5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
10 /* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
12 * a regular expression.
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
20 /* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
24 /* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
29 /* The names of the functions have been changed from regcomp and
30 * regexec to pregcomp and pregexec in order to avoid conflicts
31 * with the POSIX routines of the same names.
34 #ifdef PERL_EXT_RE_BUILD
39 * pregcomp and pregexec -- regsub and regerror are not used in perl
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
59 **** Alterations to Henry's code are...
61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
74 #define PERL_IN_REGCOMP_C
77 #ifndef PERL_IN_XSUB_RE
82 #ifdef PERL_IN_XSUB_RE
84 EXTERN_C const struct regexp_engine my_reg_engine;
89 #include "dquote_inline.h"
90 #include "invlist_inline.h"
91 #include "unicode_constants.h"
93 #define HAS_NONLATIN1_FOLD_CLOSURE(i) \
94 _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
95 #define HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(i) \
96 _HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
97 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
98 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
101 #define STATIC static
104 /* this is a chain of data about sub patterns we are processing that
105 need to be handled separately/specially in study_chunk. Its so
106 we can simulate recursion without losing state. */
108 typedef struct scan_frame {
109 regnode *last_regnode; /* last node to process in this frame */
110 regnode *next_regnode; /* next node to process when last is reached */
111 U32 prev_recursed_depth;
112 I32 stopparen; /* what stopparen do we use */
114 struct scan_frame *this_prev_frame; /* this previous frame */
115 struct scan_frame *prev_frame; /* previous frame */
116 struct scan_frame *next_frame; /* next frame */
119 /* Certain characters are output as a sequence with the first being a
121 #define isBACKSLASHED_PUNCT(c) strchr("-[]\\^", c)
124 struct RExC_state_t {
125 U32 flags; /* RXf_* are we folding, multilining? */
126 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
127 char *precomp; /* uncompiled string. */
128 char *precomp_end; /* pointer to end of uncompiled string. */
129 REGEXP *rx_sv; /* The SV that is the regexp. */
130 regexp *rx; /* perl core regexp structure */
131 regexp_internal *rxi; /* internal data for regexp object
133 char *start; /* Start of input for compile */
134 char *end; /* End of input for compile */
135 char *parse; /* Input-scan pointer. */
136 char *copy_start; /* start of copy of input within
137 constructed parse string */
138 char *copy_start_in_input; /* Position in input string
139 corresponding to copy_start */
140 SSize_t whilem_seen; /* number of WHILEM in this expr */
141 regnode *emit_start; /* Start of emitted-code area */
142 regnode *emit_bound; /* First regnode outside of the
144 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
145 implies compiling, so don't emit */
146 regnode_ssc emit_dummy; /* placeholder for emit to point to;
147 large enough for the largest
148 non-EXACTish node, so can use it as
150 I32 naughty; /* How bad is this pattern? */
151 I32 sawback; /* Did we see \1, ...? */
153 SSize_t size; /* Code size. */
154 I32 npar; /* Capture buffer count, (OPEN) plus
155 one. ("par" 0 is the whole
157 I32 nestroot; /* root parens we are in - used by
161 regnode **open_parens; /* pointers to open parens */
162 regnode **close_parens; /* pointers to close parens */
163 regnode *end_op; /* END node in program */
164 I32 utf8; /* whether the pattern is utf8 or not */
165 I32 orig_utf8; /* whether the pattern was originally in utf8 */
166 /* XXX use this for future optimisation of case
167 * where pattern must be upgraded to utf8. */
168 I32 uni_semantics; /* If a d charset modifier should use unicode
169 rules, even if the pattern is not in
171 HV *paren_names; /* Paren names */
173 regnode **recurse; /* Recurse regops */
174 I32 recurse_count; /* Number of recurse regops we have generated */
175 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
177 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
180 I32 override_recoding;
182 I32 recode_x_to_native;
184 I32 in_multi_char_class;
185 struct reg_code_blocks *code_blocks;/* positions of literal (?{})
187 int code_index; /* next code_blocks[] slot */
188 SSize_t maxlen; /* mininum possible number of chars in string to match */
189 scan_frame *frame_head;
190 scan_frame *frame_last;
193 #ifdef ADD_TO_REGEXEC
194 char *starttry; /* -Dr: where regtry was called. */
195 #define RExC_starttry (pRExC_state->starttry)
197 SV *runtime_code_qr; /* qr with the runtime code blocks */
199 const char *lastparse;
201 AV *paren_name_list; /* idx -> name */
202 U32 study_chunk_recursed_count;
205 #define RExC_lastparse (pRExC_state->lastparse)
206 #define RExC_lastnum (pRExC_state->lastnum)
207 #define RExC_paren_name_list (pRExC_state->paren_name_list)
208 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
209 #define RExC_mysv (pRExC_state->mysv1)
210 #define RExC_mysv1 (pRExC_state->mysv1)
211 #define RExC_mysv2 (pRExC_state->mysv2)
214 bool seen_unfolded_sharp_s;
220 #define RExC_flags (pRExC_state->flags)
221 #define RExC_pm_flags (pRExC_state->pm_flags)
222 #define RExC_precomp (pRExC_state->precomp)
223 #define RExC_copy_start_in_input (pRExC_state->copy_start_in_input)
224 #define RExC_copy_start_in_constructed (pRExC_state->copy_start)
225 #define RExC_precomp_end (pRExC_state->precomp_end)
226 #define RExC_rx_sv (pRExC_state->rx_sv)
227 #define RExC_rx (pRExC_state->rx)
228 #define RExC_rxi (pRExC_state->rxi)
229 #define RExC_start (pRExC_state->start)
230 #define RExC_end (pRExC_state->end)
231 #define RExC_parse (pRExC_state->parse)
232 #define RExC_whilem_seen (pRExC_state->whilem_seen)
234 /* Set during the sizing pass when there is a LATIN SMALL LETTER SHARP S in any
235 * EXACTF node, hence was parsed under /di rules. If later in the parse,
236 * something forces the pattern into using /ui rules, the sharp s should be
237 * folded into the sequence 'ss', which takes up more space than previously
238 * calculated. This means that the sizing pass needs to be restarted. (The
239 * node also becomes an EXACTFU_SS.) For all other characters, an EXACTF node
240 * that gets converted to /ui (and EXACTFU) occupies the same amount of space,
241 * so there is no need to resize [perl #125990]. */
242 #define RExC_seen_unfolded_sharp_s (pRExC_state->seen_unfolded_sharp_s)
244 #ifdef RE_TRACK_PATTERN_OFFSETS
245 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
248 #define RExC_emit (pRExC_state->emit)
249 #define RExC_emit_dummy (pRExC_state->emit_dummy)
250 #define RExC_emit_start (pRExC_state->emit_start)
251 #define RExC_emit_bound (pRExC_state->emit_bound)
252 #define RExC_sawback (pRExC_state->sawback)
253 #define RExC_seen (pRExC_state->seen)
254 #define RExC_size (pRExC_state->size)
255 #define RExC_maxlen (pRExC_state->maxlen)
256 #define RExC_npar (pRExC_state->npar)
257 #define RExC_nestroot (pRExC_state->nestroot)
258 #define RExC_extralen (pRExC_state->extralen)
259 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
260 #define RExC_utf8 (pRExC_state->utf8)
261 #define RExC_uni_semantics (pRExC_state->uni_semantics)
262 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
263 #define RExC_open_parens (pRExC_state->open_parens)
264 #define RExC_close_parens (pRExC_state->close_parens)
265 #define RExC_end_op (pRExC_state->end_op)
266 #define RExC_paren_names (pRExC_state->paren_names)
267 #define RExC_recurse (pRExC_state->recurse)
268 #define RExC_recurse_count (pRExC_state->recurse_count)
269 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
270 #define RExC_study_chunk_recursed_bytes \
271 (pRExC_state->study_chunk_recursed_bytes)
272 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
273 #define RExC_contains_locale (pRExC_state->contains_locale)
275 # define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
277 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
278 #define RExC_frame_head (pRExC_state->frame_head)
279 #define RExC_frame_last (pRExC_state->frame_last)
280 #define RExC_frame_count (pRExC_state->frame_count)
281 #define RExC_strict (pRExC_state->strict)
282 #define RExC_study_started (pRExC_state->study_started)
283 #define RExC_warn_text (pRExC_state->warn_text)
284 #define RExC_in_script_run (pRExC_state->in_script_run)
286 /* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
287 * a flag to disable back-off on the fixed/floating substrings - if it's
288 * a high complexity pattern we assume the benefit of avoiding a full match
289 * is worth the cost of checking for the substrings even if they rarely help.
291 #define RExC_naughty (pRExC_state->naughty)
292 #define TOO_NAUGHTY (10)
293 #define MARK_NAUGHTY(add) \
294 if (RExC_naughty < TOO_NAUGHTY) \
295 RExC_naughty += (add)
296 #define MARK_NAUGHTY_EXP(exp, add) \
297 if (RExC_naughty < TOO_NAUGHTY) \
298 RExC_naughty += RExC_naughty / (exp) + (add)
300 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
301 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
302 ((*s) == '{' && regcurly(s)))
305 * Flags to be passed up and down.
307 #define WORST 0 /* Worst case. */
308 #define HASWIDTH 0x01 /* Known to match non-null strings. */
310 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
311 * character. (There needs to be a case: in the switch statement in regexec.c
312 * for any node marked SIMPLE.) Note that this is not the same thing as
315 #define SPSTART 0x04 /* Starts with * or + */
316 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
317 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
318 #define RESTART_PASS1 0x20 /* Need to restart sizing pass */
319 #define NEED_UTF8 0x40 /* In conjunction with RESTART_PASS1, need to
320 calcuate sizes as UTF-8 */
322 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
324 /* whether trie related optimizations are enabled */
325 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
326 #define TRIE_STUDY_OPT
327 #define FULL_TRIE_STUDY
333 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
334 #define PBITVAL(paren) (1 << ((paren) & 7))
335 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
336 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
337 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
339 #define REQUIRE_UTF8(flagp) STMT_START { \
342 *flagp = RESTART_PASS1|NEED_UTF8; \
347 /* Change from /d into /u rules, and restart the parse if we've already seen
348 * something whose size would increase as a result, by setting *flagp and
349 * returning 'restart_retval'. RExC_uni_semantics is a flag that indicates
350 * we've changed to /u during the parse. */
351 #define REQUIRE_UNI_RULES(flagp, restart_retval) \
353 if (DEPENDS_SEMANTICS) { \
355 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
356 RExC_uni_semantics = 1; \
357 if (RExC_seen_unfolded_sharp_s) { \
358 *flagp |= RESTART_PASS1; \
359 return restart_retval; \
364 /* Executes a return statement with the value 'X', if 'flags' contains any of
365 * 'RESTART_PASS1', 'NEED_UTF8', or 'extra'. If so, *flagp is set to those
367 #define RETURN_X_ON_RESTART_OR_FLAGS(X, flags, flagp, extra) \
369 if ((flags) & (RESTART_PASS1|NEED_UTF8|(extra))) { \
370 *(flagp) = (flags) & (RESTART_PASS1|NEED_UTF8|(extra)); \
375 #define RETURN_FAIL_ON_RESTART_OR_FLAGS(flags,flagp,extra) \
376 RETURN_X_ON_RESTART_OR_FLAGS(NULL,flags,flagp,extra)
378 #define RETURN_X_ON_RESTART(X, flags,flagp) \
379 RETURN_X_ON_RESTART_OR_FLAGS( X, flags, flagp, 0)
382 #define RETURN_FAIL_ON_RESTART_FLAGP_OR_FLAGS(flagp,extra) \
383 if (*(flagp) & (RESTART_PASS1|(extra))) return NULL
385 #define MUST_RESTART(flags) ((flags) & (RESTART_PASS1))
387 #define RETURN_FAIL_ON_RESTART(flags,flagp) \
388 RETURN_X_ON_RESTART(NULL, flags,flagp)
389 #define RETURN_FAIL_ON_RESTART_FLAGP(flagp) \
390 RETURN_FAIL_ON_RESTART_FLAGP_OR_FLAGS(flagp,0)
392 /* This converts the named class defined in regcomp.h to its equivalent class
393 * number defined in handy.h. */
394 #define namedclass_to_classnum(class) ((int) ((class) / 2))
395 #define classnum_to_namedclass(classnum) ((classnum) * 2)
397 #define _invlist_union_complement_2nd(a, b, output) \
398 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
399 #define _invlist_intersection_complement_2nd(a, b, output) \
400 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
402 /* About scan_data_t.
404 During optimisation we recurse through the regexp program performing
405 various inplace (keyhole style) optimisations. In addition study_chunk
406 and scan_commit populate this data structure with information about
407 what strings MUST appear in the pattern. We look for the longest
408 string that must appear at a fixed location, and we look for the
409 longest string that may appear at a floating location. So for instance
414 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
415 strings (because they follow a .* construct). study_chunk will identify
416 both FOO and BAR as being the longest fixed and floating strings respectively.
418 The strings can be composites, for instance
422 will result in a composite fixed substring 'foo'.
