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_offset emit; /* Code-emit pointer */
143 I32 naughty; /* How bad is this pattern? */
144 I32 sawback; /* Did we see \1, ...? */
146 SSize_t size; /* Number of regnode equivalents in
149 /* position beyond 'precomp' of the warning message furthest away from
150 * 'precomp'. During the parse, no warnings are raised for any problems
151 * earlier in the parse than this position. This works if warnings are
152 * raised the first time a given spot is parsed, and if only one
153 * independent warning is raised for any given spot */
154 Size_t latest_warn_offset;
156 I32 npar; /* Capture buffer count so far in the
157 parse, (OPEN) plus one. ("par" 0 is
159 I32 total_par; /* During initial parse, is either 0,
160 or -1; the latter indicating a
161 reparse is needed. After that pass,
162 it is what 'npar' became after the
163 pass. Hence, it being > 0 indicates
164 we are in a reparse situation */
165 I32 nestroot; /* root parens we are in - used by
168 regnode_offset *open_parens; /* offsets to open parens */
169 regnode_offset *close_parens; /* offsets to close parens */
170 regnode *end_op; /* END node in program */
171 I32 utf8; /* whether the pattern is utf8 or not */
172 I32 orig_utf8; /* whether the pattern was originally in utf8 */
173 /* XXX use this for future optimisation of case
174 * where pattern must be upgraded to utf8. */
175 I32 uni_semantics; /* If a d charset modifier should use unicode
176 rules, even if the pattern is not in
178 HV *paren_names; /* Paren names */
180 regnode **recurse; /* Recurse regops */
181 I32 recurse_count; /* Number of recurse regops we have generated */
182 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
184 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
187 I32 override_recoding;
189 I32 recode_x_to_native;
191 I32 in_multi_char_class;
192 struct reg_code_blocks *code_blocks;/* positions of literal (?{})
194 int code_index; /* next code_blocks[] slot */
195 SSize_t maxlen; /* mininum possible number of chars in string to match */
196 scan_frame *frame_head;
197 scan_frame *frame_last;
200 #ifdef ADD_TO_REGEXEC
201 char *starttry; /* -Dr: where regtry was called. */
202 #define RExC_starttry (pRExC_state->starttry)
204 SV *runtime_code_qr; /* qr with the runtime code blocks */
206 const char *lastparse;
208 AV *paren_name_list; /* idx -> name */
209 U32 study_chunk_recursed_count;
213 #define RExC_lastparse (pRExC_state->lastparse)
214 #define RExC_lastnum (pRExC_state->lastnum)
215 #define RExC_paren_name_list (pRExC_state->paren_name_list)
216 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
217 #define RExC_mysv (pRExC_state->mysv1)
218 #define RExC_mysv1 (pRExC_state->mysv1)
219 #define RExC_mysv2 (pRExC_state->mysv2)
229 #define RExC_flags (pRExC_state->flags)
230 #define RExC_pm_flags (pRExC_state->pm_flags)
231 #define RExC_precomp (pRExC_state->precomp)
232 #define RExC_copy_start_in_input (pRExC_state->copy_start_in_input)
233 #define RExC_copy_start_in_constructed (pRExC_state->copy_start)
234 #define RExC_precomp_end (pRExC_state->precomp_end)
235 #define RExC_rx_sv (pRExC_state->rx_sv)
236 #define RExC_rx (pRExC_state->rx)
237 #define RExC_rxi (pRExC_state->rxi)
238 #define RExC_start (pRExC_state->start)
239 #define RExC_end (pRExC_state->end)
240 #define RExC_parse (pRExC_state->parse)
241 #define RExC_latest_warn_offset (pRExC_state->latest_warn_offset )
242 #define RExC_whilem_seen (pRExC_state->whilem_seen)
243 #define RExC_seen_d_op (pRExC_state->seen_d_op) /* Seen something that differs
244 under /d from /u ? */
247 #ifdef RE_TRACK_PATTERN_OFFSETS
248 # define RExC_offsets (RExC_rxi->u.offsets) /* I am not like the
251 #define RExC_emit (pRExC_state->emit)
252 #define RExC_emit_start (pRExC_state->emit_start)
253 #define RExC_sawback (pRExC_state->sawback)
254 #define RExC_seen (pRExC_state->seen)
255 #define RExC_size (pRExC_state->size)
256 #define RExC_maxlen (pRExC_state->maxlen)
257 #define RExC_npar (pRExC_state->npar)
258 #define RExC_total_parens (pRExC_state->total_par)
259 #define RExC_nestroot (pRExC_state->nestroot)
260 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
261 #define RExC_utf8 (pRExC_state->utf8)
262 #define RExC_uni_semantics (pRExC_state->uni_semantics)
263 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
264 #define RExC_open_parens (pRExC_state->open_parens)
265 #define RExC_close_parens (pRExC_state->close_parens)
266 #define RExC_end_op (pRExC_state->end_op)
267 #define RExC_paren_names (pRExC_state->paren_names)
268 #define RExC_recurse (pRExC_state->recurse)
269 #define RExC_recurse_count (pRExC_state->recurse_count)
270 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
271 #define RExC_study_chunk_recursed_bytes \
272 (pRExC_state->study_chunk_recursed_bytes)
273 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
274 #define RExC_contains_locale (pRExC_state->contains_locale)
276 # define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
278 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
279 #define RExC_frame_head (pRExC_state->frame_head)
280 #define RExC_frame_last (pRExC_state->frame_last)
281 #define RExC_frame_count (pRExC_state->frame_count)
282 #define RExC_strict (pRExC_state->strict)
283 #define RExC_study_started (pRExC_state->study_started)
284 #define RExC_warn_text (pRExC_state->warn_text)
285 #define RExC_in_script_run (pRExC_state->in_script_run)
286 #define RExC_use_BRANCHJ (pRExC_state->use_BRANCHJ)
288 /* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
289 * a flag to disable back-off on the fixed/floating substrings - if it's
290 * a high complexity pattern we assume the benefit of avoiding a full match
291 * is worth the cost of checking for the substrings even if they rarely help.
293 #define RExC_naughty (pRExC_state->naughty)
294 #define TOO_NAUGHTY (10)
295 #define MARK_NAUGHTY(add) \
296 if (RExC_naughty < TOO_NAUGHTY) \
297 RExC_naughty += (add)
298 #define MARK_NAUGHTY_EXP(exp, add) \
299 if (RExC_naughty < TOO_NAUGHTY) \
300 RExC_naughty += RExC_naughty / (exp) + (add)
302 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
303 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
304 ((*s) == '{' && regcurly(s)))
307 * Flags to be passed up and down.
309 #define WORST 0 /* Worst case. */
310 #define HASWIDTH 0x01 /* Known to not match null strings, could match
313 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
314 * character. (There needs to be a case: in the switch statement in regexec.c
315 * for any node marked SIMPLE.) Note that this is not the same thing as
318 #define SPSTART 0x04 /* Starts with * or + */
319 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
320 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
321 #define RESTART_PARSE 0x20 /* Need to redo the parse */
322 #define NEED_UTF8 0x40 /* In conjunction with RESTART_PARSE, need to
323 calcuate sizes as UTF-8 */
325 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
327 /* whether trie related optimizations are enabled */
328 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
329 #define TRIE_STUDY_OPT
330 #define FULL_TRIE_STUDY
336 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
337 #define PBITVAL(paren) (1 << ((paren) & 7))
338 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
339 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
340 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
342 #define REQUIRE_UTF8(flagp) STMT_START { \
344 *flagp = RESTART_PARSE|NEED_UTF8; \
349 /* Change from /d into /u rules, and restart the parse. RExC_uni_semantics is
350 * a flag that indicates we've changed to /u during the parse. */
351 #define REQUIRE_UNI_RULES(flagp, restart_retval) \
353 if (DEPENDS_SEMANTICS) { \
354 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
355 RExC_uni_semantics = 1; \
356 if (RExC_seen_d_op && LIKELY(RExC_total_parens >= 0)) { \
357 /* No need to restart the parse if we haven't seen \
358 * anything that differs between /u and /d, and no need \
359 * to restart immediately if we're going to reparse \
360 * anyway to count parens */ \
361 *flagp |= RESTART_PARSE; \
362 return restart_retval; \
367 #define BRANCH_MAX_OFFSET U16_MAX
368 #define REQUIRE_BRANCHJ(flagp, restart_retval) \
370 RExC_use_BRANCHJ = 1; \
371 if (LIKELY(RExC_total_parens >= 0)) { \
372 /* No need to restart the parse immediately if we're \
373 * going to reparse anyway to count parens */ \
374 *flagp |= RESTART_PARSE; \
375 return restart_retval; \
379 #define REQUIRE_PARENS_PASS \
381 if (RExC_total_parens == 0) RExC_total_parens = -1; \
384 /* This is used to return failure (zero) early from the calling function if
385 * various flags in 'flags' are set. Two flags always cause a return:
386 * 'RESTART_PARSE' and 'NEED_UTF8'. 'extra' can be used to specify any
387 * additional flags that should cause a return; 0 if none. If the return will
388 * be done, '*flagp' is first set to be all of the flags that caused the
390 #define RETURN_FAIL_ON_RESTART_OR_FLAGS(flags,flagp,extra) \
392 if ((flags) & (RESTART_PARSE|NEED_UTF8|(extra))) { \
393 *(flagp) = (flags) & (RESTART_PARSE|NEED_UTF8|(extra)); \
398 #define MUST_RESTART(flags) ((flags) & (RESTART_PARSE))
400 #define RETURN_FAIL_ON_RESTART(flags,flagp) \
401 RETURN_FAIL_ON_RESTART_OR_FLAGS( flags, flagp, 0)
402 #define RETURN_FAIL_ON_RESTART_FLAGP(flagp) \
403 if (MUST_RESTART(*(flagp))) return 0
405 /* This converts the named class defined in regcomp.h to its equivalent class
406 * number defined in handy.h. */
407 #define namedclass_to_classnum(class) ((int) ((class) / 2))
408 #define classnum_to_namedclass(classnum) ((classnum) * 2)
410 #define _invlist_union_complement_2nd(a, b, output) \
411 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
412 #define _invlist_intersection_complement_2nd(a, b, output) \
413 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
415 /* About scan_data_t.
417 During optimisation we recurse through the regexp program performing
418 various inplace (keyhole style) optimisations. In addition study_chunk
419 and scan_commit populate this data structure with information about
420 what strings MUST appear in the pattern. We look for the longest
421 string that must appear at a fixed location, and we look for the
422 longest string that may appear at a floating location. So for instance
427 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
428 strings (because they follow a .* construct). study_chunk will identify
429 both FOO and BAR as being the longest fixed and floating strings respectively.
431 The strings can be composites, for instance
435 will result in a composite fixed substring 'foo'.
437 For each string some basic information is maintained:
440 This is the position the string must appear at, or not before.
441 It also implicitly (when combined with minlenp) tells us how many
442 characters must match before the string we are searching for.
443 Likewise when combined with minlenp and the length of the string it
444 tells us how many characters must appear after the string we have
448 Only used for floating strings. This is the rightmost point that
449 the string can appear at. If set to SSize_t_MAX it indicates that the
450 string can occur infinitely far to the right.
451 For fixed strings, it is equal to min_offset.
454 A pointer to the minimum number of characters of the pattern that the
455 string was found inside. This is important as in the case of positive
456 lookahead or positive lookbehind we can have multiple patterns
461 The minimum length of the pattern overall is 3, the minimum length
462 of the lookahead part is 3, but the minimum length of the part that
463 will actually match is 1. So 'FOO's minimum length is 3, but the
464 minimum length for the F is 1. This is important as the minimum length
465 is used to determine offsets in front of and behind the string being
466 looked for. Since strings can be composites this is the length of the
467 pattern at the time it was committed with a scan_commit. Note that
468 the length is calculated by study_chunk, so that the minimum lengths
469 are not known until the full pattern has been compiled, thus the
470 pointer to the value.
474 In the case of lookbehind the string being searched for can be
475 offset past the start point of the final matching string.
476 If this value was just blithely removed from the min_offset it would
477 invalidate some of the calculations for how many chars must match
478 before or after (as they are derived from min_offset and minlen and
479 the length of the string being searched for).
480 When the final pattern is compiled and the data is moved from the
481 scan_data_t structure into the regexp structure the information
482 about lookbehind is factored in, with the information that would
483 have been lost precalculated in the end_shift field for the
486 The fields pos_min and pos_delta are used to store the minimum offset
487 and the delta to the maximum offset at the current point in the pattern.
