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 /* Note on debug output:
76 * This is set up so that -Dr turns on debugging like all other flags that are
77 * enabled by -DDEBUGGING. -Drv gives more verbose output. This applies to
78 * all regular expressions encountered in a program, and gives a huge amount of
79 * output for all but the shortest programs.
81 * The ability to output pattern debugging information lexically, and with much
82 * finer grained control was added, with 'use re qw(Debug ....);' available even
83 * in non-DEBUGGING builds. This is accomplished by copying the contents of
84 * regcomp.c to ext/re/re_comp.c, and regexec.c is copied to ext/re/re_exec.c.
85 * Those files are compiled and linked into the perl executable, and they are
86 * compiled essentially as if DEBUGGING were enabled, and controlled by calls
89 * That would normally mean linking errors when two functions of the same name
90 * are attempted to be placed into the same executable. That is solved in one
92 * 1) Static functions aren't known outside the file they are in, so for the
93 * many functions of that type in this file, it just isn't a problem.
94 * 2) Most externally known functions are enclosed in
95 * #ifndef PERL_IN_XSUB_RE
98 * blocks, so there is only one defintion for them in the whole
99 * executable, the one in regcomp.c (or regexec.c). The implication of
100 * that is any debugging info that comes from them is controlled only by
101 * -Dr. Further, any static function they call will also be the version
102 * in regcomp.c (or regexec.c), so its debugging will also be by -Dr.
103 * 3) About a dozen external functions are re-#defined in ext/re/re_top.h, to
104 * have different names, so that what gets loaded in the executable is
105 * 'Perl_foo' from regcomp.c (and regexec.c), and the identical function
106 * from re_comp.c (and re_exec.c), but with the name 'my_foo' Debugging
107 * in the 'Perl_foo' versions is controlled by -Dr, but the 'my_foo'
108 * versions and their callees are under control of re.pm. The catch is
109 * that references to all these go through the regexp_engine structure,
110 * which is initialized in regcomp.h to the Perl_foo versions, and
111 * substituted out in lexical scopes where 'use re' is in effect to the
112 * 'my_foo' ones. That structure is public API, so it would be a hard
113 * sell to add any additional members.
114 * 4) For functions in regcomp.c and re_comp.c that are called only from,
115 * respectively, regexec.c and re_exec.c, they can have two different
116 * names, depending on #ifdef'ing PERL_IN_XSUB_RE, in both regexec.c and
119 * The bottom line is that if you add code to one of the public functions
120 * listed in ext/re/re_top.h, debugging automagically works. But if you write
121 * a new function that needs to do debugging or there is a chain of calls from
122 * it that need to do debugging, all functions in the chain should use options
125 * A function may have to be split so that debugging stuff is static, but it
126 * calls out to some other function that only gets compiled in regcomp.c to
127 * access data that we don't want to duplicate.
131 #define PERL_IN_REGCOMP_C
135 #ifdef PERL_IN_XSUB_RE
136 # include "re_comp.h"
137 EXTERN_C const struct regexp_engine my_reg_engine;
138 EXTERN_C const struct regexp_engine wild_reg_engine;
140 # include "regcomp.h"
143 #include "invlist_inline.h"
144 #include "unicode_constants.h"
147 #define STATIC static
150 /* this is a chain of data about sub patterns we are processing that
151 need to be handled separately/specially in study_chunk. Its so
152 we can simulate recursion without losing state. */
154 typedef struct scan_frame {
155 regnode *last_regnode; /* last node to process in this frame */
156 regnode *next_regnode; /* next node to process when last is reached */
157 U32 prev_recursed_depth;
158 I32 stopparen; /* what stopparen do we use */
159 bool in_gosub; /* this or an outer frame is for GOSUB */
161 struct scan_frame *this_prev_frame; /* this previous frame */
162 struct scan_frame *prev_frame; /* previous frame */
163 struct scan_frame *next_frame; /* next frame */
166 /* Certain characters are output as a sequence with the first being a
168 #define isBACKSLASHED_PUNCT(c) memCHRs("-[]\\^", c)
171 struct RExC_state_t {
172 U32 flags; /* RXf_* are we folding, multilining? */
173 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
174 char *precomp; /* uncompiled string. */
175 char *precomp_end; /* pointer to end of uncompiled string. */
176 REGEXP *rx_sv; /* The SV that is the regexp. */
177 regexp *rx; /* perl core regexp structure */
178 regexp_internal *rxi; /* internal data for regexp object
180 char *start; /* Start of input for compile */
181 char *end; /* End of input for compile */
182 char *parse; /* Input-scan pointer. */
183 char *copy_start; /* start of copy of input within
184 constructed parse string */
185 char *save_copy_start; /* Provides one level of saving
186 and restoring 'copy_start' */
187 char *copy_start_in_input; /* Position in input string
188 corresponding to copy_start */
189 SSize_t whilem_seen; /* number of WHILEM in this expr */
190 regnode *emit_start; /* Start of emitted-code area */
191 regnode_offset emit; /* Code-emit pointer */
192 I32 naughty; /* How bad is this pattern? */
193 I32 sawback; /* Did we see \1, ...? */
194 SSize_t size; /* Number of regnode equivalents in
196 Size_t sets_depth; /* Counts recursion depth of already-
197 compiled regex set patterns */
200 I32 parens_buf_size; /* #slots malloced open/close_parens */
201 regnode_offset *open_parens; /* offsets to open parens */
202 regnode_offset *close_parens; /* offsets to close parens */
203 HV *paren_names; /* Paren names */
205 /* position beyond 'precomp' of the warning message furthest away from
206 * 'precomp'. During the parse, no warnings are raised for any problems
207 * earlier in the parse than this position. This works if warnings are
208 * raised the first time a given spot is parsed, and if only one
209 * independent warning is raised for any given spot */
210 Size_t latest_warn_offset;
212 I32 npar; /* Capture buffer count so far in the
213 parse, (OPEN) plus one. ("par" 0 is
215 I32 total_par; /* During initial parse, is either 0,
216 or -1; the latter indicating a
217 reparse is needed. After that pass,
218 it is what 'npar' became after the
219 pass. Hence, it being > 0 indicates
220 we are in a reparse situation */
221 I32 nestroot; /* root parens we are in - used by
224 regnode *end_op; /* END node in program */
225 I32 utf8; /* whether the pattern is utf8 or not */
226 I32 orig_utf8; /* whether the pattern was originally in utf8 */
227 /* XXX use this for future optimisation of case
228 * where pattern must be upgraded to utf8. */
229 I32 uni_semantics; /* If a d charset modifier should use unicode
230 rules, even if the pattern is not in
233 I32 recurse_count; /* Number of recurse regops we have generated */
234 regnode **recurse; /* Recurse regops */
235 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
237 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
240 I32 override_recoding;
241 I32 recode_x_to_native;
242 I32 in_multi_char_class;
243 int code_index; /* next code_blocks[] slot */
244 struct reg_code_blocks *code_blocks;/* positions of literal (?{})
246 SSize_t maxlen; /* mininum possible number of chars in string to match */
247 scan_frame *frame_head;
248 scan_frame *frame_last;
252 SV *runtime_code_qr; /* qr with the runtime code blocks */
254 const char *lastparse;
256 U32 study_chunk_recursed_count;
257 AV *paren_name_list; /* idx -> name */
261 #define RExC_lastparse (pRExC_state->lastparse)
262 #define RExC_lastnum (pRExC_state->lastnum)
263 #define RExC_paren_name_list (pRExC_state->paren_name_list)
264 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
265 #define RExC_mysv (pRExC_state->mysv1)
266 #define RExC_mysv1 (pRExC_state->mysv1)
267 #define RExC_mysv2 (pRExC_state->mysv2)
275 bool sWARN_EXPERIMENTAL__VLB;
276 bool sWARN_EXPERIMENTAL__REGEX_SETS;
279 #define RExC_flags (pRExC_state->flags)
280 #define RExC_pm_flags (pRExC_state->pm_flags)
281 #define RExC_precomp (pRExC_state->precomp)
282 #define RExC_copy_start_in_input (pRExC_state->copy_start_in_input)
283 #define RExC_copy_start_in_constructed (pRExC_state->copy_start)
284 #define RExC_save_copy_start_in_constructed (pRExC_state->save_copy_start)
285 #define RExC_precomp_end (pRExC_state->precomp_end)
286 #define RExC_rx_sv (pRExC_state->rx_sv)
287 #define RExC_rx (pRExC_state->rx)
288 #define RExC_rxi (pRExC_state->rxi)
289 #define RExC_start (pRExC_state->start)
290 #define RExC_end (pRExC_state->end)
291 #define RExC_parse (pRExC_state->parse)
292 #define RExC_latest_warn_offset (pRExC_state->latest_warn_offset )
293 #define RExC_whilem_seen (pRExC_state->whilem_seen)
294 #define RExC_seen_d_op (pRExC_state->seen_d_op) /* Seen something that differs
295 under /d from /u ? */
297 #ifdef RE_TRACK_PATTERN_OFFSETS
298 # define RExC_offsets (RExC_rxi->u.offsets) /* I am not like the
301 #define RExC_emit (pRExC_state->emit)
302 #define RExC_emit_start (pRExC_state->emit_start)
303 #define RExC_sawback (pRExC_state->sawback)
304 #define RExC_seen (pRExC_state->seen)
305 #define RExC_size (pRExC_state->size)
306 #define RExC_maxlen (pRExC_state->maxlen)
307 #define RExC_npar (pRExC_state->npar)
308 #define RExC_total_parens (pRExC_state->total_par)
309 #define RExC_parens_buf_size (pRExC_state->parens_buf_size)
310 #define RExC_nestroot (pRExC_state->nestroot)
311 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
312 #define RExC_utf8 (pRExC_state->utf8)
313 #define RExC_uni_semantics (pRExC_state->uni_semantics)
314 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
315 #define RExC_open_parens (pRExC_state->open_parens)
316 #define RExC_close_parens (pRExC_state->close_parens)
317 #define RExC_end_op (pRExC_state->end_op)
318 #define RExC_paren_names (pRExC_state->paren_names)
319 #define RExC_recurse (pRExC_state->recurse)
320 #define RExC_recurse_count (pRExC_state->recurse_count)
321 #define RExC_sets_depth (pRExC_state->sets_depth)
322 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
323 #define RExC_study_chunk_recursed_bytes \
324 (pRExC_state->study_chunk_recursed_bytes)
325 #define RExC_in_lookaround (pRExC_state->in_lookaround)
326 #define RExC_contains_locale (pRExC_state->contains_locale)
327 #define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
330 # define SET_recode_x_to_native(x) \
331 STMT_START { RExC_recode_x_to_native = (x); } STMT_END
333 # define SET_recode_x_to_native(x) NOOP
336 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
337 #define RExC_frame_head (pRExC_state->frame_head)
338 #define RExC_frame_last (pRExC_state->frame_last)
339 #define RExC_frame_count (pRExC_state->frame_count)
340 #define RExC_strict (pRExC_state->strict)
341 #define RExC_study_started (pRExC_state->study_started)
342 #define RExC_warn_text (pRExC_state->warn_text)
343 #define RExC_in_script_run (pRExC_state->in_script_run)
344 #define RExC_use_BRANCHJ (pRExC_state->use_BRANCHJ)
345 #define RExC_warned_WARN_EXPERIMENTAL__VLB (pRExC_state->sWARN_EXPERIMENTAL__VLB)
346 #define RExC_warned_WARN_EXPERIMENTAL__REGEX_SETS (pRExC_state->sWARN_EXPERIMENTAL__REGEX_SETS)
347 #define RExC_unlexed_names (pRExC_state->unlexed_names)
349 /* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
350 * a flag to disable back-off on the fixed/floating substrings - if it's
351 * a high complexity pattern we assume the benefit of avoiding a full match
352 * is worth the cost of checking for the substrings even if they rarely help.
354 #define RExC_naughty (pRExC_state->naughty)
355 #define TOO_NAUGHTY (10)
356 #define MARK_NAUGHTY(add) \
357 if (RExC_naughty < TOO_NAUGHTY) \
358 RExC_naughty += (add)
359 #define MARK_NAUGHTY_EXP(exp, add) \
360 if (RExC_naughty < TOO_NAUGHTY) \
361 RExC_naughty += RExC_naughty / (exp) + (add)
363 #define isNON_BRACE_QUANTIFIER(c) ((c) == '*' || (c) == '+' || (c) == '?')
364 #define isQUANTIFIER(s,e) ( isNON_BRACE_QUANTIFIER(*s) \
365 || ((*s) == '{' && regcurly(s, e, NULL)))
368 * Flags to be passed up and down.
370 #define HASWIDTH 0x01 /* Known to not match null strings, could match
372 #define SIMPLE 0x02 /* Exactly one character wide */
373 /* (or LNBREAK as a special case) */
374 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
375 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
376 #define RESTART_PARSE 0x20 /* Need to redo the parse */
377 #define NEED_UTF8 0x40 /* In conjunction with RESTART_PARSE, need to
378 calcuate sizes as UTF-8 */
380 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
382 /* whether trie related optimizations are enabled */
383 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
384 #define TRIE_STUDY_OPT
385 #define FULL_TRIE_STUDY
391 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
392 #define PBITVAL(paren) (1 << ((paren) & 7))
393 #define PAREN_OFFSET(depth) \
394 (RExC_study_chunk_recursed + (depth) * RExC_study_chunk_recursed_bytes)
395 #define PAREN_TEST(depth, paren) \
396 (PBYTE(PAREN_OFFSET(depth), paren) & PBITVAL(paren))
397 #define PAREN_SET(depth, paren) \
398 (PBYTE(PAREN_OFFSET(depth), paren) |= PBITVAL(paren))
399 #define PAREN_UNSET(depth, paren) \
400 (PBYTE(PAREN_OFFSET(depth), paren) &= ~PBITVAL(paren))
402 #define REQUIRE_UTF8(flagp) STMT_START { \
404 *flagp = RESTART_PARSE|NEED_UTF8; \
409 /* /u is to be chosen if we are supposed to use Unicode rules, or if the
410 * pattern is in UTF-8. This latter condition is in case the outermost rules
411 * are locale. See GH #17278 */
412 #define toUSE_UNI_CHARSET_NOT_DEPENDS (RExC_uni_semantics || UTF)
414 /* Change from /d into /u rules, and restart the parse. RExC_uni_semantics is
415 * a flag that indicates we need to override /d with /u as a result of
416 * something in the pattern. It should only be used in regards to calling
417 * set_regex_charset() or get_regex_charset() */
418 #define REQUIRE_UNI_RULES(flagp, restart_retval) \
420 if (DEPENDS_SEMANTICS) { \
421 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
422 RExC_uni_semantics = 1; \
423 if (RExC_seen_d_op && LIKELY(! IN_PARENS_PASS)) { \
424 /* No need to restart the parse if we haven't seen \
425 * anything that differs between /u and /d, and no need \
426 * to restart immediately if we're going to reparse \
427 * anyway to count parens */ \
428 *flagp |= RESTART_PARSE; \
429 return restart_retval; \
434 #define REQUIRE_BRANCHJ(flagp, restart_retval) \
436 RExC_use_BRANCHJ = 1; \
437 *flagp |= RESTART_PARSE; \
438 return restart_retval; \
441 /* Until we have completed the parse, we leave RExC_total_parens at 0 or
442 * less. After that, it must always be positive, because the whole re is
443 * considered to be surrounded by virtual parens. Setting it to negative
444 * indicates there is some construct that needs to know the actual number of
445 * parens to be properly handled. And that means an extra pass will be
446 * required after we've counted them all */
447 #define ALL_PARENS_COUNTED (RExC_total_parens > 0)
448 #define REQUIRE_PARENS_PASS \
449 STMT_START { /* No-op if have completed a pass */ \
450 if (! ALL_PARENS_COUNTED) RExC_total_parens = -1; \
452 #define IN_PARENS_PASS (RExC_total_parens < 0)
455 /* This is used to return failure (zero) early from the calling function if
456 * various flags in 'flags' are set. Two flags always cause a return:
457 * 'RESTART_PARSE' and 'NEED_UTF8'. 'extra' can be used to specify any
458 * additional flags that should cause a return; 0 if none. If the return will
459 * be done, '*flagp' is first set to be all of the flags that caused the
461 #define RETURN_FAIL_ON_RESTART_OR_FLAGS(flags,flagp,extra) \
463 if ((flags) & (RESTART_PARSE|NEED_UTF8|(extra))) { \
464 *(flagp) = (flags) & (RESTART_PARSE|NEED_UTF8|(extra)); \
469 #define MUST_RESTART(flags) ((flags) & (RESTART_PARSE))
471 #define RETURN_FAIL_ON_RESTART(flags,flagp) \
472 RETURN_FAIL_ON_RESTART_OR_FLAGS( flags, flagp, 0)
473 #define RETURN_FAIL_ON_RESTART_FLAGP(flagp) \
474 if (MUST_RESTART(*(flagp))) return 0
476 /* This converts the named class defined in regcomp.h to its equivalent class
477 * number defined in handy.h. */
478 #define namedclass_to_classnum(class) ((int) ((class) / 2))
479 #define classnum_to_namedclass(classnum) ((classnum) * 2)
481 #define _invlist_union_complement_2nd(a, b, output) \
482 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
483 #define _invlist_intersection_complement_2nd(a, b, output) \
484 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
486 /* We add a marker if we are deferring expansion of a property that is both
487 * 1) potentiallly user-defined; and
488 * 2) could also be an official Unicode property.
490 * Without this marker, any deferred expansion can only be for a user-defined
491 * one. This marker shouldn't conflict with any that could be in a legal name,
492 * and is appended to its name to indicate this. There is a string and
494 #define DEFERRED_COULD_BE_OFFICIAL_MARKERs "~"
495 #define DEFERRED_COULD_BE_OFFICIAL_MARKERc '~'
497 /* What is infinity for optimization purposes */
498 #define OPTIMIZE_INFTY SSize_t_MAX
500 /* About scan_data_t.
502 During optimisation we recurse through the regexp program performing
503 various inplace (keyhole style) optimisations. In addition study_chunk
504 and scan_commit populate this data structure with information about
505 what strings MUST appear in the pattern. We look for the longest
506 string that must appear at a fixed location, and we look for the
507 longest string that may appear at a floating location. So for instance
512 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
513 strings (because they follow a .* construct). study_chunk will identify
514 both FOO and BAR as being the longest fixed and floating strings respectively.
516 The strings can be composites, for instance
520 will result in a composite fixed substring 'foo'.
522 For each string some basic information is maintained:
525 This is the position the string must appear at, or not before.
526 It also implicitly (when combined with minlenp) tells us how many
527 characters must match before the string we are searching for.
528 Likewise when combined with minlenp and the length of the string it
529 tells us how many characters must appear after the string we have
533 Only used for floating strings. This is the rightmost point that
534 the string can appear at. If set to OPTIMIZE_INFTY it indicates that the
535 string can occur infinitely far to the right.
536 For fixed strings, it is equal to min_offset.
539 A pointer to the minimum number of characters of the pattern that the
540 string was found inside. This is important as in the case of positive
541 lookahead or positive lookbehind we can have multiple patterns
546 The minimum length of the pattern overall is 3, the minimum length
547 of the lookahead part is 3, but the minimum length of the part that
548 will actually match is 1. So 'FOO's minimum length is 3, but the
549 minimum length for the F is 1. This is important as the minimum length
550 is used to determine offsets in front of and behind the string being
551 looked for. Since strings can be composites this is the length of the
552 pattern at the time it was committed with a scan_commit. Note that
553 the length is calculated by study_chunk, so that the minimum lengths
554 are not known until the full pattern has been compiled, thus the
555 pointer to the value.
559 In the case of lookbehind the string being searched for can be
560 offset past the start point of the final matching string.
561 If this value was just blithely removed from the min_offset it would
562 invalidate some of the calculations for how many chars must match
563 before or after (as they are derived from min_offset and minlen and
564 the length of the string being searched for).
565 When the final pattern is compiled and the data is moved from the
566 scan_data_t structure into the regexp structure the information
567 about lookbehind is factored in, with the information that would
568 have been lost precalculated in the end_shift field for the
571 The fields pos_min and pos_delta are used to store the minimum offset
572 and the delta to the maximum offset at the current point in the pattern.
576 struct scan_data_substrs {
577 SV *str; /* longest substring found in pattern */
578 SSize_t min_offset; /* earliest point in string it can appear */
579 SSize_t max_offset; /* latest point in string it can appear */
580 SSize_t *minlenp; /* pointer to the minlen relevant to the string */
581 SSize_t lookbehind; /* is the pos of the string modified by LB */
582 I32 flags; /* per substring SF_* and SCF_* flags */
585 typedef struct scan_data_t {
586 /*I32 len_min; unused */
587 /*I32 len_delta; unused */
591 SSize_t last_end; /* min value, <0 unless valid. */
592 SSize_t last_start_min;
593 SSize_t last_start_max;
594 U8 cur_is_floating; /* whether the last_* values should be set as
595 * the next fixed (0) or floating (1)
598 /* [0] is longest fixed substring so far, [1] is longest float so far */
599 struct scan_data_substrs substrs[2];
601 I32 flags; /* common SF_* and SCF_* flags */
603 SSize_t *last_closep;
604 regnode_ssc *start_class;
608 * Forward declarations for pregcomp()'s friends.
611 static const scan_data_t zero_scan_data = {
612 0, 0, NULL, 0, 0, 0, 0,
614 { NULL, 0, 0, 0, 0, 0 },
615 { NULL, 0, 0, 0, 0, 0 },
622 #define SF_BEFORE_SEOL 0x0001
623 #define SF_BEFORE_MEOL 0x0002
624 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
626 #define SF_IS_INF 0x0040
627 #define SF_HAS_PAR 0x0080
628 #define SF_IN_PAR 0x0100
629 #define SF_HAS_EVAL 0x0200
632 /* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
633 * longest substring in the pattern. When it is not set the optimiser keeps
634 * track of position, but does not keep track of the actual strings seen,
636 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
639 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
640 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
641 * turned off because of the alternation (BRANCH). */
642 #define SCF_DO_SUBSTR 0x0400
644 #define SCF_DO_STCLASS_AND 0x0800
645 #define SCF_DO_STCLASS_OR 0x1000
646 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
647 #define SCF_WHILEM_VISITED_POS 0x2000
649 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
650 #define SCF_SEEN_ACCEPT 0x8000
651 #define SCF_TRIE_DOING_RESTUDY 0x10000
652 #define SCF_IN_DEFINE 0x20000
657 #define UTF cBOOL(RExC_utf8)
659 /* The enums for all these are ordered so things work out correctly */
660 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
661 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
662 == REGEX_DEPENDS_CHARSET)
663 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
664 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
665 >= REGEX_UNICODE_CHARSET)
666 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
667 == REGEX_ASCII_RESTRICTED_CHARSET)
668 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
669 >= REGEX_ASCII_RESTRICTED_CHARSET)
670 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
671 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
673 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
675 /* For programs that want to be strictly Unicode compatible by dying if any
676 * attempt is made to match a non-Unicode code point against a Unicode
678 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
680 #define OOB_NAMEDCLASS -1
682 /* There is no code point that is out-of-bounds, so this is problematic. But
683 * its only current use is to initialize a variable that is always set before
685 #define OOB_UNICODE 0xDEADBEEF
687 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
690 /* length of regex to show in messages that don't mark a position within */
691 #define RegexLengthToShowInErrorMessages 127
694 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
695 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
696 * op/pragma/warn/regcomp.
698 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
699 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
701 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
702 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
704 /* The code in this file in places uses one level of recursion with parsing
705 * rebased to an alternate string constructed by us in memory. This can take
706 * the form of something that is completely different from the input, or
707 * something that uses the input as part of the alternate. In the first case,
708 * there should be no possibility of an error, as we are in complete control of
709 * the alternate string. But in the second case we don't completely control
710 * the input portion, so there may be errors in that. Here's an example:
712 * is handled specially because \x{df} folds to a sequence of more than one
713 * character: 'ss'. What is done is to create and parse an alternate string,
714 * which looks like this:
715 * /(?:\x{DF}|[abc\x{DF}def])/ui
716 * where it uses the input unchanged in the middle of something it constructs,
717 * which is a branch for the DF outside the character class, and clustering
718 * parens around the whole thing. (It knows enough to skip the DF inside the
719 * class while in this substitute parse.) 'abc' and 'def' may have errors that
720 * need to be reported. The general situation looks like this:
722 * |<------- identical ------>|
724 * Input: ---------------------------------------------------------------
725 * Constructed: ---------------------------------------------------
727 * |<------- identical ------>|
729 * sI..eI is the portion of the input pattern we are concerned with here.
730 * sC..EC is the constructed substitute parse string.
731 * sC..tC is constructed by us
732 * tC..eC is an exact duplicate of the portion of the input pattern tI..eI.
733 * In the diagram, these are vertically aligned.
734 * eC..EC is also constructed by us.
735 * xC is the position in the substitute parse string where we found a
737 * xI is the position in the original pattern corresponding to xC.
739 * We want to display a message showing the real input string. Thus we need to
740 * translate from xC to xI. We know that xC >= tC, since the portion of the
741 * string sC..tC has been constructed by us, and so shouldn't have errors. We
743 * xI = tI + (xC - tC)
745 * When the substitute parse is constructed, the code needs to set:
748 * RExC_copy_start_in_input (tI)
749 * RExC_copy_start_in_constructed (tC)
750 * and restore them when done.
752 * During normal processing of the input pattern, both
753 * 'RExC_copy_start_in_input' and 'RExC_copy_start_in_constructed' are set to
754 * sI, so that xC equals xI.
757 #define sI RExC_precomp
758 #define eI RExC_precomp_end
759 #define sC RExC_start
761 #define tI RExC_copy_start_in_input
762 #define tC RExC_copy_start_in_constructed
763 #define xI(xC) (tI + (xC - tC))
764 #define xI_offset(xC) (xI(xC) - sI)
766 #define REPORT_LOCATION_ARGS(xC) \
768 (xI(xC) > eI) /* Don't run off end */ \
769 ? eI - sI /* Length before the <--HERE */ \
770 : ((xI_offset(xC) >= 0) \
772 : (Perl_croak(aTHX_ "panic: %s: %d: negative offset: %" \
773 IVdf " trying to output message for " \
775 __FILE__, __LINE__, (IV) xI_offset(xC), \
776 ((int) (eC - sC)), sC), 0)), \
777 sI), /* The input pattern printed up to the <--HERE */ \
779 (xI(xC) > eI) ? 0 : eI - xI(xC), /* Length after <--HERE */ \
780 (xI(xC) > eI) ? eI : xI(xC)) /* pattern after <--HERE */
782 /* Used to point after bad bytes for an error message, but avoid skipping
783 * past a nul byte. */
784 #define SKIP_IF_CHAR(s, e) (!*(s) ? 0 : UTF ? UTF8_SAFE_SKIP(s, e) : 1)
786 /* Set up to clean up after our imminent demise */
787 #define PREPARE_TO_DIE \
790 SAVEFREESV(RExC_rx_sv); \
791 if (RExC_open_parens) \
792 SAVEFREEPV(RExC_open_parens); \
793 if (RExC_close_parens) \
794 SAVEFREEPV(RExC_close_parens); \
798 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
799 * arg. Show regex, up to a maximum length. If it's too long, chop and add
802 #define _FAIL(code) STMT_START { \
803 const char *ellipses = ""; \
804 IV len = RExC_precomp_end - RExC_precomp; \
807 if (len > RegexLengthToShowInErrorMessages) { \
808 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
809 len = RegexLengthToShowInErrorMessages - 10; \
815 #define FAIL(msg) _FAIL( \
816 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
817 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
819 #define FAIL2(msg,arg) _FAIL( \
820 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
821 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
823 #define FAIL3(msg,arg1,arg2) _FAIL( \
824 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
825 arg1, arg2, UTF8fARG(UTF, len, RExC_precomp), ellipses))
828 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
830 #define Simple_vFAIL(m) STMT_START { \
831 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
832 m, REPORT_LOCATION_ARGS(RExC_parse)); \
836 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
838 #define vFAIL(m) STMT_START { \
844 * Like Simple_vFAIL(), but accepts two arguments.
846 #define Simple_vFAIL2(m,a1) STMT_START { \
847 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, \
848 REPORT_LOCATION_ARGS(RExC_parse)); \
852 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
854 #define vFAIL2(m,a1) STMT_START { \
856 Simple_vFAIL2(m, a1); \
861 * Like Simple_vFAIL(), but accepts three arguments.
863 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
864 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, a2, \
865 REPORT_LOCATION_ARGS(RExC_parse)); \
869 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
871 #define vFAIL3(m,a1,a2) STMT_START { \
873 Simple_vFAIL3(m, a1, a2); \
877 * Like Simple_vFAIL(), but accepts four arguments.
879 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
880 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, a2, a3, \
881 REPORT_LOCATION_ARGS(RExC_parse)); \
884 #define vFAIL4(m,a1,a2,a3) STMT_START { \
886 Simple_vFAIL4(m, a1, a2, a3); \
889 /* A specialized version of vFAIL2 that works with UTF8f */
890 #define vFAIL2utf8f(m, a1) STMT_START { \
892 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, \
893 REPORT_LOCATION_ARGS(RExC_parse)); \
896 #define vFAIL3utf8f(m, a1, a2) STMT_START { \
898 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, a2, \
899 REPORT_LOCATION_ARGS(RExC_parse)); \
902 /* Setting this to NULL is a signal to not output warnings */
903 #define TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE \
905 RExC_save_copy_start_in_constructed = RExC_copy_start_in_constructed;\
906 RExC_copy_start_in_constructed = NULL; \
908 #define RESTORE_WARNINGS \
909 RExC_copy_start_in_constructed = RExC_save_copy_start_in_constructed
911 /* Since a warning can be generated multiple times as the input is reparsed, we
912 * output it the first time we come to that point in the parse, but suppress it
913 * otherwise. 'RExC_copy_start_in_constructed' being NULL is a flag to not
914 * generate any warnings */
915 #define TO_OUTPUT_WARNINGS(loc) \
916 ( RExC_copy_start_in_constructed \
917 && ((xI(loc)) - RExC_precomp) > (Ptrdiff_t) RExC_latest_warn_offset)
919 /* After we've emitted a warning, we save the position in the input so we don't
921 #define UPDATE_WARNINGS_LOC(loc) \
923 if (TO_OUTPUT_WARNINGS(loc)) { \
924 RExC_latest_warn_offset = MAX(sI, MIN(eI, xI(loc))) \
929 /* 'warns' is the output of the packWARNx macro used in 'code' */
930 #define _WARN_HELPER(loc, warns, code) \
932 if (! RExC_copy_start_in_constructed) { \
933 Perl_croak( aTHX_ "panic! %s: %d: Tried to warn when none" \
934 " expected at '%s'", \
935 __FILE__, __LINE__, loc); \
937 if (TO_OUTPUT_WARNINGS(loc)) { \
941 UPDATE_WARNINGS_LOC(loc); \
945 /* m is not necessarily a "literal string", in this macro */
946 #define warn_non_literal_string(loc, packed_warn, m) \
947 _WARN_HELPER(loc, packed_warn, \
948 Perl_warner(aTHX_ packed_warn, \
949 "%s" REPORT_LOCATION, \
950 m, REPORT_LOCATION_ARGS(loc)))
951 #define reg_warn_non_literal_string(loc, m) \
952 warn_non_literal_string(loc, packWARN(WARN_REGEXP), m)
954 #define ckWARN2_non_literal_string(loc, packwarn, m, a1) \
957 Size_t format_size = strlen(m) + strlen(REPORT_LOCATION)+ 1;\
958 Newx(format, format_size, char); \
959 my_strlcpy(format, m, format_size); \
960 my_strlcat(format, REPORT_LOCATION, format_size); \
961 SAVEFREEPV(format); \
962 _WARN_HELPER(loc, packwarn, \
963 Perl_ck_warner(aTHX_ packwarn, \
965 a1, REPORT_LOCATION_ARGS(loc))); \
968 #define ckWARNreg(loc,m) \
969 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
970 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
972 REPORT_LOCATION_ARGS(loc)))
974 #define vWARN(loc, m) \
975 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
976 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
978 REPORT_LOCATION_ARGS(loc))) \
980 #define vWARN_dep(loc, m) \
981 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
982 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
984 REPORT_LOCATION_ARGS(loc)))
986 #define ckWARNdep(loc,m) \
987 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
988 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
990 REPORT_LOCATION_ARGS(loc)))
992 #define ckWARNregdep(loc,m) \
993 _WARN_HELPER(loc, packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
994 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
997 REPORT_LOCATION_ARGS(loc)))
999 #define ckWARN2reg_d(loc,m, a1) \
1000 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1001 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
1002 m REPORT_LOCATION, \
1003 a1, REPORT_LOCATION_ARGS(loc)))
1005 #define ckWARN2reg(loc, m, a1) \
1006 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1007 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
1008 m REPORT_LOCATION, \
1009 a1, REPORT_LOCATION_ARGS(loc)))
1011 #define vWARN3(loc, m, a1, a2) \
1012 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1013 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
1014 m REPORT_LOCATION, \
1015 a1, a2, REPORT_LOCATION_ARGS(loc)))
1017 #define ckWARN3reg(loc, m, a1, a2) \
1018 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1019 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
1020 m REPORT_LOCATION, \
1022 REPORT_LOCATION_ARGS(loc)))
1024 #define vWARN4(loc, m, a1, a2, a3) \
1025 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1026 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
1027 m REPORT_LOCATION, \
1029 REPORT_LOCATION_ARGS(loc)))
1031 #define ckWARN4reg(loc, m, a1, a2, a3) \
1032 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1033 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
1034 m REPORT_LOCATION, \
1036 REPORT_LOCATION_ARGS(loc)))
1038 #define vWARN5(loc, m, a1, a2, a3, a4) \
1039 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1040 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
1041 m REPORT_LOCATION, \
1043 REPORT_LOCATION_ARGS(loc)))
1045 #define ckWARNexperimental(loc, class, m) \
1047 if (! RExC_warned_ ## class) { /* warn once per compilation */ \
1048 RExC_warned_ ## class = 1; \
1049 _WARN_HELPER(loc, packWARN(class), \
1050 Perl_ck_warner_d(aTHX_ packWARN(class), \
1051 m REPORT_LOCATION, \
1052 REPORT_LOCATION_ARGS(loc)));\
1056 /* Convert between a pointer to a node and its offset from the beginning of the
1058 #define REGNODE_p(offset) (RExC_emit_start + (offset))
1059 #define REGNODE_OFFSET(node) ((node) - RExC_emit_start)
1061 /* Macros for recording node offsets. 20001227 mjd@plover.com
1062 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
1063 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
1064 * Element 0 holds the number n.
1065 * Position is 1 indexed.
1067 #ifndef RE_TRACK_PATTERN_OFFSETS
1068 #define Set_Node_Offset_To_R(offset,byte)
1069 #define Set_Node_Offset(node,byte)
1070 #define Set_Cur_Node_Offset
1071 #define Set_Node_Length_To_R(node,len)
1072 #define Set_Node_Length(node,len)
1073 #define Set_Node_Cur_Length(node,start)
1074 #define Node_Offset(n)
1075 #define Node_Length(n)
1076 #define Set_Node_Offset_Length(node,offset,len)
1077 #define ProgLen(ri) ri->u.proglen
1078 #define SetProgLen(ri,x) ri->u.proglen = x
1079 #define Track_Code(code)
1081 #define ProgLen(ri) ri->u.offsets[0]
1082 #define SetProgLen(ri,x) ri->u.offsets[0] = x
1083 #define Set_Node_Offset_To_R(offset,byte) STMT_START { \
1084 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
1085 __LINE__, (int)(offset), (int)(byte))); \
1086 if((offset) < 0) { \
1087 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
1090 RExC_offsets[2*(offset)-1] = (byte); \
1094 #define Set_Node_Offset(node,byte) \
1095 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (byte)-RExC_start)
1096 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
1098 #define Set_Node_Length_To_R(node,len) STMT_START { \
1099 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
1100 __LINE__, (int)(node), (int)(len))); \
1102 Perl_croak(aTHX_ "value of node is %d in Length macro", \
1105 RExC_offsets[2*(node)] = (len); \
1109 #define Set_Node_Length(node,len) \
1110 Set_Node_Length_To_R(REGNODE_OFFSET(node), len)
1111 #define Set_Node_Cur_Length(node, start) \
1112 Set_Node_Length(node, RExC_parse - start)
1114 /* Get offsets and lengths */
1115 #define Node_Offset(n) (RExC_offsets[2*(REGNODE_OFFSET(n))-1])
1116 #define Node_Length(n) (RExC_offsets[2*(REGNODE_OFFSET(n))])
1118 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
1119 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (offset)); \
1120 Set_Node_Length_To_R(REGNODE_OFFSET(node), (len)); \
1123 #define Track_Code(code) STMT_START { code } STMT_END
1126 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
1127 #define EXPERIMENTAL_INPLACESCAN
1128 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
1132 Perl_re_printf(pTHX_ const char *fmt, ...)
1136 PerlIO *f= Perl_debug_log;
1137 PERL_ARGS_ASSERT_RE_PRINTF;
1139 result = PerlIO_vprintf(f, fmt, ap);
1145 Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
1149 PerlIO *f= Perl_debug_log;
1150 PERL_ARGS_ASSERT_RE_INDENTF;
1151 va_start(ap, depth);
1152 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
1153 result = PerlIO_vprintf(f, fmt, ap);
1157 #endif /* DEBUGGING */
1159 #define DEBUG_RExC_seen() \
1160 DEBUG_OPTIMISE_MORE_r({ \
1161 Perl_re_printf( aTHX_ "RExC_seen: "); \
1163 if (RExC_seen & REG_ZERO_LEN_SEEN) \
1164 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
1166 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
1167 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
1169 if (RExC_seen & REG_GPOS_SEEN) \
1170 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
1172 if (RExC_seen & REG_RECURSE_SEEN) \
1173 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
1175 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
1176 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
1178 if (RExC_seen & REG_VERBARG_SEEN) \
1179 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
1181 if (RExC_seen & REG_CUTGROUP_SEEN) \
1182 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
1184 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
1185 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
1187 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
1188 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
1190 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
1191 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
1193 Perl_re_printf( aTHX_ "\n"); \
1196 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
1197 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
1202 S_debug_show_study_flags(pTHX_ U32 flags, const char *open_str,
1203 const char *close_str)
1208 Perl_re_printf( aTHX_ "%s", open_str);
1209 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_SEOL);
1210 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_MEOL);
1211 DEBUG_SHOW_STUDY_FLAG(flags, SF_IS_INF);
1212 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_PAR);
1213 DEBUG_SHOW_STUDY_FLAG(flags, SF_IN_PAR);
1214 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_EVAL);
1215 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_SUBSTR);
1216 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_AND);
1217 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_OR);
1218 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS);
1219 DEBUG_SHOW_STUDY_FLAG(flags, SCF_WHILEM_VISITED_POS);
1220 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_RESTUDY);
1221 DEBUG_SHOW_STUDY_FLAG(flags, SCF_SEEN_ACCEPT);
1222 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_DOING_RESTUDY);
1223 DEBUG_SHOW_STUDY_FLAG(flags, SCF_IN_DEFINE);
1224 Perl_re_printf( aTHX_ "%s", close_str);
1229 S_debug_studydata(pTHX_ const char *where, scan_data_t *data,
1230 U32 depth, int is_inf)
1232 DECLARE_AND_GET_RE_DEBUG_FLAGS;
1234 DEBUG_OPTIMISE_MORE_r({
1237 Perl_re_indentf(aTHX_ "%s: Pos:%" IVdf "/%" IVdf " Flags: 0x%" UVXf,
1241 (IV)data->pos_delta,
1245 S_debug_show_study_flags(aTHX_ data->flags," [","]");
1247 Perl_re_printf( aTHX_
1248 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s",
1250 (IV)(data->last_closep ? *((data)->last_closep) : -1),
1251 is_inf ? "INF " : ""
1254 if (data->last_found) {
1256 Perl_re_printf(aTHX_
1257 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf,
1258 SvPVX_const(data->last_found),
1260 (IV)data->last_start_min,
1261 (IV)data->last_start_max
1264 for (i = 0; i < 2; i++) {
1265 Perl_re_printf(aTHX_
1266 " %s%s: '%s' @ %" IVdf "/%" IVdf,
1267 data->cur_is_floating == i ? "*" : "",
1268 i ? "Float" : "Fixed",
1269 SvPVX_const(data->substrs[i].str),
1270 (IV)data->substrs[i].min_offset,
1271 (IV)data->substrs[i].max_offset
1273 S_debug_show_study_flags(aTHX_ data->substrs[i].flags," [","]");
1277 Perl_re_printf( aTHX_ "\n");
1283 S_debug_peep(pTHX_ const char *str, const RExC_state_t *pRExC_state,
1284 regnode *scan, U32 depth, U32 flags)
1286 DECLARE_AND_GET_RE_DEBUG_FLAGS;
1293 Next = regnext(scan);
1294 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
1295 Perl_re_indentf( aTHX_ "%s>%3d: %s (%d)",
1298 REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),
1299 Next ? (REG_NODE_NUM(Next)) : 0 );
1300 S_debug_show_study_flags(aTHX_ flags," [ ","]");
1301 Perl_re_printf( aTHX_ "\n");
1306 # define DEBUG_STUDYDATA(where, data, depth, is_inf) \
1307 S_debug_studydata(aTHX_ where, data, depth, is_inf)
1309 # define DEBUG_PEEP(str, scan, depth, flags) \
1310 S_debug_peep(aTHX_ str, pRExC_state, scan, depth, flags)
1313 # define DEBUG_STUDYDATA(where, data, depth, is_inf) NOOP
1314 # define DEBUG_PEEP(str, scan, depth, flags) NOOP
1318 /* =========================================================
1319 * BEGIN edit_distance stuff.
1321 * This calculates how many single character changes of any type are needed to
1322 * transform a string into another one. It is taken from version 3.1 of
1324 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1327 /* Our unsorted dictionary linked list. */
1328 /* Note we use UVs, not chars. */
1333 struct dictionary* next;
1335 typedef struct dictionary item;
1338 PERL_STATIC_INLINE item*
1339 push(UV key, item* curr)
1342 Newx(head, 1, item);
1350 PERL_STATIC_INLINE item*
1351 find(item* head, UV key)
1353 item* iterator = head;
1355 if (iterator->key == key){
1358 iterator = iterator->next;
1364 PERL_STATIC_INLINE item*
1365 uniquePush(item* head, UV key)
1367 item* iterator = head;
1370 if (iterator->key == key) {
1373 iterator = iterator->next;
1376 return push(key, head);
1379 PERL_STATIC_INLINE void
1380 dict_free(item* head)
1382 item* iterator = head;
1385 item* temp = iterator;
1386 iterator = iterator->next;
1393 /* End of Dictionary Stuff */
1395 /* All calculations/work are done here */
1397 S_edit_distance(const UV* src,
1399 const STRLEN x, /* length of src[] */
1400 const STRLEN y, /* length of tgt[] */
1401 const SSize_t maxDistance
1405 UV swapCount, swapScore, targetCharCount, i, j;
1407 UV score_ceil = x + y;
1409 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1411 /* intialize matrix start values */
1412 Newx(scores, ( (x + 2) * (y + 2)), UV);
1413 scores[0] = score_ceil;
1414 scores[1 * (y + 2) + 0] = score_ceil;
1415 scores[0 * (y + 2) + 1] = score_ceil;
1416 scores[1 * (y + 2) + 1] = 0;
1417 head = uniquePush(uniquePush(head, src[0]), tgt[0]);
1422 for (i=1;i<=x;i++) {
1424 head = uniquePush(head, src[i]);
1425 scores[(i+1) * (y + 2) + 1] = i;
1426 scores[(i+1) * (y + 2) + 0] = score_ceil;
1429 for (j=1;j<=y;j++) {
1432 head = uniquePush(head, tgt[j]);
1433 scores[1 * (y + 2) + (j + 1)] = j;
1434 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1437 targetCharCount = find(head, tgt[j-1])->value;
1438 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1440 if (src[i-1] != tgt[j-1]){
1441 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));
1445 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1449 find(head, src[i-1])->value = i;
1453 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1456 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1460 /* END of edit_distance() stuff
1461 * ========================================================= */
1463 /* Mark that we cannot extend a found fixed substring at this point.
1464 Update the longest found anchored substring or the longest found
1465 floating substrings if needed. */
1468 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1469 SSize_t *minlenp, int is_inf)
1471 const STRLEN l = CHR_SVLEN(data->last_found);
1472 SV * const longest_sv = data->substrs[data->cur_is_floating].str;
1473 const STRLEN old_l = CHR_SVLEN(longest_sv);
1474 DECLARE_AND_GET_RE_DEBUG_FLAGS;
1476 PERL_ARGS_ASSERT_SCAN_COMMIT;
1478 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
1479 const U8 i = data->cur_is_floating;
1480 SvSetMagicSV(longest_sv, data->last_found);
1481 data->substrs[i].min_offset = l ? data->last_start_min : data->pos_min;
1484 data->substrs[0].max_offset = data->substrs[0].min_offset;
1486 data->substrs[1].max_offset =
1490 ? data->last_start_max
1491 /* temporary underflow guard for 5.32 */
1492 : data->pos_delta < 0 ? OPTIMIZE_INFTY
1493 : (data->pos_delta > OPTIMIZE_INFTY - data->pos_min
1495 : data->pos_min + data->pos_delta));
1498 data->substrs[i].flags &= ~SF_BEFORE_EOL;
1499 data->substrs[i].flags |= data->flags & SF_BEFORE_EOL;
1500 data->substrs[i].minlenp = minlenp;
1501 data->substrs[i].lookbehind = 0;
1504 SvCUR_set(data->last_found, 0);
1506 SV * const sv = data->last_found;
1507 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1508 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1513 data->last_end = -1;
1514 data->flags &= ~SF_BEFORE_EOL;
1515 DEBUG_STUDYDATA("commit", data, 0, is_inf);
1518 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1519 * list that describes which code points it matches */
1522 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1524 /* Set the SSC 'ssc' to match an empty string or any code point */
1526 PERL_ARGS_ASSERT_SSC_ANYTHING;
1528 assert(is_ANYOF_SYNTHETIC(ssc));
1530 /* mortalize so won't leak */
1531 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
1532 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1536 S_ssc_is_anything(const regnode_ssc *ssc)
1538 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1539 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1540 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1541 * in any way, so there's no point in using it */
1546 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1548 assert(is_ANYOF_SYNTHETIC(ssc));
1550 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1554 /* See if the list consists solely of the range 0 - Infinity */
1555 invlist_iterinit(ssc->invlist);
1556 ret = invlist_iternext(ssc->invlist, &start, &end)
1560 invlist_iterfinish(ssc->invlist);
1566 /* If e.g., both \w and \W are set, matches everything */
1567 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1569 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1570 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1580 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1582 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1583 * string, any code point, or any posix class under locale */
1585 PERL_ARGS_ASSERT_SSC_INIT;
1587 Zero(ssc, 1, regnode_ssc);
1588 set_ANYOF_SYNTHETIC(ssc);
1589 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1592 /* If any portion of the regex is to operate under locale rules that aren't
1593 * fully known at compile time, initialization includes it. The reason
1594 * this isn't done for all regexes is that the optimizer was written under
1595 * the assumption that locale was all-or-nothing. Given the complexity and
1596 * lack of documentation in the optimizer, and that there are inadequate
1597 * test cases for locale, many parts of it may not work properly, it is
1598 * safest to avoid locale unless necessary. */
1599 if (RExC_contains_locale) {
1600 ANYOF_POSIXL_SETALL(ssc);
1603 ANYOF_POSIXL_ZERO(ssc);
1608 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1609 const regnode_ssc *ssc)
1611 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1612 * to the list of code points matched, and locale posix classes; hence does
1613 * not check its flags) */
1618 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1620 assert(is_ANYOF_SYNTHETIC(ssc));
1622 invlist_iterinit(ssc->invlist);
1623 ret = invlist_iternext(ssc->invlist, &start, &end)
1627 invlist_iterfinish(ssc->invlist);
1633 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1640 #define INVLIST_INDEX 0
1641 #define ONLY_LOCALE_MATCHES_INDEX 1
1642 #define DEFERRED_USER_DEFINED_INDEX 2
1645 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1646 const regnode_charclass* const node)
1648 /* Returns a mortal inversion list defining which code points are matched
1649 * by 'node', which is of type ANYOF. Handles complementing the result if
1650 * appropriate. If some code points aren't knowable at this time, the
1651 * returned list must, and will, contain every code point that is a
1655 SV* only_utf8_locale_invlist = NULL;
1657 const U32 n = ARG(node);
1658 bool new_node_has_latin1 = FALSE;
1659 const U8 flags = (inRANGE(OP(node), ANYOFH, ANYOFRb))
1661 : ANYOF_FLAGS(node);
1663 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1665 /* Look at the data structure created by S_set_ANYOF_arg() */
1666 if (n != ANYOF_ONLY_HAS_BITMAP) {
1667 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1668 AV * const av = MUTABLE_AV(SvRV(rv));
1669 SV **const ary = AvARRAY(av);
1670 assert(RExC_rxi->data->what[n] == 's');
1672 if (av_tindex_skip_len_mg(av) >= DEFERRED_USER_DEFINED_INDEX) {
1674 /* Here there are things that won't be known until runtime -- we
1675 * have to assume it could be anything */
1676 invlist = sv_2mortal(_new_invlist(1));
1677 return _add_range_to_invlist(invlist, 0, UV_MAX);
1679 else if (ary[INVLIST_INDEX]) {
1681 /* Use the node's inversion list */
1682 invlist = sv_2mortal(invlist_clone(ary[INVLIST_INDEX], NULL));
1685 /* Get the code points valid only under UTF-8 locales */
1686 if ( (flags & ANYOFL_FOLD)
1687 && av_tindex_skip_len_mg(av) >= ONLY_LOCALE_MATCHES_INDEX)
1689 only_utf8_locale_invlist = ary[ONLY_LOCALE_MATCHES_INDEX];
1694 invlist = sv_2mortal(_new_invlist(0));
1697 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1698 * code points, and an inversion list for the others, but if there are code
1699 * points that should match only conditionally on the target string being
1700 * UTF-8, those are placed in the inversion list, and not the bitmap.
1701 * Since there are circumstances under which they could match, they are
1702 * included in the SSC. But if the ANYOF node is to be inverted, we have
1703 * to exclude them here, so that when we invert below, the end result
1704 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1705 * have to do this here before we add the unconditionally matched code
1707 if (flags & ANYOF_INVERT) {
1708 _invlist_intersection_complement_2nd(invlist,
1713 /* Add in the points from the bit map */
1714 if (! inRANGE(OP(node), ANYOFH, ANYOFRb)) {
1715 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1716 if (ANYOF_BITMAP_TEST(node, i)) {
1717 unsigned int start = i++;
1719 for (; i < NUM_ANYOF_CODE_POINTS
1720 && ANYOF_BITMAP_TEST(node, i); ++i)
1724 invlist = _add_range_to_invlist(invlist, start, i-1);
1725 new_node_has_latin1 = TRUE;
1730 /* If this can match all upper Latin1 code points, have to add them
1731 * as well. But don't add them if inverting, as when that gets done below,
1732 * it would exclude all these characters, including the ones it shouldn't
1733 * that were added just above */
1734 if (! (flags & ANYOF_INVERT) && OP(node) == ANYOFD
1735 && (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1737 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1740 /* Similarly for these */
1741 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1742 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1745 if (flags & ANYOF_INVERT) {
1746 _invlist_invert(invlist);
1748 else if (flags & ANYOFL_FOLD) {
1749 if (new_node_has_latin1) {
1751 /* Under /li, any 0-255 could fold to any other 0-255, depending on
1752 * the locale. We can skip this if there are no 0-255 at all. */
1753 _invlist_union(invlist, PL_Latin1, &invlist);
1755 invlist = add_cp_to_invlist(invlist, LATIN_SMALL_LETTER_DOTLESS_I);
1756 invlist = add_cp_to_invlist(invlist, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
1759 if (_invlist_contains_cp(invlist, LATIN_SMALL_LETTER_DOTLESS_I)) {
1760 invlist = add_cp_to_invlist(invlist, 'I');
1762 if (_invlist_contains_cp(invlist,
1763 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE))
1765 invlist = add_cp_to_invlist(invlist, 'i');
1770 /* Similarly add the UTF-8 locale possible matches. These have to be
1771 * deferred until after the non-UTF-8 locale ones are taken care of just
1772 * above, or it leads to wrong results under ANYOF_INVERT */
1773 if (only_utf8_locale_invlist) {
1774 _invlist_union_maybe_complement_2nd(invlist,
1775 only_utf8_locale_invlist,
1776 flags & ANYOF_INVERT,
1783 /* These two functions currently do the exact same thing */
1784 #define ssc_init_zero ssc_init
1786 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1787 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1789 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1790 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1791 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1794 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1795 const regnode_charclass *and_with)
1797 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1798 * another SSC or a regular ANYOF class. Can create false positives. */
1801 U8 and_with_flags = inRANGE(OP(and_with), ANYOFH, ANYOFRb)
1803 : ANYOF_FLAGS(and_with);
1806 PERL_ARGS_ASSERT_SSC_AND;
1808 assert(is_ANYOF_SYNTHETIC(ssc));
1810 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1811 * the code point inversion list and just the relevant flags */
1812 if (is_ANYOF_SYNTHETIC(and_with)) {
1813 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1814 anded_flags = and_with_flags;
1816 /* XXX This is a kludge around what appears to be deficiencies in the
1817 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1818 * there are paths through the optimizer where it doesn't get weeded
1819 * out when it should. And if we don't make some extra provision for
1820 * it like the code just below, it doesn't get added when it should.
1821 * This solution is to add it only when AND'ing, which is here, and
1822 * only when what is being AND'ed is the pristine, original node
1823 * matching anything. Thus it is like adding it to ssc_anything() but
1824 * only when the result is to be AND'ed. Probably the same solution
1825 * could be adopted for the same problem we have with /l matching,
1826 * which is solved differently in S_ssc_init(), and that would lead to
1827 * fewer false positives than that solution has. But if this solution
1828 * creates bugs, the consequences are only that a warning isn't raised
1829 * that should be; while the consequences for having /l bugs is
1830 * incorrect matches */
1831 if (ssc_is_anything((regnode_ssc *)and_with)) {
1832 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1836 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1837 if (OP(and_with) == ANYOFD) {
1838 anded_flags = and_with_flags & ANYOF_COMMON_FLAGS;
1841 anded_flags = and_with_flags
1842 &( ANYOF_COMMON_FLAGS
1843 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1844 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1845 if (ANYOFL_UTF8_LOCALE_REQD(and_with_flags)) {
1847 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1852 ANYOF_FLAGS(ssc) &= anded_flags;
1854 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1855 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1856 * 'and_with' may be inverted. When not inverted, we have the situation of
1858 * (C1 | P1) & (C2 | P2)
1859 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1860 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1861 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1862 * <= ((C1 & C2) | P1 | P2)
1863 * Alternatively, the last few steps could be:
1864 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1865 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1866 * <= (C1 | C2 | (P1 & P2))
1867 * We favor the second approach if either P1 or P2 is non-empty. This is
1868 * because these components are a barrier to doing optimizations, as what
1869 * they match cannot be known until the moment of matching as they are
1870 * dependent on the current locale, 'AND"ing them likely will reduce or
1872 * But we can do better if we know that C1,P1 are in their initial state (a
1873 * frequent occurrence), each matching everything:
1874 * (<everything>) & (C2 | P2) = C2 | P2
1875 * Similarly, if C2,P2 are in their initial state (again a frequent
1876 * occurrence), the result is a no-op
1877 * (C1 | P1) & (<everything>) = C1 | P1
1880 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1881 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1882 * <= (C1 & ~C2) | (P1 & ~P2)
1885 if ((and_with_flags & ANYOF_INVERT)
1886 && ! is_ANYOF_SYNTHETIC(and_with))
1890 ssc_intersection(ssc,
1892 FALSE /* Has already been inverted */
1895 /* If either P1 or P2 is empty, the intersection will be also; can skip
1897 if (! (and_with_flags & ANYOF_MATCHES_POSIXL)) {
1898 ANYOF_POSIXL_ZERO(ssc);
1900 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1902 /* Note that the Posix class component P from 'and_with' actually
1904 * P = Pa | Pb | ... | Pn
1905 * where each component is one posix class, such as in [\w\s].
1907 * ~P = ~(Pa | Pb | ... | Pn)
1908 * = ~Pa & ~Pb & ... & ~Pn
1909 * <= ~Pa | ~Pb | ... | ~Pn
1910 * The last is something we can easily calculate, but unfortunately
1911 * is likely to have many false positives. We could do better
1912 * in some (but certainly not all) instances if two classes in
1913 * P have known relationships. For example
1914 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1916 * :lower: & :print: = :lower:
1917 * And similarly for classes that must be disjoint. For example,
1918 * since \s and \w can have no elements in common based on rules in
1919 * the POSIX standard,
1920 * \w & ^\S = nothing
1921 * Unfortunately, some vendor locales do not meet the Posix
1922 * standard, in particular almost everything by Microsoft.
1923 * The loop below just changes e.g., \w into \W and vice versa */
1925 regnode_charclass_posixl temp;
1926 int add = 1; /* To calculate the index of the complement */
1928 Zero(&temp, 1, regnode_charclass_posixl);
1929 ANYOF_POSIXL_ZERO(&temp);
1930 for (i = 0; i < ANYOF_MAX; i++) {
1932 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1933 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1935 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1936 ANYOF_POSIXL_SET(&temp, i + add);
1938 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1940 ANYOF_POSIXL_AND(&temp, ssc);
1942 } /* else ssc already has no posixes */
1943 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1944 in its initial state */
1945 else if (! is_ANYOF_SYNTHETIC(and_with)
1946 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1948 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1949 * copy it over 'ssc' */
1950 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1951 if (is_ANYOF_SYNTHETIC(and_with)) {
1952 StructCopy(and_with, ssc, regnode_ssc);
1955 ssc->invlist = anded_cp_list;
1956 ANYOF_POSIXL_ZERO(ssc);
1957 if (and_with_flags & ANYOF_MATCHES_POSIXL) {
1958 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1962 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1963 || (and_with_flags & ANYOF_MATCHES_POSIXL))
1965 /* One or the other of P1, P2 is non-empty. */
1966 if (and_with_flags & ANYOF_MATCHES_POSIXL) {
1967 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1969 ssc_union(ssc, anded_cp_list, FALSE);
1971 else { /* P1 = P2 = empty */
1972 ssc_intersection(ssc, anded_cp_list, FALSE);
1978 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1979 const regnode_charclass *or_with)
1981 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1982 * another SSC or a regular ANYOF class. Can create false positives if
1983 * 'or_with' is to be inverted. */
1987 U8 or_with_flags = inRANGE(OP(or_with), ANYOFH, ANYOFRb)
1989 : ANYOF_FLAGS(or_with);
1991 PERL_ARGS_ASSERT_SSC_OR;
1993 assert(is_ANYOF_SYNTHETIC(ssc));
1995 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1996 * the code point inversion list and just the relevant flags */
1997 if (is_ANYOF_SYNTHETIC(or_with)) {
1998 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1999 ored_flags = or_with_flags;
2002 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
2003 ored_flags = or_with_flags & ANYOF_COMMON_FLAGS;
2004 if (OP(or_with) != ANYOFD) {
2007 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
2008 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
2009 if (ANYOFL_UTF8_LOCALE_REQD(or_with_flags)) {
2011 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
2016 ANYOF_FLAGS(ssc) |= ored_flags;
2018 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
2019 * C2 is the list of code points in 'or-with'; P2, its posix classes.
2020 * 'or_with' may be inverted. When not inverted, we have the simple
2021 * situation of computing:
2022 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
2023 * If P1|P2 yields a situation with both a class and its complement are
2024 * set, like having both \w and \W, this matches all code points, and we
2025 * can delete these from the P component of the ssc going forward. XXX We
2026 * might be able to delete all the P components, but I (khw) am not certain
2027 * about this, and it is better to be safe.
2030 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
2031 * <= (C1 | P1) | ~C2
2032 * <= (C1 | ~C2) | P1
2033 * (which results in actually simpler code than the non-inverted case)
2036 if ((or_with_flags & ANYOF_INVERT)
2037 && ! is_ANYOF_SYNTHETIC(or_with))
2039 /* We ignore P2, leaving P1 going forward */
2040 } /* else Not inverted */
2041 else if (or_with_flags & ANYOF_MATCHES_POSIXL) {
2042 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
2043 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
2045 for (i = 0; i < ANYOF_MAX; i += 2) {
2046 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
2048 ssc_match_all_cp(ssc);
2049 ANYOF_POSIXL_CLEAR(ssc, i);
2050 ANYOF_POSIXL_CLEAR(ssc, i+1);
2058 FALSE /* Already has been inverted */
2063 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
2065 PERL_ARGS_ASSERT_SSC_UNION;
2067 assert(is_ANYOF_SYNTHETIC(ssc));
2069 _invlist_union_maybe_complement_2nd(ssc->invlist,
2076 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
2078 const bool invert2nd)
2080 PERL_ARGS_ASSERT_SSC_INTERSECTION;
2082 assert(is_ANYOF_SYNTHETIC(ssc));
2084 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
2091 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
2093 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
2095 assert(is_ANYOF_SYNTHETIC(ssc));
2097 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
2101 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
2103 /* AND just the single code point 'cp' into the SSC 'ssc' */
2105 SV* cp_list = _new_invlist(2);
2107 PERL_ARGS_ASSERT_SSC_CP_AND;
2109 assert(is_ANYOF_SYNTHETIC(ssc));
2111 cp_list = add_cp_to_invlist(cp_list, cp);
2112 ssc_intersection(ssc, cp_list,
2113 FALSE /* Not inverted */
2115 SvREFCNT_dec_NN(cp_list);
2119 S_ssc_clear_locale(regnode_ssc *ssc)
2121 /* Set the SSC 'ssc' to not match any locale things */
2122 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
2124 assert(is_ANYOF_SYNTHETIC(ssc));
2126 ANYOF_POSIXL_ZERO(ssc);
2127 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
2131 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
2133 /* The synthetic start class is used to hopefully quickly winnow down
2134 * places where a pattern could start a match in the target string. If it
2135 * doesn't really narrow things down that much, there isn't much point to
2136 * having the overhead of using it. This function uses some very crude
2137 * heuristics to decide if to use the ssc or not.
2139 * It returns TRUE if 'ssc' rules out more than half what it considers to
2140 * be the "likely" possible matches, but of course it doesn't know what the
2141 * actual things being matched are going to be; these are only guesses
2143 * For /l matches, it assumes that the only likely matches are going to be
2144 * in the 0-255 range, uniformly distributed, so half of that is 127
2145 * For /a and /d matches, it assumes that the likely matches will be just
2146 * the ASCII range, so half of that is 63
2147 * For /u and there isn't anything matching above the Latin1 range, it
2148 * assumes that that is the only range likely to be matched, and uses
2149 * half that as the cut-off: 127. If anything matches above Latin1,
2150 * it assumes that all of Unicode could match (uniformly), except for
2151 * non-Unicode code points and things in the General Category "Other"
2152 * (unassigned, private use, surrogates, controls and formats). This
2153 * is a much large number. */
2155 U32 count = 0; /* Running total of number of code points matched by
2157 UV start, end; /* Start and end points of current range in inversion
2158 XXX outdated. UTF-8 locales are common, what about invert? list */
2159 const U32 max_code_points = (LOC)
2161 : (( ! UNI_SEMANTICS
2162 || invlist_highest(ssc->invlist) < 256)
2165 const U32 max_match = max_code_points / 2;
2167 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
2169 invlist_iterinit(ssc->invlist);
2170 while (invlist_iternext(ssc->invlist, &start, &end)) {
2171 if (start >= max_code_points) {
2174 end = MIN(end, max_code_points - 1);
2175 count += end - start + 1;
2176 if (count >= max_match) {
2177 invlist_iterfinish(ssc->invlist);
2187 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
2189 /* The inversion list in the SSC is marked mortal; now we need a more
2190 * permanent copy, which is stored the same way that is done in a regular
2191 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
2194 SV* invlist = invlist_clone(ssc->invlist, NULL);
2196 PERL_ARGS_ASSERT_SSC_FINALIZE;
2198 assert(is_ANYOF_SYNTHETIC(ssc));
2200 /* The code in this file assumes that all but these flags aren't relevant
2201 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
2202 * by the time we reach here */
2203 assert(! (ANYOF_FLAGS(ssc)
2204 & ~( ANYOF_COMMON_FLAGS
2205 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
2206 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
2208 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
2210 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist, NULL, NULL);
2211 SvREFCNT_dec(invlist);
2213 /* Make sure is clone-safe */
2214 ssc->invlist = NULL;
2216 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
2217 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
2218 OP(ssc) = ANYOFPOSIXL;
2220 else if (RExC_contains_locale) {
2224 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
2227 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
2228 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
2229 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
2230 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
2231 ? (TRIE_LIST_CUR( idx ) - 1) \
2237 dump_trie(trie,widecharmap,revcharmap)
2238 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
2239 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
2241 These routines dump out a trie in a somewhat readable format.
2242 The _interim_ variants are used for debugging the interim
2243 tables that are used to generate the final compressed
2244 representation which is what dump_trie expects.
2246 Part of the reason for their existence is to provide a form
2247 of documentation as to how the different representations function.
2252 Dumps the final compressed table form of the trie to Perl_debug_log.
2253 Used for debugging make_trie().
2257 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
2258 AV *revcharmap, U32 depth)
2261 SV *sv=sv_newmortal();
2262 int colwidth= widecharmap ? 6 : 4;
2264 DECLARE_AND_GET_RE_DEBUG_FLAGS;
2266 PERL_ARGS_ASSERT_DUMP_TRIE;
2268 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
2269 depth+1, "Match","Base","Ofs" );
2271 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2272 SV ** const tmp = av_fetch( revcharmap, state, 0);
2274 Perl_re_printf( aTHX_ "%*s",
2276 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2277 PL_colors[0], PL_colors[1],
2278 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2279 PERL_PV_ESCAPE_FIRSTCHAR
2284 Perl_re_printf( aTHX_ "\n");
2285 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
2287 for( state = 0 ; state < trie->uniquecharcount ; state++ )
2288 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2289 Perl_re_printf( aTHX_ "\n");
2291 for( state = 1 ; state < trie->statecount ; state++ ) {
2292 const U32 base = trie->states[ state ].trans.base;
2294 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
2296 if ( trie->states[ state ].wordnum ) {
2297 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
2299 Perl_re_printf( aTHX_ "%6s", "" );
2302 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
2307 while( ( base + ofs < trie->uniquecharcount ) ||
2308 ( base + ofs - trie->uniquecharcount < trie->lasttrans
2309 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2313 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
2315 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2316 if ( ( base + ofs >= trie->uniquecharcount )
2317 && ( base + ofs - trie->uniquecharcount
2319 && trie->trans[ base + ofs
2320 - trie->uniquecharcount ].check == state )
2322 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
2323 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2326 Perl_re_printf( aTHX_ "%*s", colwidth," ." );
2330 Perl_re_printf( aTHX_ "]");
2333 Perl_re_printf( aTHX_ "\n" );
2335 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
2337 for (word=1; word <= trie->wordcount; word++) {
2338 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2339 (int)word, (int)(trie->wordinfo[word].prev),
2340 (int)(trie->wordinfo[word].len));
2342 Perl_re_printf( aTHX_ "\n" );
2345 Dumps a fully constructed but uncompressed trie in list form.
2346 List tries normally only are used for construction when the number of
2347 possible chars (trie->uniquecharcount) is very high.
2348 Used for debugging make_trie().
2351 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2352 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2356 SV *sv=sv_newmortal();
2357 int colwidth= widecharmap ? 6 : 4;
2358 DECLARE_AND_GET_RE_DEBUG_FLAGS;
2360 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2362 /* print out the table precompression. */
2363 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
2365 Perl_re_indentf( aTHX_ "%s",
2366 depth+1, "------:-----+-----------------\n" );
2368 for( state=1 ; state < next_alloc ; state ++ ) {
2371 Perl_re_indentf( aTHX_ " %4" UVXf " :",
2372 depth+1, (UV)state );
2373 if ( ! trie->states[ state ].wordnum ) {
2374 Perl_re_printf( aTHX_ "%5s| ","");
2376 Perl_re_printf( aTHX_ "W%4x| ",
2377 trie->states[ state ].wordnum
2380 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2381 SV ** const tmp = av_fetch( revcharmap,
2382 TRIE_LIST_ITEM(state, charid).forid, 0);
2384 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
2386 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2388 PL_colors[0], PL_colors[1],
2389 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2390 | PERL_PV_ESCAPE_FIRSTCHAR
2392 TRIE_LIST_ITEM(state, charid).forid,
2393 (UV)TRIE_LIST_ITEM(state, charid).newstate
2396 Perl_re_printf( aTHX_ "\n%*s| ",
2397 (int)((depth * 2) + 14), "");
2400 Perl_re_printf( aTHX_ "\n");
2405 Dumps a fully constructed but uncompressed trie in table form.
2406 This is the normal DFA style state transition table, with a few
2407 twists to facilitate compression later.