424 For each string some basic information is maintained:
427 This is the position the string must appear at, or not before.
428 It also implicitly (when combined with minlenp) tells us how many
429 characters must match before the string we are searching for.
430 Likewise when combined with minlenp and the length of the string it
431 tells us how many characters must appear after the string we have
435 Only used for floating strings. This is the rightmost point that
436 the string can appear at. If set to SSize_t_MAX it indicates that the
437 string can occur infinitely far to the right.
438 For fixed strings, it is equal to min_offset.
441 A pointer to the minimum number of characters of the pattern that the
442 string was found inside. This is important as in the case of positive
443 lookahead or positive lookbehind we can have multiple patterns
448 The minimum length of the pattern overall is 3, the minimum length
449 of the lookahead part is 3, but the minimum length of the part that
450 will actually match is 1. So 'FOO's minimum length is 3, but the
451 minimum length for the F is 1. This is important as the minimum length
452 is used to determine offsets in front of and behind the string being
453 looked for. Since strings can be composites this is the length of the
454 pattern at the time it was committed with a scan_commit. Note that
455 the length is calculated by study_chunk, so that the minimum lengths
456 are not known until the full pattern has been compiled, thus the
457 pointer to the value.
461 In the case of lookbehind the string being searched for can be
462 offset past the start point of the final matching string.
463 If this value was just blithely removed from the min_offset it would
464 invalidate some of the calculations for how many chars must match
465 before or after (as they are derived from min_offset and minlen and
466 the length of the string being searched for).
467 When the final pattern is compiled and the data is moved from the
468 scan_data_t structure into the regexp structure the information
469 about lookbehind is factored in, with the information that would
470 have been lost precalculated in the end_shift field for the
473 The fields pos_min and pos_delta are used to store the minimum offset
474 and the delta to the maximum offset at the current point in the pattern.
478 struct scan_data_substrs {
479 SV *str; /* longest substring found in pattern */
480 SSize_t min_offset; /* earliest point in string it can appear */
481 SSize_t max_offset; /* latest point in string it can appear */
482 SSize_t *minlenp; /* pointer to the minlen relevant to the string */
483 SSize_t lookbehind; /* is the pos of the string modified by LB */
484 I32 flags; /* per substring SF_* and SCF_* flags */
487 typedef struct scan_data_t {
488 /*I32 len_min; unused */
489 /*I32 len_delta; unused */
493 SSize_t last_end; /* min value, <0 unless valid. */
494 SSize_t last_start_min;
495 SSize_t last_start_max;
496 U8 cur_is_floating; /* whether the last_* values should be set as
497 * the next fixed (0) or floating (1)
500 /* [0] is longest fixed substring so far, [1] is longest float so far */
501 struct scan_data_substrs substrs[2];
503 I32 flags; /* common SF_* and SCF_* flags */
505 SSize_t *last_closep;
506 regnode_ssc *start_class;
510 * Forward declarations for pregcomp()'s friends.
513 static const scan_data_t zero_scan_data = {
514 0, 0, NULL, 0, 0, 0, 0,
516 { NULL, 0, 0, 0, 0, 0 },
517 { NULL, 0, 0, 0, 0, 0 },
524 #define SF_BEFORE_SEOL 0x0001
525 #define SF_BEFORE_MEOL 0x0002
526 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
528 #define SF_IS_INF 0x0040
529 #define SF_HAS_PAR 0x0080
530 #define SF_IN_PAR 0x0100
531 #define SF_HAS_EVAL 0x0200
534 /* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
535 * longest substring in the pattern. When it is not set the optimiser keeps
536 * track of position, but does not keep track of the actual strings seen,
538 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
541 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
542 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
543 * turned off because of the alternation (BRANCH). */
544 #define SCF_DO_SUBSTR 0x0400
546 #define SCF_DO_STCLASS_AND 0x0800
547 #define SCF_DO_STCLASS_OR 0x1000
548 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
549 #define SCF_WHILEM_VISITED_POS 0x2000
551 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
552 #define SCF_SEEN_ACCEPT 0x8000
553 #define SCF_TRIE_DOING_RESTUDY 0x10000
554 #define SCF_IN_DEFINE 0x20000
559 #define UTF cBOOL(RExC_utf8)
561 /* The enums for all these are ordered so things work out correctly */
562 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
563 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
564 == REGEX_DEPENDS_CHARSET)
565 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
566 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
567 >= REGEX_UNICODE_CHARSET)
568 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
569 == REGEX_ASCII_RESTRICTED_CHARSET)
570 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
571 >= REGEX_ASCII_RESTRICTED_CHARSET)
572 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
573 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
575 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
577 /* For programs that want to be strictly Unicode compatible by dying if any
578 * attempt is made to match a non-Unicode code point against a Unicode
580 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
582 #define OOB_NAMEDCLASS -1
584 /* There is no code point that is out-of-bounds, so this is problematic. But
585 * its only current use is to initialize a variable that is always set before
587 #define OOB_UNICODE 0xDEADBEEF
589 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
592 /* length of regex to show in messages that don't mark a position within */
593 #define RegexLengthToShowInErrorMessages 127
596 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
597 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
598 * op/pragma/warn/regcomp.
600 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
601 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
603 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
604 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
606 /* The code in this file in places uses one level of recursion with parsing
607 * rebased to an alternate string constructed by us in memory. This can take
608 * the form of something that is completely different from the input, or
609 * something that uses the input as part of the alternate. In the first case,
610 * there should be no possibility of an error, as we are in complete control of
611 * the alternate string. But in the second case we don't completely control
612 * the input portion, so there may be errors in that. Here's an example:
614 * is handled specially because \x{df} folds to a sequence of more than one
615 * character: 'ss'. What is done is to create and parse an alternate string,
616 * which looks like this:
617 * /(?:\x{DF}|[abc\x{DF}def])/ui
618 * where it uses the input unchanged in the middle of something it constructs,
619 * which is a branch for the DF outside the character class, and clustering
620 * parens around the whole thing. (It knows enough to skip the DF inside the
621 * class while in this substitute parse.) 'abc' and 'def' may have errors that
622 * need to be reported. The general situation looks like this:
624 * |<------- identical ------>|
626 * Input: ---------------------------------------------------------------
627 * Constructed: ---------------------------------------------------
629 * |<------- identical ------>|
631 * sI..eI is the portion of the input pattern we are concerned with here.
632 * sC..EC is the constructed substitute parse string.
633 * sC..tC is constructed by us
634 * tC..eC is an exact duplicate of the portion of the input pattern tI..eI.
635 * In the diagram, these are vertically aligned.
636 * eC..EC is also constructed by us.
637 * xC is the position in the substitute parse string where we found a
639 * xI is the position in the original pattern corresponding to xC.
641 * We want to display a message showing the real input string. Thus we need to
642 * translate from xC to xI. We know that xC >= tC, since the portion of the
643 * string sC..tC has been constructed by us, and so shouldn't have errors. We
645 * xI = tI + (xC - tC)
647 * When the substitute parse is constructed, the code needs to set:
650 * RExC_copy_start_in_input (tI)
651 * RExC_copy_start_in_constructed (tC)
652 * and restore them when done.
654 * During normal processing of the input pattern, both
655 * 'RExC_copy_start_in_input' and 'RExC_copy_start_in_constructed' are set to
656 * sI, so that xC equals xI.
659 #define sI RExC_precomp
660 #define eI RExC_precomp_end
661 #define sC RExC_start
663 #define tI RExC_copy_start_in_input
664 #define tC RExC_copy_start_in_constructed
665 #define xI(xC) (tI + (xC - tC))
666 #define xI_offset(xC) (xI(xC) - sI)
668 #define REPORT_LOCATION_ARGS(xC) \
670 (xI(xC) > eI) /* Don't run off end */ \
671 ? eC - sC /* Length before the <--HERE */ \
672 : ((xI_offset(xC) >= 0) \
674 : (Perl_croak(aTHX_ "panic: %s: %d: negative offset: %" \
675 IVdf " trying to output message for " \
677 __FILE__, __LINE__, (IV) xI_offset(xC), \
678 ((int) (eC - sC)), sC), 0)), \
679 sI), /* The input pattern printed up to the <--HERE */ \
681 (xI(xC) > eI) ? 0 : eI - xI(xC), /* Length after <--HERE */ \
682 (xI(xC) > eI) ? eI : xI(xC)) /* pattern after <--HERE */
684 /* Used to point after bad bytes for an error message, but avoid skipping
685 * past a nul byte. */
686 #define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
689 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
690 * arg. Show regex, up to a maximum length. If it's too long, chop and add
693 #define _FAIL(code) STMT_START { \
694 const char *ellipses = ""; \
695 IV len = RExC_precomp_end - RExC_precomp; \
698 SAVEFREESV(RExC_rx_sv); \
699 if (len > RegexLengthToShowInErrorMessages) { \
700 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
701 len = RegexLengthToShowInErrorMessages - 10; \
707 #define FAIL(msg) _FAIL( \
708 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
709 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
711 #define FAIL2(msg,arg) _FAIL( \
712 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
713 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
716 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
718 #define Simple_vFAIL(m) STMT_START { \
719 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
720 m, REPORT_LOCATION_ARGS(RExC_parse)); \
724 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
726 #define vFAIL(m) STMT_START { \
728 SAVEFREESV(RExC_rx_sv); \
733 * Like Simple_vFAIL(), but accepts two arguments.
735 #define Simple_vFAIL2(m,a1) STMT_START { \
736 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
737 REPORT_LOCATION_ARGS(RExC_parse)); \
741 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
743 #define vFAIL2(m,a1) STMT_START { \
745 SAVEFREESV(RExC_rx_sv); \
746 Simple_vFAIL2(m, a1); \
751 * Like Simple_vFAIL(), but accepts three arguments.
753 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
754 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
755 REPORT_LOCATION_ARGS(RExC_parse)); \
759 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
761 #define vFAIL3(m,a1,a2) STMT_START { \
763 SAVEFREESV(RExC_rx_sv); \
764 Simple_vFAIL3(m, a1, a2); \
768 * Like Simple_vFAIL(), but accepts four arguments.
770 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
771 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
772 REPORT_LOCATION_ARGS(RExC_parse)); \
775 #define vFAIL4(m,a1,a2,a3) STMT_START { \
777 SAVEFREESV(RExC_rx_sv); \
778 Simple_vFAIL4(m, a1, a2, a3); \
781 /* A specialized version of vFAIL2 that works with UTF8f */
782 #define vFAIL2utf8f(m, a1) STMT_START { \
784 SAVEFREESV(RExC_rx_sv); \
785 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
786 REPORT_LOCATION_ARGS(RExC_parse)); \
789 #define vFAIL3utf8f(m, a1, a2) STMT_START { \
791 SAVEFREESV(RExC_rx_sv); \
792 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
793 REPORT_LOCATION_ARGS(RExC_parse)); \
796 /* These have asserts in them because of [perl #122671] Many warnings in
797 * regcomp.c can occur twice. If they get output in pass1 and later in that
798 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
799 * would get output again. So they should be output in pass2, and these
800 * asserts make sure new warnings follow that paradigm. */
802 /* m is not necessarily a "literal string", in this macro */
803 #define reg_warn_non_literal_string(loc, m) STMT_START { \
804 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
805 "%s" REPORT_LOCATION, \
806 m, REPORT_LOCATION_ARGS(loc)); \
809 #define ckWARNreg(loc,m) STMT_START { \
810 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
812 REPORT_LOCATION_ARGS(loc)); \
815 #define vWARN(loc, m) STMT_START { \
816 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
818 REPORT_LOCATION_ARGS(loc)); \
821 #define vWARN_dep(loc, m) STMT_START { \
822 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
824 REPORT_LOCATION_ARGS(loc)); \
827 #define ckWARNdep(loc,m) STMT_START { \
828 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
830 REPORT_LOCATION_ARGS(loc)); \
833 #define ckWARNregdep(loc,m) STMT_START { \
834 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
837 REPORT_LOCATION_ARGS(loc)); \
840 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
841 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
843 a1, REPORT_LOCATION_ARGS(loc)); \
846 #define ckWARN2reg(loc, m, a1) STMT_START { \
847 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
849 a1, REPORT_LOCATION_ARGS(loc)); \
852 #define vWARN3(loc, m, a1, a2) STMT_START { \
853 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
855 a1, a2, REPORT_LOCATION_ARGS(loc)); \
858 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
859 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
862 REPORT_LOCATION_ARGS(loc)); \
865 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
866 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
869 REPORT_LOCATION_ARGS(loc)); \
872 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
873 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
876 REPORT_LOCATION_ARGS(loc)); \
879 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
880 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
883 REPORT_LOCATION_ARGS(loc)); \
886 /* Macros for recording node offsets. 20001227 mjd@plover.com
887 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
888 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
889 * Element 0 holds the number n.
890 * Position is 1 indexed.