491 struct scan_data_substrs {
492 SV *str; /* longest substring found in pattern */
493 SSize_t min_offset; /* earliest point in string it can appear */
494 SSize_t max_offset; /* latest point in string it can appear */
495 SSize_t *minlenp; /* pointer to the minlen relevant to the string */
496 SSize_t lookbehind; /* is the pos of the string modified by LB */
497 I32 flags; /* per substring SF_* and SCF_* flags */
500 typedef struct scan_data_t {
501 /*I32 len_min; unused */
502 /*I32 len_delta; unused */
506 SSize_t last_end; /* min value, <0 unless valid. */
507 SSize_t last_start_min;
508 SSize_t last_start_max;
509 U8 cur_is_floating; /* whether the last_* values should be set as
510 * the next fixed (0) or floating (1)
513 /* [0] is longest fixed substring so far, [1] is longest float so far */
514 struct scan_data_substrs substrs[2];
516 I32 flags; /* common SF_* and SCF_* flags */
518 SSize_t *last_closep;
519 regnode_ssc *start_class;
523 * Forward declarations for pregcomp()'s friends.
526 static const scan_data_t zero_scan_data = {
527 0, 0, NULL, 0, 0, 0, 0,
529 { NULL, 0, 0, 0, 0, 0 },
530 { NULL, 0, 0, 0, 0, 0 },
537 #define SF_BEFORE_SEOL 0x0001
538 #define SF_BEFORE_MEOL 0x0002
539 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
541 #define SF_IS_INF 0x0040
542 #define SF_HAS_PAR 0x0080
543 #define SF_IN_PAR 0x0100
544 #define SF_HAS_EVAL 0x0200
547 /* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
548 * longest substring in the pattern. When it is not set the optimiser keeps
549 * track of position, but does not keep track of the actual strings seen,
551 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
554 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
555 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
556 * turned off because of the alternation (BRANCH). */
557 #define SCF_DO_SUBSTR 0x0400
559 #define SCF_DO_STCLASS_AND 0x0800
560 #define SCF_DO_STCLASS_OR 0x1000
561 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
562 #define SCF_WHILEM_VISITED_POS 0x2000
564 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
565 #define SCF_SEEN_ACCEPT 0x8000
566 #define SCF_TRIE_DOING_RESTUDY 0x10000
567 #define SCF_IN_DEFINE 0x20000
572 #define UTF cBOOL(RExC_utf8)
574 /* The enums for all these are ordered so things work out correctly */
575 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
576 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
577 == REGEX_DEPENDS_CHARSET)
578 /* Use RExC_uni_semantics instead of this:
579 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
581 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
582 >= REGEX_UNICODE_CHARSET)
583 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
584 == REGEX_ASCII_RESTRICTED_CHARSET)
585 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
586 >= REGEX_ASCII_RESTRICTED_CHARSET)
587 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
588 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
590 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
592 /* For programs that want to be strictly Unicode compatible by dying if any
593 * attempt is made to match a non-Unicode code point against a Unicode
595 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
597 #define OOB_NAMEDCLASS -1
599 /* There is no code point that is out-of-bounds, so this is problematic. But
600 * its only current use is to initialize a variable that is always set before
602 #define OOB_UNICODE 0xDEADBEEF
604 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
607 /* length of regex to show in messages that don't mark a position within */
608 #define RegexLengthToShowInErrorMessages 127
611 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
612 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
613 * op/pragma/warn/regcomp.
615 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
616 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
618 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
619 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
621 /* The code in this file in places uses one level of recursion with parsing
622 * rebased to an alternate string constructed by us in memory. This can take
623 * the form of something that is completely different from the input, or
624 * something that uses the input as part of the alternate. In the first case,
625 * there should be no possibility of an error, as we are in complete control of
626 * the alternate string. But in the second case we don't completely control
627 * the input portion, so there may be errors in that. Here's an example:
629 * is handled specially because \x{df} folds to a sequence of more than one
630 * character: 'ss'. What is done is to create and parse an alternate string,
631 * which looks like this:
632 * /(?:\x{DF}|[abc\x{DF}def])/ui
633 * where it uses the input unchanged in the middle of something it constructs,
634 * which is a branch for the DF outside the character class, and clustering
635 * parens around the whole thing. (It knows enough to skip the DF inside the
636 * class while in this substitute parse.) 'abc' and 'def' may have errors that
637 * need to be reported. The general situation looks like this:
639 * |<------- identical ------>|
641 * Input: ---------------------------------------------------------------
642 * Constructed: ---------------------------------------------------
644 * |<------- identical ------>|
646 * sI..eI is the portion of the input pattern we are concerned with here.
647 * sC..EC is the constructed substitute parse string.
648 * sC..tC is constructed by us
649 * tC..eC is an exact duplicate of the portion of the input pattern tI..eI.
650 * In the diagram, these are vertically aligned.
651 * eC..EC is also constructed by us.
652 * xC is the position in the substitute parse string where we found a
654 * xI is the position in the original pattern corresponding to xC.
656 * We want to display a message showing the real input string. Thus we need to
657 * translate from xC to xI. We know that xC >= tC, since the portion of the
658 * string sC..tC has been constructed by us, and so shouldn't have errors. We
660 * xI = tI + (xC - tC)
662 * When the substitute parse is constructed, the code needs to set:
665 * RExC_copy_start_in_input (tI)
666 * RExC_copy_start_in_constructed (tC)
667 * and restore them when done.
669 * During normal processing of the input pattern, both
670 * 'RExC_copy_start_in_input' and 'RExC_copy_start_in_constructed' are set to
671 * sI, so that xC equals xI.
674 #define sI RExC_precomp
675 #define eI RExC_precomp_end
676 #define sC RExC_start
678 #define tI RExC_copy_start_in_input
679 #define tC RExC_copy_start_in_constructed
680 #define xI(xC) (tI + (xC - tC))
681 #define xI_offset(xC) (xI(xC) - sI)
683 #define REPORT_LOCATION_ARGS(xC) \
685 (xI(xC) > eI) /* Don't run off end */ \
686 ? eI - sI /* Length before the <--HERE */ \
687 : ((xI_offset(xC) >= 0) \
689 : (Perl_croak(aTHX_ "panic: %s: %d: negative offset: %" \
690 IVdf " trying to output message for " \
692 __FILE__, __LINE__, (IV) xI_offset(xC), \
693 ((int) (eC - sC)), sC), 0)), \
694 sI), /* The input pattern printed up to the <--HERE */ \
696 (xI(xC) > eI) ? 0 : eI - xI(xC), /* Length after <--HERE */ \
697 (xI(xC) > eI) ? eI : xI(xC)) /* pattern after <--HERE */
699 /* Used to point after bad bytes for an error message, but avoid skipping
700 * past a nul byte. */
701 #define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
703 /* Set up to clean up after our imminent demise */
704 #define PREPARE_TO_DIE \
707 SAVEFREESV(RExC_rx_sv); \
708 if (RExC_open_parens) \
709 SAVEFREEPV(RExC_open_parens); \
710 if (RExC_close_parens) \
711 SAVEFREEPV(RExC_close_parens); \
715 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
716 * arg. Show regex, up to a maximum length. If it's too long, chop and add
719 #define _FAIL(code) STMT_START { \
720 const char *ellipses = ""; \
721 IV len = RExC_precomp_end - RExC_precomp; \
724 if (len > RegexLengthToShowInErrorMessages) { \
725 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
726 len = RegexLengthToShowInErrorMessages - 10; \
732 #define FAIL(msg) _FAIL( \
733 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
734 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
736 #define FAIL2(msg,arg) _FAIL( \
737 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
738 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
741 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
743 #define Simple_vFAIL(m) STMT_START { \
744 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
745 m, REPORT_LOCATION_ARGS(RExC_parse)); \
749 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
751 #define vFAIL(m) STMT_START { \
757 * Like Simple_vFAIL(), but accepts two arguments.
759 #define Simple_vFAIL2(m,a1) STMT_START { \
760 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
761 REPORT_LOCATION_ARGS(RExC_parse)); \
765 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
767 #define vFAIL2(m,a1) STMT_START { \
769 Simple_vFAIL2(m, a1); \
774 * Like Simple_vFAIL(), but accepts three arguments.
776 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
777 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
778 REPORT_LOCATION_ARGS(RExC_parse)); \
782 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
784 #define vFAIL3(m,a1,a2) STMT_START { \
786 Simple_vFAIL3(m, a1, a2); \
790 * Like Simple_vFAIL(), but accepts four arguments.
792 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
793 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
794 REPORT_LOCATION_ARGS(RExC_parse)); \
797 #define vFAIL4(m,a1,a2,a3) STMT_START { \
799 Simple_vFAIL4(m, a1, a2, a3); \
802 /* A specialized version of vFAIL2 that works with UTF8f */
803 #define vFAIL2utf8f(m, a1) STMT_START { \
805 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
806 REPORT_LOCATION_ARGS(RExC_parse)); \
809 #define vFAIL3utf8f(m, a1, a2) STMT_START { \
811 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
812 REPORT_LOCATION_ARGS(RExC_parse)); \
815 /* Setting this to NULL is a signal to not output warnings */
816 #define TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE RExC_copy_start_in_constructed = NULL
817 #define RESTORE_WARNINGS RExC_copy_start_in_constructed = RExC_precomp
819 /* Since a warning can be generated multiple times as the input is reparsed, we
820 * output it the first time we come to that point in the parse, but suppress it
821 * otherwise. 'RExC_copy_start_in_constructed' being NULL is a flag to not
822 * generate any warnings */
823 #define TO_OUTPUT_WARNINGS(loc) \
824 ( RExC_copy_start_in_constructed \
825 && ((xI(loc)) - RExC_precomp) > (Ptrdiff_t) RExC_latest_warn_offset)
827 /* After we've emitted a warning, we save the position in the input so we don't
829 #define UPDATE_WARNINGS_LOC(loc) \
831 if (TO_OUTPUT_WARNINGS(loc)) { \
832 RExC_latest_warn_offset = (xI(loc)) - RExC_precomp; \
836 /* 'warns' is the output of the packWARNx macro used in 'code' */
837 #define _WARN_HELPER(loc, warns, code) \
839 if (! RExC_copy_start_in_constructed) { \
840 Perl_croak( aTHX_ "panic! %s: %d: Tried to warn when none" \
841 " expected at '%s'", \
842 __FILE__, __LINE__, loc); \
844 if (TO_OUTPUT_WARNINGS(loc)) { \
848 UPDATE_WARNINGS_LOC(loc); \
852 /* m is not necessarily a "literal string", in this macro */
853 #define reg_warn_non_literal_string(loc, m) \
854 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
855 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
856 "%s" REPORT_LOCATION, \
857 m, REPORT_LOCATION_ARGS(loc)))
859 #define ckWARNreg(loc,m) \
860 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
861 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
863 REPORT_LOCATION_ARGS(loc)))
865 #define vWARN(loc, m) \
866 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
867 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
869 REPORT_LOCATION_ARGS(loc))) \
871 #define vWARN_dep(loc, m) \
872 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
873 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
875 REPORT_LOCATION_ARGS(loc)))
877 #define ckWARNdep(loc,m) \
878 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
879 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
881 REPORT_LOCATION_ARGS(loc)))
883 #define ckWARNregdep(loc,m) \
884 _WARN_HELPER(loc, packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
885 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
888 REPORT_LOCATION_ARGS(loc)))
890 #define ckWARN2reg_d(loc,m, a1) \
891 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
892 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
894 a1, REPORT_LOCATION_ARGS(loc)))
896 #define ckWARN2reg(loc, m, a1) \
897 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
898 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
900 a1, REPORT_LOCATION_ARGS(loc)))
902 #define vWARN3(loc, m, a1, a2) \
903 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
904 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
906 a1, a2, REPORT_LOCATION_ARGS(loc)))
908 #define ckWARN3reg(loc, m, a1, a2) \
909 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
910 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
913 REPORT_LOCATION_ARGS(loc)))
915 #define vWARN4(loc, m, a1, a2, a3) \
916 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
917 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
920 REPORT_LOCATION_ARGS(loc)))
922 #define ckWARN4reg(loc, m, a1, a2, a3) \
923 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
924 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
927 REPORT_LOCATION_ARGS(loc)))
929 #define vWARN5(loc, m, a1, a2, a3, a4) \
930 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
931 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
934 REPORT_LOCATION_ARGS(loc)))
936 #define ckWARNexperimental(loc, class, m) \
937 _WARN_HELPER(loc, packWARN(class), \
938 Perl_ck_warner_d(aTHX_ packWARN(class), \
940 REPORT_LOCATION_ARGS(loc)))
942 /* Convert between a pointer to a node and its offset from the beginning of the
944 #define REGNODE_p(offset) (RExC_emit_start + (offset))
945 #define REGNODE_OFFSET(node) ((node) - RExC_emit_start)
947 /* Macros for recording node offsets. 20001227 mjd@plover.com
948 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
949 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
950 * Element 0 holds the number n.
951 * Position is 1 indexed.