2408 Used for debugging make_trie().
2411 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2412 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2417 SV *sv=sv_newmortal();
2418 int colwidth= widecharmap ? 6 : 4;
2419 DECLARE_AND_GET_RE_DEBUG_FLAGS;
2421 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2424 print out the table precompression so that we can do a visual check
2425 that they are identical.
2428 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
2430 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2431 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2433 Perl_re_printf( aTHX_ "%*s",
2435 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2436 PL_colors[0], PL_colors[1],
2437 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2438 PERL_PV_ESCAPE_FIRSTCHAR
2444 Perl_re_printf( aTHX_ "\n");
2445 Perl_re_indentf( aTHX_ "State+-", depth+1 );
2447 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2448 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
2451 Perl_re_printf( aTHX_ "\n" );
2453 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2455 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
2457 (UV)TRIE_NODENUM( state ) );
2459 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2460 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2462 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
2464 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
2466 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2467 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
2468 (UV)trie->trans[ state ].check );
2470 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
2471 (UV)trie->trans[ state ].check,
2472 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2480 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2481 startbranch: the first branch in the whole branch sequence
2482 first : start branch of sequence of branch-exact nodes.
2483 May be the same as startbranch
2484 last : Thing following the last branch.
2485 May be the same as tail.
2486 tail : item following the branch sequence
2487 count : words in the sequence
2488 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2489 depth : indent depth
2491 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2493 A trie is an N'ary tree where the branches are determined by digital
2494 decomposition of the key. IE, at the root node you look up the 1st character and
2495 follow that branch repeat until you find the end of the branches. Nodes can be
2496 marked as "accepting" meaning they represent a complete word. Eg:
2500 would convert into the following structure. Numbers represent states, letters
2501 following numbers represent valid transitions on the letter from that state, if
2502 the number is in square brackets it represents an accepting state, otherwise it
2503 will be in parenthesis.
2505 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2509 (1) +-i->(6)-+-s->[7]
2511 +-s->(3)-+-h->(4)-+-e->[5]
2513 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2515 This shows that when matching against the string 'hers' we will begin at state 1
2516 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2517 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2518 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2519 single traverse. We store a mapping from accepting to state to which word was
2520 matched, and then when we have multiple possibilities we try to complete the
2521 rest of the regex in the order in which they occurred in the alternation.
2523 The only prior NFA like behaviour that would be changed by the TRIE support is
2524 the silent ignoring of duplicate alternations which are of the form:
2526 / (DUPE|DUPE) X? (?{ ... }) Y /x
2528 Thus EVAL blocks following a trie may be called a different number of times with
2529 and without the optimisation. With the optimisations dupes will be silently
2530 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2531 the following demonstrates:
2533 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2535 which prints out 'word' three times, but
2537 'words'=~/(word|word|word)(?{ print $1 })S/
2539 which doesnt print it out at all. This is due to other optimisations kicking in.
2541 Example of what happens on a structural level:
2543 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2545 1: CURLYM[1] {1,32767}(18)
2556 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2557 and should turn into:
2559 1: CURLYM[1] {1,32767}(18)
2561 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2569 Cases where tail != last would be like /(?foo|bar)baz/:
2579 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2580 and would end up looking like:
2583 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2590 d = uvchr_to_utf8_flags(d, uv, 0);
2592 is the recommended Unicode-aware way of saying
2597 #define TRIE_STORE_REVCHAR(val) \
2600 SV *zlopp = newSV(UTF8_MAXBYTES); \
2601 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2602 unsigned char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2604 SvCUR_set(zlopp, kapow - flrbbbbb); \
2607 av_push(revcharmap, zlopp); \
2609 char ooooff = (char)val; \
2610 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2614 /* This gets the next character from the input, folding it if not already
2616 #define TRIE_READ_CHAR STMT_START { \
2619 /* if it is UTF then it is either already folded, or does not need \
2621 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2623 else if (folder == PL_fold_latin1) { \
2624 /* This folder implies Unicode rules, which in the range expressible \
2625 * by not UTF is the lower case, with the two exceptions, one of \
2626 * which should have been taken care of before calling this */ \
2627 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2628 uvc = toLOWER_L1(*uc); \
2629 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2632 /* raw data, will be folded later if needed */ \
2640 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2641 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2642 U32 ging = TRIE_LIST_LEN( state ) * 2; \
2643 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2644 TRIE_LIST_LEN( state ) = ging; \
2646 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2647 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2648 TRIE_LIST_CUR( state )++; \
2651 #define TRIE_LIST_NEW(state) STMT_START { \
2652 Newx( trie->states[ state ].trans.list, \
2653 4, reg_trie_trans_le ); \
2654 TRIE_LIST_CUR( state ) = 1; \
2655 TRIE_LIST_LEN( state ) = 4; \
2658 #define TRIE_HANDLE_WORD(state) STMT_START { \
2659 U16 dupe= trie->states[ state ].wordnum; \
2660 regnode * const noper_next = regnext( noper ); \
2663 /* store the word for dumping */ \
2665 if (OP(noper) != NOTHING) \
2666 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2668 tmp = newSVpvn_utf8( "", 0, UTF ); \
2669 av_push( trie_words, tmp ); \
2673 trie->wordinfo[curword].prev = 0; \
2674 trie->wordinfo[curword].len = wordlen; \
2675 trie->wordinfo[curword].accept = state; \
2677 if ( noper_next < tail ) { \
2679 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2681 trie->jump[curword] = (U16)(noper_next - convert); \
2683 jumper = noper_next; \
2685 nextbranch= regnext(cur); \
2689 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2690 /* chain, so that when the bits of chain are later */\
2691 /* linked together, the dups appear in the chain */\
2692 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2693 trie->wordinfo[dupe].prev = curword; \
2695 /* we haven't inserted this word yet. */ \
2696 trie->states[ state ].wordnum = curword; \
2701 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2702 ( ( base + charid >= ucharcount \
2703 && base + charid < ubound \
2704 && state == trie->trans[ base - ucharcount + charid ].check \
2705 && trie->trans[ base - ucharcount + charid ].next ) \
2706 ? trie->trans[ base - ucharcount + charid ].next \
2707 : ( state==1 ? special : 0 ) \
2710 #define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2712 TRIE_BITMAP_SET(trie, uvc); \
2713 /* store the folded codepoint */ \
2715 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2718 /* store first byte of utf8 representation of */ \
2719 /* variant codepoints */ \
2720 if (! UVCHR_IS_INVARIANT(uvc)) { \
2721 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2726 #define MADE_JUMP_TRIE 2
2727 #define MADE_EXACT_TRIE 4
2730 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2731 regnode *first, regnode *last, regnode *tail,
2732 U32 word_count, U32 flags, U32 depth)
2734 /* first pass, loop through and scan words */
2735 reg_trie_data *trie;
2736 HV *widecharmap = NULL;
2737 AV *revcharmap = newAV();
2743 regnode *jumper = NULL;
2744 regnode *nextbranch = NULL;
2745 regnode *convert = NULL;
2746 U32 *prev_states; /* temp array mapping each state to previous one */
2747 /* we just use folder as a flag in utf8 */
2748 const U8 * folder = NULL;
2750 /* in the below add_data call we are storing either 'tu' or 'tuaa'
2751 * which stands for one trie structure, one hash, optionally followed
2754 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuaa"));
2755 AV *trie_words = NULL;
2756 /* along with revcharmap, this only used during construction but both are
2757 * useful during debugging so we store them in the struct when debugging.
2760 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2761 STRLEN trie_charcount=0;
2763 SV *re_trie_maxbuff;
2764 DECLARE_AND_GET_RE_DEBUG_FLAGS;
2766 PERL_ARGS_ASSERT_MAKE_TRIE;
2768 PERL_UNUSED_ARG(depth);
2772 case EXACT: case EXACT_REQ8: case EXACTL: break;
2776 case EXACTFLU8: folder = PL_fold_latin1; break;
2777 case EXACTF: folder = PL_fold; break;
2778 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2781 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2783 trie->startstate = 1;
2784 trie->wordcount = word_count;
2785 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2786 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2787 if (flags == EXACT || flags == EXACT_REQ8 || flags == EXACTL)
2788 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2789 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2790 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2793 trie_words = newAV();
2796 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, GV_ADD);
2797 assert(re_trie_maxbuff);
2798 if (!SvIOK(re_trie_maxbuff)) {
2799 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2801 DEBUG_TRIE_COMPILE_r({
2802 Perl_re_indentf( aTHX_
2803 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2805 REG_NODE_NUM(startbranch), REG_NODE_NUM(first),
2806 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2809 /* Find the node we are going to overwrite */
2810 if ( first == startbranch && OP( last ) != BRANCH ) {
2811 /* whole branch chain */
2814 /* branch sub-chain */
2815 convert = NEXTOPER( first );
2818 /* -- First loop and Setup --
2820 We first traverse the branches and scan each word to determine if it
2821 contains widechars, and how many unique chars there are, this is
2822 important as we have to build a table with at least as many columns as we
2825 We use an array of integers to represent the character codes 0..255
2826 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2827 the native representation of the character value as the key and IV's for
2830 *TODO* If we keep track of how many times each character is used we can
2831 remap the columns so that the table compression later on is more
2832 efficient in terms of memory by ensuring the most common value is in the
2833 middle and the least common are on the outside. IMO this would be better
2834 than a most to least common mapping as theres a decent chance the most
2835 common letter will share a node with the least common, meaning the node
2836 will not be compressible. With a middle is most common approach the worst
2837 case is when we have the least common nodes twice.
2841 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2842 regnode *noper = NEXTOPER( cur );
2846 U32 wordlen = 0; /* required init */
2847 STRLEN minchars = 0;
2848 STRLEN maxchars = 0;
2849 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2852 if (OP(noper) == NOTHING) {
2853 /* skip past a NOTHING at the start of an alternation
2854 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2856 * If the next node is not something we are supposed to process
2857 * we will just ignore it due to the condition guarding the
2861 regnode *noper_next= regnext(noper);
2862 if (noper_next < tail)
2867 && ( OP(noper) == flags
2868 || (flags == EXACT && OP(noper) == EXACT_REQ8)
2869 || (flags == EXACTFU && ( OP(noper) == EXACTFU_REQ8
2870 || OP(noper) == EXACTFUP))))
2872 uc= (U8*)STRING(noper);
2873 e= uc + STR_LEN(noper);
2880 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2881 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2882 regardless of encoding */
2883 if (OP( noper ) == EXACTFUP) {
2884 /* false positives are ok, so just set this */
2885 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2889 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2891 TRIE_CHARCOUNT(trie)++;
2894 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2895 * is in effect. Under /i, this character can match itself, or
2896 * anything that folds to it. If not under /i, it can match just
2897 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2898 * all fold to k, and all are single characters. But some folds
2899 * expand to more than one character, so for example LATIN SMALL
2900 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2901 * the string beginning at 'uc' is 'ffi', it could be matched by
2902 * three characters, or just by the one ligature character. (It
2903 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2904 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2905 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2906 * match.) The trie needs to know the minimum and maximum number
2907 * of characters that could match so that it can use size alone to
2908 * quickly reject many match attempts. The max is simple: it is
2909 * the number of folded characters in this branch (since a fold is
2910 * never shorter than what folds to it. */
2914 /* And the min is equal to the max if not under /i (indicated by
2915 * 'folder' being NULL), or there are no multi-character folds. If
2916 * there is a multi-character fold, the min is incremented just
2917 * once, for the character that folds to the sequence. Each
2918 * character in the sequence needs to be added to the list below of
2919 * characters in the trie, but we count only the first towards the
2920 * min number of characters needed. This is done through the
2921 * variable 'foldlen', which is returned by the macros that look
2922 * for these sequences as the number of bytes the sequence
2923 * occupies. Each time through the loop, we decrement 'foldlen' by
2924 * how many bytes the current char occupies. Only when it reaches
2925 * 0 do we increment 'minchars' or look for another multi-character
2927 if (folder == NULL) {
2930 else if (foldlen > 0) {
2931 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2936 /* See if *uc is the beginning of a multi-character fold. If
2937 * so, we decrement the length remaining to look at, to account
2938 * for the current character this iteration. (We can use 'uc'
2939 * instead of the fold returned by TRIE_READ_CHAR because the
2940 * macro is smart enough to account for any unfolded
2943 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2944 foldlen -= UTF8SKIP(uc);
2947 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2952 /* The current character (and any potential folds) should be added
2953 * to the possible matching characters for this position in this
2957 U8 folded= folder[ (U8) uvc ];
2958 if ( !trie->charmap[ folded ] ) {
2959 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2960 TRIE_STORE_REVCHAR( folded );
2963 if ( !trie->charmap[ uvc ] ) {
2964 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2965 TRIE_STORE_REVCHAR( uvc );
2968 /* store the codepoint in the bitmap, and its folded
2970 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
2971 set_bit = 0; /* We've done our bit :-) */
2975 /* XXX We could come up with the list of code points that fold
2976 * to this using PL_utf8_foldclosures, except not for
2977 * multi-char folds, as there may be multiple combinations
2978 * there that could work, which needs to wait until runtime to
2979 * resolve (The comment about LIGATURE FFI above is such an
2984 widecharmap = newHV();
2986 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2989 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
2991 if ( !SvTRUE( *svpp ) ) {
2992 sv_setiv( *svpp, ++trie->uniquecharcount );
2993 TRIE_STORE_REVCHAR(uvc);
2996 } /* end loop through characters in this branch of the trie */
2998 /* We take the min and max for this branch and combine to find the min
2999 * and max for all branches processed so far */
3000 if( cur == first ) {
3001 trie->minlen = minchars;
3002 trie->maxlen = maxchars;
3003 } else if (minchars < trie->minlen) {
3004 trie->minlen = minchars;
3005 } else if (maxchars > trie->maxlen) {
3006 trie->maxlen = maxchars;
3008 } /* end first pass */
3009 DEBUG_TRIE_COMPILE_r(
3010 Perl_re_indentf( aTHX_
3011 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
3013 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
3014 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
3015 (int)trie->minlen, (int)trie->maxlen )
3019 We now know what we are dealing with in terms of unique chars and
3020 string sizes so we can calculate how much memory a naive
3021 representation using a flat table will take. If it's over a reasonable
3022 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
3023 conservative but potentially much slower representation using an array
3026 At the end we convert both representations into the same compressed
3027 form that will be used in regexec.c for matching with. The latter
3028 is a form that cannot be used to construct with but has memory
3029 properties similar to the list form and access properties similar
3030 to the table form making it both suitable for fast searches and
3031 small enough that its feasable to store for the duration of a program.
3033 See the comment in the code where the compressed table is produced
3034 inplace from the flat tabe representation for an explanation of how
3035 the compression works.
3040 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
3043 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
3044 > SvIV(re_trie_maxbuff) )
3047 Second Pass -- Array Of Lists Representation
3049 Each state will be represented by a list of charid:state records
3050 (reg_trie_trans_le) the first such element holds the CUR and LEN
3051 points of the allocated array. (See defines above).
3053 We build the initial structure using the lists, and then convert
3054 it into the compressed table form which allows faster lookups
3055 (but cant be modified once converted).
3058 STRLEN transcount = 1;
3060 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
3063 trie->states = (reg_trie_state *)
3064 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3065 sizeof(reg_trie_state) );
3069 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3071 regnode *noper = NEXTOPER( cur );
3072 U32 state = 1; /* required init */
3073 U16 charid = 0; /* sanity init */
3074 U32 wordlen = 0; /* required init */
3076 if (OP(noper) == NOTHING) {
3077 regnode *noper_next= regnext(noper);
3078 if (noper_next < tail)
3080 /* we will undo this assignment if noper does not
3081 * point at a trieable type in the else clause of
3082 * the following statement. */
3086 && ( OP(noper) == flags
3087 || (flags == EXACT && OP(noper) == EXACT_REQ8)
3088 || (flags == EXACTFU && ( OP(noper) == EXACTFU_REQ8
3089 || OP(noper) == EXACTFUP))))
3091 const U8 *uc= (U8*)STRING(noper);
3092 const U8 *e= uc + STR_LEN(noper);
3094 for ( ; uc < e ; uc += len ) {
3099 charid = trie->charmap[ uvc ];
3101 SV** const svpp = hv_fetch( widecharmap,
3108 charid=(U16)SvIV( *svpp );
3111 /* charid is now 0 if we dont know the char read, or
3112 * nonzero if we do */
3119 if ( !trie->states[ state ].trans.list ) {
3120 TRIE_LIST_NEW( state );
3123 check <= TRIE_LIST_USED( state );
3126 if ( TRIE_LIST_ITEM( state, check ).forid
3129 newstate = TRIE_LIST_ITEM( state, check ).newstate;
3134 newstate = next_alloc++;
3135 prev_states[newstate] = state;
3136 TRIE_LIST_PUSH( state, charid, newstate );
3141 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3145 /* If we end up here it is because we skipped past a NOTHING, but did not end up
3146 * on a trieable type. So we need to reset noper back to point at the first regop
3147 * in the branch before we call TRIE_HANDLE_WORD()
3149 noper= NEXTOPER(cur);
3151 TRIE_HANDLE_WORD(state);
3153 } /* end second pass */
3155 /* next alloc is the NEXT state to be allocated */
3156 trie->statecount = next_alloc;
3157 trie->states = (reg_trie_state *)
3158 PerlMemShared_realloc( trie->states,
3160 * sizeof(reg_trie_state) );
3162 /* and now dump it out before we compress it */
3163 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
3164 revcharmap, next_alloc,
3168 trie->trans = (reg_trie_trans *)
3169 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
3176 for( state=1 ; state < next_alloc ; state ++ ) {
3180 DEBUG_TRIE_COMPILE_MORE_r(
3181 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
3185 if (trie->states[state].trans.list) {
3186 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
3190 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3191 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
3192 if ( forid < minid ) {
3194 } else if ( forid > maxid ) {
3198 if ( transcount < tp + maxid - minid + 1) {
3200 trie->trans = (reg_trie_trans *)
3201 PerlMemShared_realloc( trie->trans,
3203 * sizeof(reg_trie_trans) );
3204 Zero( trie->trans + (transcount / 2),
3208 base = trie->uniquecharcount + tp - minid;
3209 if ( maxid == minid ) {
3211 for ( ; zp < tp ; zp++ ) {
3212 if ( ! trie->trans[ zp ].next ) {
3213 base = trie->uniquecharcount + zp - minid;
3214 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
3216 trie->trans[ zp ].check = state;
3222 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
3224 trie->trans[ tp ].check = state;
3229 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3230 const U32 tid = base
3231 - trie->uniquecharcount
3232 + TRIE_LIST_ITEM( state, idx ).forid;
3233 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
3235 trie->trans[ tid ].check = state;
3237 tp += ( maxid - minid + 1 );
3239 Safefree(trie->states[ state ].trans.list);
3242 DEBUG_TRIE_COMPILE_MORE_r(
3243 Perl_re_printf( aTHX_ " base: %d\n",base);
3246 trie->states[ state ].trans.base=base;
3248 trie->lasttrans = tp + 1;
3252 Second Pass -- Flat Table Representation.
3254 we dont use the 0 slot of either trans[] or states[] so we add 1 to
3255 each. We know that we will need Charcount+1 trans at most to store
3256 the data (one row per char at worst case) So we preallocate both
3257 structures assuming worst case.
3259 We then construct the trie using only the .next slots of the entry
3262 We use the .check field of the first entry of the node temporarily
3263 to make compression both faster and easier by keeping track of how
3264 many non zero fields are in the node.
3266 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
3269 There are two terms at use here: state as a TRIE_NODEIDX() which is
3270 a number representing the first entry of the node, and state as a
3271 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
3272 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
3273 if there are 2 entrys per node. eg:
3281 The table is internally in the right hand, idx form. However as we
3282 also have to deal with the states array which is indexed by nodenum
3283 we have to use TRIE_NODENUM() to convert.
3286 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
3289 trie->trans = (reg_trie_trans *)
3290 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
3291 * trie->uniquecharcount + 1,
3292 sizeof(reg_trie_trans) );
3293 trie->states = (reg_trie_state *)
3294 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3295 sizeof(reg_trie_state) );
3296 next_alloc = trie->uniquecharcount + 1;
3299 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3301 regnode *noper = NEXTOPER( cur );
3303 U32 state = 1; /* required init */
3305 U16 charid = 0; /* sanity init */
3306 U32 accept_state = 0; /* sanity init */
3308 U32 wordlen = 0; /* required init */
3310 if (OP(noper) == NOTHING) {
3311 regnode *noper_next= regnext(noper);
3312 if (noper_next < tail)
3314 /* we will undo this assignment if noper does not
3315 * point at a trieable type in the else clause of
3316 * the following statement. */
3320 && ( OP(noper) == flags
3321 || (flags == EXACT && OP(noper) == EXACT_REQ8)
3322 || (flags == EXACTFU && ( OP(noper) == EXACTFU_REQ8
3323 || OP(noper) == EXACTFUP))))
3325 const U8 *uc= (U8*)STRING(noper);
3326 const U8 *e= uc + STR_LEN(noper);
3328 for ( ; uc < e ; uc += len ) {
3333 charid = trie->charmap[ uvc ];
3335 SV* const * const svpp = hv_fetch( widecharmap,
3339 charid = svpp ? (U16)SvIV(*svpp) : 0;
3343 if ( !trie->trans[ state + charid ].next ) {
3344 trie->trans[ state + charid ].next = next_alloc;
3345 trie->trans[ state ].check++;
3346 prev_states[TRIE_NODENUM(next_alloc)]
3347 = TRIE_NODENUM(state);
3348 next_alloc += trie->uniquecharcount;
3350 state = trie->trans[ state + charid ].next;
3352 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3354 /* charid is now 0 if we dont know the char read, or
3355 * nonzero if we do */
3358 /* If we end up here it is because we skipped past a NOTHING, but did not end up
3359 * on a trieable type. So we need to reset noper back to point at the first regop
3360 * in the branch before we call TRIE_HANDLE_WORD().
3362 noper= NEXTOPER(cur);
3364 accept_state = TRIE_NODENUM( state );
3365 TRIE_HANDLE_WORD(accept_state);
3367 } /* end second pass */
3369 /* and now dump it out before we compress it */
3370 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3372 next_alloc, depth+1));
3376 * Inplace compress the table.*
3378 For sparse data sets the table constructed by the trie algorithm will
3379 be mostly 0/FAIL transitions or to put it another way mostly empty.
3380 (Note that leaf nodes will not contain any transitions.)
3382 This algorithm compresses the tables by eliminating most such
3383 transitions, at the cost of a modest bit of extra work during lookup:
3385 - Each states[] entry contains a .base field which indicates the
3386 index in the state[] array wheres its transition data is stored.
3388 - If .base is 0 there are no valid transitions from that node.
3390 - If .base is nonzero then charid is added to it to find an entry in
3393 -If trans[states[state].base+charid].check!=state then the
3394 transition is taken to be a 0/Fail transition. Thus if there are fail
3395 transitions at the front of the node then the .base offset will point
3396 somewhere inside the previous nodes data (or maybe even into a node
3397 even earlier), but the .check field determines if the transition is
3401 The following process inplace converts the table to the compressed
3402 table: We first do not compress the root node 1,and mark all its
3403 .check pointers as 1 and set its .base pointer as 1 as well. This
3404 allows us to do a DFA construction from the compressed table later,
3405 and ensures that any .base pointers we calculate later are greater
3408 - We set 'pos' to indicate the first entry of the second node.
3410 - We then iterate over the columns of the node, finding the first and
3411 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3412 and set the .check pointers accordingly, and advance pos
3413 appropriately and repreat for the next node. Note that when we copy
3414 the next pointers we have to convert them from the original
3415 NODEIDX form to NODENUM form as the former is not valid post
3418 - If a node has no transitions used we mark its base as 0 and do not
3419 advance the pos pointer.
3421 - If a node only has one transition we use a second pointer into the
3422 structure to fill in allocated fail transitions from other states.
3423 This pointer is independent of the main pointer and scans forward
3424 looking for null transitions that are allocated to a state. When it
3425 finds one it writes the single transition into the "hole". If the
3426 pointer doesnt find one the single transition is appended as normal.
3428 - Once compressed we can Renew/realloc the structures to release the
3431 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3432 specifically Fig 3.47 and the associated pseudocode.
3436 const U32 laststate = TRIE_NODENUM( next_alloc );
3439 trie->statecount = laststate;
3441 for ( state = 1 ; state < laststate ; state++ ) {
3443 const U32 stateidx = TRIE_NODEIDX( state );
3444 const U32 o_used = trie->trans[ stateidx ].check;
3445 U32 used = trie->trans[ stateidx ].check;
3446 trie->trans[ stateidx ].check = 0;
3449 used && charid < trie->uniquecharcount;
3452 if ( flag || trie->trans[ stateidx + charid ].next ) {
3453 if ( trie->trans[ stateidx + charid ].next ) {
3455 for ( ; zp < pos ; zp++ ) {
3456 if ( ! trie->trans[ zp ].next ) {
3460 trie->states[ state ].trans.base
3462 + trie->uniquecharcount
3464 trie->trans[ zp ].next
3465 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3467 trie->trans[ zp ].check = state;
3468 if ( ++zp > pos ) pos = zp;
3475 trie->states[ state ].trans.base
3476 = pos + trie->uniquecharcount - charid ;
3478 trie->trans[ pos ].next
3479 = SAFE_TRIE_NODENUM(
3480 trie->trans[ stateidx + charid ].next );
3481 trie->trans[ pos ].check = state;
3486 trie->lasttrans = pos + 1;
3487 trie->states = (reg_trie_state *)
3488 PerlMemShared_realloc( trie->states, laststate
3489 * sizeof(reg_trie_state) );
3490 DEBUG_TRIE_COMPILE_MORE_r(
3491 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
3493 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3497 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3500 } /* end table compress */
3502 DEBUG_TRIE_COMPILE_MORE_r(
3503 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
3505 (UV)trie->statecount,
3506 (UV)trie->lasttrans)
3508 /* resize the trans array to remove unused space */
3509 trie->trans = (reg_trie_trans *)
3510 PerlMemShared_realloc( trie->trans, trie->lasttrans
3511 * sizeof(reg_trie_trans) );
3513 { /* Modify the program and insert the new TRIE node */
3514 U8 nodetype =(U8)(flags & 0xFF);
3518 regnode *optimize = NULL;
3519 #ifdef RE_TRACK_PATTERN_OFFSETS
3522 U32 mjd_nodelen = 0;
3523 #endif /* RE_TRACK_PATTERN_OFFSETS */
3524 #endif /* DEBUGGING */
3526 This means we convert either the first branch or the first Exact,
3527 depending on whether the thing following (in 'last') is a branch
3528 or not and whther first is the startbranch (ie is it a sub part of
3529 the alternation or is it the whole thing.)
3530 Assuming its a sub part we convert the EXACT otherwise we convert
3531 the whole branch sequence, including the first.
3533 /* Find the node we are going to overwrite */
3534 if ( first != startbranch || OP( last ) == BRANCH ) {
3535 /* branch sub-chain */
3536 NEXT_OFF( first ) = (U16)(last - first);
3537 #ifdef RE_TRACK_PATTERN_OFFSETS
3539 mjd_offset= Node_Offset((convert));
3540 mjd_nodelen= Node_Length((convert));
3543 /* whole branch chain */
3545 #ifdef RE_TRACK_PATTERN_OFFSETS
3548 const regnode *nop = NEXTOPER( convert );
3549 mjd_offset= Node_Offset((nop));
3550 mjd_nodelen= Node_Length((nop));
3554 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
3556 (UV)mjd_offset, (UV)mjd_nodelen)
3559 /* But first we check to see if there is a common prefix we can
3560 split out as an EXACT and put in front of the TRIE node. */
3561 trie->startstate= 1;
3562 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3563 /* we want to find the first state that has more than
3564 * one transition, if that state is not the first state
3565 * then we have a common prefix which we can remove.
3568 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3570 I32 first_ofs = -1; /* keeps track of the ofs of the first
3571 transition, -1 means none */
3573 const U32 base = trie->states[ state ].trans.base;
3575 /* does this state terminate an alternation? */
3576 if ( trie->states[state].wordnum )
3579 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3580 if ( ( base + ofs >= trie->uniquecharcount ) &&
3581 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3582 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3584 if ( ++count > 1 ) {
3585 /* we have more than one transition */
3588 /* if this is the first state there is no common prefix
3589 * to extract, so we can exit */
3590 if ( state == 1 ) break;
3591 tmp = av_fetch( revcharmap, ofs, 0);
3592 ch = (U8*)SvPV_nolen_const( *tmp );
3594 /* if we are on count 2 then we need to initialize the
3595 * bitmap, and store the previous char if there was one
3598 /* clear the bitmap */
3599 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3601 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
3604 if (first_ofs >= 0) {
3605 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
3606 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3608 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3610 Perl_re_printf( aTHX_ "%s", (char*)ch)
3614 /* store the current firstchar in the bitmap */
3615 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3616 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
3622 /* This state has only one transition, its transition is part
3623 * of a common prefix - we need to concatenate the char it
3624 * represents to what we have so far. */
3625 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
3627 char *ch = SvPV( *tmp, len );
3629 SV *sv=sv_newmortal();
3630 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
3632 (UV)state, (UV)first_ofs,
3633 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3634 PL_colors[0], PL_colors[1],
3635 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3636 PERL_PV_ESCAPE_FIRSTCHAR
3641 OP( convert ) = nodetype;
3642 str=STRING(convert);
3643 setSTR_LEN(convert, 0);
3645 assert( ( STR_LEN(convert) + len ) < 256 );
3646 setSTR_LEN(convert, (U8)(STR_LEN(convert) + len));
3652 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
3657 trie->prefixlen = (state-1);
3659 regnode *n = convert+NODE_SZ_STR(convert);
3660 assert( NODE_SZ_STR(convert) <= U16_MAX );
3661 NEXT_OFF(convert) = (U16)(NODE_SZ_STR(convert));
3662 trie->startstate = state;
3663 trie->minlen -= (state - 1);
3664 trie->maxlen -= (state - 1);
3666 /* At least the UNICOS C compiler choked on this
3667 * being argument to DEBUG_r(), so let's just have
3670 #ifdef PERL_EXT_RE_BUILD
3676 regnode *fix = convert;
3677 U32 word = trie->wordcount;
3678 #ifdef RE_TRACK_PATTERN_OFFSETS
3681 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3682 while( ++fix < n ) {
3683 Set_Node_Offset_Length(fix, 0, 0);
3686 SV ** const tmp = av_fetch( trie_words, word, 0 );
3688 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3689 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3691 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3699 NEXT_OFF(convert) = (U16)(tail - convert);
3700 DEBUG_r(optimize= n);
3706 if ( trie->maxlen ) {
3707 NEXT_OFF( convert ) = (U16)(tail - convert);
3708 ARG_SET( convert, data_slot );
3709 /* Store the offset to the first unabsorbed branch in
3710 jump[0], which is otherwise unused by the jump logic.
3711 We use this when dumping a trie and during optimisation. */
3713 trie->jump[0] = (U16)(nextbranch - convert);
3715 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3716 * and there is a bitmap
3717 * and the first "jump target" node we found leaves enough room
3718 * then convert the TRIE node into a TRIEC node, with the bitmap
3719 * embedded inline in the opcode - this is hypothetically faster.
3721 if ( !trie->states[trie->startstate].wordnum
3723 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3725 OP( convert ) = TRIEC;
3726 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3727 PerlMemShared_free(trie->bitmap);
3730 OP( convert ) = TRIE;
3732 /* store the type in the flags */
3733 convert->flags = nodetype;
3737 + regarglen[ OP( convert ) ];
3739 /* XXX We really should free up the resource in trie now,
3740 as we won't use them - (which resources?) dmq */
3742 /* needed for dumping*/
3743 DEBUG_r(if (optimize) {
3744 regnode *opt = convert;
3746 while ( ++opt < optimize) {
3747 Set_Node_Offset_Length(opt, 0, 0);
3750 Try to clean up some of the debris left after the
3753 while( optimize < jumper ) {
3754 Track_Code( mjd_nodelen += Node_Length((optimize)); );
3755 OP( optimize ) = OPTIMIZED;
3756 Set_Node_Offset_Length(optimize, 0, 0);
3759 Set_Node_Offset_Length(convert, mjd_offset, mjd_nodelen);
3761 } /* end node insert */
3763 /* Finish populating the prev field of the wordinfo array. Walk back
3764 * from each accept state until we find another accept state, and if
3765 * so, point the first word's .prev field at the second word. If the
3766 * second already has a .prev field set, stop now. This will be the
3767 * case either if we've already processed that word's accept state,
3768 * or that state had multiple words, and the overspill words were
3769 * already linked up earlier.
3776 for (word=1; word <= trie->wordcount; word++) {
3778 if (trie->wordinfo[word].prev)
3780 state = trie->wordinfo[word].accept;
3782 state = prev_states[state];
3785 prev = trie->states[state].wordnum;
3789 trie->wordinfo[word].prev = prev;
3791 Safefree(prev_states);
3795 /* and now dump out the compressed format */
3796 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3798 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3800 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3801 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3803 SvREFCNT_dec_NN(revcharmap);
3807 : trie->startstate>1
3813 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3815 /* The Trie is constructed and compressed now so we can build a fail array if
3818 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3820 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3824 We find the fail state for each state in the trie, this state is the longest
3825 proper suffix of the current state's 'word' that is also a proper prefix of
3826 another word in our trie. State 1 represents the word '' and is thus the
3827 default fail state. This allows the DFA not to have to restart after its
3828 tried and failed a word at a given point, it simply continues as though it
3829 had been matching the other word in the first place.
3831 'abcdgu'=~/abcdefg|cdgu/
3832 When we get to 'd' we are still matching the first word, we would encounter
3833 'g' which would fail, which would bring us to the state representing 'd' in
3834 the second word where we would try 'g' and succeed, proceeding to match
3837 /* add a fail transition */
3838 const U32 trie_offset = ARG(source);
3839 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3841 const U32 ucharcount = trie->uniquecharcount;
3842 const U32 numstates = trie->statecount;
3843 const U32 ubound = trie->lasttrans + ucharcount;
3847 U32 base = trie->states[ 1 ].trans.base;
3850 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3852 DECLARE_AND_GET_RE_DEBUG_FLAGS;
3854 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3855 PERL_UNUSED_CONTEXT;
3857 PERL_UNUSED_ARG(depth);
3860 if ( OP(source) == TRIE ) {
3861 struct regnode_1 *op = (struct regnode_1 *)
3862 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3863 StructCopy(source, op, struct regnode_1);
3864 stclass = (regnode *)op;
3866 struct regnode_charclass *op = (struct regnode_charclass *)
3867 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3868 StructCopy(source, op, struct regnode_charclass);
3869 stclass = (regnode *)op;
3871 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3873 ARG_SET( stclass, data_slot );
3874 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3875 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3876 aho->trie=trie_offset;
3877 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3878 Copy( trie->states, aho->states, numstates, reg_trie_state );
3879 Newx( q, numstates, U32);
3880 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3883 /* initialize fail[0..1] to be 1 so that we always have
3884 a valid final fail state */
3885 fail[ 0 ] = fail[ 1 ] = 1;
3887 for ( charid = 0; charid < ucharcount ; charid++ ) {
3888 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3890 q[ q_write ] = newstate;
3891 /* set to point at the root */
3892 fail[ q[ q_write++ ] ]=1;
3895 while ( q_read < q_write) {
3896 const U32 cur = q[ q_read++ % numstates ];
3897 base = trie->states[ cur ].trans.base;
3899 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3900 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3902 U32 fail_state = cur;
3905 fail_state = fail[ fail_state ];
3906 fail_base = aho->states[ fail_state ].trans.base;
3907 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3909 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3910 fail[ ch_state ] = fail_state;
3911 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3913 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3915 q[ q_write++ % numstates] = ch_state;
3919 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3920 when we fail in state 1, this allows us to use the
3921 charclass scan to find a valid start char. This is based on the principle
3922 that theres a good chance the string being searched contains lots of stuff
3923 that cant be a start char.
3925 fail[ 0 ] = fail[ 1 ] = 0;
3926 DEBUG_TRIE_COMPILE_r({
3927 Perl_re_indentf( aTHX_ "Stclass Failtable (%" UVuf " states): 0",
3928 depth, (UV)numstates
3930 for( q_read=1; q_read<numstates; q_read++ ) {
3931 Perl_re_printf( aTHX_ ", %" UVuf, (UV)fail[q_read]);
3933 Perl_re_printf( aTHX_ "\n");
3936 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3941 /* The below joins as many adjacent EXACTish nodes as possible into a single
3942 * one. The regop may be changed if the node(s) contain certain sequences that
3943 * require special handling. The joining is only done if:
3944 * 1) there is room in the current conglomerated node to entirely contain the
3946 * 2) they are compatible node types
3948 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3949 * these get optimized out
3951 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3952 * as possible, even if that means splitting an existing node so that its first
3953 * part is moved to the preceeding node. This would maximise the efficiency of
3954 * memEQ during matching.
3956 * If a node is to match under /i (folded), the number of characters it matches
3957 * can be different than its character length if it contains a multi-character
3958 * fold. *min_subtract is set to the total delta number of characters of the
3961 * And *unfolded_multi_char is set to indicate whether or not the node contains
3962 * an unfolded multi-char fold. This happens when it won't be known until
3963 * runtime whether the fold is valid or not; namely
3964 * 1) for EXACTF nodes that contain LATIN SMALL LETTER SHARP S, as only if the
3965 * target string being matched against turns out to be UTF-8 is that fold
3967 * 2) for EXACTFL nodes whose folding rules depend on the locale in force at
3969 * (Multi-char folds whose components are all above the Latin1 range are not
3970 * run-time locale dependent, and have already been folded by the time this
3971 * function is called.)
3973 * This is as good a place as any to discuss the design of handling these
3974 * multi-character fold sequences. It's been wrong in Perl for a very long
3975 * time. There are three code points in Unicode whose multi-character folds
3976 * were long ago discovered to mess things up. The previous designs for
3977 * dealing with these involved assigning a special node for them. This
3978 * approach doesn't always work, as evidenced by this example:
3979 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3980 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3981 * would match just the \xDF, it won't be able to handle the case where a
3982 * successful match would have to cross the node's boundary. The new approach
3983 * that hopefully generally solves the problem generates an EXACTFUP node
3984 * that is "sss" in this case.
3986 * It turns out that there are problems with all multi-character folds, and not
3987 * just these three. Now the code is general, for all such cases. The
3988 * approach taken is:
3989 * 1) This routine examines each EXACTFish node that could contain multi-
3990 * character folded sequences. Since a single character can fold into
3991 * such a sequence, the minimum match length for this node is less than
3992 * the number of characters in the node. This routine returns in
3993 * *min_subtract how many characters to subtract from the actual
3994 * length of the string to get a real minimum match length; it is 0 if
3995 * there are no multi-char foldeds. This delta is used by the caller to
3996 * adjust the min length of the match, and the delta between min and max,
3997 * so that the optimizer doesn't reject these possibilities based on size
4000 * 2) For the sequence involving the LATIN SMALL LETTER SHARP S (U+00DF)
4001 * under /u, we fold it to 'ss' in regatom(), and in this routine, after
4002 * joining, we scan for occurrences of the sequence 'ss' in non-UTF-8
4003 * EXACTFU nodes. The node type of such nodes is then changed to
4004 * EXACTFUP, indicating it is problematic, and needs careful handling.
4005 * (The procedures in step 1) above are sufficient to handle this case in
4006 * UTF-8 encoded nodes.) The reason this is problematic is that this is
4007 * the only case where there is a possible fold length change in non-UTF-8
4008 * patterns. By reserving a special node type for problematic cases, the
4009 * far more common regular EXACTFU nodes can be processed faster.
4010 * regexec.c takes advantage of this.
4012 * EXACTFUP has been created as a grab-bag for (hopefully uncommon)
4013 * problematic cases. These all only occur when the pattern is not
4014 * UTF-8. In addition to the 'ss' sequence where there is a possible fold
4015 * length change, it handles the situation where the string cannot be
4016 * entirely folded. The strings in an EXACTFish node are folded as much
4017 * as possible during compilation in regcomp.c. This saves effort in
4018 * regex matching. By using an EXACTFUP node when it is not possible to
4019 * fully fold at compile time, regexec.c can know that everything in an
4020 * EXACTFU node is folded, so folding can be skipped at runtime. The only
4021 * case where folding in EXACTFU nodes can't be done at compile time is
4022 * the presumably uncommon MICRO SIGN, when the pattern isn't UTF-8. This
4023 * is because its fold requires UTF-8 to represent. Thus EXACTFUP nodes
4024 * handle two very different cases. Alternatively, there could have been
4025 * a node type where there are length changes, one for unfolded, and one
4026 * for both. If yet another special case needed to be created, the number
4027 * of required node types would have to go to 7. khw figures that even
4028 * though there are plenty of node types to spare, that the maintenance
4029 * cost wasn't worth the small speedup of doing it that way, especially
4030 * since he thinks the MICRO SIGN is rarely encountered in practice.
4032 * There are other cases where folding isn't done at compile time, but
4033 * none of them are under /u, and hence not for EXACTFU nodes. The folds
4034 * in EXACTFL nodes aren't known until runtime, and vary as the locale
4035 * changes. Some folds in EXACTF depend on if the runtime target string
4036 * is UTF-8 or not. (regatom() will create an EXACTFU node even under /di
4037 * when no fold in it depends on the UTF-8ness of the target string.)
4039 * 3) A problem remains for unfolded multi-char folds. (These occur when the
4040 * validity of the fold won't be known until runtime, and so must remain
4041 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFAA
4042 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
4043 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
4044 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
4045 * The reason this is a problem is that the optimizer part of regexec.c
4046 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
4047 * that a character in the pattern corresponds to at most a single
4048 * character in the target string. (And I do mean character, and not byte
4049 * here, unlike other parts of the documentation that have never been
4050 * updated to account for multibyte Unicode.) Sharp s in EXACTF and
4051 * EXACTFL nodes can match the two character string 'ss'; in EXACTFAA
4052 * nodes it can match "\x{17F}\x{17F}". These, along with other ones in
4053 * EXACTFL nodes, violate the assumption, and they are the only instances
4054 * where it is violated. I'm reluctant to try to change the assumption,
4055 * as the code involved is impenetrable to me (khw), so instead the code
4056 * here punts. This routine examines EXACTFL nodes, and (when the pattern
4057 * isn't UTF-8) EXACTF and EXACTFAA for such unfolded folds, and returns a
4058 * boolean indicating whether or not the node contains such a fold. When
4059 * it is true, the caller sets a flag that later causes the optimizer in
4060 * this file to not set values for the floating and fixed string lengths,
4061 * and thus avoids the optimizer code in regexec.c that makes the invalid
4062 * assumption. Thus, there is no optimization based on string lengths for
4063 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
4064 * EXACTF and EXACTFAA nodes that contain the sharp s. (The reason the
4065 * assumption is wrong only in these cases is that all other non-UTF-8
4066 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
4067 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
4068 * EXACTF nodes because we don't know at compile time if it actually
4069 * matches 'ss' or not. For EXACTF nodes it will match iff the target
4070 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
4071 * always matches; and EXACTFAA where it never does. In an EXACTFAA node
4072 * in a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
4073 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
4074 * string would require the pattern to be forced into UTF-8, the overhead
4075 * of which we want to avoid. Similarly the unfolded multi-char folds in
4076 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
4079 * Similarly, the code that generates tries doesn't currently handle
4080 * not-already-folded multi-char folds, and it looks like a pain to change
4081 * that. Therefore, trie generation of EXACTFAA nodes with the sharp s
4082 * doesn't work. Instead, such an EXACTFAA is turned into a new regnode,
4083 * EXACTFAA_NO_TRIE, which the trie code knows not to handle. Most people
4084 * using /iaa matching will be doing so almost entirely with ASCII
4085 * strings, so this should rarely be encountered in practice */
4088 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
4089 UV *min_subtract, bool *unfolded_multi_char,
4090 U32 flags, regnode *val, U32 depth)
4092 /* Merge several consecutive EXACTish nodes into one. */
4094 regnode *n = regnext(scan);
4096 regnode *next = scan + NODE_SZ_STR(scan);
4100 regnode *stop = scan;
4101 DECLARE_AND_GET_RE_DEBUG_FLAGS;
4103 PERL_UNUSED_ARG(depth);
4106 PERL_ARGS_ASSERT_JOIN_EXACT;
4107 #ifndef EXPERIMENTAL_INPLACESCAN
4108 PERL_UNUSED_ARG(flags);
4109 PERL_UNUSED_ARG(val);
4111 DEBUG_PEEP("join", scan, depth, 0);
4113 assert(PL_regkind[OP(scan)] == EXACT);
4115 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
4116 * EXACT ones that are mergeable to the current one. */
4118 && ( PL_regkind[OP(n)] == NOTHING
4119 || (stringok && PL_regkind[OP(n)] == EXACT))
4121 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
4124 if (OP(n) == TAIL || n > next)
4126 if (PL_regkind[OP(n)] == NOTHING) {
4127 DEBUG_PEEP("skip:", n, depth, 0);
4128 NEXT_OFF(scan) += NEXT_OFF(n);
4129 next = n + NODE_STEP_REGNODE;
4136 else if (stringok) {
4137 const unsigned int oldl = STR_LEN(scan);
4138 regnode * const nnext = regnext(n);
4140 /* XXX I (khw) kind of doubt that this works on platforms (should
4141 * Perl ever run on one) where U8_MAX is above 255 because of lots
4142 * of other assumptions */
4143 /* Don't join if the sum can't fit into a single node */
4144 if (oldl + STR_LEN(n) > U8_MAX)
4147 /* Joining something that requires UTF-8 with something that
4148 * doesn't, means the result requires UTF-8. */
4149 if (OP(scan) == EXACT && (OP(n) == EXACT_REQ8)) {
4150 OP(scan) = EXACT_REQ8;
4152 else if (OP(scan) == EXACT_REQ8 && (OP(n) == EXACT)) {
4153 ; /* join is compatible, no need to change OP */
4155 else if ((OP(scan) == EXACTFU) && (OP(n) == EXACTFU_REQ8)) {
4156 OP(scan) = EXACTFU_REQ8;
4158 else if ((OP(scan) == EXACTFU_REQ8) && (OP(n) == EXACTFU)) {
4159 ; /* join is compatible, no need to change OP */
4161 else if (OP(scan) == EXACTFU && OP(n) == EXACTFU) {
4162 ; /* join is compatible, no need to change OP */
4164 else if (OP(scan) == EXACTFU && OP(n) == EXACTFU_S_EDGE) {
4166 /* Under /di, temporary EXACTFU_S_EDGE nodes are generated,
4167 * which can join with EXACTFU ones. We check for this case
4168 * here. These need to be resolved to either EXACTFU or
4169 * EXACTF at joining time. They have nothing in them that
4170 * would forbid them from being the more desirable EXACTFU
4171 * nodes except that they begin and/or end with a single [Ss].
4172 * The reason this is problematic is because they could be
4173 * joined in this loop with an adjacent node that ends and/or
4174 * begins with [Ss] which would then form the sequence 'ss',
4175 * which matches differently under /di than /ui, in which case
4176 * EXACTFU can't be used. If the 'ss' sequence doesn't get
4177 * formed, the nodes get absorbed into any adjacent EXACTFU
4178 * node. And if the only adjacent node is EXACTF, they get
4179 * absorbed into that, under the theory that a longer node is
4180 * better than two shorter ones, even if one is EXACTFU. Note
4181 * that EXACTFU_REQ8 is generated only for UTF-8 patterns,
4182 * and the EXACTFU_S_EDGE ones only for non-UTF-8. */
4184 if (STRING(n)[STR_LEN(n)-1] == 's') {
4186 /* Here the joined node would end with 's'. If the node
4187 * following the combination is an EXACTF one, it's better to
4188 * join this trailing edge 's' node with that one, leaving the
4189 * current one in 'scan' be the more desirable EXACTFU */
4190 if (OP(nnext) == EXACTF) {
4194 OP(scan) = EXACTFU_S_EDGE;
4196 } /* Otherwise, the beginning 's' of the 2nd node just
4197 becomes an interior 's' in 'scan' */
4199 else if (OP(scan) == EXACTF && OP(n) == EXACTF) {
4200 ; /* join is compatible, no need to change OP */
4202 else if (OP(scan) == EXACTF && OP(n) == EXACTFU_S_EDGE) {
4204 /* EXACTF nodes are compatible for joining with EXACTFU_S_EDGE
4205 * nodes. But the latter nodes can be also joined with EXACTFU
4206 * ones, and that is a better outcome, so if the node following
4207 * 'n' is EXACTFU, quit now so that those two can be joined
4209 if (OP(nnext) == EXACTFU) {
4213 /* The join is compatible, and the combined node will be
4214 * EXACTF. (These don't care if they begin or end with 's' */
4216 else if (OP(scan) == EXACTFU_S_EDGE && OP(n) == EXACTFU_S_EDGE) {
4217 if ( STRING(scan)[STR_LEN(scan)-1] == 's'
4218 && STRING(n)[0] == 's')
4220 /* When combined, we have the sequence 'ss', which means we
4221 * have to remain /di */
4225 else if (OP(scan) == EXACTFU_S_EDGE && OP(n) == EXACTFU) {
4226 if (STRING(n)[0] == 's') {
4227 ; /* Here the join is compatible and the combined node
4228 starts with 's', no need to change OP */
4230 else { /* Now the trailing 's' is in the interior */
4234 else if (OP(scan) == EXACTFU_S_EDGE && OP(n) == EXACTF) {
4236 /* The join is compatible, and the combined node will be
4237 * EXACTF. (These don't care if they begin or end with 's' */
4240 else if (OP(scan) != OP(n)) {
4242 /* The only other compatible joinings are the same node type */
4246 DEBUG_PEEP("merg", n, depth, 0);
4249 NEXT_OFF(scan) += NEXT_OFF(n);
4250 assert( ( STR_LEN(scan) + STR_LEN(n) ) < 256 );
4251 setSTR_LEN(scan, (U8)(STR_LEN(scan) + STR_LEN(n)));
4252 next = n + NODE_SZ_STR(n);
4253 /* Now we can overwrite *n : */
4254 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
4262 #ifdef EXPERIMENTAL_INPLACESCAN
4263 if (flags && !NEXT_OFF(n)) {
4264 DEBUG_PEEP("atch", val, depth, 0);
4265 if (reg_off_by_arg[OP(n)]) {
4266 ARG_SET(n, val - n);
4269 NEXT_OFF(n) = val - n;
4276 /* This temporary node can now be turned into EXACTFU, and must, as
4277 * regexec.c doesn't handle it */
4278 if (OP(scan) == EXACTFU_S_EDGE) {
4283 *unfolded_multi_char = FALSE;
4285 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
4286 * can now analyze for sequences of problematic code points. (Prior to
4287 * this final joining, sequences could have been split over boundaries, and
4288 * hence missed). The sequences only happen in folding, hence for any
4289 * non-EXACT EXACTish node */
4290 if (OP(scan) != EXACT && OP(scan) != EXACT_REQ8 && OP(scan) != EXACTL) {
4291 U8* s0 = (U8*) STRING(scan);
4293 U8* s_end = s0 + STR_LEN(scan);
4295 int total_count_delta = 0; /* Total delta number of characters that
4296 multi-char folds expand to */
4298 /* One pass is made over the node's string looking for all the
4299 * possibilities. To avoid some tests in the loop, there are two main
4300 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
4305 if (OP(scan) == EXACTFL) {
4308 /* An EXACTFL node would already have been changed to another
4309 * node type unless there is at least one character in it that
4310 * is problematic; likely a character whose fold definition
4311 * won't be known until runtime, and so has yet to be folded.
4312 * For all but the UTF-8 locale, folds are 1-1 in length, but
4313 * to handle the UTF-8 case, we need to create a temporary
4314 * folded copy using UTF-8 locale rules in order to analyze it.
4315 * This is because our macros that look to see if a sequence is
4316 * a multi-char fold assume everything is folded (otherwise the
4317 * tests in those macros would be too complicated and slow).
4318 * Note that here, the non-problematic folds will have already
4319 * been done, so we can just copy such characters. We actually
4320 * don't completely fold the EXACTFL string. We skip the
4321 * unfolded multi-char folds, as that would just create work
4322 * below to figure out the size they already are */
4324 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
4327 STRLEN s_len = UTF8SKIP(s);
4328 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
4329 Copy(s, d, s_len, U8);
4332 else if (is_FOLDS_TO_MULTI_utf8(s)) {
4333 *unfolded_multi_char = TRUE;
4334 Copy(s, d, s_len, U8);
4337 else if (isASCII(*s)) {
4338 *(d++) = toFOLD(*s);
4342 _toFOLD_utf8_flags(s, s_end, d, &len, FOLD_FLAGS_FULL);
4348 /* Point the remainder of the routine to look at our temporary
4352 } /* End of creating folded copy of EXACTFL string */
4354 /* Examine the string for a multi-character fold sequence. UTF-8
4355 * patterns have all characters pre-folded by the time this code is
4357 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
4358 length sequence we are looking for is 2 */
4360 int count = 0; /* How many characters in a multi-char fold */
4361 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
4362 if (! len) { /* Not a multi-char fold: get next char */
4367 { /* Here is a generic multi-char fold. */
4368 U8* multi_end = s + len;
4370 /* Count how many characters are in it. In the case of
4371 * /aa, no folds which contain ASCII code points are
4372 * allowed, so check for those, and skip if found. */
4373 if (OP(scan) != EXACTFAA && OP(scan) != EXACTFAA_NO_TRIE) {
4374 count = utf8_length(s, multi_end);
4378 while (s < multi_end) {
4381 goto next_iteration;
4391 /* The delta is how long the sequence is minus 1 (1 is how long
4392 * the character that folds to the sequence is) */
4393 total_count_delta += count - 1;
4397 /* We created a temporary folded copy of the string in EXACTFL
4398 * nodes. Therefore we need to be sure it doesn't go below zero,
4399 * as the real string could be shorter */
4400 if (OP(scan) == EXACTFL) {
4401 int total_chars = utf8_length((U8*) STRING(scan),
4402 (U8*) STRING(scan) + STR_LEN(scan));
4403 if (total_count_delta > total_chars) {
4404 total_count_delta = total_chars;
4408 *min_subtract += total_count_delta;
4411 else if (OP(scan) == EXACTFAA) {
4413 /* Non-UTF-8 pattern, EXACTFAA node. There can't be a multi-char
4414 * fold to the ASCII range (and there are no existing ones in the
4415 * upper latin1 range). But, as outlined in the comments preceding
4416 * this function, we need to flag any occurrences of the sharp s.
4417 * This character forbids trie formation (because of added
4419 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
4420 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
4421 || UNICODE_DOT_DOT_VERSION > 0)
4423 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
4424 OP(scan) = EXACTFAA_NO_TRIE;
4425 *unfolded_multi_char = TRUE;
4431 else if (OP(scan) != EXACTFAA_NO_TRIE) {
4433 /* Non-UTF-8 pattern, not EXACTFAA node. Look for the multi-char
4434 * folds that are all Latin1. As explained in the comments
4435 * preceding this function, we look also for the sharp s in EXACTF
4436 * and EXACTFL nodes; it can be in the final position. Otherwise
4437 * we can stop looking 1 byte earlier because have to find at least
4438 * two characters for a multi-fold */
4439 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
4444 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
4445 if (! len) { /* Not a multi-char fold. */
4446 if (*s == LATIN_SMALL_LETTER_SHARP_S
4447 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
4449 *unfolded_multi_char = TRUE;
4456 && isALPHA_FOLD_EQ(*s, 's')
4457 && isALPHA_FOLD_EQ(*(s+1), 's'))
4460 /* EXACTF nodes need to know that the minimum length
4461 * changed so that a sharp s in the string can match this
4462 * ss in the pattern, but they remain EXACTF nodes, as they
4463 * won't match this unless the target string is in UTF-8,
4464 * which we don't know until runtime. EXACTFL nodes can't
4465 * transform into EXACTFU nodes */
4466 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
4467 OP(scan) = EXACTFUP;
4471 *min_subtract += len - 1;
4479 /* Allow dumping but overwriting the collection of skipped
4480 * ops and/or strings with fake optimized ops */
4481 n = scan + NODE_SZ_STR(scan);
4489 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl", scan, depth, 0);});
4493 /* REx optimizer. Converts nodes into quicker variants "in place".
4494 Finds fixed substrings. */
4496 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
4497 to the position after last scanned or to NULL. */
4499 #define INIT_AND_WITHP \
4500 assert(!and_withp); \
4501 Newx(and_withp, 1, regnode_ssc); \
4502 SAVEFREEPV(and_withp)
4506 S_unwind_scan_frames(pTHX_ const void *p)
4508 scan_frame *f= (scan_frame *)p;
4510 scan_frame *n= f->next_frame;
4516 /* Follow the next-chain of the current node and optimize away
4517 all the NOTHINGs from it.
4520 S_rck_elide_nothing(pTHX_ regnode *node)
4522 PERL_ARGS_ASSERT_RCK_ELIDE_NOTHING;
4524 if (OP(node) != CURLYX) {
4525 const int max = (reg_off_by_arg[OP(node)]
4527 /* I32 may be smaller than U16 on CRAYs! */
4528 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
4529 int off = (reg_off_by_arg[OP(node)] ? ARG(node) : NEXT_OFF(node));
4533 /* Skip NOTHING and LONGJMP. */
4537 (PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
4538 || ((OP(n) == LONGJMP) && (noff = ARG(n)))
4544 if (reg_off_by_arg[OP(node)])
4547 NEXT_OFF(node) = off;
4552 /* the return from this sub is the minimum length that could possibly match */
4554 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
4555 SSize_t *minlenp, SSize_t *deltap,
4560 regnode_ssc *and_withp,
4561 U32 flags, U32 depth, bool was_mutate_ok)
4562 /* scanp: Start here (read-write). */
4563 /* deltap: Write maxlen-minlen here. */
4564 /* last: Stop before this one. */
4565 /* data: string data about the pattern */
4566 /* stopparen: treat close N as END */
4567 /* recursed: which subroutines have we recursed into */
4568 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
4570 SSize_t final_minlen;
4571 /* There must be at least this number of characters to match */
4574 regnode *scan = *scanp, *next;
4576 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
4577 int is_inf_internal = 0; /* The studied chunk is infinite */
4578 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
4579 scan_data_t data_fake;
4580 SV *re_trie_maxbuff = NULL;
4581 regnode *first_non_open = scan;
4582 SSize_t stopmin = OPTIMIZE_INFTY;
4583 scan_frame *frame = NULL;
4584 DECLARE_AND_GET_RE_DEBUG_FLAGS;
4586 PERL_ARGS_ASSERT_STUDY_CHUNK;
4587 RExC_study_started= 1;
4589 Zero(&data_fake, 1, scan_data_t);
4592 while (first_non_open && OP(first_non_open) == OPEN)
4593 first_non_open=regnext(first_non_open);
4599 RExC_study_chunk_recursed_count++;
4601 DEBUG_OPTIMISE_MORE_r(
4603 Perl_re_indentf( aTHX_ "study_chunk stopparen=%ld recursed_count=%lu depth=%lu recursed_depth=%lu scan=%p last=%p",
4604 depth, (long)stopparen,
4605 (unsigned long)RExC_study_chunk_recursed_count,
4606 (unsigned long)depth, (unsigned long)recursed_depth,
4609 if (recursed_depth) {
4612 for ( j = 0 ; j < recursed_depth ; j++ ) {
4613 for ( i = 0 ; i < (U32)RExC_total_parens ; i++ ) {
4614 if (PAREN_TEST(j, i) && (!j || !PAREN_TEST(j - 1, i))) {
4615 Perl_re_printf( aTHX_ " %d",(int)i);
4619 if ( j + 1 < recursed_depth ) {
4620 Perl_re_printf( aTHX_ ",");
4624 Perl_re_printf( aTHX_ "\n");
4627 while ( scan && OP(scan) != END && scan < last ){
4628 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
4629 node length to get a real minimum (because
4630 the folded version may be shorter) */
4631 bool unfolded_multi_char = FALSE;
4632 /* avoid mutating ops if we are anywhere within the recursed or
4633 * enframed handling for a GOSUB: the outermost level will handle it.
4635 bool mutate_ok = was_mutate_ok && !(frame && frame->in_gosub);
4636 /* Peephole optimizer: */
4637 DEBUG_STUDYDATA("Peep", data, depth, is_inf);
4638 DEBUG_PEEP("Peep", scan, depth, flags);
4641 /* The reason we do this here is that we need to deal with things like
4642 * /(?:f)(?:o)(?:o)/ which cant be dealt with by the normal EXACT
4643 * parsing code, as each (?:..) is handled by a different invocation of
4646 if (PL_regkind[OP(scan)] == EXACT
4647 && OP(scan) != LEXACT
4648 && OP(scan) != LEXACT_REQ8
4651 join_exact(pRExC_state, scan, &min_subtract, &unfolded_multi_char,
4652 0, NULL, depth + 1);
4655 /* Follow the next-chain of the current node and optimize
4656 away all the NOTHINGs from it.
4658 rck_elide_nothing(scan);
4660 /* The principal pseudo-switch. Cannot be a switch, since we look into
4661 * several different things. */
4662 if ( OP(scan) == DEFINEP ) {
4664 SSize_t deltanext = 0;
4665 SSize_t fake_last_close = 0;
4666 I32 f = SCF_IN_DEFINE;
4668 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4669 scan = regnext(scan);
4670 assert( OP(scan) == IFTHEN );
4671 DEBUG_PEEP("expect IFTHEN", scan, depth, flags);
4673 data_fake.last_closep= &fake_last_close;
4675 next = regnext(scan);
4676 scan = NEXTOPER(NEXTOPER(scan));
4677 DEBUG_PEEP("scan", scan, depth, flags);
4678 DEBUG_PEEP("next", next, depth, flags);
4680 /* we suppose the run is continuous, last=next...
4681 * NOTE we dont use the return here! */
4682 /* DEFINEP study_chunk() recursion */
4683 (void)study_chunk(pRExC_state, &scan, &minlen,
4684 &deltanext, next, &data_fake, stopparen,
4685 recursed_depth, NULL, f, depth+1, mutate_ok);
4690 OP(scan) == BRANCH ||
4691 OP(scan) == BRANCHJ ||
4694 next = regnext(scan);
4697 /* The op(next)==code check below is to see if we
4698 * have "BRANCH-BRANCH", "BRANCHJ-BRANCHJ", "IFTHEN-IFTHEN"
4699 * IFTHEN is special as it might not appear in pairs.
4700 * Not sure whether BRANCH-BRANCHJ is possible, regardless
4701 * we dont handle it cleanly. */
4702 if (OP(next) == code || code == IFTHEN) {
4703 /* NOTE - There is similar code to this block below for
4704 * handling TRIE nodes on a re-study. If you change stuff here
4705 * check there too. */
4706 SSize_t max1 = 0, min1 = OPTIMIZE_INFTY, num = 0;
4708 regnode * const startbranch=scan;
4710 if (flags & SCF_DO_SUBSTR) {
4711 /* Cannot merge strings after this. */
4712 scan_commit(pRExC_state, data, minlenp, is_inf);
4715 if (flags & SCF_DO_STCLASS)
4716 ssc_init_zero(pRExC_state, &accum);
4718 while (OP(scan) == code) {
4719 SSize_t deltanext, minnext, fake;
4721 regnode_ssc this_class;
4723 DEBUG_PEEP("Branch", scan, depth, flags);
4726 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4728 data_fake.whilem_c = data->whilem_c;
4729 data_fake.last_closep = data->last_closep;
4732 data_fake.last_closep = &fake;
4734 data_fake.pos_delta = delta;
4735 next = regnext(scan);
4737 scan = NEXTOPER(scan); /* everything */
4738 if (code != BRANCH) /* everything but BRANCH */
4739 scan = NEXTOPER(scan);
4741 if (flags & SCF_DO_STCLASS) {
4742 ssc_init(pRExC_state, &this_class);
4743 data_fake.start_class = &this_class;
4744 f = SCF_DO_STCLASS_AND;
4746 if (flags & SCF_WHILEM_VISITED_POS)
4747 f |= SCF_WHILEM_VISITED_POS;
4749 /* we suppose the run is continuous, last=next...*/
4750 /* recurse study_chunk() for each BRANCH in an alternation */
4751 minnext = study_chunk(pRExC_state, &scan, minlenp,
4752 &deltanext, next, &data_fake, stopparen,
4753 recursed_depth, NULL, f, depth+1,
4758 if (deltanext == OPTIMIZE_INFTY) {
4759 is_inf = is_inf_internal = 1;
4760 max1 = OPTIMIZE_INFTY;
4761 } else if (max1 < minnext + deltanext)
4762 max1 = minnext + deltanext;
4764 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4766 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4767 if ( stopmin > minnext)
4768 stopmin = min + min1;
4769 flags &= ~SCF_DO_SUBSTR;
4771 data->flags |= SCF_SEEN_ACCEPT;
4774 if (data_fake.flags & SF_HAS_EVAL)
4775 data->flags |= SF_HAS_EVAL;
4776 data->whilem_c = data_fake.whilem_c;
4778 if (flags & SCF_DO_STCLASS)
4779 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
4781 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
4783 if (flags & SCF_DO_SUBSTR) {
4784 data->pos_min += min1;
4785 if (data->pos_delta >= OPTIMIZE_INFTY - (max1 - min1))
4786 data->pos_delta = OPTIMIZE_INFTY;
4788 data->pos_delta += max1 - min1;
4789 if (max1 != min1 || is_inf)
4790 data->cur_is_floating = 1;
4793 if (delta == OPTIMIZE_INFTY
4794 || OPTIMIZE_INFTY - delta - (max1 - min1) < 0)
4795 delta = OPTIMIZE_INFTY;
4797 delta += max1 - min1;
4798 if (flags & SCF_DO_STCLASS_OR) {
4799 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
4801 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4802 flags &= ~SCF_DO_STCLASS;
4805 else if (flags & SCF_DO_STCLASS_AND) {
4807 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
4808 flags &= ~SCF_DO_STCLASS;
4811 /* Switch to OR mode: cache the old value of
4812 * data->start_class */
4814 StructCopy(data->start_class, and_withp, regnode_ssc);
4815 flags &= ~SCF_DO_STCLASS_AND;
4816 StructCopy(&accum, data->start_class, regnode_ssc);
4817 flags |= SCF_DO_STCLASS_OR;
4821 if (PERL_ENABLE_TRIE_OPTIMISATION
4822 && OP(startbranch) == BRANCH
4827 Assuming this was/is a branch we are dealing with: 'scan'
4828 now points at the item that follows the branch sequence,
4829 whatever it is. We now start at the beginning of the
4830 sequence and look for subsequences of
4836 which would be constructed from a pattern like
4839 If we can find such a subsequence we need to turn the first
4840 element into a trie and then add the subsequent branch exact
4841 strings to the trie.
4845 1. patterns where the whole set of branches can be
4848 2. patterns where only a subset can be converted.
4850 In case 1 we can replace the whole set with a single regop
4851 for the trie. In case 2 we need to keep the start and end
4854 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
4855 becomes BRANCH TRIE; BRANCH X;
4857 There is an additional case, that being where there is a
4858 common prefix, which gets split out into an EXACT like node
4859 preceding the TRIE node.
4861 If x(1..n)==tail then we can do a simple trie, if not we make
4862 a "jump" trie, such that when we match the appropriate word
4863 we "jump" to the appropriate tail node. Essentially we turn
4864 a nested if into a case structure of sorts.
4869 if (!re_trie_maxbuff) {
4870 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
4871 if (!SvIOK(re_trie_maxbuff))
4872 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
4874 if ( SvIV(re_trie_maxbuff)>=0 ) {
4876 regnode *first = (regnode *)NULL;
4877 regnode *prev = (regnode *)NULL;
4878 regnode *tail = scan;
4882 /* var tail is used because there may be a TAIL
4883 regop in the way. Ie, the exacts will point to the
4884 thing following the TAIL, but the last branch will
4885 point at the TAIL. So we advance tail. If we
4886 have nested (?:) we may have to move through several
4890 while ( OP( tail ) == TAIL ) {
4891 /* this is the TAIL generated by (?:) */
4892 tail = regnext( tail );
4896 DEBUG_TRIE_COMPILE_r({
4897 regprop(RExC_rx, RExC_mysv, tail, NULL, pRExC_state);
4898 Perl_re_indentf( aTHX_ "%s %" UVuf ":%s\n",
4900 "Looking for TRIE'able sequences. Tail node is ",
4901 (UV) REGNODE_OFFSET(tail),
4902 SvPV_nolen_const( RExC_mysv )
4908 Step through the branches
4909 cur represents each branch,
4910 noper is the first thing to be matched as part
4912 noper_next is the regnext() of that node.
4914 We normally handle a case like this
4915 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
4916 support building with NOJUMPTRIE, which restricts
4917 the trie logic to structures like /FOO|BAR/.
4919 If noper is a trieable nodetype then the branch is
4920 a possible optimization target. If we are building
4921 under NOJUMPTRIE then we require that noper_next is
4922 the same as scan (our current position in the regex
4925 Once we have two or more consecutive such branches
4926 we can create a trie of the EXACT's contents and
4927 stitch it in place into the program.