892 #ifndef RE_TRACK_PATTERN_OFFSETS
893 #define Set_Node_Offset_To_R(node,byte)
894 #define Set_Node_Offset(node,byte)
895 #define Set_Cur_Node_Offset
896 #define Set_Node_Length_To_R(node,len)
897 #define Set_Node_Length(node,len)
898 #define Set_Node_Cur_Length(node,start)
899 #define Node_Offset(n)
900 #define Node_Length(n)
901 #define Set_Node_Offset_Length(node,offset,len)
902 #define ProgLen(ri) ri->u.proglen
903 #define SetProgLen(ri,x) ri->u.proglen = x
905 #define ProgLen(ri) ri->u.offsets[0]
906 #define SetProgLen(ri,x) ri->u.offsets[0] = x
907 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
909 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
910 __LINE__, (int)(node), (int)(byte))); \
912 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
915 RExC_offsets[2*(node)-1] = (byte); \
920 #define Set_Node_Offset(node,byte) \
921 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
922 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
924 #define Set_Node_Length_To_R(node,len) STMT_START { \
926 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
927 __LINE__, (int)(node), (int)(len))); \
929 Perl_croak(aTHX_ "value of node is %d in Length macro", \
932 RExC_offsets[2*(node)] = (len); \
937 #define Set_Node_Length(node,len) \
938 Set_Node_Length_To_R((node)-RExC_emit_start, len)
939 #define Set_Node_Cur_Length(node, start) \
940 Set_Node_Length(node, RExC_parse - start)
942 /* Get offsets and lengths */
943 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
944 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
946 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
947 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
948 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
952 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
953 #define EXPERIMENTAL_INPLACESCAN
954 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
958 Perl_re_printf(pTHX_ const char *fmt, ...)
962 PerlIO *f= Perl_debug_log;
963 PERL_ARGS_ASSERT_RE_PRINTF;
965 result = PerlIO_vprintf(f, fmt, ap);
971 Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
975 PerlIO *f= Perl_debug_log;
976 PERL_ARGS_ASSERT_RE_INDENTF;
978 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
979 result = PerlIO_vprintf(f, fmt, ap);
983 #endif /* DEBUGGING */
985 #define DEBUG_RExC_seen() \
986 DEBUG_OPTIMISE_MORE_r({ \
987 Perl_re_printf( aTHX_ "RExC_seen: "); \
989 if (RExC_seen & REG_ZERO_LEN_SEEN) \
990 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
992 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
993 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
995 if (RExC_seen & REG_GPOS_SEEN) \
996 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
998 if (RExC_seen & REG_RECURSE_SEEN) \
999 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
1001 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
1002 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
1004 if (RExC_seen & REG_VERBARG_SEEN) \
1005 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
1007 if (RExC_seen & REG_CUTGROUP_SEEN) \
1008 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
1010 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
1011 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
1013 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
1014 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
1016 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
1017 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
1019 Perl_re_printf( aTHX_ "\n"); \
1022 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
1023 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
1028 S_debug_show_study_flags(pTHX_ U32 flags, const char *open_str,
1029 const char *close_str)
1034 Perl_re_printf( aTHX_ "%s", open_str);
1035 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_SEOL);
1036 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_MEOL);
1037 DEBUG_SHOW_STUDY_FLAG(flags, SF_IS_INF);
1038 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_PAR);
1039 DEBUG_SHOW_STUDY_FLAG(flags, SF_IN_PAR);
1040 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_EVAL);
1041 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_SUBSTR);
1042 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_AND);
1043 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_OR);
1044 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS);
1045 DEBUG_SHOW_STUDY_FLAG(flags, SCF_WHILEM_VISITED_POS);
1046 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_RESTUDY);
1047 DEBUG_SHOW_STUDY_FLAG(flags, SCF_SEEN_ACCEPT);
1048 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_DOING_RESTUDY);
1049 DEBUG_SHOW_STUDY_FLAG(flags, SCF_IN_DEFINE);
1050 Perl_re_printf( aTHX_ "%s", close_str);
1055 S_debug_studydata(pTHX_ const char *where, scan_data_t *data,
1056 U32 depth, int is_inf)
1058 GET_RE_DEBUG_FLAGS_DECL;
1060 DEBUG_OPTIMISE_MORE_r({
1063 Perl_re_indentf(aTHX_ "%s: Pos:%" IVdf "/%" IVdf " Flags: 0x%" UVXf,
1067 (IV)data->pos_delta,
1071 S_debug_show_study_flags(aTHX_ data->flags," [","]");
1073 Perl_re_printf( aTHX_
1074 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s",
1076 (IV)(data->last_closep ? *((data)->last_closep) : -1),
1077 is_inf ? "INF " : ""
1080 if (data->last_found) {
1082 Perl_re_printf(aTHX_
1083 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf,
1084 SvPVX_const(data->last_found),
1086 (IV)data->last_start_min,
1087 (IV)data->last_start_max
1090 for (i = 0; i < 2; i++) {
1091 Perl_re_printf(aTHX_
1092 " %s%s: '%s' @ %" IVdf "/%" IVdf,
1093 data->cur_is_floating == i ? "*" : "",
1094 i ? "Float" : "Fixed",
1095 SvPVX_const(data->substrs[i].str),
1096 (IV)data->substrs[i].min_offset,
1097 (IV)data->substrs[i].max_offset
1099 S_debug_show_study_flags(aTHX_ data->substrs[i].flags," [","]");
1103 Perl_re_printf( aTHX_ "\n");
1109 S_debug_peep(pTHX_ const char *str, const RExC_state_t *pRExC_state,
1110 regnode *scan, U32 depth, U32 flags)
1112 GET_RE_DEBUG_FLAGS_DECL;
1119 Next = regnext(scan);
1120 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
1121 Perl_re_indentf( aTHX_ "%s>%3d: %s (%d)",
1124 REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),
1125 Next ? (REG_NODE_NUM(Next)) : 0 );
1126 S_debug_show_study_flags(aTHX_ flags," [ ","]");
1127 Perl_re_printf( aTHX_ "\n");
1132 # define DEBUG_STUDYDATA(where, data, depth, is_inf) \
1133 S_debug_studydata(aTHX_ where, data, depth, is_inf)
1135 # define DEBUG_PEEP(str, scan, depth, flags) \
1136 S_debug_peep(aTHX_ str, pRExC_state, scan, depth, flags)
1139 # define DEBUG_STUDYDATA(where, data, depth, is_inf) NOOP
1140 # define DEBUG_PEEP(str, scan, depth, flags) NOOP
1144 /* =========================================================
1145 * BEGIN edit_distance stuff.
1147 * This calculates how many single character changes of any type are needed to
1148 * transform a string into another one. It is taken from version 3.1 of
1150 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1153 /* Our unsorted dictionary linked list. */
1154 /* Note we use UVs, not chars. */
1159 struct dictionary* next;
1161 typedef struct dictionary item;
1164 PERL_STATIC_INLINE item*
1165 push(UV key,item* curr)
1168 Newx(head, 1, item);
1176 PERL_STATIC_INLINE item*
1177 find(item* head, UV key)
1179 item* iterator = head;
1181 if (iterator->key == key){
1184 iterator = iterator->next;
1190 PERL_STATIC_INLINE item*
1191 uniquePush(item* head,UV key)
1193 item* iterator = head;
1196 if (iterator->key == key) {
1199 iterator = iterator->next;
1202 return push(key,head);
1205 PERL_STATIC_INLINE void
1206 dict_free(item* head)
1208 item* iterator = head;
1211 item* temp = iterator;
1212 iterator = iterator->next;
1219 /* End of Dictionary Stuff */
1221 /* All calculations/work are done here */
1223 S_edit_distance(const UV* src,
1225 const STRLEN x, /* length of src[] */
1226 const STRLEN y, /* length of tgt[] */
1227 const SSize_t maxDistance
1231 UV swapCount,swapScore,targetCharCount,i,j;
1233 UV score_ceil = x + y;
1235 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1237 /* intialize matrix start values */
1238 Newx(scores, ( (x + 2) * (y + 2)), UV);
1239 scores[0] = score_ceil;
1240 scores[1 * (y + 2) + 0] = score_ceil;
1241 scores[0 * (y + 2) + 1] = score_ceil;
1242 scores[1 * (y + 2) + 1] = 0;
1243 head = uniquePush(uniquePush(head,src[0]),tgt[0]);
1248 for (i=1;i<=x;i++) {
1250 head = uniquePush(head,src[i]);
1251 scores[(i+1) * (y + 2) + 1] = i;
1252 scores[(i+1) * (y + 2) + 0] = score_ceil;
1255 for (j=1;j<=y;j++) {
1258 head = uniquePush(head,tgt[j]);
1259 scores[1 * (y + 2) + (j + 1)] = j;
1260 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1263 targetCharCount = find(head,tgt[j-1])->value;
1264 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1266 if (src[i-1] != tgt[j-1]){
1267 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));
1271 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1275 find(head,src[i-1])->value = i;
1279 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1282 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1286 /* END of edit_distance() stuff
1287 * ========================================================= */
1289 /* is c a control character for which we have a mnemonic? */
1290 #define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
1293 S_cntrl_to_mnemonic(const U8 c)
1295 /* Returns the mnemonic string that represents character 'c', if one
1296 * exists; NULL otherwise. The only ones that exist for the purposes of
1297 * this routine are a few control characters */
1300 case '\a': return "\\a";
1301 case '\b': return "\\b";
1302 case ESC_NATIVE: return "\\e";
1303 case '\f': return "\\f";
1304 case '\n': return "\\n";
1305 case '\r': return "\\r";
1306 case '\t': return "\\t";
1312 /* Mark that we cannot extend a found fixed substring at this point.
1313 Update the longest found anchored substring or the longest found
1314 floating substrings if needed. */
1317 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1318 SSize_t *minlenp, int is_inf)
1320 const STRLEN l = CHR_SVLEN(data->last_found);
1321 SV * const longest_sv = data->substrs[data->cur_is_floating].str;
1322 const STRLEN old_l = CHR_SVLEN(longest_sv);
1323 GET_RE_DEBUG_FLAGS_DECL;
1325 PERL_ARGS_ASSERT_SCAN_COMMIT;
1327 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
1328 const U8 i = data->cur_is_floating;
1329 SvSetMagicSV(longest_sv, data->last_found);
1330 data->substrs[i].min_offset = l ? data->last_start_min : data->pos_min;
1333 data->substrs[0].max_offset = data->substrs[0].min_offset;
1335 data->substrs[1].max_offset = (l
1336 ? data->last_start_max
1337 : (data->pos_delta > SSize_t_MAX - data->pos_min
1339 : data->pos_min + data->pos_delta));
1341 || (STRLEN)data->substrs[1].max_offset > (STRLEN)SSize_t_MAX)
1342 data->substrs[1].max_offset = SSize_t_MAX;
1345 if (data->flags & SF_BEFORE_EOL)
1346 data->substrs[i].flags |= (data->flags & SF_BEFORE_EOL);
1348 data->substrs[i].flags &= ~SF_BEFORE_EOL;
1349 data->substrs[i].minlenp = minlenp;
1350 data->substrs[i].lookbehind = 0;
1353 SvCUR_set(data->last_found, 0);
1355 SV * const sv = data->last_found;
1356 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1357 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1362 data->last_end = -1;
1363 data->flags &= ~SF_BEFORE_EOL;
1364 DEBUG_STUDYDATA("commit", data, 0, is_inf);
1367 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1368 * list that describes which code points it matches */
1371 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1373 /* Set the SSC 'ssc' to match an empty string or any code point */
1375 PERL_ARGS_ASSERT_SSC_ANYTHING;
1377 assert(is_ANYOF_SYNTHETIC(ssc));
1379 /* mortalize so won't leak */
1380 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
1381 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1385 S_ssc_is_anything(const regnode_ssc *ssc)
1387 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1388 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1389 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1390 * in any way, so there's no point in using it */
1395 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1397 assert(is_ANYOF_SYNTHETIC(ssc));
1399 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1403 /* See if the list consists solely of the range 0 - Infinity */
1404 invlist_iterinit(ssc->invlist);
1405 ret = invlist_iternext(ssc->invlist, &start, &end)
1409 invlist_iterfinish(ssc->invlist);
1415 /* If e.g., both \w and \W are set, matches everything */
1416 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1418 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1419 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1429 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1431 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1432 * string, any code point, or any posix class under locale */
1434 PERL_ARGS_ASSERT_SSC_INIT;
1436 Zero(ssc, 1, regnode_ssc);
1437 set_ANYOF_SYNTHETIC(ssc);
1438 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1441 /* If any portion of the regex is to operate under locale rules that aren't
1442 * fully known at compile time, initialization includes it. The reason
1443 * this isn't done for all regexes is that the optimizer was written under
1444 * the assumption that locale was all-or-nothing. Given the complexity and
1445 * lack of documentation in the optimizer, and that there are inadequate
1446 * test cases for locale, many parts of it may not work properly, it is
1447 * safest to avoid locale unless necessary. */
1448 if (RExC_contains_locale) {
1449 ANYOF_POSIXL_SETALL(ssc);
1452 ANYOF_POSIXL_ZERO(ssc);
1457 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1458 const regnode_ssc *ssc)
1460 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1461 * to the list of code points matched, and locale posix classes; hence does
1462 * not check its flags) */
1467 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1469 assert(is_ANYOF_SYNTHETIC(ssc));
1471 invlist_iterinit(ssc->invlist);
1472 ret = invlist_iternext(ssc->invlist, &start, &end)
1476 invlist_iterfinish(ssc->invlist);
1482 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1490 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1491 const regnode_charclass* const node)
1493 /* Returns a mortal inversion list defining which code points are matched
1494 * by 'node', which is of type ANYOF. Handles complementing the result if
1495 * appropriate. If some code points aren't knowable at this time, the
1496 * returned list must, and will, contain every code point that is a
1500 SV* only_utf8_locale_invlist = NULL;
1502 const U32 n = ARG(node);
1503 bool new_node_has_latin1 = FALSE;
1505 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1507 /* Look at the data structure created by S_set_ANYOF_arg() */
1508 if (n != ANYOF_ONLY_HAS_BITMAP) {
1509 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1510 AV * const av = MUTABLE_AV(SvRV(rv));
1511 SV **const ary = AvARRAY(av);
1512 assert(RExC_rxi->data->what[n] == 's');
1514 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1515 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1]), NULL));
1517 else if (ary[0] && ary[0] != &PL_sv_undef) {
1519 /* Here, no compile-time swash, and there are things that won't be
1520 * known until runtime -- we have to assume it could be anything */
1521 invlist = sv_2mortal(_new_invlist(1));
1522 return _add_range_to_invlist(invlist, 0, UV_MAX);
1524 else if (ary[3] && ary[3] != &PL_sv_undef) {
1526 /* Here no compile-time swash, and no run-time only data. Use the
1527 * node's inversion list */
1528 invlist = sv_2mortal(invlist_clone(ary[3], NULL));
1531 /* Get the code points valid only under UTF-8 locales */
1532 if ((ANYOF_FLAGS(node) & ANYOFL_FOLD)
1533 && ary[2] && ary[2] != &PL_sv_undef)
1535 only_utf8_locale_invlist = ary[2];
1540 invlist = sv_2mortal(_new_invlist(0));
1543 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1544 * code points, and an inversion list for the others, but if there are code
1545 * points that should match only conditionally on the target string being
1546 * UTF-8, those are placed in the inversion list, and not the bitmap.