953 #ifndef RE_TRACK_PATTERN_OFFSETS
954 #define Set_Node_Offset_To_R(offset,byte)
955 #define Set_Node_Offset(node,byte)
956 #define Set_Cur_Node_Offset
957 #define Set_Node_Length_To_R(node,len)
958 #define Set_Node_Length(node,len)
959 #define Set_Node_Cur_Length(node,start)
960 #define Node_Offset(n)
961 #define Node_Length(n)
962 #define Set_Node_Offset_Length(node,offset,len)
963 #define ProgLen(ri) ri->u.proglen
964 #define SetProgLen(ri,x) ri->u.proglen = x
965 #define Track_Code(code)
967 #define ProgLen(ri) ri->u.offsets[0]
968 #define SetProgLen(ri,x) ri->u.offsets[0] = x
969 #define Set_Node_Offset_To_R(offset,byte) STMT_START { \
970 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
971 __LINE__, (int)(offset), (int)(byte))); \
973 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
976 RExC_offsets[2*(offset)-1] = (byte); \
980 #define Set_Node_Offset(node,byte) \
981 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (byte)-RExC_start)
982 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
984 #define Set_Node_Length_To_R(node,len) STMT_START { \
985 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
986 __LINE__, (int)(node), (int)(len))); \
988 Perl_croak(aTHX_ "value of node is %d in Length macro", \
991 RExC_offsets[2*(node)] = (len); \
995 #define Set_Node_Length(node,len) \
996 Set_Node_Length_To_R(REGNODE_OFFSET(node), len)
997 #define Set_Node_Cur_Length(node, start) \
998 Set_Node_Length(node, RExC_parse - start)
1000 /* Get offsets and lengths */
1001 #define Node_Offset(n) (RExC_offsets[2*(REGNODE_OFFSET(n))-1])
1002 #define Node_Length(n) (RExC_offsets[2*(REGNODE_OFFSET(n))])
1004 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
1005 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (offset)); \
1006 Set_Node_Length_To_R(REGNODE_OFFSET(node), (len)); \
1009 #define Track_Code(code) STMT_START { code } STMT_END
1012 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
1013 #define EXPERIMENTAL_INPLACESCAN
1014 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
1018 Perl_re_printf(pTHX_ const char *fmt, ...)
1022 PerlIO *f= Perl_debug_log;
1023 PERL_ARGS_ASSERT_RE_PRINTF;
1025 result = PerlIO_vprintf(f, fmt, ap);
1031 Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
1035 PerlIO *f= Perl_debug_log;
1036 PERL_ARGS_ASSERT_RE_INDENTF;
1037 va_start(ap, depth);
1038 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
1039 result = PerlIO_vprintf(f, fmt, ap);
1043 #endif /* DEBUGGING */
1045 #define DEBUG_RExC_seen() \
1046 DEBUG_OPTIMISE_MORE_r({ \
1047 Perl_re_printf( aTHX_ "RExC_seen: "); \
1049 if (RExC_seen & REG_ZERO_LEN_SEEN) \
1050 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
1052 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
1053 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
1055 if (RExC_seen & REG_GPOS_SEEN) \
1056 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
1058 if (RExC_seen & REG_RECURSE_SEEN) \
1059 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
1061 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
1062 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
1064 if (RExC_seen & REG_VERBARG_SEEN) \
1065 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
1067 if (RExC_seen & REG_CUTGROUP_SEEN) \
1068 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
1070 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
1071 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
1073 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
1074 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
1076 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
1077 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
1079 Perl_re_printf( aTHX_ "\n"); \
1082 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
1083 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
1088 S_debug_show_study_flags(pTHX_ U32 flags, const char *open_str,
1089 const char *close_str)
1094 Perl_re_printf( aTHX_ "%s", open_str);
1095 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_SEOL);
1096 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_MEOL);
1097 DEBUG_SHOW_STUDY_FLAG(flags, SF_IS_INF);
1098 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_PAR);
1099 DEBUG_SHOW_STUDY_FLAG(flags, SF_IN_PAR);
1100 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_EVAL);
1101 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_SUBSTR);
1102 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_AND);
1103 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_OR);
1104 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS);
1105 DEBUG_SHOW_STUDY_FLAG(flags, SCF_WHILEM_VISITED_POS);
1106 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_RESTUDY);
1107 DEBUG_SHOW_STUDY_FLAG(flags, SCF_SEEN_ACCEPT);
1108 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_DOING_RESTUDY);
1109 DEBUG_SHOW_STUDY_FLAG(flags, SCF_IN_DEFINE);
1110 Perl_re_printf( aTHX_ "%s", close_str);
1115 S_debug_studydata(pTHX_ const char *where, scan_data_t *data,
1116 U32 depth, int is_inf)
1118 GET_RE_DEBUG_FLAGS_DECL;
1120 DEBUG_OPTIMISE_MORE_r({
1123 Perl_re_indentf(aTHX_ "%s: Pos:%" IVdf "/%" IVdf " Flags: 0x%" UVXf,
1127 (IV)data->pos_delta,
1131 S_debug_show_study_flags(aTHX_ data->flags," [","]");
1133 Perl_re_printf( aTHX_
1134 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s",
1136 (IV)(data->last_closep ? *((data)->last_closep) : -1),
1137 is_inf ? "INF " : ""
1140 if (data->last_found) {
1142 Perl_re_printf(aTHX_
1143 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf,
1144 SvPVX_const(data->last_found),
1146 (IV)data->last_start_min,
1147 (IV)data->last_start_max
1150 for (i = 0; i < 2; i++) {
1151 Perl_re_printf(aTHX_
1152 " %s%s: '%s' @ %" IVdf "/%" IVdf,
1153 data->cur_is_floating == i ? "*" : "",
1154 i ? "Float" : "Fixed",
1155 SvPVX_const(data->substrs[i].str),
1156 (IV)data->substrs[i].min_offset,
1157 (IV)data->substrs[i].max_offset
1159 S_debug_show_study_flags(aTHX_ data->substrs[i].flags," [","]");
1163 Perl_re_printf( aTHX_ "\n");
1169 S_debug_peep(pTHX_ const char *str, const RExC_state_t *pRExC_state,
1170 regnode *scan, U32 depth, U32 flags)
1172 GET_RE_DEBUG_FLAGS_DECL;
1179 Next = regnext(scan);
1180 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
1181 Perl_re_indentf( aTHX_ "%s>%3d: %s (%d)",
1184 REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),
1185 Next ? (REG_NODE_NUM(Next)) : 0 );
1186 S_debug_show_study_flags(aTHX_ flags," [ ","]");
1187 Perl_re_printf( aTHX_ "\n");
1192 # define DEBUG_STUDYDATA(where, data, depth, is_inf) \
1193 S_debug_studydata(aTHX_ where, data, depth, is_inf)
1195 # define DEBUG_PEEP(str, scan, depth, flags) \
1196 S_debug_peep(aTHX_ str, pRExC_state, scan, depth, flags)
1199 # define DEBUG_STUDYDATA(where, data, depth, is_inf) NOOP
1200 # define DEBUG_PEEP(str, scan, depth, flags) NOOP
1204 /* =========================================================
1205 * BEGIN edit_distance stuff.
1207 * This calculates how many single character changes of any type are needed to
1208 * transform a string into another one. It is taken from version 3.1 of
1210 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1213 /* Our unsorted dictionary linked list. */
1214 /* Note we use UVs, not chars. */
1219 struct dictionary* next;
1221 typedef struct dictionary item;
1224 PERL_STATIC_INLINE item*
1225 push(UV key, item* curr)
1228 Newx(head, 1, item);
1236 PERL_STATIC_INLINE item*
1237 find(item* head, UV key)
1239 item* iterator = head;
1241 if (iterator->key == key){
1244 iterator = iterator->next;
1250 PERL_STATIC_INLINE item*
1251 uniquePush(item* head, UV key)
1253 item* iterator = head;
1256 if (iterator->key == key) {
1259 iterator = iterator->next;
1262 return push(key, head);
1265 PERL_STATIC_INLINE void
1266 dict_free(item* head)
1268 item* iterator = head;
1271 item* temp = iterator;
1272 iterator = iterator->next;
1279 /* End of Dictionary Stuff */
1281 /* All calculations/work are done here */
1283 S_edit_distance(const UV* src,
1285 const STRLEN x, /* length of src[] */
1286 const STRLEN y, /* length of tgt[] */
1287 const SSize_t maxDistance
1291 UV swapCount, swapScore, targetCharCount, i, j;
1293 UV score_ceil = x + y;
1295 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1297 /* intialize matrix start values */
1298 Newx(scores, ( (x + 2) * (y + 2)), UV);
1299 scores[0] = score_ceil;
1300 scores[1 * (y + 2) + 0] = score_ceil;
1301 scores[0 * (y + 2) + 1] = score_ceil;
1302 scores[1 * (y + 2) + 1] = 0;
1303 head = uniquePush(uniquePush(head, src[0]), tgt[0]);
1308 for (i=1;i<=x;i++) {
1310 head = uniquePush(head, src[i]);
1311 scores[(i+1) * (y + 2) + 1] = i;
1312 scores[(i+1) * (y + 2) + 0] = score_ceil;
1315 for (j=1;j<=y;j++) {
1318 head = uniquePush(head, tgt[j]);
1319 scores[1 * (y + 2) + (j + 1)] = j;
1320 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1323 targetCharCount = find(head, tgt[j-1])->value;
1324 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1326 if (src[i-1] != tgt[j-1]){
1327 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));
1331 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1335 find(head, src[i-1])->value = i;
1339 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1342 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1346 /* END of edit_distance() stuff
1347 * ========================================================= */
1349 /* is c a control character for which we have a mnemonic? */
1350 #define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
1353 S_cntrl_to_mnemonic(const U8 c)
1355 /* Returns the mnemonic string that represents character 'c', if one
1356 * exists; NULL otherwise. The only ones that exist for the purposes of
1357 * this routine are a few control characters */
1360 case '\a': return "\\a";
1361 case '\b': return "\\b";
1362 case ESC_NATIVE: return "\\e";
1363 case '\f': return "\\f";
1364 case '\n': return "\\n";
1365 case '\r': return "\\r";
1366 case '\t': return "\\t";
1372 /* Mark that we cannot extend a found fixed substring at this point.