4929 If the sequence represents all of the branches in
4930 the alternation we replace the entire thing with a
4933 Otherwise when it is a subsequence we need to
4934 stitch it in place and replace only the relevant
4935 branches. This means the first branch has to remain
4936 as it is used by the alternation logic, and its
4937 next pointer, and needs to be repointed at the item
4938 on the branch chain following the last branch we
4939 have optimized away.
4941 This could be either a BRANCH, in which case the
4942 subsequence is internal, or it could be the item
4943 following the branch sequence in which case the
4944 subsequence is at the end (which does not
4945 necessarily mean the first node is the start of the
4948 TRIE_TYPE(X) is a define which maps the optype to a
4952 ----------------+-----------
4957 EXACTFU_REQ8 | EXACTFU
4961 EXACTFLU8 | EXACTFLU8
4965 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) \
4967 : ( EXACT == (X) || EXACT_REQ8 == (X) ) \
4969 : ( EXACTFU == (X) \
4970 || EXACTFU_REQ8 == (X) \
4971 || EXACTFUP == (X) ) \
4973 : ( EXACTFAA == (X) ) \
4975 : ( EXACTL == (X) ) \
4977 : ( EXACTFLU8 == (X) ) \
4981 /* dont use tail as the end marker for this traverse */
4982 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
4983 regnode * const noper = NEXTOPER( cur );
4984 U8 noper_type = OP( noper );
4985 U8 noper_trietype = TRIE_TYPE( noper_type );
4986 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
4987 regnode * const noper_next = regnext( noper );
4988 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4989 U8 noper_next_trietype = (noper_next && noper_next < tail) ? TRIE_TYPE( noper_next_type ) :0;
4992 DEBUG_TRIE_COMPILE_r({
4993 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4994 Perl_re_indentf( aTHX_ "- %d:%s (%d)",
4996 REG_NODE_NUM(cur), SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur) );
4998 regprop(RExC_rx, RExC_mysv, noper, NULL, pRExC_state);
4999 Perl_re_printf( aTHX_ " -> %d:%s",
5000 REG_NODE_NUM(noper), SvPV_nolen_const(RExC_mysv));
5003 regprop(RExC_rx, RExC_mysv, noper_next, NULL, pRExC_state);
5004 Perl_re_printf( aTHX_ "\t=> %d:%s\t",
5005 REG_NODE_NUM(noper_next), SvPV_nolen_const(RExC_mysv));
5007 Perl_re_printf( aTHX_ "(First==%d,Last==%d,Cur==%d,tt==%s,ntt==%s,nntt==%s)\n",
5008 REG_NODE_NUM(first), REG_NODE_NUM(prev), REG_NODE_NUM(cur),
5009 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
5013 /* Is noper a trieable nodetype that can be merged
5014 * with the current trie (if there is one)? */
5018 ( noper_trietype == NOTHING )
5019 || ( trietype == NOTHING )
5020 || ( trietype == noper_trietype )
5023 && noper_next >= tail
5027 /* Handle mergable triable node Either we are
5028 * the first node in a new trieable sequence,
5029 * in which case we do some bookkeeping,
5030 * otherwise we update the end pointer. */
5033 if ( noper_trietype == NOTHING ) {
5034 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
5035 regnode * const noper_next = regnext( noper );
5036 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
5037 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
5040 if ( noper_next_trietype ) {
5041 trietype = noper_next_trietype;
5042 } else if (noper_next_type) {
5043 /* a NOTHING regop is 1 regop wide.
5044 * We need at least two for a trie
5045 * so we can't merge this in */
5049 trietype = noper_trietype;
5052 if ( trietype == NOTHING )
5053 trietype = noper_trietype;
5058 } /* end handle mergable triable node */
5060 /* handle unmergable node -
5061 * noper may either be a triable node which can
5062 * not be tried together with the current trie,
5063 * or a non triable node */
5065 /* If last is set and trietype is not
5066 * NOTHING then we have found at least two
5067 * triable branch sequences in a row of a
5068 * similar trietype so we can turn them
5069 * into a trie. If/when we allow NOTHING to
5070 * start a trie sequence this condition
5071 * will be required, and it isn't expensive
5072 * so we leave it in for now. */
5073 if ( trietype && trietype != NOTHING )
5074 make_trie( pRExC_state,
5075 startbranch, first, cur, tail,
5076 count, trietype, depth+1 );
5077 prev = NULL; /* note: we clear/update
5078 first, trietype etc below,
5079 so we dont do it here */
5083 && noper_next >= tail
5086 /* noper is triable, so we can start a new
5090 trietype = noper_trietype;
5092 /* if we already saw a first but the
5093 * current node is not triable then we have
5094 * to reset the first information. */
5099 } /* end handle unmergable node */
5100 } /* loop over branches */
5101 DEBUG_TRIE_COMPILE_r({
5102 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
5103 Perl_re_indentf( aTHX_ "- %s (%d) <SCAN FINISHED> ",
5104 depth+1, SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur));
5105 Perl_re_printf( aTHX_ "(First==%d, Last==%d, Cur==%d, tt==%s)\n",
5106 REG_NODE_NUM(first), REG_NODE_NUM(prev), REG_NODE_NUM(cur),
5107 PL_reg_name[trietype]
5111 if ( prev && trietype ) {
5112 if ( trietype != NOTHING ) {
5113 /* the last branch of the sequence was part of
5114 * a trie, so we have to construct it here
5115 * outside of the loop */
5116 made= make_trie( pRExC_state, startbranch,
5117 first, scan, tail, count,
5118 trietype, depth+1 );
5119 #ifdef TRIE_STUDY_OPT
5120 if ( ((made == MADE_EXACT_TRIE &&
5121 startbranch == first)
5122 || ( first_non_open == first )) &&
5124 flags |= SCF_TRIE_RESTUDY;
5125 if ( startbranch == first
5128 RExC_seen &=~REG_TOP_LEVEL_BRANCHES_SEEN;
5133 /* at this point we know whatever we have is a
5134 * NOTHING sequence/branch AND if 'startbranch'
5135 * is 'first' then we can turn the whole thing
5138 if ( startbranch == first ) {
5140 /* the entire thing is a NOTHING sequence,
5141 * something like this: (?:|) So we can
5142 * turn it into a plain NOTHING op. */
5143 DEBUG_TRIE_COMPILE_r({
5144 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
5145 Perl_re_indentf( aTHX_ "- %s (%d) <NOTHING BRANCH SEQUENCE>\n",
5147 SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur));
5150 OP(startbranch)= NOTHING;
5151 NEXT_OFF(startbranch)= tail - startbranch;
5152 for ( opt= startbranch + 1; opt < tail ; opt++ )
5156 } /* end if ( prev) */
5157 } /* TRIE_MAXBUF is non zero */
5161 else if ( code == BRANCHJ ) { /* single branch is optimized. */
5162 scan = NEXTOPER(NEXTOPER(scan));
5163 } else /* single branch is optimized. */
5164 scan = NEXTOPER(scan);
5166 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB) {
5168 regnode *start = NULL;
5169 regnode *end = NULL;
5170 U32 my_recursed_depth= recursed_depth;
5172 if (OP(scan) != SUSPEND) { /* GOSUB */
5173 /* Do setup, note this code has side effects beyond
5174 * the rest of this block. Specifically setting
5175 * RExC_recurse[] must happen at least once during
5178 RExC_recurse[ARG2L(scan)] = scan;
5179 start = REGNODE_p(RExC_open_parens[paren]);
5180 end = REGNODE_p(RExC_close_parens[paren]);
5182 /* NOTE we MUST always execute the above code, even
5183 * if we do nothing with a GOSUB */
5185 ( flags & SCF_IN_DEFINE )
5188 (is_inf_internal || is_inf || (data && data->flags & SF_IS_INF))
5190 ( (flags & (SCF_DO_STCLASS | SCF_DO_SUBSTR)) == 0 )
5193 /* no need to do anything here if we are in a define. */
5194 /* or we are after some kind of infinite construct
5195 * so we can skip recursing into this item.
5196 * Since it is infinite we will not change the maxlen
5197 * or delta, and if we miss something that might raise
5198 * the minlen it will merely pessimise a little.
5200 * Iow /(?(DEFINE)(?<foo>foo|food))a+(?&foo)/
5201 * might result in a minlen of 1 and not of 4,
5202 * but this doesn't make us mismatch, just try a bit
5203 * harder than we should.
5205 * However we must assume this GOSUB is infinite, to
5206 * avoid wrongly applying other optimizations in the
5207 * enclosing scope - see GH 18096, for example.
5209 is_inf = is_inf_internal = 1;
5210 scan= regnext(scan);
5216 || !PAREN_TEST(recursed_depth - 1, paren)
5218 /* it is quite possible that there are more efficient ways
5219 * to do this. We maintain a bitmap per level of recursion
5220 * of which patterns we have entered so we can detect if a
5221 * pattern creates a possible infinite loop. When we
5222 * recurse down a level we copy the previous levels bitmap
5223 * down. When we are at recursion level 0 we zero the top
5224 * level bitmap. It would be nice to implement a different
5225 * more efficient way of doing this. In particular the top
5226 * level bitmap may be unnecessary.
5228 if (!recursed_depth) {
5229 Zero(RExC_study_chunk_recursed, RExC_study_chunk_recursed_bytes, U8);
5231 Copy(PAREN_OFFSET(recursed_depth - 1),
5232 PAREN_OFFSET(recursed_depth),
5233 RExC_study_chunk_recursed_bytes, U8);
5235 /* we havent recursed into this paren yet, so recurse into it */
5236 DEBUG_STUDYDATA("gosub-set", data, depth, is_inf);
5237 PAREN_SET(recursed_depth, paren);
5238 my_recursed_depth= recursed_depth + 1;
5240 DEBUG_STUDYDATA("gosub-inf", data, depth, is_inf);
5241 /* some form of infinite recursion, assume infinite length
5243 if (flags & SCF_DO_SUBSTR) {
5244 scan_commit(pRExC_state, data, minlenp, is_inf);
5245 data->cur_is_floating = 1;
5247 is_inf = is_inf_internal = 1;
5248 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5249 ssc_anything(data->start_class);
5250 flags &= ~SCF_DO_STCLASS;
5252 start= NULL; /* reset start so we dont recurse later on. */
5257 end = regnext(scan);
5260 scan_frame *newframe;
5262 if (!RExC_frame_last) {
5263 Newxz(newframe, 1, scan_frame);
5264 SAVEDESTRUCTOR_X(S_unwind_scan_frames, newframe);
5265 RExC_frame_head= newframe;
5267 } else if (!RExC_frame_last->next_frame) {
5268 Newxz(newframe, 1, scan_frame);
5269 RExC_frame_last->next_frame= newframe;
5270 newframe->prev_frame= RExC_frame_last;
5273 newframe= RExC_frame_last->next_frame;
5275 RExC_frame_last= newframe;
5277 newframe->next_regnode = regnext(scan);
5278 newframe->last_regnode = last;
5279 newframe->stopparen = stopparen;
5280 newframe->prev_recursed_depth = recursed_depth;
5281 newframe->this_prev_frame= frame;
5282 newframe->in_gosub = (
5283 (frame && frame->in_gosub) || OP(scan) == GOSUB
5286 DEBUG_STUDYDATA("frame-new", data, depth, is_inf);
5287 DEBUG_PEEP("fnew", scan, depth, flags);
5294 recursed_depth= my_recursed_depth;
5299 else if (PL_regkind[OP(scan)] == EXACT && ! isEXACTFish(OP(scan))) {
5300 SSize_t bytelen = STR_LEN(scan), charlen;
5304 const U8 * const s = (U8*)STRING(scan);
5305 uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
5306 charlen = utf8_length(s, s + bytelen);
5308 uc = *((U8*)STRING(scan));
5312 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
5313 /* The code below prefers earlier match for fixed
5314 offset, later match for variable offset. */
5315 if (data->last_end == -1) { /* Update the start info. */
5316 data->last_start_min = data->pos_min;
5317 data->last_start_max =
5318 is_inf ? OPTIMIZE_INFTY
5319 : (data->pos_delta > OPTIMIZE_INFTY - data->pos_min)
5320 ? OPTIMIZE_INFTY : data->pos_min + data->pos_delta;
5322 sv_catpvn(data->last_found, STRING(scan), bytelen);
5324 SvUTF8_on(data->last_found);
5326 SV * const sv = data->last_found;
5327 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
5328 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5329 if (mg && mg->mg_len >= 0)
5330 mg->mg_len += charlen;
5332 data->last_end = data->pos_min + charlen;
5333 data->pos_min += charlen; /* As in the first entry. */
5334 data->flags &= ~SF_BEFORE_EOL;
5337 /* ANDing the code point leaves at most it, and not in locale, and
5338 * can't match null string */
5339 if (flags & SCF_DO_STCLASS_AND) {
5340 ssc_cp_and(data->start_class, uc);
5341 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5342 ssc_clear_locale(data->start_class);
5344 else if (flags & SCF_DO_STCLASS_OR) {
5345 ssc_add_cp(data->start_class, uc);
5346 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5348 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5349 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5351 flags &= ~SCF_DO_STCLASS;
5353 else if (PL_regkind[OP(scan)] == EXACT) {
5354 /* But OP != EXACT!, so is EXACTFish */
5355 SSize_t bytelen = STR_LEN(scan), charlen;
5356 const U8 * s = (U8*)STRING(scan);
5358 /* Replace a length 1 ASCII fold pair node with an ANYOFM node,
5359 * with the mask set to the complement of the bit that differs
5360 * between upper and lower case, and the lowest code point of the
5361 * pair (which the '&' forces) */
5364 && ( OP(scan) == EXACTFAA
5365 || ( OP(scan) == EXACTFU
5366 && ! HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(*s)))
5369 U8 mask = ~ ('A' ^ 'a'); /* These differ in just one bit */
5372 ARG_SET(scan, *s & mask);
5374 /* we're not EXACTFish any more, so restudy */
5378 /* Search for fixed substrings supports EXACT only. */
5379 if (flags & SCF_DO_SUBSTR) {
5381 scan_commit(pRExC_state, data, minlenp, is_inf);
5383 charlen = UTF ? (SSize_t) utf8_length(s, s + bytelen) : bytelen;
5384 if (unfolded_multi_char) {
5385 RExC_seen |= REG_UNFOLDED_MULTI_SEEN;
5387 min += charlen - min_subtract;
5389 if ((SSize_t)min_subtract < OPTIMIZE_INFTY
5390 && delta < OPTIMIZE_INFTY - (SSize_t)min_subtract
5392 delta += min_subtract;
5394 delta = OPTIMIZE_INFTY;
5396 if (flags & SCF_DO_SUBSTR) {
5397 data->pos_min += charlen - min_subtract;
5398 if (data->pos_min < 0) {
5401 if ((SSize_t)min_subtract < OPTIMIZE_INFTY
5402 && data->pos_delta < OPTIMIZE_INFTY - (SSize_t)min_subtract
5404 data->pos_delta += min_subtract;
5406 data->pos_delta = OPTIMIZE_INFTY;
5409 data->cur_is_floating = 1; /* float */
5413 if (flags & SCF_DO_STCLASS) {
5414 SV* EXACTF_invlist = make_exactf_invlist(pRExC_state, scan);
5416 assert(EXACTF_invlist);
5417 if (flags & SCF_DO_STCLASS_AND) {
5418 if (OP(scan) != EXACTFL)
5419 ssc_clear_locale(data->start_class);
5420 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5421 ANYOF_POSIXL_ZERO(data->start_class);
5422 ssc_intersection(data->start_class, EXACTF_invlist, FALSE);
5424 else { /* SCF_DO_STCLASS_OR */
5425 ssc_union(data->start_class, EXACTF_invlist, FALSE);
5426 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5428 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5429 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5431 flags &= ~SCF_DO_STCLASS;
5432 SvREFCNT_dec(EXACTF_invlist);
5435 else if (REGNODE_VARIES(OP(scan))) {
5436 SSize_t mincount, maxcount, minnext, deltanext, pos_before = 0;
5437 I32 fl = 0, f = flags;
5438 regnode * const oscan = scan;
5439 regnode_ssc this_class;
5440 regnode_ssc *oclass = NULL;
5441 I32 next_is_eval = 0;
5443 switch (PL_regkind[OP(scan)]) {
5444 case WHILEM: /* End of (?:...)* . */
5445 scan = NEXTOPER(scan);
5448 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
5449 next = NEXTOPER(scan);
5450 if ( ( PL_regkind[OP(next)] == EXACT
5451 && ! isEXACTFish(OP(next)))
5452 || (flags & SCF_DO_STCLASS))
5455 maxcount = REG_INFTY;
5456 next = regnext(scan);
5457 scan = NEXTOPER(scan);
5461 if (flags & SCF_DO_SUBSTR)
5463 /* This will bypass the formal 'min += minnext * mincount'
5464 * calculation in the do_curly path, so assumes min width
5465 * of the PLUS payload is exactly one. */
5469 next = NEXTOPER(scan);
5471 /* This temporary node can now be turned into EXACTFU, and
5472 * must, as regexec.c doesn't handle it */
5473 if (OP(next) == EXACTFU_S_EDGE && mutate_ok) {
5477 if ( STR_LEN(next) == 1
5478 && isALPHA_A(* STRING(next))
5479 && ( OP(next) == EXACTFAA
5480 || ( OP(next) == EXACTFU
5481 && ! HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(* STRING(next))))
5484 /* These differ in just one bit */
5485 U8 mask = ~ ('A' ^ 'a');
5487 assert(isALPHA_A(* STRING(next)));
5489 /* Then replace it by an ANYOFM node, with
5490 * the mask set to the complement of the
5491 * bit that differs between upper and lower
5492 * case, and the lowest code point of the
5493 * pair (which the '&' forces) */
5495 ARG_SET(next, *STRING(next) & mask);
5499 if (flags & SCF_DO_STCLASS) {
5501 maxcount = REG_INFTY;
5502 next = regnext(scan);
5503 scan = NEXTOPER(scan);
5506 if (flags & SCF_DO_SUBSTR) {
5507 scan_commit(pRExC_state, data, minlenp, is_inf);
5508 /* Cannot extend fixed substrings */
5509 data->cur_is_floating = 1; /* float */
5511 is_inf = is_inf_internal = 1;
5512 scan = regnext(scan);
5513 goto optimize_curly_tail;
5515 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
5516 && (scan->flags == stopparen))
5521 mincount = ARG1(scan);
5522 maxcount = ARG2(scan);
5524 next = regnext(scan);
5525 if (OP(scan) == CURLYX) {
5526 I32 lp = (data ? *(data->last_closep) : 0);
5527 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
5529 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
5530 next_is_eval = (OP(scan) == EVAL);
5532 if (flags & SCF_DO_SUBSTR) {
5534 scan_commit(pRExC_state, data, minlenp, is_inf);
5535 /* Cannot extend fixed substrings */
5536 pos_before = data->pos_min;
5540 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
5542 data->flags |= SF_IS_INF;
5544 if (flags & SCF_DO_STCLASS) {
5545 ssc_init(pRExC_state, &this_class);
5546 oclass = data->start_class;
5547 data->start_class = &this_class;
5548 f |= SCF_DO_STCLASS_AND;
5549 f &= ~SCF_DO_STCLASS_OR;
5551 /* Exclude from super-linear cache processing any {n,m}
5552 regops for which the combination of input pos and regex
5553 pos is not enough information to determine if a match
5556 For example, in the regex /foo(bar\s*){4,8}baz/ with the
5557 regex pos at the \s*, the prospects for a match depend not
5558 only on the input position but also on how many (bar\s*)
5559 repeats into the {4,8} we are. */
5560 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
5561 f &= ~SCF_WHILEM_VISITED_POS;
5563 /* This will finish on WHILEM, setting scan, or on NULL: */
5564 /* recurse study_chunk() on loop bodies */
5565 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
5566 last, data, stopparen, recursed_depth, NULL,
5568 ? (f & ~SCF_DO_SUBSTR)
5570 , depth+1, mutate_ok);
5572 if (flags & SCF_DO_STCLASS)
5573 data->start_class = oclass;
5574 if (mincount == 0 || minnext == 0) {
5575 if (flags & SCF_DO_STCLASS_OR) {
5576 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5578 else if (flags & SCF_DO_STCLASS_AND) {
5579 /* Switch to OR mode: cache the old value of
5580 * data->start_class */
5582 StructCopy(data->start_class, and_withp, regnode_ssc);
5583 flags &= ~SCF_DO_STCLASS_AND;
5584 StructCopy(&this_class, data->start_class, regnode_ssc);
5585 flags |= SCF_DO_STCLASS_OR;
5586 ANYOF_FLAGS(data->start_class)
5587 |= SSC_MATCHES_EMPTY_STRING;
5589 } else { /* Non-zero len */
5590 if (flags & SCF_DO_STCLASS_OR) {
5591 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5592 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5594 else if (flags & SCF_DO_STCLASS_AND)
5595 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5596 flags &= ~SCF_DO_STCLASS;
5598 if (!scan) /* It was not CURLYX, but CURLY. */
5600 if (((flags & (SCF_TRIE_DOING_RESTUDY|SCF_DO_SUBSTR))==SCF_DO_SUBSTR)
5601 /* ? quantifier ok, except for (?{ ... }) */
5602 && (next_is_eval || !(mincount == 0 && maxcount == 1))
5603 && (minnext == 0) && (deltanext == 0)
5604 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
5605 && maxcount <= REG_INFTY/3) /* Complement check for big
5608 _WARN_HELPER(RExC_precomp_end, packWARN(WARN_REGEXP),
5609 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP),
5610 "Quantifier unexpected on zero-length expression "
5611 "in regex m/%" UTF8f "/",
5612 UTF8fARG(UTF, RExC_precomp_end - RExC_precomp,
5616 if ( ( minnext > 0 && mincount >= SSize_t_MAX / minnext )
5617 || min >= SSize_t_MAX - minnext * mincount )
5619 FAIL("Regexp out of space");
5622 min += minnext * mincount;
5623 is_inf_internal |= deltanext == OPTIMIZE_INFTY
5624 || (maxcount == REG_INFTY && minnext + deltanext > 0);
5625 is_inf |= is_inf_internal;
5627 delta = OPTIMIZE_INFTY;
5629 delta += (minnext + deltanext) * maxcount
5630 - minnext * mincount;
5632 /* Try powerful optimization CURLYX => CURLYN. */
5633 if ( OP(oscan) == CURLYX && data
5634 && data->flags & SF_IN_PAR
5635 && !(data->flags & SF_HAS_EVAL)
5636 && !deltanext && minnext == 1
5639 /* Try to optimize to CURLYN. */
5640 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
5641 regnode * const nxt1 = nxt;
5648 if (!REGNODE_SIMPLE(OP(nxt))
5649 && !(PL_regkind[OP(nxt)] == EXACT
5650 && STR_LEN(nxt) == 1))
5656 if (OP(nxt) != CLOSE)
5658 if (RExC_open_parens) {
5661 RExC_open_parens[ARG(nxt1)] = REGNODE_OFFSET(oscan);
5664 RExC_close_parens[ARG(nxt1)] = REGNODE_OFFSET(nxt) + 2;
5666 /* Now we know that nxt2 is the only contents: */
5667 oscan->flags = (U8)ARG(nxt);
5669 OP(nxt1) = NOTHING; /* was OPEN. */
5672 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5673 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
5674 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
5675 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5676 OP(nxt + 1) = OPTIMIZED; /* was count. */
5677 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
5682 /* Try optimization CURLYX => CURLYM. */
5683 if ( OP(oscan) == CURLYX && data
5684 && !(data->flags & SF_HAS_PAR)
5685 && !(data->flags & SF_HAS_EVAL)
5686 && !deltanext /* atom is fixed width */
5687 && minnext != 0 /* CURLYM can't handle zero width */
5688 /* Nor characters whose fold at run-time may be
5689 * multi-character */
5690 && ! (RExC_seen & REG_UNFOLDED_MULTI_SEEN)
5693 /* XXXX How to optimize if data == 0? */
5694 /* Optimize to a simpler form. */
5695 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
5699 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
5700 && (OP(nxt2) != WHILEM))
5702 OP(nxt2) = SUCCEED; /* Whas WHILEM */
5703 /* Need to optimize away parenths. */
5704 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
5705 /* Set the parenth number. */
5706 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
5708 oscan->flags = (U8)ARG(nxt);
5709 if (RExC_open_parens) {
5711 RExC_open_parens[ARG(nxt1)] = REGNODE_OFFSET(oscan);
5714 RExC_close_parens[ARG(nxt1)] = REGNODE_OFFSET(nxt2)
5717 OP(nxt1) = OPTIMIZED; /* was OPEN. */
5718 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5721 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5722 OP(nxt + 1) = OPTIMIZED; /* was count. */
5723 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
5724 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
5727 while ( nxt1 && (OP(nxt1) != WHILEM)) {
5728 regnode *nnxt = regnext(nxt1);
5730 if (reg_off_by_arg[OP(nxt1)])
5731 ARG_SET(nxt1, nxt2 - nxt1);
5732 else if (nxt2 - nxt1 < U16_MAX)
5733 NEXT_OFF(nxt1) = nxt2 - nxt1;
5735 OP(nxt) = NOTHING; /* Cannot beautify */
5740 /* Optimize again: */
5741 /* recurse study_chunk() on optimised CURLYX => CURLYM */
5742 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
5743 NULL, stopparen, recursed_depth, NULL, 0,
5744 depth+1, mutate_ok);
5749 else if ((OP(oscan) == CURLYX)
5750 && (flags & SCF_WHILEM_VISITED_POS)
5751 /* See the comment on a similar expression above.
5752 However, this time it's not a subexpression
5753 we care about, but the expression itself. */
5754 && (maxcount == REG_INFTY)
5756 /* This stays as CURLYX, we can put the count/of pair. */
5757 /* Find WHILEM (as in regexec.c) */
5758 regnode *nxt = oscan + NEXT_OFF(oscan);
5760 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
5762 nxt = PREVOPER(nxt);
5763 if (nxt->flags & 0xf) {
5764 /* we've already set whilem count on this node */
5765 } else if (++data->whilem_c < 16) {
5766 assert(data->whilem_c <= RExC_whilem_seen);
5767 nxt->flags = (U8)(data->whilem_c
5768 | (RExC_whilem_seen << 4)); /* On WHILEM */
5771 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
5773 if (flags & SCF_DO_SUBSTR) {
5774 SV *last_str = NULL;
5775 STRLEN last_chrs = 0;
5776 int counted = mincount != 0;
5778 if (data->last_end > 0 && mincount != 0) { /* Ends with a
5780 SSize_t b = pos_before >= data->last_start_min
5781 ? pos_before : data->last_start_min;
5783 const char * const s = SvPV_const(data->last_found, l);
5784 SSize_t old = b - data->last_start_min;
5788 old = utf8_hop_forward((U8*)s, old,
5789 (U8 *) SvEND(data->last_found))
5792 /* Get the added string: */
5793 last_str = newSVpvn_utf8(s + old, l, UTF);
5794 last_chrs = UTF ? utf8_length((U8*)(s + old),
5795 (U8*)(s + old + l)) : l;
5796 if (deltanext == 0 && pos_before == b) {
5797 /* What was added is a constant string */
5800 SvGROW(last_str, (mincount * l) + 1);
5801 repeatcpy(SvPVX(last_str) + l,
5802 SvPVX_const(last_str), l,
5804 SvCUR_set(last_str, SvCUR(last_str) * mincount);
5805 /* Add additional parts. */
5806 SvCUR_set(data->last_found,
5807 SvCUR(data->last_found) - l);
5808 sv_catsv(data->last_found, last_str);
5810 SV * sv = data->last_found;
5812 SvUTF8(sv) && SvMAGICAL(sv) ?
5813 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5814 if (mg && mg->mg_len >= 0)
5815 mg->mg_len += last_chrs * (mincount-1);
5817 last_chrs *= mincount;
5818 data->last_end += l * (mincount - 1);
5821 /* start offset must point into the last copy */
5822 data->last_start_min += minnext * (mincount - 1);
5823 data->last_start_max =
5826 : data->last_start_max +
5827 (maxcount - 1) * (minnext + data->pos_delta);
5830 /* It is counted once already... */
5831 data->pos_min += minnext * (mincount - counted);
5833 Perl_re_printf( aTHX_ "counted=%" UVuf " deltanext=%" UVuf
5834 " OPTIMIZE_INFTY=%" UVuf " minnext=%" UVuf
5835 " maxcount=%" UVuf " mincount=%" UVuf
5836 " data->pos_delta=%" UVuf "\n",
5837 (UV)counted, (UV)deltanext, (UV)OPTIMIZE_INFTY, (UV)minnext, (UV)maxcount,
5838 (UV)mincount, (UV)data->pos_delta);
5839 if (deltanext != OPTIMIZE_INFTY)
5840 Perl_re_printf( aTHX_ "LHS=%" UVuf " RHS=%" UVuf "\n",
5841 (UV)(-counted * deltanext + (minnext + deltanext) * maxcount
5842 - minnext * mincount), (UV)(OPTIMIZE_INFTY - data->pos_delta));
5844 if (deltanext == OPTIMIZE_INFTY
5845 || data->pos_delta == OPTIMIZE_INFTY
5846 || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= OPTIMIZE_INFTY - data->pos_delta)
5847 data->pos_delta = OPTIMIZE_INFTY;
5849 data->pos_delta += - counted * deltanext +
5850 (minnext + deltanext) * maxcount - minnext * mincount;
5851 if (mincount != maxcount) {
5852 /* Cannot extend fixed substrings found inside
5854 scan_commit(pRExC_state, data, minlenp, is_inf);
5855 if (mincount && last_str) {
5856 SV * const sv = data->last_found;
5857 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
5858 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5862 sv_setsv(sv, last_str);
5863 data->last_end = data->pos_min;
5864 data->last_start_min = data->pos_min - last_chrs;
5865 data->last_start_max = is_inf
5867 : data->pos_min + data->pos_delta - last_chrs;
5869 data->cur_is_floating = 1; /* float */
5871 SvREFCNT_dec(last_str);
5873 if (data && (fl & SF_HAS_EVAL))
5874 data->flags |= SF_HAS_EVAL;
5875 optimize_curly_tail:
5876 rck_elide_nothing(oscan);
5880 Perl_croak(aTHX_ "panic: unexpected varying REx opcode %d",
5884 if (flags & SCF_DO_SUBSTR) {
5885 /* Cannot expect anything... */
5886 scan_commit(pRExC_state, data, minlenp, is_inf);
5887 data->cur_is_floating = 1; /* float */
5889 is_inf = is_inf_internal = 1;
5890 if (flags & SCF_DO_STCLASS_OR) {
5891 if (OP(scan) == CLUMP) {
5892 /* Actually is any start char, but very few code points
5893 * aren't start characters */
5894 ssc_match_all_cp(data->start_class);
5897 ssc_anything(data->start_class);
5900 flags &= ~SCF_DO_STCLASS;
5904 else if (OP(scan) == LNBREAK) {
5905 if (flags & SCF_DO_STCLASS) {
5906 if (flags & SCF_DO_STCLASS_AND) {
5907 ssc_intersection(data->start_class,
5908 PL_XPosix_ptrs[_CC_VERTSPACE], FALSE);
5909 ssc_clear_locale(data->start_class);
5910 ANYOF_FLAGS(data->start_class)
5911 &= ~SSC_MATCHES_EMPTY_STRING;
5913 else if (flags & SCF_DO_STCLASS_OR) {
5914 ssc_union(data->start_class,
5915 PL_XPosix_ptrs[_CC_VERTSPACE],
5917 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5919 /* See commit msg for
5920 * 749e076fceedeb708a624933726e7989f2302f6a */
5921 ANYOF_FLAGS(data->start_class)
5922 &= ~SSC_MATCHES_EMPTY_STRING;
5924 flags &= ~SCF_DO_STCLASS;
5927 if (delta != OPTIMIZE_INFTY)
5928 delta++; /* Because of the 2 char string cr-lf */
5929 if (flags & SCF_DO_SUBSTR) {
5930 /* Cannot expect anything... */
5931 scan_commit(pRExC_state, data, minlenp, is_inf);
5933 if (data->pos_delta != OPTIMIZE_INFTY) {
5934 data->pos_delta += 1;
5936 data->cur_is_floating = 1; /* float */
5939 else if (REGNODE_SIMPLE(OP(scan))) {
5941 if (flags & SCF_DO_SUBSTR) {
5942 scan_commit(pRExC_state, data, minlenp, is_inf);
5946 if (flags & SCF_DO_STCLASS) {
5948 SV* my_invlist = NULL;
5951 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5952 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5954 /* Some of the logic below assumes that switching
5955 locale on will only add false positives. */
5960 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d",
5964 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5965 ssc_match_all_cp(data->start_class);
5970 SV* REG_ANY_invlist = _new_invlist(2);
5971 REG_ANY_invlist = add_cp_to_invlist(REG_ANY_invlist,
5973 if (flags & SCF_DO_STCLASS_OR) {
5974 ssc_union(data->start_class,
5976 TRUE /* TRUE => invert, hence all but \n
5980 else if (flags & SCF_DO_STCLASS_AND) {
5981 ssc_intersection(data->start_class,
5983 TRUE /* TRUE => invert */
5985 ssc_clear_locale(data->start_class);
5987 SvREFCNT_dec_NN(REG_ANY_invlist);
5999 if (flags & SCF_DO_STCLASS_AND)
6000 ssc_and(pRExC_state, data->start_class,
6001 (regnode_charclass *) scan);
6003 ssc_or(pRExC_state, data->start_class,
6004 (regnode_charclass *) scan);
6007 case NANYOFM: /* NANYOFM already contains the inversion of the
6008 input ANYOF data, so, unlike things like
6009 NPOSIXA, don't change 'invert' to TRUE */
6013 SV* cp_list = get_ANYOFM_contents(scan);
6015 if (flags & SCF_DO_STCLASS_OR) {
6016 ssc_union(data->start_class, cp_list, invert);
6018 else if (flags & SCF_DO_STCLASS_AND) {
6019 ssc_intersection(data->start_class, cp_list, invert);
6022 SvREFCNT_dec_NN(cp_list);
6031 cp_list = _add_range_to_invlist(cp_list,
6033 ANYOFRbase(scan) + ANYOFRdelta(scan));
6035 if (flags & SCF_DO_STCLASS_OR) {
6036 ssc_union(data->start_class, cp_list, invert);
6038 else if (flags & SCF_DO_STCLASS_AND) {
6039 ssc_intersection(data->start_class, cp_list, invert);
6042 SvREFCNT_dec_NN(cp_list);
6051 namedclass = classnum_to_namedclass(FLAGS(scan)) + invert;
6052 if (flags & SCF_DO_STCLASS_AND) {
6053 bool was_there = cBOOL(
6054 ANYOF_POSIXL_TEST(data->start_class,
6056 ANYOF_POSIXL_ZERO(data->start_class);
6057 if (was_there) { /* Do an AND */
6058 ANYOF_POSIXL_SET(data->start_class, namedclass);
6060 /* No individual code points can now match */
6061 data->start_class->invlist
6062 = sv_2mortal(_new_invlist(0));
6065 int complement = namedclass + ((invert) ? -1 : 1);
6067 assert(flags & SCF_DO_STCLASS_OR);
6069 /* If the complement of this class was already there,
6070 * the result is that they match all code points,
6071 * (\d + \D == everything). Remove the classes from
6072 * future consideration. Locale is not relevant in
6074 if (ANYOF_POSIXL_TEST(data->start_class, complement)) {
6075 ssc_match_all_cp(data->start_class);
6076 ANYOF_POSIXL_CLEAR(data->start_class, namedclass);
6077 ANYOF_POSIXL_CLEAR(data->start_class, complement);
6079 else { /* The usual case; just add this class to the
6081 ANYOF_POSIXL_SET(data->start_class, namedclass);
6086 case NPOSIXA: /* For these, we always know the exact set of
6091 my_invlist = invlist_clone(PL_Posix_ptrs[FLAGS(scan)], NULL);
6092 goto join_posix_and_ascii;
6100 my_invlist = invlist_clone(PL_XPosix_ptrs[FLAGS(scan)], NULL);
6102 /* NPOSIXD matches all upper Latin1 code points unless the
6103 * target string being matched is UTF-8, which is
6104 * unknowable until match time. Since we are going to
6105 * invert, we want to get rid of all of them so that the
6106 * inversion will match all */
6107 if (OP(scan) == NPOSIXD) {
6108 _invlist_subtract(my_invlist, PL_UpperLatin1,
6112 join_posix_and_ascii:
6114 if (flags & SCF_DO_STCLASS_AND) {
6115 ssc_intersection(data->start_class, my_invlist, invert);
6116 ssc_clear_locale(data->start_class);
6119 assert(flags & SCF_DO_STCLASS_OR);
6120 ssc_union(data->start_class, my_invlist, invert);
6122 SvREFCNT_dec(my_invlist);
6124 if (flags & SCF_DO_STCLASS_OR)
6125 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6126 flags &= ~SCF_DO_STCLASS;
6129 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
6130 data->flags |= (OP(scan) == MEOL
6133 scan_commit(pRExC_state, data, minlenp, is_inf);
6136 else if ( PL_regkind[OP(scan)] == BRANCHJ
6137 /* Lookbehind, or need to calculate parens/evals/stclass: */
6138 && (scan->flags || data || (flags & SCF_DO_STCLASS))
6139 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM))
6141 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
6142 || OP(scan) == UNLESSM )
6144 /* Negative Lookahead/lookbehind
6145 In this case we can't do fixed string optimisation.
6148 SSize_t deltanext, minnext, fake = 0;
6153 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
6155 data_fake.whilem_c = data->whilem_c;
6156 data_fake.last_closep = data->last_closep;
6159 data_fake.last_closep = &fake;
6160 data_fake.pos_delta = delta;
6161 if ( flags & SCF_DO_STCLASS && !scan->flags
6162 && OP(scan) == IFMATCH ) { /* Lookahead */
6163 ssc_init(pRExC_state, &intrnl);
6164 data_fake.start_class = &intrnl;
6165 f |= SCF_DO_STCLASS_AND;
6167 if (flags & SCF_WHILEM_VISITED_POS)
6168 f |= SCF_WHILEM_VISITED_POS;
6169 next = regnext(scan);
6170 nscan = NEXTOPER(NEXTOPER(scan));
6172 /* recurse study_chunk() for lookahead body */
6173 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
6174 last, &data_fake, stopparen,
6175 recursed_depth, NULL, f, depth+1,
6179 || deltanext > (I32) U8_MAX
6180 || minnext > (I32)U8_MAX
6181 || minnext + deltanext > (I32)U8_MAX)
6183 FAIL2("Lookbehind longer than %" UVuf " not implemented",
6187 /* The 'next_off' field has been repurposed to count the
6188 * additional starting positions to try beyond the initial
6189 * one. (This leaves it at 0 for non-variable length
6190 * matches to avoid breakage for those not using this
6193 scan->next_off = deltanext;
6194 ckWARNexperimental(RExC_parse,
6195 WARN_EXPERIMENTAL__VLB,
6196 "Variable length lookbehind is experimental");
6198 scan->flags = (U8)minnext + deltanext;
6201 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
6203 if (data_fake.flags & SF_HAS_EVAL)
6204 data->flags |= SF_HAS_EVAL;
6205 data->whilem_c = data_fake.whilem_c;
6207 if (f & SCF_DO_STCLASS_AND) {
6208 if (flags & SCF_DO_STCLASS_OR) {
6209 /* OR before, AND after: ideally we would recurse with
6210 * data_fake to get the AND applied by study of the
6211 * remainder of the pattern, and then derecurse;
6212 * *** HACK *** for now just treat as "no information".
6213 * See [perl #56690].
6215 ssc_init(pRExC_state, data->start_class);
6217 /* AND before and after: combine and continue. These
6218 * assertions are zero-length, so can match an EMPTY
6220 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
6221 ANYOF_FLAGS(data->start_class)
6222 |= SSC_MATCHES_EMPTY_STRING;
6226 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
6228 /* Positive Lookahead/lookbehind
6229 In this case we can do fixed string optimisation,
6230 but we must be careful about it. Note in the case of
6231 lookbehind the positions will be offset by the minimum
6232 length of the pattern, something we won't know about
6233 until after the recurse.
6235 SSize_t deltanext, fake = 0;
6239 /* We use SAVEFREEPV so that when the full compile
6240 is finished perl will clean up the allocated
6241 minlens when it's all done. This way we don't
6242 have to worry about freeing them when we know
6243 they wont be used, which would be a pain.
6246 Newx( minnextp, 1, SSize_t );
6247 SAVEFREEPV(minnextp);
6250 StructCopy(data, &data_fake, scan_data_t);
6251 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
6254 scan_commit(pRExC_state, &data_fake, minlenp, is_inf);
6255 data_fake.last_found=newSVsv(data->last_found);
6259 data_fake.last_closep = &fake;
6260 data_fake.flags = 0;
6261 data_fake.substrs[0].flags = 0;
6262 data_fake.substrs[1].flags = 0;
6263 data_fake.pos_delta = delta;
6265 data_fake.flags |= SF_IS_INF;
6266 if ( flags & SCF_DO_STCLASS && !scan->flags
6267 && OP(scan) == IFMATCH ) { /* Lookahead */
6268 ssc_init(pRExC_state, &intrnl);
6269 data_fake.start_class = &intrnl;
6270 f |= SCF_DO_STCLASS_AND;
6272 if (flags & SCF_WHILEM_VISITED_POS)
6273 f |= SCF_WHILEM_VISITED_POS;
6274 next = regnext(scan);
6275 nscan = NEXTOPER(NEXTOPER(scan));
6277 /* positive lookahead study_chunk() recursion */
6278 *minnextp = study_chunk(pRExC_state, &nscan, minnextp,
6279 &deltanext, last, &data_fake,
6280 stopparen, recursed_depth, NULL,
6281 f, depth+1, mutate_ok);
6283 assert(0); /* This code has never been tested since this
6284 is normally not compiled */
6286 || deltanext > (I32) U8_MAX
6287 || *minnextp > (I32)U8_MAX
6288 || *minnextp + deltanext > (I32)U8_MAX)
6290 FAIL2("Lookbehind longer than %" UVuf " not implemented",
6295 scan->next_off = deltanext;
6297 scan->flags = (U8)*minnextp + deltanext;
6302 if (f & SCF_DO_STCLASS_AND) {
6303 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
6304 ANYOF_FLAGS(data->start_class) |= SSC_MATCHES_EMPTY_STRING;
6307 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
6309 if (data_fake.flags & SF_HAS_EVAL)
6310 data->flags |= SF_HAS_EVAL;
6311 data->whilem_c = data_fake.whilem_c;
6312 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
6314 if (RExC_rx->minlen<*minnextp)
6315 RExC_rx->minlen=*minnextp;
6316 scan_commit(pRExC_state, &data_fake, minnextp, is_inf);
6317 SvREFCNT_dec_NN(data_fake.last_found);
6319 for (i = 0; i < 2; i++) {
6320 if (data_fake.substrs[i].minlenp != minlenp) {
6321 data->substrs[i].min_offset =
6322 data_fake.substrs[i].min_offset;
6323 data->substrs[i].max_offset =
6324 data_fake.substrs[i].max_offset;
6325 data->substrs[i].minlenp =
6326 data_fake.substrs[i].minlenp;
6327 data->substrs[i].lookbehind += scan->flags;
6335 else if (OP(scan) == OPEN) {
6336 if (stopparen != (I32)ARG(scan))
6339 else if (OP(scan) == CLOSE) {
6340 if (stopparen == (I32)ARG(scan)) {
6343 if ((I32)ARG(scan) == is_par) {
6344 next = regnext(scan);
6346 if ( next && (OP(next) != WHILEM) && next < last)
6347 is_par = 0; /* Disable optimization */
6350 *(data->last_closep) = ARG(scan);
6352 else if (OP(scan) == EVAL) {
6354 data->flags |= SF_HAS_EVAL;
6356 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
6357 if (flags & SCF_DO_SUBSTR) {
6358 scan_commit(pRExC_state, data, minlenp, is_inf);
6359 flags &= ~SCF_DO_SUBSTR;
6361 if (data && OP(scan)==ACCEPT) {
6362 data->flags |= SCF_SEEN_ACCEPT;
6367 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
6369 if (flags & SCF_DO_SUBSTR) {
6370 scan_commit(pRExC_state, data, minlenp, is_inf);
6371 data->cur_is_floating = 1; /* float */
6373 is_inf = is_inf_internal = 1;
6374 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
6375 ssc_anything(data->start_class);
6376 flags &= ~SCF_DO_STCLASS;
6378 else if (OP(scan) == GPOS) {
6379 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
6380 !(delta || is_inf || (data && data->pos_delta)))
6382 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
6383 RExC_rx->intflags |= PREGf_ANCH_GPOS;
6384 if (RExC_rx->gofs < (STRLEN)min)
6385 RExC_rx->gofs = min;
6387 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
6391 #ifdef TRIE_STUDY_OPT
6392 #ifdef FULL_TRIE_STUDY
6393 else if (PL_regkind[OP(scan)] == TRIE) {
6394 /* NOTE - There is similar code to this block above for handling
6395 BRANCH nodes on the initial study. If you change stuff here
6397 regnode *trie_node= scan;
6398 regnode *tail= regnext(scan);
6399 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
6400 SSize_t max1 = 0, min1 = OPTIMIZE_INFTY;
6403 if (flags & SCF_DO_SUBSTR) { /* XXXX Add !SUSPEND? */
6404 /* Cannot merge strings after this. */
6405 scan_commit(pRExC_state, data, minlenp, is_inf);
6407 if (flags & SCF_DO_STCLASS)
6408 ssc_init_zero(pRExC_state, &accum);
6414 const regnode *nextbranch= NULL;
6417 for ( word=1 ; word <= trie->wordcount ; word++)
6419 SSize_t deltanext=0, minnext=0, f = 0, fake;
6420 regnode_ssc this_class;
6422 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
6424 data_fake.whilem_c = data->whilem_c;
6425 data_fake.last_closep = data->last_closep;
6428 data_fake.last_closep = &fake;
6429 data_fake.pos_delta = delta;
6430 if (flags & SCF_DO_STCLASS) {
6431 ssc_init(pRExC_state, &this_class);
6432 data_fake.start_class = &this_class;
6433 f = SCF_DO_STCLASS_AND;
6435 if (flags & SCF_WHILEM_VISITED_POS)
6436 f |= SCF_WHILEM_VISITED_POS;
6438 if (trie->jump[word]) {
6440 nextbranch = trie_node + trie->jump[0];
6441 scan= trie_node + trie->jump[word];
6442 /* We go from the jump point to the branch that follows
6443 it. Note this means we need the vestigal unused
6444 branches even though they arent otherwise used. */
6445 /* optimise study_chunk() for TRIE */
6446 minnext = study_chunk(pRExC_state, &scan, minlenp,
6447 &deltanext, (regnode *)nextbranch, &data_fake,
6448 stopparen, recursed_depth, NULL, f, depth+1,
6451 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
6452 nextbranch= regnext((regnode*)nextbranch);
6454 if (min1 > (SSize_t)(minnext + trie->minlen))
6455 min1 = minnext + trie->minlen;
6456 if (deltanext == OPTIMIZE_INFTY) {
6457 is_inf = is_inf_internal = 1;
6458 max1 = OPTIMIZE_INFTY;
6459 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
6460 max1 = minnext + deltanext + trie->maxlen;
6462 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
6464 if (data_fake.flags & SCF_SEEN_ACCEPT) {
6465 if ( stopmin > min + min1)
6466 stopmin = min + min1;
6467 flags &= ~SCF_DO_SUBSTR;
6469 data->flags |= SCF_SEEN_ACCEPT;
6472 if (data_fake.flags & SF_HAS_EVAL)
6473 data->flags |= SF_HAS_EVAL;
6474 data->whilem_c = data_fake.whilem_c;
6476 if (flags & SCF_DO_STCLASS)
6477 ssc_or(pRExC_state, &accum, (regnode_charclass *) &this_class);
6480 if (flags & SCF_DO_SUBSTR) {
6481 data->pos_min += min1;
6482 data->pos_delta += max1 - min1;
6483 if (max1 != min1 || is_inf)
6484 data->cur_is_floating = 1; /* float */
6487 if (delta != OPTIMIZE_INFTY) {
6488 if (OPTIMIZE_INFTY - (max1 - min1) >= delta)
6489 delta += max1 - min1;
6491 delta = OPTIMIZE_INFTY;
6493 if (flags & SCF_DO_STCLASS_OR) {
6494 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &accum);
6496 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6497 flags &= ~SCF_DO_STCLASS;
6500 else if (flags & SCF_DO_STCLASS_AND) {
6502 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
6503 flags &= ~SCF_DO_STCLASS;
6506 /* Switch to OR mode: cache the old value of
6507 * data->start_class */
6509 StructCopy(data->start_class, and_withp, regnode_ssc);
6510 flags &= ~SCF_DO_STCLASS_AND;
6511 StructCopy(&accum, data->start_class, regnode_ssc);
6512 flags |= SCF_DO_STCLASS_OR;
6519 else if (PL_regkind[OP(scan)] == TRIE) {
6520 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
6523 min += trie->minlen;
6524 delta += (trie->maxlen - trie->minlen);
6525 flags &= ~SCF_DO_STCLASS; /* xxx */
6526 if (flags & SCF_DO_SUBSTR) {
6527 /* Cannot expect anything... */
6528 scan_commit(pRExC_state, data, minlenp, is_inf);
6529 data->pos_min += trie->minlen;
6530 data->pos_delta += (trie->maxlen - trie->minlen);
6531 if (trie->maxlen != trie->minlen)
6532 data->cur_is_floating = 1; /* float */
6534 if (trie->jump) /* no more substrings -- for now /grr*/
6535 flags &= ~SCF_DO_SUBSTR;
6537 else if (OP(scan) == REGEX_SET) {
6538 Perl_croak(aTHX_ "panic: %s regnode should be resolved"
6539 " before optimization", reg_name[REGEX_SET]);
6542 #endif /* old or new */
6543 #endif /* TRIE_STUDY_OPT */
6545 /* Else: zero-length, ignore. */
6546 scan = regnext(scan);
6551 /* we need to unwind recursion. */
6554 DEBUG_STUDYDATA("frame-end", data, depth, is_inf);
6555 DEBUG_PEEP("fend", scan, depth, flags);
6557 /* restore previous context */
6558 last = frame->last_regnode;
6559 scan = frame->next_regnode;
6560 stopparen = frame->stopparen;
6561 recursed_depth = frame->prev_recursed_depth;
6563 RExC_frame_last = frame->prev_frame;
6564 frame = frame->this_prev_frame;
6565 goto fake_study_recurse;
6569 DEBUG_STUDYDATA("pre-fin", data, depth, is_inf);
6572 *deltap = is_inf_internal ? OPTIMIZE_INFTY : delta;
6574 if (flags & SCF_DO_SUBSTR && is_inf)
6575 data->pos_delta = OPTIMIZE_INFTY - data->pos_min;
6576 if (is_par > (I32)U8_MAX)
6578 if (is_par && pars==1 && data) {
6579 data->flags |= SF_IN_PAR;
6580 data->flags &= ~SF_HAS_PAR;
6582 else if (pars && data) {
6583 data->flags |= SF_HAS_PAR;
6584 data->flags &= ~SF_IN_PAR;
6586 if (flags & SCF_DO_STCLASS_OR)
6587 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6588 if (flags & SCF_TRIE_RESTUDY)
6589 data->flags |= SCF_TRIE_RESTUDY;
6591 DEBUG_STUDYDATA("post-fin", data, depth, is_inf);
6593 final_minlen = min < stopmin
6596 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN)) {
6597 if (final_minlen > OPTIMIZE_INFTY - delta)
6598 RExC_maxlen = OPTIMIZE_INFTY;
6599 else if (RExC_maxlen < final_minlen + delta)
6600 RExC_maxlen = final_minlen + delta;
6602 return final_minlen;
6606 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
6608 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
6610 PERL_ARGS_ASSERT_ADD_DATA;
6612 Renewc(RExC_rxi->data,
6613 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
6614 char, struct reg_data);
6616 Renew(RExC_rxi->data->what, count + n, U8);
6618 Newx(RExC_rxi->data->what, n, U8);
6619 RExC_rxi->data->count = count + n;
6620 Copy(s, RExC_rxi->data->what + count, n, U8);
6624 /*XXX: todo make this not included in a non debugging perl, but appears to be
6625 * used anyway there, in 'use re' */
6626 #ifndef PERL_IN_XSUB_RE
6628 Perl_reginitcolors(pTHX)
6630 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
6632 char *t = savepv(s);
6636 t = strchr(t, '\t');
6642 PL_colors[i] = t = (char *)"";
6647 PL_colors[i++] = (char *)"";
6654 #ifdef TRIE_STUDY_OPT
6655 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
6658 (data.flags & SCF_TRIE_RESTUDY) \
6666 #define CHECK_RESTUDY_GOTO_butfirst
6670 * pregcomp - compile a regular expression into internal code
6672 * Decides which engine's compiler to call based on the hint currently in
6676 #ifndef PERL_IN_XSUB_RE
6678 /* return the currently in-scope regex engine (or the default if none) */
6680 regexp_engine const *
6681 Perl_current_re_engine(pTHX)
6683 if (IN_PERL_COMPILETIME) {
6684 HV * const table = GvHV(PL_hintgv);
6687 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
6688 return &PL_core_reg_engine;
6689 ptr = hv_fetchs(table, "regcomp", FALSE);
6690 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
6691 return &PL_core_reg_engine;
6692 return INT2PTR(regexp_engine*, SvIV(*ptr));
6696 if (!PL_curcop->cop_hints_hash)
6697 return &PL_core_reg_engine;
6698 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
6699 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
6700 return &PL_core_reg_engine;
6701 return INT2PTR(regexp_engine*, SvIV(ptr));
6707 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
6709 regexp_engine const *eng = current_re_engine();
6710 DECLARE_AND_GET_RE_DEBUG_FLAGS;
6712 PERL_ARGS_ASSERT_PREGCOMP;
6714 /* Dispatch a request to compile a regexp to correct regexp engine. */
6716 Perl_re_printf( aTHX_ "Using engine %" UVxf "\n",
6719 return CALLREGCOMP_ENG(eng, pattern, flags);
6723 /* public(ish) entry point for the perl core's own regex compiling code.
6724 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
6725 * pattern rather than a list of OPs, and uses the internal engine rather
6726 * than the current one */
6729 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
6731 SV *pat = pattern; /* defeat constness! */
6733 PERL_ARGS_ASSERT_RE_COMPILE;
6735 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
6736 #ifdef PERL_IN_XSUB_RE
6739 &PL_core_reg_engine,
6741 NULL, NULL, rx_flags, 0);
6745 S_free_codeblocks(pTHX_ struct reg_code_blocks *cbs)
6749 if (--cbs->refcnt > 0)
6751 for (n = 0; n < cbs->count; n++) {
6752 REGEXP *rx = cbs->cb[n].src_regex;
6754 cbs->cb[n].src_regex = NULL;
6755 SvREFCNT_dec_NN(rx);
6763 static struct reg_code_blocks *
6764 S_alloc_code_blocks(pTHX_ int ncode)
6766 struct reg_code_blocks *cbs;
6767 Newx(cbs, 1, struct reg_code_blocks);
6770 SAVEDESTRUCTOR_X(S_free_codeblocks, cbs);
6772 Newx(cbs->cb, ncode, struct reg_code_block);
6779 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
6780 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
6781 * point to the realloced string and length.
6783 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
6787 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
6788 char **pat_p, STRLEN *plen_p, int num_code_blocks)
6790 U8 *const src = (U8*)*pat_p;
6795 DECLARE_AND_GET_RE_DEBUG_FLAGS;
6797 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6798 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
6800 /* 1 for each byte + 1 for each byte that expands to two, + trailing NUL */
6801 Newx(dst, *plen_p + variant_under_utf8_count(src, src + *plen_p) + 1, U8);
6804 while (s < *plen_p) {
6805 append_utf8_from_native_byte(src[s], &d);
6807 if (n < num_code_blocks) {
6808 assert(pRExC_state->code_blocks);
6809 if (!do_end && pRExC_state->code_blocks->cb[n].start == s) {
6810 pRExC_state->code_blocks->cb[n].start = d - dst - 1;
6811 assert(*(d - 1) == '(');
6814 else if (do_end && pRExC_state->code_blocks->cb[n].end == s) {
6815 pRExC_state->code_blocks->cb[n].end = d - dst - 1;
6816 assert(*(d - 1) == ')');
6825 *pat_p = (char*) dst;
6827 RExC_orig_utf8 = RExC_utf8 = 1;
6832 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
6833 * while recording any code block indices, and handling overloading,
6834 * nested qr// objects etc. If pat is null, it will allocate a new
6835 * string, or just return the first arg, if there's only one.
6837 * Returns the malloced/updated pat.
6838 * patternp and pat_count is the array of SVs to be concatted;
6839 * oplist is the optional list of ops that generated the SVs;
6840 * recompile_p is a pointer to a boolean that will be set if
6841 * the regex will need to be recompiled.
6842 * delim, if non-null is an SV that will be inserted between each element
6846 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
6847 SV *pat, SV ** const patternp, int pat_count,
6848 OP *oplist, bool *recompile_p, SV *delim)
6852 bool use_delim = FALSE;
6853 bool alloced = FALSE;
6855 /* if we know we have at least two args, create an empty string,
6856 * then concatenate args to that. For no args, return an empty string */
6857 if (!pat && pat_count != 1) {
6863 for (svp = patternp; svp < patternp + pat_count; svp++) {
6866 STRLEN orig_patlen = 0;
6868 SV *msv = use_delim ? delim : *svp;
6869 if (!msv) msv = &PL_sv_undef;
6871 /* if we've got a delimiter, we go round the loop twice for each
6872 * svp slot (except the last), using the delimiter the second
6881 if (SvTYPE(msv) == SVt_PVAV) {
6882 /* we've encountered an interpolated array within
6883 * the pattern, e.g. /...@a..../. Expand the list of elements,
6884 * then recursively append elements.
6885 * The code in this block is based on S_pushav() */
6887 AV *const av = (AV*)msv;
6888 const SSize_t maxarg = AvFILL(av) + 1;
6892 assert(oplist->op_type == OP_PADAV
6893 || oplist->op_type == OP_RV2AV);
6894 oplist = OpSIBLING(oplist);
6897 if (SvRMAGICAL(av)) {
6900 Newx(array, maxarg, SV*);
6902 for (i=0; i < maxarg; i++) {
6903 SV ** const svp = av_fetch(av, i, FALSE);
6904 array[i] = svp ? *svp : &PL_sv_undef;
6908 array = AvARRAY(av);
6910 pat = S_concat_pat(aTHX_ pRExC_state, pat,
6911 array, maxarg, NULL, recompile_p,
6913 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
6919 /* we make the assumption here that each op in the list of
6920 * op_siblings maps to one SV pushed onto the stack,
6921 * except for code blocks, with have both an OP_NULL and
6923 * This allows us to match up the list of SVs against the
6924 * list of OPs to find the next code block.
6926 * Note that PUSHMARK PADSV PADSV ..
6928 * PADRANGE PADSV PADSV ..
6929 * so the alignment still works. */
6932 if (oplist->op_type == OP_NULL
6933 && (oplist->op_flags & OPf_SPECIAL))
6935 assert(n < pRExC_state->code_blocks->count);
6936 pRExC_state->code_blocks->cb[n].start = pat ? SvCUR(pat) : 0;
6937 pRExC_state->code_blocks->cb[n].block = oplist;
6938 pRExC_state->code_blocks->cb[n].src_regex = NULL;
6941 oplist = OpSIBLING(oplist); /* skip CONST */
6944 oplist = OpSIBLING(oplist);;
6947 /* apply magic and QR overloading to arg */
6950 if (SvROK(msv) && SvAMAGIC(msv)) {
6951 SV *sv = AMG_CALLunary(msv, regexp_amg);
6955 if (SvTYPE(sv) != SVt_REGEXP)
6956 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
6961 /* try concatenation overload ... */
6962 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
6963 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
6966 /* overloading involved: all bets are off over literal
6967 * code. Pretend we haven't seen it */
6969 pRExC_state->code_blocks->count -= n;
6973 /* ... or failing that, try "" overload */
6974 while (SvAMAGIC(msv)
6975 && (sv = AMG_CALLunary(msv, string_amg))
6979 && SvRV(msv) == SvRV(sv))
6984 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
6988 /* this is a partially unrolled
6989 * sv_catsv_nomg(pat, msv);
6990 * that allows us to adjust code block indices if
6993 char *dst = SvPV_force_nomg(pat, dlen);
6995 if (SvUTF8(msv) && !SvUTF8(pat)) {
6996 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
6997 sv_setpvn(pat, dst, dlen);
7000 sv_catsv_nomg(pat, msv);
7004 /* We have only one SV to process, but we need to verify
7005 * it is properly null terminated or we will fail asserts
7006 * later. In theory we probably shouldn't get such SV's,
7007 * but if we do we should handle it gracefully. */
7008 if ( SvTYPE(msv) != SVt_PV || (SvLEN(msv) > SvCUR(msv) && *(SvEND(msv)) == 0) || SvIsCOW_shared_hash(msv) ) {
7009 /* not a string, or a string with a trailing null */
7012 /* a string with no trailing null, we need to copy it
7013 * so it has a trailing null */
7014 pat = sv_2mortal(newSVsv(msv));
7019 pRExC_state->code_blocks->cb[n-1].end = SvCUR(pat)-1;
7022 /* extract any code blocks within any embedded qr//'s */
7023 if (rx && SvTYPE(rx) == SVt_REGEXP
7024 && RX_ENGINE((REGEXP*)rx)->op_comp)
7027 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
7028 if (ri->code_blocks && ri->code_blocks->count) {
7030 /* the presence of an embedded qr// with code means
7031 * we should always recompile: the text of the
7032 * qr// may not have changed, but it may be a
7033 * different closure than last time */
7035 if (pRExC_state->code_blocks) {
7036 int new_count = pRExC_state->code_blocks->count
7037 + ri->code_blocks->count;
7038 Renew(pRExC_state->code_blocks->cb,
7039 new_count, struct reg_code_block);
7040 pRExC_state->code_blocks->count = new_count;
7043 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_
7044 ri->code_blocks->count);
7046 for (i=0; i < ri->code_blocks->count; i++) {
7047 struct reg_code_block *src, *dst;
7048 STRLEN offset = orig_patlen
7049 + ReANY((REGEXP *)rx)->pre_prefix;
7050 assert(n < pRExC_state->code_blocks->count);
7051 src = &ri->code_blocks->cb[i];
7052 dst = &pRExC_state->code_blocks->cb[n];
7053 dst->start = src->start + offset;
7054 dst->end = src->end + offset;
7055 dst->block = src->block;
7056 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
7065 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
7074 /* see if there are any run-time code blocks in the pattern.
7075 * False positives are allowed */
7078 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
7079 char *pat, STRLEN plen)
7084 PERL_UNUSED_CONTEXT;
7086 for (s = 0; s < plen; s++) {
7087 if ( pRExC_state->code_blocks
7088 && n < pRExC_state->code_blocks->count
7089 && s == pRExC_state->code_blocks->cb[n].start)
7091 s = pRExC_state->code_blocks->cb[n].end;
7095 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
7097 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
7099 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
7106 /* Handle run-time code blocks. We will already have compiled any direct
7107 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
7108 * copy of it, but with any literal code blocks blanked out and
7109 * appropriate chars escaped; then feed it into
7111 * eval "qr'modified_pattern'"
7115 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
7119 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
7121 * After eval_sv()-ing that, grab any new code blocks from the returned qr
7122 * and merge them with any code blocks of the original regexp.
7124 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
7125 * instead, just save the qr and return FALSE; this tells our caller that
7126 * the original pattern needs upgrading to utf8.
7130 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
7131 char *pat, STRLEN plen)
7135 DECLARE_AND_GET_RE_DEBUG_FLAGS;
7137 if (pRExC_state->runtime_code_qr) {
7138 /* this is the second time we've been called; this should
7139 * only happen if the main pattern got upgraded to utf8
7140 * during compilation; re-use the qr we compiled first time
7141 * round (which should be utf8 too)
7143 qr = pRExC_state->runtime_code_qr;
7144 pRExC_state->runtime_code_qr = NULL;
7145 assert(RExC_utf8 && SvUTF8(qr));
7151 int newlen = plen + 7; /* allow for "qr''xx\0" extra chars */
7155 /* determine how many extra chars we need for ' and \ escaping */
7156 for (s = 0; s < plen; s++) {
7157 if (pat[s] == '\'' || pat[s] == '\\')
7161 Newx(newpat, newlen, char);
7163 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
7165 for (s = 0; s < plen; s++) {
7166 if ( pRExC_state->code_blocks
7167 && n < pRExC_state->code_blocks->count
7168 && s == pRExC_state->code_blocks->cb[n].start)
7170 /* blank out literal code block so that they aren't
7171 * recompiled: eg change from/to:
7181 assert(pat[s] == '(');
7182 assert(pat[s+1] == '?');
7186 while (s < pRExC_state->code_blocks->cb[n].end) {
7194 if (pat[s] == '\'' || pat[s] == '\\')
7199 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED) {
7201 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED_MORE) {
7207 Perl_re_printf( aTHX_
7208 "%sre-parsing pattern for runtime code:%s %s\n",
7209 PL_colors[4], PL_colors[5], newpat);
7212 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
7218 PUSHSTACKi(PERLSI_REQUIRE);
7219 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
7220 * parsing qr''; normally only q'' does this. It also alters
7222 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
7223 SvREFCNT_dec_NN(sv);
7228 SV * const errsv = ERRSV;
7229 if (SvTRUE_NN(errsv))
7230 /* use croak_sv ? */
7231 Perl_croak_nocontext("%" SVf, SVfARG(errsv));
7233 assert(SvROK(qr_ref));
7235 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
7236 /* the leaving below frees the tmp qr_ref.
7237 * Give qr a life of its own */
7245 if (!RExC_utf8 && SvUTF8(qr)) {
7246 /* first time through; the pattern got upgraded; save the
7247 * qr for the next time through */
7248 assert(!pRExC_state->runtime_code_qr);
7249 pRExC_state->runtime_code_qr = qr;
7254 /* extract any code blocks within the returned qr// */
7257 /* merge the main (r1) and run-time (r2) code blocks into one */
7259 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
7260 struct reg_code_block *new_block, *dst;
7261 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
7265 if (!r2->code_blocks || !r2->code_blocks->count) /* we guessed wrong */
7267 SvREFCNT_dec_NN(qr);
7271 if (!r1->code_blocks)
7272 r1->code_blocks = S_alloc_code_blocks(aTHX_ 0);
7274 r1c = r1->code_blocks->count;
7275 r2c = r2->code_blocks->count;
7277 Newx(new_block, r1c + r2c, struct reg_code_block);
7281 while (i1 < r1c || i2 < r2c) {
7282 struct reg_code_block *src;
7286 src = &r2->code_blocks->cb[i2++];
7290 src = &r1->code_blocks->cb[i1++];
7291 else if ( r1->code_blocks->cb[i1].start
7292 < r2->code_blocks->cb[i2].start)
7294 src = &r1->code_blocks->cb[i1++];
7295 assert(src->end < r2->code_blocks->cb[i2].start);
7298 assert( r1->code_blocks->cb[i1].start
7299 > r2->code_blocks->cb[i2].start);
7300 src = &r2->code_blocks->cb[i2++];
7302 assert(src->end < r1->code_blocks->cb[i1].start);
7305 assert(pat[src->start] == '(');
7306 assert(pat[src->end] == ')');
7307 dst->start = src->start;
7308 dst->end = src->end;
7309 dst->block = src->block;
7310 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
7314 r1->code_blocks->count += r2c;
7315 Safefree(r1->code_blocks->cb);
7316 r1->code_blocks->cb = new_block;
7319 SvREFCNT_dec_NN(qr);
7325 S_setup_longest(pTHX_ RExC_state_t *pRExC_state,
7326 struct reg_substr_datum *rsd,
7327 struct scan_data_substrs *sub,
7328 STRLEN longest_length)
7330 /* This is the common code for setting up the floating and fixed length
7331 * string data extracted from Perl_re_op_compile() below. Returns a boolean
7332 * as to whether succeeded or not */
7336 bool eol = cBOOL(sub->flags & SF_BEFORE_EOL);
7337 bool meol = cBOOL(sub->flags & SF_BEFORE_MEOL);
7339 if (! (longest_length
7340 || (eol /* Can't have SEOL and MULTI */
7341 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
7343 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
7344 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
7349 /* copy the information about the longest from the reg_scan_data
7350 over to the program. */
7351 if (SvUTF8(sub->str)) {
7353 rsd->utf8_substr = sub->str;
7355 rsd->substr = sub->str;
7356 rsd->utf8_substr = NULL;
7358 /* end_shift is how many chars that must be matched that
7359 follow this item. We calculate it ahead of time as once the
7360 lookbehind offset is added in we lose the ability to correctly
7362 ml = sub->minlenp ? *(sub->minlenp) : (SSize_t)longest_length;
7363 rsd->end_shift = ml - sub->min_offset
7365 /* XXX SvTAIL is always false here - did you mean FBMcf_TAIL
7367 + (SvTAIL(sub->str) != 0)
7371 t = (eol/* Can't have SEOL and MULTI */
7372 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
7373 fbm_compile(sub->str, t ? FBMcf_TAIL : 0);
7379 S_set_regex_pv(pTHX_ RExC_state_t *pRExC_state, REGEXP *Rx)
7381 /* Calculates and sets in the compiled pattern 'Rx' the string to compile,
7382 * properly wrapped with the right modifiers */
7384 bool has_p = ((RExC_rx->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
7385 bool has_charset = RExC_utf8 || (get_regex_charset(RExC_rx->extflags)
7386 != REGEX_DEPENDS_CHARSET);
7388 /* The caret is output if there are any defaults: if not all the STD
7389 * flags are set, or if no character set specifier is needed */
7391 (((RExC_rx->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
7393 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
7394 == REG_RUN_ON_COMMENT_SEEN);
7395 U8 reganch = (U8)((RExC_rx->extflags & RXf_PMf_STD_PMMOD)
7396 >> RXf_PMf_STD_PMMOD_SHIFT);
7397 const char *fptr = STD_PAT_MODS; /*"msixxn"*/
7399 STRLEN pat_len = RExC_precomp_end - RExC_precomp;
7401 /* We output all the necessary flags; we never output a minus, as all
7402 * those are defaults, so are
7403 * covered by the caret */
7404 const STRLEN wraplen = pat_len + has_p + has_runon
7405 + has_default /* If needs a caret */
7406 + PL_bitcount[reganch] /* 1 char for each set standard flag */
7408 /* If needs a character set specifier */
7409 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
7410 + (sizeof("(?:)") - 1);
7412 PERL_ARGS_ASSERT_SET_REGEX_PV;
7414 /* make sure PL_bitcount bounds not exceeded */
7415 STATIC_ASSERT_STMT(sizeof(STD_PAT_MODS) <= 8);
7417 p = sv_grow(MUTABLE_SV(Rx), wraplen + 1); /* +1 for the ending NUL */
7420 SvFLAGS(Rx) |= SVf_UTF8;
7423 /* If a default, cover it using the caret */
7425 *p++= DEFAULT_PAT_MOD;
7431 name = get_regex_charset_name(RExC_rx->extflags, &len);
7432 if (strEQ(name, DEPENDS_PAT_MODS)) { /* /d under UTF-8 => /u */
7434 name = UNICODE_PAT_MODS;
7435 len = sizeof(UNICODE_PAT_MODS) - 1;
7437 Copy(name, p, len, char);
7441 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
7444 while((ch = *fptr++)) {
7452 Copy(RExC_precomp, p, pat_len, char);
7453 assert ((RX_WRAPPED(Rx) - p) < 16);
7454 RExC_rx->pre_prefix = p - RX_WRAPPED(Rx);
7457 /* Adding a trailing \n causes this to compile properly:
7458 my $R = qr / A B C # D E/x; /($R)/
7459 Otherwise the parens are considered part of the comment */
7464 SvCUR_set(Rx, p - RX_WRAPPED(Rx));
7468 * Perl_re_op_compile - the perl internal RE engine's function to compile a
7469 * regular expression into internal code.