1547 * Since there are circumstances under which they could match, they are
1548 * included in the SSC. But if the ANYOF node is to be inverted, we have
1549 * to exclude them here, so that when we invert below, the end result
1550 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1551 * have to do this here before we add the unconditionally matched code
1553 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1554 _invlist_intersection_complement_2nd(invlist,
1559 /* Add in the points from the bit map */
1560 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1561 if (ANYOF_BITMAP_TEST(node, i)) {
1562 unsigned int start = i++;
1564 for (; i < NUM_ANYOF_CODE_POINTS && ANYOF_BITMAP_TEST(node, i); ++i) {
1567 invlist = _add_range_to_invlist(invlist, start, i-1);
1568 new_node_has_latin1 = TRUE;
1572 /* If this can match all upper Latin1 code points, have to add them
1573 * as well. But don't add them if inverting, as when that gets done below,
1574 * it would exclude all these characters, including the ones it shouldn't
1575 * that were added just above */
1576 if (! (ANYOF_FLAGS(node) & ANYOF_INVERT) && OP(node) == ANYOFD
1577 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1579 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1582 /* Similarly for these */
1583 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1584 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1587 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1588 _invlist_invert(invlist);
1590 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOFL_FOLD) {
1592 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1593 * locale. We can skip this if there are no 0-255 at all. */
1594 _invlist_union(invlist, PL_Latin1, &invlist);
1597 /* Similarly add the UTF-8 locale possible matches. These have to be
1598 * deferred until after the non-UTF-8 locale ones are taken care of just
1599 * above, or it leads to wrong results under ANYOF_INVERT */
1600 if (only_utf8_locale_invlist) {
1601 _invlist_union_maybe_complement_2nd(invlist,
1602 only_utf8_locale_invlist,
1603 ANYOF_FLAGS(node) & ANYOF_INVERT,
1610 /* These two functions currently do the exact same thing */
1611 #define ssc_init_zero ssc_init
1613 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1614 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1616 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1617 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1618 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1621 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1622 const regnode_charclass *and_with)
1624 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1625 * another SSC or a regular ANYOF class. Can create false positives. */
1630 PERL_ARGS_ASSERT_SSC_AND;
1632 assert(is_ANYOF_SYNTHETIC(ssc));
1634 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1635 * the code point inversion list and just the relevant flags */
1636 if (is_ANYOF_SYNTHETIC(and_with)) {
1637 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1638 anded_flags = ANYOF_FLAGS(and_with);
1640 /* XXX This is a kludge around what appears to be deficiencies in the
1641 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1642 * there are paths through the optimizer where it doesn't get weeded
1643 * out when it should. And if we don't make some extra provision for
1644 * it like the code just below, it doesn't get added when it should.
1645 * This solution is to add it only when AND'ing, which is here, and
1646 * only when what is being AND'ed is the pristine, original node
1647 * matching anything. Thus it is like adding it to ssc_anything() but
1648 * only when the result is to be AND'ed. Probably the same solution
1649 * could be adopted for the same problem we have with /l matching,
1650 * which is solved differently in S_ssc_init(), and that would lead to
1651 * fewer false positives than that solution has. But if this solution
1652 * creates bugs, the consequences are only that a warning isn't raised
1653 * that should be; while the consequences for having /l bugs is
1654 * incorrect matches */
1655 if (ssc_is_anything((regnode_ssc *)and_with)) {
1656 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1660 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1661 if (OP(and_with) == ANYOFD) {
1662 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1665 anded_flags = ANYOF_FLAGS(and_with)
1666 &( ANYOF_COMMON_FLAGS
1667 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1668 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1669 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(and_with))) {
1671 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1676 ANYOF_FLAGS(ssc) &= anded_flags;
1678 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1679 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1680 * 'and_with' may be inverted. When not inverted, we have the situation of
1682 * (C1 | P1) & (C2 | P2)
1683 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1684 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1685 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1686 * <= ((C1 & C2) | P1 | P2)
1687 * Alternatively, the last few steps could be:
1688 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1689 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1690 * <= (C1 | C2 | (P1 & P2))
1691 * We favor the second approach if either P1 or P2 is non-empty. This is
1692 * because these components are a barrier to doing optimizations, as what
1693 * they match cannot be known until the moment of matching as they are
1694 * dependent on the current locale, 'AND"ing them likely will reduce or
1696 * But we can do better if we know that C1,P1 are in their initial state (a
1697 * frequent occurrence), each matching everything:
1698 * (<everything>) & (C2 | P2) = C2 | P2
1699 * Similarly, if C2,P2 are in their initial state (again a frequent
1700 * occurrence), the result is a no-op
1701 * (C1 | P1) & (<everything>) = C1 | P1
1704 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1705 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1706 * <= (C1 & ~C2) | (P1 & ~P2)
1709 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1710 && ! is_ANYOF_SYNTHETIC(and_with))
1714 ssc_intersection(ssc,
1716 FALSE /* Has already been inverted */
1719 /* If either P1 or P2 is empty, the intersection will be also; can skip
1721 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
1722 ANYOF_POSIXL_ZERO(ssc);
1724 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1726 /* Note that the Posix class component P from 'and_with' actually
1728 * P = Pa | Pb | ... | Pn
1729 * where each component is one posix class, such as in [\w\s].
1731 * ~P = ~(Pa | Pb | ... | Pn)
1732 * = ~Pa & ~Pb & ... & ~Pn
1733 * <= ~Pa | ~Pb | ... | ~Pn
1734 * The last is something we can easily calculate, but unfortunately
1735 * is likely to have many false positives. We could do better
1736 * in some (but certainly not all) instances if two classes in
1737 * P have known relationships. For example
1738 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1740 * :lower: & :print: = :lower:
1741 * And similarly for classes that must be disjoint. For example,
1742 * since \s and \w can have no elements in common based on rules in
1743 * the POSIX standard,
1744 * \w & ^\S = nothing
1745 * Unfortunately, some vendor locales do not meet the Posix
1746 * standard, in particular almost everything by Microsoft.
1747 * The loop below just changes e.g., \w into \W and vice versa */
1749 regnode_charclass_posixl temp;
1750 int add = 1; /* To calculate the index of the complement */
1752 Zero(&temp, 1, regnode_charclass_posixl);
1753 ANYOF_POSIXL_ZERO(&temp);
1754 for (i = 0; i < ANYOF_MAX; i++) {
1756 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1757 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1759 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1760 ANYOF_POSIXL_SET(&temp, i + add);
1762 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1764 ANYOF_POSIXL_AND(&temp, ssc);
1766 } /* else ssc already has no posixes */
1767 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1768 in its initial state */
1769 else if (! is_ANYOF_SYNTHETIC(and_with)
1770 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1772 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1773 * copy it over 'ssc' */
1774 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1775 if (is_ANYOF_SYNTHETIC(and_with)) {
1776 StructCopy(and_with, ssc, regnode_ssc);
1779 ssc->invlist = anded_cp_list;
1780 ANYOF_POSIXL_ZERO(ssc);
1781 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1782 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1786 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1787 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
1789 /* One or the other of P1, P2 is non-empty. */
1790 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1791 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1793 ssc_union(ssc, anded_cp_list, FALSE);
1795 else { /* P1 = P2 = empty */
1796 ssc_intersection(ssc, anded_cp_list, FALSE);
1802 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1803 const regnode_charclass *or_with)
1805 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1806 * another SSC or a regular ANYOF class. Can create false positives if
1807 * 'or_with' is to be inverted. */
1812 PERL_ARGS_ASSERT_SSC_OR;
1814 assert(is_ANYOF_SYNTHETIC(ssc));
1816 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1817 * the code point inversion list and just the relevant flags */
1818 if (is_ANYOF_SYNTHETIC(or_with)) {
1819 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1820 ored_flags = ANYOF_FLAGS(or_with);
1823 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1824 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1825 if (OP(or_with) != ANYOFD) {
1827 |= ANYOF_FLAGS(or_with)
1828 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1829 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1830 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(or_with))) {
1832 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1837 ANYOF_FLAGS(ssc) |= ored_flags;
1839 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1840 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1841 * 'or_with' may be inverted. When not inverted, we have the simple
1842 * situation of computing:
1843 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1844 * If P1|P2 yields a situation with both a class and its complement are
1845 * set, like having both \w and \W, this matches all code points, and we
1846 * can delete these from the P component of the ssc going forward. XXX We
1847 * might be able to delete all the P components, but I (khw) am not certain
1848 * about this, and it is better to be safe.
1851 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1852 * <= (C1 | P1) | ~C2
1853 * <= (C1 | ~C2) | P1
1854 * (which results in actually simpler code than the non-inverted case)
1857 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1858 && ! is_ANYOF_SYNTHETIC(or_with))
1860 /* We ignore P2, leaving P1 going forward */
1861 } /* else Not inverted */
1862 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
1863 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1864 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1866 for (i = 0; i < ANYOF_MAX; i += 2) {
1867 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1869 ssc_match_all_cp(ssc);
1870 ANYOF_POSIXL_CLEAR(ssc, i);
1871 ANYOF_POSIXL_CLEAR(ssc, i+1);
1879 FALSE /* Already has been inverted */
1883 PERL_STATIC_INLINE void
1884 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1886 PERL_ARGS_ASSERT_SSC_UNION;
1888 assert(is_ANYOF_SYNTHETIC(ssc));
1890 _invlist_union_maybe_complement_2nd(ssc->invlist,
1896 PERL_STATIC_INLINE void
1897 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1899 const bool invert2nd)
1901 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1903 assert(is_ANYOF_SYNTHETIC(ssc));
1905 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1911 PERL_STATIC_INLINE void
1912 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1914 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1916 assert(is_ANYOF_SYNTHETIC(ssc));
1918 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1921 PERL_STATIC_INLINE void
1922 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1924 /* AND just the single code point 'cp' into the SSC 'ssc' */
1926 SV* cp_list = _new_invlist(2);
1928 PERL_ARGS_ASSERT_SSC_CP_AND;
1930 assert(is_ANYOF_SYNTHETIC(ssc));
1932 cp_list = add_cp_to_invlist(cp_list, cp);
1933 ssc_intersection(ssc, cp_list,
1934 FALSE /* Not inverted */
1936 SvREFCNT_dec_NN(cp_list);
1939 PERL_STATIC_INLINE void
1940 S_ssc_clear_locale(regnode_ssc *ssc)
1942 /* Set the SSC 'ssc' to not match any locale things */
1943 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1945 assert(is_ANYOF_SYNTHETIC(ssc));
1947 ANYOF_POSIXL_ZERO(ssc);
1948 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1951 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1954 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1956 /* The synthetic start class is used to hopefully quickly winnow down
1957 * places where a pattern could start a match in the target string. If it
1958 * doesn't really narrow things down that much, there isn't much point to
1959 * having the overhead of using it. This function uses some very crude
1960 * heuristics to decide if to use the ssc or not.