1373 Update the longest found anchored substring or the longest found
1374 floating substrings if needed. */
1377 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1378 SSize_t *minlenp, int is_inf)
1380 const STRLEN l = CHR_SVLEN(data->last_found);
1381 SV * const longest_sv = data->substrs[data->cur_is_floating].str;
1382 const STRLEN old_l = CHR_SVLEN(longest_sv);
1383 GET_RE_DEBUG_FLAGS_DECL;
1385 PERL_ARGS_ASSERT_SCAN_COMMIT;
1387 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
1388 const U8 i = data->cur_is_floating;
1389 SvSetMagicSV(longest_sv, data->last_found);
1390 data->substrs[i].min_offset = l ? data->last_start_min : data->pos_min;
1393 data->substrs[0].max_offset = data->substrs[0].min_offset;
1395 data->substrs[1].max_offset = (l
1396 ? data->last_start_max
1397 : (data->pos_delta > SSize_t_MAX - data->pos_min
1399 : data->pos_min + data->pos_delta));
1401 || (STRLEN)data->substrs[1].max_offset > (STRLEN)SSize_t_MAX)
1402 data->substrs[1].max_offset = SSize_t_MAX;
1405 if (data->flags & SF_BEFORE_EOL)
1406 data->substrs[i].flags |= (data->flags & SF_BEFORE_EOL);
1408 data->substrs[i].flags &= ~SF_BEFORE_EOL;
1409 data->substrs[i].minlenp = minlenp;
1410 data->substrs[i].lookbehind = 0;
1413 SvCUR_set(data->last_found, 0);
1415 SV * const sv = data->last_found;
1416 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1417 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1422 data->last_end = -1;
1423 data->flags &= ~SF_BEFORE_EOL;
1424 DEBUG_STUDYDATA("commit", data, 0, is_inf);
1427 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1428 * list that describes which code points it matches */
1431 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1433 /* Set the SSC 'ssc' to match an empty string or any code point */
1435 PERL_ARGS_ASSERT_SSC_ANYTHING;
1437 assert(is_ANYOF_SYNTHETIC(ssc));
1439 /* mortalize so won't leak */
1440 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
1441 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1445 S_ssc_is_anything(const regnode_ssc *ssc)
1447 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1448 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1449 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1450 * in any way, so there's no point in using it */
1455 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1457 assert(is_ANYOF_SYNTHETIC(ssc));
1459 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1463 /* See if the list consists solely of the range 0 - Infinity */
1464 invlist_iterinit(ssc->invlist);
1465 ret = invlist_iternext(ssc->invlist, &start, &end)
1469 invlist_iterfinish(ssc->invlist);
1475 /* If e.g., both \w and \W are set, matches everything */
1476 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1478 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1479 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1489 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1491 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1492 * string, any code point, or any posix class under locale */
1494 PERL_ARGS_ASSERT_SSC_INIT;
1496 Zero(ssc, 1, regnode_ssc);
1497 set_ANYOF_SYNTHETIC(ssc);
1498 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1501 /* If any portion of the regex is to operate under locale rules that aren't
1502 * fully known at compile time, initialization includes it. The reason
1503 * this isn't done for all regexes is that the optimizer was written under
1504 * the assumption that locale was all-or-nothing. Given the complexity and
1505 * lack of documentation in the optimizer, and that there are inadequate
1506 * test cases for locale, many parts of it may not work properly, it is
1507 * safest to avoid locale unless necessary. */
1508 if (RExC_contains_locale) {
1509 ANYOF_POSIXL_SETALL(ssc);
1512 ANYOF_POSIXL_ZERO(ssc);
1517 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1518 const regnode_ssc *ssc)
1520 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1521 * to the list of code points matched, and locale posix classes; hence does
1522 * not check its flags) */
1527 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1529 assert(is_ANYOF_SYNTHETIC(ssc));
1531 invlist_iterinit(ssc->invlist);
1532 ret = invlist_iternext(ssc->invlist, &start, &end)
1536 invlist_iterfinish(ssc->invlist);
1542 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1550 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1551 const regnode_charclass* const node)
1553 /* Returns a mortal inversion list defining which code points are matched
1554 * by 'node', which is of type ANYOF. Handles complementing the result if
1555 * appropriate. If some code points aren't knowable at this time, the
1556 * returned list must, and will, contain every code point that is a
1560 SV* only_utf8_locale_invlist = NULL;
1562 const U32 n = ARG(node);
1563 bool new_node_has_latin1 = FALSE;
1565 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1567 /* Look at the data structure created by S_set_ANYOF_arg() */
1568 if (n != ANYOF_ONLY_HAS_BITMAP) {
1569 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1570 AV * const av = MUTABLE_AV(SvRV(rv));
1571 SV **const ary = AvARRAY(av);
1572 assert(RExC_rxi->data->what[n] == 's');
1574 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1575 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1]), NULL));
1577 else if (ary[0] && ary[0] != &PL_sv_undef) {
1579 /* Here, no compile-time swash, and there are things that won't be
1580 * known until runtime -- we have to assume it could be anything */
1581 invlist = sv_2mortal(_new_invlist(1));
1582 return _add_range_to_invlist(invlist, 0, UV_MAX);
1584 else if (ary[3] && ary[3] != &PL_sv_undef) {
1586 /* Here no compile-time swash, and no run-time only data. Use the
1587 * node's inversion list */
1588 invlist = sv_2mortal(invlist_clone(ary[3], NULL));
1591 /* Get the code points valid only under UTF-8 locales */
1592 if ((ANYOF_FLAGS(node) & ANYOFL_FOLD)
1593 && ary[2] && ary[2] != &PL_sv_undef)
1595 only_utf8_locale_invlist = ary[2];
1600 invlist = sv_2mortal(_new_invlist(0));
1603 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1604 * code points, and an inversion list for the others, but if there are code
1605 * points that should match only conditionally on the target string being
1606 * UTF-8, those are placed in the inversion list, and not the bitmap.
1607 * Since there are circumstances under which they could match, they are
1608 * included in the SSC. But if the ANYOF node is to be inverted, we have
1609 * to exclude them here, so that when we invert below, the end result
1610 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1611 * have to do this here before we add the unconditionally matched code
1613 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1614 _invlist_intersection_complement_2nd(invlist,
1619 /* Add in the points from the bit map */
1620 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1621 if (ANYOF_BITMAP_TEST(node, i)) {
1622 unsigned int start = i++;
1624 for (; i < NUM_ANYOF_CODE_POINTS && ANYOF_BITMAP_TEST(node, i); ++i) {
1627 invlist = _add_range_to_invlist(invlist, start, i-1);
1628 new_node_has_latin1 = TRUE;
1632 /* If this can match all upper Latin1 code points, have to add them
1633 * as well. But don't add them if inverting, as when that gets done below,
1634 * it would exclude all these characters, including the ones it shouldn't
1635 * that were added just above */
1636 if (! (ANYOF_FLAGS(node) & ANYOF_INVERT) && OP(node) == ANYOFD
1637 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1639 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1642 /* Similarly for these */
1643 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1644 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1647 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1648 _invlist_invert(invlist);
1650 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOFL_FOLD) {
1652 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1653 * locale. We can skip this if there are no 0-255 at all. */
1654 _invlist_union(invlist, PL_Latin1, &invlist);
1657 /* Similarly add the UTF-8 locale possible matches. These have to be
1658 * deferred until after the non-UTF-8 locale ones are taken care of just
1659 * above, or it leads to wrong results under ANYOF_INVERT */
1660 if (only_utf8_locale_invlist) {
1661 _invlist_union_maybe_complement_2nd(invlist,
1662 only_utf8_locale_invlist,
1663 ANYOF_FLAGS(node) & ANYOF_INVERT,
1670 /* These two functions currently do the exact same thing */
1671 #define ssc_init_zero ssc_init
1673 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1674 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1676 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1677 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1678 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1681 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1682 const regnode_charclass *and_with)
1684 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1685 * another SSC or a regular ANYOF class. Can create false positives. */
1690 PERL_ARGS_ASSERT_SSC_AND;
1692 assert(is_ANYOF_SYNTHETIC(ssc));
1694 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1695 * the code point inversion list and just the relevant flags */
1696 if (is_ANYOF_SYNTHETIC(and_with)) {
1697 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1698 anded_flags = ANYOF_FLAGS(and_with);
1700 /* XXX This is a kludge around what appears to be deficiencies in the
1701 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1702 * there are paths through the optimizer where it doesn't get weeded
1703 * out when it should. And if we don't make some extra provision for
1704 * it like the code just below, it doesn't get added when it should.
1705 * This solution is to add it only when AND'ing, which is here, and
1706 * only when what is being AND'ed is the pristine, original node
1707 * matching anything. Thus it is like adding it to ssc_anything() but
1708 * only when the result is to be AND'ed. Probably the same solution
1709 * could be adopted for the same problem we have with /l matching,
1710 * which is solved differently in S_ssc_init(), and that would lead to
1711 * fewer false positives than that solution has. But if this solution
1712 * creates bugs, the consequences are only that a warning isn't raised
1713 * that should be; while the consequences for having /l bugs is
1714 * incorrect matches */
1715 if (ssc_is_anything((regnode_ssc *)and_with)) {
1716 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1720 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1721 if (OP(and_with) == ANYOFD) {
1722 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1725 anded_flags = ANYOF_FLAGS(and_with)
1726 &( ANYOF_COMMON_FLAGS
1727 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1728 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1729 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(and_with))) {
1731 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1736 ANYOF_FLAGS(ssc) &= anded_flags;
1738 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1739 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1740 * 'and_with' may be inverted. When not inverted, we have the situation of
1742 * (C1 | P1) & (C2 | P2)
1743 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1744 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1745 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1746 * <= ((C1 & C2) | P1 | P2)
1747 * Alternatively, the last few steps could be:
1748 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1749 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1750 * <= (C1 | C2 | (P1 & P2))
1751 * We favor the second approach if either P1 or P2 is non-empty. This is
1752 * because these components are a barrier to doing optimizations, as what
1753 * they match cannot be known until the moment of matching as they are
1754 * dependent on the current locale, 'AND"ing them likely will reduce or
1756 * But we can do better if we know that C1,P1 are in their initial state (a
1757 * frequent occurrence), each matching everything:
1758 * (<everything>) & (C2 | P2) = C2 | P2
1759 * Similarly, if C2,P2 are in their initial state (again a frequent
1760 * occurrence), the result is a no-op
1761 * (C1 | P1) & (<everything>) = C1 | P1
1764 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1765 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1766 * <= (C1 & ~C2) | (P1 & ~P2)
1769 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1770 && ! is_ANYOF_SYNTHETIC(and_with))
1774 ssc_intersection(ssc,
1776 FALSE /* Has already been inverted */
1779 /* If either P1 or P2 is empty, the intersection will be also; can skip
1781 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
1782 ANYOF_POSIXL_ZERO(ssc);
1784 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1786 /* Note that the Posix class component P from 'and_with' actually
1788 * P = Pa | Pb | ... | Pn
1789 * where each component is one posix class, such as in [\w\s].
1791 * ~P = ~(Pa | Pb | ... | Pn)
1792 * = ~Pa & ~Pb & ... & ~Pn
1793 * <= ~Pa | ~Pb | ... | ~Pn
1794 * The last is something we can easily calculate, but unfortunately
1795 * is likely to have many false positives. We could do better
1796 * in some (but certainly not all) instances if two classes in
1797 * P have known relationships. For example
1798 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1800 * :lower: & :print: = :lower:
1801 * And similarly for classes that must be disjoint. For example,
1802 * since \s and \w can have no elements in common based on rules in
1803 * the POSIX standard,
1804 * \w & ^\S = nothing
1805 * Unfortunately, some vendor locales do not meet the Posix
1806 * standard, in particular almost everything by Microsoft.
1807 * The loop below just changes e.g., \w into \W and vice versa */
1809 regnode_charclass_posixl temp;
1810 int add = 1; /* To calculate the index of the complement */
1812 Zero(&temp, 1, regnode_charclass_posixl);
1813 ANYOF_POSIXL_ZERO(&temp);
1814 for (i = 0; i < ANYOF_MAX; i++) {
1816 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1817 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1819 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1820 ANYOF_POSIXL_SET(&temp, i + add);
1822 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1824 ANYOF_POSIXL_AND(&temp, ssc);
1826 } /* else ssc already has no posixes */
1827 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1828 in its initial state */
1829 else if (! is_ANYOF_SYNTHETIC(and_with)
1830 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1832 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1833 * copy it over 'ssc' */
1834 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1835 if (is_ANYOF_SYNTHETIC(and_with)) {
1836 StructCopy(and_with, ssc, regnode_ssc);
1839 ssc->invlist = anded_cp_list;
1840 ANYOF_POSIXL_ZERO(ssc);
1841 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1842 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1846 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1847 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
1849 /* One or the other of P1, P2 is non-empty. */
1850 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1851 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1853 ssc_union(ssc, anded_cp_list, FALSE);
1855 else { /* P1 = P2 = empty */
1856 ssc_intersection(ssc, anded_cp_list, FALSE);
1862 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1863 const regnode_charclass *or_with)
1865 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1866 * another SSC or a regular ANYOF class. Can create false positives if
1867 * 'or_with' is to be inverted. */
1872 PERL_ARGS_ASSERT_SSC_OR;
1874 assert(is_ANYOF_SYNTHETIC(ssc));
1876 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1877 * the code point inversion list and just the relevant flags */
1878 if (is_ANYOF_SYNTHETIC(or_with)) {
1879 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1880 ored_flags = ANYOF_FLAGS(or_with);
1883 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1884 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1885 if (OP(or_with) != ANYOFD) {
1887 |= ANYOF_FLAGS(or_with)
1888 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1889 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1890 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(or_with))) {
1892 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1897 ANYOF_FLAGS(ssc) |= ored_flags;
1899 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1900 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1901 * 'or_with' may be inverted. When not inverted, we have the simple
1902 * situation of computing:
1903 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1904 * If P1|P2 yields a situation with both a class and its complement are
1905 * set, like having both \w and \W, this matches all code points, and we
1906 * can delete these from the P component of the ssc going forward. XXX We
1907 * might be able to delete all the P components, but I (khw) am not certain
1908 * about this, and it is better to be safe.
1911 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1912 * <= (C1 | P1) | ~C2
1913 * <= (C1 | ~C2) | P1
1914 * (which results in actually simpler code than the non-inverted case)
1917 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1918 && ! is_ANYOF_SYNTHETIC(or_with))
1920 /* We ignore P2, leaving P1 going forward */
1921 } /* else Not inverted */
1922 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
1923 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1924 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1926 for (i = 0; i < ANYOF_MAX; i += 2) {
1927 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1929 ssc_match_all_cp(ssc);
1930 ANYOF_POSIXL_CLEAR(ssc, i);
1931 ANYOF_POSIXL_CLEAR(ssc, i+1);
1939 FALSE /* Already has been inverted */
1943 PERL_STATIC_INLINE void
1944 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1946 PERL_ARGS_ASSERT_SSC_UNION;
1948 assert(is_ANYOF_SYNTHETIC(ssc));
1950 _invlist_union_maybe_complement_2nd(ssc->invlist,
1956 PERL_STATIC_INLINE void
1957 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1959 const bool invert2nd)
1961 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1963 assert(is_ANYOF_SYNTHETIC(ssc));
1965 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1971 PERL_STATIC_INLINE void
1972 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1974 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1976 assert(is_ANYOF_SYNTHETIC(ssc));
1978 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1981 PERL_STATIC_INLINE void
1982 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1984 /* AND just the single code point 'cp' into the SSC 'ssc' */
1986 SV* cp_list = _new_invlist(2);
1988 PERL_ARGS_ASSERT_SSC_CP_AND;
1990 assert(is_ANYOF_SYNTHETIC(ssc));
1992 cp_list = add_cp_to_invlist(cp_list, cp);
1993 ssc_intersection(ssc, cp_list,
1994 FALSE /* Not inverted */
1996 SvREFCNT_dec_NN(cp_list);
1999 PERL_STATIC_INLINE void
2000 S_ssc_clear_locale(regnode_ssc *ssc)
2002 /* Set the SSC 'ssc' to not match any locale things */
2003 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
2005 assert(is_ANYOF_SYNTHETIC(ssc));
2007 ANYOF_POSIXL_ZERO(ssc);
2008 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
2011 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
2014 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
2016 /* The synthetic start class is used to hopefully quickly winnow down
2017 * places where a pattern could start a match in the target string. If it
2018 * doesn't really narrow things down that much, there isn't much point to
2019 * having the overhead of using it. This function uses some very crude
2020 * heuristics to decide if to use the ssc or not.