7470 * The pattern may be passed either as:
7471 * a list of SVs (patternp plus pat_count)
7472 * a list of OPs (expr)
7473 * If both are passed, the SV list is used, but the OP list indicates
7474 * which SVs are actually pre-compiled code blocks
7476 * The SVs in the list have magic and qr overloading applied to them (and
7477 * the list may be modified in-place with replacement SVs in the latter
7480 * If the pattern hasn't changed from old_re, then old_re will be
7483 * eng is the current engine. If that engine has an op_comp method, then
7484 * handle directly (i.e. we assume that op_comp was us); otherwise, just
7485 * do the initial concatenation of arguments and pass on to the external
7488 * If is_bare_re is not null, set it to a boolean indicating whether the
7489 * arg list reduced (after overloading) to a single bare regex which has
7490 * been returned (i.e. /$qr/).
7492 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
7494 * pm_flags contains the PMf_* flags, typically based on those from the
7495 * pm_flags field of the related PMOP. Currently we're only interested in
7496 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL, PMf_WILDCARD.
7498 * For many years this code had an initial sizing pass that calculated
7499 * (sometimes incorrectly, leading to security holes) the size needed for the
7500 * compiled pattern. That was changed by commit
7501 * 7c932d07cab18751bfc7515b4320436273a459e2 in 5.29, which reallocs the size, a
7502 * node at a time, as parsing goes along. Patches welcome to fix any obsolete
7503 * references to this sizing pass.
7505 * Now, an initial crude guess as to the size needed is made, based on the
7506 * length of the pattern. Patches welcome to improve that guess. That amount
7507 * of space is malloc'd and then immediately freed, and then clawed back node
7508 * by node. This design is to minimze, to the extent possible, memory churn
7509 * when doing the reallocs.
7511 * A separate parentheses counting pass may be needed in some cases.
7512 * (Previously the sizing pass did this.) Patches welcome to reduce the number
7515 * The existence of a sizing pass necessitated design decisions that are no
7516 * longer needed. There are potential areas of simplification.
7518 * Beware that the optimization-preparation code in here knows about some
7519 * of the structure of the compiled regexp. [I'll say.]
7523 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
7524 OP *expr, const regexp_engine* eng, REGEXP *old_re,
7525 bool *is_bare_re, const U32 orig_rx_flags, const U32 pm_flags)
7527 REGEXP *Rx; /* Capital 'R' means points to a REGEXP */
7535 SV** new_patternp = patternp;
7537 /* these are all flags - maybe they should be turned
7538 * into a single int with different bit masks */
7539 I32 sawlookahead = 0;
7544 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
7546 bool runtime_code = 0;
7548 RExC_state_t RExC_state;
7549 RExC_state_t * const pRExC_state = &RExC_state;
7550 #ifdef TRIE_STUDY_OPT
7552 RExC_state_t copyRExC_state;
7554 DECLARE_AND_GET_RE_DEBUG_FLAGS;
7556 PERL_ARGS_ASSERT_RE_OP_COMPILE;
7558 DEBUG_r(if (!PL_colorset) reginitcolors());
7561 pRExC_state->warn_text = NULL;
7562 pRExC_state->unlexed_names = NULL;
7563 pRExC_state->code_blocks = NULL;
7566 *is_bare_re = FALSE;
7568 if (expr && (expr->op_type == OP_LIST ||
7569 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
7570 /* allocate code_blocks if needed */
7574 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o))
7575 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
7576 ncode++; /* count of DO blocks */
7579 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_ ncode);
7583 /* compile-time pattern with just OP_CONSTs and DO blocks */
7588 /* find how many CONSTs there are */
7591 if (expr->op_type == OP_CONST)
7594 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
7595 if (o->op_type == OP_CONST)
7599 /* fake up an SV array */
7601 assert(!new_patternp);
7602 Newx(new_patternp, n, SV*);
7603 SAVEFREEPV(new_patternp);
7607 if (expr->op_type == OP_CONST)
7608 new_patternp[n] = cSVOPx_sv(expr);
7610 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
7611 if (o->op_type == OP_CONST)
7612 new_patternp[n++] = cSVOPo_sv;
7617 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7618 "Assembling pattern from %d elements%s\n", pat_count,
7619 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
7621 /* set expr to the first arg op */
7623 if (pRExC_state->code_blocks && pRExC_state->code_blocks->count
7624 && expr->op_type != OP_CONST)
7626 expr = cLISTOPx(expr)->op_first;
7627 assert( expr->op_type == OP_PUSHMARK
7628 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
7629 || expr->op_type == OP_PADRANGE);
7630 expr = OpSIBLING(expr);
7633 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
7634 expr, &recompile, NULL);
7636 /* handle bare (possibly after overloading) regex: foo =~ $re */
7641 if (SvTYPE(re) == SVt_REGEXP) {
7645 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7646 "Precompiled pattern%s\n",
7647 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
7653 exp = SvPV_nomg(pat, plen);
7655 if (!eng->op_comp) {
7656 if ((SvUTF8(pat) && IN_BYTES)
7657 || SvGMAGICAL(pat) || SvAMAGIC(pat))
7659 /* make a temporary copy; either to convert to bytes,
7660 * or to avoid repeating get-magic / overloaded stringify */
7661 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
7662 (IN_BYTES ? 0 : SvUTF8(pat)));
7664 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
7667 /* ignore the utf8ness if the pattern is 0 length */
7668 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
7669 RExC_uni_semantics = 0;
7670 RExC_contains_locale = 0;
7671 RExC_strict = cBOOL(pm_flags & RXf_PMf_STRICT);
7672 RExC_in_script_run = 0;
7673 RExC_study_started = 0;
7674 pRExC_state->runtime_code_qr = NULL;
7675 RExC_frame_head= NULL;
7676 RExC_frame_last= NULL;
7677 RExC_frame_count= 0;
7678 RExC_latest_warn_offset = 0;
7679 RExC_use_BRANCHJ = 0;
7680 RExC_warned_WARN_EXPERIMENTAL__VLB = 0;
7681 RExC_warned_WARN_EXPERIMENTAL__REGEX_SETS = 0;
7682 RExC_total_parens = 0;
7683 RExC_open_parens = NULL;
7684 RExC_close_parens = NULL;
7685 RExC_paren_names = NULL;
7687 RExC_seen_d_op = FALSE;
7689 RExC_paren_name_list = NULL;
7693 RExC_mysv1= sv_newmortal();
7694 RExC_mysv2= sv_newmortal();
7698 SV *dsv= sv_newmortal();
7699 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, PL_dump_re_max_len);
7700 Perl_re_printf( aTHX_ "%sCompiling REx%s %s\n",
7701 PL_colors[4], PL_colors[5], s);
7704 /* we jump here if we have to recompile, e.g., from upgrading the pattern
7707 if ((pm_flags & PMf_USE_RE_EVAL)
7708 /* this second condition covers the non-regex literal case,
7709 * i.e. $foo =~ '(?{})'. */
7710 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
7712 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
7715 /* return old regex if pattern hasn't changed */
7716 /* XXX: note in the below we have to check the flags as well as the
7719 * Things get a touch tricky as we have to compare the utf8 flag
7720 * independently from the compile flags. */
7724 && !!RX_UTF8(old_re) == !!RExC_utf8
7725 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
7726 && RX_PRECOMP(old_re)
7727 && RX_PRELEN(old_re) == plen
7728 && memEQ(RX_PRECOMP(old_re), exp, plen)
7729 && !runtime_code /* with runtime code, always recompile */ )
7732 SV *dsv= sv_newmortal();
7733 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, PL_dump_re_max_len);
7734 Perl_re_printf( aTHX_ "%sSkipping recompilation of unchanged REx%s %s\n",
7735 PL_colors[4], PL_colors[5], s);
7740 /* Allocate the pattern's SV */
7741 RExC_rx_sv = Rx = (REGEXP*) newSV_type(SVt_REGEXP);
7742 RExC_rx = ReANY(Rx);
7743 if ( RExC_rx == NULL )
7744 FAIL("Regexp out of space");
7746 rx_flags = orig_rx_flags;
7748 if ( toUSE_UNI_CHARSET_NOT_DEPENDS
7749 && initial_charset == REGEX_DEPENDS_CHARSET)
7752 /* Set to use unicode semantics if the pattern is in utf8 and has the
7753 * 'depends' charset specified, as it means unicode when utf8 */
7754 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
7755 RExC_uni_semantics = 1;
7758 RExC_pm_flags = pm_flags;
7761 assert(TAINTING_get || !TAINT_get);
7763 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
7765 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
7766 /* whoops, we have a non-utf8 pattern, whilst run-time code
7767 * got compiled as utf8. Try again with a utf8 pattern */
7768 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7769 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7773 assert(!pRExC_state->runtime_code_qr);
7779 RExC_in_lookaround = 0;
7780 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
7781 RExC_recode_x_to_native = 0;
7782 RExC_in_multi_char_class = 0;
7784 RExC_start = RExC_copy_start_in_constructed = RExC_copy_start_in_input = RExC_precomp = exp;
7785 RExC_precomp_end = RExC_end = exp + plen;
7787 RExC_whilem_seen = 0;
7789 RExC_recurse = NULL;
7790 RExC_study_chunk_recursed = NULL;
7791 RExC_study_chunk_recursed_bytes= 0;
7792 RExC_recurse_count = 0;
7793 RExC_sets_depth = 0;
7794 pRExC_state->code_index = 0;
7796 /* Initialize the string in the compiled pattern. This is so that there is
7797 * something to output if necessary */
7798 set_regex_pv(pRExC_state, Rx);
7801 Perl_re_printf( aTHX_
7802 "Starting parse and generation\n");
7804 RExC_lastparse=NULL;
7807 /* Allocate space and zero-initialize. Note, the two step process
7808 of zeroing when in debug mode, thus anything assigned has to
7809 happen after that */
7812 /* On the first pass of the parse, we guess how big this will be. Then
7813 * we grow in one operation to that amount and then give it back. As
7814 * we go along, we re-allocate what we need.
7816 * XXX Currently the guess is essentially that the pattern will be an
7817 * EXACT node with one byte input, one byte output. This is crude, and
7818 * better heuristics are welcome.
7820 * On any subsequent passes, we guess what we actually computed in the
7821 * latest earlier pass. Such a pass probably didn't complete so is
7822 * missing stuff. We could improve those guesses by knowing where the
7823 * parse stopped, and use the length so far plus apply the above
7824 * assumption to what's left. */
7825 RExC_size = STR_SZ(RExC_end - RExC_start);
7828 Newxc(RExC_rxi, sizeof(regexp_internal) + RExC_size, char, regexp_internal);
7829 if ( RExC_rxi == NULL )
7830 FAIL("Regexp out of space");
7832 Zero(RExC_rxi, sizeof(regexp_internal) + RExC_size, char);
7833 RXi_SET( RExC_rx, RExC_rxi );
7835 /* We start from 0 (over from 0 in the case this is a reparse. The first
7836 * node parsed will give back any excess memory we have allocated so far).
7840 /* non-zero initialization begins here */
7841 RExC_rx->engine= eng;
7842 RExC_rx->extflags = rx_flags;
7843 RXp_COMPFLAGS(RExC_rx) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
7845 if (pm_flags & PMf_IS_QR) {
7846 RExC_rxi->code_blocks = pRExC_state->code_blocks;
7847 if (RExC_rxi->code_blocks) {
7848 RExC_rxi->code_blocks->refcnt++;
7852 RExC_rx->intflags = 0;
7854 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
7857 /* This NUL is guaranteed because the pattern comes from an SV*, and the sv
7858 * code makes sure the final byte is an uncounted NUL. But should this
7859 * ever not be the case, lots of things could read beyond the end of the
7860 * buffer: loops like
7861 * while(isFOO(*RExC_parse)) RExC_parse++;
7862 * strchr(RExC_parse, "foo");
7863 * etc. So it is worth noting. */
7864 assert(*RExC_end == '\0');
7868 RExC_parens_buf_size = 0;
7869 RExC_emit_start = RExC_rxi->program;
7870 pRExC_state->code_index = 0;
7872 *((char*) RExC_emit_start) = (char) REG_MAGIC;
7876 if (reg(pRExC_state, 0, &flags, 1)) {
7878 /* Success!, But we may need to redo the parse knowing how many parens
7879 * there actually are */
7880 if (IN_PARENS_PASS) {
7881 flags |= RESTART_PARSE;
7884 /* We have that number in RExC_npar */
7885 RExC_total_parens = RExC_npar;
7887 else if (! MUST_RESTART(flags)) {
7889 Perl_croak(aTHX_ "panic: reg returned failure to re_op_compile, flags=%#" UVxf, (UV) flags);
7892 /* Here, we either have success, or we have to redo the parse for some reason */
7893 if (MUST_RESTART(flags)) {
7895 /* It's possible to write a regexp in ascii that represents Unicode
7896 codepoints outside of the byte range, such as via \x{100}. If we
7897 detect such a sequence we have to convert the entire pattern to utf8
7898 and then recompile, as our sizing calculation will have been based
7899 on 1 byte == 1 character, but we will need to use utf8 to encode
7900 at least some part of the pattern, and therefore must convert the whole
7903 if (flags & NEED_UTF8) {
7905 /* We have stored the offset of the final warning output so far.
7906 * That must be adjusted. Any variant characters between the start
7907 * of the pattern and this warning count for 2 bytes in the final,
7908 * so just add them again */
7909 if (UNLIKELY(RExC_latest_warn_offset > 0)) {
7910 RExC_latest_warn_offset +=
7911 variant_under_utf8_count((U8 *) exp, (U8 *) exp
7912 + RExC_latest_warn_offset);
7914 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7915 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7916 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "Need to redo parse after upgrade\n"));
7919 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "Need to redo parse\n"));
7922 if (ALL_PARENS_COUNTED) {
7923 /* Make enough room for all the known parens, and zero it */
7924 Renew(RExC_open_parens, RExC_total_parens, regnode_offset);
7925 Zero(RExC_open_parens, RExC_total_parens, regnode_offset);
7926 RExC_open_parens[0] = 1; /* +1 for REG_MAGIC */
7928 Renew(RExC_close_parens, RExC_total_parens, regnode_offset);
7929 Zero(RExC_close_parens, RExC_total_parens, regnode_offset);
7931 else { /* Parse did not complete. Reinitialize the parentheses
7933 RExC_total_parens = 0;
7934 if (RExC_open_parens) {
7935 Safefree(RExC_open_parens);
7936 RExC_open_parens = NULL;
7938 if (RExC_close_parens) {
7939 Safefree(RExC_close_parens);
7940 RExC_close_parens = NULL;
7944 /* Clean up what we did in this parse */
7945 SvREFCNT_dec_NN(RExC_rx_sv);
7950 /* Here, we have successfully parsed and generated the pattern's program
7951 * for the regex engine. We are ready to finish things up and look for
7954 /* Update the string to compile, with correct modifiers, etc */
7955 set_regex_pv(pRExC_state, Rx);
7957 RExC_rx->nparens = RExC_total_parens - 1;
7959 /* Uses the upper 4 bits of the FLAGS field, so keep within that size */
7960 if (RExC_whilem_seen > 15)
7961 RExC_whilem_seen = 15;
7964 Perl_re_printf( aTHX_
7965 "Required size %" IVdf " nodes\n", (IV)RExC_size);
7967 RExC_lastparse=NULL;
7970 #ifdef RE_TRACK_PATTERN_OFFSETS
7971 DEBUG_OFFSETS_r(Perl_re_printf( aTHX_
7972 "%s %" UVuf " bytes for offset annotations.\n",
7973 RExC_offsets ? "Got" : "Couldn't get",
7974 (UV)((RExC_offsets[0] * 2 + 1))));
7975 DEBUG_OFFSETS_r(if (RExC_offsets) {
7976 const STRLEN len = RExC_offsets[0];
7978 DECLARE_AND_GET_RE_DEBUG_FLAGS;
7979 Perl_re_printf( aTHX_
7980 "Offsets: [%" UVuf "]\n\t", (UV)RExC_offsets[0]);
7981 for (i = 1; i <= len; i++) {
7982 if (RExC_offsets[i*2-1] || RExC_offsets[i*2])
7983 Perl_re_printf( aTHX_ "%" UVuf ":%" UVuf "[%" UVuf "] ",
7984 (UV)i, (UV)RExC_offsets[i*2-1], (UV)RExC_offsets[i*2]);
7986 Perl_re_printf( aTHX_ "\n");
7990 SetProgLen(RExC_rxi,RExC_size);
7993 DEBUG_DUMP_PRE_OPTIMIZE_r({
7994 SV * const sv = sv_newmortal();
7995 RXi_GET_DECL(RExC_rx, ri);
7997 Perl_re_printf( aTHX_ "Program before optimization:\n");
7999 (void)dumpuntil(RExC_rx, ri->program, ri->program + 1, NULL, NULL,
8004 Perl_re_printf( aTHX_ "Starting post parse optimization\n");
8007 /* XXXX To minimize changes to RE engine we always allocate
8008 3-units-long substrs field. */
8009 Newx(RExC_rx->substrs, 1, struct reg_substr_data);
8010 if (RExC_recurse_count) {
8011 Newx(RExC_recurse, RExC_recurse_count, regnode *);
8012 SAVEFREEPV(RExC_recurse);
8015 if (RExC_seen & REG_RECURSE_SEEN) {
8016 /* Note, RExC_total_parens is 1 + the number of parens in a pattern.
8017 * So its 1 if there are no parens. */
8018 RExC_study_chunk_recursed_bytes= (RExC_total_parens >> 3) +
8019 ((RExC_total_parens & 0x07) != 0);
8020 Newx(RExC_study_chunk_recursed,
8021 RExC_study_chunk_recursed_bytes * RExC_total_parens, U8);
8022 SAVEFREEPV(RExC_study_chunk_recursed);
8026 RExC_rx->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
8028 RExC_study_chunk_recursed_count= 0;
8030 Zero(RExC_rx->substrs, 1, struct reg_substr_data);
8031 if (RExC_study_chunk_recursed) {
8032 Zero(RExC_study_chunk_recursed,
8033 RExC_study_chunk_recursed_bytes * RExC_total_parens, U8);
8037 #ifdef TRIE_STUDY_OPT
8039 StructCopy(&zero_scan_data, &data, scan_data_t);
8040 copyRExC_state = RExC_state;
8043 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "Restudying\n"));
8045 RExC_state = copyRExC_state;
8046 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
8047 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
8049 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
8050 StructCopy(&zero_scan_data, &data, scan_data_t);
8053 StructCopy(&zero_scan_data, &data, scan_data_t);
8056 /* Dig out information for optimizations. */
8057 RExC_rx->extflags = RExC_flags; /* was pm_op */
8058 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
8061 SvUTF8_on(Rx); /* Unicode in it? */
8062 RExC_rxi->regstclass = NULL;
8063 if (RExC_naughty >= TOO_NAUGHTY) /* Probably an expensive pattern. */
8064 RExC_rx->intflags |= PREGf_NAUGHTY;
8065 scan = RExC_rxi->program + 1; /* First BRANCH. */
8067 /* testing for BRANCH here tells us whether there is "must appear"
8068 data in the pattern. If there is then we can use it for optimisations */
8069 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
8072 STRLEN longest_length[2];
8073 regnode_ssc ch_class; /* pointed to by data */
8075 SSize_t last_close = 0; /* pointed to by data */
8076 regnode *first= scan;
8077 regnode *first_next= regnext(first);
8081 * Skip introductions and multiplicators >= 1
8082 * so that we can extract the 'meat' of the pattern that must
8083 * match in the large if() sequence following.
8084 * NOTE that EXACT is NOT covered here, as it is normally
8085 * picked up by the optimiser separately.
8087 * This is unfortunate as the optimiser isnt handling lookahead
8088 * properly currently.
8091 while ((OP(first) == OPEN && (sawopen = 1)) ||
8092 /* An OR of *one* alternative - should not happen now. */
8093 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
8094 /* for now we can't handle lookbehind IFMATCH*/
8095 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
8096 (OP(first) == PLUS) ||
8097 (OP(first) == MINMOD) ||
8098 /* An {n,m} with n>0 */
8099 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
8100 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
8103 * the only op that could be a regnode is PLUS, all the rest
8104 * will be regnode_1 or regnode_2.
8106 * (yves doesn't think this is true)
8108 if (OP(first) == PLUS)
8111 if (OP(first) == MINMOD)
8113 first += regarglen[OP(first)];
8115 first = NEXTOPER(first);
8116 first_next= regnext(first);
8119 /* Starting-point info. */
8121 DEBUG_PEEP("first:", first, 0, 0);
8122 /* Ignore EXACT as we deal with it later. */
8123 if (PL_regkind[OP(first)] == EXACT) {
8124 if (! isEXACTFish(OP(first))) {
8125 NOOP; /* Empty, get anchored substr later. */
8128 RExC_rxi->regstclass = first;
8131 else if (PL_regkind[OP(first)] == TRIE &&
8132 ((reg_trie_data *)RExC_rxi->data->data[ ARG(first) ])->minlen>0)
8134 /* this can happen only on restudy */
8135 RExC_rxi->regstclass = construct_ahocorasick_from_trie(pRExC_state, (regnode *)first, 0);
8138 else if (REGNODE_SIMPLE(OP(first)))
8139 RExC_rxi->regstclass = first;
8140 else if (PL_regkind[OP(first)] == BOUND ||
8141 PL_regkind[OP(first)] == NBOUND)
8142 RExC_rxi->regstclass = first;
8143 else if (PL_regkind[OP(first)] == BOL) {
8144 RExC_rx->intflags |= (OP(first) == MBOL
8147 first = NEXTOPER(first);
8150 else if (OP(first) == GPOS) {
8151 RExC_rx->intflags |= PREGf_ANCH_GPOS;
8152 first = NEXTOPER(first);
8155 else if ((!sawopen || !RExC_sawback) &&
8157 (OP(first) == STAR &&
8158 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
8159 !(RExC_rx->intflags & PREGf_ANCH) && !pRExC_state->code_blocks)
8161 /* turn .* into ^.* with an implied $*=1 */
8163 (OP(NEXTOPER(first)) == REG_ANY)
8166 RExC_rx->intflags |= (type | PREGf_IMPLICIT);
8167 first = NEXTOPER(first);
8170 if (sawplus && !sawminmod && !sawlookahead
8171 && (!sawopen || !RExC_sawback)
8172 && !pRExC_state->code_blocks) /* May examine pos and $& */
8173 /* x+ must match at the 1st pos of run of x's */
8174 RExC_rx->intflags |= PREGf_SKIP;
8176 /* Scan is after the zeroth branch, first is atomic matcher. */
8177 #ifdef TRIE_STUDY_OPT
8180 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
8181 (IV)(first - scan + 1))
8185 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
8186 (IV)(first - scan + 1))
8192 * If there's something expensive in the r.e., find the
8193 * longest literal string that must appear and make it the
8194 * regmust. Resolve ties in favor of later strings, since
8195 * the regstart check works with the beginning of the r.e.
8196 * and avoiding duplication strengthens checking. Not a
8197 * strong reason, but sufficient in the absence of others.
8198 * [Now we resolve ties in favor of the earlier string if
8199 * it happens that c_offset_min has been invalidated, since the
8200 * earlier string may buy us something the later one won't.]
8203 data.substrs[0].str = newSVpvs("");
8204 data.substrs[1].str = newSVpvs("");
8205 data.last_found = newSVpvs("");
8206 data.cur_is_floating = 0; /* initially any found substring is fixed */
8207 ENTER_with_name("study_chunk");
8208 SAVEFREESV(data.substrs[0].str);
8209 SAVEFREESV(data.substrs[1].str);
8210 SAVEFREESV(data.last_found);
8212 if (!RExC_rxi->regstclass) {
8213 ssc_init(pRExC_state, &ch_class);
8214 data.start_class = &ch_class;
8215 stclass_flag = SCF_DO_STCLASS_AND;
8216 } else /* XXXX Check for BOUND? */
8218 data.last_closep = &last_close;
8222 * MAIN ENTRY FOR study_chunk() FOR m/PATTERN/
8223 * (NO top level branches)
8225 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
8226 scan + RExC_size, /* Up to end */
8228 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
8229 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
8233 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
8236 if ( RExC_total_parens == 1 && !data.cur_is_floating
8237 && data.last_start_min == 0 && data.last_end > 0
8238 && !RExC_seen_zerolen
8239 && !(RExC_seen & REG_VERBARG_SEEN)
8240 && !(RExC_seen & REG_GPOS_SEEN)
8242 RExC_rx->extflags |= RXf_CHECK_ALL;
8244 scan_commit(pRExC_state, &data,&minlen, 0);
8247 /* XXX this is done in reverse order because that's the way the
8248 * code was before it was parameterised. Don't know whether it
8249 * actually needs doing in reverse order. DAPM */
8250 for (i = 1; i >= 0; i--) {
8251 longest_length[i] = CHR_SVLEN(data.substrs[i].str);
8254 && SvCUR(data.substrs[0].str) /* ok to leave SvCUR */
8255 && data.substrs[0].min_offset
8256 == data.substrs[1].min_offset
8257 && SvCUR(data.substrs[0].str)
8258 == SvCUR(data.substrs[1].str)
8260 && S_setup_longest (aTHX_ pRExC_state,
8261 &(RExC_rx->substrs->data[i]),
8265 RExC_rx->substrs->data[i].min_offset =
8266 data.substrs[i].min_offset - data.substrs[i].lookbehind;
8268 RExC_rx->substrs->data[i].max_offset = data.substrs[i].max_offset;
8269 /* Don't offset infinity */
8270 if (data.substrs[i].max_offset < OPTIMIZE_INFTY)
8271 RExC_rx->substrs->data[i].max_offset -= data.substrs[i].lookbehind;
8272 SvREFCNT_inc_simple_void_NN(data.substrs[i].str);
8275 RExC_rx->substrs->data[i].substr = NULL;
8276 RExC_rx->substrs->data[i].utf8_substr = NULL;
8277 longest_length[i] = 0;
8281 LEAVE_with_name("study_chunk");
8283 if (RExC_rxi->regstclass
8284 && (OP(RExC_rxi->regstclass) == REG_ANY || OP(RExC_rxi->regstclass) == SANY))
8285 RExC_rxi->regstclass = NULL;
8287 if ((!(RExC_rx->substrs->data[0].substr || RExC_rx->substrs->data[0].utf8_substr)
8288 || RExC_rx->substrs->data[0].min_offset)
8290 && ! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
8291 && is_ssc_worth_it(pRExC_state, data.start_class))
8293 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
8295 ssc_finalize(pRExC_state, data.start_class);
8297 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
8298 StructCopy(data.start_class,
8299 (regnode_ssc*)RExC_rxi->data->data[n],
8301 RExC_rxi->regstclass = (regnode*)RExC_rxi->data->data[n];
8302 RExC_rx->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
8303 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
8304 regprop(RExC_rx, sv, (regnode*)data.start_class, NULL, pRExC_state);
8305 Perl_re_printf( aTHX_
8306 "synthetic stclass \"%s\".\n",
8307 SvPVX_const(sv));});
8308 data.start_class = NULL;
8311 /* A temporary algorithm prefers floated substr to fixed one of
8312 * same length to dig more info. */
8313 i = (longest_length[0] <= longest_length[1]);
8314 RExC_rx->substrs->check_ix = i;
8315 RExC_rx->check_end_shift = RExC_rx->substrs->data[i].end_shift;
8316 RExC_rx->check_substr = RExC_rx->substrs->data[i].substr;
8317 RExC_rx->check_utf8 = RExC_rx->substrs->data[i].utf8_substr;
8318 RExC_rx->check_offset_min = RExC_rx->substrs->data[i].min_offset;
8319 RExC_rx->check_offset_max = RExC_rx->substrs->data[i].max_offset;
8320 if (!i && (RExC_rx->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS)))
8321 RExC_rx->intflags |= PREGf_NOSCAN;
8323 if ((RExC_rx->check_substr || RExC_rx->check_utf8) ) {
8324 RExC_rx->extflags |= RXf_USE_INTUIT;
8325 if (SvTAIL(RExC_rx->check_substr ? RExC_rx->check_substr : RExC_rx->check_utf8))
8326 RExC_rx->extflags |= RXf_INTUIT_TAIL;
8329 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
8330 if ( (STRLEN)minlen < longest_length[1] )
8331 minlen= longest_length[1];
8332 if ( (STRLEN)minlen < longest_length[0] )
8333 minlen= longest_length[0];
8337 /* Several toplevels. Best we can is to set minlen. */
8339 regnode_ssc ch_class;
8340 SSize_t last_close = 0;
8342 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "\nMulti Top Level\n"));
8344 scan = RExC_rxi->program + 1;
8345 ssc_init(pRExC_state, &ch_class);
8346 data.start_class = &ch_class;
8347 data.last_closep = &last_close;
8351 * MAIN ENTRY FOR study_chunk() FOR m/P1|P2|.../
8352 * (patterns WITH top level branches)
8354 minlen = study_chunk(pRExC_state,
8355 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
8356 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
8357 ? SCF_TRIE_DOING_RESTUDY
8361 CHECK_RESTUDY_GOTO_butfirst(NOOP);
8363 RExC_rx->check_substr = NULL;
8364 RExC_rx->check_utf8 = NULL;
8365 RExC_rx->substrs->data[0].substr = NULL;
8366 RExC_rx->substrs->data[0].utf8_substr = NULL;
8367 RExC_rx->substrs->data[1].substr = NULL;
8368 RExC_rx->substrs->data[1].utf8_substr = NULL;
8370 if (! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
8371 && is_ssc_worth_it(pRExC_state, data.start_class))
8373 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
8375 ssc_finalize(pRExC_state, data.start_class);
8377 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
8378 StructCopy(data.start_class,
8379 (regnode_ssc*)RExC_rxi->data->data[n],
8381 RExC_rxi->regstclass = (regnode*)RExC_rxi->data->data[n];
8382 RExC_rx->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
8383 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
8384 regprop(RExC_rx, sv, (regnode*)data.start_class, NULL, pRExC_state);
8385 Perl_re_printf( aTHX_
8386 "synthetic stclass \"%s\".\n",
8387 SvPVX_const(sv));});
8388 data.start_class = NULL;
8392 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
8393 RExC_rx->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
8394 RExC_rx->maxlen = REG_INFTY;
8397 RExC_rx->maxlen = RExC_maxlen;
8400 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
8401 the "real" pattern. */
8403 Perl_re_printf( aTHX_ "minlen: %" IVdf " RExC_rx->minlen:%" IVdf " maxlen:%" IVdf "\n",
8404 (IV)minlen, (IV)RExC_rx->minlen, (IV)RExC_maxlen);
8406 RExC_rx->minlenret = minlen;
8407 if (RExC_rx->minlen < minlen)
8408 RExC_rx->minlen = minlen;
8410 if (RExC_seen & REG_RECURSE_SEEN ) {
8411 RExC_rx->intflags |= PREGf_RECURSE_SEEN;
8412 Newx(RExC_rx->recurse_locinput, RExC_rx->nparens + 1, char *);
8414 if (RExC_seen & REG_GPOS_SEEN)
8415 RExC_rx->intflags |= PREGf_GPOS_SEEN;
8416 if (RExC_seen & REG_LOOKBEHIND_SEEN)
8417 RExC_rx->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
8419 if (pRExC_state->code_blocks)
8420 RExC_rx->extflags |= RXf_EVAL_SEEN;
8421 if (RExC_seen & REG_VERBARG_SEEN)
8423 RExC_rx->intflags |= PREGf_VERBARG_SEEN;
8424 RExC_rx->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
8426 if (RExC_seen & REG_CUTGROUP_SEEN)
8427 RExC_rx->intflags |= PREGf_CUTGROUP_SEEN;
8428 if (pm_flags & PMf_USE_RE_EVAL)
8429 RExC_rx->intflags |= PREGf_USE_RE_EVAL;
8430 if (RExC_paren_names)
8431 RXp_PAREN_NAMES(RExC_rx) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
8433 RXp_PAREN_NAMES(RExC_rx) = NULL;
8435 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
8436 * so it can be used in pp.c */
8437 if (RExC_rx->intflags & PREGf_ANCH)
8438 RExC_rx->extflags |= RXf_IS_ANCHORED;
8442 /* this is used to identify "special" patterns that might result
8443 * in Perl NOT calling the regex engine and instead doing the match "itself",
8444 * particularly special cases in split//. By having the regex compiler
8445 * do this pattern matching at a regop level (instead of by inspecting the pattern)
8446 * we avoid weird issues with equivalent patterns resulting in different behavior,
8447 * AND we allow non Perl engines to get the same optimizations by the setting the
8448 * flags appropriately - Yves */
8449 regnode *first = RExC_rxi->program + 1;
8451 regnode *next = NEXTOPER(first);
8452 /* It's safe to read through *next only if OP(first) is a regop of
8453 * the right type (not EXACT, for example).
8455 U8 nop = (fop == NOTHING || fop == MBOL || fop == SBOL || fop == PLUS)
8458 if (PL_regkind[fop] == NOTHING && nop == END)
8459 RExC_rx->extflags |= RXf_NULL;
8460 else if ((fop == MBOL || (fop == SBOL && !first->flags)) && nop == END)
8461 /* when fop is SBOL first->flags will be true only when it was
8462 * produced by parsing /\A/, and not when parsing /^/. This is
8463 * very important for the split code as there we want to
8464 * treat /^/ as /^/m, but we do not want to treat /\A/ as /^/m.
8465 * See rt #122761 for more details. -- Yves */
8466 RExC_rx->extflags |= RXf_START_ONLY;
8467 else if (fop == PLUS
8468 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
8469 && OP(regnext(first)) == END)
8470 RExC_rx->extflags |= RXf_WHITE;
8471 else if ( RExC_rx->extflags & RXf_SPLIT
8472 && (PL_regkind[fop] == EXACT && ! isEXACTFish(fop))
8473 && STR_LEN(first) == 1
8474 && *(STRING(first)) == ' '
8475 && OP(regnext(first)) == END )
8476 RExC_rx->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
8480 if (RExC_contains_locale) {
8481 RXp_EXTFLAGS(RExC_rx) |= RXf_TAINTED;
8485 if (RExC_paren_names) {
8486 RExC_rxi->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
8487 RExC_rxi->data->data[RExC_rxi->name_list_idx]
8488 = (void*)SvREFCNT_inc(RExC_paren_name_list);
8491 RExC_rxi->name_list_idx = 0;
8493 while ( RExC_recurse_count > 0 ) {
8494 const regnode *scan = RExC_recurse[ --RExC_recurse_count ];
8496 * This data structure is set up in study_chunk() and is used
8497 * to calculate the distance between a GOSUB regopcode and
8498 * the OPEN/CURLYM (CURLYM's are special and can act like OPEN's)
8501 * If for some reason someone writes code that optimises
8502 * away a GOSUB opcode then the assert should be changed to
8503 * an if(scan) to guard the ARG2L_SET() - Yves
8506 assert(scan && OP(scan) == GOSUB);
8507 ARG2L_SET( scan, RExC_open_parens[ARG(scan)] - REGNODE_OFFSET(scan));
8510 Newxz(RExC_rx->offs, RExC_total_parens, regexp_paren_pair);
8511 /* assume we don't need to swap parens around before we match */
8513 Perl_re_printf( aTHX_ "study_chunk_recursed_count: %lu\n",
8514 (unsigned long)RExC_study_chunk_recursed_count);
8518 Perl_re_printf( aTHX_ "Final program:\n");
8522 if (RExC_open_parens) {
8523 Safefree(RExC_open_parens);
8524 RExC_open_parens = NULL;
8526 if (RExC_close_parens) {
8527 Safefree(RExC_close_parens);
8528 RExC_close_parens = NULL;
8532 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
8533 * by setting the regexp SV to readonly-only instead. If the
8534 * pattern's been recompiled, the USEDness should remain. */
8535 if (old_re && SvREADONLY(old_re))
8543 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
8546 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
8548 PERL_UNUSED_ARG(value);
8550 if (flags & RXapif_FETCH) {
8551 return reg_named_buff_fetch(rx, key, flags);
8552 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
8553 Perl_croak_no_modify();
8555 } else if (flags & RXapif_EXISTS) {
8556 return reg_named_buff_exists(rx, key, flags)
8559 } else if (flags & RXapif_REGNAMES) {
8560 return reg_named_buff_all(rx, flags);
8561 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
8562 return reg_named_buff_scalar(rx, flags);
8564 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
8570 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
8573 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
8574 PERL_UNUSED_ARG(lastkey);
8576 if (flags & RXapif_FIRSTKEY)
8577 return reg_named_buff_firstkey(rx, flags);
8578 else if (flags & RXapif_NEXTKEY)
8579 return reg_named_buff_nextkey(rx, flags);
8581 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
8588 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
8592 struct regexp *const rx = ReANY(r);
8594 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
8596 if (rx && RXp_PAREN_NAMES(rx)) {
8597 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
8600 SV* sv_dat=HeVAL(he_str);
8601 I32 *nums=(I32*)SvPVX(sv_dat);
8602 AV * const retarray = (flags & RXapif_ALL) ? newAV() : NULL;
8603 for ( i=0; i<SvIVX(sv_dat); i++ ) {
8604 if ((I32)(rx->nparens) >= nums[i]
8605 && rx->offs[nums[i]].start != -1
8606 && rx->offs[nums[i]].end != -1)
8609 CALLREG_NUMBUF_FETCH(r, nums[i], ret);
8614 ret = newSVsv(&PL_sv_undef);
8617 av_push(retarray, ret);
8620 return newRV_noinc(MUTABLE_SV(retarray));
8627 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
8630 struct regexp *const rx = ReANY(r);
8632 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
8634 if (rx && RXp_PAREN_NAMES(rx)) {
8635 if (flags & RXapif_ALL) {
8636 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
8638 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
8640 SvREFCNT_dec_NN(sv);
8652 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
8654 struct regexp *const rx = ReANY(r);
8656 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
8658 if ( rx && RXp_PAREN_NAMES(rx) ) {
8659 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
8661 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
8668 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
8670 struct regexp *const rx = ReANY(r);
8671 DECLARE_AND_GET_RE_DEBUG_FLAGS;
8673 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
8675 if (rx && RXp_PAREN_NAMES(rx)) {
8676 HV *hv = RXp_PAREN_NAMES(rx);
8678 while ( (temphe = hv_iternext_flags(hv, 0)) ) {
8681 SV* sv_dat = HeVAL(temphe);
8682 I32 *nums = (I32*)SvPVX(sv_dat);
8683 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8684 if ((I32)(rx->lastparen) >= nums[i] &&
8685 rx->offs[nums[i]].start != -1 &&
8686 rx->offs[nums[i]].end != -1)
8692 if (parno || flags & RXapif_ALL) {
8693 return newSVhek(HeKEY_hek(temphe));
8701 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
8706 struct regexp *const rx = ReANY(r);
8708 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
8710 if (rx && RXp_PAREN_NAMES(rx)) {
8711 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
8712 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
8713 } else if (flags & RXapif_ONE) {
8714 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
8715 av = MUTABLE_AV(SvRV(ret));
8716 length = av_count(av);
8717 SvREFCNT_dec_NN(ret);
8718 return newSViv(length);
8720 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
8725 return &PL_sv_undef;
8729 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
8731 struct regexp *const rx = ReANY(r);
8734 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
8736 if (rx && RXp_PAREN_NAMES(rx)) {
8737 HV *hv= RXp_PAREN_NAMES(rx);
8739 (void)hv_iterinit(hv);
8740 while ( (temphe = hv_iternext_flags(hv, 0)) ) {
8743 SV* sv_dat = HeVAL(temphe);
8744 I32 *nums = (I32*)SvPVX(sv_dat);
8745 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8746 if ((I32)(rx->lastparen) >= nums[i] &&
8747 rx->offs[nums[i]].start != -1 &&
8748 rx->offs[nums[i]].end != -1)
8754 if (parno || flags & RXapif_ALL) {
8755 av_push(av, newSVhek(HeKEY_hek(temphe)));
8760 return newRV_noinc(MUTABLE_SV(av));
8764 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
8767 struct regexp *const rx = ReANY(r);
8773 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
8775 if ( n == RX_BUFF_IDX_CARET_PREMATCH
8776 || n == RX_BUFF_IDX_CARET_FULLMATCH
8777 || n == RX_BUFF_IDX_CARET_POSTMATCH
8780 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8782 /* on something like
8785 * the KEEPCOPY is set on the PMOP rather than the regex */
8786 if (PL_curpm && r == PM_GETRE(PL_curpm))
8787 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8796 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
8797 /* no need to distinguish between them any more */
8798 n = RX_BUFF_IDX_FULLMATCH;
8800 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
8801 && rx->offs[0].start != -1)
8803 /* $`, ${^PREMATCH} */
8804 i = rx->offs[0].start;
8808 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
8809 && rx->offs[0].end != -1)
8811 /* $', ${^POSTMATCH} */
8812 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
8813 i = rx->sublen + rx->suboffset - rx->offs[0].end;
8816 if (inRANGE(n, 0, (I32)rx->nparens) &&
8817 (s1 = rx->offs[n].start) != -1 &&
8818 (t1 = rx->offs[n].end) != -1)
8820 /* $&, ${^MATCH}, $1 ... */
8822 s = rx->subbeg + s1 - rx->suboffset;
8827 assert(s >= rx->subbeg);
8828 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
8830 #ifdef NO_TAINT_SUPPORT
8831 sv_setpvn(sv, s, i);
8833 const int oldtainted = TAINT_get;
8835 sv_setpvn(sv, s, i);
8836 TAINT_set(oldtainted);
8838 if (RXp_MATCH_UTF8(rx))
8843 if (RXp_MATCH_TAINTED(rx)) {
8844 if (SvTYPE(sv) >= SVt_PVMG) {
8845 MAGIC* const mg = SvMAGIC(sv);
8848 SvMAGIC_set(sv, mg->mg_moremagic);
8850 if ((mgt = SvMAGIC(sv))) {
8851 mg->mg_moremagic = mgt;
8852 SvMAGIC_set(sv, mg);
8869 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
8870 SV const * const value)
8872 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
8874 PERL_UNUSED_ARG(rx);
8875 PERL_UNUSED_ARG(paren);
8876 PERL_UNUSED_ARG(value);
8879 Perl_croak_no_modify();
8883 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
8886 struct regexp *const rx = ReANY(r);
8890 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
8892 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
8893 || paren == RX_BUFF_IDX_CARET_FULLMATCH
8894 || paren == RX_BUFF_IDX_CARET_POSTMATCH
8897 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8899 /* on something like
8902 * the KEEPCOPY is set on the PMOP rather than the regex */
8903 if (PL_curpm && r == PM_GETRE(PL_curpm))
8904 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8910 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
8912 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
8913 case RX_BUFF_IDX_PREMATCH: /* $` */
8914 if (rx->offs[0].start != -1) {
8915 i = rx->offs[0].start;
8924 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
8925 case RX_BUFF_IDX_POSTMATCH: /* $' */
8926 if (rx->offs[0].end != -1) {
8927 i = rx->sublen - rx->offs[0].end;
8929 s1 = rx->offs[0].end;
8936 default: /* $& / ${^MATCH}, $1, $2, ... */
8937 if (paren <= (I32)rx->nparens &&
8938 (s1 = rx->offs[paren].start) != -1 &&
8939 (t1 = rx->offs[paren].end) != -1)
8945 if (ckWARN(WARN_UNINITIALIZED))
8946 report_uninit((const SV *)sv);
8951 if (i > 0 && RXp_MATCH_UTF8(rx)) {
8952 const char * const s = rx->subbeg - rx->suboffset + s1;
8957 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
8964 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
8966 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
8967 PERL_UNUSED_ARG(rx);
8971 return newSVpvs("Regexp");
8974 /* Scans the name of a named buffer from the pattern.
8975 * If flags is REG_RSN_RETURN_NULL returns null.
8976 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
8977 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
8978 * to the parsed name as looked up in the RExC_paren_names hash.
8979 * If there is an error throws a vFAIL().. type exception.
8982 #define REG_RSN_RETURN_NULL 0
8983 #define REG_RSN_RETURN_NAME 1
8984 #define REG_RSN_RETURN_DATA 2
8987 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
8989 char *name_start = RExC_parse;
8992 PERL_ARGS_ASSERT_REG_SCAN_NAME;
8994 assert (RExC_parse <= RExC_end);
8995 if (RExC_parse == RExC_end) NOOP;
8996 else if (isIDFIRST_lazy_if_safe(RExC_parse, RExC_end, UTF)) {
8997 /* Note that the code here assumes well-formed UTF-8. Skip IDFIRST by
8998 * using do...while */
9001 RExC_parse += UTF8SKIP(RExC_parse);
9002 } while ( RExC_parse < RExC_end
9003 && isWORDCHAR_utf8_safe((U8*)RExC_parse, (U8*) RExC_end));
9007 } while (RExC_parse < RExC_end && isWORDCHAR(*RExC_parse));
9009 RExC_parse++; /* so the <- from the vFAIL is after the offending
9011 vFAIL("Group name must start with a non-digit word character");
9013 sv_name = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
9014 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
9015 if ( flags == REG_RSN_RETURN_NAME)
9017 else if (flags==REG_RSN_RETURN_DATA) {
9020 if ( ! sv_name ) /* should not happen*/
9021 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
9022 if (RExC_paren_names)
9023 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
9025 sv_dat = HeVAL(he_str);
9026 if ( ! sv_dat ) { /* Didn't find group */
9028 /* It might be a forward reference; we can't fail until we
9029 * know, by completing the parse to get all the groups, and
9031 if (ALL_PARENS_COUNTED) {
9032 vFAIL("Reference to nonexistent named group");
9035 REQUIRE_PARENS_PASS;
9041 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
9042 (unsigned long) flags);
9045 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
9046 if (RExC_lastparse!=RExC_parse) { \
9047 Perl_re_printf( aTHX_ "%s", \
9048 Perl_pv_pretty(aTHX_ RExC_mysv1, RExC_parse, \
9049 RExC_end - RExC_parse, 16, \
9051 PERL_PV_ESCAPE_UNI_DETECT | \
9052 PERL_PV_PRETTY_ELLIPSES | \
9053 PERL_PV_PRETTY_LTGT | \
9054 PERL_PV_ESCAPE_RE | \
9055 PERL_PV_PRETTY_EXACTSIZE \
9059 Perl_re_printf( aTHX_ "%16s",""); \
9061 if (RExC_lastnum!=RExC_emit) \
9062 Perl_re_printf( aTHX_ "|%4zu", RExC_emit); \
9064 Perl_re_printf( aTHX_ "|%4s",""); \
9065 Perl_re_printf( aTHX_ "|%*s%-4s", \
9066 (int)((depth*2)), "", \
9069 RExC_lastnum=RExC_emit; \
9070 RExC_lastparse=RExC_parse; \
9075 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
9076 DEBUG_PARSE_MSG((funcname)); \
9077 Perl_re_printf( aTHX_ "%4s","\n"); \
9079 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({\
9080 DEBUG_PARSE_MSG((funcname)); \
9081 Perl_re_printf( aTHX_ fmt "\n",args); \
9084 /* This section of code defines the inversion list object and its methods. The
9085 * interfaces are highly subject to change, so as much as possible is static to
9086 * this file. An inversion list is here implemented as a malloc'd C UV array
9087 * as an SVt_INVLIST scalar.
9089 * An inversion list for Unicode is an array of code points, sorted by ordinal
9090 * number. Each element gives the code point that begins a range that extends
9091 * up-to but not including the code point given by the next element. The final
9092 * element gives the first code point of a range that extends to the platform's
9093 * infinity. The even-numbered elements (invlist[0], invlist[2], invlist[4],
9094 * ...) give ranges whose code points are all in the inversion list. We say
9095 * that those ranges are in the set. The odd-numbered elements give ranges
9096 * whose code points are not in the inversion list, and hence not in the set.
9097 * Thus, element [0] is the first code point in the list. Element [1]
9098 * is the first code point beyond that not in the list; and element [2] is the
9099 * first code point beyond that that is in the list. In other words, the first
9100 * range is invlist[0]..(invlist[1]-1), and all code points in that range are
9101 * in the inversion list. The second range is invlist[1]..(invlist[2]-1), and
9102 * all code points in that range are not in the inversion list. The third
9103 * range invlist[2]..(invlist[3]-1) gives code points that are in the inversion
9104 * list, and so forth. Thus every element whose index is divisible by two
9105 * gives the beginning of a range that is in the list, and every element whose
9106 * index is not divisible by two gives the beginning of a range not in the
9107 * list. If the final element's index is divisible by two, the inversion list
9108 * extends to the platform's infinity; otherwise the highest code point in the
9109 * inversion list is the contents of that element minus 1.
9111 * A range that contains just a single code point N will look like
9113 * invlist[i+1] == N+1
9115 * If N is UV_MAX (the highest representable code point on the machine), N+1 is
9116 * impossible to represent, so element [i+1] is omitted. The single element
9118 * invlist[0] == UV_MAX
9119 * contains just UV_MAX, but is interpreted as matching to infinity.
9121 * Taking the complement (inverting) an inversion list is quite simple, if the
9122 * first element is 0, remove it; otherwise add a 0 element at the beginning.
9123 * This implementation reserves an element at the beginning of each inversion
9124 * list to always contain 0; there is an additional flag in the header which
9125 * indicates if the list begins at the 0, or is offset to begin at the next
9126 * element. This means that the inversion list can be inverted without any
9127 * copying; just flip the flag.
9129 * More about inversion lists can be found in "Unicode Demystified"
9130 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
9132 * The inversion list data structure is currently implemented as an SV pointing
9133 * to an array of UVs that the SV thinks are bytes. This allows us to have an
9134 * array of UV whose memory management is automatically handled by the existing
9135 * facilities for SV's.
9137 * Some of the methods should always be private to the implementation, and some
9138 * should eventually be made public */
9140 /* The header definitions are in F<invlist_inline.h> */
9142 #ifndef PERL_IN_XSUB_RE
9144 PERL_STATIC_INLINE UV*
9145 S__invlist_array_init(SV* const invlist, const bool will_have_0)
9147 /* Returns a pointer to the first element in the inversion list's array.
9148 * This is called upon initialization of an inversion list. Where the
9149 * array begins depends on whether the list has the code point U+0000 in it
9150 * or not. The other parameter tells it whether the code that follows this
9151 * call is about to put a 0 in the inversion list or not. The first
9152 * element is either the element reserved for 0, if TRUE, or the element
9153 * after it, if FALSE */
9155 bool* offset = get_invlist_offset_addr(invlist);
9156 UV* zero_addr = (UV *) SvPVX(invlist);
9158 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
9161 assert(! _invlist_len(invlist));
9165 /* 1^1 = 0; 1^0 = 1 */
9166 *offset = 1 ^ will_have_0;
9167 return zero_addr + *offset;
9171 S_invlist_replace_list_destroys_src(pTHX_ SV * dest, SV * src)
9173 /* Replaces the inversion list in 'dest' with the one from 'src'. It
9174 * steals the list from 'src', so 'src' is made to have a NULL list. This
9175 * is similar to what SvSetMagicSV() would do, if it were implemented on
9176 * inversion lists, though this routine avoids a copy */
9178 const UV src_len = _invlist_len(src);
9179 const bool src_offset = *get_invlist_offset_addr(src);
9180 const STRLEN src_byte_len = SvLEN(src);
9181 char * array = SvPVX(src);
9183 const int oldtainted = TAINT_get;
9185 PERL_ARGS_ASSERT_INVLIST_REPLACE_LIST_DESTROYS_SRC;
9187 assert(is_invlist(src));
9188 assert(is_invlist(dest));
9189 assert(! invlist_is_iterating(src));
9190 assert(SvCUR(src) == 0 || SvCUR(src) < SvLEN(src));
9192 /* Make sure it ends in the right place with a NUL, as our inversion list
9193 * manipulations aren't careful to keep this true, but sv_usepvn_flags()
9195 array[src_byte_len - 1] = '\0';
9197 TAINT_NOT; /* Otherwise it breaks */
9198 sv_usepvn_flags(dest,
9202 /* This flag is documented to cause a copy to be avoided */
9203 SV_HAS_TRAILING_NUL);
9204 TAINT_set(oldtainted);
9209 /* Finish up copying over the other fields in an inversion list */
9210 *get_invlist_offset_addr(dest) = src_offset;
9211 invlist_set_len(dest, src_len, src_offset);
9212 *get_invlist_previous_index_addr(dest) = 0;
9213 invlist_iterfinish(dest);
9216 PERL_STATIC_INLINE IV*
9217 S_get_invlist_previous_index_addr(SV* invlist)
9219 /* Return the address of the IV that is reserved to hold the cached index
9221 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
9223 assert(is_invlist(invlist));
9225 return &(((XINVLIST*) SvANY(invlist))->prev_index);
9228 PERL_STATIC_INLINE IV
9229 S_invlist_previous_index(SV* const invlist)
9231 /* Returns cached index of previous search */
9233 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
9235 return *get_invlist_previous_index_addr(invlist);
9238 PERL_STATIC_INLINE void
9239 S_invlist_set_previous_index(SV* const invlist, const IV index)
9241 /* Caches <index> for later retrieval */
9243 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
9245 assert(index == 0 || index < (int) _invlist_len(invlist));
9247 *get_invlist_previous_index_addr(invlist) = index;
9250 PERL_STATIC_INLINE void
9251 S_invlist_trim(SV* invlist)
9253 /* Free the not currently-being-used space in an inversion list */
9255 /* But don't free up the space needed for the 0 UV that is always at the
9256 * beginning of the list, nor the trailing NUL */
9257 const UV min_size = TO_INTERNAL_SIZE(1) + 1;
9259 PERL_ARGS_ASSERT_INVLIST_TRIM;
9261 assert(is_invlist(invlist));
9263 SvPV_renew(invlist, MAX(min_size, SvCUR(invlist) + 1));
9266 PERL_STATIC_INLINE void
9267 S_invlist_clear(pTHX_ SV* invlist) /* Empty the inversion list */
9269 PERL_ARGS_ASSERT_INVLIST_CLEAR;
9271 assert(is_invlist(invlist));
9273 invlist_set_len(invlist, 0, 0);
9274 invlist_trim(invlist);
9277 #endif /* ifndef PERL_IN_XSUB_RE */
9279 PERL_STATIC_INLINE bool
9280 S_invlist_is_iterating(SV* const invlist)
9282 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
9284 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
9287 #ifndef PERL_IN_XSUB_RE
9289 PERL_STATIC_INLINE UV
9290 S_invlist_max(SV* const invlist)
9292 /* Returns the maximum number of elements storable in the inversion list's
9293 * array, without having to realloc() */
9295 PERL_ARGS_ASSERT_INVLIST_MAX;
9297 assert(is_invlist(invlist));
9299 /* Assumes worst case, in which the 0 element is not counted in the
9300 * inversion list, so subtracts 1 for that */
9301 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
9302 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
9303 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
9307 S_initialize_invlist_guts(pTHX_ SV* invlist, const Size_t initial_size)
9309 PERL_ARGS_ASSERT_INITIALIZE_INVLIST_GUTS;
9311 /* First 1 is in case the zero element isn't in the list; second 1 is for
9313 SvGROW(invlist, TO_INTERNAL_SIZE(initial_size + 1) + 1);
9314 invlist_set_len(invlist, 0, 0);
9316 /* Force iterinit() to be used to get iteration to work */
9317 invlist_iterfinish(invlist);
9319 *get_invlist_previous_index_addr(invlist) = 0;
9320 SvPOK_on(invlist); /* This allows B to extract the PV */
9324 Perl__new_invlist(pTHX_ IV initial_size)
9327 /* Return a pointer to a newly constructed inversion list, with enough
9328 * space to store 'initial_size' elements. If that number is negative, a
9329 * system default is used instead */
9333 if (initial_size < 0) {
9337 new_list = newSV_type(SVt_INVLIST);
9338 initialize_invlist_guts(new_list, initial_size);
9344 Perl__new_invlist_C_array(pTHX_ const UV* const list)
9346 /* Return a pointer to a newly constructed inversion list, initialized to
9347 * point to <list>, which has to be in the exact correct inversion list
9348 * form, including internal fields. Thus this is a dangerous routine that
9349 * should not be used in the wrong hands. The passed in 'list' contains
9350 * several header fields at the beginning that are not part of the
9351 * inversion list body proper */
9353 const STRLEN length = (STRLEN) list[0];
9354 const UV version_id = list[1];
9355 const bool offset = cBOOL(list[2]);
9356 #define HEADER_LENGTH 3
9357 /* If any of the above changes in any way, you must change HEADER_LENGTH
9358 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
9359 * perl -E 'say int(rand 2**31-1)'
9361 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
9362 data structure type, so that one being
9363 passed in can be validated to be an
9364 inversion list of the correct vintage.
9367 SV* invlist = newSV_type(SVt_INVLIST);
9369 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
9371 if (version_id != INVLIST_VERSION_ID) {
9372 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
9375 /* The generated array passed in includes header elements that aren't part
9376 * of the list proper, so start it just after them */
9377 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
9379 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
9380 shouldn't touch it */
9382 *(get_invlist_offset_addr(invlist)) = offset;
9384 /* The 'length' passed to us is the physical number of elements in the
9385 * inversion list. But if there is an offset the logical number is one
9387 invlist_set_len(invlist, length - offset, offset);
9389 invlist_set_previous_index(invlist, 0);
9391 /* Initialize the iteration pointer. */
9392 invlist_iterfinish(invlist);
9394 SvREADONLY_on(invlist);
9401 S__append_range_to_invlist(pTHX_ SV* const invlist,
9402 const UV start, const UV end)
9404 /* Subject to change or removal. Append the range from 'start' to 'end' at
9405 * the end of the inversion list. The range must be above any existing
9409 UV max = invlist_max(invlist);
9410 UV len = _invlist_len(invlist);
9413 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
9415 if (len == 0) { /* Empty lists must be initialized */
9416 offset = start != 0;
9417 array = _invlist_array_init(invlist, ! offset);
9420 /* Here, the existing list is non-empty. The current max entry in the
9421 * list is generally the first value not in the set, except when the
9422 * set extends to the end of permissible values, in which case it is
9423 * the first entry in that final set, and so this call is an attempt to
9424 * append out-of-order */
9426 UV final_element = len - 1;
9427 array = invlist_array(invlist);
9428 if ( array[final_element] > start
9429 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
9431 Perl_croak(aTHX_ "panic: attempting to append to an inversion list, but wasn't at the end of the list, final=%" UVuf ", start=%" UVuf ", match=%c",
9432 array[final_element], start,
9433 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
9436 /* Here, it is a legal append. If the new range begins 1 above the end
9437 * of the range below it, it is extending the range below it, so the
9438 * new first value not in the set is one greater than the newly
9439 * extended range. */
9440 offset = *get_invlist_offset_addr(invlist);
9441 if (array[final_element] == start) {
9442 if (end != UV_MAX) {
9443 array[final_element] = end + 1;
9446 /* But if the end is the maximum representable on the machine,
9447 * assume that infinity was actually what was meant. Just let
9448 * the range that this would extend to have no end */
9449 invlist_set_len(invlist, len - 1, offset);
9455 /* Here the new range doesn't extend any existing set. Add it */
9457 len += 2; /* Includes an element each for the start and end of range */
9459 /* If wll overflow the existing space, extend, which may cause the array to
9462 invlist_extend(invlist, len);
9464 /* Have to set len here to avoid assert failure in invlist_array() */
9465 invlist_set_len(invlist, len, offset);
9467 array = invlist_array(invlist);
9470 invlist_set_len(invlist, len, offset);
9473 /* The next item on the list starts the range, the one after that is
9474 * one past the new range. */
9475 array[len - 2] = start;
9476 if (end != UV_MAX) {
9477 array[len - 1] = end + 1;
9480 /* But if the end is the maximum representable on the machine, just let
9481 * the range have no end */
9482 invlist_set_len(invlist, len - 1, offset);
9487 Perl__invlist_search(SV* const invlist, const UV cp)
9489 /* Searches the inversion list for the entry that contains the input code
9490 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
9491 * return value is the index into the list's array of the range that
9492 * contains <cp>, that is, 'i' such that
9493 * array[i] <= cp < array[i+1]
9498 IV high = _invlist_len(invlist);
9499 const IV highest_element = high - 1;
9502 PERL_ARGS_ASSERT__INVLIST_SEARCH;
9504 /* If list is empty, return failure. */
9509 /* (We can't get the array unless we know the list is non-empty) */
9510 array = invlist_array(invlist);
9512 mid = invlist_previous_index(invlist);
9514 if (mid > highest_element) {
9515 mid = highest_element;
9518 /* <mid> contains the cache of the result of the previous call to this
9519 * function (0 the first time). See if this call is for the same result,
9520 * or if it is for mid-1. This is under the theory that calls to this
9521 * function will often be for related code points that are near each other.
9522 * And benchmarks show that caching gives better results. We also test
9523 * here if the code point is within the bounds of the list. These tests
9524 * replace others that would have had to be made anyway to make sure that
9525 * the array bounds were not exceeded, and these give us extra information
9526 * at the same time */
9527 if (cp >= array[mid]) {
9528 if (cp >= array[highest_element]) {
9529 return highest_element;
9532 /* Here, array[mid] <= cp < array[highest_element]. This means that
9533 * the final element is not the answer, so can exclude it; it also
9534 * means that <mid> is not the final element, so can refer to 'mid + 1'
9536 if (cp < array[mid + 1]) {
9542 else { /* cp < aray[mid] */
9543 if (cp < array[0]) { /* Fail if outside the array */
9547 if (cp >= array[mid - 1]) {
9552 /* Binary search. What we are looking for is <i> such that
9553 * array[i] <= cp < array[i+1]
9554 * The loop below converges on the i+1. Note that there may not be an
9555 * (i+1)th element in the array, and things work nonetheless */
9556 while (low < high) {
9557 mid = (low + high) / 2;
9558 assert(mid <= highest_element);
9559 if (array[mid] <= cp) { /* cp >= array[mid] */
9562 /* We could do this extra test to exit the loop early.
9563 if (cp < array[low]) {
9568 else { /* cp < array[mid] */
9575 invlist_set_previous_index(invlist, high);
9580 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9581 const bool complement_b, SV** output)
9583 /* Take the union of two inversion lists and point '*output' to it. On
9584 * input, '*output' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9585 * even 'a' or 'b'). If to an inversion list, the contents of the original
9586 * list will be replaced by the union. The first list, 'a', may be
9587 * NULL, in which case a copy of the second list is placed in '*output'.
9588 * If 'complement_b' is TRUE, the union is taken of the complement
9589 * (inversion) of 'b' instead of b itself.
9591 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9592 * Richard Gillam, published by Addison-Wesley, and explained at some
9593 * length there. The preface says to incorporate its examples into your
9594 * code at your own risk.
9596 * The algorithm is like a merge sort. */
9598 const UV* array_a; /* a's array */
9600 UV len_a; /* length of a's array */
9603 SV* u; /* the resulting union */
9607 UV i_a = 0; /* current index into a's array */
9611 /* running count, as explained in the algorithm source book; items are
9612 * stopped accumulating and are output when the count changes to/from 0.
9613 * The count is incremented when we start a range that's in an input's set,
9614 * and decremented when we start a range that's not in a set. So this
9615 * variable can be 0, 1, or 2. When it is 0 neither input is in their set,
9616 * and hence nothing goes into the union; 1, just one of the inputs is in
9617 * its set (and its current range gets added to the union); and 2 when both
9618 * inputs are in their sets. */
9621 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
9623 assert(*output == NULL || is_invlist(*output));
9625 len_b = _invlist_len(b);
9628 /* Here, 'b' is empty, hence it's complement is all possible code
9629 * points. So if the union includes the complement of 'b', it includes
9630 * everything, and we need not even look at 'a'. It's easiest to
9631 * create a new inversion list that matches everything. */
9633 SV* everything = _add_range_to_invlist(NULL, 0, UV_MAX);
9635 if (*output == NULL) { /* If the output didn't exist, just point it
9637 *output = everything;
9639 else { /* Otherwise, replace its contents with the new list */
9640 invlist_replace_list_destroys_src(*output, everything);
9641 SvREFCNT_dec_NN(everything);
9647 /* Here, we don't want the complement of 'b', and since 'b' is empty,
9648 * the union will come entirely from 'a'. If 'a' is NULL or empty, the
9649 * output will be empty */
9651 if (a == NULL || _invlist_len(a) == 0) {
9652 if (*output == NULL) {
9653 *output = _new_invlist(0);
9656 invlist_clear(*output);
9661 /* Here, 'a' is not empty, but 'b' is, so 'a' entirely determines the
9662 * union. We can just return a copy of 'a' if '*output' doesn't point
9663 * to an existing list */
9664 if (*output == NULL) {
9665 *output = invlist_clone(a, NULL);
9669 /* If the output is to overwrite 'a', we have a no-op, as it's
9675 /* Here, '*output' is to be overwritten by 'a' */
9676 u = invlist_clone(a, NULL);
9677 invlist_replace_list_destroys_src(*output, u);
9683 /* Here 'b' is not empty. See about 'a' */
9685 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
9687 /* Here, 'a' is empty (and b is not). That means the union will come
9688 * entirely from 'b'. If '*output' is NULL, we can directly return a
9689 * clone of 'b'. Otherwise, we replace the contents of '*output' with
9692 SV ** dest = (*output == NULL) ? output : &u;
9693 *dest = invlist_clone(b, NULL);
9695 _invlist_invert(*dest);
9699 invlist_replace_list_destroys_src(*output, u);
9706 /* Here both lists exist and are non-empty */
9707 array_a = invlist_array(a);
9708 array_b = invlist_array(b);
9710 /* If are to take the union of 'a' with the complement of b, set it
9711 * up so are looking at b's complement. */
9714 /* To complement, we invert: if the first element is 0, remove it. To
9715 * do this, we just pretend the array starts one later */
9716 if (array_b[0] == 0) {
9722 /* But if the first element is not zero, we pretend the list starts
9723 * at the 0 that is always stored immediately before the array. */
9729 /* Size the union for the worst case: that the sets are completely
9731 u = _new_invlist(len_a + len_b);
9733 /* Will contain U+0000 if either component does */
9734 array_u = _invlist_array_init(u, ( len_a > 0 && array_a[0] == 0)
9735 || (len_b > 0 && array_b[0] == 0));
9737 /* Go through each input list item by item, stopping when have exhausted
9739 while (i_a < len_a && i_b < len_b) {
9740 UV cp; /* The element to potentially add to the union's array */
9741 bool cp_in_set; /* is it in the input list's set or not */
9743 /* We need to take one or the other of the two inputs for the union.