1962 * It returns TRUE if 'ssc' rules out more than half what it considers to
1963 * be the "likely" possible matches, but of course it doesn't know what the
1964 * actual things being matched are going to be; these are only guesses
1966 * For /l matches, it assumes that the only likely matches are going to be
1967 * in the 0-255 range, uniformly distributed, so half of that is 127
1968 * For /a and /d matches, it assumes that the likely matches will be just
1969 * the ASCII range, so half of that is 63
1970 * For /u and there isn't anything matching above the Latin1 range, it
1971 * assumes that that is the only range likely to be matched, and uses
1972 * half that as the cut-off: 127. If anything matches above Latin1,
1973 * it assumes that all of Unicode could match (uniformly), except for
1974 * non-Unicode code points and things in the General Category "Other"
1975 * (unassigned, private use, surrogates, controls and formats). This
1976 * is a much large number. */
1978 U32 count = 0; /* Running total of number of code points matched by
1980 UV start, end; /* Start and end points of current range in inversion
1982 const U32 max_code_points = (LOC)
1984 : (( ! UNI_SEMANTICS
1985 || invlist_highest(ssc->invlist) < 256)
1988 const U32 max_match = max_code_points / 2;
1990 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1992 invlist_iterinit(ssc->invlist);
1993 while (invlist_iternext(ssc->invlist, &start, &end)) {
1994 if (start >= max_code_points) {
1997 end = MIN(end, max_code_points - 1);
1998 count += end - start + 1;
1999 if (count >= max_match) {
2000 invlist_iterfinish(ssc->invlist);
2010 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
2012 /* The inversion list in the SSC is marked mortal; now we need a more
2013 * permanent copy, which is stored the same way that is done in a regular
2014 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
2017 SV* invlist = invlist_clone(ssc->invlist, NULL);
2019 PERL_ARGS_ASSERT_SSC_FINALIZE;
2021 assert(is_ANYOF_SYNTHETIC(ssc));
2023 /* The code in this file assumes that all but these flags aren't relevant
2024 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
2025 * by the time we reach here */
2026 assert(! (ANYOF_FLAGS(ssc)
2027 & ~( ANYOF_COMMON_FLAGS
2028 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
2029 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
2031 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
2033 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
2034 NULL, NULL, NULL, FALSE);
2036 /* Make sure is clone-safe */
2037 ssc->invlist = NULL;
2039 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
2040 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
2043 if (RExC_contains_locale) {
2047 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
2050 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
2051 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
2052 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
2053 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
2054 ? (TRIE_LIST_CUR( idx ) - 1) \
2060 dump_trie(trie,widecharmap,revcharmap)
2061 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
2062 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
2064 These routines dump out a trie in a somewhat readable format.
2065 The _interim_ variants are used for debugging the interim
2066 tables that are used to generate the final compressed
2067 representation which is what dump_trie expects.
2069 Part of the reason for their existence is to provide a form
2070 of documentation as to how the different representations function.
2075 Dumps the final compressed table form of the trie to Perl_debug_log.
2076 Used for debugging make_trie().
2080 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
2081 AV *revcharmap, U32 depth)
2084 SV *sv=sv_newmortal();
2085 int colwidth= widecharmap ? 6 : 4;
2087 GET_RE_DEBUG_FLAGS_DECL;
2089 PERL_ARGS_ASSERT_DUMP_TRIE;
2091 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
2092 depth+1, "Match","Base","Ofs" );
2094 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2095 SV ** const tmp = av_fetch( revcharmap, state, 0);
2097 Perl_re_printf( aTHX_ "%*s",
2099 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2100 PL_colors[0], PL_colors[1],
2101 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2102 PERL_PV_ESCAPE_FIRSTCHAR
2107 Perl_re_printf( aTHX_ "\n");
2108 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
2110 for( state = 0 ; state < trie->uniquecharcount ; state++ )
2111 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2112 Perl_re_printf( aTHX_ "\n");
2114 for( state = 1 ; state < trie->statecount ; state++ ) {
2115 const U32 base = trie->states[ state ].trans.base;
2117 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
2119 if ( trie->states[ state ].wordnum ) {
2120 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
2122 Perl_re_printf( aTHX_ "%6s", "" );
2125 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
2130 while( ( base + ofs < trie->uniquecharcount ) ||
2131 ( base + ofs - trie->uniquecharcount < trie->lasttrans
2132 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2136 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
2138 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2139 if ( ( base + ofs >= trie->uniquecharcount )
2140 && ( base + ofs - trie->uniquecharcount
2142 && trie->trans[ base + ofs
2143 - trie->uniquecharcount ].check == state )
2145 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
2146 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2149 Perl_re_printf( aTHX_ "%*s",colwidth," ." );
2153 Perl_re_printf( aTHX_ "]");
2156 Perl_re_printf( aTHX_ "\n" );
2158 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
2160 for (word=1; word <= trie->wordcount; word++) {
2161 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2162 (int)word, (int)(trie->wordinfo[word].prev),
2163 (int)(trie->wordinfo[word].len));
2165 Perl_re_printf( aTHX_ "\n" );
2168 Dumps a fully constructed but uncompressed trie in list form.
2169 List tries normally only are used for construction when the number of
2170 possible chars (trie->uniquecharcount) is very high.
2171 Used for debugging make_trie().
2174 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2175 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2179 SV *sv=sv_newmortal();
2180 int colwidth= widecharmap ? 6 : 4;
2181 GET_RE_DEBUG_FLAGS_DECL;
2183 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2185 /* print out the table precompression. */
2186 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
2188 Perl_re_indentf( aTHX_ "%s",
2189 depth+1, "------:-----+-----------------\n" );
2191 for( state=1 ; state < next_alloc ; state ++ ) {
2194 Perl_re_indentf( aTHX_ " %4" UVXf " :",
2195 depth+1, (UV)state );
2196 if ( ! trie->states[ state ].wordnum ) {
2197 Perl_re_printf( aTHX_ "%5s| ","");
2199 Perl_re_printf( aTHX_ "W%4x| ",
2200 trie->states[ state ].wordnum
2203 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2204 SV ** const tmp = av_fetch( revcharmap,
2205 TRIE_LIST_ITEM(state,charid).forid, 0);
2207 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
2209 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2211 PL_colors[0], PL_colors[1],
2212 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2213 | PERL_PV_ESCAPE_FIRSTCHAR
2215 TRIE_LIST_ITEM(state,charid).forid,
2216 (UV)TRIE_LIST_ITEM(state,charid).newstate
2219 Perl_re_printf( aTHX_ "\n%*s| ",
2220 (int)((depth * 2) + 14), "");
2223 Perl_re_printf( aTHX_ "\n");
2228 Dumps a fully constructed but uncompressed trie in table form.
2229 This is the normal DFA style state transition table, with a few
2230 twists to facilitate compression later.
2231 Used for debugging make_trie().
2234 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2235 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2240 SV *sv=sv_newmortal();
2241 int colwidth= widecharmap ? 6 : 4;
2242 GET_RE_DEBUG_FLAGS_DECL;
2244 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2247 print out the table precompression so that we can do a visual check
2248 that they are identical.
2251 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
2253 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2254 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2256 Perl_re_printf( aTHX_ "%*s",
2258 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2259 PL_colors[0], PL_colors[1],
2260 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2261 PERL_PV_ESCAPE_FIRSTCHAR
2267 Perl_re_printf( aTHX_ "\n");
2268 Perl_re_indentf( aTHX_ "State+-", depth+1 );
2270 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2271 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
2274 Perl_re_printf( aTHX_ "\n" );
2276 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2278 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
2280 (UV)TRIE_NODENUM( state ) );
2282 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2283 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2285 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
2287 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
2289 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2290 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
2291 (UV)trie->trans[ state ].check );
2293 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
2294 (UV)trie->trans[ state ].check,
2295 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2303 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2304 startbranch: the first branch in the whole branch sequence
2305 first : start branch of sequence of branch-exact nodes.
2306 May be the same as startbranch
2307 last : Thing following the last branch.
2308 May be the same as tail.
2309 tail : item following the branch sequence
2310 count : words in the sequence
2311 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2312 depth : indent depth
2314 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2316 A trie is an N'ary tree where the branches are determined by digital
2317 decomposition of the key. IE, at the root node you look up the 1st character and
2318 follow that branch repeat until you find the end of the branches. Nodes can be
2319 marked as "accepting" meaning they represent a complete word. Eg:
2323 would convert into the following structure. Numbers represent states, letters
2324 following numbers represent valid transitions on the letter from that state, if
2325 the number is in square brackets it represents an accepting state, otherwise it
2326 will be in parenthesis.
2328 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2332 (1) +-i->(6)-+-s->[7]
2334 +-s->(3)-+-h->(4)-+-e->[5]
2336 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2338 This shows that when matching against the string 'hers' we will begin at state 1
2339 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2340 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2341 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2342 single traverse. We store a mapping from accepting to state to which word was
2343 matched, and then when we have multiple possibilities we try to complete the
2344 rest of the regex in the order in which they occurred in the alternation.
2346 The only prior NFA like behaviour that would be changed by the TRIE support is
2347 the silent ignoring of duplicate alternations which are of the form:
2349 / (DUPE|DUPE) X? (?{ ... }) Y /x
2351 Thus EVAL blocks following a trie may be called a different number of times with
2352 and without the optimisation. With the optimisations dupes will be silently
2353 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2354 the following demonstrates:
2356 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2358 which prints out 'word' three times, but
2360 'words'=~/(word|word|word)(?{ print $1 })S/
2362 which doesnt print it out at all. This is due to other optimisations kicking in.
2364 Example of what happens on a structural level:
2366 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2368 1: CURLYM[1] {1,32767}(18)
2379 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2380 and should turn into:
2382 1: CURLYM[1] {1,32767}(18)
2384 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2392 Cases where tail != last would be like /(?foo|bar)baz/:
2402 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2403 and would end up looking like:
2406 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2413 d = uvchr_to_utf8_flags(d, uv, 0);
2415 is the recommended Unicode-aware way of saying
2420 #define TRIE_STORE_REVCHAR(val) \
2423 SV *zlopp = newSV(UTF8_MAXBYTES); \
2424 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2425 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2426 SvCUR_set(zlopp, kapow - flrbbbbb); \
2429 av_push(revcharmap, zlopp); \
2431 char ooooff = (char)val; \
2432 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2436 /* This gets the next character from the input, folding it if not already
2438 #define TRIE_READ_CHAR STMT_START { \
2441 /* if it is UTF then it is either already folded, or does not need \
2443 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2445 else if (folder == PL_fold_latin1) { \
2446 /* This folder implies Unicode rules, which in the range expressible \
2447 * by not UTF is the lower case, with the two exceptions, one of \
2448 * which should have been taken care of before calling this */ \
2449 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2450 uvc = toLOWER_L1(*uc); \
2451 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2454 /* raw data, will be folded later if needed */ \
2462 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2463 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2464 U32 ging = TRIE_LIST_LEN( state ) * 2; \
2465 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2466 TRIE_LIST_LEN( state ) = ging; \
2468 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2469 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2470 TRIE_LIST_CUR( state )++; \
2473 #define TRIE_LIST_NEW(state) STMT_START { \
2474 Newx( trie->states[ state ].trans.list, \
2475 4, reg_trie_trans_le ); \
2476 TRIE_LIST_CUR( state ) = 1; \
2477 TRIE_LIST_LEN( state ) = 4; \
2480 #define TRIE_HANDLE_WORD(state) STMT_START { \
2481 U16 dupe= trie->states[ state ].wordnum; \
2482 regnode * const noper_next = regnext( noper ); \
2485 /* store the word for dumping */ \
2487 if (OP(noper) != NOTHING) \
2488 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2490 tmp = newSVpvn_utf8( "", 0, UTF ); \
2491 av_push( trie_words, tmp ); \
2495 trie->wordinfo[curword].prev = 0; \
2496 trie->wordinfo[curword].len = wordlen; \
2497 trie->wordinfo[curword].accept = state; \
2499 if ( noper_next < tail ) { \
2501 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2503 trie->jump[curword] = (U16)(noper_next - convert); \
2505 jumper = noper_next; \
2507 nextbranch= regnext(cur); \
2511 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2512 /* chain, so that when the bits of chain are later */\
2513 /* linked together, the dups appear in the chain */\
2514 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2515 trie->wordinfo[dupe].prev = curword; \
2517 /* we haven't inserted this word yet. */ \
2518 trie->states[ state ].wordnum = curword; \
2523 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2524 ( ( base + charid >= ucharcount \
2525 && base + charid < ubound \
2526 && state == trie->trans[ base - ucharcount + charid ].check \
2527 && trie->trans[ base - ucharcount + charid ].next ) \
2528 ? trie->trans[ base - ucharcount + charid ].next \
2529 : ( state==1 ? special : 0 ) \
2532 #define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2534 TRIE_BITMAP_SET(trie, uvc); \
2535 /* store the folded codepoint */ \
2537 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2540 /* store first byte of utf8 representation of */ \
2541 /* variant codepoints */ \
2542 if (! UVCHR_IS_INVARIANT(uvc)) { \
2543 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2548 #define MADE_JUMP_TRIE 2
2549 #define MADE_EXACT_TRIE 4
2552 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2553 regnode *first, regnode *last, regnode *tail,
2554 U32 word_count, U32 flags, U32 depth)
2556 /* first pass, loop through and scan words */
2557 reg_trie_data *trie;
2558 HV *widecharmap = NULL;
2559 AV *revcharmap = newAV();
2565 regnode *jumper = NULL;
2566 regnode *nextbranch = NULL;
2567 regnode *convert = NULL;
2568 U32 *prev_states; /* temp array mapping each state to previous one */
2569 /* we just use folder as a flag in utf8 */
2570 const U8 * folder = NULL;
2572 /* in the below add_data call we are storing either 'tu' or 'tuaa'
2573 * which stands for one trie structure, one hash, optionally followed
2576 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuaa"));
2577 AV *trie_words = NULL;
2578 /* along with revcharmap, this only used during construction but both are
2579 * useful during debugging so we store them in the struct when debugging.