2022 * It returns TRUE if 'ssc' rules out more than half what it considers to
2023 * be the "likely" possible matches, but of course it doesn't know what the
2024 * actual things being matched are going to be; these are only guesses
2026 * For /l matches, it assumes that the only likely matches are going to be
2027 * in the 0-255 range, uniformly distributed, so half of that is 127
2028 * For /a and /d matches, it assumes that the likely matches will be just
2029 * the ASCII range, so half of that is 63
2030 * For /u and there isn't anything matching above the Latin1 range, it
2031 * assumes that that is the only range likely to be matched, and uses
2032 * half that as the cut-off: 127. If anything matches above Latin1,
2033 * it assumes that all of Unicode could match (uniformly), except for
2034 * non-Unicode code points and things in the General Category "Other"
2035 * (unassigned, private use, surrogates, controls and formats). This
2036 * is a much large number. */
2038 U32 count = 0; /* Running total of number of code points matched by
2040 UV start, end; /* Start and end points of current range in inversion
2042 const U32 max_code_points = (LOC)
2044 : (( ! RExC_uni_semantics
2045 || invlist_highest(ssc->invlist) < 256)
2048 const U32 max_match = max_code_points / 2;
2050 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
2052 invlist_iterinit(ssc->invlist);
2053 while (invlist_iternext(ssc->invlist, &start, &end)) {
2054 if (start >= max_code_points) {
2057 end = MIN(end, max_code_points - 1);
2058 count += end - start + 1;
2059 if (count >= max_match) {
2060 invlist_iterfinish(ssc->invlist);
2070 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
2072 /* The inversion list in the SSC is marked mortal; now we need a more
2073 * permanent copy, which is stored the same way that is done in a regular
2074 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
2077 SV* invlist = invlist_clone(ssc->invlist, NULL);
2079 PERL_ARGS_ASSERT_SSC_FINALIZE;
2081 assert(is_ANYOF_SYNTHETIC(ssc));
2083 /* The code in this file assumes that all but these flags aren't relevant
2084 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
2085 * by the time we reach here */
2086 assert(! (ANYOF_FLAGS(ssc)
2087 & ~( ANYOF_COMMON_FLAGS
2088 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
2089 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
2091 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
2093 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
2094 NULL, NULL, NULL, FALSE);
2096 /* Make sure is clone-safe */
2097 ssc->invlist = NULL;
2099 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
2100 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
2101 OP(ssc) = ANYOFPOSIXL;
2103 else if (RExC_contains_locale) {
2107 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
2110 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
2111 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
2112 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
2113 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
2114 ? (TRIE_LIST_CUR( idx ) - 1) \
2120 dump_trie(trie,widecharmap,revcharmap)
2121 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
2122 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
2124 These routines dump out a trie in a somewhat readable format.
2125 The _interim_ variants are used for debugging the interim
2126 tables that are used to generate the final compressed
2127 representation which is what dump_trie expects.
2129 Part of the reason for their existence is to provide a form
2130 of documentation as to how the different representations function.
2135 Dumps the final compressed table form of the trie to Perl_debug_log.
2136 Used for debugging make_trie().
2140 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
2141 AV *revcharmap, U32 depth)
2144 SV *sv=sv_newmortal();
2145 int colwidth= widecharmap ? 6 : 4;
2147 GET_RE_DEBUG_FLAGS_DECL;
2149 PERL_ARGS_ASSERT_DUMP_TRIE;
2151 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
2152 depth+1, "Match","Base","Ofs" );
2154 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2155 SV ** const tmp = av_fetch( revcharmap, state, 0);
2157 Perl_re_printf( aTHX_ "%*s",
2159 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2160 PL_colors[0], PL_colors[1],
2161 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2162 PERL_PV_ESCAPE_FIRSTCHAR
2167 Perl_re_printf( aTHX_ "\n");
2168 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
2170 for( state = 0 ; state < trie->uniquecharcount ; state++ )
2171 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2172 Perl_re_printf( aTHX_ "\n");
2174 for( state = 1 ; state < trie->statecount ; state++ ) {
2175 const U32 base = trie->states[ state ].trans.base;
2177 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
2179 if ( trie->states[ state ].wordnum ) {
2180 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
2182 Perl_re_printf( aTHX_ "%6s", "" );
2185 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
2190 while( ( base + ofs < trie->uniquecharcount ) ||
2191 ( base + ofs - trie->uniquecharcount < trie->lasttrans
2192 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2196 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
2198 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2199 if ( ( base + ofs >= trie->uniquecharcount )
2200 && ( base + ofs - trie->uniquecharcount
2202 && trie->trans[ base + ofs
2203 - trie->uniquecharcount ].check == state )
2205 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
2206 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2209 Perl_re_printf( aTHX_ "%*s", colwidth," ." );
2213 Perl_re_printf( aTHX_ "]");
2216 Perl_re_printf( aTHX_ "\n" );
2218 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
2220 for (word=1; word <= trie->wordcount; word++) {
2221 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2222 (int)word, (int)(trie->wordinfo[word].prev),
2223 (int)(trie->wordinfo[word].len));
2225 Perl_re_printf( aTHX_ "\n" );
2228 Dumps a fully constructed but uncompressed trie in list form.
2229 List tries normally only are used for construction when the number of
2230 possible chars (trie->uniquecharcount) is very high.
2231 Used for debugging make_trie().
2234 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2235 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2239 SV *sv=sv_newmortal();
2240 int colwidth= widecharmap ? 6 : 4;
2241 GET_RE_DEBUG_FLAGS_DECL;
2243 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2245 /* print out the table precompression. */
2246 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
2248 Perl_re_indentf( aTHX_ "%s",
2249 depth+1, "------:-----+-----------------\n" );
2251 for( state=1 ; state < next_alloc ; state ++ ) {
2254 Perl_re_indentf( aTHX_ " %4" UVXf " :",
2255 depth+1, (UV)state );
2256 if ( ! trie->states[ state ].wordnum ) {
2257 Perl_re_printf( aTHX_ "%5s| ","");
2259 Perl_re_printf( aTHX_ "W%4x| ",
2260 trie->states[ state ].wordnum
2263 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2264 SV ** const tmp = av_fetch( revcharmap,
2265 TRIE_LIST_ITEM(state, charid).forid, 0);
2267 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
2269 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2271 PL_colors[0], PL_colors[1],
2272 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2273 | PERL_PV_ESCAPE_FIRSTCHAR
2275 TRIE_LIST_ITEM(state, charid).forid,
2276 (UV)TRIE_LIST_ITEM(state, charid).newstate
2279 Perl_re_printf( aTHX_ "\n%*s| ",
2280 (int)((depth * 2) + 14), "");
2283 Perl_re_printf( aTHX_ "\n");
2288 Dumps a fully constructed but uncompressed trie in table form.
2289 This is the normal DFA style state transition table, with a few
2290 twists to facilitate compression later.
2291 Used for debugging make_trie().
2294 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2295 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2300 SV *sv=sv_newmortal();
2301 int colwidth= widecharmap ? 6 : 4;
2302 GET_RE_DEBUG_FLAGS_DECL;
2304 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2307 print out the table precompression so that we can do a visual check
2308 that they are identical.
2311 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
2313 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2314 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2316 Perl_re_printf( aTHX_ "%*s",
2318 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2319 PL_colors[0], PL_colors[1],
2320 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2321 PERL_PV_ESCAPE_FIRSTCHAR
2327 Perl_re_printf( aTHX_ "\n");
2328 Perl_re_indentf( aTHX_ "State+-", depth+1 );
2330 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2331 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
2334 Perl_re_printf( aTHX_ "\n" );
2336 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2338 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
2340 (UV)TRIE_NODENUM( state ) );
2342 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2343 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2345 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
2347 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
2349 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2350 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
2351 (UV)trie->trans[ state ].check );
2353 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
2354 (UV)trie->trans[ state ].check,
2355 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2363 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2364 startbranch: the first branch in the whole branch sequence
2365 first : start branch of sequence of branch-exact nodes.
2366 May be the same as startbranch
2367 last : Thing following the last branch.
2368 May be the same as tail.
2369 tail : item following the branch sequence
2370 count : words in the sequence
2371 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2372 depth : indent depth
2374 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2376 A trie is an N'ary tree where the branches are determined by digital
2377 decomposition of the key. IE, at the root node you look up the 1st character and
2378 follow that branch repeat until you find the end of the branches. Nodes can be
2379 marked as "accepting" meaning they represent a complete word. Eg:
2383 would convert into the following structure. Numbers represent states, letters
2384 following numbers represent valid transitions on the letter from that state, if
2385 the number is in square brackets it represents an accepting state, otherwise it
2386 will be in parenthesis.
2388 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2392 (1) +-i->(6)-+-s->[7]
2394 +-s->(3)-+-h->(4)-+-e->[5]
2396 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2398 This shows that when matching against the string 'hers' we will begin at state 1
2399 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2400 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2401 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2402 single traverse. We store a mapping from accepting to state to which word was
2403 matched, and then when we have multiple possibilities we try to complete the
2404 rest of the regex in the order in which they occurred in the alternation.
2406 The only prior NFA like behaviour that would be changed by the TRIE support is
2407 the silent ignoring of duplicate alternations which are of the form:
2409 / (DUPE|DUPE) X? (?{ ... }) Y /x
2411 Thus EVAL blocks following a trie may be called a different number of times with
2412 and without the optimisation. With the optimisations dupes will be silently
2413 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2414 the following demonstrates:
2416 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2418 which prints out 'word' three times, but
2420 'words'=~/(word|word|word)(?{ print $1 })S/
2422 which doesnt print it out at all. This is due to other optimisations kicking in.
2424 Example of what happens on a structural level:
2426 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2428 1: CURLYM[1] {1,32767}(18)
2439 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2440 and should turn into:
2442 1: CURLYM[1] {1,32767}(18)
2444 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2452 Cases where tail != last would be like /(?foo|bar)baz/:
2462 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2463 and would end up looking like:
2466 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2473 d = uvchr_to_utf8_flags(d, uv, 0);
2475 is the recommended Unicode-aware way of saying
2480 #define TRIE_STORE_REVCHAR(val) \
2483 SV *zlopp = newSV(UTF8_MAXBYTES); \
2484 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2485 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2486 SvCUR_set(zlopp, kapow - flrbbbbb); \
2489 av_push(revcharmap, zlopp); \
2491 char ooooff = (char)val; \
2492 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2496 /* This gets the next character from the input, folding it if not already
2498 #define TRIE_READ_CHAR STMT_START { \
2501 /* if it is UTF then it is either already folded, or does not need \
2503 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2505 else if (folder == PL_fold_latin1) { \
2506 /* This folder implies Unicode rules, which in the range expressible \
2507 * by not UTF is the lower case, with the two exceptions, one of \
2508 * which should have been taken care of before calling this */ \
2509 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2510 uvc = toLOWER_L1(*uc); \
2511 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2514 /* raw data, will be folded later if needed */ \
2522 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2523 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2524 U32 ging = TRIE_LIST_LEN( state ) * 2; \
2525 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2526 TRIE_LIST_LEN( state ) = ging; \
2528 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2529 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2530 TRIE_LIST_CUR( state )++; \
2533 #define TRIE_LIST_NEW(state) STMT_START { \
2534 Newx( trie->states[ state ].trans.list, \
2535 4, reg_trie_trans_le ); \
2536 TRIE_LIST_CUR( state ) = 1; \
2537 TRIE_LIST_LEN( state ) = 4; \
2540 #define TRIE_HANDLE_WORD(state) STMT_START { \
2541 U16 dupe= trie->states[ state ].wordnum; \
2542 regnode * const noper_next = regnext( noper ); \
2545 /* store the word for dumping */ \
2547 if (OP(noper) != NOTHING) \
2548 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2550 tmp = newSVpvn_utf8( "", 0, UTF ); \
2551 av_push( trie_words, tmp ); \
2555 trie->wordinfo[curword].prev = 0; \
2556 trie->wordinfo[curword].len = wordlen; \
2557 trie->wordinfo[curword].accept = state; \
2559 if ( noper_next < tail ) { \
2561 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2563 trie->jump[curword] = (U16)(noper_next - convert); \
2565 jumper = noper_next; \
2567 nextbranch= regnext(cur); \
2571 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2572 /* chain, so that when the bits of chain are later */\
2573 /* linked together, the dups appear in the chain */\
2574 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2575 trie->wordinfo[dupe].prev = curword; \
2577 /* we haven't inserted this word yet. */ \
2578 trie->states[ state ].wordnum = curword; \
2583 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2584 ( ( base + charid >= ucharcount \
2585 && base + charid < ubound \
2586 && state == trie->trans[ base - ucharcount + charid ].check \
2587 && trie->trans[ base - ucharcount + charid ].next ) \
2588 ? trie->trans[ base - ucharcount + charid ].next \
2589 : ( state==1 ? special : 0 ) \
2592 #define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2594 TRIE_BITMAP_SET(trie, uvc); \
2595 /* store the folded codepoint */ \
2597 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2600 /* store first byte of utf8 representation of */ \
2601 /* variant codepoints */ \
2602 if (! UVCHR_IS_INVARIANT(uvc)) { \
2603 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2608 #define MADE_JUMP_TRIE 2
2609 #define MADE_EXACT_TRIE 4
2612 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2613 regnode *first, regnode *last, regnode *tail,
2614 U32 word_count, U32 flags, U32 depth)
2616 /* first pass, loop through and scan words */
2617 reg_trie_data *trie;
2618 HV *widecharmap = NULL;
2619 AV *revcharmap = newAV();
2625 regnode *jumper = NULL;
2626 regnode *nextbranch = NULL;
2627 regnode *convert = NULL;
2628 U32 *prev_states; /* temp array mapping each state to previous one */
2629 /* we just use folder as a flag in utf8 */
2630 const U8 * folder = NULL;
2632 /* in the below add_data call we are storing either 'tu' or 'tuaa'
2633 * which stands for one trie structure, one hash, optionally followed
2636 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuaa"));
2637 AV *trie_words = NULL;
2638 /* along with revcharmap, this only used during construction but both are
2639 * useful during debugging so we store them in the struct when debugging.
