters pre-folded by the time this code is
3859 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
3860 length sequence we are looking for is 2 */
3862 int count = 0; /* How many characters in a multi-char fold */
3863 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
3864 if (! len) { /* Not a multi-char fold: get next char */
3869 /* Nodes with 'ss' require special handling, except for
3870 * EXACTFA-ish for which there is no multi-char fold to this */
3871 if (len == 2 && *s == 's' && *(s+1) == 's'
3872 && OP(scan) != EXACTFA
3873 && OP(scan) != EXACTFA_NO_TRIE)
3876 if (OP(scan) != EXACTFL) {
3877 OP(scan) = EXACTFU_SS;
3881 else { /* Here is a generic multi-char fold. */
3882 U8* multi_end = s + len;
3884 /* Count how many characters are in it. In the case of
3885 * /aa, no folds which contain ASCII code points are
3886 * allowed, so check for those, and skip if found. */
3887 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE) {
3888 count = utf8_length(s, multi_end);
3892 while (s < multi_end) {
3895 goto next_iteration;
3905 /* The delta is how long the sequence is minus 1 (1 is how long
3906 * the character that folds to the sequence is) */
3907 total_count_delta += count - 1;
3911 /* We created a temporary folded copy of the string in EXACTFL
3912 * nodes. Therefore we need to be sure it doesn't go below zero,
3913 * as the real string could be shorter */
3914 if (OP(scan) == EXACTFL) {
3915 int total_chars = utf8_length((U8*) STRING(scan),
3916 (U8*) STRING(scan) + STR_LEN(scan));
3917 if (total_count_delta > total_chars) {
3918 total_count_delta = total_chars;
3922 *min_subtract += total_count_delta;
3925 else if (OP(scan) == EXACTFA) {
3927 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
3928 * fold to the ASCII range (and there are no existing ones in the
3929 * upper latin1 range). But, as outlined in the comments preceding
3930 * this function, we need to flag any occurrences of the sharp s.
3931 * This character forbids trie formation (because of added
3933 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
3934 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
3935 || UNICODE_DOT_DOT_VERSION > 0)
3937 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
3938 OP(scan) = EXACTFA_NO_TRIE;
3939 *unfolded_multi_char = TRUE;
3947 /* Non-UTF-8 pattern, not EXACTFA node. Look for the multi-char
3948 * folds that are all Latin1. As explained in the comments
3949 * preceding this function, we look also for the sharp s in EXACTF
3950 * and EXACTFL nodes; it can be in the final position. Otherwise
3951 * we can stop looking 1 byte earlier because have to find at least
3952 * two characters for a multi-fold */
3953 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
3958 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
3959 if (! len) { /* Not a multi-char fold. */
3960 if (*s == LATIN_SMALL_LETTER_SHARP_S
3961 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
3963 *unfolded_multi_char = TRUE;
3970 && isALPHA_FOLD_EQ(*s, 's')
3971 && isALPHA_FOLD_EQ(*(s+1), 's'))
https://perl5.git.perl.org / perl5.git / blob