2582 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2583 STRLEN trie_charcount=0;
2585 SV *re_trie_maxbuff;
2586 GET_RE_DEBUG_FLAGS_DECL;
2588 PERL_ARGS_ASSERT_MAKE_TRIE;
2590 PERL_UNUSED_ARG(depth);
2594 case EXACT: case EXACTL: break;
2598 case EXACTFLU8: folder = PL_fold_latin1; break;
2599 case EXACTF: folder = PL_fold; break;
2600 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2603 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2605 trie->startstate = 1;
2606 trie->wordcount = word_count;
2607 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2608 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2609 if (flags == EXACT || flags == EXACTL)
2610 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2611 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2612 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2615 trie_words = newAV();
2618 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2619 assert(re_trie_maxbuff);
2620 if (!SvIOK(re_trie_maxbuff)) {
2621 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2623 DEBUG_TRIE_COMPILE_r({
2624 Perl_re_indentf( aTHX_
2625 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2627 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2628 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2631 /* Find the node we are going to overwrite */
2632 if ( first == startbranch && OP( last ) != BRANCH ) {
2633 /* whole branch chain */
2636 /* branch sub-chain */
2637 convert = NEXTOPER( first );
2640 /* -- First loop and Setup --
2642 We first traverse the branches and scan each word to determine if it
2643 contains widechars, and how many unique chars there are, this is
2644 important as we have to build a table with at least as many columns as we
2647 We use an array of integers to represent the character codes 0..255
2648 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2649 the native representation of the character value as the key and IV's for
2652 *TODO* If we keep track of how many times each character is used we can
2653 remap the columns so that the table compression later on is more
2654 efficient in terms of memory by ensuring the most common value is in the
2655 middle and the least common are on the outside. IMO this would be better
2656 than a most to least common mapping as theres a decent chance the most
2657 common letter will share a node with the least common, meaning the node
2658 will not be compressible. With a middle is most common approach the worst
2659 case is when we have the least common nodes twice.
2663 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2664 regnode *noper = NEXTOPER( cur );
2668 U32 wordlen = 0; /* required init */
2669 STRLEN minchars = 0;
2670 STRLEN maxchars = 0;
2671 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2674 if (OP(noper) == NOTHING) {
2675 /* skip past a NOTHING at the start of an alternation
2676 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2678 regnode *noper_next= regnext(noper);
2679 if (noper_next < tail)
2683 if ( noper < tail &&
2685 OP(noper) == flags ||
2688 OP(noper) == EXACTFU_SS
2692 uc= (U8*)STRING(noper);
2693 e= uc + STR_LEN(noper);
2700 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2701 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2702 regardless of encoding */
2703 if (OP( noper ) == EXACTFU_SS) {
2704 /* false positives are ok, so just set this */
2705 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2709 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2711 TRIE_CHARCOUNT(trie)++;
2714 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2715 * is in effect. Under /i, this character can match itself, or
2716 * anything that folds to it. If not under /i, it can match just
2717 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2718 * all fold to k, and all are single characters. But some folds
2719 * expand to more than one character, so for example LATIN SMALL
2720 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2721 * the string beginning at 'uc' is 'ffi', it could be matched by
2722 * three characters, or just by the one ligature character. (It
2723 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2724 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2725 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2726 * match.) The trie needs to know the minimum and maximum number
2727 * of characters that could match so that it can use size alone to
2728 * quickly reject many match attempts. The max is simple: it is
2729 * the number of folded characters in this branch (since a fold is
2730 * never shorter than what folds to it. */
2734 /* And the min is equal to the max if not under /i (indicated by
2735 * 'folder' being NULL), or there are no multi-character folds. If
2736 * there is a multi-character fold, the min is incremented just
2737 * once, for the character that folds to the sequence. Each
2738 * character in the sequence needs to be added to the list below of
2739 * characters in the trie, but we count only the first towards the
2740 * min number of characters needed. This is done through the
2741 * variable 'foldlen', which is returned by the macros that look
2742 * for these sequences as the number of bytes the sequence
2743 * occupies. Each time through the loop, we decrement 'foldlen' by
2744 * how many bytes the current char occupies. Only when it reaches
2745 * 0 do we increment 'minchars' or look for another multi-character
2747 if (folder == NULL) {
2750 else if (foldlen > 0) {
2751 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2756 /* See if *uc is the beginning of a multi-character fold. If
2757 * so, we decrement the length remaining to look at, to account
2758 * for the current character this iteration. (We can use 'uc'
2759 * instead of the fold returned by TRIE_READ_CHAR because for
2760 * non-UTF, the latin1_safe macro is smart enough to account
2761 * for all the unfolded characters, and because for UTF, the
2762 * string will already have been folded earlier in the
2763 * compilation process */
2765 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2766 foldlen -= UTF8SKIP(uc);
2769 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2774 /* The current character (and any potential folds) should be added
2775 * to the possible matching characters for this position in this
2779 U8 folded= folder[ (U8) uvc ];
2780 if ( !trie->charmap[ folded ] ) {
2781 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2782 TRIE_STORE_REVCHAR( folded );
2785 if ( !trie->charmap[ uvc ] ) {
2786 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2787 TRIE_STORE_REVCHAR( uvc );
2790 /* store the codepoint in the bitmap, and its folded
2792 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
2793 set_bit = 0; /* We've done our bit :-) */
2797 /* XXX We could come up with the list of code points that fold
2798 * to this using PL_utf8_foldclosures, except not for
2799 * multi-char folds, as there may be multiple combinations
2800 * there that could work, which needs to wait until runtime to
2801 * resolve (The comment about LIGATURE FFI above is such an
2806 widecharmap = newHV();
2808 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2811 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
2813 if ( !SvTRUE( *svpp ) ) {
2814 sv_setiv( *svpp, ++trie->uniquecharcount );
2815 TRIE_STORE_REVCHAR(uvc);
2818 } /* end loop through characters in this branch of the trie */
2820 /* We take the min and max for this branch and combine to find the min
2821 * and max for all branches processed so far */
2822 if( cur == first ) {
2823 trie->minlen = minchars;
2824 trie->maxlen = maxchars;
2825 } else if (minchars < trie->minlen) {
2826 trie->minlen = minchars;
2827 } else if (maxchars > trie->maxlen) {
2828 trie->maxlen = maxchars;
2830 } /* end first pass */
2831 DEBUG_TRIE_COMPILE_r(
2832 Perl_re_indentf( aTHX_
2833 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2835 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2836 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2837 (int)trie->minlen, (int)trie->maxlen )
2841 We now know what we are dealing with in terms of unique chars and
2842 string sizes so we can calculate how much memory a naive
2843 representation using a flat table will take. If it's over a reasonable
2844 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2845 conservative but potentially much slower representation using an array
2848 At the end we convert both representations into the same compressed
2849 form that will be used in regexec.c for matching with. The latter
2850 is a form that cannot be used to construct with but has memory
2851 properties similar to the list form and access properties similar
2852 to the table form making it both suitable for fast searches and
2853 small enough that its feasable to store for the duration of a program.
2855 See the comment in the code where the compressed table is produced
2856 inplace from the flat tabe representation for an explanation of how
2857 the compression works.
2862 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2865 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2866 > SvIV(re_trie_maxbuff) )
2869 Second Pass -- Array Of Lists Representation
2871 Each state will be represented by a list of charid:state records
2872 (reg_trie_trans_le) the first such element holds the CUR and LEN
2873 points of the allocated array. (See defines above).
2875 We build the initial structure using the lists, and then convert
2876 it into the compressed table form which allows faster lookups
2877 (but cant be modified once converted).
2880 STRLEN transcount = 1;
2882 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
2885 trie->states = (reg_trie_state *)
2886 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2887 sizeof(reg_trie_state) );
2891 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2893 regnode *noper = NEXTOPER( cur );
2894 U32 state = 1; /* required init */
2895 U16 charid = 0; /* sanity init */
2896 U32 wordlen = 0; /* required init */
2898 if (OP(noper) == NOTHING) {
2899 regnode *noper_next= regnext(noper);
2900 if (noper_next < tail)
2904 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
2905 const U8 *uc= (U8*)STRING(noper);
2906 const U8 *e= uc + STR_LEN(noper);
2908 for ( ; uc < e ; uc += len ) {
2913 charid = trie->charmap[ uvc ];
2915 SV** const svpp = hv_fetch( widecharmap,
2922 charid=(U16)SvIV( *svpp );
2925 /* charid is now 0 if we dont know the char read, or
2926 * nonzero if we do */
2933 if ( !trie->states[ state ].trans.list ) {
2934 TRIE_LIST_NEW( state );
2937 check <= TRIE_LIST_USED( state );
2940 if ( TRIE_LIST_ITEM( state, check ).forid
2943 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2948 newstate = next_alloc++;
2949 prev_states[newstate] = state;
2950 TRIE_LIST_PUSH( state, charid, newstate );
2955 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
2959 TRIE_HANDLE_WORD(state);
2961 } /* end second pass */
2963 /* next alloc is the NEXT state to be allocated */
2964 trie->statecount = next_alloc;
2965 trie->states = (reg_trie_state *)
2966 PerlMemShared_realloc( trie->states,
2968 * sizeof(reg_trie_state) );
2970 /* and now dump it out before we compress it */
2971 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2972 revcharmap, next_alloc,
2976 trie->trans = (reg_trie_trans *)
2977 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2984 for( state=1 ; state < next_alloc ; state ++ ) {
2988 DEBUG_TRIE_COMPILE_MORE_r(
2989 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
2993 if (trie->states[state].trans.list) {
2994 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2998 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2999 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
3000 if ( forid < minid ) {
3002 } else if ( forid > maxid ) {
3006 if ( transcount < tp + maxid - minid + 1) {
3008 trie->trans = (reg_trie_trans *)
3009 PerlMemShared_realloc( trie->trans,
3011 * sizeof(reg_trie_trans) );
3012 Zero( trie->trans + (transcount / 2),
3016 base = trie->uniquecharcount + tp - minid;
3017 if ( maxid == minid ) {
3019 for ( ; zp < tp ; zp++ ) {
3020 if ( ! trie->trans[ zp ].next ) {
3021 base = trie->uniquecharcount + zp - minid;
3022 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
3024 trie->trans[ zp ].check = state;
3030 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
3032 trie->trans[ tp ].check = state;
3037 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3038 const U32 tid = base
3039 - trie->uniquecharcount
3040 + TRIE_LIST_ITEM( state, idx ).forid;
3041 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
3043 trie->trans[ tid ].check = state;
3045 tp += ( maxid - minid + 1 );
3047 Safefree(trie->states[ state ].trans.list);
3050 DEBUG_TRIE_COMPILE_MORE_r(
3051 Perl_re_printf( aTHX_ " base: %d\n",base);
3054 trie->states[ state ].trans.base=base;
3056 trie->lasttrans = tp + 1;
3060 Second Pass -- Flat Table Representation.
3062 we dont use the 0 slot of either trans[] or states[] so we add 1 to
3063 each. We know that we will need Charcount+1 trans at most to store
3064 the data (one row per char at worst case) So we preallocate both
3065 structures assuming worst case.
3067 We then construct the trie using only the .next slots of the entry
3070 We use the .check field of the first entry of the node temporarily
3071 to make compression both faster and easier by keeping track of how
3072 many non zero fields are in the node.
3074 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
3077 There are two terms at use here: state as a TRIE_NODEIDX() which is
3078 a number representing the first entry of the node, and state as a
3079 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
3080 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
3081 if there are 2 entrys per node. eg:
3089 The table is internally in the right hand, idx form. However as we
3090 also have to deal with the states array which is indexed by nodenum
3091 we have to use TRIE_NODENUM() to convert.