2642 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2643 STRLEN trie_charcount=0;
2645 SV *re_trie_maxbuff;
2646 GET_RE_DEBUG_FLAGS_DECL;
2648 PERL_ARGS_ASSERT_MAKE_TRIE;
2650 PERL_UNUSED_ARG(depth);
2654 case EXACT: case EXACT_ONLY8: case EXACTL: break;
2658 case EXACTFLU8: folder = PL_fold_latin1; break;
2659 case EXACTF: folder = PL_fold; break;
2660 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2663 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2665 trie->startstate = 1;
2666 trie->wordcount = word_count;
2667 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2668 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2669 if (flags == EXACT || flags == EXACT_ONLY8 || flags == EXACTL)
2670 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2671 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2672 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2675 trie_words = newAV();
2678 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2679 assert(re_trie_maxbuff);
2680 if (!SvIOK(re_trie_maxbuff)) {
2681 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2683 DEBUG_TRIE_COMPILE_r({
2684 Perl_re_indentf( aTHX_
2685 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2687 REG_NODE_NUM(startbranch), REG_NODE_NUM(first),
2688 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2691 /* Find the node we are going to overwrite */
2692 if ( first == startbranch && OP( last ) != BRANCH ) {
2693 /* whole branch chain */
2696 /* branch sub-chain */
2697 convert = NEXTOPER( first );
2700 /* -- First loop and Setup --
2702 We first traverse the branches and scan each word to determine if it
2703 contains widechars, and how many unique chars there are, this is
2704 important as we have to build a table with at least as many columns as we
2707 We use an array of integers to represent the character codes 0..255
2708 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2709 the native representation of the character value as the key and IV's for
2712 *TODO* If we keep track of how many times each character is used we can
2713 remap the columns so that the table compression later on is more
2714 efficient in terms of memory by ensuring the most common value is in the
2715 middle and the least common are on the outside. IMO this would be better
2716 than a most to least common mapping as theres a decent chance the most
2717 common letter will share a node with the least common, meaning the node
2718 will not be compressible. With a middle is most common approach the worst
2719 case is when we have the least common nodes twice.
2723 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2724 regnode *noper = NEXTOPER( cur );
2728 U32 wordlen = 0; /* required init */
2729 STRLEN minchars = 0;
2730 STRLEN maxchars = 0;
2731 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2734 if (OP(noper) == NOTHING) {
2735 /* skip past a NOTHING at the start of an alternation
2736 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2738 regnode *noper_next= regnext(noper);
2739 if (noper_next < tail)
2744 && ( OP(noper) == flags
2745 || (flags == EXACT && OP(noper) == EXACT_ONLY8)
2746 || (flags == EXACTFU && ( OP(noper) == EXACTFU_ONLY8
2747 || OP(noper) == EXACTFU_SS))) )
2749 uc= (U8*)STRING(noper);
2750 e= uc + STR_LEN(noper);
2757 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2758 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2759 regardless of encoding */
2760 if (OP( noper ) == EXACTFU_SS) {
2761 /* false positives are ok, so just set this */
2762 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2766 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2768 TRIE_CHARCOUNT(trie)++;
2771 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2772 * is in effect. Under /i, this character can match itself, or
2773 * anything that folds to it. If not under /i, it can match just
2774 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2775 * all fold to k, and all are single characters. But some folds
2776 * expand to more than one character, so for example LATIN SMALL
2777 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2778 * the string beginning at 'uc' is 'ffi', it could be matched by
2779 * three characters, or just by the one ligature character. (It
2780 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2781 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2782 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2783 * match.) The trie needs to know the minimum and maximum number
2784 * of characters that could match so that it can use size alone to
2785 * quickly reject many match attempts. The max is simple: it is
2786 * the number of folded characters in this branch (since a fold is
2787 * never shorter than what folds to it. */
2791 /* And the min is equal to the max if not under /i (indicated by
2792 * 'folder' being NULL), or there are no multi-character folds. If
2793 * there is a multi-character fold, the min is incremented just
2794 * once, for the character that folds to the sequence. Each
2795 * character in the sequence needs to be added to the list below of
2796 * characters in the trie, but we count only the first towards the
2797 * min number of characters needed. This is done through the
2798 * variable 'foldlen', which is returned by the macros that look
2799 * for these sequences as the number of bytes the sequence
2800 * occupies. Each time through the loop, we decrement 'foldlen' by
2801 * how many bytes the current char occupies. Only when it reaches
2802 * 0 do we increment 'minchars' or look for another multi-character
2804 if (folder == NULL) {
2807 else if (foldlen > 0) {
2808 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2813 /* See if *uc is the beginning of a multi-character fold. If
2814 * so, we decrement the length remaining to look at, to account
2815 * for the current character this iteration. (We can use 'uc'
2816 * instead of the fold returned by TRIE_READ_CHAR because for
2817 * non-UTF, the latin1_safe macro is smart enough to account
2818 * for all the unfolded characters, and because for UTF, the
2819 * string will already have been folded earlier in the
2820 * compilation process */
2822 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2823 foldlen -= UTF8SKIP(uc);
2826 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2831 /* The current character (and any potential folds) should be added
2832 * to the possible matching characters for this position in this
2836 U8 folded= folder[ (U8) uvc ];
2837 if ( !trie->charmap[ folded ] ) {
2838 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2839 TRIE_STORE_REVCHAR( folded );
2842 if ( !trie->charmap[ uvc ] ) {
2843 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2844 TRIE_STORE_REVCHAR( uvc );
2847 /* store the codepoint in the bitmap, and its folded
2849 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
2850 set_bit = 0; /* We've done our bit :-) */
2854 /* XXX We could come up with the list of code points that fold
2855 * to this using PL_utf8_foldclosures, except not for
2856 * multi-char folds, as there may be multiple combinations
2857 * there that could work, which needs to wait until runtime to
2858 * resolve (The comment about LIGATURE FFI above is such an
2863 widecharmap = newHV();
2865 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2868 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
2870 if ( !SvTRUE( *svpp ) ) {
2871 sv_setiv( *svpp, ++trie->uniquecharcount );
2872 TRIE_STORE_REVCHAR(uvc);
2875 } /* end loop through characters in this branch of the trie */
2877 /* We take the min and max for this branch and combine to find the min
2878 * and max for all branches processed so far */
2879 if( cur == first ) {
2880 trie->minlen = minchars;
2881 trie->maxlen = maxchars;
2882 } else if (minchars < trie->minlen) {
2883 trie->minlen = minchars;
2884 } else if (maxchars > trie->maxlen) {
2885 trie->maxlen = maxchars;
2887 } /* end first pass */
2888 DEBUG_TRIE_COMPILE_r(
2889 Perl_re_indentf( aTHX_
2890 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2892 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2893 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2894 (int)trie->minlen, (int)trie->maxlen )
2898 We now know what we are dealing with in terms of unique chars and
2899 string sizes so we can calculate how much memory a naive
2900 representation using a flat table will take. If it's over a reasonable
2901 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2902 conservative but potentially much slower representation using an array
2905 At the end we convert both representations into the same compressed
2906 form that will be used in regexec.c for matching with. The latter
2907 is a form that cannot be used to construct with but has memory
2908 properties similar to the list form and access properties similar
2909 to the table form making it both suitable for fast searches and
2910 small enough that its feasable to store for the duration of a program.
2912 See the comment in the code where the compressed table is produced
2913 inplace from the flat tabe representation for an explanation of how
2914 the compression works.
2919 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2922 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2923 > SvIV(re_trie_maxbuff) )
2926 Second Pass -- Array Of Lists Representation
2928 Each state will be represented by a list of charid:state records
2929 (reg_trie_trans_le) the first such element holds the CUR and LEN
2930 points of the allocated array. (See defines above).
2932 We build the initial structure using the lists, and then convert
2933 it into the compressed table form which allows faster lookups
2934 (but cant be modified once converted).
2937 STRLEN transcount = 1;
2939 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
2942 trie->states = (reg_trie_state *)
2943 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2944 sizeof(reg_trie_state) );
2948 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2950 regnode *noper = NEXTOPER( cur );
2951 U32 state = 1; /* required init */
2952 U16 charid = 0; /* sanity init */
2953 U32 wordlen = 0; /* required init */
2955 if (OP(noper) == NOTHING) {
2956 regnode *noper_next= regnext(noper);
2957 if (noper_next < tail)
2962 && ( OP(noper) == flags
2963 || (flags == EXACT && OP(noper) == EXACT_ONLY8)
2964 || (flags == EXACTFU && ( OP(noper) == EXACTFU_ONLY8
2965 || OP(noper) == EXACTFU_SS))) )
2967 const U8 *uc= (U8*)STRING(noper);
2968 const U8 *e= uc + STR_LEN(noper);
2970 for ( ; uc < e ; uc += len ) {
2975 charid = trie->charmap[ uvc ];
2977 SV** const svpp = hv_fetch( widecharmap,
2984 charid=(U16)SvIV( *svpp );
2987 /* charid is now 0 if we dont know the char read, or
2988 * nonzero if we do */
2995 if ( !trie->states[ state ].trans.list ) {
2996 TRIE_LIST_NEW( state );
2999 check <= TRIE_LIST_USED( state );
3002 if ( TRIE_LIST_ITEM( state, check ).forid
3005 newstate = TRIE_LIST_ITEM( state, check ).newstate;
3010 newstate = next_alloc++;
3011 prev_states[newstate] = state;
3012 TRIE_LIST_PUSH( state, charid, newstate );
3017 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3021 TRIE_HANDLE_WORD(state);
3023 } /* end second pass */
3025 /* next alloc is the NEXT state to be allocated */
3026 trie->statecount = next_alloc;
3027 trie->states = (reg_trie_state *)
3028 PerlMemShared_realloc( trie->states,
3030 * sizeof(reg_trie_state) );
3032 /* and now dump it out before we compress it */
3033 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
3034 revcharmap, next_alloc,
3038 trie->trans = (reg_trie_trans *)
3039 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
3046 for( state=1 ; state < next_alloc ; state ++ ) {
3050 DEBUG_TRIE_COMPILE_MORE_r(
3051 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
3055 if (trie->states[state].trans.list) {
3056 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
3060 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3061 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
3062 if ( forid < minid ) {
3064 } else if ( forid > maxid ) {
3068 if ( transcount < tp + maxid - minid + 1) {
3070 trie->trans = (reg_trie_trans *)
3071 PerlMemShared_realloc( trie->trans,
3073 * sizeof(reg_trie_trans) );
3074 Zero( trie->trans + (transcount / 2),
3078 base = trie->uniquecharcount + tp - minid;
3079 if ( maxid == minid ) {
3081 for ( ; zp < tp ; zp++ ) {
3082 if ( ! trie->trans[ zp ].next ) {
3083 base = trie->uniquecharcount + zp - minid;
3084 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
3086 trie->trans[ zp ].check = state;
3092 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
3094 trie->trans[ tp ].check = state;
3099 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3100 const U32 tid = base
3101 - trie->uniquecharcount
3102 + TRIE_LIST_ITEM( state, idx ).forid;
3103 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
3105 trie->trans[ tid ].check = state;
3107 tp += ( maxid - minid + 1 );
3109 Safefree(trie->states[ state ].trans.list);
3112 DEBUG_TRIE_COMPILE_MORE_r(
3113 Perl_re_printf( aTHX_ " base: %d\n",base);
3116 trie->states[ state ].trans.base=base;
3118 trie->lasttrans = tp + 1;
3122 Second Pass -- Flat Table Representation.
3124 we dont use the 0 slot of either trans[] or states[] so we add 1 to
3125 each. We know that we will need Charcount+1 trans at most to store
3126 the data (one row per char at worst case) So we preallocate both
3127 structures assuming worst case.