9744 * Since we are merging two sorted lists, we take the smaller of the
9745 * next items. In case of a tie, we take first the one that is in its
9746 * set. If we first took the one not in its set, it would decrement
9747 * the count, possibly to 0 which would cause it to be output as ending
9748 * the range, and the next time through we would take the same number,
9749 * and output it again as beginning the next range. By doing it the
9750 * opposite way, there is no possibility that the count will be
9751 * momentarily decremented to 0, and thus the two adjoining ranges will
9752 * be seamlessly merged. (In a tie and both are in the set or both not
9753 * in the set, it doesn't matter which we take first.) */
9754 if ( array_a[i_a] < array_b[i_b]
9755 || ( array_a[i_a] == array_b[i_b]
9756 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9758 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9759 cp = array_a[i_a++];
9762 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9763 cp = array_b[i_b++];
9766 /* Here, have chosen which of the two inputs to look at. Only output
9767 * if the running count changes to/from 0, which marks the
9768 * beginning/end of a range that's in the set */
9771 array_u[i_u++] = cp;
9778 array_u[i_u++] = cp;
9784 /* The loop above increments the index into exactly one of the input lists
9785 * each iteration, and ends when either index gets to its list end. That
9786 * means the other index is lower than its end, and so something is
9787 * remaining in that one. We decrement 'count', as explained below, if
9788 * that list is in its set. (i_a and i_b each currently index the element
9789 * beyond the one we care about.) */
9790 if ( (i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9791 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9796 /* Above we decremented 'count' if the list that had unexamined elements in
9797 * it was in its set. This has made it so that 'count' being non-zero
9798 * means there isn't anything left to output; and 'count' equal to 0 means
9799 * that what is left to output is precisely that which is left in the
9800 * non-exhausted input list.
9802 * To see why, note first that the exhausted input obviously has nothing
9803 * left to add to the union. If it was in its set at its end, that means
9804 * the set extends from here to the platform's infinity, and hence so does
9805 * the union and the non-exhausted set is irrelevant. The exhausted set
9806 * also contributed 1 to 'count'. If 'count' was 2, it got decremented to
9807 * 1, but if it was 1, the non-exhausted set wasn't in its set, and so
9808 * 'count' remains at 1. This is consistent with the decremented 'count'
9809 * != 0 meaning there's nothing left to add to the union.
9811 * But if the exhausted input wasn't in its set, it contributed 0 to
9812 * 'count', and the rest of the union will be whatever the other input is.
9813 * If 'count' was 0, neither list was in its set, and 'count' remains 0;
9814 * otherwise it gets decremented to 0. This is consistent with 'count'
9815 * == 0 meaning the remainder of the union is whatever is left in the
9816 * non-exhausted list. */
9821 IV copy_count = len_a - i_a;
9822 if (copy_count > 0) { /* The non-exhausted input is 'a' */
9823 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
9825 else { /* The non-exhausted input is b */
9826 copy_count = len_b - i_b;
9827 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
9829 len_u = i_u + copy_count;
9832 /* Set the result to the final length, which can change the pointer to
9833 * array_u, so re-find it. (Note that it is unlikely that this will
9834 * change, as we are shrinking the space, not enlarging it) */
9835 if (len_u != _invlist_len(u)) {
9836 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
9838 array_u = invlist_array(u);
9841 if (*output == NULL) { /* Simply return the new inversion list */
9845 /* Otherwise, overwrite the inversion list that was in '*output'. We
9846 * could instead free '*output', and then set it to 'u', but experience
9847 * has shown [perl #127392] that if the input is a mortal, we can get a
9848 * huge build-up of these during regex compilation before they get
9850 invlist_replace_list_destroys_src(*output, u);
9858 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9859 const bool complement_b, SV** i)
9861 /* Take the intersection of two inversion lists and point '*i' to it. On
9862 * input, '*i' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9863 * even 'a' or 'b'). If to an inversion list, the contents of the original
9864 * list will be replaced by the intersection. The first list, 'a', may be
9865 * NULL, in which case '*i' will be an empty list. If 'complement_b' is
9866 * TRUE, the result will be the intersection of 'a' and the complement (or
9867 * inversion) of 'b' instead of 'b' directly.
9869 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9870 * Richard Gillam, published by Addison-Wesley, and explained at some
9871 * length there. The preface says to incorporate its examples into your
9872 * code at your own risk. In fact, it had bugs
9874 * The algorithm is like a merge sort, and is essentially the same as the
9878 const UV* array_a; /* a's array */
9880 UV len_a; /* length of a's array */
9883 SV* r; /* the resulting intersection */
9887 UV i_a = 0; /* current index into a's array */
9891 /* running count of how many of the two inputs are postitioned at ranges
9892 * that are in their sets. As explained in the algorithm source book,
9893 * items are stopped accumulating and are output when the count changes
9894 * to/from 2. The count is incremented when we start a range that's in an
9895 * input's set, and decremented when we start a range that's not in a set.
9896 * Only when it is 2 are we in the intersection. */
9899 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
9901 assert(*i == NULL || is_invlist(*i));
9903 /* Special case if either one is empty */
9904 len_a = (a == NULL) ? 0 : _invlist_len(a);
9905 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
9906 if (len_a != 0 && complement_b) {
9908 /* Here, 'a' is not empty, therefore from the enclosing 'if', 'b'
9909 * must be empty. Here, also we are using 'b's complement, which
9910 * hence must be every possible code point. Thus the intersection
9913 if (*i == a) { /* No-op */
9918 *i = invlist_clone(a, NULL);
9922 r = invlist_clone(a, NULL);
9923 invlist_replace_list_destroys_src(*i, r);
9928 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
9929 * intersection must be empty */
9931 *i = _new_invlist(0);
9939 /* Here both lists exist and are non-empty */
9940 array_a = invlist_array(a);
9941 array_b = invlist_array(b);
9943 /* If are to take the intersection of 'a' with the complement of b, set it
9944 * up so are looking at b's complement. */
9947 /* To complement, we invert: if the first element is 0, remove it. To
9948 * do this, we just pretend the array starts one later */
9949 if (array_b[0] == 0) {
9955 /* But if the first element is not zero, we pretend the list starts
9956 * at the 0 that is always stored immediately before the array. */
9962 /* Size the intersection for the worst case: that the intersection ends up
9963 * fragmenting everything to be completely disjoint */
9964 r= _new_invlist(len_a + len_b);
9966 /* Will contain U+0000 iff both components do */
9967 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
9968 && len_b > 0 && array_b[0] == 0);
9970 /* Go through each list item by item, stopping when have exhausted one of
9972 while (i_a < len_a && i_b < len_b) {
9973 UV cp; /* The element to potentially add to the intersection's
9975 bool cp_in_set; /* Is it in the input list's set or not */
9977 /* We need to take one or the other of the two inputs for the
9978 * intersection. Since we are merging two sorted lists, we take the
9979 * smaller of the next items. In case of a tie, we take first the one
9980 * that is not in its set (a difference from the union algorithm). If
9981 * we first took the one in its set, it would increment the count,
9982 * possibly to 2 which would cause it to be output as starting a range
9983 * in the intersection, and the next time through we would take that
9984 * same number, and output it again as ending the set. By doing the
9985 * opposite of this, there is no possibility that the count will be
9986 * momentarily incremented to 2. (In a tie and both are in the set or
9987 * both not in the set, it doesn't matter which we take first.) */
9988 if ( array_a[i_a] < array_b[i_b]
9989 || ( array_a[i_a] == array_b[i_b]
9990 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9992 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9993 cp = array_a[i_a++];
9996 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
10000 /* Here, have chosen which of the two inputs to look at. Only output
10001 * if the running count changes to/from 2, which marks the
10002 * beginning/end of a range that's in the intersection */
10006 array_r[i_r++] = cp;
10011 array_r[i_r++] = cp;
10018 /* The loop above increments the index into exactly one of the input lists
10019 * each iteration, and ends when either index gets to its list end. That
10020 * means the other index is lower than its end, and so something is
10021 * remaining in that one. We increment 'count', as explained below, if the
10022 * exhausted list was in its set. (i_a and i_b each currently index the
10023 * element beyond the one we care about.) */
10024 if ( (i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
10025 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
10030 /* Above we incremented 'count' if the exhausted list was in its set. This
10031 * has made it so that 'count' being below 2 means there is nothing left to
10032 * output; otheriwse what's left to add to the intersection is precisely
10033 * that which is left in the non-exhausted input list.
10035 * To see why, note first that the exhausted input obviously has nothing
10036 * left to affect the intersection. If it was in its set at its end, that
10037 * means the set extends from here to the platform's infinity, and hence
10038 * anything in the non-exhausted's list will be in the intersection, and
10039 * anything not in it won't be. Hence, the rest of the intersection is
10040 * precisely what's in the non-exhausted list The exhausted set also
10041 * contributed 1 to 'count', meaning 'count' was at least 1. Incrementing
10042 * it means 'count' is now at least 2. This is consistent with the
10043 * incremented 'count' being >= 2 means to add the non-exhausted list to
10044 * the intersection.
10046 * But if the exhausted input wasn't in its set, it contributed 0 to
10047 * 'count', and the intersection can't include anything further; the
10048 * non-exhausted set is irrelevant. 'count' was at most 1, and doesn't get
10049 * incremented. This is consistent with 'count' being < 2 meaning nothing
10050 * further to add to the intersection. */
10051 if (count < 2) { /* Nothing left to put in the intersection. */
10054 else { /* copy the non-exhausted list, unchanged. */
10055 IV copy_count = len_a - i_a;
10056 if (copy_count > 0) { /* a is the one with stuff left */
10057 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
10059 else { /* b is the one with stuff left */
10060 copy_count = len_b - i_b;
10061 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
10063 len_r = i_r + copy_count;
10066 /* Set the result to the final length, which can change the pointer to
10067 * array_r, so re-find it. (Note that it is unlikely that this will
10068 * change, as we are shrinking the space, not enlarging it) */
10069 if (len_r != _invlist_len(r)) {
10070 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
10072 array_r = invlist_array(r);
10075 if (*i == NULL) { /* Simply return the calculated intersection */
10078 else { /* Otherwise, replace the existing inversion list in '*i'. We could
10079 instead free '*i', and then set it to 'r', but experience has
10080 shown [perl #127392] that if the input is a mortal, we can get a
10081 huge build-up of these during regex compilation before they get
10084 invlist_replace_list_destroys_src(*i, r);
10089 SvREFCNT_dec_NN(r);
10096 Perl__add_range_to_invlist(pTHX_ SV* invlist, UV start, UV end)
10098 /* Add the range from 'start' to 'end' inclusive to the inversion list's
10099 * set. A pointer to the inversion list is returned. This may actually be
10100 * a new list, in which case the passed in one has been destroyed. The
10101 * passed-in inversion list can be NULL, in which case a new one is created
10102 * with just the one range in it. The new list is not necessarily
10103 * NUL-terminated. Space is not freed if the inversion list shrinks as a
10104 * result of this function. The gain would not be large, and in many
10105 * cases, this is called multiple times on a single inversion list, so
10106 * anything freed may almost immediately be needed again.
10108 * This used to mostly call the 'union' routine, but that is much more
10109 * heavyweight than really needed for a single range addition */
10111 UV* array; /* The array implementing the inversion list */
10112 UV len; /* How many elements in 'array' */
10113 SSize_t i_s; /* index into the invlist array where 'start'
10115 SSize_t i_e = 0; /* And the index where 'end' should go */
10116 UV cur_highest; /* The highest code point in the inversion list
10117 upon entry to this function */
10119 /* This range becomes the whole inversion list if none already existed */
10120 if (invlist == NULL) {
10121 invlist = _new_invlist(2);
10122 _append_range_to_invlist(invlist, start, end);
10126 /* Likewise, if the inversion list is currently empty */
10127 len = _invlist_len(invlist);
10129 _append_range_to_invlist(invlist, start, end);
10133 /* Starting here, we have to know the internals of the list */
10134 array = invlist_array(invlist);
10136 /* If the new range ends higher than the current highest ... */
10137 cur_highest = invlist_highest(invlist);
10138 if (end > cur_highest) {
10140 /* If the whole range is higher, we can just append it */
10141 if (start > cur_highest) {
10142 _append_range_to_invlist(invlist, start, end);
10146 /* Otherwise, add the portion that is higher ... */
10147 _append_range_to_invlist(invlist, cur_highest + 1, end);
10149 /* ... and continue on below to handle the rest. As a result of the
10150 * above append, we know that the index of the end of the range is the
10151 * final even numbered one of the array. Recall that the final element
10152 * always starts a range that extends to infinity. If that range is in
10153 * the set (meaning the set goes from here to infinity), it will be an
10154 * even index, but if it isn't in the set, it's odd, and the final
10155 * range in the set is one less, which is even. */
10156 if (end == UV_MAX) {
10164 /* We have dealt with appending, now see about prepending. If the new
10165 * range starts lower than the current lowest ... */
10166 if (start < array[0]) {
10168 /* Adding something which has 0 in it is somewhat tricky, and uncommon.
10169 * Let the union code handle it, rather than having to know the
10170 * trickiness in two code places. */
10171 if (UNLIKELY(start == 0)) {
10174 range_invlist = _new_invlist(2);
10175 _append_range_to_invlist(range_invlist, start, end);
10177 _invlist_union(invlist, range_invlist, &invlist);
10179 SvREFCNT_dec_NN(range_invlist);
10184 /* If the whole new range comes before the first entry, and doesn't
10185 * extend it, we have to insert it as an additional range */
10186 if (end < array[0] - 1) {
10188 goto splice_in_new_range;
10191 /* Here the new range adjoins the existing first range, extending it
10195 /* And continue on below to handle the rest. We know that the index of
10196 * the beginning of the range is the first one of the array */
10199 else { /* Not prepending any part of the new range to the existing list.
10200 * Find where in the list it should go. This finds i_s, such that:
10201 * invlist[i_s] <= start < array[i_s+1]
10203 i_s = _invlist_search(invlist, start);
10206 /* At this point, any extending before the beginning of the inversion list
10207 * and/or after the end has been done. This has made it so that, in the
10208 * code below, each endpoint of the new range is either in a range that is
10209 * in the set, or is in a gap between two ranges that are. This means we
10210 * don't have to worry about exceeding the array bounds.
10212 * Find where in the list the new range ends (but we can skip this if we
10213 * have already determined what it is, or if it will be the same as i_s,
10214 * which we already have computed) */
10216 i_e = (start == end)
10218 : _invlist_search(invlist, end);
10221 /* Here generally invlist[i_e] <= end < array[i_e+1]. But if invlist[i_e]
10222 * is a range that goes to infinity there is no element at invlist[i_e+1],
10223 * so only the first relation holds. */
10225 if ( ! ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
10227 /* Here, the ranges on either side of the beginning of the new range
10228 * are in the set, and this range starts in the gap between them.
10230 * The new range extends the range above it downwards if the new range
10231 * ends at or above that range's start */
10232 const bool extends_the_range_above = ( end == UV_MAX
10233 || end + 1 >= array[i_s+1]);
10235 /* The new range extends the range below it upwards if it begins just
10236 * after where that range ends */
10237 if (start == array[i_s]) {
10239 /* If the new range fills the entire gap between the other ranges,
10240 * they will get merged together. Other ranges may also get
10241 * merged, depending on how many of them the new range spans. In
10242 * the general case, we do the merge later, just once, after we
10243 * figure out how many to merge. But in the case where the new
10244 * range exactly spans just this one gap (possibly extending into
10245 * the one above), we do the merge here, and an early exit. This
10246 * is done here to avoid having to special case later. */
10247 if (i_e - i_s <= 1) {
10249 /* If i_e - i_s == 1, it means that the new range terminates
10250 * within the range above, and hence 'extends_the_range_above'
10251 * must be true. (If the range above it extends to infinity,
10252 * 'i_s+2' will be above the array's limit, but 'len-i_s-2'
10253 * will be 0, so no harm done.) */
10254 if (extends_the_range_above) {
10255 Move(array + i_s + 2, array + i_s, len - i_s - 2, UV);
10256 invlist_set_len(invlist,
10258 *(get_invlist_offset_addr(invlist)));
10262 /* Here, i_e must == i_s. We keep them in sync, as they apply
10263 * to the same range, and below we are about to decrement i_s
10268 /* Here, the new range is adjacent to the one below. (It may also
10269 * span beyond the range above, but that will get resolved later.)
10270 * Extend the range below to include this one. */
10271 array[i_s] = (end == UV_MAX) ? UV_MAX : end + 1;
10273 start = array[i_s];
10275 else if (extends_the_range_above) {
10277 /* Here the new range only extends the range above it, but not the
10278 * one below. It merges with the one above. Again, we keep i_e
10279 * and i_s in sync if they point to the same range */
10284 array[i_s] = start;
10288 /* Here, we've dealt with the new range start extending any adjoining
10291 * If the new range extends to infinity, it is now the final one,
10292 * regardless of what was there before */
10293 if (UNLIKELY(end == UV_MAX)) {
10294 invlist_set_len(invlist, i_s + 1, *(get_invlist_offset_addr(invlist)));
10298 /* If i_e started as == i_s, it has also been dealt with,
10299 * and been updated to the new i_s, which will fail the following if */
10300 if (! ELEMENT_RANGE_MATCHES_INVLIST(i_e)) {
10302 /* Here, the ranges on either side of the end of the new range are in
10303 * the set, and this range ends in the gap between them.
10305 * If this range is adjacent to (hence extends) the range above it, it
10306 * becomes part of that range; likewise if it extends the range below,
10307 * it becomes part of that range */
10308 if (end + 1 == array[i_e+1]) {
10310 array[i_e] = start;
10312 else if (start <= array[i_e]) {
10313 array[i_e] = end + 1;
10320 /* If the range fits entirely in an existing range (as possibly already
10321 * extended above), it doesn't add anything new */
10322 if (ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
10326 /* Here, no part of the range is in the list. Must add it. It will
10327 * occupy 2 more slots */
10328 splice_in_new_range:
10330 invlist_extend(invlist, len + 2);
10331 array = invlist_array(invlist);
10332 /* Move the rest of the array down two slots. Don't include any
10334 Move(array + i_e + 1, array + i_e + 3, len - i_e - 1, UV);
10336 /* Do the actual splice */
10337 array[i_e+1] = start;
10338 array[i_e+2] = end + 1;
10339 invlist_set_len(invlist, len + 2, *(get_invlist_offset_addr(invlist)));
10343 /* Here the new range crossed the boundaries of a pre-existing range. The
10344 * code above has adjusted things so that both ends are in ranges that are
10345 * in the set. This means everything in between must also be in the set.
10346 * Just squash things together */
10347 Move(array + i_e + 1, array + i_s + 1, len - i_e - 1, UV);
10348 invlist_set_len(invlist,
10350 *(get_invlist_offset_addr(invlist)));
10356 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
10357 UV** other_elements_ptr)
10359 /* Create and return an inversion list whose contents are to be populated
10360 * by the caller. The caller gives the number of elements (in 'size') and
10361 * the very first element ('element0'). This function will set
10362 * '*other_elements_ptr' to an array of UVs, where the remaining elements
10363 * are to be placed.
10365 * Obviously there is some trust involved that the caller will properly
10366 * fill in the other elements of the array.
10368 * (The first element needs to be passed in, as the underlying code does
10369 * things differently depending on whether it is zero or non-zero) */
10371 SV* invlist = _new_invlist(size);
10374 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
10376 invlist = add_cp_to_invlist(invlist, element0);
10377 offset = *get_invlist_offset_addr(invlist);
10379 invlist_set_len(invlist, size, offset);
10380 *other_elements_ptr = invlist_array(invlist) + 1;
10386 #ifndef PERL_IN_XSUB_RE
10388 Perl__invlist_invert(pTHX_ SV* const invlist)
10390 /* Complement the input inversion list. This adds a 0 if the list didn't
10391 * have a zero; removes it otherwise. As described above, the data
10392 * structure is set up so that this is very efficient */
10394 PERL_ARGS_ASSERT__INVLIST_INVERT;
10396 assert(! invlist_is_iterating(invlist));
10398 /* The inverse of matching nothing is matching everything */
10399 if (_invlist_len(invlist) == 0) {
10400 _append_range_to_invlist(invlist, 0, UV_MAX);
10404 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
10408 Perl_invlist_clone(pTHX_ SV* const invlist, SV* new_invlist)
10410 /* Return a new inversion list that is a copy of the input one, which is
10411 * unchanged. The new list will not be mortal even if the old one was. */
10413 const STRLEN nominal_length = _invlist_len(invlist);
10414 const STRLEN physical_length = SvCUR(invlist);
10415 const bool offset = *(get_invlist_offset_addr(invlist));
10417 PERL_ARGS_ASSERT_INVLIST_CLONE;
10419 if (new_invlist == NULL) {
10420 new_invlist = _new_invlist(nominal_length);
10423 sv_upgrade(new_invlist, SVt_INVLIST);
10424 initialize_invlist_guts(new_invlist, nominal_length);
10427 *(get_invlist_offset_addr(new_invlist)) = offset;
10428 invlist_set_len(new_invlist, nominal_length, offset);
10429 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
10431 return new_invlist;
10436 PERL_STATIC_INLINE UV
10437 S_invlist_lowest(SV* const invlist)
10439 /* Returns the lowest code point that matches an inversion list. This API
10440 * has an ambiguity, as it returns 0 under either the lowest is actually
10441 * 0, or if the list is empty. If this distinction matters to you, check
10442 * for emptiness before calling this function */
10444 UV len = _invlist_len(invlist);
10447 PERL_ARGS_ASSERT_INVLIST_LOWEST;
10453 array = invlist_array(invlist);
10459 S_invlist_contents(pTHX_ SV* const invlist, const bool traditional_style)
10461 /* Get the contents of an inversion list into a string SV so that they can
10462 * be printed out. If 'traditional_style' is TRUE, it uses the format
10463 * traditionally done for debug tracing; otherwise it uses a format
10464 * suitable for just copying to the output, with blanks between ranges and
10465 * a dash between range components */
10469 const char intra_range_delimiter = (traditional_style ? '\t' : '-');
10470 const char inter_range_delimiter = (traditional_style ? '\n' : ' ');
10472 if (traditional_style) {
10473 output = newSVpvs("\n");
10476 output = newSVpvs("");
10479 PERL_ARGS_ASSERT_INVLIST_CONTENTS;
10481 assert(! invlist_is_iterating(invlist));
10483 invlist_iterinit(invlist);
10484 while (invlist_iternext(invlist, &start, &end)) {
10485 if (end == UV_MAX) {
10486 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%cINFTY%c",
10487 start, intra_range_delimiter,
10488 inter_range_delimiter);
10490 else if (end != start) {
10491 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c%04" UVXf "%c",
10493 intra_range_delimiter,
10494 end, inter_range_delimiter);
10497 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c",
10498 start, inter_range_delimiter);
10502 if (SvCUR(output) && ! traditional_style) {/* Get rid of trailing blank */
10503 SvCUR_set(output, SvCUR(output) - 1);
10509 #ifndef PERL_IN_XSUB_RE
10511 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
10512 const char * const indent, SV* const invlist)
10514 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
10515 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
10516 * the string 'indent'. The output looks like this:
10517 [0] 0x000A .. 0x000D
10519 [4] 0x2028 .. 0x2029
10520 [6] 0x3104 .. INFTY
10521 * This means that the first range of code points matched by the list are
10522 * 0xA through 0xD; the second range contains only the single code point
10523 * 0x85, etc. An inversion list is an array of UVs. Two array elements
10524 * are used to define each range (except if the final range extends to
10525 * infinity, only a single element is needed). The array index of the
10526 * first element for the corresponding range is given in brackets. */
10531 PERL_ARGS_ASSERT__INVLIST_DUMP;
10533 if (invlist_is_iterating(invlist)) {
10534 Perl_dump_indent(aTHX_ level, file,
10535 "%sCan't dump inversion list because is in middle of iterating\n",
10540 invlist_iterinit(invlist);
10541 while (invlist_iternext(invlist, &start, &end)) {
10542 if (end == UV_MAX) {
10543 Perl_dump_indent(aTHX_ level, file,
10544 "%s[%" UVuf "] 0x%04" UVXf " .. INFTY\n",
10545 indent, (UV)count, start);
10547 else if (end != start) {
10548 Perl_dump_indent(aTHX_ level, file,
10549 "%s[%" UVuf "] 0x%04" UVXf " .. 0x%04" UVXf "\n",
10550 indent, (UV)count, start, end);
10553 Perl_dump_indent(aTHX_ level, file, "%s[%" UVuf "] 0x%04" UVXf "\n",
10554 indent, (UV)count, start);
10562 #if defined(PERL_ARGS_ASSERT__INVLISTEQ) && !defined(PERL_IN_XSUB_RE)
10564 Perl__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
10566 /* Return a boolean as to if the two passed in inversion lists are
10567 * identical. The final argument, if TRUE, says to take the complement of
10568 * the second inversion list before doing the comparison */
10570 const UV len_a = _invlist_len(a);
10571 UV len_b = _invlist_len(b);
10573 const UV* array_a = NULL;
10574 const UV* array_b = NULL;
10576 PERL_ARGS_ASSERT__INVLISTEQ;
10578 /* This code avoids accessing the arrays unless it knows the length is
10583 return ! complement_b;
10587 array_a = invlist_array(a);
10591 array_b = invlist_array(b);
10594 /* If are to compare 'a' with the complement of b, set it
10595 * up so are looking at b's complement. */
10596 if (complement_b) {
10598 /* The complement of nothing is everything, so <a> would have to have
10599 * just one element, starting at zero (ending at infinity) */
10601 return (len_a == 1 && array_a[0] == 0);
10603 if (array_b[0] == 0) {
10605 /* Otherwise, to complement, we invert. Here, the first element is
10606 * 0, just remove it. To do this, we just pretend the array starts
10614 /* But if the first element is not zero, we pretend the list starts
10615 * at the 0 that is always stored immediately before the array. */
10621 return len_a == len_b
10622 && memEQ(array_a, array_b, len_a * sizeof(array_a[0]));
10628 * As best we can, determine the characters that can match the start of
10629 * the given EXACTF-ish node. This is for use in creating ssc nodes, so there
10630 * can be false positive matches
10632 * Returns the invlist as a new SV*; it is the caller's responsibility to
10633 * call SvREFCNT_dec() when done with it.
10636 S_make_exactf_invlist(pTHX_ RExC_state_t *pRExC_state, regnode *node)
10638 const U8 * s = (U8*)STRING(node);
10639 SSize_t bytelen = STR_LEN(node);
10641 /* Start out big enough for 2 separate code points */
10642 SV* invlist = _new_invlist(4);
10644 PERL_ARGS_ASSERT_MAKE_EXACTF_INVLIST;
10649 /* We punt and assume can match anything if the node begins
10650 * with a multi-character fold. Things are complicated. For
10651 * example, /ffi/i could match any of:
10652 * "\N{LATIN SMALL LIGATURE FFI}"
10653 * "\N{LATIN SMALL LIGATURE FF}I"
10654 * "F\N{LATIN SMALL LIGATURE FI}"
10655 * plus several other things; and making sure we have all the
10656 * possibilities is hard. */
10657 if (is_MULTI_CHAR_FOLD_latin1_safe(s, s + bytelen)) {
10658 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10661 /* Any Latin1 range character can potentially match any
10662 * other depending on the locale, and in Turkic locales, U+130 and
10664 if (OP(node) == EXACTFL) {
10665 _invlist_union(invlist, PL_Latin1, &invlist);
10666 invlist = add_cp_to_invlist(invlist,
10667 LATIN_SMALL_LETTER_DOTLESS_I);
10668 invlist = add_cp_to_invlist(invlist,
10669 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
10672 /* But otherwise, it matches at least itself. We can
10673 * quickly tell if it has a distinct fold, and if so,
10674 * it matches that as well */
10675 invlist = add_cp_to_invlist(invlist, uc);
10676 if (IS_IN_SOME_FOLD_L1(uc))
10677 invlist = add_cp_to_invlist(invlist, PL_fold_latin1[uc]);
10680 /* Some characters match above-Latin1 ones under /i. This
10681 * is true of EXACTFL ones when the locale is UTF-8 */
10682 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(uc)
10683 && (! isASCII(uc) || ! inRANGE(OP(node), EXACTFAA,
10684 EXACTFAA_NO_TRIE)))
10686 add_above_Latin1_folds(pRExC_state, (U8) uc, &invlist);
10690 else { /* Pattern is UTF-8 */
10691 U8 folded[UTF8_MAX_FOLD_CHAR_EXPAND * UTF8_MAXBYTES_CASE + 1] = { '\0' };
10692 const U8* e = s + bytelen;
10695 fc = uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
10697 /* The only code points that aren't folded in a UTF EXACTFish
10698 * node are the problematic ones in EXACTFL nodes */
10699 if (OP(node) == EXACTFL && is_PROBLEMATIC_LOCALE_FOLDEDS_START_cp(uc)) {
10700 /* We need to check for the possibility that this EXACTFL
10701 * node begins with a multi-char fold. Therefore we fold
10702 * the first few characters of it so that we can make that
10708 for (i = 0; i < UTF8_MAX_FOLD_CHAR_EXPAND && s < e; i++) {
10710 *(d++) = (U8) toFOLD(*s);
10711 if (fc < 0) { /* Save the first fold */
10718 UV fold = toFOLD_utf8_safe(s, e, d, &len);
10719 if (fc < 0) { /* Save the first fold */
10727 /* And set up so the code below that looks in this folded
10728 * buffer instead of the node's string */
10733 /* When we reach here 's' points to the fold of the first
10734 * character(s) of the node; and 'e' points to far enough along
10735 * the folded string to be just past any possible multi-char
10738 * Like the non-UTF case above, we punt if the node begins with a
10739 * multi-char fold */
10741 if (is_MULTI_CHAR_FOLD_utf8_safe(s, e)) {
10742 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10744 else { /* Single char fold */
10747 const U32 * remaining_folds;
10748 Size_t folds_count;
10750 /* It matches itself */
10751 invlist = add_cp_to_invlist(invlist, fc);
10753 /* ... plus all the things that fold to it, which are found in
10754 * PL_utf8_foldclosures */
10755 folds_count = _inverse_folds(fc, &first_fold,
10757 for (k = 0; k < folds_count; k++) {
10758 UV c = (k == 0) ? first_fold : remaining_folds[k-1];
10760 /* /aa doesn't allow folds between ASCII and non- */
10761 if ( inRANGE(OP(node), EXACTFAA, EXACTFAA_NO_TRIE)
10762 && isASCII(c) != isASCII(fc))
10767 invlist = add_cp_to_invlist(invlist, c);
10770 if (OP(node) == EXACTFL) {
10772 /* If either [iI] are present in an EXACTFL node the above code
10773 * should have added its normal case pair, but under a Turkish
10774 * locale they could match instead the case pairs from it. Add
10775 * those as potential matches as well */
10776 if (isALPHA_FOLD_EQ(fc, 'I')) {
10777 invlist = add_cp_to_invlist(invlist,
10778 LATIN_SMALL_LETTER_DOTLESS_I);
10779 invlist = add_cp_to_invlist(invlist,
10780 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
10782 else if (fc == LATIN_SMALL_LETTER_DOTLESS_I) {
10783 invlist = add_cp_to_invlist(invlist, 'I');
10785 else if (fc == LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE) {
10786 invlist = add_cp_to_invlist(invlist, 'i');
10795 #undef HEADER_LENGTH
10796 #undef TO_INTERNAL_SIZE
10797 #undef FROM_INTERNAL_SIZE
10798 #undef INVLIST_VERSION_ID
10800 /* End of inversion list object */
10803 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
10805 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
10806 * constructs, and updates RExC_flags with them. On input, RExC_parse
10807 * should point to the first flag; it is updated on output to point to the
10808 * final ')' or ':'. There needs to be at least one flag, or this will
10811 /* for (?g), (?gc), and (?o) warnings; warning
10812 about (?c) will warn about (?g) -- japhy */
10814 #define WASTED_O 0x01
10815 #define WASTED_G 0x02
10816 #define WASTED_C 0x04
10817 #define WASTED_GC (WASTED_G|WASTED_C)
10818 I32 wastedflags = 0x00;
10819 U32 posflags = 0, negflags = 0;
10820 U32 *flagsp = &posflags;
10821 char has_charset_modifier = '\0';
10823 bool has_use_defaults = FALSE;
10824 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
10825 int x_mod_count = 0;
10827 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
10829 /* '^' as an initial flag sets certain defaults */
10830 if (UCHARAT(RExC_parse) == '^') {
10832 has_use_defaults = TRUE;
10833 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
10834 cs = (toUSE_UNI_CHARSET_NOT_DEPENDS)
10835 ? REGEX_UNICODE_CHARSET
10836 : REGEX_DEPENDS_CHARSET;
10837 set_regex_charset(&RExC_flags, cs);
10840 cs = get_regex_charset(RExC_flags);
10841 if ( cs == REGEX_DEPENDS_CHARSET
10842 && (toUSE_UNI_CHARSET_NOT_DEPENDS))
10844 cs = REGEX_UNICODE_CHARSET;
10848 while (RExC_parse < RExC_end) {
10849 /* && memCHRs("iogcmsx", *RExC_parse) */
10850 /* (?g), (?gc) and (?o) are useless here
10851 and must be globally applied -- japhy */
10852 if ((RExC_pm_flags & PMf_WILDCARD)) {
10853 if (flagsp == & negflags) {
10854 if (*RExC_parse == 'm') {
10856 /* diag_listed_as: Use of %s is not allowed in Unicode
10857 property wildcard subpatterns in regex; marked by <--
10859 vFAIL("Use of modifier '-m' is not allowed in Unicode"
10860 " property wildcard subpatterns");
10864 if (*RExC_parse == 's') {
10865 goto modifier_illegal_in_wildcard;
10870 switch (*RExC_parse) {
10872 /* Code for the imsxn flags */
10873 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp, x_mod_count);
10875 case LOCALE_PAT_MOD:
10876 if (has_charset_modifier) {
10877 goto excess_modifier;
10879 else if (flagsp == &negflags) {
10882 cs = REGEX_LOCALE_CHARSET;
10883 has_charset_modifier = LOCALE_PAT_MOD;
10885 case UNICODE_PAT_MOD:
10886 if (has_charset_modifier) {
10887 goto excess_modifier;
10889 else if (flagsp == &negflags) {
10892 cs = REGEX_UNICODE_CHARSET;
10893 has_charset_modifier = UNICODE_PAT_MOD;
10895 case ASCII_RESTRICT_PAT_MOD:
10896 if (flagsp == &negflags) {
10899 if (has_charset_modifier) {
10900 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
10901 goto excess_modifier;
10903 /* Doubled modifier implies more restricted */
10904 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
10907 cs = REGEX_ASCII_RESTRICTED_CHARSET;
10909 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
10911 case DEPENDS_PAT_MOD:
10912 if (has_use_defaults) {
10913 goto fail_modifiers;
10915 else if (flagsp == &negflags) {
10918 else if (has_charset_modifier) {
10919 goto excess_modifier;
10922 /* The dual charset means unicode semantics if the
10923 * pattern (or target, not known until runtime) are
10924 * utf8, or something in the pattern indicates unicode
10926 cs = (toUSE_UNI_CHARSET_NOT_DEPENDS)
10927 ? REGEX_UNICODE_CHARSET
10928 : REGEX_DEPENDS_CHARSET;
10929 has_charset_modifier = DEPENDS_PAT_MOD;
10933 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
10934 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
10936 else if (has_charset_modifier == *(RExC_parse - 1)) {
10937 vFAIL2("Regexp modifier \"%c\" may not appear twice",
10938 *(RExC_parse - 1));
10941 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
10943 NOT_REACHED; /*NOTREACHED*/
10946 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
10947 *(RExC_parse - 1));
10948 NOT_REACHED; /*NOTREACHED*/
10949 case GLOBAL_PAT_MOD: /* 'g' */
10950 if (RExC_pm_flags & PMf_WILDCARD) {
10951 goto modifier_illegal_in_wildcard;
10954 case ONCE_PAT_MOD: /* 'o' */
10955 if (ckWARN(WARN_REGEXP)) {
10956 const I32 wflagbit = *RExC_parse == 'o'
10959 if (! (wastedflags & wflagbit) ) {
10960 wastedflags |= wflagbit;
10961 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10964 "Useless (%s%c) - %suse /%c modifier",
10965 flagsp == &negflags ? "?-" : "?",
10967 flagsp == &negflags ? "don't " : "",
10974 case CONTINUE_PAT_MOD: /* 'c' */
10975 if (RExC_pm_flags & PMf_WILDCARD) {
10976 goto modifier_illegal_in_wildcard;
10978 if (ckWARN(WARN_REGEXP)) {
10979 if (! (wastedflags & WASTED_C) ) {
10980 wastedflags |= WASTED_GC;
10981 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10984 "Useless (%sc) - %suse /gc modifier",
10985 flagsp == &negflags ? "?-" : "?",
10986 flagsp == &negflags ? "don't " : ""
10991 case KEEPCOPY_PAT_MOD: /* 'p' */
10992 if (RExC_pm_flags & PMf_WILDCARD) {
10993 goto modifier_illegal_in_wildcard;
10995 if (flagsp == &negflags) {
10996 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
10998 *flagsp |= RXf_PMf_KEEPCOPY;
11002 /* A flag is a default iff it is following a minus, so
11003 * if there is a minus, it means will be trying to
11004 * re-specify a default which is an error */
11005 if (has_use_defaults || flagsp == &negflags) {
11006 goto fail_modifiers;
11008 flagsp = &negflags;
11009 wastedflags = 0; /* reset so (?g-c) warns twice */
11015 if ( (RExC_pm_flags & PMf_WILDCARD)
11016 && cs != REGEX_ASCII_MORE_RESTRICTED_CHARSET)
11019 /* diag_listed_as: Use of %s is not allowed in Unicode
11020 property wildcard subpatterns in regex; marked by <--
11022 vFAIL2("Use of modifier '%c' is not allowed in Unicode"
11023 " property wildcard subpatterns",
11024 has_charset_modifier);
11027 if ((posflags & (RXf_PMf_EXTENDED|RXf_PMf_EXTENDED_MORE)) == RXf_PMf_EXTENDED) {
11028 negflags |= RXf_PMf_EXTENDED_MORE;
11030 RExC_flags |= posflags;
11032 if (negflags & RXf_PMf_EXTENDED) {
11033 negflags |= RXf_PMf_EXTENDED_MORE;
11035 RExC_flags &= ~negflags;
11036 set_regex_charset(&RExC_flags, cs);
11041 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
11042 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11043 vFAIL2utf8f("Sequence (%" UTF8f "...) not recognized",
11044 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
11045 NOT_REACHED; /*NOTREACHED*/
11048 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11051 vFAIL("Sequence (?... not terminated");
11053 modifier_illegal_in_wildcard:
11055 /* diag_listed_as: Use of %s is not allowed in Unicode property wildcard
11056 subpatterns in regex; marked by <-- HERE in m/%s/ */
11057 vFAIL2("Use of modifier '%c' is not allowed in Unicode property wildcard"
11058 " subpatterns", *(RExC_parse - 1));
11062 - reg - regular expression, i.e. main body or parenthesized thing
11064 * Caller must absorb opening parenthesis.
11066 * Combining parenthesis handling with the base level of regular expression
11067 * is a trifle forced, but the need to tie the tails of the branches to what
11068 * follows makes it hard to avoid.
11070 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
11072 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
11074 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
11077 STATIC regnode_offset
11078 S_handle_named_backref(pTHX_ RExC_state_t *pRExC_state,
11080 char * parse_start,
11084 regnode_offset ret;
11085 char* name_start = RExC_parse;
11087 SV *sv_dat = reg_scan_name(pRExC_state, REG_RSN_RETURN_DATA);
11088 DECLARE_AND_GET_RE_DEBUG_FLAGS;
11090 PERL_ARGS_ASSERT_HANDLE_NAMED_BACKREF;
11092 if (RExC_parse != name_start && ch == '}') {
11093 while (isBLANK(*RExC_parse)) {
11097 if (RExC_parse == name_start || *RExC_parse != ch) {
11098 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
11099 vFAIL2("Sequence %.3s... not terminated", parse_start);
11103 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11104 RExC_rxi->data->data[num]=(void*)sv_dat;
11105 SvREFCNT_inc_simple_void_NN(sv_dat);
11108 ret = reganode(pRExC_state,
11111 : (ASCII_FOLD_RESTRICTED)
11113 : (AT_LEAST_UNI_SEMANTICS)
11119 *flagp |= HASWIDTH;
11121 Set_Node_Offset(REGNODE_p(ret), parse_start+1);
11122 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
11124 nextchar(pRExC_state);
11128 /* On success, returns the offset at which any next node should be placed into
11129 * the regex engine program being compiled.
11131 * Returns 0 otherwise, with *flagp set to indicate why:
11132 * TRYAGAIN at the end of (?) that only sets flags.
11133 * RESTART_PARSE if the parse needs to be restarted, or'd with
11134 * NEED_UTF8 if the pattern needs to be upgraded to UTF-8.
11135 * Otherwise would only return 0 if regbranch() returns 0, which cannot
11137 STATIC regnode_offset
11138 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp, U32 depth)
11139 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
11140 * 2 is like 1, but indicates that nextchar() has been called to advance
11141 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
11142 * this flag alerts us to the need to check for that */
11144 regnode_offset ret = 0; /* Will be the head of the group. */
11146 regnode_offset lastbr;
11147 regnode_offset ender = 0;
11150 U32 oregflags = RExC_flags;
11151 bool have_branch = 0;
11153 I32 freeze_paren = 0;
11154 I32 after_freeze = 0;
11155 I32 num; /* numeric backreferences */
11156 SV * max_open; /* Max number of unclosed parens */
11157 I32 was_in_lookaround = RExC_in_lookaround;
11159 char * parse_start = RExC_parse; /* MJD */
11160 char * const oregcomp_parse = RExC_parse;
11162 DECLARE_AND_GET_RE_DEBUG_FLAGS;
11164 PERL_ARGS_ASSERT_REG;
11165 DEBUG_PARSE("reg ");
11167 max_open = get_sv(RE_COMPILE_RECURSION_LIMIT, GV_ADD);
11169 if (!SvIOK(max_open)) {
11170 sv_setiv(max_open, RE_COMPILE_RECURSION_INIT);
11172 if (depth > 4 * (UV) SvIV(max_open)) { /* We increase depth by 4 for each
11174 vFAIL("Too many nested open parens");
11177 *flagp = 0; /* Initialize. */
11179 /* Having this true makes it feasible to have a lot fewer tests for the
11180 * parse pointer being in scope. For example, we can write
11181 * while(isFOO(*RExC_parse)) RExC_parse++;
11183 * while(RExC_parse < RExC_end && isFOO(*RExC_parse)) RExC_parse++;
11185 assert(*RExC_end == '\0');
11187 /* Make an OPEN node, if parenthesized. */
11190 /* Under /x, space and comments can be gobbled up between the '(' and
11191 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
11192 * intervening space, as the sequence is a token, and a token should be
11194 bool has_intervening_patws = (paren == 2)
11195 && *(RExC_parse - 1) != '(';
11197 if (RExC_parse >= RExC_end) {
11198 vFAIL("Unmatched (");
11201 if (paren == 'r') { /* Atomic script run */
11205 else if ( *RExC_parse == '*') { /* (*VERB:ARG), (*construct:...) */
11206 char *start_verb = RExC_parse + 1;
11208 char *start_arg = NULL;
11209 unsigned char op = 0;
11210 int arg_required = 0;
11211 int internal_argval = -1; /* if >-1 we are not allowed an argument*/
11212 bool has_upper = FALSE;
11214 if (has_intervening_patws) {
11215 RExC_parse++; /* past the '*' */
11217 /* For strict backwards compatibility, don't change the message
11218 * now that we also have lowercase operands */
11219 if (isUPPER(*RExC_parse)) {
11220 vFAIL("In '(*VERB...)', the '(' and '*' must be adjacent");
11223 vFAIL("In '(*...)', the '(' and '*' must be adjacent");
11226 while (RExC_parse < RExC_end && *RExC_parse != ')' ) {
11227 if ( *RExC_parse == ':' ) {
11228 start_arg = RExC_parse + 1;
11232 if (isUPPER(*RExC_parse)) {
11238 RExC_parse += UTF8SKIP(RExC_parse);
11241 verb_len = RExC_parse - start_verb;
11243 if (RExC_parse >= RExC_end) {
11244 goto unterminated_verb_pattern;
11247 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11248 while ( RExC_parse < RExC_end && *RExC_parse != ')' ) {
11249 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11251 if ( RExC_parse >= RExC_end || *RExC_parse != ')' ) {
11252 unterminated_verb_pattern:
11254 vFAIL("Unterminated verb pattern argument");
11257 vFAIL("Unterminated '(*...' argument");
11261 if ( RExC_parse >= RExC_end || *RExC_parse != ')' ) {
11263 vFAIL("Unterminated verb pattern");
11266 vFAIL("Unterminated '(*...' construct");
11271 /* Here, we know that RExC_parse < RExC_end */
11273 switch ( *start_verb ) {
11274 case 'A': /* (*ACCEPT) */
11275 if ( memEQs(start_verb, verb_len,"ACCEPT") ) {
11277 internal_argval = RExC_nestroot;
11280 case 'C': /* (*COMMIT) */
11281 if ( memEQs(start_verb, verb_len,"COMMIT") )
11284 case 'F': /* (*FAIL) */
11285 if ( verb_len==1 || memEQs(start_verb, verb_len,"FAIL") ) {
11289 case ':': /* (*:NAME) */
11290 case 'M': /* (*MARK:NAME) */
11291 if ( verb_len==0 || memEQs(start_verb, verb_len,"MARK") ) {
11296 case 'P': /* (*PRUNE) */
11297 if ( memEQs(start_verb, verb_len,"PRUNE") )
11300 case 'S': /* (*SKIP) */
11301 if ( memEQs(start_verb, verb_len,"SKIP") )
11304 case 'T': /* (*THEN) */
11305 /* [19:06] <TimToady> :: is then */
11306 if ( memEQs(start_verb, verb_len,"THEN") ) {
11308 RExC_seen |= REG_CUTGROUP_SEEN;
11312 if ( memEQs(start_verb, verb_len, "asr")
11313 || memEQs(start_verb, verb_len, "atomic_script_run"))
11315 paren = 'r'; /* Mnemonic: recursed run */
11318 else if (memEQs(start_verb, verb_len, "atomic")) {
11319 paren = 't'; /* AtOMIC */
11320 goto alpha_assertions;
11324 if ( memEQs(start_verb, verb_len, "plb")
11325 || memEQs(start_verb, verb_len, "positive_lookbehind"))
11328 goto lookbehind_alpha_assertions;
11330 else if ( memEQs(start_verb, verb_len, "pla")
11331 || memEQs(start_verb, verb_len, "positive_lookahead"))
11334 goto alpha_assertions;
11338 if ( memEQs(start_verb, verb_len, "nlb")
11339 || memEQs(start_verb, verb_len, "negative_lookbehind"))
11342 goto lookbehind_alpha_assertions;
11344 else if ( memEQs(start_verb, verb_len, "nla")
11345 || memEQs(start_verb, verb_len, "negative_lookahead"))
11348 goto alpha_assertions;
11352 if ( memEQs(start_verb, verb_len, "sr")
11353 || memEQs(start_verb, verb_len, "script_run"))
11355 regnode_offset atomic;
11361 /* This indicates Unicode rules. */
11362 REQUIRE_UNI_RULES(flagp, 0);
11368 RExC_parse = start_arg;
11370 if (RExC_in_script_run) {
11372 /* Nested script runs are treated as no-ops, because
11373 * if the nested one fails, the outer one must as
11374 * well. It could fail sooner, and avoid (??{} with
11375 * side effects, but that is explicitly documented as
11376 * undefined behavior. */
11380 if (paren == 's') {
11385 /* But, the atomic part of a nested atomic script run
11386 * isn't a no-op, but can be treated just like a '(?>'
11392 if (paren == 's') {
11393 /* Here, we're starting a new regular script run */
11394 ret = reg_node(pRExC_state, SROPEN);
11395 RExC_in_script_run = 1;
11400 /* Here, we are starting an atomic script run. This is
11401 * handled by recursing to deal with the atomic portion
11402 * separately, enclosed in SROPEN ... SRCLOSE nodes */
11404 ret = reg_node(pRExC_state, SROPEN);
11406 RExC_in_script_run = 1;
11408 atomic = reg(pRExC_state, 'r', &flags, depth);
11409 if (flags & (RESTART_PARSE|NEED_UTF8)) {
11410 *flagp = flags & (RESTART_PARSE|NEED_UTF8);
11414 if (! REGTAIL(pRExC_state, ret, atomic)) {
11415 REQUIRE_BRANCHJ(flagp, 0);
11418 if (! REGTAIL(pRExC_state, atomic, reg_node(pRExC_state,
11421 REQUIRE_BRANCHJ(flagp, 0);
11424 RExC_in_script_run = 0;
11430 lookbehind_alpha_assertions:
11431 RExC_seen |= REG_LOOKBEHIND_SEEN;
11436 RExC_in_lookaround++;
11437 RExC_seen_zerolen++;
11443 /* An empty negative lookahead assertion simply is failure */
11444 if (paren == 'A' && RExC_parse == start_arg) {
11445 ret=reganode(pRExC_state, OPFAIL, 0);
11446 nextchar(pRExC_state);
11450 RExC_parse = start_arg;
11455 "'(*%" UTF8f "' requires a terminating ':'",
11456 UTF8fARG(UTF, verb_len, start_verb));
11457 NOT_REACHED; /*NOTREACHED*/
11459 } /* End of switch */
11462 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
11464 if (has_upper || verb_len == 0) {
11466 "Unknown verb pattern '%" UTF8f "'",
11467 UTF8fARG(UTF, verb_len, start_verb));
11471 "Unknown '(*...)' construct '%" UTF8f "'",
11472 UTF8fARG(UTF, verb_len, start_verb));
11475 if ( RExC_parse == start_arg ) {
11478 if ( arg_required && !start_arg ) {
11479 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
11480 (int) verb_len, start_verb);
11482 if (internal_argval == -1) {
11483 ret = reganode(pRExC_state, op, 0);
11485 ret = reg2Lanode(pRExC_state, op, 0, internal_argval);
11487 RExC_seen |= REG_VERBARG_SEEN;
11489 SV *sv = newSVpvn( start_arg,
11490 RExC_parse - start_arg);
11491 ARG(REGNODE_p(ret)) = add_data( pRExC_state,
11492 STR_WITH_LEN("S"));
11493 RExC_rxi->data->data[ARG(REGNODE_p(ret))]=(void*)sv;
11494 FLAGS(REGNODE_p(ret)) = 1;
11496 FLAGS(REGNODE_p(ret)) = 0;
11498 if ( internal_argval != -1 )
11499 ARG2L_SET(REGNODE_p(ret), internal_argval);
11500 nextchar(pRExC_state);
11503 else if (*RExC_parse == '?') { /* (?...) */
11504 bool is_logical = 0;
11505 const char * const seqstart = RExC_parse;
11506 const char * endptr;
11507 const char non_existent_group_msg[]
11508 = "Reference to nonexistent group";
11509 const char impossible_group[] = "Invalid reference to group";
11511 if (has_intervening_patws) {
11513 vFAIL("In '(?...)', the '(' and '?' must be adjacent");
11516 RExC_parse++; /* past the '?' */
11517 paren = *RExC_parse; /* might be a trailing NUL, if not
11519 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11520 if (RExC_parse > RExC_end) {
11523 ret = 0; /* For look-ahead/behind. */
11526 case 'P': /* (?P...) variants for those used to PCRE/Python */
11527 paren = *RExC_parse;
11528 if ( paren == '<') { /* (?P<...>) named capture */
11530 if (RExC_parse >= RExC_end) {
11531 vFAIL("Sequence (?P<... not terminated");
11533 goto named_capture;
11535 else if (paren == '>') { /* (?P>name) named recursion */
11537 if (RExC_parse >= RExC_end) {
11538 vFAIL("Sequence (?P>... not terminated");
11540 goto named_recursion;
11542 else if (paren == '=') { /* (?P=...) named backref */
11544 return handle_named_backref(pRExC_state, flagp,
11547 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
11548 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11549 vFAIL3("Sequence (%.*s...) not recognized",
11550 (int) (RExC_parse - seqstart), seqstart);
11551 NOT_REACHED; /*NOTREACHED*/
11552 case '<': /* (?<...) */
11553 /* If you want to support (?<*...), first reconcile with GH #17363 */
11554 if (*RExC_parse == '!')
11556 else if (*RExC_parse != '=')
11563 case '\'': /* (?'...') */
11564 name_start = RExC_parse;
11565 svname = reg_scan_name(pRExC_state, REG_RSN_RETURN_NAME);
11566 if ( RExC_parse == name_start
11567 || RExC_parse >= RExC_end
11568 || *RExC_parse != paren)
11570 vFAIL2("Sequence (?%c... not terminated",
11571 paren=='>' ? '<' : (char) paren);
11576 if (!svname) /* shouldn't happen */
11578 "panic: reg_scan_name returned NULL");
11579 if (!RExC_paren_names) {
11580 RExC_paren_names= newHV();
11581 sv_2mortal(MUTABLE_SV(RExC_paren_names));
11583 RExC_paren_name_list= newAV();
11584 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
11587 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
11589 sv_dat = HeVAL(he_str);
11591 /* croak baby croak */
11593 "panic: paren_name hash element allocation failed");
11594 } else if ( SvPOK(sv_dat) ) {
11595 /* (?|...) can mean we have dupes so scan to check
11596 its already been stored. Maybe a flag indicating
11597 we are inside such a construct would be useful,
11598 but the arrays are likely to be quite small, so
11599 for now we punt -- dmq */
11600 IV count = SvIV(sv_dat);
11601 I32 *pv = (I32*)SvPVX(sv_dat);
11603 for ( i = 0 ; i < count ; i++ ) {
11604 if ( pv[i] == RExC_npar ) {
11610 pv = (I32*)SvGROW(sv_dat,
11611 SvCUR(sv_dat) + sizeof(I32)+1);
11612 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
11613 pv[count] = RExC_npar;
11614 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
11617 (void)SvUPGRADE(sv_dat, SVt_PVNV);
11618 sv_setpvn(sv_dat, (char *)&(RExC_npar),
11621 SvIV_set(sv_dat, 1);
11624 /* Yes this does cause a memory leak in debugging Perls
11626 if (!av_store(RExC_paren_name_list,
11627 RExC_npar, SvREFCNT_inc_NN(svname)))
11628 SvREFCNT_dec_NN(svname);
11631 /*sv_dump(sv_dat);*/
11633 nextchar(pRExC_state);
11635 goto capturing_parens;
11638 RExC_seen |= REG_LOOKBEHIND_SEEN;
11639 RExC_in_lookaround++;
11641 if (RExC_parse >= RExC_end) {
11642 vFAIL("Sequence (?... not terminated");
11644 RExC_seen_zerolen++;
11646 case '=': /* (?=...) */
11647 RExC_seen_zerolen++;
11648 RExC_in_lookaround++;
11650 case '!': /* (?!...) */
11651 RExC_seen_zerolen++;
11652 /* check if we're really just a "FAIL" assertion */
11653 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
11654 FALSE /* Don't force to /x */ );
11655 if (*RExC_parse == ')') {
11656 ret=reganode(pRExC_state, OPFAIL, 0);
11657 nextchar(pRExC_state);
11660 RExC_in_lookaround++;
11662 case '|': /* (?|...) */
11663 /* branch reset, behave like a (?:...) except that
11664 buffers in alternations share the same numbers */
11666 after_freeze = freeze_paren = RExC_npar;
11668 /* XXX This construct currently requires an extra pass.
11669 * Investigation would be required to see if that could be
11671 REQUIRE_PARENS_PASS;
11673 case ':': /* (?:...) */
11674 case '>': /* (?>...) */
11676 case '$': /* (?$...) */
11677 case '@': /* (?@...) */
11678 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
11680 case '0' : /* (?0) */
11681 case 'R' : /* (?R) */
11682 if (RExC_parse == RExC_end || *RExC_parse != ')')
11683 FAIL("Sequence (?R) not terminated");
11685 RExC_seen |= REG_RECURSE_SEEN;
11687 /* XXX These constructs currently require an extra pass.
11688 * It probably could be changed */
11689 REQUIRE_PARENS_PASS;
11691 *flagp |= POSTPONED;
11692 goto gen_recurse_regop;
11694 /* named and numeric backreferences */
11695 case '&': /* (?&NAME) */
11696 parse_start = RExC_parse - 1;
11699 SV *sv_dat = reg_scan_name(pRExC_state,
11700 REG_RSN_RETURN_DATA);
11701 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
11703 if (RExC_parse >= RExC_end || *RExC_parse != ')')
11704 vFAIL("Sequence (?&... not terminated");
11705 goto gen_recurse_regop;
11708 if (! inRANGE(RExC_parse[0], '1', '9')) {
11710 vFAIL("Illegal pattern");
11712 goto parse_recursion;
11714 case '-': /* (?-1) */
11715 if (! inRANGE(RExC_parse[0], '1', '9')) {
11716 RExC_parse--; /* rewind to let it be handled later */
11720 case '1': case '2': case '3': case '4': /* (?1) */
11721 case '5': case '6': case '7': case '8': case '9':
11722 RExC_parse = (char *) seqstart + 1; /* Point to the digit */
11725 bool is_neg = FALSE;
11727 parse_start = RExC_parse - 1; /* MJD */
11728 if (*RExC_parse == '-') {
11733 if (grok_atoUV(RExC_parse, &unum, &endptr)
11737 RExC_parse = (char*)endptr;
11739 else { /* Overflow, or something like that. Position
11740 beyond all digits for the message */
11741 while (RExC_parse < RExC_end && isDIGIT(*RExC_parse)) {
11744 vFAIL(impossible_group);
11747 /* -num is always representable on 1 and 2's complement
11752 if (*RExC_parse!=')')
11753 vFAIL("Expecting close bracket");
11756 if (paren == '-' || paren == '+') {
11758 /* Don't overflow */
11759 if (UNLIKELY(I32_MAX - RExC_npar < num)) {
11761 vFAIL(impossible_group);
11765 Diagram of capture buffer numbering.
11766 Top line is the normal capture buffer numbers
11767 Bottom line is the negative indexing as from
11771 /(a(x)y)(a(b(c(?+2)d)e)f)(g(h))/
11772 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
11773 - 5 4 3 2 1 X Y x x
11775 Resolve to absolute group. Recall that RExC_npar is +1 of
11776 the actual parenthesis group number. For lookahead, we
11777 have to compensate for that. Using the above example, when
11778 we get to Y in the parse, num is 2 and RExC_npar is 6. We
11779 want 7 for +2, and 4 for -2.
11781 if ( paren == '+' ) {
11787 if (paren == '-' && num < 1) {
11789 vFAIL(non_existent_group_msg);
11793 if (num >= RExC_npar) {
11795 /* It might be a forward reference; we can't fail until we
11796 * know, by completing the parse to get all the groups, and
11797 * then reparsing */
11798 if (ALL_PARENS_COUNTED) {
11799 if (num >= RExC_total_parens) {
11801 vFAIL(non_existent_group_msg);
11805 REQUIRE_PARENS_PASS;
11809 /* We keep track how many GOSUB items we have produced.
11810 To start off the ARG2L() of the GOSUB holds its "id",
11811 which is used later in conjunction with RExC_recurse
11812 to calculate the offset we need to jump for the GOSUB,
11813 which it will store in the final representation.
11814 We have to defer the actual calculation until much later
11815 as the regop may move.
11817 ret = reg2Lanode(pRExC_state, GOSUB, num, RExC_recurse_count);
11818 RExC_recurse_count++;
11819 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11820 "%*s%*s Recurse #%" UVuf " to %" IVdf "\n",
11821 22, "| |", (int)(depth * 2 + 1), "",
11822 (UV)ARG(REGNODE_p(ret)),
11823 (IV)ARG2L(REGNODE_p(ret))));
11824 RExC_seen |= REG_RECURSE_SEEN;
11826 Set_Node_Length(REGNODE_p(ret),
11827 1 + regarglen[OP(REGNODE_p(ret))]); /* MJD */
11828 Set_Node_Offset(REGNODE_p(ret), parse_start); /* MJD */
11830 *flagp |= POSTPONED;
11831 assert(*RExC_parse == ')');
11832 nextchar(pRExC_state);
11837 case '?': /* (??...) */
11839 if (*RExC_parse != '{') {
11840 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
11841 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11843 "Sequence (%" UTF8f "...) not recognized",
11844 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
11845 NOT_REACHED; /*NOTREACHED*/
11847 *flagp |= POSTPONED;
11851 case '{': /* (?{...}) */
11854 struct reg_code_block *cb;
11857 RExC_seen_zerolen++;
11859 if ( !pRExC_state->code_blocks
11860 || pRExC_state->code_index
11861 >= pRExC_state->code_blocks->count
11862 || pRExC_state->code_blocks->cb[pRExC_state->code_index].start
11863 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
11866 if (RExC_pm_flags & PMf_USE_RE_EVAL)
11867 FAIL("panic: Sequence (?{...}): no code block found\n");
11868 FAIL("Eval-group not allowed at runtime, use re 'eval'");
11870 /* this is a pre-compiled code block (?{...}) */
11871 cb = &pRExC_state->code_blocks->cb[pRExC_state->code_index];
11872 RExC_parse = RExC_start + cb->end;
11874 if (cb->src_regex) {
11875 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
11876 RExC_rxi->data->data[n] =
11877 (void*)SvREFCNT_inc((SV*)cb->src_regex);
11878 RExC_rxi->data->data[n+1] = (void*)o;
11881 n = add_data(pRExC_state,
11882 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
11883 RExC_rxi->data->data[n] = (void*)o;
11885 pRExC_state->code_index++;
11886 nextchar(pRExC_state);
11889 regnode_offset eval;
11890 ret = reg_node(pRExC_state, LOGICAL);
11892 eval = reg2Lanode(pRExC_state, EVAL,
11895 /* for later propagation into (??{})
11897 RExC_flags & RXf_PMf_COMPILETIME
11899 FLAGS(REGNODE_p(ret)) = 2;
11900 if (! REGTAIL(pRExC_state, ret, eval)) {
11901 REQUIRE_BRANCHJ(flagp, 0);
11903 /* deal with the length of this later - MJD */
11906 ret = reg2Lanode(pRExC_state, EVAL, n, 0);
11907 Set_Node_Length(REGNODE_p(ret), RExC_parse - parse_start + 1);
11908 Set_Node_Offset(REGNODE_p(ret), parse_start);
11911 case '(': /* (?(?{...})...) and (?(?=...)...) */
11914 const int DEFINE_len = sizeof("DEFINE") - 1;
11915 if ( RExC_parse < RExC_end - 1
11916 && ( ( RExC_parse[0] == '?' /* (?(?...)) */
11917 && ( RExC_parse[1] == '='
11918 || RExC_parse[1] == '!'