3094 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
3097 trie->trans = (reg_trie_trans *)
3098 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
3099 * trie->uniquecharcount + 1,
3100 sizeof(reg_trie_trans) );
3101 trie->states = (reg_trie_state *)
3102 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3103 sizeof(reg_trie_state) );
3104 next_alloc = trie->uniquecharcount + 1;
3107 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3109 regnode *noper = NEXTOPER( cur );
3111 U32 state = 1; /* required init */
3113 U16 charid = 0; /* sanity init */
3114 U32 accept_state = 0; /* sanity init */
3116 U32 wordlen = 0; /* required init */
3118 if (OP(noper) == NOTHING) {
3119 regnode *noper_next= regnext(noper);
3120 if (noper_next < tail)
3124 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
3125 const U8 *uc= (U8*)STRING(noper);
3126 const U8 *e= uc + STR_LEN(noper);
3128 for ( ; uc < e ; uc += len ) {
3133 charid = trie->charmap[ uvc ];
3135 SV* const * const svpp = hv_fetch( widecharmap,
3139 charid = svpp ? (U16)SvIV(*svpp) : 0;
3143 if ( !trie->trans[ state + charid ].next ) {
3144 trie->trans[ state + charid ].next = next_alloc;
3145 trie->trans[ state ].check++;
3146 prev_states[TRIE_NODENUM(next_alloc)]
3147 = TRIE_NODENUM(state);
3148 next_alloc += trie->uniquecharcount;
3150 state = trie->trans[ state + charid ].next;
3152 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3154 /* charid is now 0 if we dont know the char read, or
3155 * nonzero if we do */
3158 accept_state = TRIE_NODENUM( state );
3159 TRIE_HANDLE_WORD(accept_state);
3161 } /* end second pass */
3163 /* and now dump it out before we compress it */
3164 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3166 next_alloc, depth+1));
3170 * Inplace compress the table.*
3172 For sparse data sets the table constructed by the trie algorithm will
3173 be mostly 0/FAIL transitions or to put it another way mostly empty.
3174 (Note that leaf nodes will not contain any transitions.)
3176 This algorithm compresses the tables by eliminating most such
3177 transitions, at the cost of a modest bit of extra work during lookup:
3179 - Each states[] entry contains a .base field which indicates the
3180 index in the state[] array wheres its transition data is stored.
3182 - If .base is 0 there are no valid transitions from that node.
3184 - If .base is nonzero then charid is added to it to find an entry in
3187 -If trans[states[state].base+charid].check!=state then the
3188 transition is taken to be a 0/Fail transition. Thus if there are fail
3189 transitions at the front of the node then the .base offset will point
3190 somewhere inside the previous nodes data (or maybe even into a node
3191 even earlier), but the .check field determines if the transition is
3195 The following process inplace converts the table to the compressed
3196 table: We first do not compress the root node 1,and mark all its
3197 .check pointers as 1 and set its .base pointer as 1 as well. This
3198 allows us to do a DFA construction from the compressed table later,
3199 and ensures that any .base pointers we calculate later are greater
3202 - We set 'pos' to indicate the first entry of the second node.
3204 - We then iterate over the columns of the node, finding the first and
3205 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3206 and set the .check pointers accordingly, and advance pos
3207 appropriately and repreat for the next node. Note that when we copy
3208 the next pointers we have to convert them from the original
3209 NODEIDX form to NODENUM form as the former is not valid post
3212 - If a node has no transitions used we mark its base as 0 and do not
3213 advance the pos pointer.
3215 - If a node only has one transition we use a second pointer into the
3216 structure to fill in allocated fail transitions from other states.
3217 This pointer is independent of the main pointer and scans forward
3218 looking for null transitions that are allocated to a state. When it
3219 finds one it writes the single transition into the "hole". If the
3220 pointer doesnt find one the single transition is appended as normal.
3222 - Once compressed we can Renew/realloc the structures to release the
3225 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3226 specifically Fig 3.47 and the associated pseudocode.
3230 const U32 laststate = TRIE_NODENUM( next_alloc );
3233 trie->statecount = laststate;
3235 for ( state = 1 ; state < laststate ; state++ ) {
3237 const U32 stateidx = TRIE_NODEIDX( state );
3238 const U32 o_used = trie->trans[ stateidx ].check;
3239 U32 used = trie->trans[ stateidx ].check;
3240 trie->trans[ stateidx ].check = 0;
3243 used && charid < trie->uniquecharcount;
3246 if ( flag || trie->trans[ stateidx + charid ].next ) {
3247 if ( trie->trans[ stateidx + charid ].next ) {
3249 for ( ; zp < pos ; zp++ ) {
3250 if ( ! trie->trans[ zp ].next ) {
3254 trie->states[ state ].trans.base
3256 + trie->uniquecharcount
3258 trie->trans[ zp ].next
3259 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3261 trie->trans[ zp ].check = state;
3262 if ( ++zp > pos ) pos = zp;
3269 trie->states[ state ].trans.base
3270 = pos + trie->uniquecharcount - charid ;
3272 trie->trans[ pos ].next
3273 = SAFE_TRIE_NODENUM(
3274 trie->trans[ stateidx + charid ].next );
3275 trie->trans[ pos ].check = state;
3280 trie->lasttrans = pos + 1;
3281 trie->states = (reg_trie_state *)
3282 PerlMemShared_realloc( trie->states, laststate
3283 * sizeof(reg_trie_state) );
3284 DEBUG_TRIE_COMPILE_MORE_r(
3285 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
3287 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3291 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3294 } /* end table compress */
3296 DEBUG_TRIE_COMPILE_MORE_r(
3297 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
3299 (UV)trie->statecount,
3300 (UV)trie->lasttrans)
3302 /* resize the trans array to remove unused space */
3303 trie->trans = (reg_trie_trans *)
3304 PerlMemShared_realloc( trie->trans, trie->lasttrans
3305 * sizeof(reg_trie_trans) );
3307 { /* Modify the program and insert the new TRIE node */
3308 U8 nodetype =(U8)(flags & 0xFF);
3312 regnode *optimize = NULL;
3313 #ifdef RE_TRACK_PATTERN_OFFSETS
3316 U32 mjd_nodelen = 0;
3317 #endif /* RE_TRACK_PATTERN_OFFSETS */
3318 #endif /* DEBUGGING */
3320 This means we convert either the first branch or the first Exact,
3321 depending on whether the thing following (in 'last') is a branch
3322 or not and whther first is the startbranch (ie is it a sub part of
3323 the alternation or is it the whole thing.)
3324 Assuming its a sub part we convert the EXACT otherwise we convert
3325 the whole branch sequence, including the first.
3327 /* Find the node we are going to overwrite */
3328 if ( first != startbranch || OP( last ) == BRANCH ) {
3329 /* branch sub-chain */
3330 NEXT_OFF( first ) = (U16)(last - first);
3331 #ifdef RE_TRACK_PATTERN_OFFSETS
3333 mjd_offset= Node_Offset((convert));
3334 mjd_nodelen= Node_Length((convert));
3337 /* whole branch chain */
3339 #ifdef RE_TRACK_PATTERN_OFFSETS
3342 const regnode *nop = NEXTOPER( convert );
3343 mjd_offset= Node_Offset((nop));
3344 mjd_nodelen= Node_Length((nop));
3348 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
3350 (UV)mjd_offset, (UV)mjd_nodelen)
3353 /* But first we check to see if there is a common prefix we can
3354 split out as an EXACT and put in front of the TRIE node. */
3355 trie->startstate= 1;
3356 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3357 /* we want to find the first state that has more than
3358 * one transition, if that state is not the first state
3359 * then we have a common prefix which we can remove.
3362 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3364 I32 first_ofs = -1; /* keeps track of the ofs of the first
3365 transition, -1 means none */
3367 const U32 base = trie->states[ state ].trans.base;
3369 /* does this state terminate an alternation? */
3370 if ( trie->states[state].wordnum )
3373 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3374 if ( ( base + ofs >= trie->uniquecharcount ) &&
3375 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3376 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3378 if ( ++count > 1 ) {
3379 /* we have more than one transition */
3382 /* if this is the first state there is no common prefix
3383 * to extract, so we can exit */
3384 if ( state == 1 ) break;
3385 tmp = av_fetch( revcharmap, ofs, 0);
3386 ch = (U8*)SvPV_nolen_const( *tmp );
3388 /* if we are on count 2 then we need to initialize the
3389 * bitmap, and store the previous char if there was one
3392 /* clear the bitmap */
3393 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3395 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
3398 if (first_ofs >= 0) {
3399 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
3400 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3402 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
3404 Perl_re_printf( aTHX_ "%s", (char*)ch)
3408 /* store the current firstchar in the bitmap */
3409 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
3410 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
3416 /* This state has only one transition, its transition is part
3417 * of a common prefix - we need to concatenate the char it
3418 * represents to what we have so far. */
3419 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
3421 char *ch = SvPV( *tmp, len );
3423 SV *sv=sv_newmortal();
3424 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
3426 (UV)state, (UV)first_ofs,
3427 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3428 PL_colors[0], PL_colors[1],
3429 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3430 PERL_PV_ESCAPE_FIRSTCHAR
3435 OP( convert ) = nodetype;
3436 str=STRING(convert);
3439 STR_LEN(convert) += len;
3445 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
3450 trie->prefixlen = (state-1);
3452 regnode *n = convert+NODE_SZ_STR(convert);
3453 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3454 trie->startstate = state;
3455 trie->minlen -= (state - 1);
3456 trie->maxlen -= (state - 1);
3458 /* At least the UNICOS C compiler choked on this
3459 * being argument to DEBUG_r(), so let's just have
3462 #ifdef PERL_EXT_RE_BUILD
3468 regnode *fix = convert;
3469 U32 word = trie->wordcount;
3470 #ifdef RE_TRACK_PATTERN_OFFSETS
3473 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3474 while( ++fix < n ) {
3475 Set_Node_Offset_Length(fix, 0, 0);
3478 SV ** const tmp = av_fetch( trie_words, word, 0 );
3480 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3481 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3483 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3491 NEXT_OFF(convert) = (U16)(tail - convert);
3492 DEBUG_r(optimize= n);
3498 if ( trie->maxlen ) {
3499 NEXT_OFF( convert ) = (U16)(tail - convert);
3500 ARG_SET( convert, data_slot );
3501 /* Store the offset to the first unabsorbed branch in
3502 jump[0], which is otherwise unused by the jump logic.
3503 We use this when dumping a trie and during optimisation. */
3505 trie->jump[0] = (U16)(nextbranch - convert);
3507 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3508 * and there is a bitmap
3509 * and the first "jump target" node we found leaves enough room
3510 * then convert the TRIE node into a TRIEC node, with the bitmap
3511 * embedded inline in the opcode - this is hypothetically faster.
3513 if ( !trie->states[trie->startstate].wordnum
3515 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3517 OP( convert ) = TRIEC;
3518 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3519 PerlMemShared_free(trie->bitmap);
3522 OP( convert ) = TRIE;
3524 /* store the type in the flags */
3525 convert->flags = nodetype;
3529 + regarglen[ OP( convert ) ];
3531 /* XXX We really should free up the resource in trie now,
3532 as we won't use them - (which resources?) dmq */
3534 /* needed for dumping*/
3535 DEBUG_r(if (optimize) {
3536 regnode *opt = convert;
3538 while ( ++opt < optimize) {
3539 Set_Node_Offset_Length(opt,0,0);
3542 Try to clean up some of the debris left after the
3545 while( optimize < jumper ) {
3546 #ifdef RE_TRACK_PATTERN_OFFSETS
3547 mjd_nodelen += Node_Length((optimize));
3549 OP( optimize ) = OPTIMIZED;
3550 Set_Node_Offset_Length(optimize,0,0);
3553 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
3555 } /* end node insert */
3557 /* Finish populating the prev field of the wordinfo array. Walk back
3558 * from each accept state until we find another accept state, and if
3559 * so, point the first word's .prev field at the second word. If the
3560 * second already has a .prev field set, stop now. This will be the
3561 * case either if we've already processed that word's accept state,
3562 * or that state had multiple words, and the overspill words were
3563 * already linked up earlier.
3570 for (word=1; word <= trie->wordcount; word++) {
3572 if (trie->wordinfo[word].prev)
3574 state = trie->wordinfo[word].accept;
3576 state = prev_states[state];
3579 prev = trie->states[state].wordnum;
3583 trie->wordinfo[word].prev = prev;
3585 Safefree(prev_states);
3589 /* and now dump out the compressed format */
3590 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3592 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3594 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3595 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3597 SvREFCNT_dec_NN(revcharmap);
3601 : trie->startstate>1
3607 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3609 /* The Trie is constructed and compressed now so we can build a fail array if
3612 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3614 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3618 We find the fail state for each state in the trie, this state is the longest
3619 proper suffix of the current state's 'word' that is also a proper prefix of
3620 another word in our trie. State 1 represents the word '' and is thus the
3621 default fail state. This allows the DFA not to have to restart after its
3622 tried and failed a word at a given point, it simply continues as though it
3623 had been matching the other word in the first place.