3129 We then construct the trie using only the .next slots of the entry
3132 We use the .check field of the first entry of the node temporarily
3133 to make compression both faster and easier by keeping track of how
3134 many non zero fields are in the node.
3136 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
3139 There are two terms at use here: state as a TRIE_NODEIDX() which is
3140 a number representing the first entry of the node, and state as a
3141 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
3142 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
3143 if there are 2 entrys per node. eg:
3151 The table is internally in the right hand, idx form. However as we
3152 also have to deal with the states array which is indexed by nodenum
3153 we have to use TRIE_NODENUM() to convert.
3156 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
3159 trie->trans = (reg_trie_trans *)
3160 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
3161 * trie->uniquecharcount + 1,
3162 sizeof(reg_trie_trans) );
3163 trie->states = (reg_trie_state *)
3164 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3165 sizeof(reg_trie_state) );
3166 next_alloc = trie->uniquecharcount + 1;
3169 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3171 regnode *noper = NEXTOPER( cur );
3173 U32 state = 1; /* required init */
3175 U16 charid = 0; /* sanity init */
3176 U32 accept_state = 0; /* sanity init */
3178 U32 wordlen = 0; /* required init */
3180 if (OP(noper) == NOTHING) {
3181 regnode *noper_next= regnext(noper);
3182 if (noper_next < tail)
3187 && ( OP(noper) == flags
3188 || (flags == EXACT && OP(noper) == EXACT_ONLY8)
3189 || (flags == EXACTFU && ( OP(noper) == EXACTFU_ONLY8
3190 || OP(noper) == EXACTFU_SS))) )
3192 const U8 *uc= (U8*)STRING(noper);
3193 const U8 *e= uc + STR_LEN(noper);
3195 for ( ; uc < e ; uc += len ) {
3200 charid = trie->charmap[ uvc ];
3202 SV* const * const svpp = hv_fetch( widecharmap,
3206 charid = svpp ? (U16)SvIV(*svpp) : 0;
3210 if ( !trie->trans[ state + charid ].next ) {
3211 trie->trans[ state + charid ].next = next_alloc;
3212 trie->trans[ state ].check++;
3213 prev_states[TRIE_NODENUM(next_alloc)]
3214 = TRIE_NODENUM(state);
3215 next_alloc += trie->uniquecharcount;
3217 state = trie->trans[ state + charid ].next;
3219 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3221 /* charid is now 0 if we dont know the char read, or
3222 * nonzero if we do */
3225 accept_state = TRIE_NODENUM( state );
3226 TRIE_HANDLE_WORD(accept_state);
3228 } /* end second pass */
3230 /* and now dump it out before we compress it */
3231 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3233 next_alloc, depth+1));
3237 * Inplace compress the table.*
3239 For sparse data sets the table constructed by the trie algorithm will
3240 be mostly 0/FAIL transitions or to put it another way mostly empty.
3241 (Note that leaf nodes will not contain any transitions.)
3243 This algorithm compresses the tables by eliminating most such
3244 transitions, at the cost of a modest bit of extra work during lookup:
3246 - Each states[] entry contains a .base field which indicates the
3247 index in the state[] array wheres its transition data is stored.
3249 - If .base is 0 there are no valid transitions from that node.
3251 - If .base is nonzero then charid is added to it to find an entry in
3254 -If trans[states[state].base+charid].check!=state then the
3255 transition is taken to be a 0/Fail transition. Thus if there are fail
3256 transitions at the front of the node then the .base offset will point
3257 somewhere inside the previous nodes data (or maybe even into a node
3258 even earlier), but the .check field determines if the transition is
3262 The following process inplace converts the table to the compressed
3263 table: We first do not compress the root node 1,and mark all its
3264 .check pointers as 1 and set its .base pointer as 1 as well. This
3265 allows us to do a DFA construction from the compressed table later,
3266 and ensures that any .base pointers we calculate later are greater
3269 - We set 'pos' to indicate the first entry of the second node.
3271 - We then iterate over the columns of the node, finding the first and
3272 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3273 and set the .check pointers accordingly, and advance pos
3274 appropriately and repreat for the next node. Note that when we copy
3275 the next pointers we have to convert them from the original
3276 NODEIDX form to NODENUM form as the former is not valid post
3279 - If a node has no transitions used we mark its base as 0 and do not
3280 advance the pos pointer.
3282 - If a node only has one transition we use a second pointer into the
3283 structure to fill in allocated fail transitions from other states.
3284 This pointer is independent of the main pointer and scans forward
3285 looking for null transitions that are allocated to a state. When it
3286 finds one it writes the single transition into the "hole". If the
3287 pointer doesnt find one the single transition is appended as normal.
3289 - Once compressed we can Renew/realloc the structures to release the
3292 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3293 specifically Fig 3.47 and the associated pseudocode.
3297 const U32 laststate = TRIE_NODENUM( next_alloc );
3300 trie->statecount = laststate;
3302 for ( state = 1 ; state < laststate ; state++ ) {
3304 const U32 stateidx = TRIE_NODEIDX( state );
3305 const U32 o_used = trie->trans[ stateidx ].check;
3306 U32 used = trie->trans[ stateidx ].check;
3307 trie->trans[ stateidx ].check = 0;
3310 used && charid < trie->uniquecharcount;
3313 if ( flag || trie->trans[ stateidx + charid ].next ) {
3314 if ( trie->trans[ stateidx + charid ].next ) {
3316 for ( ; zp < pos ; zp++ ) {
3317 if ( ! trie->trans[ zp ].next ) {
3321 trie->states[ state ].trans.base
3323 + trie->uniquecharcount
3325 trie->trans[ zp ].next
3326 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3328 trie->trans[ zp ].check = state;
3329 if ( ++zp > pos ) pos = zp;
3336 trie->states[ state ].trans.base
3337 = pos + trie->uniquecharcount - charid ;
3339 trie->trans[ pos ].next
3340 = SAFE_TRIE_NODENUM(
3341 trie->trans[ stateidx + charid ].next );
3342 trie->trans[ pos ].check = state;
3347 trie->lasttrans = pos + 1;
3348 trie->states = (reg_trie_state *)
3349 PerlMemShared_realloc( trie->states, laststate
3350 * sizeof(reg_trie_state) );
3351 DEBUG_TRIE_COMPILE_MORE_r(
3352 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
3354 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3358 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3361 } /* end table compress */
3363 DEBUG_TRIE_COMPILE_MORE_r(
3364 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
3366 (UV)trie->statecount,
3367 (UV)trie->lasttrans)
3369 /* resize the trans array to remove unused space */
3370 trie->trans = (reg_trie_trans *)
3371 PerlMemShared_realloc( trie->trans, trie->lasttrans
3372 * sizeof(reg_trie_trans) );
3374 { /* Modify the program and insert the new TRIE node */
3375 U8 nodetype =(U8)(flags & 0xFF);
3379 regnode *optimize = NULL;
3380 #ifdef RE_TRACK_PATTERN_OFFSETS
3383 U32 mjd_nodelen = 0;
3384 #endif /* RE_TRACK_PATTERN_OFFSETS */
3385 #endif /* DEBUGGING */
3387 This means we convert either the first branch or the first Exact,
3388 depending on whether the thing following (in 'last') is a branch
3389 or not and whther first is the startbranch (ie is it a sub part of
3390 the alternation or is it the whole thing.)
3391 Assuming its a sub part we convert the EXACT otherwise we convert
3392 the whole branch sequence, including the first.
3394 /* Find the node we are going to overwrite */
3395 if ( first != startbranch || OP( last ) == BRANCH ) {
3396 /* branch sub-chain */
3397 NEXT_OFF( first ) = (U16)(last - first);
3398 #ifdef RE_TRACK_PATTERN_OFFSETS
3400 mjd_offset= Node_Offset((convert));
3401 mjd_nodelen= Node_Length((convert));
3404 /* whole branch chain */
3406 #ifdef RE_TRACK_PATTERN_OFFSETS
3409 const regnode *nop = NEXTOPER( convert );
3410 mjd_offset= Node_Offset((nop));
3411 mjd_nodelen= Node_Length((nop));
3415 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
3417 (UV)mjd_offset, (UV)mjd_nodelen)
3420 /* But first we check to see if there is a common prefix we can
3421 split out as an EXACT and put in front of the TRIE node. */
3422 trie->startstate= 1;
3423 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3424 /* we want to find the first state that has more than
3425 * one transition, if that state is not the first state
3426 * then we have a common prefix which we can remove.
3429 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3431 I32 first_ofs = -1; /* keeps track of the ofs of the first
3432 transition, -1 means none */
3434 const U32 base = trie->states[ state ].trans.base;
3436 /* does this state terminate an alternation? */
3437 if ( trie->states[state].wordnum )
3440 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3441 if ( ( base + ofs >= trie->uniquecharcount ) &&
3442 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3443 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3445 if ( ++count > 1 ) {
3446 /* we have more than one transition */
3449 /* if this is the first state there is no common prefix
3450 * to extract, so we can exit */
3451 if ( state == 1 ) break;
3452 tmp = av_fetch( revcharmap, ofs, 0);
3453 ch = (U8*)SvPV_nolen_const( *tmp );
3455 /* if we are on count 2 then we need to initialize the
3456 * bitmap, and store the previous char if there was one
3459 /* clear the bitmap */
3460 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3462 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
3465 if (first_ofs >= 0) {
3466 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
3467 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3469 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3471 Perl_re_printf( aTHX_ "%s", (char*)ch)
3475 /* store the current firstchar in the bitmap */
3476 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3477 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
3483 /* This state has only one transition, its transition is part
3484 * of a common prefix - we need to concatenate the char it
3485 * represents to what we have so far. */
3486 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
3488 char *ch = SvPV( *tmp, len );
3490 SV *sv=sv_newmortal();
3491 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
3493 (UV)state, (UV)first_ofs,
3494 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3495 PL_colors[0], PL_colors[1],
3496 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3497 PERL_PV_ESCAPE_FIRSTCHAR
3502 OP( convert ) = nodetype;
3503 str=STRING(convert);
3506 STR_LEN(convert) += len;
3512 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
3517 trie->prefixlen = (state-1);
3519 regnode *n = convert+NODE_SZ_STR(convert);
3520 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3521 trie->startstate = state;
3522 trie->minlen -= (state - 1);
3523 trie->maxlen -= (state - 1);
3525 /* At least the UNICOS C compiler choked on this
3526 * being argument to DEBUG_r(), so let's just have
3529 #ifdef PERL_EXT_RE_BUILD
3535 regnode *fix = convert;
3536 U32 word = trie->wordcount;
3537 #ifdef RE_TRACK_PATTERN_OFFSETS
3540 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3541 while( ++fix < n ) {
3542 Set_Node_Offset_Length(fix, 0, 0);
3545 SV ** const tmp = av_fetch( trie_words, word, 0 );
3547 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3548 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3550 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3558 NEXT_OFF(convert) = (U16)(tail - convert);
3559 DEBUG_r(optimize= n);
3565 if ( trie->maxlen ) {
3566 NEXT_OFF( convert ) = (U16)(tail - convert);
3567 ARG_SET( convert, data_slot );
3568 /* Store the offset to the first unabsorbed branch in
3569 jump[0], which is otherwise unused by the jump logic.
3570 We use this when dumping a trie and during optimisation. */
3572 trie->jump[0] = (U16)(nextbranch - convert);
3574 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3575 * and there is a bitmap
3576 * and the first "jump target" node we found leaves enough room
3577 * then convert the TRIE node into a TRIEC node, with the bitmap
3578 * embedded inline in the opcode - this is hypothetically faster.