11919 || RExC_parse[1] == '<'
11920 || RExC_parse[1] == '{'))
11921 || ( RExC_parse[0] == '*' /* (?(*...)) */
11922 && ( memBEGINs(RExC_parse + 1,
11923 (Size_t) (RExC_end - (RExC_parse + 1)),
11925 || memBEGINs(RExC_parse + 1,
11926 (Size_t) (RExC_end - (RExC_parse + 1)),
11928 || memBEGINs(RExC_parse + 1,
11929 (Size_t) (RExC_end - (RExC_parse + 1)),
11931 || memBEGINs(RExC_parse + 1,
11932 (Size_t) (RExC_end - (RExC_parse + 1)),
11934 || memBEGINs(RExC_parse + 1,
11935 (Size_t) (RExC_end - (RExC_parse + 1)),
11936 "positive_lookahead:")
11937 || memBEGINs(RExC_parse + 1,
11938 (Size_t) (RExC_end - (RExC_parse + 1)),
11939 "positive_lookbehind:")
11940 || memBEGINs(RExC_parse + 1,
11941 (Size_t) (RExC_end - (RExC_parse + 1)),
11942 "negative_lookahead:")
11943 || memBEGINs(RExC_parse + 1,
11944 (Size_t) (RExC_end - (RExC_parse + 1)),
11945 "negative_lookbehind:"))))
11946 ) { /* Lookahead or eval. */
11948 regnode_offset tail;
11950 ret = reg_node(pRExC_state, LOGICAL);
11951 FLAGS(REGNODE_p(ret)) = 1;
11953 tail = reg(pRExC_state, 1, &flag, depth+1);
11954 RETURN_FAIL_ON_RESTART(flag, flagp);
11955 if (! REGTAIL(pRExC_state, ret, tail)) {
11956 REQUIRE_BRANCHJ(flagp, 0);
11960 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
11961 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
11963 char ch = RExC_parse[0] == '<' ? '>' : '\'';
11964 char *name_start= RExC_parse++;
11966 SV *sv_dat=reg_scan_name(pRExC_state, REG_RSN_RETURN_DATA);
11967 if ( RExC_parse == name_start
11968 || RExC_parse >= RExC_end
11969 || *RExC_parse != ch)
11971 vFAIL2("Sequence (?(%c... not terminated",
11972 (ch == '>' ? '<' : ch));
11976 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11977 RExC_rxi->data->data[num]=(void*)sv_dat;
11978 SvREFCNT_inc_simple_void_NN(sv_dat);
11980 ret = reganode(pRExC_state, GROUPPN, num);
11981 goto insert_if_check_paren;
11983 else if (memBEGINs(RExC_parse,
11984 (STRLEN) (RExC_end - RExC_parse),
11987 ret = reganode(pRExC_state, DEFINEP, 0);
11988 RExC_parse += DEFINE_len;
11990 goto insert_if_check_paren;
11992 else if (RExC_parse[0] == 'R') {
11994 /* parno == 0 => /(?(R)YES|NO)/ "in any form of recursion OR eval"
11995 * parno == 1 => /(?(R0)YES|NO)/ "in GOSUB (?0) / (?R)"
11996 * parno == 2 => /(?(R1)YES|NO)/ "in GOSUB (?1) (parno-1)"
11999 if (RExC_parse[0] == '0') {
12003 else if (inRANGE(RExC_parse[0], '1', '9')) {
12006 if (grok_atoUV(RExC_parse, &uv, &endptr)
12009 parno = (I32)uv + 1;
12010 RExC_parse = (char*)endptr;
12012 /* else "Switch condition not recognized" below */
12013 } else if (RExC_parse[0] == '&') {
12016 sv_dat = reg_scan_name(pRExC_state,
12017 REG_RSN_RETURN_DATA);
12019 parno = 1 + *((I32 *)SvPVX(sv_dat));
12021 ret = reganode(pRExC_state, INSUBP, parno);
12022 goto insert_if_check_paren;
12024 else if (inRANGE(RExC_parse[0], '1', '9')) {
12029 if (grok_atoUV(RExC_parse, &uv, &endptr)
12033 RExC_parse = (char*)endptr;
12036 vFAIL("panic: grok_atoUV returned FALSE");
12038 ret = reganode(pRExC_state, GROUPP, parno);
12040 insert_if_check_paren:
12041 if (UCHARAT(RExC_parse) != ')') {
12043 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
12045 vFAIL("Switch condition not recognized");
12047 nextchar(pRExC_state);
12049 if (! REGTAIL(pRExC_state, ret, reganode(pRExC_state,
12052 REQUIRE_BRANCHJ(flagp, 0);
12054 br = regbranch(pRExC_state, &flags, 1, depth+1);
12056 RETURN_FAIL_ON_RESTART(flags,flagp);
12057 FAIL2("panic: regbranch returned failure, flags=%#" UVxf,
12060 if (! REGTAIL(pRExC_state, br, reganode(pRExC_state,
12063 REQUIRE_BRANCHJ(flagp, 0);
12065 c = UCHARAT(RExC_parse);
12066 nextchar(pRExC_state);
12067 if (flags&HASWIDTH)
12068 *flagp |= HASWIDTH;
12071 vFAIL("(?(DEFINE)....) does not allow branches");
12073 /* Fake one for optimizer. */
12074 lastbr = reganode(pRExC_state, IFTHEN, 0);
12076 if (!regbranch(pRExC_state, &flags, 1, depth+1)) {
12077 RETURN_FAIL_ON_RESTART(flags, flagp);
12078 FAIL2("panic: regbranch returned failure, flags=%#" UVxf,
12081 if (! REGTAIL(pRExC_state, ret, lastbr)) {
12082 REQUIRE_BRANCHJ(flagp, 0);
12084 if (flags&HASWIDTH)
12085 *flagp |= HASWIDTH;
12086 c = UCHARAT(RExC_parse);
12087 nextchar(pRExC_state);
12092 if (RExC_parse >= RExC_end)
12093 vFAIL("Switch (?(condition)... not terminated");
12095 vFAIL("Switch (?(condition)... contains too many branches");
12097 ender = reg_node(pRExC_state, TAIL);
12098 if (! REGTAIL(pRExC_state, br, ender)) {
12099 REQUIRE_BRANCHJ(flagp, 0);
12102 if (! REGTAIL(pRExC_state, lastbr, ender)) {
12103 REQUIRE_BRANCHJ(flagp, 0);
12105 if (! REGTAIL(pRExC_state,
12108 NEXTOPER(REGNODE_p(lastbr)))),
12111 REQUIRE_BRANCHJ(flagp, 0);
12115 if (! REGTAIL(pRExC_state, ret, ender)) {
12116 REQUIRE_BRANCHJ(flagp, 0);
12118 #if 0 /* Removing this doesn't cause failures in the test suite -- khw */
12119 RExC_size++; /* XXX WHY do we need this?!!
12120 For large programs it seems to be required
12121 but I can't figure out why. -- dmq*/
12126 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
12128 vFAIL("Unknown switch condition (?(...))");
12130 case '[': /* (?[ ... ]) */
12131 return handle_regex_sets(pRExC_state, NULL, flagp, depth+1,
12133 case 0: /* A NUL */
12134 RExC_parse--; /* for vFAIL to print correctly */
12135 vFAIL("Sequence (? incomplete");
12139 if (RExC_strict) { /* [perl #132851] */
12140 ckWARNreg(RExC_parse, "Empty (?) without any modifiers");
12143 case '*': /* If you want to support (?*...), first reconcile with GH #17363 */
12145 default: /* e.g., (?i) */
12146 RExC_parse = (char *) seqstart + 1;
12148 parse_lparen_question_flags(pRExC_state);
12149 if (UCHARAT(RExC_parse) != ':') {
12150 if (RExC_parse < RExC_end)
12151 nextchar(pRExC_state);
12156 nextchar(pRExC_state);
12161 else if (!(RExC_flags & RXf_PMf_NOCAPTURE)) { /* (...) */
12165 if (! ALL_PARENS_COUNTED) {
12166 /* If we are in our first pass through (and maybe only pass),
12167 * we need to allocate memory for the capturing parentheses
12171 if (!RExC_parens_buf_size) {
12172 /* first guess at number of parens we might encounter */
12173 RExC_parens_buf_size = 10;
12175 /* setup RExC_open_parens, which holds the address of each
12176 * OPEN tag, and to make things simpler for the 0 index the
12177 * start of the program - this is used later for offsets */
12178 Newxz(RExC_open_parens, RExC_parens_buf_size,
12180 RExC_open_parens[0] = 1; /* +1 for REG_MAGIC */
12182 /* setup RExC_close_parens, which holds the address of each
12183 * CLOSE tag, and to make things simpler for the 0 index
12184 * the end of the program - this is used later for offsets
12186 Newxz(RExC_close_parens, RExC_parens_buf_size,
12188 /* we dont know where end op starts yet, so we dont need to
12189 * set RExC_close_parens[0] like we do RExC_open_parens[0]
12192 else if (RExC_npar > RExC_parens_buf_size) {
12193 I32 old_size = RExC_parens_buf_size;
12195 RExC_parens_buf_size *= 2;
12197 Renew(RExC_open_parens, RExC_parens_buf_size,
12199 Zero(RExC_open_parens + old_size,
12200 RExC_parens_buf_size - old_size, regnode_offset);
12202 Renew(RExC_close_parens, RExC_parens_buf_size,
12204 Zero(RExC_close_parens + old_size,
12205 RExC_parens_buf_size - old_size, regnode_offset);
12209 ret = reganode(pRExC_state, OPEN, parno);
12210 if (!RExC_nestroot)
12211 RExC_nestroot = parno;
12212 if (RExC_open_parens && !RExC_open_parens[parno])
12214 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12215 "%*s%*s Setting open paren #%" IVdf " to %zu\n",
12216 22, "| |", (int)(depth * 2 + 1), "",
12218 RExC_open_parens[parno]= ret;
12221 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
12222 Set_Node_Offset(REGNODE_p(ret), RExC_parse); /* MJD */
12225 /* with RXf_PMf_NOCAPTURE treat (...) as (?:...) */
12234 /* Pick up the branches, linking them together. */
12235 parse_start = RExC_parse; /* MJD */
12236 br = regbranch(pRExC_state, &flags, 1, depth+1);
12238 /* branch_len = (paren != 0); */
12241 RETURN_FAIL_ON_RESTART(flags, flagp);
12242 FAIL2("panic: regbranch returned failure, flags=%#" UVxf, (UV) flags);
12244 if (*RExC_parse == '|') {
12245 if (RExC_use_BRANCHJ) {
12246 reginsert(pRExC_state, BRANCHJ, br, depth+1);
12249 reginsert(pRExC_state, BRANCH, br, depth+1);
12250 Set_Node_Length(REGNODE_p(br), paren != 0);
12251 Set_Node_Offset_To_R(br, parse_start-RExC_start);
12255 else if (paren == ':') {
12256 *flagp |= flags&SIMPLE;
12258 if (is_open) { /* Starts with OPEN. */
12259 if (! REGTAIL(pRExC_state, ret, br)) { /* OPEN -> first. */
12260 REQUIRE_BRANCHJ(flagp, 0);
12263 else if (paren != '?') /* Not Conditional */
12265 *flagp |= flags & (HASWIDTH | POSTPONED);
12267 while (*RExC_parse == '|') {
12268 if (RExC_use_BRANCHJ) {
12271 ender = reganode(pRExC_state, LONGJMP, 0);
12273 /* Append to the previous. */
12274 shut_gcc_up = REGTAIL(pRExC_state,
12275 REGNODE_OFFSET(NEXTOPER(NEXTOPER(REGNODE_p(lastbr)))),
12277 PERL_UNUSED_VAR(shut_gcc_up);
12279 nextchar(pRExC_state);
12280 if (freeze_paren) {
12281 if (RExC_npar > after_freeze)
12282 after_freeze = RExC_npar;
12283 RExC_npar = freeze_paren;
12285 br = regbranch(pRExC_state, &flags, 0, depth+1);
12288 RETURN_FAIL_ON_RESTART(flags, flagp);
12289 FAIL2("panic: regbranch returned failure, flags=%#" UVxf, (UV) flags);
12291 if (! REGTAIL(pRExC_state, lastbr, br)) { /* BRANCH -> BRANCH. */
12292 REQUIRE_BRANCHJ(flagp, 0);
12295 *flagp |= flags & (HASWIDTH | POSTPONED);
12298 if (have_branch || paren != ':') {
12301 /* Make a closing node, and hook it on the end. */
12304 ender = reg_node(pRExC_state, TAIL);
12307 ender = reganode(pRExC_state, CLOSE, parno);
12308 if ( RExC_close_parens ) {
12309 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12310 "%*s%*s Setting close paren #%" IVdf " to %zu\n",
12311 22, "| |", (int)(depth * 2 + 1), "",
12312 (IV)parno, ender));
12313 RExC_close_parens[parno]= ender;
12314 if (RExC_nestroot == parno)
12317 Set_Node_Offset(REGNODE_p(ender), RExC_parse+1); /* MJD */
12318 Set_Node_Length(REGNODE_p(ender), 1); /* MJD */
12321 ender = reg_node(pRExC_state, SRCLOSE);
12322 RExC_in_script_run = 0;
12332 *flagp &= ~HASWIDTH;
12334 case 't': /* aTomic */
12336 ender = reg_node(pRExC_state, SUCCEED);
12339 ender = reg_node(pRExC_state, END);
12340 assert(!RExC_end_op); /* there can only be one! */
12341 RExC_end_op = REGNODE_p(ender);
12342 if (RExC_close_parens) {
12343 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12344 "%*s%*s Setting close paren #0 (END) to %zu\n",
12345 22, "| |", (int)(depth * 2 + 1), "",
12348 RExC_close_parens[0]= ender;
12353 DEBUG_PARSE_MSG("lsbr");
12354 regprop(RExC_rx, RExC_mysv1, REGNODE_p(lastbr), NULL, pRExC_state);
12355 regprop(RExC_rx, RExC_mysv2, REGNODE_p(ender), NULL, pRExC_state);
12356 Perl_re_printf( aTHX_ "~ tying lastbr %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
12357 SvPV_nolen_const(RExC_mysv1),
12359 SvPV_nolen_const(RExC_mysv2),
12361 (IV)(ender - lastbr)
12364 if (! REGTAIL(pRExC_state, lastbr, ender)) {
12365 REQUIRE_BRANCHJ(flagp, 0);
12369 char is_nothing= 1;
12371 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
12373 /* Hook the tails of the branches to the closing node. */
12374 for (br = REGNODE_p(ret); br; br = regnext(br)) {
12375 const U8 op = PL_regkind[OP(br)];
12376 if (op == BRANCH) {
12377 if (! REGTAIL_STUDY(pRExC_state,
12378 REGNODE_OFFSET(NEXTOPER(br)),
12381 REQUIRE_BRANCHJ(flagp, 0);
12383 if ( OP(NEXTOPER(br)) != NOTHING
12384 || regnext(NEXTOPER(br)) != REGNODE_p(ender))
12387 else if (op == BRANCHJ) {
12388 bool shut_gcc_up = REGTAIL_STUDY(pRExC_state,
12389 REGNODE_OFFSET(NEXTOPER(NEXTOPER(br))),
12391 PERL_UNUSED_VAR(shut_gcc_up);
12392 /* for now we always disable this optimisation * /
12393 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
12394 || regnext(NEXTOPER(NEXTOPER(br))) != REGNODE_p(ender))
12400 regnode * ret_as_regnode = REGNODE_p(ret);
12401 br= PL_regkind[OP(ret_as_regnode)] != BRANCH
12402 ? regnext(ret_as_regnode)
12405 DEBUG_PARSE_MSG("NADA");
12406 regprop(RExC_rx, RExC_mysv1, ret_as_regnode,
12407 NULL, pRExC_state);
12408 regprop(RExC_rx, RExC_mysv2, REGNODE_p(ender),
12409 NULL, pRExC_state);
12410 Perl_re_printf( aTHX_ "~ converting ret %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
12411 SvPV_nolen_const(RExC_mysv1),
12412 (IV)REG_NODE_NUM(ret_as_regnode),
12413 SvPV_nolen_const(RExC_mysv2),
12419 if (OP(REGNODE_p(ender)) == TAIL) {
12421 RExC_emit= REGNODE_OFFSET(br) + 1;
12424 for ( opt= br + 1; opt < REGNODE_p(ender) ; opt++ )
12425 OP(opt)= OPTIMIZED;
12426 NEXT_OFF(br)= REGNODE_p(ender) - br;
12434 /* Even/odd or x=don't care: 010101x10x */
12435 static const char parens[] = "=!aA<,>Bbt";
12436 /* flag below is set to 0 up through 'A'; 1 for larger */
12438 if (paren && (p = strchr(parens, paren))) {
12439 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
12440 int flag = (p - parens) > 3;
12442 if (paren == '>' || paren == 't') {
12443 node = SUSPEND, flag = 0;
12446 reginsert(pRExC_state, node, ret, depth+1);
12447 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
12448 Set_Node_Offset(REGNODE_p(ret), parse_start + 1);
12449 FLAGS(REGNODE_p(ret)) = flag;
12450 if (! REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL)))
12452 REQUIRE_BRANCHJ(flagp, 0);
12457 /* Check for proper termination. */
12459 /* restore original flags, but keep (?p) and, if we've encountered
12460 * something in the parse that changes /d rules into /u, keep the /u */
12461 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
12462 if (DEPENDS_SEMANTICS && toUSE_UNI_CHARSET_NOT_DEPENDS) {
12463 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
12465 if (RExC_parse >= RExC_end || UCHARAT(RExC_parse) != ')') {
12466 RExC_parse = oregcomp_parse;
12467 vFAIL("Unmatched (");
12469 nextchar(pRExC_state);
12471 else if (!paren && RExC_parse < RExC_end) {
12472 if (*RExC_parse == ')') {
12474 vFAIL("Unmatched )");
12477 FAIL("Junk on end of regexp"); /* "Can't happen". */
12478 NOT_REACHED; /* NOTREACHED */
12481 if (after_freeze > RExC_npar)
12482 RExC_npar = after_freeze;
12484 RExC_in_lookaround = was_in_lookaround;
12490 - regbranch - one alternative of an | operator
12492 * Implements the concatenation operator.
12494 * On success, returns the offset at which any next node should be placed into
12495 * the regex engine program being compiled.
12497 * Returns 0 otherwise, setting flagp to RESTART_PARSE if the parse needs
12498 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to
12501 STATIC regnode_offset
12502 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
12504 regnode_offset ret;
12505 regnode_offset chain = 0;
12506 regnode_offset latest;
12507 I32 flags = 0, c = 0;
12508 DECLARE_AND_GET_RE_DEBUG_FLAGS;
12510 PERL_ARGS_ASSERT_REGBRANCH;
12512 DEBUG_PARSE("brnc");
12517 if (RExC_use_BRANCHJ)
12518 ret = reganode(pRExC_state, BRANCHJ, 0);
12520 ret = reg_node(pRExC_state, BRANCH);
12521 Set_Node_Length(REGNODE_p(ret), 1);
12525 *flagp = 0; /* Initialize. */
12527 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
12528 FALSE /* Don't force to /x */ );
12529 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
12530 flags &= ~TRYAGAIN;
12531 latest = regpiece(pRExC_state, &flags, depth+1);
12533 if (flags & TRYAGAIN)
12535 RETURN_FAIL_ON_RESTART(flags, flagp);
12536 FAIL2("panic: regpiece returned failure, flags=%#" UVxf, (UV) flags);
12540 *flagp |= flags&(HASWIDTH|POSTPONED);
12542 /* FIXME adding one for every branch after the first is probably
12543 * excessive now we have TRIE support. (hv) */
12545 if (! REGTAIL(pRExC_state, chain, latest)) {
12546 /* XXX We could just redo this branch, but figuring out what
12547 * bookkeeping needs to be reset is a pain, and it's likely
12548 * that other branches that goto END will also be too large */
12549 REQUIRE_BRANCHJ(flagp, 0);
12555 if (chain == 0) { /* Loop ran zero times. */
12556 chain = reg_node(pRExC_state, NOTHING);
12561 *flagp |= flags&SIMPLE;
12573 #ifndef PERL_IN_XSUB_RE
12575 Perl_regcurly(const char *s, const char *e, const char * result[5])
12577 /* This function matches a {m,n} quantifier. When called with a NULL final
12578 * argument, it simply parses the input from 's' up through 'e-1', and
12579 * returns a boolean as to whether or not this input is syntactically a
12580 * {m,n} quantifier.
12582 * When called with a non-NULL final parameter, and when the function
12583 * returns TRUE, it additionally stores information into the array
12584 * specified by that parameter about what it found in the parse. The
12585 * parameter must be a pointer into a 5 element array of 'const char *'
12586 * elements. The returned information is as follows:
12587 * result[RBRACE] points to the closing brace
12588 * result[MIN_S] points to the first byte of the lower bound
12589 * result[MIN_E] points to one beyond the final byte of the lower bound
12590 * result[MAX_S] points to the first byte of the upper bound
12591 * result[MAX_E] points to one beyond the final byte of the upper bound
12593 * If the quantifier is of the form {m,} (meaning an infinite upper
12594 * bound), result[MAX_E] is set to result[MAX_S]; what they actually point
12595 * to is irrelevant, just that it's the same place
12597 * If instead the quantifier is of the form {m} there is actually only
12598 * one bound, and both the upper and lower result[] elements are set to
12601 * This function checks only for syntactic validity; it leaves checking for
12602 * semantic validity and raising any diagnostics to the caller. This
12603 * function is called in multiple places to check for syntax, but only from
12604 * one for semantics. It makes it as simple as possible for the
12605 * syntax-only callers, while furnishing just enough information for the
12609 const char * min_start = NULL;
12610 const char * max_start = NULL;
12611 const char * min_end = NULL;
12612 const char * max_end = NULL;
12614 bool has_comma = FALSE;
12616 PERL_ARGS_ASSERT_REGCURLY;
12618 if (s >= e || *s++ != '{')
12621 while (s < e && isBLANK(*s)) {
12629 } while (s < e && isDIGIT(*s));
12633 while (s < e && isBLANK(*s)) {
12641 while (s < e && isBLANK(*s)) {
12649 } while (s < e && isDIGIT(*s));
12654 while (s < e && isBLANK(*s)) {
12657 /* Need at least one number */
12658 if (s >= e || *s != '}' || (! min_start && ! max_end)) {
12664 result[RBRACE] = s;
12666 result[MIN_S] = min_start;
12667 result[MIN_E] = min_end;
12670 result[MAX_S] = max_start;
12671 result[MAX_E] = max_end;
12674 /* Having no value after the comma is signalled by setting
12675 * start and end to the same value. What that value is isn't
12676 * relevant; NULL is chosen simply because it will fail if the
12677 * caller mistakenly uses it */
12678 result[MAX_S] = result[MAX_E] = NULL;
12681 else { /* No comma means lower and upper bounds are the same */
12682 result[MAX_S] = min_start;
12683 result[MAX_E] = min_end;
12692 S_get_quantifier_value(pTHX_ RExC_state_t *pRExC_state,
12693 const char * start, const char * end)
12695 /* This is a helper function for regpiece() to compute, given the
12696 * quantifier {m,n}, the value of either m or n, based on the starting
12697 * position 'start' in the string, through the byte 'end-1', returning it
12698 * if valid, and failing appropriately if not. It knows the restrictions
12699 * imposed on quantifier values */
12702 STATIC_ASSERT_DECL(REG_INFTY <= U32_MAX);
12704 PERL_ARGS_ASSERT_GET_QUANTIFIER_VALUE;
12706 if (grok_atoUV(start, &uv, &end)) {
12707 if (uv < REG_INFTY) { /* A valid, small-enough number */
12711 else if (*start == '0') { /* grok_atoUV() fails for only two reasons:
12712 leading zeros or overflow */
12713 RExC_parse = (char * ) end;
12715 /* Perhaps too generic a msg for what is only failure from having
12716 * leading zeros, but this is how it's always behaved. */
12717 vFAIL("Invalid quantifier in {,}");
12718 NOT_REACHED; /*NOTREACHED*/
12721 /* Here, found a quantifier, but was too large; either it overflowed or was
12722 * too big a legal number */
12723 RExC_parse = (char * ) end;
12724 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
12726 NOT_REACHED; /*NOTREACHED*/
12727 return U32_MAX; /* Perhaps some compilers will be expecting a return */
12731 - regpiece - something followed by possible quantifier * + ? {n,m}
12733 * Note that the branching code sequences used for ? and the general cases
12734 * of * and + are somewhat optimized: they use the same NOTHING node as
12735 * both the endmarker for their branch list and the body of the last branch.
12736 * It might seem that this node could be dispensed with entirely, but the
12737 * endmarker role is not redundant.
12739 * On success, returns the offset at which any next node should be placed into
12740 * the regex engine program being compiled.
12742 * Returns 0 otherwise, with *flagp set to indicate why:
12743 * TRYAGAIN if regatom() returns 0 with TRYAGAIN.
12744 * RESTART_PARSE if the parse needs to be restarted, or'd with
12745 * NEED_UTF8 if the pattern needs to be upgraded to UTF-8.
12747 STATIC regnode_offset
12748 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
12750 regnode_offset ret;
12753 const char * const origparse = RExC_parse;
12755 I32 max = REG_INFTY;
12756 #ifdef RE_TRACK_PATTERN_OFFSETS
12760 /* Save the original in case we change the emitted regop to a FAIL. */
12761 const regnode_offset orig_emit = RExC_emit;
12763 DECLARE_AND_GET_RE_DEBUG_FLAGS;
12765 PERL_ARGS_ASSERT_REGPIECE;
12767 DEBUG_PARSE("piec");
12769 ret = regatom(pRExC_state, &flags, depth+1);
12771 RETURN_FAIL_ON_RESTART_OR_FLAGS(flags, flagp, TRYAGAIN);
12772 FAIL2("panic: regatom returned failure, flags=%#" UVxf, (UV) flags);
12775 #ifdef RE_TRACK_PATTERN_OFFSETS
12776 parse_start = RExC_parse;
12781 const char * regcurly_return[5];
12784 nextchar(pRExC_state);
12789 nextchar(pRExC_state);
12794 nextchar(pRExC_state);
12798 case '{': /* A '{' may or may not indicate a quantifier; call regcurly()
12799 to determine which */
12800 if (regcurly(RExC_parse, RExC_end, regcurly_return)) {
12801 const char * min_start = regcurly_return[MIN_S];
12802 const char * min_end = regcurly_return[MIN_E];
12803 const char * max_start = regcurly_return[MAX_S];
12804 const char * max_end = regcurly_return[MAX_E];
12807 min = get_quantifier_value(pRExC_state, min_start, min_end);
12813 if (max_start == max_end) { /* Was of the form {m,} */
12816 else if (max_start == min_start) { /* Was of the form {m} */
12819 else { /* Was of the form {m,n} */
12820 assert(max_end >= max_start);
12822 max = get_quantifier_value(pRExC_state, max_start, max_end);
12825 RExC_parse = (char *) regcurly_return[RBRACE];
12826 nextchar(pRExC_state);
12828 if (max < min) { /* If can't match, warn and optimize to fail
12830 reginsert(pRExC_state, OPFAIL, orig_emit, depth+1);
12831 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
12832 NEXT_OFF(REGNODE_p(orig_emit)) =
12833 regarglen[OPFAIL] + NODE_STEP_REGNODE;
12836 else if (min == max && *RExC_parse == '?') {
12837 ckWARN2reg(RExC_parse + 1,
12838 "Useless use of greediness modifier '%c'",
12843 } /* End of is {m,n} */
12845 /* Here was a '{', but what followed it didn't form a quantifier. */
12851 NOT_REACHED; /*NOTREACHED*/
12854 /* Here we have a quantifier, and have calculated 'min' and 'max'.
12856 * Check and possibly adjust a zero width operand */
12857 if (! (flags & (HASWIDTH|POSTPONED))) {
12858 if (max > REG_INFTY/3) {
12859 if (origparse[0] == '\\' && origparse[1] == 'K') {
12861 "%" UTF8f " is forbidden - matches null string"
12863 UTF8fARG(UTF, (RExC_parse >= origparse
12864 ? RExC_parse - origparse
12868 ckWARN2reg(RExC_parse,
12869 "%" UTF8f " matches null string many times",
12870 UTF8fARG(UTF, (RExC_parse >= origparse
12871 ? RExC_parse - origparse
12877 /* There's no point in trying to match something 0 length more than
12878 * once except for extra side effects, which we don't have here since
12888 /* If this is a code block pass it up */
12889 *flagp |= (flags & POSTPONED);
12892 *flagp |= (flags & HASWIDTH);
12893 if (max == REG_INFTY)
12894 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
12897 /* 'SIMPLE' operands don't require full generality */
12898 if ((flags&SIMPLE)) {
12899 if (max == REG_INFTY) {
12901 if (UNLIKELY(RExC_pm_flags & PMf_WILDCARD)) {
12902 goto min0_maxINF_wildcard_forbidden;
12905 reginsert(pRExC_state, STAR, ret, depth+1);
12909 else if (min == 1) {
12910 reginsert(pRExC_state, PLUS, ret, depth+1);
12916 /* Here, SIMPLE, but not the '*' and '+' special cases */
12918 MARK_NAUGHTY_EXP(2, 2);
12919 reginsert(pRExC_state, CURLY, ret, depth+1);
12920 Set_Node_Offset(REGNODE_p(ret), parse_start+1); /* MJD */
12921 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
12923 else { /* not SIMPLE */
12924 const regnode_offset w = reg_node(pRExC_state, WHILEM);
12926 FLAGS(REGNODE_p(w)) = 0;
12927 if (! REGTAIL(pRExC_state, ret, w)) {
12928 REQUIRE_BRANCHJ(flagp, 0);
12930 if (RExC_use_BRANCHJ) {
12931 reginsert(pRExC_state, LONGJMP, ret, depth+1);
12932 reginsert(pRExC_state, NOTHING, ret, depth+1);
12933 NEXT_OFF(REGNODE_p(ret)) = 3; /* Go over LONGJMP. */
12935 reginsert(pRExC_state, CURLYX, ret, depth+1);
12937 Set_Node_Offset(REGNODE_p(ret), parse_start+1);
12938 Set_Node_Length(REGNODE_p(ret),
12939 op == '{' ? (RExC_parse - parse_start) : 1);
12941 if (RExC_use_BRANCHJ)
12942 NEXT_OFF(REGNODE_p(ret)) = 3; /* Go over NOTHING to
12944 if (! REGTAIL(pRExC_state, ret, reg_node(pRExC_state,
12947 REQUIRE_BRANCHJ(flagp, 0);
12949 RExC_whilem_seen++;
12950 MARK_NAUGHTY_EXP(1, 4); /* compound interest */
12953 /* Finish up the CURLY/CURLYX case */
12954 FLAGS(REGNODE_p(ret)) = 0;
12956 ARG1_SET(REGNODE_p(ret), (U16)min);
12957 ARG2_SET(REGNODE_p(ret), (U16)max);
12961 /* Process any greediness modifiers */
12962 if (*RExC_parse == '?') {
12963 nextchar(pRExC_state);
12964 reginsert(pRExC_state, MINMOD, ret, depth+1);
12965 if (! REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE)) {
12966 REQUIRE_BRANCHJ(flagp, 0);
12969 else if (*RExC_parse == '+') {
12970 regnode_offset ender;
12971 nextchar(pRExC_state);
12972 ender = reg_node(pRExC_state, SUCCEED);
12973 if (! REGTAIL(pRExC_state, ret, ender)) {
12974 REQUIRE_BRANCHJ(flagp, 0);
12976 reginsert(pRExC_state, SUSPEND, ret, depth+1);
12977 ender = reg_node(pRExC_state, TAIL);
12978 if (! REGTAIL(pRExC_state, ret, ender)) {
12979 REQUIRE_BRANCHJ(flagp, 0);
12983 /* Forbid extra quantifiers */
12984 if (isQUANTIFIER(RExC_parse, RExC_end)) {
12986 vFAIL("Nested quantifiers");
12991 min0_maxINF_wildcard_forbidden:
12993 /* Here we are in a wildcard match, and the minimum match length is 0, and
12994 * the max could be infinity. This is currently forbidden. The only
12995 * reason is to make it harder to write patterns that take a long long time
12996 * to halt, and because the use of this construct isn't necessary in
12997 * matching Unicode property values */
12999 /* diag_listed_as: Use of %s is not allowed in Unicode property wildcard
13000 subpatterns in regex; marked by <-- HERE in m/%s/
13002 vFAIL("Use of quantifier '*' is not allowed in Unicode property wildcard"
13005 /* Note, don't need to worry about the input being '{0,}', as a '}' isn't
13006 * legal at all in wildcards, so can't get this far */
13008 NOT_REACHED; /*NOTREACHED*/
13012 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state,
13013 regnode_offset * node_p,
13021 /* This routine teases apart the various meanings of \N and returns
13022 * accordingly. The input parameters constrain which meaning(s) is/are valid
13023 * in the current context.
13025 * Exactly one of <node_p> and <code_point_p> must be non-NULL.
13027 * If <code_point_p> is not NULL, the context is expecting the result to be a
13028 * single code point. If this \N instance turns out to a single code point,
13029 * the function returns TRUE and sets *code_point_p to that code point.
13031 * If <node_p> is not NULL, the context is expecting the result to be one of
13032 * the things representable by a regnode. If this \N instance turns out to be
13033 * one such, the function generates the regnode, returns TRUE and sets *node_p
13034 * to point to the offset of that regnode into the regex engine program being
13037 * If this instance of \N isn't legal in any context, this function will
13038 * generate a fatal error and not return.
13040 * On input, RExC_parse should point to the first char following the \N at the
13041 * time of the call. On successful return, RExC_parse will have been updated
13042 * to point to just after the sequence identified by this routine. Also
13043 * *flagp has been updated as needed.
13045 * When there is some problem with the current context and this \N instance,
13046 * the function returns FALSE, without advancing RExC_parse, nor setting
13047 * *node_p, nor *code_point_p, nor *flagp.
13049 * If <cp_count> is not NULL, the caller wants to know the length (in code
13050 * points) that this \N sequence matches. This is set, and the input is
13051 * parsed for errors, even if the function returns FALSE, as detailed below.
13053 * There are 6 possibilities here, as detailed in the next 6 paragraphs.
13055 * Probably the most common case is for the \N to specify a single code point.
13056 * *cp_count will be set to 1, and *code_point_p will be set to that code
13059 * Another possibility is for the input to be an empty \N{}. This is no
13060 * longer accepted, and will generate a fatal error.
13062 * Another possibility is for a custom charnames handler to be in effect which
13063 * translates the input name to an empty string. *cp_count will be set to 0.
13064 * *node_p will be set to a generated NOTHING node.
13066 * Still another possibility is for the \N to mean [^\n]. *cp_count will be
13067 * set to 0. *node_p will be set to a generated REG_ANY node.
13069 * The fifth possibility is that \N resolves to a sequence of more than one
13070 * code points. *cp_count will be set to the number of code points in the
13071 * sequence. *node_p will be set to a generated node returned by this
13072 * function calling S_reg().
13074 * The sixth and final possibility is that it is premature to be calling this
13075 * function; the parse needs to be restarted. This can happen when this
13076 * changes from /d to /u rules, or when the pattern needs to be upgraded to
13077 * UTF-8. The latter occurs only when the fifth possibility would otherwise
13078 * be in effect, and is because one of those code points requires the pattern
13079 * to be recompiled as UTF-8. The function returns FALSE, and sets the
13080 * RESTART_PARSE and NEED_UTF8 flags in *flagp, as appropriate. When this
13081 * happens, the caller needs to desist from continuing parsing, and return
13082 * this information to its caller. This is not set for when there is only one
13083 * code point, as this can be called as part of an ANYOF node, and they can
13084 * store above-Latin1 code points without the pattern having to be in UTF-8.
13086 * For non-single-quoted regexes, the tokenizer has resolved character and
13087 * sequence names inside \N{...} into their Unicode values, normalizing the
13088 * result into what we should see here: '\N{U+c1.c2...}', where c1... are the
13089 * hex-represented code points in the sequence. This is done there because
13090 * the names can vary based on what charnames pragma is in scope at the time,
13091 * so we need a way to take a snapshot of what they resolve to at the time of
13092 * the original parse. [perl #56444].
13094 * That parsing is skipped for single-quoted regexes, so here we may get
13095 * '\N{NAME}', which is parsed now. If the single-quoted regex is something
13096 * like '\N{U+41}', that code point is Unicode, and has to be translated into
13097 * the native character set for non-ASCII platforms. The other possibilities
13098 * are already native, so no translation is done. */
13100 char * endbrace; /* points to '}' following the name */
13101 char * e; /* points to final non-blank before endbrace */
13102 char* p = RExC_parse; /* Temporary */
13104 SV * substitute_parse = NULL;
13109 DECLARE_AND_GET_RE_DEBUG_FLAGS;
13111 PERL_ARGS_ASSERT_GROK_BSLASH_N;
13113 assert(cBOOL(node_p) ^ cBOOL(code_point_p)); /* Exactly one should be set */
13114 assert(! (node_p && cp_count)); /* At most 1 should be set */
13116 if (cp_count) { /* Initialize return for the most common case */
13120 /* The [^\n] meaning of \N ignores spaces and comments under the /x
13121 * modifier. The other meanings do not (except blanks adjacent to and
13122 * within the braces), so use a temporary until we find out which we are
13123 * being called with */
13124 skip_to_be_ignored_text(pRExC_state, &p,
13125 FALSE /* Don't force to /x */ );
13127 /* Disambiguate between \N meaning a named character versus \N meaning
13128 * [^\n]. The latter is assumed when the {...} following the \N is a legal
13129 * quantifier, or if there is no '{' at all */
13130 if (*p != '{' || regcurly(p, RExC_end, NULL)) {
13140 *node_p = reg_node(pRExC_state, REG_ANY);
13141 *flagp |= HASWIDTH|SIMPLE;
13143 Set_Node_Length(REGNODE_p(*(node_p)), 1); /* MJD */
13147 /* The test above made sure that the next real character is a '{', but
13148 * under the /x modifier, it could be separated by space (or a comment and
13149 * \n) and this is not allowed (for consistency with \x{...} and the
13150 * tokenizer handling of \N{NAME}). */
13151 if (*RExC_parse != '{') {
13152 vFAIL("Missing braces on \\N{}");
13155 RExC_parse++; /* Skip past the '{' */
13157 endbrace = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
13158 if (! endbrace) { /* no trailing brace */
13159 vFAIL2("Missing right brace on \\%c{}", 'N');
13162 /* Here, we have decided it should be a named character or sequence. These
13163 * imply Unicode semantics */
13164 REQUIRE_UNI_RULES(flagp, FALSE);
13166 /* \N{_} is what toke.c returns to us to indicate a name that evaluates to
13167 * nothing at all (not allowed under strict) */
13168 if (endbrace - RExC_parse == 1 && *RExC_parse == '_') {
13169 RExC_parse = endbrace;
13171 RExC_parse++; /* Position after the "}" */
13172 vFAIL("Zero length \\N{}");
13178 nextchar(pRExC_state);
13183 *node_p = reg_node(pRExC_state, NOTHING);
13187 while (isBLANK(*RExC_parse)) {
13192 while (RExC_parse < e && isBLANK(*(e-1))) {
13196 if (e - RExC_parse < 2 || ! strBEGINs(RExC_parse, "U+")) {
13198 /* Here, the name isn't of the form U+.... This can happen if the
13199 * pattern is single-quoted, so didn't get evaluated in toke.c. Now
13200 * is the time to find out what the name means */
13202 const STRLEN name_len = e - RExC_parse;
13203 SV * value_sv; /* What does this name evaluate to */
13205 const U8 * value; /* string of name's value */
13206 STRLEN value_len; /* and its length */
13208 /* RExC_unlexed_names is a hash of names that weren't evaluated by
13209 * toke.c, and their values. Make sure is initialized */
13210 if (! RExC_unlexed_names) {
13211 RExC_unlexed_names = newHV();
13214 /* If we have already seen this name in this pattern, use that. This
13215 * allows us to only call the charnames handler once per name per
13216 * pattern. A broken or malicious handler could return something
13217 * different each time, which could cause the results to vary depending
13218 * on if something gets added or subtracted from the pattern that
13219 * causes the number of passes to change, for example */
13220 if ((value_svp = hv_fetch(RExC_unlexed_names, RExC_parse,
13223 value_sv = *value_svp;
13225 else { /* Otherwise we have to go out and get the name */
13226 const char * error_msg = NULL;
13227 value_sv = get_and_check_backslash_N_name(RExC_parse, e,
13231 RExC_parse = endbrace;
13235 /* If no error message, should have gotten a valid return */
13238 /* Save the name's meaning for later use */
13239 if (! hv_store(RExC_unlexed_names, RExC_parse, name_len,
13242 Perl_croak(aTHX_ "panic: hv_store() unexpectedly failed");
13246 /* Here, we have the value the name evaluates to in 'value_sv' */
13247 value = (U8 *) SvPV(value_sv, value_len);
13249 /* See if the result is one code point vs 0 or multiple */
13250 if (inRANGE(value_len, 1, ((UV) SvUTF8(value_sv)
13254 /* Here, exactly one code point. If that isn't what is wanted,
13256 if (! code_point_p) {
13261 /* Convert from string to numeric code point */
13262 *code_point_p = (SvUTF8(value_sv))
13263 ? valid_utf8_to_uvchr(value, NULL)
13266 /* Have parsed this entire single code point \N{...}. *cp_count
13267 * has already been set to 1, so don't do it again. */
13268 RExC_parse = endbrace;
13269 nextchar(pRExC_state);
13271 } /* End of is a single code point */
13273 /* Count the code points, if caller desires. The API says to do this
13274 * even if we will later return FALSE */
13278 *cp_count = (SvUTF8(value_sv))
13279 ? utf8_length(value, value + value_len)
13283 /* Fail if caller doesn't want to handle a multi-code-point sequence.
13284 * But don't back the pointer up if the caller wants to know how many
13285 * code points there are (they need to handle it themselves in this
13294 /* Convert this to a sub-pattern of the form "(?: ... )", and then call
13295 * reg recursively to parse it. That way, it retains its atomicness,
13296 * while not having to worry about any special handling that some code
13297 * points may have. */
13299 substitute_parse = newSVpvs("?:");
13300 sv_catsv(substitute_parse, value_sv);
13301 sv_catpv(substitute_parse, ")");
13303 /* The value should already be native, so no need to convert on EBCDIC
13305 assert(! RExC_recode_x_to_native);
13308 else { /* \N{U+...} */
13309 Size_t count = 0; /* code point count kept internally */
13311 /* We can get to here when the input is \N{U+...} or when toke.c has
13312 * converted a name to the \N{U+...} form. This include changing a
13313 * name that evaluates to multiple code points to \N{U+c1.c2.c3 ...} */
13315 RExC_parse += 2; /* Skip past the 'U+' */
13317 /* Code points are separated by dots. The '}' terminates the whole
13320 do { /* Loop until the ending brace */
13321 I32 flags = PERL_SCAN_SILENT_OVERFLOW
13322 | PERL_SCAN_SILENT_ILLDIGIT
13323 | PERL_SCAN_NOTIFY_ILLDIGIT
13324 | PERL_SCAN_ALLOW_MEDIAL_UNDERSCORES
13325 | PERL_SCAN_DISALLOW_PREFIX;
13326 STRLEN len = e - RExC_parse;
13328 char * start_digit = RExC_parse;
13329 UV cp = grok_hex(RExC_parse, &len, &flags, &overflow_value);
13334 vFAIL("Invalid hexadecimal number in \\N{U+...}");
13339 if (cp > MAX_LEGAL_CP) {
13340 vFAIL(form_cp_too_large_msg(16, start_digit, len, 0));
13343 if (RExC_parse >= e) { /* Got to the closing '}' */
13348 /* Here, is a single code point; fail if doesn't want that */
13349 if (! code_point_p) {
13354 /* A single code point is easy to handle; just return it */
13355 *code_point_p = UNI_TO_NATIVE(cp);
13356 RExC_parse = endbrace;
13357 nextchar(pRExC_state);
13361 /* Here, the parse stopped bfore the ending brace. This is legal
13362 * only if that character is a dot separating code points, like a
13363 * multiple character sequence (of the form "\N{U+c1.c2. ... }".
13364 * So the next character must be a dot (and the one after that
13365 * can't be the ending brace, or we'd have something like
13368 if (*RExC_parse != '.' || RExC_parse + 1 >= e) {
13369 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
13370 ? UTF8SKIP(RExC_parse)
13372 RExC_parse = MIN(e, RExC_parse);/* Guard against malformed utf8
13377 /* Here, looks like its really a multiple character sequence. Fail
13378 * if that's not what the caller wants. But continue with counting
13379 * and error checking if they still want a count */
13380 if (! node_p && ! cp_count) {
13384 /* What is done here is to convert this to a sub-pattern of the
13385 * form \x{char1}\x{char2}... and then call reg recursively to
13386 * parse it (enclosing in "(?: ... )" ). That way, it retains its
13387 * atomicness, while not having to worry about special handling
13388 * that some code points may have. We don't create a subpattern,
13389 * but go through the motions of code point counting and error
13390 * checking, if the caller doesn't want a node returned. */
13392 if (node_p && ! substitute_parse) {
13393 substitute_parse = newSVpvs("?:");
13399 /* Convert to notation the rest of the code understands */
13400 sv_catpvs(substitute_parse, "\\x{");
13401 sv_catpvn(substitute_parse, start_digit,
13402 RExC_parse - start_digit);
13403 sv_catpvs(substitute_parse, "}");
13406 /* Move to after the dot (or ending brace the final time through.)
13411 } while (RExC_parse < e);
13413 if (! node_p) { /* Doesn't want the node */
13420 sv_catpvs(substitute_parse, ")");
13422 /* The values are Unicode, and therefore have to be converted to native
13423 * on a non-Unicode (meaning non-ASCII) platform. */
13424 SET_recode_x_to_native(1);
13427 /* Here, we have the string the name evaluates to, ready to be parsed,
13428 * stored in 'substitute_parse' as a series of valid "\x{...}\x{...}"
13429 * constructs. This can be called from within a substitute parse already.
13430 * The error reporting mechanism doesn't work for 2 levels of this, but the
13431 * code above has validated this new construct, so there should be no
13432 * errors generated by the below. And this isn' an exact copy, so the
13433 * mechanism to seamlessly deal with this won't work, so turn off warnings
13435 save_start = RExC_start;
13436 orig_end = RExC_end;
13438 RExC_parse = RExC_start = SvPVX(substitute_parse);
13439 RExC_end = RExC_parse + SvCUR(substitute_parse);
13440 TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE;
13442 *node_p = reg(pRExC_state, 1, &flags, depth+1);
13444 /* Restore the saved values */
13446 RExC_start = save_start;
13447 RExC_parse = endbrace;
13448 RExC_end = orig_end;
13449 SET_recode_x_to_native(0);
13451 SvREFCNT_dec_NN(substitute_parse);
13454 RETURN_FAIL_ON_RESTART(flags, flagp);
13455 FAIL2("panic: reg returned failure to grok_bslash_N, flags=%#" UVxf,
13458 *flagp |= flags&(HASWIDTH|SIMPLE|POSTPONED);
13460 nextchar(pRExC_state);
13467 S_compute_EXACTish(RExC_state_t *pRExC_state)
13471 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
13479 op = get_regex_charset(RExC_flags);
13480 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
13481 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
13482 been, so there is no hole */
13485 return op + EXACTF;
13488 /* Parse backref decimal value, unless it's too big to sensibly be a backref,
13489 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
13492 S_backref_value(char *p, char *e)
13494 const char* endptr = e;
13496 if (grok_atoUV(p, &val, &endptr) && val <= I32_MAX)
13503 - regatom - the lowest level
13505 Try to identify anything special at the start of the current parse position.
13506 If there is, then handle it as required. This may involve generating a
13507 single regop, such as for an assertion; or it may involve recursing, such as
13508 to handle a () structure.
13510 If the string doesn't start with something special then we gobble up
13511 as much literal text as we can. If we encounter a quantifier, we have to
13512 back off the final literal character, as that quantifier applies to just it
13513 and not to the whole string of literals.
13515 Once we have been able to handle whatever type of thing started the
13516 sequence, we return the offset into the regex engine program being compiled
13517 at which any next regnode should be placed.
13519 Returns 0, setting *flagp to TRYAGAIN if reg() returns 0 with TRYAGAIN.
13520 Returns 0, setting *flagp to RESTART_PARSE if the parse needs to be
13521 restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
13522 Otherwise does not return 0.
13524 Note: we have to be careful with escapes, as they can be both literal
13525 and special, and in the case of \10 and friends, context determines which.
13527 A summary of the code structure is:
13529 switch (first_byte) {
13530 cases for each special:
13531 handle this special;
13534 switch (2nd byte) {
13535 cases for each unambiguous special:
13536 handle this special;
13538 cases for each ambigous special/literal:
13540 if (special) handle here
13542 default: // unambiguously literal:
13545 default: // is a literal char
13548 create EXACTish node for literal;
13549 while (more input and node isn't full) {
13550 switch (input_byte) {
13551 cases for each special;
13552 make sure parse pointer is set so that the next call to
13553 regatom will see this special first
13554 goto loopdone; // EXACTish node terminated by prev. char
13556 append char to EXACTISH node;
13558 get next input byte;
13562 return the generated node;
13564 Specifically there are two separate switches for handling
13565 escape sequences, with the one for handling literal escapes requiring
13566 a dummy entry for all of the special escapes that are actually handled
13571 STATIC regnode_offset
13572 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
13574 regnode_offset ret = 0;
13580 DECLARE_AND_GET_RE_DEBUG_FLAGS;
13582 *flagp = 0; /* Initialize. */
13584 DEBUG_PARSE("atom");
13586 PERL_ARGS_ASSERT_REGATOM;
13589 parse_start = RExC_parse;
13590 assert(RExC_parse < RExC_end);
13591 switch ((U8)*RExC_parse) {
13593 RExC_seen_zerolen++;
13594 nextchar(pRExC_state);
13595 if (RExC_flags & RXf_PMf_MULTILINE)
13596 ret = reg_node(pRExC_state, MBOL);
13598 ret = reg_node(pRExC_state, SBOL);
13599 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13602 nextchar(pRExC_state);
13604 RExC_seen_zerolen++;
13605 if (RExC_flags & RXf_PMf_MULTILINE)
13606 ret = reg_node(pRExC_state, MEOL);
13608 ret = reg_node(pRExC_state, SEOL);
13609 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13612 nextchar(pRExC_state);
13613 if (RExC_flags & RXf_PMf_SINGLELINE)
13614 ret = reg_node(pRExC_state, SANY);
13616 ret = reg_node(pRExC_state, REG_ANY);
13617 *flagp |= HASWIDTH|SIMPLE;
13619 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13623 char * const oregcomp_parse = ++RExC_parse;
13624 ret = regclass(pRExC_state, flagp, depth+1,
13625 FALSE, /* means parse the whole char class */
13626 TRUE, /* allow multi-char folds */
13627 FALSE, /* don't silence non-portable warnings. */
13628 (bool) RExC_strict,
13629 TRUE, /* Allow an optimized regnode result */
13632 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13633 FAIL2("panic: regclass returned failure to regatom, flags=%#" UVxf,
13636 if (*RExC_parse != ']') {
13637 RExC_parse = oregcomp_parse;
13638 vFAIL("Unmatched [");
13640 nextchar(pRExC_state);
13641 Set_Node_Length(REGNODE_p(ret), RExC_parse - oregcomp_parse + 1); /* MJD */
13645 nextchar(pRExC_state);
13646 ret = reg(pRExC_state, 2, &flags, depth+1);
13648 if (flags & TRYAGAIN) {
13649 if (RExC_parse >= RExC_end) {
13650 /* Make parent create an empty node if needed. */
13651 *flagp |= TRYAGAIN;
13656 RETURN_FAIL_ON_RESTART(flags, flagp);
13657 FAIL2("panic: reg returned failure to regatom, flags=%#" UVxf,
13660 *flagp |= flags&(HASWIDTH|SIMPLE|POSTPONED);
13664 if (flags & TRYAGAIN) {
13665 *flagp |= TRYAGAIN;
13668 vFAIL("Internal urp");
13669 /* Supposed to be caught earlier. */
13675 vFAIL("Quantifier follows nothing");
13680 This switch handles escape sequences that resolve to some kind
13681 of special regop and not to literal text. Escape sequences that
13682 resolve to literal text are handled below in the switch marked
13685 Every entry in this switch *must* have a corresponding entry
13686 in the literal escape switch. However, the opposite is not
13687 required, as the default for this switch is to jump to the
13688 literal text handling code.
13691 switch ((U8)*RExC_parse) {
13692 /* Special Escapes */
13694 RExC_seen_zerolen++;
13695 /* Under wildcards, this is changed to match \n; should be
13696 * invisible to the user, as they have to compile under /m */
13697 if (RExC_pm_flags & PMf_WILDCARD) {
13698 ret = reg_node(pRExC_state, MBOL);
13701 ret = reg_node(pRExC_state, SBOL);
13702 /* SBOL is shared with /^/ so we set the flags so we can tell
13703 * /\A/ from /^/ in split. */
13704 FLAGS(REGNODE_p(ret)) = 1;
13706 goto finish_meta_pat;
13708 if (RExC_pm_flags & PMf_WILDCARD) {
13710 /* diag_listed_as: Use of %s is not allowed in Unicode property
13711 wildcard subpatterns in regex; marked by <-- HERE in m/%s/
13713 vFAIL("Use of '\\G' is not allowed in Unicode property"
13714 " wildcard subpatterns");
13716 ret = reg_node(pRExC_state, GPOS);
13717 RExC_seen |= REG_GPOS_SEEN;
13718 goto finish_meta_pat;
13720 if (!RExC_in_lookaround) {
13721 RExC_seen_zerolen++;
13722 ret = reg_node(pRExC_state, KEEPS);
13723 /* XXX:dmq : disabling in-place substitution seems to
13724 * be necessary here to avoid cases of memory corruption, as
13725 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
13727 RExC_seen |= REG_LOOKBEHIND_SEEN;
13728 goto finish_meta_pat;
13731 ++RExC_parse; /* advance past the 'K' */
13732 vFAIL("\\K not permitted in lookahead/lookbehind");
13735 if (RExC_pm_flags & PMf_WILDCARD) {
13736 /* See comment under \A above */
13737 ret = reg_node(pRExC_state, MEOL);
13740 ret = reg_node(pRExC_state, SEOL);
13742 RExC_seen_zerolen++; /* Do not optimize RE away */
13743 goto finish_meta_pat;
13745 if (RExC_pm_flags & PMf_WILDCARD) {
13746 /* See comment under \A above */
13747 ret = reg_node(pRExC_state, MEOL);
13750 ret = reg_node(pRExC_state, EOS);
13752 RExC_seen_zerolen++; /* Do not optimize RE away */
13753 goto finish_meta_pat;
13755 vFAIL("\\C no longer supported");
13757 ret = reg_node(pRExC_state, CLUMP);
13758 *flagp |= HASWIDTH;
13759 goto finish_meta_pat;
13767 regex_charset charset = get_regex_charset(RExC_flags);
13769 RExC_seen_zerolen++;
13770 RExC_seen |= REG_LOOKBEHIND_SEEN;
13771 op = BOUND + charset;
13773 if (RExC_parse >= RExC_end || *(RExC_parse + 1) != '{') {
13774 flags = TRADITIONAL_BOUND;
13775 if (op > BOUNDA) { /* /aa is same as /a */
13781 char name = *RExC_parse;
13782 char * endbrace = (char *) memchr(RExC_parse, '}',
13783 RExC_end - RExC_parse);
13784 char * e = endbrace;
13789 vFAIL2("Missing right brace on \\%c{}", name);
13792 while (isBLANK(*RExC_parse)) {
13796 while (RExC_parse < e && isBLANK(*(e - 1))) {
13800 if (e == RExC_parse) {
13801 RExC_parse = endbrace + 1; /* After the '}' */
13802 vFAIL2("Empty \\%c{}", name);
13805 length = e - RExC_parse;
13807 switch (*RExC_parse) {
13810 && (memNEs(RExC_parse + 1, length - 1, "cb")))
13812 goto bad_bound_type;
13817 if (length != 2 || *(RExC_parse + 1) != 'b') {
13818 goto bad_bound_type;
13823 if (length != 2 || *(RExC_parse + 1) != 'b') {
13824 goto bad_bound_type;
13829 if (length != 2 || *(RExC_parse + 1) != 'b') {
13830 goto bad_bound_type;
13838 "'%" UTF8f "' is an unknown bound type",
13839 UTF8fARG(UTF, length, e - length));
13840 NOT_REACHED; /*NOTREACHED*/
13842 RExC_parse = endbrace;
13843 REQUIRE_UNI_RULES(flagp, 0);
13848 else if (op >= BOUNDA) { /* /aa is same as /a */
13852 /* Don't have to worry about UTF-8, in this message because
13853 * to get here the contents of the \b must be ASCII */
13854 ckWARN4reg(RExC_parse + 1, /* Include the '}' in msg */
13855 "Using /u for '%.*s' instead of /%s",
13857 endbrace - length + 1,
13858 (charset == REGEX_ASCII_RESTRICTED_CHARSET)
13859 ? ASCII_RESTRICT_PAT_MODS
13860 : ASCII_MORE_RESTRICT_PAT_MODS);
13865 RExC_seen_d_op = TRUE;
13867 else if (op == BOUNDL) {
13868 RExC_contains_locale = 1;
13872 op += NBOUND - BOUND;
13875 ret = reg_node(pRExC_state, op);
13876 FLAGS(REGNODE_p(ret)) = flags;
13878 goto finish_meta_pat;
13882 ret = reg_node(pRExC_state, LNBREAK);
13883 *flagp |= HASWIDTH|SIMPLE;
13884 goto finish_meta_pat;
13898 /* These all have the same meaning inside [brackets], and it knows
13899 * how to do the best optimizations for them. So, pretend we found
13900 * these within brackets, and let it do the work */
13903 ret = regclass(pRExC_state, flagp, depth+1,
13904 TRUE, /* means just parse this element */
13905 FALSE, /* don't allow multi-char folds */
13906 FALSE, /* don't silence non-portable warnings. It
13907 would be a bug if these returned
13909 (bool) RExC_strict,
13910 TRUE, /* Allow an optimized regnode result */
13912 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13913 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
13914 * multi-char folds are allowed. */
13916 FAIL2("panic: regclass returned failure to regatom, flags=%#" UVxf,
13919 RExC_parse--; /* regclass() leaves this one too far ahead */
13922 /* The escapes above that don't take a parameter can't be
13923 * followed by a '{'. But 'pX', 'p{foo}' and
13924 * correspondingly 'P' can be */
13925 if ( RExC_parse - parse_start == 1
13926 && UCHARAT(RExC_parse + 1) == '{'
13927 && UNLIKELY(! regcurly(RExC_parse + 1, RExC_end, NULL)))
13930 vFAIL("Unescaped left brace in regex is illegal here");
13932 Set_Node_Offset(REGNODE_p(ret), parse_start);
13933 Set_Node_Length(REGNODE_p(ret), RExC_parse - parse_start + 1); /* MJD */
13934 nextchar(pRExC_state);
13937 /* Handle \N, \N{} and \N{NAMED SEQUENCE} (the latter meaning the
13938 * \N{...} evaluates to a sequence of more than one code points).
13939 * The function call below returns a regnode, which is our result.
13940 * The parameters cause it to fail if the \N{} evaluates to a
13941 * single code point; we handle those like any other literal. The
13942 * reason that the multicharacter case is handled here and not as
13943 * part of the EXACtish code is because of quantifiers. In
13944 * /\N{BLAH}+/, the '+' applies to the whole thing, and doing it
13945 * this way makes that Just Happen. dmq.
13946 * join_exact() will join this up with adjacent EXACTish nodes
13947 * later on, if appropriate. */
13949 if (grok_bslash_N(pRExC_state,
13950 &ret, /* Want a regnode returned */
13951 NULL, /* Fail if evaluates to a single code
13953 NULL, /* Don't need a count of how many code
13962 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13964 /* Here, evaluates to a single code point. Go get that */
13965 RExC_parse = parse_start;
13968 case 'k': /* Handle \k<NAME> and \k'NAME' and \k{NAME} */
13969 parse_named_seq: /* Also handle non-numeric \g{...} */
13972 if ( RExC_parse >= RExC_end - 1
13973 || (( ch = RExC_parse[1]) != '<'
13978 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
13979 vFAIL2("Sequence %.2s... not terminated", parse_start);
13983 while (isBLANK(*RExC_parse)) {
13987 ret = handle_named_backref(pRExC_state,
13999 case '1': case '2': case '3': case '4':
14000 case '5': case '6': case '7': case '8': case '9':
14003 char * endbrace = NULL;
14004 char * s = RExC_parse;
14005 char * e = RExC_end;
14012 endbrace = (char *) memchr(s, '}', RExC_end - s);
14015 /* Missing '}'. Position after the number to give
14016 * a better indication to the user of where the
14023 /* If it looks to be a name and not a number, go
14024 * handle it there */
14025 if (! isDIGIT(*s)) {
14026 goto parse_named_seq;
14031 } while isDIGIT(*s);
14034 vFAIL("Unterminated \\g{...} pattern");
14037 s++; /* Past the '{' */
14039 while (isBLANK(*s)) {
14043 /* Ignore trailing blanks */
14045 while (s < e && isBLANK(*(e - 1))) {
14050 /* Here, have isolated the meat of the construct from any
14051 * surrounding braces */
14058 if (endbrace && !isDIGIT(*s)) {
14059 goto parse_named_seq;
14063 num = S_backref_value(RExC_parse, RExC_end);
14065 vFAIL("Reference to invalid group 0");
14066 else if (num == I32_MAX) {
14067 if (isDIGIT(*RExC_parse))
14068 vFAIL("Reference to nonexistent group");
14070 vFAIL("Unterminated \\g... pattern");
14074 num = RExC_npar - num;
14076 vFAIL("Reference to nonexistent or unclosed group");
14080 num = S_backref_value(RExC_parse, RExC_end);
14081 /* bare \NNN might be backref or octal - if it is larger
14082 * than or equal RExC_npar then it is assumed to be an
14083 * octal escape. Note RExC_npar is +1 from the actual
14084 * number of parens. */
14085 /* Note we do NOT check if num == I32_MAX here, as that is
14086 * handled by the RExC_npar check */
14088 if ( /* any numeric escape < 10 is always a backref */
14090 /* any numeric escape < RExC_npar is a backref */
14091 && num >= RExC_npar
14092 /* cannot be an octal escape if it starts with [89]
14094 && ! inRANGE(*RExC_parse, '8', '9')
14096 /* Probably not meant to be a backref, instead likely
14097 * to be an octal character escape, e.g. \35 or \777.
14098 * The above logic should make it obvious why using
14099 * octal escapes in patterns is problematic. - Yves */
14100 RExC_parse = parse_start;
14105 /* At this point RExC_parse points at a numeric escape like
14106 * \12 or \88 or the digits in \g{34} or \g34 or something
14107 * similar, which we should NOT treat as an octal escape. It
14108 * may or may not be a valid backref escape. For instance
14109 * \88888888 is unlikely to be a valid backref.
14111 * We've already figured out what value the digits represent.
14112 * Now, move the parse to beyond them. */
14114 RExC_parse = endbrace + 1;
14116 else while (isDIGIT(*RExC_parse)) {
14120 if (num >= (I32)RExC_npar) {
14122 /* It might be a forward reference; we can't fail until we
14123 * know, by completing the parse to get all the groups, and
14124 * then reparsing */
14125 if (ALL_PARENS_COUNTED) {
14126 if (num >= RExC_total_parens) {
14127 vFAIL("Reference to nonexistent group");
14131 REQUIRE_PARENS_PASS;
14135 ret = reganode(pRExC_state,
14138 : (ASCII_FOLD_RESTRICTED)
14140 : (AT_LEAST_UNI_SEMANTICS)
14146 if (OP(REGNODE_p(ret)) == REFF) {
14147 RExC_seen_d_op = TRUE;
14149 *flagp |= HASWIDTH;
14151 /* override incorrect value set in reganode MJD */
14152 Set_Node_Offset(REGNODE_p(ret), parse_start);
14153 Set_Node_Cur_Length(REGNODE_p(ret), parse_start-1);
14154 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14155 FALSE /* Don't force to /x */ );
14159 if (RExC_parse >= RExC_end)
14160 FAIL("Trailing \\");
14163 /* Do not generate "unrecognized" warnings here, we fall
14164 back into the quick-grab loop below */
14165 RExC_parse = parse_start;
14167 } /* end of switch on a \foo sequence */
14172 /* '#' comments should have been spaced over before this function was
14174 assert((RExC_flags & RXf_PMf_EXTENDED) == 0);
14176 if (RExC_flags & RXf_PMf_EXTENDED) {
14177 RExC_parse = reg_skipcomment( pRExC_state, RExC_parse );
14178 if (RExC_parse < RExC_end)
14188 /* Here, we have determined that the next thing is probably a
14189 * literal character. RExC_parse points to the first byte of its
14190 * definition. (It still may be an escape sequence that evaluates
14191 * to a single character) */
14196 char *s, *old_s = NULL, *old_old_s = NULL;
14198 U32 max_string_len = 255;
14200 /* We may have to reparse the node, artificially stopping filling
14201 * it early, based on info gleaned in the first parse. This
14202 * variable gives where we stop. Make it above the normal stopping
14203 * place first time through; otherwise it would stop too early */
14204 U32 upper_fill = max_string_len + 1;
14206 /* We start out as an EXACT node, even if under /i, until we find a
14207 * character which is in a fold. The algorithm now segregates into
14208 * separate nodes, characters that fold from those that don't under
14209 * /i. (This hopefully will create nodes that are fixed strings
14210 * even under /i, giving the optimizer something to grab on to.)
14211 * So, if a node has something in it and the next character is in
14212 * the opposite category, that node is closed up, and the function
14213 * returns. Then regatom is called again, and a new node is
14214 * created for the new category. */
14215 U8 node_type = EXACT;
14217 /* Assume the node will be fully used; the excess is given back at
14218 * the end. Under /i, we may need to temporarily add the fold of
14219 * an extra character or two at the end to check for splitting
14220 * multi-char folds, so allocate extra space for that. We can't
14221 * make any other length assumptions, as a byte input sequence
14222 * could shrink down. */
14223 Ptrdiff_t current_string_nodes = STR_SZ(max_string_len
14227 ? UTF8_MAXBYTES_CASE
14228 /* Max non-UTF-8 expansion is 2 */ : 2)));
14230 bool next_is_quantifier;
14231 char * oldp = NULL;
14233 /* We can convert EXACTF nodes to EXACTFU if they contain only
14234 * characters that match identically regardless of the target
14235 * string's UTF8ness. The reason to do this is that EXACTF is not
14236 * trie-able, EXACTFU is, and EXACTFU requires fewer operations at
14239 * Similarly, we can convert EXACTFL nodes to EXACTFLU8 if they
14240 * contain only above-Latin1 characters (hence must be in UTF8),
14241 * which don't participate in folds with Latin1-range characters,
14242 * as the latter's folds aren't known until runtime. */
14243 bool maybe_exactfu = FOLD && (DEPENDS_SEMANTICS || LOC);
14245 /* Single-character EXACTish nodes are almost always SIMPLE. This
14246 * allows us to override this as encountered */
14247 U8 maybe_SIMPLE = SIMPLE;
14249 /* Does this node contain something that can't match unless the
14250 * target string is (also) in UTF-8 */
14251 bool requires_utf8_target = FALSE;
14253 /* The sequence 'ss' is problematic in non-UTF-8 patterns. */
14254 bool has_ss = FALSE;
14256 /* So is the MICRO SIGN */
14257 bool has_micro_sign = FALSE;
14259 /* Set when we fill up the current node and there is still more
14260 * text to process */
14263 /* Allocate an EXACT node. The node_type may change below to
14264 * another EXACTish node, but since the size of the node doesn't
14265 * change, it works */
14266 ret = regnode_guts(pRExC_state, node_type, current_string_nodes,
14268 FILL_NODE(ret, node_type);
14271 s = STRING(REGNODE_p(ret));
14282 maybe_exactfu = FOLD && (DEPENDS_SEMANTICS || LOC);
14283 maybe_SIMPLE = SIMPLE;
14284 requires_utf8_target = FALSE;
14286 has_micro_sign = FALSE;
14290 /* This breaks under rare circumstances. If folding, we do not
14291 * want to split a node at a character that is a non-final in a
14292 * multi-char fold, as an input string could just happen to want to
14293 * match across the node boundary. The code at the end of the loop
14294 * looks for this, and backs off until it finds not such a
14295 * character, but it is possible (though extremely, extremely
14296 * unlikely) for all characters in the node to be non-final fold
14297 * ones, in which case we just leave the node fully filled, and
14298 * hope that it doesn't match the string in just the wrong place */
14300 assert( ! UTF /* Is at the beginning of a character */
14301 || UTF8_IS_INVARIANT(UCHARAT(RExC_parse))
14302 || UTF8_IS_START(UCHARAT(RExC_parse)));
14304 overflowed = FALSE;
14306 /* Here, we have a literal character. Find the maximal string of
14307 * them in the input that we can fit into a single EXACTish node.
14308 * We quit at the first non-literal or when the node gets full, or
14309 * under /i the categorization of folding/non-folding character
14311 while (p < RExC_end && len < upper_fill) {
14313 /* In most cases each iteration adds one byte to the output.
14314 * The exceptions override this */
14315 Size_t added_len = 1;
14321 /* White space has already been ignored */
14322 assert( (RExC_flags & RXf_PMf_EXTENDED) == 0
14323 || ! is_PATWS_safe((p), RExC_end, UTF));
14326 const char* message;
14339 /* Literal Escapes Switch
14341 This switch is meant to handle escape sequences that
14342 resolve to a literal character.
14344 Every escape sequence that represents something
14345 else, like an assertion or a char class, is handled
14346 in the switch marked 'Special Escapes' above in this
14347 routine, but also has an entry here as anything that
14348 isn't explicitly mentioned here will be treated as
14349 an unescaped equivalent literal.
14352 switch ((U8)*++p) {
14354 /* These are all the special escapes. */
14355 case 'A': /* Start assertion */
14356 case 'b': case 'B': /* Word-boundary assertion*/
14357 case 'C': /* Single char !DANGEROUS! */
14358 case 'd': case 'D': /* digit class */
14359 case 'g': case 'G': /* generic-backref, pos assertion */
14360 case 'h': case 'H': /* HORIZWS */
14361 case 'k': case 'K': /* named backref, keep marker */
14362 case 'p': case 'P': /* Unicode property */
14363 case 'R': /* LNBREAK */
14364 case 's': case 'S': /* space class */
14365 case 'v': case 'V': /* VERTWS */
14366 case 'w': case 'W': /* word class */
14367 case 'X': /* eXtended Unicode "combining
14368 character sequence" */
14369 case 'z': case 'Z': /* End of line/string assertion */
14373 /* Anything after here is an escape that resolves to a
14374 literal. (Except digits, which may or may not)
14380 case 'N': /* Handle a single-code point named character. */
14381 RExC_parse = p + 1;
14382 if (! grok_bslash_N(pRExC_state,
14383 NULL, /* Fail if evaluates to
14384 anything other than a
14385 single code point */
14386 &ender, /* The returned single code
14388 NULL, /* Don't need a count of
14389 how many code points */
14394 if (*flagp & NEED_UTF8)
14395 FAIL("panic: grok_bslash_N set NEED_UTF8");
14396 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
14398 /* Here, it wasn't a single code point. Go close
14399 * up this EXACTish node. The switch() prior to
14400 * this switch handles the other cases */
14401 RExC_parse = p = oldp;
14405 RExC_parse = parse_start;
14407 /* The \N{} means the pattern, if previously /d,
14408 * becomes /u. That means it can't be an EXACTF node,
14409 * but an EXACTFU */
14410 if (node_type == EXACTF) {
14411 node_type = EXACTFU;
14413 /* If the node already contains something that
14414 * differs between EXACTF and EXACTFU, reparse it
14416 if (! maybe_exactfu) {
14437 ender = ESC_NATIVE;
14445 if (! grok_bslash_o(&p,
14450 (bool) RExC_strict,
14451 FALSE, /* No illegal cp's */
14454 RExC_parse = p; /* going to die anyway; point to
14455 exact spot of failure */
14459 if (message && TO_OUTPUT_WARNINGS(p)) {
14460 warn_non_literal_string(p, packed_warn, message);
14464 if (! grok_bslash_x(&p,
14469 (bool) RExC_strict,
14470 FALSE, /* No illegal cp's */
14473 RExC_parse = p; /* going to die anyway; point
14474 to exact spot of failure */
14478 if (message && TO_OUTPUT_WARNINGS(p)) {
14479 warn_non_literal_string(p, packed_warn, message);
14483 if (ender < 0x100) {
14484 if (RExC_recode_x_to_native) {
14485 ender = LATIN1_TO_NATIVE(ender);
14492 if (! grok_bslash_c(*p, &grok_c_char,
14493 &message, &packed_warn))
14495 /* going to die anyway; point to exact spot of
14497 RExC_parse = p + ((UTF)
14498 ? UTF8_SAFE_SKIP(p, RExC_end)
14503 ender = grok_c_char;
14505 if (message && TO_OUTPUT_WARNINGS(p)) {
14506 warn_non_literal_string(p, packed_warn, message);
14510 case '8': case '9': /* must be a backreference */
14512 /* we have an escape like \8 which cannot be an octal escape
14513 * so we exit the loop, and let the outer loop handle this
14514 * escape which may or may not be a legitimate backref. */
14516 case '1': case '2': case '3':case '4':
14517 case '5': case '6': case '7':
14519 /* When we parse backslash escapes there is ambiguity
14520 * between backreferences and octal escapes. Any escape
14521 * from \1 - \9 is a backreference, any multi-digit
14522 * escape which does not start with 0 and which when
14523 * evaluated as decimal could refer to an already
14524 * parsed capture buffer is a back reference. Anything
14527 * Note this implies that \118 could be interpreted as
14528 * 118 OR as "\11" . "8" depending on whether there
14529 * were 118 capture buffers defined already in the
14532 /* NOTE, RExC_npar is 1 more than the actual number of
14533 * parens we have seen so far, hence the "<" as opposed
14535 if ( !isDIGIT(p[1]) || S_backref_value(p, RExC_end) < RExC_npar)
14536 { /* Not to be treated as an octal constant, go
14544 I32 flags = PERL_SCAN_SILENT_ILLDIGIT
14545 | PERL_SCAN_NOTIFY_ILLDIGIT;
14547 ender = grok_oct(p, &numlen, &flags, NULL);
14549 if ( (flags & PERL_SCAN_NOTIFY_ILLDIGIT)
14550 && isDIGIT(*p) /* like \08, \178 */
14551 && ckWARN(WARN_REGEXP))
14553 reg_warn_non_literal_string(
14555 form_alien_digit_msg(8, numlen, p,
14556 RExC_end, UTF, FALSE));
14562 FAIL("Trailing \\");
14565 if (isALPHANUMERIC(*p)) {
14566 /* An alpha followed by '{' is going to fail next
14567 * iteration, so don't output this warning in that
14569 if (! isALPHA(*p) || *(p + 1) != '{') {
14570 ckWARN2reg(p + 1, "Unrecognized escape \\%.1s"
14571 " passed through", p);
14574 goto normal_default;
14575 } /* End of switch on '\' */
14578 /* Trying to gain new uses for '{' without breaking too
14579 * much existing code is hard. The solution currently
14581 * 1) If there is no ambiguity that a '{' should always
14582 * be taken literally, at the start of a construct, we
14584 * 2) If the literal '{' conflicts with our desired use
14585 * of it as a metacharacter, we die. The deprecation
14586 * cycles for this have come and gone.