3625 'abcdgu'=~/abcdefg|cdgu/
3626 When we get to 'd' we are still matching the first word, we would encounter
3627 'g' which would fail, which would bring us to the state representing 'd' in
3628 the second word where we would try 'g' and succeed, proceeding to match
3631 /* add a fail transition */
3632 const U32 trie_offset = ARG(source);
3633 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3635 const U32 ucharcount = trie->uniquecharcount;
3636 const U32 numstates = trie->statecount;
3637 const U32 ubound = trie->lasttrans + ucharcount;
3641 U32 base = trie->states[ 1 ].trans.base;
3644 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3646 GET_RE_DEBUG_FLAGS_DECL;
3648 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3649 PERL_UNUSED_CONTEXT;
3651 PERL_UNUSED_ARG(depth);
3654 if ( OP(source) == TRIE ) {
3655 struct regnode_1 *op = (struct regnode_1 *)
3656 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3657 StructCopy(source,op,struct regnode_1);
3658 stclass = (regnode *)op;
3660 struct regnode_charclass *op = (struct regnode_charclass *)
3661 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3662 StructCopy(source,op,struct regnode_charclass);
3663 stclass = (regnode *)op;
3665 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3667 ARG_SET( stclass, data_slot );
3668 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3669 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3670 aho->trie=trie_offset;
3671 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3672 Copy( trie->states, aho->states, numstates, reg_trie_state );
3673 Newx( q, numstates, U32);
3674 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3677 /* initialize fail[0..1] to be 1 so that we always have
3678 a valid final fail state */
3679 fail[ 0 ] = fail[ 1 ] = 1;
3681 for ( charid = 0; charid < ucharcount ; charid++ ) {
3682 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3684 q[ q_write ] = newstate;
3685 /* set to point at the root */
3686 fail[ q[ q_write++ ] ]=1;
3689 while ( q_read < q_write) {
3690 const U32 cur = q[ q_read++ % numstates ];
3691 base = trie->states[ cur ].trans.base;
3693 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3694 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3696 U32 fail_state = cur;
3699 fail_state = fail[ fail_state ];
3700 fail_base = aho->states[ fail_state ].trans.base;
3701 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3703 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3704 fail[ ch_state ] = fail_state;
3705 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3707 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3709 q[ q_write++ % numstates] = ch_state;
3713 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3714 when we fail in state 1, this allows us to use the
3715 charclass scan to find a valid start char. This is based on the principle
3716 that theres a good chance the string being searched contains lots of stuff
3717 that cant be a start char.
3719 fail[ 0 ] = fail[ 1 ] = 0;
3720 DEBUG_TRIE_COMPILE_r({
3721 Perl_re_indentf( aTHX_ "Stclass Failtable (%" UVuf " states): 0",
3722 depth, (UV)numstates
3724 for( q_read=1; q_read<numstates; q_read++ ) {
3725 Perl_re_printf( aTHX_ ", %" UVuf, (UV)fail[q_read]);
3727 Perl_re_printf( aTHX_ "\n");
3730 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3735 /* The below joins as many adjacent EXACTish nodes as possible into a single
3736 * one. The regop may be changed if the node(s) contain certain sequences that
3737 * require special handling. The joining is only done if:
3738 * 1) there is room in the current conglomerated node to entirely contain the
3740 * 2) they are the exact same node type
3742 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3743 * these get optimized out
3745 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3746 * as possible, even if that means splitting an existing node so that its first
3747 * part is moved to the preceeding node. This would maximise the efficiency of
3748 * memEQ during matching.
3750 * If a node is to match under /i (folded), the number of characters it matches
3751 * can be different than its character length if it contains a multi-character
3752 * fold. *min_subtract is set to the total delta number of characters of the
3755 * And *unfolded_multi_char is set to indicate whether or not the node contains
3756 * an unfolded multi-char fold. This happens when it won't be known until
3757 * runtime whether the fold is valid or not; namely
3758 * 1) for EXACTF nodes that contain LATIN SMALL LETTER SHARP S, as only if the
3759 * target string being matched against turns out to be UTF-8 is that fold
3761 * 2) for EXACTFL nodes whose folding rules depend on the locale in force at
3763 * (Multi-char folds whose components are all above the Latin1 range are not
3764 * run-time locale dependent, and have already been folded by the time this
3765 * function is called.)
3767 * This is as good a place as any to discuss the design of handling these
3768 * multi-character fold sequences. It's been wrong in Perl for a very long
3769 * time. There are three code points in Unicode whose multi-character folds
3770 * were long ago discovered to mess things up. The previous designs for
3771 * dealing with these involved assigning a special node for them. This
3772 * approach doesn't always work, as evidenced by this example:
3773 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3774 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3775 * would match just the \xDF, it won't be able to handle the case where a
3776 * successful match would have to cross the node's boundary. The new approach
3777 * that hopefully generally solves the problem generates an EXACTFU_SS node
3778 * that is "sss" in this case.
3780 * It turns out that there are problems with all multi-character folds, and not
3781 * just these three. Now the code is general, for all such cases. The
3782 * approach taken is:
3783 * 1) This routine examines each EXACTFish node that could contain multi-
3784 * character folded sequences. Since a single character can fold into
3785 * such a sequence, the minimum match length for this node is less than
3786 * the number of characters in the node. This routine returns in
3787 * *min_subtract how many characters to subtract from the the actual
3788 * length of the string to get a real minimum match length; it is 0 if
3789 * there are no multi-char foldeds. This delta is used by the caller to
3790 * adjust the min length of the match, and the delta between min and max,
3791 * so that the optimizer doesn't reject these possibilities based on size
3793 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3794 * is used for an EXACTFU node that contains at least one "ss" sequence in
3795 * it. For non-UTF-8 patterns and strings, this is the only case where
3796 * there is a possible fold length change. That means that a regular
3797 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3798 * with length changes, and so can be processed faster. regexec.c takes
3799 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3800 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3801 * known until runtime). This saves effort in regex matching. However,
3802 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3803 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3804 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3805 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3806 * possibilities for the non-UTF8 patterns are quite simple, except for
3807 * the sharp s. All the ones that don't involve a UTF-8 target string are
3808 * members of a fold-pair, and arrays are set up for all of them so that
3809 * the other member of the pair can be found quickly. Code elsewhere in
3810 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3811 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3812 * described in the next item.
3813 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3814 * validity of the fold won't be known until runtime, and so must remain
3815 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFAA
3816 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3817 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3818 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3819 * The reason this is a problem is that the optimizer part of regexec.c
3820 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3821 * that a character in the pattern corresponds to at most a single
3822 * character in the target string. (And I do mean character, and not byte
3823 * here, unlike other parts of the documentation that have never been
3824 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3825 * EXACTFL nodes can match the two character string 'ss'; in EXACTFAA
3826 * nodes it can match "\x{17F}\x{17F}". These, along with other ones in
3827 * EXACTFL nodes, violate the assumption, and they are the only instances
3828 * where it is violated. I'm reluctant to try to change the assumption,
3829 * as the code involved is impenetrable to me (khw), so instead the code
3830 * here punts. This routine examines EXACTFL nodes, and (when the pattern
3831 * isn't UTF-8) EXACTF and EXACTFAA for such unfolded folds, and returns a
3832 * boolean indicating whether or not the node contains such a fold. When
3833 * it is true, the caller sets a flag that later causes the optimizer in
3834 * this file to not set values for the floating and fixed string lengths,
3835 * and thus avoids the optimizer code in regexec.c that makes the invalid
3836 * assumption. Thus, there is no optimization based on string lengths for
3837 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3838 * EXACTF and EXACTFAA nodes that contain the sharp s. (The reason the
3839 * assumption is wrong only in these cases is that all other non-UTF-8
3840 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3841 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3842 * EXACTF nodes because we don't know at compile time if it actually
3843 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3844 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3845 * always matches; and EXACTFAA where it never does. In an EXACTFAA node
3846 * in a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3847 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3848 * string would require the pattern to be forced into UTF-8, the overhead
3849 * of which we want to avoid. Similarly the unfolded multi-char folds in
3850 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3853 * Similarly, the code that generates tries doesn't currently handle
3854 * not-already-folded multi-char folds, and it looks like a pain to change
3855 * that. Therefore, trie generation of EXACTFAA nodes with the sharp s
3856 * doesn't work. Instead, such an EXACTFAA is turned into a new regnode,
3857 * EXACTFAA_NO_TRIE, which the trie code knows not to handle. Most people
3858 * using /iaa matching will be doing so almost entirely with ASCII
3859 * strings, so this should rarely be encountered in practice */
3861 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3862 if (PL_regkind[OP(scan)] == EXACT) \
3863 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3866 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3867 UV *min_subtract, bool *unfolded_multi_char,
3868 U32 flags,regnode *val, U32 depth)
3870 /* Merge several consecutive EXACTish nodes into one. */
3871 regnode *n = regnext(scan);
3873 regnode *next = scan + NODE_SZ_STR(scan);
3877 regnode *stop = scan;
3878 GET_RE_DEBUG_FLAGS_DECL;
3880 PERL_UNUSED_ARG(depth);
3883 PERL_ARGS_ASSERT_JOIN_EXACT;
3884 #ifndef EXPERIMENTAL_INPLACESCAN
3885 PERL_UNUSED_ARG(flags);
3886 PERL_UNUSED_ARG(val);
3888 DEBUG_PEEP("join", scan, depth, 0);
3890 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3891 * EXACT ones that are mergeable to the current one. */
3893 && (PL_regkind[OP(n)] == NOTHING
3894 || (stringok && OP(n) == OP(scan)))
3896 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3899 if (OP(n) == TAIL || n > next)
3901 if (PL_regkind[OP(n)] == NOTHING) {
3902 DEBUG_PEEP("skip:", n, depth, 0);
3903 NEXT_OFF(scan) += NEXT_OFF(n);
3904 next = n + NODE_STEP_REGNODE;
3911 else if (stringok) {
3912 const unsigned int oldl = STR_LEN(scan);
3913 regnode * const nnext = regnext(n);
3915 /* XXX I (khw) kind of doubt that this works on platforms (should
3916 * Perl ever run on one) where U8_MAX is above 255 because of lots
3917 * of other assumptions */
3918 /* Don't join if the sum can't fit into a single node */
3919 if (oldl + STR_LEN(n) > U8_MAX)
3922 DEBUG_PEEP("merg", n, depth, 0);
3925 NEXT_OFF(scan) += NEXT_OFF(n);
3926 STR_LEN(scan) += STR_LEN(n);
3927 next = n + NODE_SZ_STR(n);
3928 /* Now we can overwrite *n : */
3929 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3937 #ifdef EXPERIMENTAL_INPLACESCAN
3938 if (flags && !NEXT_OFF(n)) {
3939 DEBUG_PEEP("atch", val, depth, 0);
3940 if (reg_off_by_arg[OP(n)]) {
3941 ARG_SET(n, val - n);
3944 NEXT_OFF(n) = val - n;
3952 *unfolded_multi_char = FALSE;
3954 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3955 * can now analyze for sequences of problematic code points. (Prior to
3956 * this final joining, sequences could have been split over boundaries, and
3957 * hence missed). The sequences only happen in folding, hence for any
3958 * non-EXACT EXACTish node */
3959 if (OP(scan) != EXACT && OP(scan) != EXACTL) {
3960 U8* s0 = (U8*) STRING(scan);
3962 U8* s_end = s0 + STR_LEN(scan);
3964 int total_count_delta = 0; /* Total delta number of characters that
3965 multi-char folds expand to */
3967 /* One pass is made over the node's string looking for all the
3968 * possibilities. To avoid some tests in the loop, there are two main
3969 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3974 if (OP(scan) == EXACTFL) {
3977 /* An EXACTFL node would already have been changed to another
3978 * node type unless there is at least one character in it that
3979 * is problematic; likely a character whose fold definition
3980 * won't be known until runtime, and so has yet to be folded.
3981 * For all but the UTF-8 locale, folds are 1-1 in length, but
3982 * to handle the UTF-8 case, we need to create a temporary
3983 * folded copy using UTF-8 locale rules in order to analyze it.
3984 * This is because our macros that look to see if a sequence is
3985 * a multi-char fold assume everything is folded (otherwise the
3986 * tests in those macros would be too complicated and slow).
3987 * Note that here, the non-problematic folds will have already
3988 * been done, so we can just copy such characters. We actually
3989 * don't completely fold the EXACTFL string. We skip the
3990 * unfolded multi-char folds, as that would just create work
3991 * below to figure out the size they already are */
3993 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3996 STRLEN s_len = UTF8SKIP(s);
3997 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3998 Copy(s, d, s_len, U8);
4001 else if (is_FOLDS_TO_MULTI_utf8(s)) {
4002 *unfolded_multi_char = TRUE;
4003 Copy(s, d, s_len, U8);
4006 else if (isASCII(*s)) {
4007 *(d++) = toFOLD(*s);
4011 _toFOLD_utf8_flags(s, s_end, d, &len, FOLD_FLAGS_FULL);
4017 /* Point the remainder of the routine to look at our temporary
4021 } /* End of creating folded copy of EXACTFL string */
4023 /* Examine the string for a multi-character fold sequence. UTF-8
4024 * patterns have all characters pre-folded by the time this code is
4026 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
4027 length sequence we are looking for is 2 */
4029 int count = 0; /* How many characters in a multi-char fold */
4030 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
4031 if (! len) { /* Not a multi-char fold: get next char */
https://perl5.git.perl.org / perl5.git / blob