3580 if ( !trie->states[trie->startstate].wordnum
3582 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3584 OP( convert ) = TRIEC;
3585 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3586 PerlMemShared_free(trie->bitmap);
3589 OP( convert ) = TRIE;
3591 /* store the type in the flags */
3592 convert->flags = nodetype;
3596 + regarglen[ OP( convert ) ];
3598 /* XXX We really should free up the resource in trie now,
3599 as we won't use them - (which resources?) dmq */
3601 /* needed for dumping*/
3602 DEBUG_r(if (optimize) {
3603 regnode *opt = convert;
3605 while ( ++opt < optimize) {
3606 Set_Node_Offset_Length(opt, 0, 0);
3609 Try to clean up some of the debris left after the
3612 while( optimize < jumper ) {
3613 Track_Code( mjd_nodelen += Node_Length((optimize)); );
3614 OP( optimize ) = OPTIMIZED;
3615 Set_Node_Offset_Length(optimize, 0, 0);
3618 Set_Node_Offset_Length(convert, mjd_offset, mjd_nodelen);
3620 } /* end node insert */
3622 /* Finish populating the prev field of the wordinfo array. Walk back
3623 * from each accept state until we find another accept state, and if
3624 * so, point the first word's .prev field at the second word. If the
3625 * second already has a .prev field set, stop now. This will be the
3626 * case either if we've already processed that word's accept state,
3627 * or that state had multiple words, and the overspill words were
3628 * already linked up earlier.
3635 for (word=1; word <= trie->wordcount; word++) {
3637 if (trie->wordinfo[word].prev)
3639 state = trie->wordinfo[word].accept;
3641 state = prev_states[state];
3644 prev = trie->states[state].wordnum;
3648 trie->wordinfo[word].prev = prev;
3650 Safefree(prev_states);
3654 /* and now dump out the compressed format */
3655 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3657 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3659 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3660 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3662 SvREFCNT_dec_NN(revcharmap);
3666 : trie->startstate>1
3672 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3674 /* The Trie is constructed and compressed now so we can build a fail array if
3677 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3679 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3683 We find the fail state for each state in the trie, this state is the longest
3684 proper suffix of the current state's 'word' that is also a proper prefix of
3685 another word in our trie. State 1 represents the word '' and is thus the
3686 default fail state. This allows the DFA not to have to restart after its
3687 tried and failed a word at a given point, it simply continues as though it
3688 had been matching the other word in the first place.
3690 'abcdgu'=~/abcdefg|cdgu/
3691 When we get to 'd' we are still matching the first word, we would encounter
3692 'g' which would fail, which would bring us to the state representing 'd' in
3693 the second word where we would try 'g' and succeed, proceeding to match
3696 /* add a fail transition */
3697 const U32 trie_offset = ARG(source);
3698 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3700 const U32 ucharcount = trie->uniquecharcount;
3701 const U32 numstates = trie->statecount;
3702 const U32 ubound = trie->lasttrans + ucharcount;
3706 U32 base = trie->states[ 1 ].trans.base;
3709 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3711 GET_RE_DEBUG_FLAGS_DECL;
3713 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3714 PERL_UNUSED_CONTEXT;
3716 PERL_UNUSED_ARG(depth);
3719 if ( OP(source) == TRIE ) {
3720 struct regnode_1 *op = (struct regnode_1 *)
3721 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3722 StructCopy(source, op, struct regnode_1);
3723 stclass = (regnode *)op;
3725 struct regnode_charclass *op = (struct regnode_charclass *)
3726 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3727 StructCopy(source, op, struct regnode_charclass);
3728 stclass = (regnode *)op;
3730 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3732 ARG_SET( stclass, data_slot );
3733 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3734 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3735 aho->trie=trie_offset;
3736 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3737 Copy( trie->states, aho->states, numstates, reg_trie_state );
3738 Newx( q, numstates, U32);
3739 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3742 /* initialize fail[0..1] to be 1 so that we always have
3743 a valid final fail state */
3744 fail[ 0 ] = fail[ 1 ] = 1;
3746 for ( charid = 0; charid < ucharcount ; charid++ ) {
3747 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3749 q[ q_write ] = newstate;
3750 /* set to point at the root */
3751 fail[ q[ q_write++ ] ]=1;
3754 while ( q_read < q_write) {
3755 const U32 cur = q[ q_read++ % numstates ];
3756 base = trie->states[ cur ].trans.base;
3758 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3759 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3761 U32 fail_state = cur;
3764 fail_state = fail[ fail_state ];
3765 fail_base = aho->states[ fail_state ].trans.base;
3766 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3768 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3769 fail[ ch_state ] = fail_state;
3770 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3772 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3774 q[ q_write++ % numstates] = ch_state;
3778 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3779 when we fail in state 1, this allows us to use the
3780 charclass scan to find a valid start char. This is based on the principle
3781 that theres a good chance the string being searched contains lots of stuff
3782 that cant be a start char.
3784 fail[ 0 ] = fail[ 1 ] = 0;
3785 DEBUG_TRIE_COMPILE_r({
3786 Perl_re_indentf( aTHX_ "Stclass Failtable (%" UVuf " states): 0",
3787 depth, (UV)numstates
3789 for( q_read=1; q_read<numstates; q_read++ ) {
3790 Perl_re_printf( aTHX_ ", %" UVuf, (UV)fail[q_read]);
3792 Perl_re_printf( aTHX_ "\n");
3795 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3800 /* The below joins as many adjacent EXACTish nodes as possible into a single
3801 * one. The regop may be changed if the node(s) contain certain sequences that
3802 * require special handling. The joining is only done if:
3803 * 1) there is room in the current conglomerated node to entirely contain the
3805 * 2) they are compatible node types
3807 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3808 * these get optimized out
3810 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3811 * as possible, even if that means splitting an existing node so that its first
3812 * part is moved to the preceeding node. This would maximise the efficiency of
3813 * memEQ during matching.
3815 * If a node is to match under /i (folded), the number of characters it matches
3816 * can be different than its character length if it contains a multi-character
3817 * fold. *min_subtract is set to the total delta number of characters of the
3820 * And *unfolded_multi_char is set to indicate whether or not the node contains
3821 * an unfolded multi-char fold. This happens when it won't be known until
3822 * runtime whether the fold is valid or not; namely
3823 * 1) for EXACTF nodes that contain LATIN SMALL LETTER SHARP S, as only if the
3824 * target string being matched against turns out to be UTF-8 is that fold
3826 * 2) for EXACTFL nodes whose folding rules depend on the locale in force at
3828 * (Multi-char folds whose components are all above the Latin1 range are not
3829 * run-time locale dependent, and have already been folded by the time this
3830 * function is called.)
3832 * This is as good a place as any to discuss the design of handling these
3833 * multi-character fold sequences. It's been wrong in Perl for a very long
3834 * time. There are three code points in Unicode whose multi-character folds
3835 * were long ago discovered to mess things up. The previous designs for
3836 * dealing with these involved assigning a special node for them. This
3837 * approach doesn't always work, as evidenced by this example:
3838 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3839 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3840 * would match just the \xDF, it won't be able to handle the case where a
3841 * successful match would have to cross the node's boundary. The new approach
3842 * that hopefully generally solves the problem generates an EXACTFU_SS node
3843 * that is "sss" in this case.
3845 * It turns out that there are problems with all multi-character folds, and not
3846 * just these three. Now the code is general, for all such cases. The
3847 * approach taken is:
3848 * 1) This routine examines each EXACTFish node that could contain multi-
3849 * character folded sequences. Since a single character can fold into
3850 * such a sequence, the minimum match length for this node is less than
3851 * the number of characters in the node. This routine returns in
3852 * *min_subtract how many characters to subtract from the the actual
3853 * length of the string to get a real minimum match length; it is 0 if
3854 * there are no multi-char foldeds. This delta is used by the caller to
3855 * adjust the min length of the match, and the delta between min and max,
3856 * so that the optimizer doesn't reject these possibilities based on size
3858 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3859 * is used for an EXACTFU node that contains at least one "ss" sequence in
3860 * it. For non-UTF-8 patterns and strings, this is the only case where
3861 * there is a possible fold length change. That means that a regular
3862 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3863 * with length changes, and so can be processed faster. regexec.c takes
3864 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3865 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3866 * known until runtime). This saves effort in regex matching. However,
3867 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3868 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3869 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3870 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3871 * possibilities for the non-UTF8 patterns are quite simple, except for
3872 * the sharp s. All the ones that don't involve a UTF-8 target string are
3873 * members of a fold-pair, and arrays are set up for all of them so that
3874 * the other member of the pair can be found quickly. Code elsewhere in
3875 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3876 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3877 * described in the next item.
3878 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3879 * validity of the fold won't be known until runtime, and so must remain
3880 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFAA
3881 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3882 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3883 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3884 * The reason this is a problem is that the optimizer part of regexec.c
3885 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3886 * that a character in the pattern corresponds to at most a single
3887 * character in the target string. (And I do mean character, and not byte
3888 * here, unlike other parts of the documentation that have never been
3889 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3890 * EXACTFL nodes can match the two character string 'ss'; in EXACTFAA
3891 * nodes it can match "\x{17F}\x{17F}". These, along with other ones in
3892 * EXACTFL nodes, violate the assumption, and they are the only instances
3893 * where it is violated. I'm reluctant to try to change the assumption,
3894 * as the code involved is impenetrable to me (khw), so instead the code
3895 * here punts. This routine examines EXACTFL nodes, and (when the pattern
3896 * isn't UTF-8) EXACTF and EXACTFAA for such unfolded folds, and returns a
3897 * boolean indicating whether or not the node contains such a fold. When
3898 * it is true, the caller sets a flag that later causes the optimizer in
3899 * this file to not set values for the floating and fixed string lengths,
3900 * and thus avoids the optimizer code in regexec.c that makes the invalid
3901 * assumption. Thus, there is no optimization based on string lengths for
3902 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3903 * EXACTF and EXACTFAA nodes that contain the sharp s. (The reason the
3904 * assumption is wrong only in these cases is that all other non-UTF-8
3905 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3906 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3907 * EXACTF nodes because we don't know at compile time if it actually
3908 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3909 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3910 * always matches; and EXACTFAA where it never does. In an EXACTFAA node
3911 * in a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3912 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3913 * string would require the pattern to be forced into UTF-8, the overhead
3914 * of which we want to avoid. Similarly the unfolded multi-char folds in
3915 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3918 * Similarly, the code that generates tries doesn't currently handle
3919 * not-already-folded multi-char folds, and it looks like a pain to change
3920 * that. Therefore, trie generation of EXACTFAA nodes with the sharp s
3921 * doesn't work. Instead, such an EXACTFAA is turned into a new regnode,
3922 * EXACTFAA_NO_TRIE, which the trie code knows not to handle. Most people
3923 * using /iaa matching will be doing so almost entirely with ASCII
3924 * strings, so this should rarely be encountered in practice */
3926 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3927 if (PL_regkind[OP(scan)] == EXACT) \
3928 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags), NULL, depth+1)
3931 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3932 UV *min_subtract, bool *unfolded_multi_char,
3933 U32 flags, regnode *val, U32 depth)
3935 /* Merge several consecutive EXACTish nodes into one. */
3937 regnode *n = regnext(scan);
3939 regnode *next = scan + NODE_SZ_STR(scan);
3943 regnode *stop = scan;
3944 GET_RE_DEBUG_FLAGS_DECL;
3946 PERL_UNUSED_ARG(depth);
3949 PERL_ARGS_ASSERT_JOIN_EXACT;
3950 #ifndef EXPERIMENTAL_INPLACESCAN
3951 PERL_UNUSED_ARG(flags);
3952 PERL_UNUSED_ARG(val);
3954 DEBUG_PEEP("join", scan, depth, 0);
3956 assert(PL_regkind[OP(scan)] == EXACT);
3958 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3959 * EXACT ones that are mergeable to the current one. */
3961 && ( PL_regkind[OP(n)] == NOTHING
3962 || (stringok && PL_regkind[OP(n)] == EXACT))
3964 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3967 if (OP(n) == TAIL || n > next)
3969 if (PL_regkind[OP(n)] == NOTHING) {
3970 DEBUG_PEEP("skip:", n, depth, 0);
3971 NEXT_OFF(scan) += NEXT_OFF(n);
3972 next = n + NODE_STEP_REGNODE;
3979 else if (stringok) {
3980 const unsigned int oldl = STR_LEN(scan);
3981 regnode * const nnext = regnext(n);
3983 /* XXX I (khw) kind of doubt that this works on platforms (should
3984 * Perl ever run on one) where U8_MAX is above 255 because of lots
3985 * of other assumptions */
3986 /* Don't join if the sum can't fit into a single node */
3987 if (oldl + STR_LEN(n) > U8_MAX)
3990 /* Joining something that requires UTF-8 with something that
3991 * doesn't, means the result requires UTF-8. */
3992 if (OP(scan) == EXACT && (OP(n) == EXACT_ONLY8)) {
3993 OP(scan) = EXACT_ONLY8;
3995 else if (OP(scan) == EXACT_ONLY8 && (OP(n) == EXACT)) {
3996 ; /* join is compatible, no need to change OP */
https://perl5.git.perl.org / perl5.git / blob