14587 * 3) If there is ambiguity, we raise a simple warning.
14588 * This could happen, for example, if the user
14589 * intended it to introduce a quantifier, but slightly
14590 * misspelled the quantifier. Without this warning,
14591 * the quantifier would silently be taken as a literal
14592 * string of characters instead of a meta construct */
14593 if (len || (p > RExC_start && isALPHA_A(*(p - 1)))) {
14595 || ( p > parse_start + 1
14596 && isALPHA_A(*(p - 1))
14597 && *(p - 2) == '\\'))
14599 RExC_parse = p + 1;
14600 vFAIL("Unescaped left brace in regex is "
14603 ckWARNreg(p + 1, "Unescaped left brace in regex is"
14604 " passed through");
14606 goto normal_default;
14609 if (p > RExC_parse && RExC_strict) {
14610 ckWARN2reg(p + 1, "Unescaped literal '%c'", *p);
14613 default: /* A literal character */
14615 if (! UTF8_IS_INVARIANT(*p) && UTF) {
14617 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
14618 &numlen, UTF8_ALLOW_DEFAULT);
14624 } /* End of switch on the literal */
14626 /* Here, have looked at the literal character, and <ender>
14627 * contains its ordinal; <p> points to the character after it.
14631 REQUIRE_UTF8(flagp);
14632 if ( UNICODE_IS_PERL_EXTENDED(ender)
14633 && TO_OUTPUT_WARNINGS(p))
14635 ckWARN2_non_literal_string(p,
14636 packWARN(WARN_PORTABLE),
14637 PL_extended_cp_format,
14642 /* We need to check if the next non-ignored thing is a
14643 * quantifier. Move <p> to after anything that should be
14644 * ignored, which, as a side effect, positions <p> for the next
14645 * loop iteration */
14646 skip_to_be_ignored_text(pRExC_state, &p,
14647 FALSE /* Don't force to /x */ );
14649 /* If the next thing is a quantifier, it applies to this
14650 * character only, which means that this character has to be in
14651 * its own node and can't just be appended to the string in an
14652 * existing node, so if there are already other characters in
14653 * the node, close the node with just them, and set up to do
14654 * this character again next time through, when it will be the
14655 * only thing in its new node */
14657 next_is_quantifier = LIKELY(p < RExC_end)
14658 && UNLIKELY(isQUANTIFIER(p, RExC_end));
14660 if (next_is_quantifier && LIKELY(len)) {
14665 /* Ready to add 'ender' to the node */
14667 if (! FOLD) { /* The simple case, just append the literal */
14670 /* Don't output if it would overflow */
14671 if (UNLIKELY(len > max_string_len - ((UTF)
14672 ? UVCHR_SKIP(ender)
14679 if (UVCHR_IS_INVARIANT(ender) || ! UTF) {
14680 *(s++) = (char) ender;
14683 U8 * new_s = uvchr_to_utf8((U8*)s, ender);
14684 added_len = (char *) new_s - s;
14685 s = (char *) new_s;
14688 requires_utf8_target = TRUE;
14692 else if (LOC && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)) {
14694 /* Here are folding under /l, and the code point is
14695 * problematic. If this is the first character in the
14696 * node, change the node type to folding. Otherwise, if
14697 * this is the first problematic character, close up the
14698 * existing node, so can start a new node with this one */
14700 node_type = EXACTFL;
14701 RExC_contains_locale = 1;
14703 else if (node_type == EXACT) {
14708 /* This problematic code point means we can't simplify
14710 maybe_exactfu = FALSE;
14712 /* Although these two characters have folds that are
14713 * locale-problematic, they also have folds to above Latin1
14714 * that aren't a problem. Doing these now helps at
14716 if (UNLIKELY( ender == GREEK_CAPITAL_LETTER_MU
14717 || ender == LATIN_CAPITAL_LETTER_SHARP_S))
14722 /* Here, we are adding a problematic fold character.
14723 * "Problematic" in this context means that its fold isn't
14724 * known until runtime. (The non-problematic code points
14725 * are the above-Latin1 ones that fold to also all
14726 * above-Latin1. Their folds don't vary no matter what the
14727 * locale is.) But here we have characters whose fold
14728 * depends on the locale. We just add in the unfolded
14729 * character, and wait until runtime to fold it */
14730 goto not_fold_common;
14732 else /* regular fold; see if actually is in a fold */
14733 if ( (ender < 256 && ! IS_IN_SOME_FOLD_L1(ender))
14735 && ! _invlist_contains_cp(PL_in_some_fold, ender)))
14737 /* Here, folding, but the character isn't in a fold.
14739 * Start a new node if previous characters in the node were
14741 if (len && node_type != EXACT) {
14746 /* Here, continuing a node with non-folded characters. Add
14748 goto not_fold_common;
14750 else { /* Here, does participate in some fold */
14752 /* If this is the first character in the node, change its
14753 * type to folding. Otherwise, if this is the first
14754 * folding character in the node, close up the existing
14755 * node, so can start a new node with this one. */
14757 node_type = compute_EXACTish(pRExC_state);
14759 else if (node_type == EXACT) {
14764 if (UTF) { /* Alway use the folded value for UTF-8
14766 if (UVCHR_IS_INVARIANT(ender)) {
14767 if (UNLIKELY(len + 1 > max_string_len)) {
14772 *(s)++ = (U8) toFOLD(ender);
14778 folded = _to_uni_fold_flags(
14780 (U8 *) s, /* We have allocated extra space
14781 in 's' so can't run off the
14785 | (( ASCII_FOLD_RESTRICTED
14786 || node_type == EXACTFL)
14787 ? FOLD_FLAGS_NOMIX_ASCII
14789 if (UNLIKELY(len + added_len > max_string_len)) {
14797 && LIKELY(folded != GREEK_SMALL_LETTER_MU))
14799 /* U+B5 folds to the MU, so its possible for a
14800 * non-UTF-8 target to match it */
14801 requires_utf8_target = TRUE;
14805 else { /* Here is non-UTF8. */
14807 /* The fold will be one or (rarely) two characters.
14808 * Check that there's room for at least a single one
14809 * before setting any flags, etc. Because otherwise an
14810 * overflowing character could cause a flag to be set
14811 * even though it doesn't end up in this node. (For
14812 * the two character fold, we check again, before
14813 * setting any flags) */
14814 if (UNLIKELY(len + 1 > max_string_len)) {
14819 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
14820 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
14821 || UNICODE_DOT_DOT_VERSION > 0)
14823 /* On non-ancient Unicodes, check for the only possible
14824 * multi-char fold */
14825 if (UNLIKELY(ender == LATIN_SMALL_LETTER_SHARP_S)) {
14827 /* This potential multi-char fold means the node
14828 * can't be simple (because it could match more
14829 * than a single char). And in some cases it will
14830 * match 'ss', so set that flag */
14834 /* It can't change to be an EXACTFU (unless already
14835 * is one). We fold it iff under /u rules. */
14836 if (node_type != EXACTFU) {
14837 maybe_exactfu = FALSE;
14840 if (UNLIKELY(len + 2 > max_string_len)) {
14849 goto done_with_this_char;
14852 else if ( UNLIKELY(isALPHA_FOLD_EQ(ender, 's'))
14854 && UNLIKELY(isALPHA_FOLD_EQ(*(s-1), 's')))
14856 /* Also, the sequence 'ss' is special when not
14857 * under /u. If the target string is UTF-8, it
14858 * should match SHARP S; otherwise it won't. So,
14859 * here we have to exclude the possibility of this
14860 * node moving to /u.*/
14862 maybe_exactfu = FALSE;
14865 /* Here, the fold will be a single character */
14867 if (UNLIKELY(ender == MICRO_SIGN)) {
14868 has_micro_sign = TRUE;
14870 else if (PL_fold[ender] != PL_fold_latin1[ender]) {
14872 /* If the character's fold differs between /d and
14873 * /u, this can't change to be an EXACTFU node */
14874 maybe_exactfu = FALSE;
14877 *(s++) = (DEPENDS_SEMANTICS)
14878 ? (char) toFOLD(ender)
14880 /* Under /u, the fold of any character in
14881 * the 0-255 range happens to be its
14882 * lowercase equivalent, except for LATIN
14883 * SMALL LETTER SHARP S, which was handled
14884 * above, and the MICRO SIGN, whose fold
14885 * requires UTF-8 to represent. */
14886 : (char) toLOWER_L1(ender);
14888 } /* End of adding current character to the node */
14890 done_with_this_char:
14894 if (next_is_quantifier) {
14896 /* Here, the next input is a quantifier, and to get here,
14897 * the current character is the only one in the node. */
14901 } /* End of loop through literal characters */
14903 /* Here we have either exhausted the input or run out of room in
14904 * the node. If the former, we are done. (If we encountered a
14905 * character that can't be in the node, transfer is made directly
14906 * to <loopdone>, and so we wouldn't have fallen off the end of the
14908 if (LIKELY(! overflowed)) {
14912 /* Here we have run out of room. We can grow plain EXACT and
14913 * LEXACT nodes. If the pattern is gigantic enough, though,
14914 * eventually we'll have to artificially chunk the pattern into
14915 * multiple nodes. */
14916 if (! LOC && (node_type == EXACT || node_type == LEXACT)) {
14917 Size_t overhead = 1 + regarglen[OP(REGNODE_p(ret))];
14918 Size_t overhead_expansion = 0;
14920 Size_t max_nodes_for_string;
14924 /* Here we couldn't fit the final character in the current
14925 * node, so it will have to be reparsed, no matter what else we
14929 /* If would have overflowed a regular EXACT node, switch
14930 * instead to an LEXACT. The code below is structured so that
14931 * the actual growing code is common to changing from an EXACT
14932 * or just increasing the LEXACT size. This means that we have
14933 * to save the string in the EXACT case before growing, and
14934 * then copy it afterwards to its new location */
14935 if (node_type == EXACT) {
14936 overhead_expansion = regarglen[LEXACT] - regarglen[EXACT];
14937 RExC_emit += overhead_expansion;
14938 Copy(s0, temp, len, char);
14941 /* Ready to grow. If it was a plain EXACT, the string was
14942 * saved, and the first few bytes of it overwritten by adding
14943 * an argument field. We assume, as we do elsewhere in this
14944 * file, that one byte of remaining input will translate into
14945 * one byte of output, and if that's too small, we grow again,
14946 * if too large the excess memory is freed at the end */
14948 max_nodes_for_string = U16_MAX - overhead - overhead_expansion;
14949 achievable = MIN(max_nodes_for_string,
14950 current_string_nodes + STR_SZ(RExC_end - p));
14951 delta = achievable - current_string_nodes;
14953 /* If there is just no more room, go finish up this chunk of
14959 change_engine_size(pRExC_state, delta + overhead_expansion);
14960 current_string_nodes += delta;
14962 = sizeof(struct regnode) * current_string_nodes;
14963 upper_fill = max_string_len + 1;
14965 /* If the length was small, we know this was originally an
14966 * EXACT node now converted to LEXACT, and the string has to be
14967 * restored. Otherwise the string was untouched. 260 is just
14968 * a number safely above 255 so don't have to worry about
14969 * getting it precise */
14971 node_type = LEXACT;
14972 FILL_NODE(ret, node_type);
14973 s0 = STRING(REGNODE_p(ret));
14974 Copy(temp, s0, len, char);
14978 goto continue_parse;
14981 bool splittable = FALSE;
14982 bool backed_up = FALSE;
14983 char * e; /* should this be U8? */
14984 char * s_start; /* should this be U8? */
14986 /* Here is /i. Running out of room creates a problem if we are
14987 * folding, and the split happens in the middle of a
14988 * multi-character fold, as a match that should have occurred,
14989 * won't, due to the way nodes are matched, and our artificial
14990 * boundary. So back off until we aren't splitting such a
14991 * fold. If there is no such place to back off to, we end up
14992 * taking the entire node as-is. This can happen if the node
14993 * consists entirely of 'f' or entirely of 's' characters (or
14994 * things that fold to them) as 'ff' and 'ss' are
14995 * multi-character folds.
14997 * The Unicode standard says that multi character folds consist
14998 * of either two or three characters. That means we would be
14999 * splitting one if the final character in the node is at the
15000 * beginning of either type, or is the second of a three
15004 * ender is the code point of the character that won't fit
15006 * s points to just beyond the final byte in the node.
15007 * It's where we would place ender if there were
15008 * room, and where in fact we do place ender's fold
15009 * in the code below, as we've over-allocated space
15010 * for s0 (hence s) to allow for this
15011 * e starts at 's' and advances as we append things.
15012 * old_s is the same as 's'. (If ender had fit, 's' would
15013 * have been advanced to beyond it).
15014 * old_old_s points to the beginning byte of the final
15015 * character in the node
15016 * p points to the beginning byte in the input of the
15017 * character beyond 'ender'.
15018 * oldp points to the beginning byte in the input of
15021 * In the case of /il, we haven't folded anything that could be
15022 * affected by the locale. That means only above-Latin1
15023 * characters that fold to other above-latin1 characters get
15024 * folded at compile time. To check where a good place to
15025 * split nodes is, everything in it will have to be folded.
15026 * The boolean 'maybe_exactfu' keeps track in /il if there are
15027 * any unfolded characters in the node. */
15028 bool need_to_fold_loc = LOC && ! maybe_exactfu;
15030 /* If we do need to fold the node, we need a place to store the
15031 * folded copy, and a way to map back to the unfolded original
15033 char * locfold_buf = NULL;
15034 Size_t * loc_correspondence = NULL;
15036 if (! need_to_fold_loc) { /* The normal case. Just
15037 initialize to the actual node */
15040 s = old_old_s; /* Point to the beginning of the final char
15041 that fits in the node */
15045 /* Here, we have filled a /il node, and there are unfolded
15046 * characters in it. If the runtime locale turns out to be
15047 * UTF-8, there are possible multi-character folds, just
15048 * like when not under /l. The node hence can't terminate
15049 * in the middle of such a fold. To determine this, we
15050 * have to create a folded copy of this node. That means
15051 * reparsing the node, folding everything assuming a UTF-8
15052 * locale. (If at runtime it isn't such a locale, the
15053 * actions here wouldn't have been necessary, but we have
15054 * to assume the worst case.) If we find we need to back
15055 * off the folded string, we do so, and then map that
15056 * position back to the original unfolded node, which then
15057 * gets output, truncated at that spot */
15059 char * redo_p = RExC_parse;
15063 /* Allow enough space assuming a single byte input folds to
15064 * a single byte output, plus assume that the two unparsed
15065 * characters (that we may need) fold to the largest number
15066 * of bytes possible, plus extra for one more worst case
15067 * scenario. In the loop below, if we start eating into
15068 * that final spare space, we enlarge this initial space */
15069 Size_t size = max_string_len + (3 * UTF8_MAXBYTES_CASE) + 1;
15071 Newxz(locfold_buf, size, char);
15072 Newxz(loc_correspondence, size, Size_t);
15074 /* Redo this node's parse, folding into 'locfold_buf' */
15075 redo_p = RExC_parse;
15076 old_redo_e = redo_e = locfold_buf;
15077 while (redo_p <= oldp) {
15079 old_redo_e = redo_e;
15080 loc_correspondence[redo_e - locfold_buf]
15081 = redo_p - RExC_parse;
15086 (void) _to_utf8_fold_flags((U8 *) redo_p,
15091 redo_e += added_len;
15092 redo_p += UTF8SKIP(redo_p);
15096 /* Note that if this code is run on some ancient
15097 * Unicode versions, SHARP S doesn't fold to 'ss',
15098 * but rather than clutter the code with #ifdef's,
15099 * as is done above, we ignore that possibility.
15100 * This is ok because this code doesn't affect what
15101 * gets matched, but merely where the node gets
15103 if (UCHARAT(redo_p) != LATIN_SMALL_LETTER_SHARP_S) {
15104 *redo_e++ = toLOWER_L1(UCHARAT(redo_p));
15114 /* If we're getting so close to the end that a
15115 * worst-case fold in the next character would cause us
15116 * to overflow, increase, assuming one byte output byte
15117 * per one byte input one, plus room for another worst
15119 if ( redo_p <= oldp
15120 && redo_e > locfold_buf + size
15121 - (UTF8_MAXBYTES_CASE + 1))
15123 Size_t new_size = size
15125 + UTF8_MAXBYTES_CASE + 1;
15126 Ptrdiff_t e_offset = redo_e - locfold_buf;
15128 Renew(locfold_buf, new_size, char);
15129 Renew(loc_correspondence, new_size, Size_t);
15132 redo_e = locfold_buf + e_offset;
15136 /* Set so that things are in terms of the folded, temporary
15139 s_start = locfold_buf;
15144 /* Here, we have 's', 's_start' and 'e' set up to point to the
15145 * input that goes into the node, folded.
15147 * If the final character of the node and the fold of ender
15148 * form the first two characters of a three character fold, we
15149 * need to peek ahead at the next (unparsed) character in the
15150 * input to determine if the three actually do form such a
15151 * fold. Just looking at that character is not generally
15152 * sufficient, as it could be, for example, an escape sequence
15153 * that evaluates to something else, and it needs to be folded.
15155 * khw originally thought to just go through the parse loop one
15156 * extra time, but that doesn't work easily as that iteration
15157 * could cause things to think that the parse is over and to
15158 * goto loopdone. The character could be a '$' for example, or
15159 * the character beyond could be a quantifier, and other
15160 * glitches as well.
15162 * The solution used here for peeking ahead is to look at that
15163 * next character. If it isn't ASCII punctuation, then it will
15164 * be something that would continue on in an EXACTish node if
15165 * there were space. We append the fold of it to s, having
15166 * reserved enough room in s0 for the purpose. If we can't
15167 * reasonably peek ahead, we instead assume the worst case:
15168 * that it is something that would form the completion of a
15171 * If we can't split between s and ender, we work backwards
15172 * character-by-character down to s0. At each current point
15173 * see if we are at the beginning of a multi-char fold. If so,
15174 * that means we would be splitting the fold across nodes, and
15175 * so we back up one and try again.
15177 * If we're not at the beginning, we still could be at the
15178 * final two characters of a (rare) three character fold. We
15179 * check if the sequence starting at the character before the
15180 * current position (and including the current and next
15181 * characters) is a three character fold. If not, the node can
15182 * be split here. If it is, we have to backup two characters
15185 * Otherwise, the node can be split at the current position.
15187 * The same logic is used for UTF-8 patterns and not */
15191 /* Append the fold of ender */
15192 (void) _to_uni_fold_flags(
15196 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
15197 ? FOLD_FLAGS_NOMIX_ASCII
15201 /* 's' and the character folded to by ender may be the
15202 * first two of a three-character fold, in which case the
15203 * node should not be split here. That may mean examining
15204 * the so-far unparsed character starting at 'p'. But if
15205 * ender folded to more than one character, we already have
15206 * three characters to look at. Also, we first check if
15207 * the sequence consisting of s and the next character form
15208 * the first two of some three character fold. If not,
15209 * there's no need to peek ahead. */
15210 if ( added_len <= UTF8SKIP(e - added_len)
15211 && UNLIKELY(is_THREE_CHAR_FOLD_HEAD_utf8_safe(s, e)))
15213 /* Here, the two do form the beginning of a potential
15214 * three character fold. The unexamined character may
15215 * or may not complete it. Peek at it. It might be
15216 * something that ends the node or an escape sequence,
15217 * in which case we don't know without a lot of work
15218 * what it evaluates to, so we have to assume the worst
15219 * case: that it does complete the fold, and so we
15220 * can't split here. All such instances will have
15221 * that character be an ASCII punctuation character,
15222 * like a backslash. So, for that case, backup one and
15223 * drop down to try at that position */
15225 s = (char *) utf8_hop_back((U8 *) s, -1,
15230 /* Here, since it's not punctuation, it must be a
15231 * real character, and we can append its fold to
15232 * 'e' (having deliberately reserved enough space
15233 * for this eventuality) and drop down to check if
15234 * the three actually do form a folded sequence */
15235 (void) _to_utf8_fold_flags(
15236 (U8 *) p, (U8 *) RExC_end,
15239 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
15240 ? FOLD_FLAGS_NOMIX_ASCII
15246 /* Here, we either have three characters available in
15247 * sequence starting at 's', or we have two characters and
15248 * know that the following one can't possibly be part of a
15249 * three character fold. We go through the node backwards
15250 * until we find a place where we can split it without
15251 * breaking apart a multi-character fold. At any given
15252 * point we have to worry about if such a fold begins at
15253 * the current 's', and also if a three-character fold
15254 * begins at s-1, (containing s and s+1). Splitting in
15255 * either case would break apart a fold */
15257 char *prev_s = (char *) utf8_hop_back((U8 *) s, -1,
15260 /* If is a multi-char fold, can't split here. Backup
15261 * one char and try again */
15262 if (UNLIKELY(is_MULTI_CHAR_FOLD_utf8_safe(s, e))) {
15268 /* If the two characters beginning at 's' are part of a
15269 * three character fold starting at the character
15270 * before s, we can't split either before or after s.
15271 * Backup two chars and try again */
15272 if ( LIKELY(s > s_start)
15273 && UNLIKELY(is_THREE_CHAR_FOLD_utf8_safe(prev_s, e)))
15276 s = (char *) utf8_hop_back((U8 *) s, -1, (U8 *) s_start);
15281 /* Here there's no multi-char fold between s and the
15282 * next character following it. We can split */
15286 } while (s > s_start); /* End of loops backing up through the node */
15288 /* Here we either couldn't find a place to split the node,
15289 * or else we broke out of the loop setting 'splittable' to
15290 * true. In the latter case, the place to split is between
15291 * the first and second characters in the sequence starting
15297 else { /* Pattern not UTF-8 */
15298 if ( ender != LATIN_SMALL_LETTER_SHARP_S
15299 || ASCII_FOLD_RESTRICTED)
15301 assert( toLOWER_L1(ender) < 256 );
15302 *e++ = (char)(toLOWER_L1(ender)); /* should e and the cast be U8? */
15310 && UNLIKELY(is_THREE_CHAR_FOLD_HEAD_latin1_safe(s, e)))
15317 if ( UCHARAT(p) != LATIN_SMALL_LETTER_SHARP_S
15318 || ASCII_FOLD_RESTRICTED)
15320 assert( toLOWER_L1(ender) < 256 );
15321 *e++ = (char)(toLOWER_L1(ender)); /* should e and the cast be U8? */
15331 if (UNLIKELY(is_MULTI_CHAR_FOLD_latin1_safe(s, e))) {
15337 if ( LIKELY(s > s_start)
15338 && UNLIKELY(is_THREE_CHAR_FOLD_latin1_safe(s - 1, e)))
15348 } while (s > s_start);
15355 /* Here, we are done backing up. If we didn't backup at all
15356 * (the likely case), just proceed */
15359 /* If we did find a place to split, reparse the entire node
15360 * stopping where we have calculated. */
15363 /* If we created a temporary folded string under /l, we
15364 * have to map that back to the original */
15365 if (need_to_fold_loc) {
15366 upper_fill = loc_correspondence[s - s_start];
15367 if (upper_fill == 0) {
15368 FAIL2("panic: loc_correspondence[%d] is 0",
15369 (int) (s - s_start));
15371 Safefree(locfold_buf);
15372 Safefree(loc_correspondence);
15375 upper_fill = s - s0;
15380 /* Here the node consists entirely of non-final multi-char
15381 * folds. (Likely it is all 'f's or all 's's.) There's no
15382 * decent place to split it, so give up and just take the
15387 if (need_to_fold_loc) {
15388 Safefree(locfold_buf);
15389 Safefree(loc_correspondence);
15391 } /* End of verifying node ends with an appropriate char */
15393 /* We need to start the next node at the character that didn't fit
15397 loopdone: /* Jumped to when encounters something that shouldn't be
15400 /* Free up any over-allocated space; cast is to silence bogus
15401 * warning in MS VC */
15402 change_engine_size(pRExC_state,
15403 - (Ptrdiff_t) (current_string_nodes - STR_SZ(len)));
15405 /* I (khw) don't know if you can get here with zero length, but the
15406 * old code handled this situation by creating a zero-length EXACT
15407 * node. Might as well be NOTHING instead */
15409 OP(REGNODE_p(ret)) = NOTHING;
15413 /* If the node type is EXACT here, check to see if it
15414 * should be EXACTL, or EXACT_REQ8. */
15415 if (node_type == EXACT) {
15417 node_type = EXACTL;
15419 else if (requires_utf8_target) {
15420 node_type = EXACT_REQ8;
15423 else if (node_type == LEXACT) {
15424 if (requires_utf8_target) {
15425 node_type = LEXACT_REQ8;
15429 if ( UNLIKELY(has_micro_sign || has_ss)
15430 && (node_type == EXACTFU || ( node_type == EXACTF
15431 && maybe_exactfu)))
15432 { /* These two conditions are problematic in non-UTF-8
15435 node_type = EXACTFUP;
15437 else if (node_type == EXACTFL) {
15439 /* 'maybe_exactfu' is deliberately set above to
15440 * indicate this node type, where all code points in it
15442 if (maybe_exactfu) {
15443 node_type = EXACTFLU8;
15446 _invlist_contains_cp(PL_HasMultiCharFold, ender)))
15448 /* A character that folds to more than one will
15449 * match multiple characters, so can't be SIMPLE.
15450 * We don't have to worry about this with EXACTFLU8
15451 * nodes just above, as they have already been
15452 * folded (since the fold doesn't vary at run
15453 * time). Here, if the final character in the node
15454 * folds to multiple, it can't be simple. (This
15455 * only has an effect if the node has only a single
15456 * character, hence the final one, as elsewhere we
15457 * turn off simple for nodes whose length > 1 */
15461 else if (node_type == EXACTF) { /* Means is /di */
15463 /* This intermediate variable is needed solely because
15464 * the asserts in the macro where used exceed Win32's
15465 * literal string capacity */
15466 char first_char = * STRING(REGNODE_p(ret));
15468 /* If 'maybe_exactfu' is clear, then we need to stay
15469 * /di. If it is set, it means there are no code
15470 * points that match differently depending on UTF8ness
15471 * of the target string, so it can become an EXACTFU
15473 if (! maybe_exactfu) {
15474 RExC_seen_d_op = TRUE;
15476 else if ( isALPHA_FOLD_EQ(first_char, 's')
15477 || isALPHA_FOLD_EQ(ender, 's'))
15479 /* But, if the node begins or ends in an 's' we
15480 * have to defer changing it into an EXACTFU, as
15481 * the node could later get joined with another one
15482 * that ends or begins with 's' creating an 'ss'
15483 * sequence which would then wrongly match the
15484 * sharp s without the target being UTF-8. We
15485 * create a special node that we resolve later when
15486 * we join nodes together */
15488 node_type = EXACTFU_S_EDGE;
15491 node_type = EXACTFU;
15495 if (requires_utf8_target && node_type == EXACTFU) {
15496 node_type = EXACTFU_REQ8;
15500 OP(REGNODE_p(ret)) = node_type;
15501 setSTR_LEN(REGNODE_p(ret), len);
15502 RExC_emit += STR_SZ(len);
15504 /* If the node isn't a single character, it can't be SIMPLE */
15505 if (len > (Size_t) ((UTF) ? UTF8SKIP(STRING(REGNODE_p(ret))) : 1)) {
15509 *flagp |= HASWIDTH | maybe_SIMPLE;
15512 Set_Node_Length(REGNODE_p(ret), p - parse_start - 1);
15516 /* len is STRLEN which is unsigned, need to copy to signed */
15519 vFAIL("Internal disaster");
15522 } /* End of label 'defchar:' */
15524 } /* End of giant switch on input character */
15526 /* Position parse to next real character */
15527 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
15528 FALSE /* Don't force to /x */ );
15529 if ( *RExC_parse == '{'
15530 && OP(REGNODE_p(ret)) != SBOL && ! regcurly(RExC_parse, RExC_end, NULL))
15534 vFAIL("Unescaped left brace in regex is illegal here");
15536 ckWARNreg(RExC_parse + 1, "Unescaped left brace in regex is"
15537 " passed through");
15545 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
15547 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
15548 * sets up the bitmap and any flags, removing those code points from the
15549 * inversion list, setting it to NULL should it become completely empty */
15552 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
15553 assert(PL_regkind[OP(node)] == ANYOF);
15555 /* There is no bitmap for this node type */
15556 if (inRANGE(OP(node), ANYOFH, ANYOFRb)) {
15560 ANYOF_BITMAP_ZERO(node);
15561 if (*invlist_ptr) {
15563 /* This gets set if we actually need to modify things */
15564 bool change_invlist = FALSE;
15568 /* Start looking through *invlist_ptr */
15569 invlist_iterinit(*invlist_ptr);
15570 while (invlist_iternext(*invlist_ptr, &start, &end)) {
15574 if (end == UV_MAX && start <= NUM_ANYOF_CODE_POINTS) {
15575 ANYOF_FLAGS(node) |= ANYOF_MATCHES_ALL_ABOVE_BITMAP;
15578 /* Quit if are above what we should change */
15579 if (start >= NUM_ANYOF_CODE_POINTS) {
15583 change_invlist = TRUE;
15585 /* Set all the bits in the range, up to the max that we are doing */
15586 high = (end < NUM_ANYOF_CODE_POINTS - 1)
15588 : NUM_ANYOF_CODE_POINTS - 1;
15589 for (i = start; i <= (int) high; i++) {
15590 ANYOF_BITMAP_SET(node, i);
15593 invlist_iterfinish(*invlist_ptr);
15595 /* Done with loop; remove any code points that are in the bitmap from
15596 * *invlist_ptr; similarly for code points above the bitmap if we have
15597 * a flag to match all of them anyways */
15598 if (change_invlist) {
15599 _invlist_subtract(*invlist_ptr, PL_InBitmap, invlist_ptr);
15601 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
15602 _invlist_intersection(*invlist_ptr, PL_InBitmap, invlist_ptr);
15605 /* If have completely emptied it, remove it completely */
15606 if (_invlist_len(*invlist_ptr) == 0) {
15607 SvREFCNT_dec_NN(*invlist_ptr);
15608 *invlist_ptr = NULL;
15613 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
15614 Character classes ([:foo:]) can also be negated ([:^foo:]).
15615 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
15616 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
15617 but trigger failures because they are currently unimplemented. */
15619 #define POSIXCC_DONE(c) ((c) == ':')
15620 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
15621 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
15622 #define MAYBE_POSIXCC(c) (POSIXCC(c) || (c) == '^' || (c) == ';')
15624 #define WARNING_PREFIX "Assuming NOT a POSIX class since "
15625 #define NO_BLANKS_POSIX_WARNING "no blanks are allowed in one"
15626 #define SEMI_COLON_POSIX_WARNING "a semi-colon was found instead of a colon"
15628 #define NOT_MEANT_TO_BE_A_POSIX_CLASS (OOB_NAMEDCLASS - 1)
15630 /* 'posix_warnings' and 'warn_text' are names of variables in the following
15632 #define ADD_POSIX_WARNING(p, text) STMT_START { \
15633 if (posix_warnings) { \
15634 if (! RExC_warn_text ) RExC_warn_text = \
15635 (AV *) sv_2mortal((SV *) newAV()); \
15636 av_push(RExC_warn_text, Perl_newSVpvf(aTHX_ \
15640 REPORT_LOCATION_ARGS(p))); \
15643 #define CLEAR_POSIX_WARNINGS() \
15645 if (posix_warnings && RExC_warn_text) \
15646 av_clear(RExC_warn_text); \
15649 #define CLEAR_POSIX_WARNINGS_AND_RETURN(ret) \
15651 CLEAR_POSIX_WARNINGS(); \
15656 S_handle_possible_posix(pTHX_ RExC_state_t *pRExC_state,
15658 const char * const s, /* Where the putative posix class begins.
15659 Normally, this is one past the '['. This
15660 parameter exists so it can be somewhere
15661 besides RExC_parse. */
15662 char ** updated_parse_ptr, /* Where to set the updated parse pointer, or
15664 AV ** posix_warnings, /* Where to place any generated warnings, or
15666 const bool check_only /* Don't die if error */
15669 /* This parses what the caller thinks may be one of the three POSIX
15671 * 1) a character class, like [:blank:]
15672 * 2) a collating symbol, like [. .]
15673 * 3) an equivalence class, like [= =]
15674 * In the latter two cases, it croaks if it finds a syntactically legal
15675 * one, as these are not handled by Perl.
15677 * The main purpose is to look for a POSIX character class. It returns:
15678 * a) the class number
15679 * if it is a completely syntactically and semantically legal class.
15680 * 'updated_parse_ptr', if not NULL, is set to point to just after the
15681 * closing ']' of the class
15682 * b) OOB_NAMEDCLASS
15683 * if it appears that one of the three POSIX constructs was meant, but
15684 * its specification was somehow defective. 'updated_parse_ptr', if
15685 * not NULL, is set to point to the character just after the end
15686 * character of the class. See below for handling of warnings.
15687 * c) NOT_MEANT_TO_BE_A_POSIX_CLASS
15688 * if it doesn't appear that a POSIX construct was intended.
15689 * 'updated_parse_ptr' is not changed. No warnings nor errors are
15692 * In b) there may be errors or warnings generated. If 'check_only' is
15693 * TRUE, then any errors are discarded. Warnings are returned to the
15694 * caller via an AV* created into '*posix_warnings' if it is not NULL. If
15695 * instead it is NULL, warnings are suppressed.
15697 * The reason for this function, and its complexity is that a bracketed
15698 * character class can contain just about anything. But it's easy to
15699 * mistype the very specific posix class syntax but yielding a valid
15700 * regular bracketed class, so it silently gets compiled into something
15701 * quite unintended.
15703 * The solution adopted here maintains backward compatibility except that
15704 * it adds a warning if it looks like a posix class was intended but
15705 * improperly specified. The warning is not raised unless what is input
15706 * very closely resembles one of the 14 legal posix classes. To do this,
15707 * it uses fuzzy parsing. It calculates how many single-character edits it
15708 * would take to transform what was input into a legal posix class. Only
15709 * if that number is quite small does it think that the intention was a
15710 * posix class. Obviously these are heuristics, and there will be cases
15711 * where it errs on one side or another, and they can be tweaked as
15712 * experience informs.
15714 * The syntax for a legal posix class is:
15716 * qr/(?xa: \[ : \^? [[:lower:]]{4,6} : \] )/
15718 * What this routine considers syntactically to be an intended posix class
15719 * is this (the comments indicate some restrictions that the pattern
15722 * qr/(?x: \[? # The left bracket, possibly
15724 * \h* # possibly followed by blanks
15725 * (?: \^ \h* )? # possibly a misplaced caret
15726 * [:;]? # The opening class character,
15727 * # possibly omitted. A typo
15728 * # semi-colon can also be used.
15730 * \^? # possibly a correctly placed
15731 * # caret, but not if there was also
15732 * # a misplaced one
15734 * .{3,15} # The class name. If there are
15735 * # deviations from the legal syntax,
15736 * # its edit distance must be close
15737 * # to a real class name in order
15738 * # for it to be considered to be
15739 * # an intended posix class.
15741 * [[:punct:]]? # The closing class character,
15742 * # possibly omitted. If not a colon
15743 * # nor semi colon, the class name
15744 * # must be even closer to a valid
15747 * \]? # The right bracket, possibly
15751 * In the above, \h must be ASCII-only.
15753 * These are heuristics, and can be tweaked as field experience dictates.
15754 * There will be cases when someone didn't intend to specify a posix class
15755 * that this warns as being so. The goal is to minimize these, while
15756 * maximizing the catching of things intended to be a posix class that
15757 * aren't parsed as such.
15761 const char * const e = RExC_end;
15762 unsigned complement = 0; /* If to complement the class */
15763 bool found_problem = FALSE; /* Assume OK until proven otherwise */
15764 bool has_opening_bracket = FALSE;
15765 bool has_opening_colon = FALSE;
15766 int class_number = OOB_NAMEDCLASS; /* Out-of-bounds until find
15768 const char * possible_end = NULL; /* used for a 2nd parse pass */
15769 const char* name_start; /* ptr to class name first char */
15771 /* If the number of single-character typos the input name is away from a
15772 * legal name is no more than this number, it is considered to have meant
15773 * the legal name */
15774 int max_distance = 2;
15776 /* to store the name. The size determines the maximum length before we
15777 * decide that no posix class was intended. Should be at least
15778 * sizeof("alphanumeric") */
15780 STATIC_ASSERT_DECL(C_ARRAY_LENGTH(input_text) >= sizeof "alphanumeric");
15782 PERL_ARGS_ASSERT_HANDLE_POSSIBLE_POSIX;
15784 CLEAR_POSIX_WARNINGS();
15787 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
15790 if (*(p - 1) != '[') {
15791 ADD_POSIX_WARNING(p, "it doesn't start with a '['");
15792 found_problem = TRUE;
15795 has_opening_bracket = TRUE;
15798 /* They could be confused and think you can put spaces between the
15801 found_problem = TRUE;
15805 } while (p < e && isBLANK(*p));
15807 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15810 /* For [. .] and [= =]. These are quite different internally from [: :],
15811 * so they are handled separately. */
15812 if (POSIXCC_NOTYET(*p) && p < e - 3) /* 1 for the close, and 1 for the ']'
15813 and 1 for at least one char in it
15816 const char open_char = *p;
15817 const char * temp_ptr = p + 1;
15819 /* These two constructs are not handled by perl, and if we find a
15820 * syntactically valid one, we croak. khw, who wrote this code, finds
15821 * this explanation of them very unclear:
15822 * http://pubs.opengroup.org/onlinepubs/009696899/basedefs/xbd_chap09.html
15823 * And searching the rest of the internet wasn't very helpful either.
15824 * It looks like just about any byte can be in these constructs,
15825 * depending on the locale. But unless the pattern is being compiled
15826 * under /l, which is very rare, Perl runs under the C or POSIX locale.
15827 * In that case, it looks like [= =] isn't allowed at all, and that
15828 * [. .] could be any single code point, but for longer strings the
15829 * constituent characters would have to be the ASCII alphabetics plus
15830 * the minus-hyphen. Any sensible locale definition would limit itself
15831 * to these. And any portable one definitely should. Trying to parse
15832 * the general case is a nightmare (see [perl #127604]). So, this code
15833 * looks only for interiors of these constructs that match:
15835 * Using \w relaxes the apparent rules a little, without adding much
15836 * danger of mistaking something else for one of these constructs.
15838 * [. .] in some implementations described on the internet is usable to
15839 * escape a character that otherwise is special in bracketed character
15840 * classes. For example [.].] means a literal right bracket instead of
15841 * the ending of the class
15843 * [= =] can legitimately contain a [. .] construct, but we don't
15844 * handle this case, as that [. .] construct will later get parsed
15845 * itself and croak then. And [= =] is checked for even when not under
15846 * /l, as Perl has long done so.
15848 * The code below relies on there being a trailing NUL, so it doesn't
15849 * have to keep checking if the parse ptr < e.
15851 if (temp_ptr[1] == open_char) {
15854 else while ( temp_ptr < e
15855 && (isWORDCHAR(*temp_ptr) || *temp_ptr == '-'))
15860 if (*temp_ptr == open_char) {
15862 if (*temp_ptr == ']') {
15864 if (! found_problem && ! check_only) {
15865 RExC_parse = (char *) temp_ptr;
15866 vFAIL3("POSIX syntax [%c %c] is reserved for future "
15867 "extensions", open_char, open_char);
15870 /* Here, the syntax wasn't completely valid, or else the call
15871 * is to check-only */
15872 if (updated_parse_ptr) {
15873 *updated_parse_ptr = (char *) temp_ptr;
15876 CLEAR_POSIX_WARNINGS_AND_RETURN(OOB_NAMEDCLASS);
15880 /* If we find something that started out to look like one of these
15881 * constructs, but isn't, we continue below so that it can be checked
15882 * for being a class name with a typo of '.' or '=' instead of a colon.
15886 /* Here, we think there is a possibility that a [: :] class was meant, and
15887 * we have the first real character. It could be they think the '^' comes
15890 found_problem = TRUE;
15891 ADD_POSIX_WARNING(p + 1, "the '^' must come after the colon");
15896 found_problem = TRUE;
15900 } while (p < e && isBLANK(*p));
15902 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15906 /* But the first character should be a colon, which they could have easily
15907 * mistyped on a qwerty keyboard as a semi-colon (and which may be hard to
15908 * distinguish from a colon, so treat that as a colon). */
15911 has_opening_colon = TRUE;
15913 else if (*p == ';') {
15914 found_problem = TRUE;
15916 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
15917 has_opening_colon = TRUE;
15920 found_problem = TRUE;
15921 ADD_POSIX_WARNING(p, "there must be a starting ':'");
15923 /* Consider an initial punctuation (not one of the recognized ones) to
15924 * be a left terminator */
15925 if (*p != '^' && *p != ']' && isPUNCT(*p)) {
15930 /* They may think that you can put spaces between the components */
15932 found_problem = TRUE;
15936 } while (p < e && isBLANK(*p));
15938 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15943 /* We consider something like [^:^alnum:]] to not have been intended to
15944 * be a posix class, but XXX maybe we should */
15946 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15953 /* Again, they may think that you can put spaces between the components */
15955 found_problem = TRUE;
15959 } while (p < e && isBLANK(*p));
15961 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15966 /* XXX This ']' may be a typo, and something else was meant. But
15967 * treating it as such creates enough complications, that that
15968 * possibility isn't currently considered here. So we assume that the
15969 * ']' is what is intended, and if we've already found an initial '[',
15970 * this leaves this construct looking like [:] or [:^], which almost
15971 * certainly weren't intended to be posix classes */
15972 if (has_opening_bracket) {
15973 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15976 /* But this function can be called when we parse the colon for
15977 * something like qr/[alpha:]]/, so we back up to look for the
15982 found_problem = TRUE;
15983 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
15985 else if (*p != ':') {
15987 /* XXX We are currently very restrictive here, so this code doesn't
15988 * consider the possibility that, say, /[alpha.]]/ was intended to
15989 * be a posix class. */
15990 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15993 /* Here we have something like 'foo:]'. There was no initial colon,
15994 * and we back up over 'foo. XXX Unlike the going forward case, we
15995 * don't handle typos of non-word chars in the middle */
15996 has_opening_colon = FALSE;
15999 while (p > RExC_start && isWORDCHAR(*p)) {
16004 /* Here, we have positioned ourselves to where we think the first
16005 * character in the potential class is */
16008 /* Now the interior really starts. There are certain key characters that
16009 * can end the interior, or these could just be typos. To catch both
16010 * cases, we may have to do two passes. In the first pass, we keep on
16011 * going unless we come to a sequence that matches
16012 * qr/ [[:punct:]] [[:blank:]]* \] /xa
16013 * This means it takes a sequence to end the pass, so two typos in a row if
16014 * that wasn't what was intended. If the class is perfectly formed, just
16015 * this one pass is needed. We also stop if there are too many characters
16016 * being accumulated, but this number is deliberately set higher than any
16017 * real class. It is set high enough so that someone who thinks that
16018 * 'alphanumeric' is a correct name would get warned that it wasn't.
16019 * While doing the pass, we keep track of where the key characters were in
16020 * it. If we don't find an end to the class, and one of the key characters
16021 * was found, we redo the pass, but stop when we get to that character.
16022 * Thus the key character was considered a typo in the first pass, but a
16023 * terminator in the second. If two key characters are found, we stop at
16024 * the second one in the first pass. Again this can miss two typos, but
16025 * catches a single one
16027 * In the first pass, 'possible_end' starts as NULL, and then gets set to
16028 * point to the first key character. For the second pass, it starts as -1.
16034 bool has_blank = FALSE;
16035 bool has_upper = FALSE;
16036 bool has_terminating_colon = FALSE;
16037 bool has_terminating_bracket = FALSE;
16038 bool has_semi_colon = FALSE;
16039 unsigned int name_len = 0;
16040 int punct_count = 0;
16044 /* Squeeze out blanks when looking up the class name below */
16045 if (isBLANK(*p) ) {
16047 found_problem = TRUE;
16052 /* The name will end with a punctuation */
16054 const char * peek = p + 1;
16056 /* Treat any non-']' punctuation followed by a ']' (possibly
16057 * with intervening blanks) as trying to terminate the class.
16058 * ']]' is very likely to mean a class was intended (but
16059 * missing the colon), but the warning message that gets
16060 * generated shows the error position better if we exit the
16061 * loop at the bottom (eventually), so skip it here. */
16063 if (peek < e && isBLANK(*peek)) {
16065 found_problem = TRUE;
16068 } while (peek < e && isBLANK(*peek));
16071 if (peek < e && *peek == ']') {
16072 has_terminating_bracket = TRUE;
16074 has_terminating_colon = TRUE;
16076 else if (*p == ';') {
16077 has_semi_colon = TRUE;
16078 has_terminating_colon = TRUE;
16081 found_problem = TRUE;
16088 /* Here we have punctuation we thought didn't end the class.
16089 * Keep track of the position of the key characters that are
16090 * more likely to have been class-enders */
16091 if (*p == ']' || *p == '[' || *p == ':' || *p == ';') {
16093 /* Allow just one such possible class-ender not actually
16094 * ending the class. */
16095 if (possible_end) {
16101 /* If we have too many punctuation characters, no use in
16103 if (++punct_count > max_distance) {
16107 /* Treat the punctuation as a typo. */
16108 input_text[name_len++] = *p;
16111 else if (isUPPER(*p)) { /* Use lowercase for lookup */
16112 input_text[name_len++] = toLOWER(*p);
16114 found_problem = TRUE;
16116 } else if (! UTF || UTF8_IS_INVARIANT(*p)) {
16117 input_text[name_len++] = *p;
16121 input_text[name_len++] = utf8_to_uvchr_buf((U8 *) p, e, NULL);
16125 /* The declaration of 'input_text' is how long we allow a potential
16126 * class name to be, before saying they didn't mean a class name at
16128 if (name_len >= C_ARRAY_LENGTH(input_text)) {
16133 /* We get to here when the possible class name hasn't been properly
16134 * terminated before:
16135 * 1) we ran off the end of the pattern; or
16136 * 2) found two characters, each of which might have been intended to
16137 * be the name's terminator
16138 * 3) found so many punctuation characters in the purported name,
16139 * that the edit distance to a valid one is exceeded
16140 * 4) we decided it was more characters than anyone could have
16141 * intended to be one. */
16143 found_problem = TRUE;
16145 /* In the final two cases, we know that looking up what we've
16146 * accumulated won't lead to a match, even a fuzzy one. */
16147 if ( name_len >= C_ARRAY_LENGTH(input_text)
16148 || punct_count > max_distance)
16150 /* If there was an intermediate key character that could have been
16151 * an intended end, redo the parse, but stop there */
16152 if (possible_end && possible_end != (char *) -1) {
16153 possible_end = (char *) -1; /* Special signal value to say
16154 we've done a first pass */
16159 /* Otherwise, it can't have meant to have been a class */
16160 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
16163 /* If we ran off the end, and the final character was a punctuation
16164 * one, back up one, to look at that final one just below. Later, we
16165 * will restore the parse pointer if appropriate */
16166 if (name_len && p == e && isPUNCT(*(p-1))) {
16171 if (p < e && isPUNCT(*p)) {
16173 has_terminating_bracket = TRUE;
16175 /* If this is a 2nd ']', and the first one is just below this
16176 * one, consider that to be the real terminator. This gives a
16177 * uniform and better positioning for the warning message */
16179 && possible_end != (char *) -1
16180 && *possible_end == ']'
16181 && name_len && input_text[name_len - 1] == ']')
16186 /* And this is actually equivalent to having done the 2nd
16187 * pass now, so set it to not try again */
16188 possible_end = (char *) -1;
16193 has_terminating_colon = TRUE;
16195 else if (*p == ';') {
16196 has_semi_colon = TRUE;
16197 has_terminating_colon = TRUE;
16205 /* Here, we have a class name to look up. We can short circuit the
16206 * stuff below for short names that can't possibly be meant to be a
16207 * class name. (We can do this on the first pass, as any second pass
16208 * will yield an even shorter name) */
16209 if (name_len < 3) {
16210 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
16213 /* Find which class it is. Initially switch on the length of the name.
16215 switch (name_len) {
16217 if (memEQs(name_start, 4, "word")) {
16218 /* this is not POSIX, this is the Perl \w */
16219 class_number = ANYOF_WORDCHAR;
16223 /* Names all of length 5: alnum alpha ascii blank cntrl digit
16224 * graph lower print punct space upper
16225 * Offset 4 gives the best switch position. */
16226 switch (name_start[4]) {
16228 if (memBEGINs(name_start, 5, "alph")) /* alpha */
16229 class_number = ANYOF_ALPHA;
16232 if (memBEGINs(name_start, 5, "spac")) /* space */
16233 class_number = ANYOF_SPACE;
16236 if (memBEGINs(name_start, 5, "grap")) /* graph */
16237 class_number = ANYOF_GRAPH;
16240 if (memBEGINs(name_start, 5, "asci")) /* ascii */
16241 class_number = ANYOF_ASCII;
16244 if (memBEGINs(name_start, 5, "blan")) /* blank */
16245 class_number = ANYOF_BLANK;
16248 if (memBEGINs(name_start, 5, "cntr")) /* cntrl */
16249 class_number = ANYOF_CNTRL;
16252 if (memBEGINs(name_start, 5, "alnu")) /* alnum */
16253 class_number = ANYOF_ALPHANUMERIC;
16256 if (memBEGINs(name_start, 5, "lowe")) /* lower */
16257 class_number = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
16258 else if (memBEGINs(name_start, 5, "uppe")) /* upper */
16259 class_number = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
16262 if (memBEGINs(name_start, 5, "digi")) /* digit */
16263 class_number = ANYOF_DIGIT;
16264 else if (memBEGINs(name_start, 5, "prin")) /* print */
16265 class_number = ANYOF_PRINT;
16266 else if (memBEGINs(name_start, 5, "punc")) /* punct */
16267 class_number = ANYOF_PUNCT;
16272 if (memEQs(name_start, 6, "xdigit"))
16273 class_number = ANYOF_XDIGIT;
16277 /* If the name exactly matches a posix class name the class number will
16278 * here be set to it, and the input almost certainly was meant to be a
16279 * posix class, so we can skip further checking. If instead the syntax
16280 * is exactly correct, but the name isn't one of the legal ones, we
16281 * will return that as an error below. But if neither of these apply,
16282 * it could be that no posix class was intended at all, or that one
16283 * was, but there was a typo. We tease these apart by doing fuzzy
16284 * matching on the name */
16285 if (class_number == OOB_NAMEDCLASS && found_problem) {
16286 const UV posix_names[][6] = {
16287 { 'a', 'l', 'n', 'u', 'm' },
16288 { 'a', 'l', 'p', 'h', 'a' },
16289 { 'a', 's', 'c', 'i', 'i' },
16290 { 'b', 'l', 'a', 'n', 'k' },
16291 { 'c', 'n', 't', 'r', 'l' },
16292 { 'd', 'i', 'g', 'i', 't' },
16293 { 'g', 'r', 'a', 'p', 'h' },
16294 { 'l', 'o', 'w', 'e', 'r' },
16295 { 'p', 'r', 'i', 'n', 't' },
16296 { 'p', 'u', 'n', 'c', 't' },
16297 { 's', 'p', 'a', 'c', 'e' },
16298 { 'u', 'p', 'p', 'e', 'r' },
16299 { 'w', 'o', 'r', 'd' },
16300 { 'x', 'd', 'i', 'g', 'i', 't' }
16302 /* The names of the above all have added NULs to make them the same
16303 * size, so we need to also have the real lengths */
16304 const UV posix_name_lengths[] = {
16305 sizeof("alnum") - 1,
16306 sizeof("alpha") - 1,
16307 sizeof("ascii") - 1,
16308 sizeof("blank") - 1,
16309 sizeof("cntrl") - 1,
16310 sizeof("digit") - 1,
16311 sizeof("graph") - 1,
16312 sizeof("lower") - 1,
16313 sizeof("print") - 1,
16314 sizeof("punct") - 1,
16315 sizeof("space") - 1,
16316 sizeof("upper") - 1,
16317 sizeof("word") - 1,
16318 sizeof("xdigit")- 1
16321 int temp_max = max_distance; /* Use a temporary, so if we
16322 reparse, we haven't changed the
16325 /* Use a smaller max edit distance if we are missing one of the
16327 if ( has_opening_bracket + has_opening_colon < 2
16328 || has_terminating_bracket + has_terminating_colon < 2)
16333 /* See if the input name is close to a legal one */
16334 for (i = 0; i < C_ARRAY_LENGTH(posix_names); i++) {
16336 /* Short circuit call if the lengths are too far apart to be
16338 if (abs( (int) (name_len - posix_name_lengths[i]))
16344 if (edit_distance(input_text,
16347 posix_name_lengths[i],
16351 { /* If it is close, it probably was intended to be a class */
16352 goto probably_meant_to_be;
16356 /* Here the input name is not close enough to a valid class name
16357 * for us to consider it to be intended to be a posix class. If
16358 * we haven't already done so, and the parse found a character that
16359 * could have been terminators for the name, but which we absorbed
16360 * as typos during the first pass, repeat the parse, signalling it
16361 * to stop at that character */
16362 if (possible_end && possible_end != (char *) -1) {
16363 possible_end = (char *) -1;
16368 /* Here neither pass found a close-enough class name */
16369 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
16372 probably_meant_to_be:
16374 /* Here we think that a posix specification was intended. Update any
16376 if (updated_parse_ptr) {
16377 *updated_parse_ptr = (char *) p;
16380 /* If a posix class name was intended but incorrectly specified, we
16381 * output or return the warnings */
16382 if (found_problem) {
16384 /* We set flags for these issues in the parse loop above instead of
16385 * adding them to the list of warnings, because we can parse it
16386 * twice, and we only want one warning instance */
16388 ADD_POSIX_WARNING(p, "the name must be all lowercase letters");
16391 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
16393 if (has_semi_colon) {
16394 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
16396 else if (! has_terminating_colon) {
16397 ADD_POSIX_WARNING(p, "there is no terminating ':'");
16399 if (! has_terminating_bracket) {
16400 ADD_POSIX_WARNING(p, "there is no terminating ']'");
16403 if ( posix_warnings
16405 && av_count(RExC_warn_text) > 0)
16407 *posix_warnings = RExC_warn_text;
16410 else if (class_number != OOB_NAMEDCLASS) {
16411 /* If it is a known class, return the class. The class number
16412 * #defines are structured so each complement is +1 to the normal
16414 CLEAR_POSIX_WARNINGS_AND_RETURN(class_number + complement);
16416 else if (! check_only) {
16418 /* Here, it is an unrecognized class. This is an error (unless the
16419 * call is to check only, which we've already handled above) */
16420 const char * const complement_string = (complement)
16423 RExC_parse = (char *) p;
16424 vFAIL3utf8f("POSIX class [:%s%" UTF8f ":] unknown",
16426 UTF8fARG(UTF, RExC_parse - name_start - 2, name_start));
16430 return OOB_NAMEDCLASS;
16432 #undef ADD_POSIX_WARNING
16434 STATIC unsigned int
16435 S_regex_set_precedence(const U8 my_operator) {
16437 /* Returns the precedence in the (?[...]) construct of the input operator,
16438 * specified by its character representation. The precedence follows
16439 * general Perl rules, but it extends this so that ')' and ']' have (low)
16440 * precedence even though they aren't really operators */
16442 switch (my_operator) {
16458 NOT_REACHED; /* NOTREACHED */
16459 return 0; /* Silence compiler warning */
16462 STATIC regnode_offset
16463 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
16464 I32 *flagp, U32 depth,
16465 char * const oregcomp_parse)
16467 /* Handle the (?[...]) construct to do set operations */
16469 U8 curchar; /* Current character being parsed */
16470 UV start, end; /* End points of code point ranges */
16471 SV* final = NULL; /* The end result inversion list */
16472 SV* result_string; /* 'final' stringified */
16473 AV* stack; /* stack of operators and operands not yet
16475 AV* fence_stack = NULL; /* A stack containing the positions in
16476 'stack' of where the undealt-with left
16477 parens would be if they were actually
16479 /* The 'volatile' is a workaround for an optimiser bug
16480 * in Solaris Studio 12.3. See RT #127455 */
16481 volatile IV fence = 0; /* Position of where most recent undealt-
16482 with left paren in stack is; -1 if none.
16484 STRLEN len; /* Temporary */
16485 regnode_offset node; /* Temporary, and final regnode returned by
16487 const bool save_fold = FOLD; /* Temporary */
16488 char *save_end, *save_parse; /* Temporaries */
16489 const bool in_locale = LOC; /* we turn off /l during processing */
16491 DECLARE_AND_GET_RE_DEBUG_FLAGS;
16493 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
16494 PERL_UNUSED_ARG(oregcomp_parse); /* Only for Set_Node_Length */
16496 DEBUG_PARSE("xcls");
16499 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
16502 /* The use of this operator implies /u. This is required so that the
16503 * compile time values are valid in all runtime cases */
16504 REQUIRE_UNI_RULES(flagp, 0);
16506 ckWARNexperimental(RExC_parse,
16507 WARN_EXPERIMENTAL__REGEX_SETS,
16508 "The regex_sets feature is experimental");
16510 /* Everything in this construct is a metacharacter. Operands begin with
16511 * either a '\' (for an escape sequence), or a '[' for a bracketed
16512 * character class. Any other character should be an operator, or
16513 * parenthesis for grouping. Both types of operands are handled by calling
16514 * regclass() to parse them. It is called with a parameter to indicate to
16515 * return the computed inversion list. The parsing here is implemented via
16516 * a stack. Each entry on the stack is a single character representing one
16517 * of the operators; or else a pointer to an operand inversion list. */
16519 #define IS_OPERATOR(a) SvIOK(a)
16520 #define IS_OPERAND(a) (! IS_OPERATOR(a))
16522 /* The stack is kept in Łukasiewicz order. (That's pronounced similar
16523 * to luke-a-shave-itch (or -itz), but people who didn't want to bother
16524 * with pronouncing it called it Reverse Polish instead, but now that YOU
16525 * know how to pronounce it you can use the correct term, thus giving due
16526 * credit to the person who invented it, and impressing your geek friends.
16527 * Wikipedia says that the pronounciation of "Ł" has been changing so that
16528 * it is now more like an English initial W (as in wonk) than an L.)
16530 * This means that, for example, 'a | b & c' is stored on the stack as
16538 * where the numbers in brackets give the stack [array] element number.
16539 * In this implementation, parentheses are not stored on the stack.
16540 * Instead a '(' creates a "fence" so that the part of the stack below the
16541 * fence is invisible except to the corresponding ')' (this allows us to
16542 * replace testing for parens, by using instead subtraction of the fence
16543 * position). As new operands are processed they are pushed onto the stack
16544 * (except as noted in the next paragraph). New operators of higher
16545 * precedence than the current final one are inserted on the stack before
16546 * the lhs operand (so that when the rhs is pushed next, everything will be
16547 * in the correct positions shown above. When an operator of equal or
16548 * lower precedence is encountered in parsing, all the stacked operations
16549 * of equal or higher precedence are evaluated, leaving the result as the
16550 * top entry on the stack. This makes higher precedence operations
16551 * evaluate before lower precedence ones, and causes operations of equal
16552 * precedence to left associate.
16554 * The only unary operator '!' is immediately pushed onto the stack when
16555 * encountered. When an operand is encountered, if the top of the stack is
16556 * a '!", the complement is immediately performed, and the '!' popped. The
16557 * resulting value is treated as a new operand, and the logic in the
16558 * previous paragraph is executed. Thus in the expression
16560 * the stack looks like
16566 * as 'b' gets parsed, the latter gets evaluated to '!b', and the stack
16573 * A ')' is treated as an operator with lower precedence than all the
16574 * aforementioned ones, which causes all operations on the stack above the
16575 * corresponding '(' to be evaluated down to a single resultant operand.
16576 * Then the fence for the '(' is removed, and the operand goes through the
16577 * algorithm above, without the fence.
16579 * A separate stack is kept of the fence positions, so that the position of
16580 * the latest so-far unbalanced '(' is at the top of it.
16582 * The ']' ending the construct is treated as the lowest operator of all,
16583 * so that everything gets evaluated down to a single operand, which is the
16586 sv_2mortal((SV *)(stack = newAV()));
16587 sv_2mortal((SV *)(fence_stack = newAV()));
16589 while (RExC_parse < RExC_end) {
16590 I32 top_index; /* Index of top-most element in 'stack' */
16591 SV** top_ptr; /* Pointer to top 'stack' element */
16592 SV* current = NULL; /* To contain the current inversion list
16594 SV* only_to_avoid_leaks;
16596 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
16597 TRUE /* Force /x */ );
16598 if (RExC_parse >= RExC_end) { /* Fail */
16602 curchar = UCHARAT(RExC_parse);
16606 #ifdef ENABLE_REGEX_SETS_DEBUGGING
16607 /* Enable with -Accflags=-DENABLE_REGEX_SETS_DEBUGGING */
16608 DEBUG_U(dump_regex_sets_structures(pRExC_state,
16609 stack, fence, fence_stack));
16612 top_index = av_tindex_skip_len_mg(stack);
16615 SV** stacked_ptr; /* Ptr to something already on 'stack' */
16616 char stacked_operator; /* The topmost operator on the 'stack'. */
16617 SV* lhs; /* Operand to the left of the operator */
16618 SV* rhs; /* Operand to the right of the operator */
16619 SV* fence_ptr; /* Pointer to top element of the fence
16623 if ( RExC_parse < RExC_end - 2
16624 && UCHARAT(RExC_parse + 1) == '?'
16625 && UCHARAT(RExC_parse + 2) == '^')
16627 const regnode_offset orig_emit = RExC_emit;
16628 SV * resultant_invlist;
16630 /* If is a '(?^', could be an embedded '(?^flags:(?[...])'.
16631 * This happens when we have some thing like
16633 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
16635 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
16637 * Here we would be handling the interpolated
16638 * '$thai_or_lao'. We handle this by a recursive call to
16639 * reg which returns the inversion list the
16640 * interpolated expression evaluates to. Actually, the
16641 * return is a special regnode containing a pointer to that
16642 * inversion list. If the return isn't that regnode alone,
16643 * we know that this wasn't such an interpolation, which is
16644 * an error: we need to get a single inversion list back
16645 * from the recursion */
16650 node = reg(pRExC_state, 2, flagp, depth+1);
16651 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16653 if ( OP(REGNODE_p(node)) != REGEX_SET
16654 /* If more than a single node returned, the nested
16655 * parens evaluated to more than just a (?[...]),
16656 * which isn't legal */
16657 || RExC_emit != orig_emit
16658 + NODE_STEP_REGNODE
16659 + regarglen[REGEX_SET])
16661 vFAIL("Expecting interpolated extended charclass");
16663 resultant_invlist = (SV *) ARGp(REGNODE_p(node));
16664 current = invlist_clone(resultant_invlist, NULL);
16665 SvREFCNT_dec(resultant_invlist);
16668 RExC_emit = orig_emit;
16669 goto handle_operand;
16672 /* A regular '('. Look behind for illegal syntax */
16673 if (top_index - fence >= 0) {
16674 /* If the top entry on the stack is an operator, it had
16675 * better be a '!', otherwise the entry below the top
16676 * operand should be an operator */
16677 if ( ! (top_ptr = av_fetch(stack, top_index, FALSE))
16678 || (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) != '!')
16679 || ( IS_OPERAND(*top_ptr)
16680 && ( top_index - fence < 1
16681 || ! (stacked_ptr = av_fetch(stack,
16684 || ! IS_OPERATOR(*stacked_ptr))))
16687 vFAIL("Unexpected '(' with no preceding operator");
16691 /* Stack the position of this undealt-with left paren */
16692 av_push(fence_stack, newSViv(fence));
16693 fence = top_index + 1;
16697 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
16698 * multi-char folds are allowed. */
16699 if (!regclass(pRExC_state, flagp, depth+1,
16700 TRUE, /* means parse just the next thing */
16701 FALSE, /* don't allow multi-char folds */
16702 FALSE, /* don't silence non-portable warnings. */
16704 FALSE, /* Require return to be an ANYOF */
16707 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16708 goto regclass_failed;
16713 /* regclass() will return with parsing just the \ sequence,
16714 * leaving the parse pointer at the next thing to parse */
16716 goto handle_operand;
16718 case '[': /* Is a bracketed character class */
16720 /* See if this is a [:posix:] class. */
16721 bool is_posix_class = (OOB_NAMEDCLASS
16722 < handle_possible_posix(pRExC_state,
16726 TRUE /* checking only */));
16727 /* If it is a posix class, leave the parse pointer at the '['
16728 * to fool regclass() into thinking it is part of a
16729 * '[[:posix:]]'. */
16730 if (! is_posix_class) {
16734 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
16735 * multi-char folds are allowed. */
16736 if (!regclass(pRExC_state, flagp, depth+1,
16737 is_posix_class, /* parse the whole char
16738 class only if not a
16740 FALSE, /* don't allow multi-char folds */
16741 TRUE, /* silence non-portable warnings. */
16743 FALSE, /* Require return to be an ANYOF */
16746 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16747 goto regclass_failed;
16752 /* function call leaves parse pointing to the ']', except if we
16754 if (is_posix_class) {
16758 goto handle_operand;
16762 if (top_index >= 1) {
16763 goto join_operators;
16766 /* Only a single operand on the stack: are done */
16770 if (av_tindex_skip_len_mg(fence_stack) < 0) {
16771 if (UCHARAT(RExC_parse - 1) == ']') {
16775 vFAIL("Unexpected ')'");
16778 /* If nothing after the fence, is missing an operand */
16779 if (top_index - fence < 0) {
16783 /* If at least two things on the stack, treat this as an
16785 if (top_index - fence >= 1) {
16786 goto join_operators;
16789 /* Here only a single thing on the fenced stack, and there is a
16790 * fence. Get rid of it */
16791 fence_ptr = av_pop(fence_stack);
16793 fence = SvIV(fence_ptr);
16794 SvREFCNT_dec_NN(fence_ptr);
16801 /* Having gotten rid of the fence, we pop the operand at the
16802 * stack top and process it as a newly encountered operand */
16803 current = av_pop(stack);
16804 if (IS_OPERAND(current)) {
16805 goto handle_operand;
16817 /* These binary operators should have a left operand already
16819 if ( top_index - fence < 0
16820 || top_index - fence == 1
16821 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
16822 || ! IS_OPERAND(*top_ptr))
16824 goto unexpected_binary;
16827 /* If only the one operand is on the part of the stack visible
16828 * to us, we just place this operator in the proper position */
16829 if (top_index - fence < 2) {
16831 /* Place the operator before the operand */
16833 SV* lhs = av_pop(stack);
16834 av_push(stack, newSVuv(curchar));
16835 av_push(stack, lhs);
16839 /* But if there is something else on the stack, we need to
16840 * process it before this new operator if and only if the
16841 * stacked operation has equal or higher precedence than the
16846 /* The operator on the stack is supposed to be below both its
16848 if ( ! (stacked_ptr = av_fetch(stack, top_index - 2, FALSE))
16849 || IS_OPERAND(*stacked_ptr))
16851 /* But if not, it's legal and indicates we are completely
16852 * done if and only if we're currently processing a ']',
16853 * which should be the final thing in the expression */
16854 if (curchar == ']') {
16860 vFAIL2("Unexpected binary operator '%c' with no "
16861 "preceding operand", curchar);
16863 stacked_operator = (char) SvUV(*stacked_ptr);
16865 if (regex_set_precedence(curchar)
16866 > regex_set_precedence(stacked_operator))
16868 /* Here, the new operator has higher precedence than the
16869 * stacked one. This means we need to add the new one to
16870 * the stack to await its rhs operand (and maybe more
16871 * stuff). We put it before the lhs operand, leaving
16872 * untouched the stacked operator and everything below it
16874 lhs = av_pop(stack);
16875 assert(IS_OPERAND(lhs));
16877 av_push(stack, newSVuv(curchar));
16878 av_push(stack, lhs);
16882 /* Here, the new operator has equal or lower precedence than
16883 * what's already there. This means the operation already
16884 * there should be performed now, before the new one. */
16886 rhs = av_pop(stack);
16887 if (! IS_OPERAND(rhs)) {
16889 /* This can happen when a ! is not followed by an operand,
16890 * like in /(?[\t &!])/ */
16894 lhs = av_pop(stack);
16896 if (! IS_OPERAND(lhs)) {
16898 /* This can happen when there is an empty (), like in
16899 * /(?[[0]+()+])/ */
16903 switch (stacked_operator) {
16905 _invlist_intersection(lhs, rhs, &rhs);
16910 _invlist_union(lhs, rhs, &rhs);
16914 _invlist_subtract(lhs, rhs, &rhs);
16917 case '^': /* The union minus the intersection */
16922 _invlist_union(lhs, rhs, &u);
16923 _invlist_intersection(lhs, rhs, &i);
16924 _invlist_subtract(u, i, &rhs);
16925 SvREFCNT_dec_NN(i);
16926 SvREFCNT_dec_NN(u);
16932 /* Here, the higher precedence operation has been done, and the
16933 * result is in 'rhs'. We overwrite the stacked operator with
16934 * the result. Then we redo this code to either push the new
16935 * operator onto the stack or perform any higher precedence
16936 * stacked operation */
16937 only_to_avoid_leaks = av_pop(stack);
16938 SvREFCNT_dec(only_to_avoid_leaks);
16939 av_push(stack, rhs);
16942 case '!': /* Highest priority, right associative */
16944 /* If what's already at the top of the stack is another '!",
16945 * they just cancel each other out */
16946 if ( (top_ptr = av_fetch(stack, top_index, FALSE))
16947 && (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) == '!'))
16949 only_to_avoid_leaks = av_pop(stack);
16950 SvREFCNT_dec(only_to_avoid_leaks);
16952 else { /* Otherwise, since it's right associative, just push
16954 av_push(stack, newSVuv(curchar));
16959 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16960 if (RExC_parse >= RExC_end) {
16963 vFAIL("Unexpected character");
16967 /* Here 'current' is the operand. If something is already on the
16968 * stack, we have to check if it is a !. But first, the code above
16969 * may have altered the stack in the time since we earlier set
16972 top_index = av_tindex_skip_len_mg(stack);
16973 if (top_index - fence >= 0) {
16974 /* If the top entry on the stack is an operator, it had better
16975 * be a '!', otherwise the entry below the top operand should
16976 * be an operator */
16977 top_ptr = av_fetch(stack, top_index, FALSE);
16979 if (IS_OPERATOR(*top_ptr)) {
16981 /* The only permissible operator at the top of the stack is
16982 * '!', which is applied immediately to this operand. */
16983 curchar = (char) SvUV(*top_ptr);
16984 if (curchar != '!') {
16985 SvREFCNT_dec(current);
16986 vFAIL2("Unexpected binary operator '%c' with no "
16987 "preceding operand", curchar);
16990 _invlist_invert(current);
16992 only_to_avoid_leaks = av_pop(stack);
16993 SvREFCNT_dec(only_to_avoid_leaks);
16995 /* And we redo with the inverted operand. This allows
16996 * handling multiple ! in a row */
16997 goto handle_operand;
16999 /* Single operand is ok only for the non-binary ')'
17001 else if ((top_index - fence == 0 && curchar != ')')
17002 || (top_index - fence > 0
17003 && (! (stacked_ptr = av_fetch(stack,
17006 || IS_OPERAND(*stacked_ptr))))
17008 SvREFCNT_dec(current);
17009 vFAIL("Operand with no preceding operator");
17013 /* Here there was nothing on the stack or the top element was
17014 * another operand. Just add this new one */
17015 av_push(stack, current);
17017 } /* End of switch on next parse token */
17019 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
17020 } /* End of loop parsing through the construct */
17022 vFAIL("Syntax error in (?[...])");
17026 if (RExC_parse >= RExC_end || RExC_parse[1] != ')') {
17027 if (RExC_parse < RExC_end) {
17031 vFAIL("Unexpected ']' with no following ')' in (?[...");
17034 if (av_tindex_skip_len_mg(fence_stack) >= 0) {
17035 vFAIL("Unmatched (");
17038 if (av_tindex_skip_len_mg(stack) < 0 /* Was empty */
17039 || ((final = av_pop(stack)) == NULL)
17040 || ! IS_OPERAND(final)
17041 || ! is_invlist(final)
17042 || av_tindex_skip_len_mg(stack) >= 0) /* More left on stack */
17045 SvREFCNT_dec(final);
17046 vFAIL("Incomplete expression within '(?[ ])'");
17049 /* Here, 'final' is the resultant inversion list from evaluating the
17050 * expression. Return it if so requested */
17051 if (return_invlist) {
17052 *return_invlist = final;
17056 if (RExC_sets_depth) { /* If within a recursive call, return in a special
17059 node = regpnode(pRExC_state, REGEX_SET, final);
17063 /* Otherwise generate a resultant node, based on 'final'. regclass()
17064 * is expecting a string of ranges and individual code points */
17065 invlist_iterinit(final);
17066 result_string = newSVpvs("");
17067 while (invlist_iternext(final, &start, &end)) {
17068 if (start == end) {
17069 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}", start);
17072 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}-\\x{%"
17073 UVXf "}", start, end);
17077 /* About to generate an ANYOF (or similar) node from the inversion list
17078 * we have calculated */
17079 save_parse = RExC_parse;
17080 RExC_parse = SvPV(result_string, len);
17081 save_end = RExC_end;
17082 RExC_end = RExC_parse + len;
17083 TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE;
17085 /* We turn off folding around the call, as the class we have
17086 * constructed already has all folding taken into consideration, and we
17087 * don't want regclass() to add to that */
17088 RExC_flags &= ~RXf_PMf_FOLD;
17089 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if multi-char
17090 * folds are allowed. */
17091 node = regclass(pRExC_state, flagp, depth+1,
17092 FALSE, /* means parse the whole char class */
17093 FALSE, /* don't allow multi-char folds */
17094 TRUE, /* silence non-portable warnings. The above may
17095 very well have generated non-portable code
17096 points, but they're valid on this machine */
17097 FALSE, /* similarly, no need for strict */
17099 /* We can optimize into something besides an ANYOF,
17100 * except under /l, which needs to be ANYOF because of
17101 * runtime checks for locale sanity, etc */
17107 RExC_parse = save_parse + 1;
17108 RExC_end = save_end;
17109 SvREFCNT_dec_NN(final);
17110 SvREFCNT_dec_NN(result_string);
17113 RExC_flags |= RXf_PMf_FOLD;
17117 RETURN_FAIL_ON_RESTART(*flagp, flagp);
17118 goto regclass_failed;
17121 /* Fix up the node type if we are in locale. (We have pretended we are
17122 * under /u for the purposes of regclass(), as this construct will only
17123 * work under UTF-8 locales. But now we change the opcode to be ANYOFL
17124 * (so as to cause any warnings about bad locales to be output in
17125 * regexec.c), and add the flag that indicates to check if not in a
17126 * UTF-8 locale. The reason we above forbid optimization into
17127 * something other than an ANYOF node is simply to minimize the number
17128 * of code changes in regexec.c. Otherwise we would have to create new
17129 * EXACTish node types and deal with them. This decision could be
17130 * revisited should this construct become popular.
17132 * (One might think we could look at the resulting ANYOF node and
17133 * suppress the flag if everything is above 255, as those would be
17134 * UTF-8 only, but this isn't true, as the components that led to that
17135 * result could have been locale-affected, and just happen to cancel
17136 * each other out under UTF-8 locales.) */
17138 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
17140 assert(OP(REGNODE_p(node)) == ANYOF);
17142 OP(REGNODE_p(node)) = ANYOFL;
17143 ANYOF_FLAGS(REGNODE_p(node))
17144 |= ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
17148 nextchar(pRExC_state);
17149 Set_Node_Length(REGNODE_p(node), RExC_parse - oregcomp_parse + 1); /* MJD */
17153 FAIL2("panic: regclass returned failure to handle_sets, " "flags=%#" UVxf,
17157 #ifdef ENABLE_REGEX_SETS_DEBUGGING
17160 S_dump_regex_sets_structures(pTHX_ RExC_state_t *pRExC_state,
17161 AV * stack, const IV fence, AV * fence_stack)
17162 { /* Dumps the stacks in handle_regex_sets() */
17164 const SSize_t stack_top = av_tindex_skip_len_mg(stack);
17165 const SSize_t fence_stack_top = av_tindex_skip_len_mg(fence_stack);
17168 PERL_ARGS_ASSERT_DUMP_REGEX_SETS_STRUCTURES;
17170 PerlIO_printf(Perl_debug_log, "\nParse position is:%s\n", RExC_parse);
17172 if (stack_top < 0) {
17173 PerlIO_printf(Perl_debug_log, "Nothing on stack\n");
17176 PerlIO_printf(Perl_debug_log, "Stack: (fence=%d)\n", (int) fence);
17177 for (i = stack_top; i >= 0; i--) {
17178 SV ** element_ptr = av_fetch(stack, i, FALSE);
17179 if (! element_ptr) {
17182 if (IS_OPERATOR(*element_ptr)) {
17183 PerlIO_printf(Perl_debug_log, "[%d]: %c\n",
17184 (int) i, (int) SvIV(*element_ptr));
17187 PerlIO_printf(Perl_debug_log, "[%d] ", (int) i);
17188 sv_dump(*element_ptr);
17193 if (fence_stack_top < 0) {
17194 PerlIO_printf(Perl_debug_log, "Nothing on fence_stack\n");
17197 PerlIO_printf(Perl_debug_log, "Fence_stack: \n");
17198 for (i = fence_stack_top; i >= 0; i--) {
17199 SV ** element_ptr = av_fetch(fence_stack, i, FALSE);
17200 if (! element_ptr) {
17203 PerlIO_printf(Perl_debug_log, "[%d]: %d\n",
17204 (int) i, (int) SvIV(*element_ptr));
17215 S_add_above_Latin1_folds(pTHX_ RExC_state_t *pRExC_state, const U8 cp, SV** invlist)
17217 /* This adds the Latin1/above-Latin1 folding rules.
17219 * This should be called only for a Latin1-range code points, cp, which is
17220 * known to be involved in a simple fold with other code points above
17221 * Latin1. It would give false results if /aa has been specified.
17222 * Multi-char folds are outside the scope of this, and must be handled
17225 PERL_ARGS_ASSERT_ADD_ABOVE_LATIN1_FOLDS;
17227 assert(HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(cp));
17229 /* The rules that are valid for all Unicode versions are hard-coded in */
17234 add_cp_to_invlist(*invlist, KELVIN_SIGN);
17238 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_LONG_S);
17241 *invlist = add_cp_to_invlist(*invlist, GREEK_CAPITAL_LETTER_MU);
17242 *invlist = add_cp_to_invlist(*invlist, GREEK_SMALL_LETTER_MU);
17244 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
17245 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
17246 *invlist = add_cp_to_invlist(*invlist, ANGSTROM_SIGN);
17248 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
17249 *invlist = add_cp_to_invlist(*invlist,
17250 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
17253 default: /* Other code points are checked against the data for the
17254 current Unicode version */
17256 Size_t folds_count;
17258 const U32 * remaining_folds;
17262 folded_cp = toFOLD(cp);
17265 U8 dummy_fold[UTF8_MAXBYTES_CASE+1];
17267 folded_cp = _to_fold_latin1(cp, dummy_fold, &dummy_len, 0);
17270 if (folded_cp > 255) {
17271 *invlist = add_cp_to_invlist(*invlist, folded_cp);
17274 folds_count = _inverse_folds(folded_cp, &first_fold,
17276 if (folds_count == 0) {
17278 /* Use deprecated warning to increase the chances of this being
17280 ckWARN2reg_d(RExC_parse,
17281 "Perl folding rules are not up-to-date for 0x%02X;"
17282 " please use the perlbug utility to report;", cp);
17287 if (first_fold > 255) {
17288 *invlist = add_cp_to_invlist(*invlist, first_fold);
17290 for (i = 0; i < folds_count - 1; i++) {
17291 if (remaining_folds[i] > 255) {
17292 *invlist = add_cp_to_invlist(*invlist,
17293 remaining_folds[i]);
17303 S_output_posix_warnings(pTHX_ RExC_state_t *pRExC_state, AV* posix_warnings)
17305 /* Output the elements of the array given by '*posix_warnings' as REGEXP
17309 const bool first_is_fatal = ckDEAD(packWARN(WARN_REGEXP));
17311 PERL_ARGS_ASSERT_OUTPUT_POSIX_WARNINGS;
17313 if (! TO_OUTPUT_WARNINGS(RExC_parse)) {
17314 CLEAR_POSIX_WARNINGS();
17318 while ((msg = av_shift(posix_warnings)) != &PL_sv_undef) {
17319 if (first_is_fatal) { /* Avoid leaking this */
17320 av_undef(posix_warnings); /* This isn't necessary if the
17321 array is mortal, but is a
17323 (void) sv_2mortal(msg);
17326 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s", SvPVX(msg));
17327 SvREFCNT_dec_NN(msg);
17330 UPDATE_WARNINGS_LOC(RExC_parse);
17333 PERL_STATIC_INLINE Size_t
17334 S_find_first_differing_byte_pos(const U8 * s1, const U8 * s2, const Size_t max)
17336 const U8 * const start = s1;
17337 const U8 * const send = start + max;
17339 PERL_ARGS_ASSERT_FIND_FIRST_DIFFERING_BYTE_POS;
17341 while (s1 < send && *s1 == *s2) {
17350 S_add_multi_match(pTHX_ AV* multi_char_matches, SV* multi_string, const STRLEN cp_count)
17352 /* This adds the string scalar <multi_string> to the array
17353 * <multi_char_matches>. <multi_string> is known to have exactly
17354 * <cp_count> code points in it. This is used when constructing a
17355 * bracketed character class and we find something that needs to match more
17356 * than a single character.
17358 * <multi_char_matches> is actually an array of arrays. Each top-level
17359 * element is an array that contains all the strings known so far that are
17360 * the same length. And that length (in number of code points) is the same
17361 * as the index of the top-level array. Hence, the [2] element is an
17362 * array, each element thereof is a string containing TWO code points;
17363 * while element [3] is for strings of THREE characters, and so on. Since
17364 * this is for multi-char strings there can never be a [0] nor [1] element.
17366 * When we rewrite the character class below, we will do so such that the
17367 * longest strings are written first, so that it prefers the longest
17368 * matching strings first. This is done even if it turns out that any
17369 * quantifier is non-greedy, out of this programmer's (khw) laziness. Tom
17370 * Christiansen has agreed that this is ok. This makes the test for the
17371 * ligature 'ffi' come before the test for 'ff', for example */
17374 AV** this_array_ptr;
17376 PERL_ARGS_ASSERT_ADD_MULTI_MATCH;
17378 if (! multi_char_matches) {
17379 multi_char_matches = newAV();
17382 if (av_exists(multi_char_matches, cp_count)) {
17383 this_array_ptr = (AV**) av_fetch(multi_char_matches, cp_count, FALSE);
17384 this_array = *this_array_ptr;
17387 this_array = newAV();
17388 av_store(multi_char_matches, cp_count,
17391 av_push(this_array, multi_string);
17393 return multi_char_matches;
17396 /* The names of properties whose definitions are not known at compile time are
17397 * stored in this SV, after a constant heading. So if the length has been
17398 * changed since initialization, then there is a run-time definition. */
17399 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
17400 (SvCUR(listsv) != initial_listsv_len)
17402 /* There is a restricted set of white space characters that are legal when
17403 * ignoring white space in a bracketed character class. This generates the
17404 * code to skip them.
17406 * There is a line below that uses the same white space criteria but is outside
17407 * this macro. Both here and there must use the same definition */
17408 #define SKIP_BRACKETED_WHITE_SPACE(do_skip, p, stop_p) \
17411 while (p < stop_p && isBLANK_A(UCHARAT(p))) \
17418 STATIC regnode_offset
17419 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
17420 const bool stop_at_1, /* Just parse the next thing, don't
17421 look for a full character class */
17422 bool allow_mutiple_chars,
17423 const bool silence_non_portable, /* Don't output warnings
17427 bool optimizable, /* ? Allow a non-ANYOF return
17429 SV** ret_invlist /* Return an inversion list, not a node */
17432 /* parse a bracketed class specification. Most of these will produce an
17433 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
17434 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
17435 * under /i with multi-character folds: it will be rewritten following the
17436 * paradigm of this example, where the <multi-fold>s are characters which
17437 * fold to multiple character sequences:
17438 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
17439 * gets effectively rewritten as:
17440 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
17441 * reg() gets called (recursively) on the rewritten version, and this
17442 * function will return what it constructs. (Actually the <multi-fold>s
17443 * aren't physically removed from the [abcdefghi], it's just that they are
17444 * ignored in the recursion by means of a flag:
17445 * <RExC_in_multi_char_class>.)
17447 * ANYOF nodes contain a bit map for the first NUM_ANYOF_CODE_POINTS
17448 * characters, with the corresponding bit set if that character is in the
17449 * list. For characters above this, an inversion list is used. There
17450 * are extra bits for \w, etc. in locale ANYOFs, as what these match is not
17451 * determinable at compile time
17453 * On success, returns the offset at which any next node should be placed
17454 * into the regex engine program being compiled.
17456 * Returns 0 otherwise, setting flagp to RESTART_PARSE if the parse needs
17457 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to
17461 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
17463 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
17464 regnode_offset ret = -1; /* Initialized to an illegal value */
17466 int namedclass = OOB_NAMEDCLASS;
17467 char *rangebegin = NULL;
17468 SV *listsv = NULL; /* List of \p{user-defined} whose definitions
17469 aren't available at the time this was called */
17470 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
17471 than just initialized. */
17472 SV* properties = NULL; /* Code points that match \p{} \P{} */
17473 SV* posixes = NULL; /* Code points that match classes like [:word:],
17474 extended beyond the Latin1 range. These have to
17475 be kept separate from other code points for much
17476 of this function because their handling is
17477 different under /i, and for most classes under
17479 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
17480 separate for a while from the non-complemented
17481 versions because of complications with /d
17483 SV* simple_posixes = NULL; /* But under some conditions, the classes can be
17484 treated more simply than the general case,
17485 leading to less compilation and execution
17487 UV element_count = 0; /* Number of distinct elements in the class.
17488 Optimizations may be possible if this is tiny */
17489 AV * multi_char_matches = NULL; /* Code points that fold to more than one
17490 character; used under /i */
17492 char * stop_ptr = RExC_end; /* where to stop parsing */
17494 /* ignore unescaped whitespace? */
17495 const bool skip_white = cBOOL( ret_invlist
17496 || (RExC_flags & RXf_PMf_EXTENDED_MORE));
17498 /* inversion list of code points this node matches only when the target
17499 * string is in UTF-8. These are all non-ASCII, < 256. (Because is under
17501 SV* upper_latin1_only_utf8_matches = NULL;
17503 /* Inversion list of code points this node matches regardless of things
17504 * like locale, folding, utf8ness of the target string */
17505 SV* cp_list = NULL;
17507 /* Like cp_list, but code points on this list need to be checked for things
17508 * that fold to/from them under /i */
17509 SV* cp_foldable_list = NULL;
17511 /* Like cp_list, but code points on this list are valid only when the
17512 * runtime locale is UTF-8 */
17513 SV* only_utf8_locale_list = NULL;
17515 /* In a range, if one of the endpoints is non-character-set portable,
17516 * meaning that it hard-codes a code point that may mean a different
17517 * charactger in ASCII vs. EBCDIC, as opposed to, say, a literal 'A' or a
17518 * mnemonic '\t' which each mean the same character no matter which
17519 * character set the platform is on. */
17520 unsigned int non_portable_endpoint = 0;
17522 /* Is the range unicode? which means on a platform that isn't 1-1 native
17523 * to Unicode (i.e. non-ASCII), each code point in it should be considered
17524 * to be a Unicode value. */
17525 bool unicode_range = FALSE;
17526 bool invert = FALSE; /* Is this class to be complemented */
17528 bool warn_super = ALWAYS_WARN_SUPER;
17530 const char * orig_parse = RExC_parse;
17532 /* This variable is used to mark where the end in the input is of something
17533 * that looks like a POSIX construct but isn't. During the parse, when
17534 * something looks like it could be such a construct is encountered, it is
17535 * checked for being one, but not if we've already checked this area of the
17536 * input. Only after this position is reached do we check again */
17537 char *not_posix_region_end = RExC_parse - 1;
17539 AV* posix_warnings = NULL;
17540 const bool do_posix_warnings = ckWARN(WARN_REGEXP);
17541 U8 op = ANYOF; /* The returned node-type, initialized the expected type.
17543 U8 anyof_flags = 0; /* flag bits if the node is an ANYOF-type */
17544 U32 posixl = 0; /* bit field of posix classes matched under /l */
17547 /* Flags as to what things aren't knowable until runtime. (Note that these are
17548 * mutually exclusive.) */
17549 #define HAS_USER_DEFINED_PROPERTY 0x01 /* /u any user-defined properties that
17550 haven't been defined as of yet */
17551 #define HAS_D_RUNTIME_DEPENDENCY 0x02 /* /d if the target being matched is
17553 #define HAS_L_RUNTIME_DEPENDENCY 0x04 /* /l what the posix classes match and
17554 what gets folded */
17555 U32 has_runtime_dependency = 0; /* OR of the above flags */
17557 DECLARE_AND_GET_RE_DEBUG_FLAGS;
17559 PERL_ARGS_ASSERT_REGCLASS;
17561 PERL_UNUSED_ARG(depth);
17564 assert(! (ret_invlist && allow_mutiple_chars));
17566 /* If wants an inversion list returned, we can't optimize to something
17569 optimizable = FALSE;
17572 DEBUG_PARSE("clas");
17574 #if UNICODE_MAJOR_VERSION < 3 /* no multifolds in early Unicode */ \
17575 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0 \
17576 && UNICODE_DOT_DOT_VERSION == 0)
17577 allow_mutiple_chars = FALSE;
17580 /* We include the /i status at the beginning of this so that we can
17581 * know it at runtime */
17582 listsv = sv_2mortal(Perl_newSVpvf(aTHX_ "#%d\n", cBOOL(FOLD)));
17583 initial_listsv_len = SvCUR(listsv);
17584 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
17586 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse, RExC_end);
17588 assert(RExC_parse <= RExC_end);
17590 if (UCHARAT(RExC_parse) == '^') { /* Complement the class */
17593 allow_mutiple_chars = FALSE;
17595 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse, RExC_end);
17598 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
17599 if (! ret_invlist && MAYBE_POSIXCC(UCHARAT(RExC_parse))) {
17600 int maybe_class = handle_possible_posix(pRExC_state,
17602 ¬_posix_region_end,
17604 TRUE /* checking only */);
17605 if (maybe_class >= OOB_NAMEDCLASS && do_posix_warnings) {
17606 ckWARN4reg(not_posix_region_end,
17607 "POSIX syntax [%c %c] belongs inside character classes%s",
17608 *RExC_parse, *RExC_parse,
17609 (maybe_class == OOB_NAMEDCLASS)
17610 ? ((POSIXCC_NOTYET(*RExC_parse))
17611 ? " (but this one isn't implemented)"
17612 : " (but this one isn't fully valid)")
17618 /* If the caller wants us to just parse a single element, accomplish this
17619 * by faking the loop ending condition */
17620 if (stop_at_1 && RExC_end > RExC_parse) {
17621 stop_ptr = RExC_parse + 1;
17624 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
17625 if (UCHARAT(RExC_parse) == ']')
17626 goto charclassloop;
17630 if ( posix_warnings
17631 && av_tindex_skip_len_mg(posix_warnings) >= 0
17632 && RExC_parse > not_posix_region_end)
17634 /* Warnings about posix class issues are considered tentative until
17635 * we are far enough along in the parse that we can no longer
17636 * change our mind, at which point we output them. This is done
17637 * each time through the loop so that a later class won't zap them
17638 * before they have been dealt with. */
17639 output_posix_warnings(pRExC_state, posix_warnings);
17642 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse, RExC_end);
17644 if (RExC_parse >= stop_ptr) {
17648 if (UCHARAT(RExC_parse) == ']') {
17654 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
17655 save_value = value;
17656 save_prevvalue = prevvalue;
17659 rangebegin = RExC_parse;
17661 non_portable_endpoint = 0;
17663 if (UTF && ! UTF8_IS_INVARIANT(* RExC_parse)) {
17664 value = utf8n_to_uvchr((U8*)RExC_parse,
17665 RExC_end - RExC_parse,
17666 &numlen, UTF8_ALLOW_DEFAULT);
17667 RExC_parse += numlen;
17670 value = UCHARAT(RExC_parse++);
17672 if (value == '[') {
17673 char * posix_class_end;
17674 namedclass = handle_possible_posix(pRExC_state,
17677 do_posix_warnings ? &posix_warnings : NULL,
17678 FALSE /* die if error */);
17679 if (namedclass > OOB_NAMEDCLASS) {
17681 /* If there was an earlier attempt to parse this particular
17682 * posix class, and it failed, it was a false alarm, as this
17683 * successful one proves */
17684 if ( posix_warnings
17685 && av_tindex_skip_len_mg(posix_warnings) >= 0
17686 && not_posix_region_end >= RExC_parse
17687 && not_posix_region_end <= posix_class_end)
17689 av_undef(posix_warnings);
17692 RExC_parse = posix_class_end;
17694 else if (namedclass == OOB_NAMEDCLASS) {
17695 not_posix_region_end = posix_class_end;
17698 namedclass = OOB_NAMEDCLASS;
17701 else if ( RExC_parse - 1 > not_posix_region_end
17702 && MAYBE_POSIXCC(value))
17704 (void) handle_possible_posix(
17706 RExC_parse - 1, /* -1 because parse has already been
17708 ¬_posix_region_end,
17709 do_posix_warnings ? &posix_warnings : NULL,
17710 TRUE /* checking only */);
17712 else if ( strict && ! skip_white
17713 && ( _generic_isCC(value, _CC_VERTSPACE)
17714 || is_VERTWS_cp_high(value)))
17716 vFAIL("Literal vertical space in [] is illegal except under /x");
17718 else if (value == '\\') {
17719 /* Is a backslash; get the code point of the char after it */
17721 if (RExC_parse >= RExC_end) {
17722 vFAIL("Unmatched [");
17725 if (UTF && ! UTF8_IS_INVARIANT(UCHARAT(RExC_parse))) {
17726 value = utf8n_to_uvchr((U8*)RExC_parse,
17727 RExC_end - RExC_parse,
17728 &numlen, UTF8_ALLOW_DEFAULT);
17729 RExC_parse += numlen;
17732 value = UCHARAT(RExC_parse++);
17734 /* Some compilers cannot handle switching on 64-bit integer
17735 * values, therefore value cannot be an UV. Yes, this will
17736 * be a problem later if we want switch on Unicode.
17737 * A similar issue a little bit later when switching on
17738 * namedclass. --jhi */
17740 /* If the \ is escaping white space when white space is being
17741 * skipped, it means that that white space is wanted literally, and
17742 * is already in 'value'. Otherwise, need to translate the escape
17743 * into what it signifies. */
17744 if (! skip_white || ! isBLANK_A(value)) switch ((I32)value) {
17745 const char * message;
17749 case 'w': namedclass = ANYOF_WORDCHAR; break;
17750 case 'W': namedclass = ANYOF_NWORDCHAR; break;
17751 case 's': namedclass = ANYOF_SPACE; break;
17752 case 'S': namedclass = ANYOF_NSPACE; break;
17753 case 'd': namedclass = ANYOF_DIGIT; break;
17754 case 'D': namedclass = ANYOF_NDIGIT; break;
17755 case 'v': namedclass = ANYOF_VERTWS; break;
17756 case 'V': namedclass = ANYOF_NVERTWS; break;
17757 case 'h': namedclass = ANYOF_HORIZWS; break;
17758 case 'H': namedclass = ANYOF_NHORIZWS; break;
17759 case 'N': /* Handle \N{NAME} in class */
17761 const char * const backslash_N_beg = RExC_parse - 2;
17764 if (! grok_bslash_N(pRExC_state,
17765 NULL, /* No regnode */
17766 &value, /* Yes single value */
17767 &cp_count, /* Multiple code pt count */
17773 if (*flagp & NEED_UTF8)
17774 FAIL("panic: grok_bslash_N set NEED_UTF8");
17776 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
17778 if (cp_count < 0) {
17779 vFAIL("\\N in a character class must be a named character: \\N{...}");
17781 else if (cp_count == 0) {
17782 ckWARNreg(RExC_parse,
17783 "Ignoring zero length \\N{} in character class");
17785 else { /* cp_count > 1 */
17786 assert(cp_count > 1);
17787 if (! RExC_in_multi_char_class) {
17788 if ( ! allow_mutiple_chars
17791 || *RExC_parse == '-')
17795 vFAIL("\\N{} here is restricted to one character");
17797 ckWARNreg(RExC_parse, "Using just the first character returned by \\N{} in character class");
17798 break; /* <value> contains the first code
17799 point. Drop out of the switch to
17803 SV * multi_char_N = newSVpvn(backslash_N_beg,
17804 RExC_parse - backslash_N_beg);
17806 = add_multi_match(multi_char_matches,
17811 } /* End of cp_count != 1 */
17813 /* This element should not be processed further in this
17816 value = save_value;
17817 prevvalue = save_prevvalue;
17818 continue; /* Back to top of loop to get next char */
17821 /* Here, is a single code point, and <value> contains it */
17822 unicode_range = TRUE; /* \N{} are Unicode */
17830 if (RExC_pm_flags & PMf_WILDCARD) {
17832 /* diag_listed_as: Use of %s is not allowed in Unicode
17833 property wildcard subpatterns in regex; marked by <--
17835 vFAIL3("Use of '\\%c%c' is not allowed in Unicode property"
17836 " wildcard subpatterns", (char) value, *(RExC_parse - 1));
17839 /* \p means they want Unicode semantics */
17840 REQUIRE_UNI_RULES(flagp, 0);
17842 if (RExC_parse >= RExC_end)
17843 vFAIL2("Empty \\%c", (U8)value);
17844 if (*RExC_parse == '{') {
17845 const U8 c = (U8)value;
17846 e = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
17849 vFAIL2("Missing right brace on \\%c{}", c);
17854 /* White space is allowed adjacent to the braces and after
17855 * any '^', even when not under /x */
17856 while (isSPACE(*RExC_parse)) {
17860 if (UCHARAT(RExC_parse) == '^') {
17862 /* toggle. (The rhs xor gets the single bit that
17863 * differs between P and p; the other xor inverts just
17865 value ^= 'P' ^ 'p';
17868 while (isSPACE(*RExC_parse)) {
17873 if (e == RExC_parse)
17874 vFAIL2("Empty \\%c{}", c);
17876 n = e - RExC_parse;
17877 while (isSPACE(*(RExC_parse + n - 1)))
17880 } /* The \p isn't immediately followed by a '{' */
17881 else if (! isALPHA(*RExC_parse)) {
17882 RExC_parse += (UTF)
17883 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
17885 vFAIL2("Character following \\%c must be '{' or a "
17886 "single-character Unicode property name",
17894 char* name = RExC_parse;
17896 /* Any message returned about expanding the definition */
17897 SV* msg = newSVpvs_flags("", SVs_TEMP);
17899 /* If set TRUE, the property is user-defined as opposed to
17900 * official Unicode */
17901 bool user_defined = FALSE;
17902 AV * strings = NULL;
17904 SV * prop_definition = parse_uniprop_string(
17905 name, n, UTF, FOLD,
17906 FALSE, /* This is compile-time */
17908 /* We can't defer this defn when
17909 * the full result is required in
17911 ! cBOOL(ret_invlist),
17918 if (SvCUR(msg)) { /* Assumes any error causes a msg */
17919 assert(prop_definition == NULL);
17920 RExC_parse = e + 1;
17921 if (SvUTF8(msg)) { /* msg being UTF-8 makes the whole
17922 thing so, or else the display is
17926 /* diag_listed_as: Can't find Unicode property definition "%s" in regex; marked by <-- HERE in m/%s/ */
17927 vFAIL2utf8f("%" UTF8f, UTF8fARG(SvUTF8(msg),
17928 SvCUR(msg), SvPVX(msg)));
17931 assert(prop_definition || strings);
17935 if (! prop_definition) {
17936 RExC_parse = e + 1;
17937 vFAIL("Unicode string properties are not implemented in (?[...])");
17941 "Using just the single character results"
17942 " returned by \\p{} in (?[...])");
17945 else if (! RExC_in_multi_char_class) {
17946 if (invert ^ (value == 'P')) {
17947 RExC_parse = e + 1;
17948 vFAIL("Inverting a character class which contains"
17949 " a multi-character sequence is illegal");
17952 /* For each multi-character string ... */
17953 while (av_count(strings) > 0) {
17954 /* ... Each entry is itself an array of code
17956 AV * this_string = (AV *) av_shift( strings);
17957 STRLEN cp_count = av_count(this_string);
17958 SV * final = newSV(cp_count * 4);
17961 /* Create another string of sequences of \x{...} */
17962 while (av_count(this_string) > 0) {
17963 SV * character = av_shift(this_string);
17964 UV cp = SvUV(character);
17967 REQUIRE_UTF8(flagp);
17969 Perl_sv_catpvf(aTHX_ final, "\\x{%" UVXf "}",
17971 SvREFCNT_dec_NN(character);
17973 SvREFCNT_dec_NN(this_string);
17975 /* And add that to the list of such things */
17977 = add_multi_match(multi_char_matches,
17982 SvREFCNT_dec_NN(strings);
17985 if (! prop_definition) { /* If we got only a string,
17986 this iteration didn't really
17987 find a character */
17990 else if (! is_invlist(prop_definition)) {
17992 /* Here, the definition isn't known, so we have gotten
17993 * returned a string that will be evaluated if and when
17994 * encountered at runtime. We add it to the list of
17995 * such properties, along with whether it should be
17996 * complemented or not */
17997 if (value == 'P') {
17998 sv_catpvs(listsv, "!");
18001 sv_catpvs(listsv, "+");
18003 sv_catsv(listsv, prop_definition);
18005 has_runtime_dependency |= HAS_USER_DEFINED_PROPERTY;
18007 /* We don't know yet what this matches, so have to flag
18009 anyof_flags |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
18012 assert (prop_definition && is_invlist(prop_definition));
18014 /* Here we do have the complete property definition
18016 * Temporary workaround for [perl #133136]. For this
18017 * precise input that is in the .t that is failing,
18018 * load utf8.pm, which is what the test wants, so that
18019 * that .t passes */
18020 if ( memEQs(RExC_start, e + 1 - RExC_start,
18022 && ! hv_common(GvHVn(PL_incgv),
18024 "utf8.pm", sizeof("utf8.pm") - 1,
18025 0, HV_FETCH_ISEXISTS, NULL, 0))
18027 require_pv("utf8.pm");
18030 if (! user_defined &&
18031 /* We warn on matching an above-Unicode code point
18032 * if the match would return true, except don't
18033 * warn for \p{All}, which has exactly one element
18035 (_invlist_contains_cp(prop_definition, 0x110000)
18036 && (! (_invlist_len(prop_definition) == 1
18037 && *invlist_array(prop_definition) == 0))))
18042 /* Invert if asking for the complement */
18043 if (value == 'P') {
18044 _invlist_union_complement_2nd(properties,
18049 _invlist_union(properties, prop_definition, &properties);
18054 RExC_parse = e + 1;
18055 namedclass = ANYOF_UNIPROP; /* no official name, but it's
18059 case 'n': value = '\n'; break;
18060 case 'r': value = '\r'; break;
18061 case 't': value = '\t'; break;
18062 case 'f': value = '\f'; break;
18063 case 'b': value = '\b'; break;
18064 case 'e': value = ESC_NATIVE; break;
18065 case 'a': value = '\a'; break;
18067 RExC_parse--; /* function expects to be pointed at the 'o' */
18068 if (! grok_bslash_o(&RExC_parse,
18074 cBOOL(range), /* MAX_UV allowed for range
18080 else if (message && TO_OUTPUT_WARNINGS(RExC_parse)) {
18081 warn_non_literal_string(RExC_parse, packed_warn, message);
18085 non_portable_endpoint++;
18089 RExC_parse--; /* function expects to be pointed at the 'x' */
18090 if (! grok_bslash_x(&RExC_parse,
18096 cBOOL(range), /* MAX_UV allowed for range
18102 else if (message && TO_OUTPUT_WARNINGS(RExC_parse)) {
18103 warn_non_literal_string(RExC_parse, packed_warn, message);
18107 non_portable_endpoint++;
18111 if (! grok_bslash_c(*RExC_parse, &grok_c_char, &message,
18114 /* going to die anyway; point to exact spot of
18116 RExC_parse += (UTF)
18117 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
18122 value = grok_c_char;
18124 if (message && TO_OUTPUT_WARNINGS(RExC_parse)) {
18125 warn_non_literal_string(RExC_parse, packed_warn, message);
18128 non_portable_endpoint++;
18130 case '0': case '1': case '2': case '3': case '4':
18131 case '5': case '6': case '7':
18133 /* Take 1-3 octal digits */
18134 I32 flags = PERL_SCAN_SILENT_ILLDIGIT
18135 | PERL_SCAN_NOTIFY_ILLDIGIT;
18136 numlen = (strict) ? 4 : 3;
18137 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
18138 RExC_parse += numlen;
18141 RExC_parse += (UTF)
18142 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
18144 vFAIL("Need exactly 3 octal digits");
18146 else if ( (flags & PERL_SCAN_NOTIFY_ILLDIGIT)
18147 && RExC_parse < RExC_end
18148 && isDIGIT(*RExC_parse)
18149 && ckWARN(WARN_REGEXP))
18151 reg_warn_non_literal_string(
18153 form_alien_digit_msg(8, numlen, RExC_parse,
18154 RExC_end, UTF, FALSE));
18158 non_portable_endpoint++;
18163 /* Allow \_ to not give an error */
18164 if (isWORDCHAR(value) && value != '_') {
18166 vFAIL2("Unrecognized escape \\%c in character class",
18170 ckWARN2reg(RExC_parse,
18171 "Unrecognized escape \\%c in character class passed through",
18176 } /* End of switch on char following backslash */
18177 } /* end of handling backslash escape sequences */
18179 /* Here, we have the current token in 'value' */
18181 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
18184 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
18185 * literal, as is the character that began the false range, i.e.
18186 * the 'a' in the examples */
18188 const int w = (RExC_parse >= rangebegin)
18189 ? RExC_parse - rangebegin
18193 "False [] range \"%" UTF8f "\"",
18194 UTF8fARG(UTF, w, rangebegin));
18197 ckWARN2reg(RExC_parse,
18198 "False [] range \"%" UTF8f "\"",
18199 UTF8fARG(UTF, w, rangebegin));
18200 cp_list = add_cp_to_invlist(cp_list, '-');
18201 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
18205 range = 0; /* this was not a true range */
18206 element_count += 2; /* So counts for three values */
18209 classnum = namedclass_to_classnum(namedclass);
18211 if (LOC && namedclass < ANYOF_POSIXL_MAX
18212 #ifndef HAS_ISASCII
18213 && classnum != _CC_ASCII
18216 SV* scratch_list = NULL;
18218 /* What the Posix classes (like \w, [:space:]) match isn't
18219 * generally knowable under locale until actual match time. A
18220 * special node is used for these which has extra space for a
18221 * bitmap, with a bit reserved for each named class that is to
18222 * be matched against. (This isn't needed for \p{} and
18223 * pseudo-classes, as they are not affected by locale, and
18224 * hence are dealt with separately.) However, if a named class
18225 * and its complement are both present, then it matches
18226 * everything, and there is no runtime dependency. Odd numbers
18227 * are the complements of the next lower number, so xor works.
18228 * (Note that something like [\w\D] should match everything,
18229 * because \d should be a proper subset of \w. But rather than
18230 * trust that the locale is well behaved, we leave this to
18231 * runtime to sort out) */
18232 if (POSIXL_TEST(posixl, namedclass ^ 1)) {
18233 cp_list = _add_range_to_invlist(cp_list, 0, UV_MAX);
18234 POSIXL_ZERO(posixl);
18235 has_runtime_dependency &= ~HAS_L_RUNTIME_DEPENDENCY;
18236 anyof_flags &= ~ANYOF_MATCHES_POSIXL;
18237 continue; /* We could ignore the rest of the class, but
18238 best to parse it for any errors */
18240 else { /* Here, isn't the complement of any already parsed
18242 POSIXL_SET(posixl, namedclass);
18243 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
18244 anyof_flags |= ANYOF_MATCHES_POSIXL;
18246 /* The above-Latin1 characters are not subject to locale
18247 * rules. Just add them to the unconditionally-matched
18250 /* Get the list of the above-Latin1 code points this
18252 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
18253 PL_XPosix_ptrs[classnum],
18255 /* Odd numbers are complements,
18256 * like NDIGIT, NASCII, ... */
18257 namedclass % 2 != 0,
18259 /* Checking if 'cp_list' is NULL first saves an extra
18260 * clone. Its reference count will be decremented at the
18261 * next union, etc, or if this is the only instance, at the
18262 * end of the routine */
18264 cp_list = scratch_list;
18267 _invlist_union(cp_list, scratch_list, &cp_list);
18268 SvREFCNT_dec_NN(scratch_list);
18270 continue; /* Go get next character */
18275 /* Here, is not /l, or is a POSIX class for which /l doesn't
18276 * matter (or is a Unicode property, which is skipped here). */
18277 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
18278 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
18280 /* Here, should be \h, \H, \v, or \V. None of /d, /i
18281 * nor /l make a difference in what these match,
18282 * therefore we just add what they match to cp_list. */
18283 if (classnum != _CC_VERTSPACE) {
18284 assert( namedclass == ANYOF_HORIZWS
18285 || namedclass == ANYOF_NHORIZWS);
18287 /* It turns out that \h is just a synonym for
18289 classnum = _CC_BLANK;
18292 _invlist_union_maybe_complement_2nd(
18294 PL_XPosix_ptrs[classnum],
18295 namedclass % 2 != 0, /* Complement if odd
18296 (NHORIZWS, NVERTWS)
18301 else if ( AT_LEAST_UNI_SEMANTICS
18302 || classnum == _CC_ASCII
18303 || (DEPENDS_SEMANTICS && ( classnum == _CC_DIGIT
18304 || classnum == _CC_XDIGIT)))
18306 /* We usually have to worry about /d affecting what POSIX
18307 * classes match, with special code needed because we won't
18308 * know until runtime what all matches. But there is no
18309 * extra work needed under /u and /a; and [:ascii:] is
18310 * unaffected by /d; and :digit: and :xdigit: don't have
18311 * runtime differences under /d. So we can special case
18312 * these, and avoid some extra work below, and at runtime.
18314 _invlist_union_maybe_complement_2nd(
18316 ((AT_LEAST_ASCII_RESTRICTED)
18317 ? PL_Posix_ptrs[classnum]
18318 : PL_XPosix_ptrs[classnum]),
18319 namedclass % 2 != 0,
18322 else { /* Garden variety class. If is NUPPER, NALPHA, ...
18323 complement and use nposixes */
18324 SV** posixes_ptr = namedclass % 2 == 0
18327 _invlist_union_maybe_complement_2nd(
18329 PL_XPosix_ptrs[classnum],
18330 namedclass % 2 != 0,
18334 } /* end of namedclass \blah */
18336 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse, RExC_end);
18338 /* If 'range' is set, 'value' is the ending of a range--check its
18339 * validity. (If value isn't a single code point in the case of a
18340 * range, we should have figured that out above in the code that
18341 * catches false ranges). Later, we will handle each individual code
18342 * point in the range. If 'range' isn't set, this could be the
18343 * beginning of a range, so check for that by looking ahead to see if
18344 * the next real character to be processed is the range indicator--the
18349 /* For unicode ranges, we have to test that the Unicode as opposed
18350 * to the native values are not decreasing. (Above 255, there is
18351 * no difference between native and Unicode) */
18352 if (unicode_range && prevvalue < 255 && value < 255) {
18353 if (NATIVE_TO_LATIN1(prevvalue) > NATIVE_TO_LATIN1(value)) {
18354 goto backwards_range;
18359 if (prevvalue > value) /* b-a */ {
18364 w = RExC_parse - rangebegin;
18366 "Invalid [] range \"%" UTF8f "\"",
18367 UTF8fARG(UTF, w, rangebegin));
18368 NOT_REACHED; /* NOTREACHED */
18372 prevvalue = value; /* save the beginning of the potential range */
18373 if (! stop_at_1 /* Can't be a range if parsing just one thing */
18374 && *RExC_parse == '-')
18376 char* next_char_ptr = RExC_parse + 1;
18378 /* Get the next real char after the '-' */
18379 SKIP_BRACKETED_WHITE_SPACE(skip_white, next_char_ptr, RExC_end);
18381 /* If the '-' is at the end of the class (just before the ']',
18382 * it is a literal minus; otherwise it is a range */
18383 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
18384 RExC_parse = next_char_ptr;
18386 /* a bad range like \w-, [:word:]- ? */
18387 if (namedclass > OOB_NAMEDCLASS) {
18388 if (strict || ckWARN(WARN_REGEXP)) {
18389 const int w = RExC_parse >= rangebegin
18390 ? RExC_parse - rangebegin
18393 vFAIL4("False [] range \"%*.*s\"",
18398 "False [] range \"%*.*s\"",
18402 cp_list = add_cp_to_invlist(cp_list, '-');
18405 range = 1; /* yeah, it's a range! */
18406 continue; /* but do it the next time */
18411 if (namedclass > OOB_NAMEDCLASS) {
18415 /* Here, we have a single value this time through the loop, and
18416 * <prevvalue> is the beginning of the range, if any; or <value> if
18419 /* non-Latin1 code point implies unicode semantics. */
18421 if (value > MAX_LEGAL_CP && ( value != UV_MAX
18422 || prevvalue > MAX_LEGAL_CP))
18424 vFAIL(form_cp_too_large_msg(16, NULL, 0, value));
18426 REQUIRE_UNI_RULES(flagp, 0);
18427 if ( ! silence_non_portable
18428 && UNICODE_IS_PERL_EXTENDED(value)
18429 && TO_OUTPUT_WARNINGS(RExC_parse))
18431 ckWARN2_non_literal_string(RExC_parse,
18432 packWARN(WARN_PORTABLE),
18433 PL_extended_cp_format,
18438 /* Ready to process either the single value, or the completed range.
18439 * For single-valued non-inverted ranges, we consider the possibility
18440 * of multi-char folds. (We made a conscious decision to not do this
18441 * for the other cases because it can often lead to non-intuitive
18442 * results. For example, you have the peculiar case that:
18443 * "s s" =~ /^[^\xDF]+$/i => Y
18444 * "ss" =~ /^[^\xDF]+$/i => N
18446 * See [perl #89750] */
18447 if (FOLD && allow_mutiple_chars && value == prevvalue) {
18448 if ( value == LATIN_SMALL_LETTER_SHARP_S
18449 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
18452 /* Here <value> is indeed a multi-char fold. Get what it is */
18454 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
18457 UV folded = _to_uni_fold_flags(
18461 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
18462 ? FOLD_FLAGS_NOMIX_ASCII
18466 /* Here, <folded> should be the first character of the
18467 * multi-char fold of <value>, with <foldbuf> containing the
18468 * whole thing. But, if this fold is not allowed (because of
18469 * the flags), <fold> will be the same as <value>, and should
18470 * be processed like any other character, so skip the special
18472 if (folded != value) {
18474 /* Skip if we are recursed, currently parsing the class
18475 * again. Otherwise add this character to the list of
18476 * multi-char folds. */
18477 if (! RExC_in_multi_char_class) {
18478 STRLEN cp_count = utf8_length(foldbuf,
18479 foldbuf + foldlen);
18480 SV* multi_fold = sv_2mortal(newSVpvs(""));
18482 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%" UVXf "}", value);
18485 = add_multi_match(multi_char_matches,
18491 /* This element should not be processed further in this
18494 value = save_value;
18495 prevvalue = save_prevvalue;
18501 if (strict && ckWARN(WARN_REGEXP)) {
18504 /* If the range starts above 255, everything is portable and
18505 * likely to be so for any forseeable character set, so don't
18507 if (unicode_range && non_portable_endpoint && prevvalue < 256) {
18508 vWARN(RExC_parse, "Both or neither range ends should be Unicode");
18510 else if (prevvalue != value) {
18512 /* Under strict, ranges that stop and/or end in an ASCII
18513 * printable should have each end point be a portable value
18514 * for it (preferably like 'A', but we don't warn if it is
18515 * a (portable) Unicode name or code point), and the range
18516 * must be all digits or all letters of the same case.
18517 * Otherwise, the range is non-portable and unclear as to
18518 * what it contains */
18519 if ( (isPRINT_A(prevvalue) || isPRINT_A(value))
18520 && ( non_portable_endpoint
18521 || ! ( (isDIGIT_A(prevvalue) && isDIGIT_A(value))
18522 || (isLOWER_A(prevvalue) && isLOWER_A(value))
18523 || (isUPPER_A(prevvalue) && isUPPER_A(value))
18525 vWARN(RExC_parse, "Ranges of ASCII printables should"
18526 " be some subset of \"0-9\","
18527 " \"A-Z\", or \"a-z\"");
18529 else if (prevvalue >= FIRST_NON_ASCII_DECIMAL_DIGIT) {
18530 SSize_t index_start;
18531 SSize_t index_final;
18533 /* But the nature of Unicode and languages mean we
18534 * can't do the same checks for above-ASCII ranges,
18535 * except in the case of digit ones. These should
18536 * contain only digits from the same group of 10. The
18537 * ASCII case is handled just above. Hence here, the
18538 * range could be a range of digits. First some
18539 * unlikely special cases. Grandfather in that a range
18540 * ending in 19DA (NEW TAI LUE THAM DIGIT ONE) is bad
18541 * if its starting value is one of the 10 digits prior
18542 * to it. This is because it is an alternate way of
18543 * writing 19D1, and some people may expect it to be in
18544 * that group. But it is bad, because it won't give
18545 * the expected results. In Unicode 5.2 it was
18546 * considered to be in that group (of 11, hence), but
18547 * this was fixed in the next version */
18549 if (UNLIKELY(value == 0x19DA && prevvalue >= 0x19D0)) {
18550 goto warn_bad_digit_range;
18552 else if (UNLIKELY( prevvalue >= 0x1D7CE
18553 && value <= 0x1D7FF))
18555 /* This is the only other case currently in Unicode
18556 * where the algorithm below fails. The code
18557 * points just above are the end points of a single
18558 * range containing only decimal digits. It is 5
18559 * different series of 0-9. All other ranges of
18560 * digits currently in Unicode are just a single
18561 * series. (And mktables will notify us if a later
18562 * Unicode version breaks this.)
18564 * If the range being checked is at most 9 long,
18565 * and the digit values represented are in
18566 * numerical order, they are from the same series.
18568 if ( value - prevvalue > 9
18569 || ((( value - 0x1D7CE) % 10)
18570 <= (prevvalue - 0x1D7CE) % 10))
18572 goto warn_bad_digit_range;
18577 /* For all other ranges of digits in Unicode, the
18578 * algorithm is just to check if both end points
18579 * are in the same series, which is the same range.
18581 index_start = _invlist_search(
18582 PL_XPosix_ptrs[_CC_DIGIT],
18585 /* Warn if the range starts and ends with a digit,
18586 * and they are not in the same group of 10. */
18587 if ( index_start >= 0
18588 && ELEMENT_RANGE_MATCHES_INVLIST(index_start)
18590 _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
18591 value)) != index_start
18592 && index_final >= 0
18593 && ELEMENT_RANGE_MATCHES_INVLIST(index_final))
18595 warn_bad_digit_range:
18596 vWARN(RExC_parse, "Ranges of digits should be"
18597 " from the same group of"
18604 if ((! range || prevvalue == value) && non_portable_endpoint) {
18605 if (isPRINT_A(value)) {
18608 if (isBACKSLASHED_PUNCT(value)) {
18609 literal[d++] = '\\';
18611 literal[d++] = (char) value;
18612 literal[d++] = '\0';
18615 "\"%.*s\" is more clearly written simply as \"%s\"",
18616 (int) (RExC_parse - rangebegin),
18621 else if (isMNEMONIC_CNTRL(value)) {
18623 "\"%.*s\" is more clearly written simply as \"%s\"",
18624 (int) (RExC_parse - rangebegin),
18626 cntrl_to_mnemonic((U8) value)
18632 /* Deal with this element of the class */
18635 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
18638 /* On non-ASCII platforms, for ranges that span all of 0..255, and ones
18639 * that don't require special handling, we can just add the range like
18640 * we do for ASCII platforms */
18641 if ((UNLIKELY(prevvalue == 0) && value >= 255)
18642 || ! (prevvalue < 256
18644 || (! non_portable_endpoint
18645 && ((isLOWER_A(prevvalue) && isLOWER_A(value))
18646 || (isUPPER_A(prevvalue)
18647 && isUPPER_A(value)))))))
18649 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
18653 /* Here, requires special handling. This can be because it is a
18654 * range whose code points are considered to be Unicode, and so
18655 * must be individually translated into native, or because its a
18656 * subrange of 'A-Z' or 'a-z' which each aren't contiguous in
18657 * EBCDIC, but we have defined them to include only the "expected"
18658 * upper or lower case ASCII alphabetics. Subranges above 255 are
18659 * the same in native and Unicode, so can be added as a range */
18660 U8 start = NATIVE_TO_LATIN1(prevvalue);
18662 U8 end = (value < 256) ? NATIVE_TO_LATIN1(value) : 255;
18663 for (j = start; j <= end; j++) {
18664 cp_foldable_list = add_cp_to_invlist(cp_foldable_list, LATIN1_TO_NATIVE(j));
18667 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
18673 range = 0; /* this range (if it was one) is done now */
18674 } /* End of loop through all the text within the brackets */
18676 if ( posix_warnings && av_tindex_skip_len_mg(posix_warnings) >= 0) {
18677 output_posix_warnings(pRExC_state, posix_warnings);
18680 /* If anything in the class expands to more than one character, we have to
18681 * deal with them by building up a substitute parse string, and recursively
18682 * calling reg() on it, instead of proceeding */
18683 if (multi_char_matches) {
18684 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
18687 char *save_end = RExC_end;
18688 char *save_parse = RExC_parse;
18689 char *save_start = RExC_start;
18690 Size_t constructed_prefix_len = 0; /* This gives the length of the
18691 constructed portion of the
18692 substitute parse. */
18693 bool first_time = TRUE; /* First multi-char occurrence doesn't get
18698 /* Only one level of recursion allowed */
18699 assert(RExC_copy_start_in_constructed == RExC_precomp);
18701 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
18702 because too confusing */
18704 sv_catpvs(substitute_parse, "(?:");
18708 /* Look at the longest strings first */
18709 for (cp_count = av_tindex_skip_len_mg(multi_char_matches);
18714 if (av_exists(multi_char_matches, cp_count)) {
18715 AV** this_array_ptr;
18718 this_array_ptr = (AV**) av_fetch(multi_char_matches,
18720 while ((this_sequence = av_pop(*this_array_ptr)) !=
18723 if (! first_time) {
18724 sv_catpvs(substitute_parse, "|");
18726 first_time = FALSE;
18728 sv_catpv(substitute_parse, SvPVX(this_sequence));
18733 /* If the character class contains anything else besides these
18734 * multi-character strings, have to include it in recursive parsing */
18735 if (element_count) {
18736 bool has_l_bracket = orig_parse > RExC_start && *(orig_parse - 1) == '[';
18738 sv_catpvs(substitute_parse, "|");
18739 if (has_l_bracket) { /* Add an [ if the original had one */
18740 sv_catpvs(substitute_parse, "[");
18742 constructed_prefix_len = SvCUR(substitute_parse);
18743 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
18745 /* Put in a closing ']' to match any opening one, but not if going
18746 * off the end, as otherwise we are adding something that really
18748 if (has_l_bracket && RExC_parse < RExC_end) {
18749 sv_catpvs(substitute_parse, "]");
18753 sv_catpvs(substitute_parse, ")");
18756 /* This is a way to get the parse to skip forward a whole named
18757 * sequence instead of matching the 2nd character when it fails the
18759 sv_catpvs(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
18763 /* Set up the data structure so that any errors will be properly
18764 * reported. See the comments at the definition of
18765 * REPORT_LOCATION_ARGS for details */
18766 RExC_copy_start_in_input = (char *) orig_parse;
18767 RExC_start = RExC_parse = SvPV(substitute_parse, len);
18768 RExC_copy_start_in_constructed = RExC_start + constructed_prefix_len;
18769 RExC_end = RExC_parse + len;
18770 RExC_in_multi_char_class = 1;
18772 ret = reg(pRExC_state, 1, ®_flags, depth+1);
18774 *flagp |= reg_flags & (HASWIDTH|SIMPLE|POSTPONED|RESTART_PARSE|NEED_UTF8);
18776 /* And restore so can parse the rest of the pattern */
18777 RExC_parse = save_parse;
18778 RExC_start = RExC_copy_start_in_constructed = RExC_copy_start_in_input = save_start;
18779 RExC_end = save_end;
18780 RExC_in_multi_char_class = 0;
18781 SvREFCNT_dec_NN(multi_char_matches);
18782 SvREFCNT_dec(properties);
18783 SvREFCNT_dec(cp_list);
18784 SvREFCNT_dec(simple_posixes);
18785 SvREFCNT_dec(posixes);
18786 SvREFCNT_dec(nposixes);
18787 SvREFCNT_dec(cp_foldable_list);
18791 /* If folding, we calculate all characters that could fold to or from the
18792 * ones already on the list */
18793 if (cp_foldable_list) {
18795 UV start, end; /* End points of code point ranges */
18797 SV* fold_intersection = NULL;
18800 /* Our calculated list will be for Unicode rules. For locale
18801 * matching, we have to keep a separate list that is consulted at
18802 * runtime only when the locale indicates Unicode rules (and we
18803 * don't include potential matches in the ASCII/Latin1 range, as
18804 * any code point could fold to any other, based on the run-time
18805 * locale). For non-locale, we just use the general list */
18807 use_list = &only_utf8_locale_list;
18810 use_list = &cp_list;
18813 /* Only the characters in this class that participate in folds need
18814 * be checked. Get the intersection of this class and all the
18815 * possible characters that are foldable. This can quickly narrow
18816 * down a large class */
18817 _invlist_intersection(PL_in_some_fold, cp_foldable_list,
18818 &fold_intersection);
18820 /* Now look at the foldable characters in this class individually */
18821 invlist_iterinit(fold_intersection);
18822 while (invlist_iternext(fold_intersection, &start, &end)) {
18826 /* Look at every character in the range */
18827 for (j = start; j <= end; j++) {
18828 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
18831 Size_t folds_count;
18833 const U32 * remaining_folds;
18837 /* Under /l, we don't know what code points below 256
18838 * fold to, except we do know the MICRO SIGN folds to
18839 * an above-255 character if the locale is UTF-8, so we
18840 * add it to the special list (in *use_list) Otherwise
18841 * we know now what things can match, though some folds
18842 * are valid under /d only if the target is UTF-8.
18843 * Those go in a separate list */
18844 if ( IS_IN_SOME_FOLD_L1(j)
18845 && ! (LOC && j != MICRO_SIGN))
18848 /* ASCII is always matched; non-ASCII is matched
18849 * only under Unicode rules (which could happen
18850 * under /l if the locale is a UTF-8 one */
18851 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
18852 *use_list = add_cp_to_invlist(*use_list,
18853 PL_fold_latin1[j]);
18855 else if (j != PL_fold_latin1[j]) {
18856 upper_latin1_only_utf8_matches
18857 = add_cp_to_invlist(
18858 upper_latin1_only_utf8_matches,
18859 PL_fold_latin1[j]);
18863 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(j)
18864 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
18866 add_above_Latin1_folds(pRExC_state,
18873 /* Here is an above Latin1 character. We don't have the
18874 * rules hard-coded for it. First, get its fold. This is
18875 * the simple fold, as the multi-character folds have been
18876 * handled earlier and separated out */
18877 folded = _to_uni_fold_flags(j, foldbuf, &foldlen,
18878 (ASCII_FOLD_RESTRICTED)
18879 ? FOLD_FLAGS_NOMIX_ASCII
18882 /* Single character fold of above Latin1. Add everything
18883 * in its fold closure to the list that this node should
18885 folds_count = _inverse_folds(folded, &first_fold,
18887 for (k = 0; k <= folds_count; k++) {
18888 UV c = (k == 0) /* First time through use itself */
18890 : (k == 1) /* 2nd time use, the first fold */
18893 /* Then the remaining ones */
18894 : remaining_folds[k-2];
18896 /* /aa doesn't allow folds between ASCII and non- */
18897 if (( ASCII_FOLD_RESTRICTED
18898 && (isASCII(c) != isASCII(j))))
18903 /* Folds under /l which cross the 255/256 boundary are
18904 * added to a separate list. (These are valid only
18905 * when the locale is UTF-8.) */
18906 if (c < 256 && LOC) {
18907 *use_list = add_cp_to_invlist(*use_list, c);
18911 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
18913 cp_list = add_cp_to_invlist(cp_list, c);
18916 /* Similarly folds involving non-ascii Latin1
18917 * characters under /d are added to their list */
18918 upper_latin1_only_utf8_matches
18919 = add_cp_to_invlist(
18920 upper_latin1_only_utf8_matches,
18926 SvREFCNT_dec_NN(fold_intersection);
18929 /* Now that we have finished adding all the folds, there is no reason
18930 * to keep the foldable list separate */
18931 _invlist_union(cp_list, cp_foldable_list, &cp_list);
18932 SvREFCNT_dec_NN(cp_foldable_list);
18935 /* And combine the result (if any) with any inversion lists from posix
18936 * classes. The lists are kept separate up to now because we don't want to
18937 * fold the classes */
18938 if (simple_posixes) { /* These are the classes known to be unaffected by
18941 _invlist_union(cp_list, simple_posixes, &cp_list);
18942 SvREFCNT_dec_NN(simple_posixes);
18945 cp_list = simple_posixes;
18948 if (posixes || nposixes) {
18949 if (! DEPENDS_SEMANTICS) {
18951 /* For everything but /d, we can just add the current 'posixes' and
18952 * 'nposixes' to the main list */
18955 _invlist_union(cp_list, posixes, &cp_list);
18956 SvREFCNT_dec_NN(posixes);
18964 _invlist_union(cp_list, nposixes, &cp_list);
18965 SvREFCNT_dec_NN(nposixes);
18968 cp_list = nposixes;
18973 /* Under /d, things like \w match upper Latin1 characters only if
18974 * the target string is in UTF-8. But things like \W match all the
18975 * upper Latin1 characters if the target string is not in UTF-8.
18977 * Handle the case with something like \W separately */
18979 SV* only_non_utf8_list = invlist_clone(PL_UpperLatin1, NULL);
18981 /* A complemented posix class matches all upper Latin1
18982 * characters if not in UTF-8. And it matches just certain
18983 * ones when in UTF-8. That means those certain ones are
18984 * matched regardless, so can just be added to the
18985 * unconditional list */
18987 _invlist_union(cp_list, nposixes, &cp_list);
18988 SvREFCNT_dec_NN(nposixes);
18992 cp_list = nposixes;
18995 /* Likewise for 'posixes' */
18996 _invlist_union(posixes, cp_list, &cp_list);
18997 SvREFCNT_dec(posixes);
18999 /* Likewise for anything else in the range that matched only
19001 if (upper_latin1_only_utf8_matches) {
19002 _invlist_union(cp_list,
19003 upper_latin1_only_utf8_matches,
19005 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
19006 upper_latin1_only_utf8_matches = NULL;
19009 /* If we don't match all the upper Latin1 characters regardless
19010 * of UTF-8ness, we have to set a flag to match the rest when
19012 _invlist_subtract(only_non_utf8_list, cp_list,
19013 &only_non_utf8_list);
19014 if (_invlist_len(only_non_utf8_list) != 0) {
19015 anyof_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
19017 SvREFCNT_dec_NN(only_non_utf8_list);
19020 /* Here there were no complemented posix classes. That means
19021 * the upper Latin1 characters in 'posixes' match only when the
19022 * target string is in UTF-8. So we have to add them to the
19023 * list of those types of code points, while adding the
19024 * remainder to the unconditional list.
19026 * First calculate what they are */
19027 SV* nonascii_but_latin1_properties = NULL;
19028 _invlist_intersection(posixes, PL_UpperLatin1,
19029 &nonascii_but_latin1_properties);
19031 /* And add them to the final list of such characters. */
19032 _invlist_union(upper_latin1_only_utf8_matches,
19033 nonascii_but_latin1_properties,
19034 &upper_latin1_only_utf8_matches);
19036 /* Remove them from what now becomes the unconditional list */
19037 _invlist_subtract(posixes, nonascii_but_latin1_properties,
19040 /* And add those unconditional ones to the final list */
19042 _invlist_union(cp_list, posixes, &cp_list);
19043 SvREFCNT_dec_NN(posixes);
19050 SvREFCNT_dec(nonascii_but_latin1_properties);
19052 /* Get rid of any characters from the conditional list that we
19053 * now know are matched unconditionally, which may make that
19055 _invlist_subtract(upper_latin1_only_utf8_matches,
19057 &upper_latin1_only_utf8_matches);
19058 if (_invlist_len(upper_latin1_only_utf8_matches) == 0) {
19059 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
19060 upper_latin1_only_utf8_matches = NULL;
19066 /* And combine the result (if any) with any inversion list from properties.
19067 * The lists are kept separate up to now so that we can distinguish the two
19068 * in regards to matching above-Unicode. A run-time warning is generated
19069 * if a Unicode property is matched against a non-Unicode code point. But,
19070 * we allow user-defined properties to match anything, without any warning,
19071 * and we also suppress the warning if there is a portion of the character
19072 * class that isn't a Unicode property, and which matches above Unicode, \W
19073 * or [\x{110000}] for example.
19074 * (Note that in this case, unlike the Posix one above, there is no
19075 * <upper_latin1_only_utf8_matches>, because having a Unicode property
19076 * forces Unicode semantics */
19080 /* If it matters to the final outcome, see if a non-property
19081 * component of the class matches above Unicode. If so, the
19082 * warning gets suppressed. This is true even if just a single
19083 * such code point is specified, as, though not strictly correct if
19084 * another such code point is matched against, the fact that they
19085 * are using above-Unicode code points indicates they should know
19086 * the issues involved */
19088 warn_super = ! (invert
19089 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
19092 _invlist_union(properties, cp_list, &cp_list);
19093 SvREFCNT_dec_NN(properties);
19096 cp_list = properties;
19101 |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
19103 /* Because an ANYOF node is the only one that warns, this node
19104 * can't be optimized into something else */
19105 optimizable = FALSE;
19109 /* Here, we have calculated what code points should be in the character
19112 * Now we can see about various optimizations. Fold calculation (which we
19113 * did above) needs to take place before inversion. Otherwise /[^k]/i
19114 * would invert to include K, which under /i would match k, which it
19115 * shouldn't. Therefore we can't invert folded locale now, as it won't be
19116 * folded until runtime */
19118 /* If we didn't do folding, it's because some information isn't available
19119 * until runtime; set the run-time fold flag for these We know to set the
19120 * flag if we have a non-NULL list for UTF-8 locales, or the class matches
19121 * at least one 0-255 range code point */
19124 /* Some things on the list might be unconditionally included because of
19125 * other components. Remove them, and clean up the list if it goes to
19127 if (only_utf8_locale_list && cp_list) {
19128 _invlist_subtract(only_utf8_locale_list, cp_list,
19129 &only_utf8_locale_list);
19131 if (_invlist_len(only_utf8_locale_list) == 0) {
19132 SvREFCNT_dec_NN(only_utf8_locale_list);
19133 only_utf8_locale_list = NULL;
19136 if ( only_utf8_locale_list
19137 || (cp_list && ( _invlist_contains_cp(cp_list, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE)
19138 || _invlist_contains_cp(cp_list, LATIN_SMALL_LETTER_DOTLESS_I))))
19140 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
19143 | ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
19145 else if (cp_list && invlist_lowest(cp_list) < 256) {
19146 /* If nothing is below 256, has no locale dependency; otherwise it
19148 anyof_flags |= ANYOFL_FOLD;
19149 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
19152 else if ( DEPENDS_SEMANTICS
19153 && ( upper_latin1_only_utf8_matches
19154 || (anyof_flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)))
19156 RExC_seen_d_op = TRUE;
19157 has_runtime_dependency |= HAS_D_RUNTIME_DEPENDENCY;
19160 /* Optimize inverted patterns (e.g. [^a-z]) when everything is known at
19164 && ! has_runtime_dependency)
19166 _invlist_invert(cp_list);
19168 /* Clear the invert flag since have just done it here */
19172 /* All possible optimizations below still have these characteristics.
19173 * (Multi-char folds aren't SIMPLE, but they don't get this far in this
19175 *flagp |= HASWIDTH|SIMPLE;
19178 *ret_invlist = cp_list;
19180 return (cp_list) ? RExC_emit : 0;
19183 if (anyof_flags & ANYOF_LOCALE_FLAGS) {
19184 RExC_contains_locale = 1;
19189 /* Some character classes are equivalent to other nodes. Such nodes
19190 * take up less room, and some nodes require fewer operations to
19191 * execute, than ANYOF nodes. EXACTish nodes may be joinable with
19192 * adjacent nodes to improve efficiency. */
19193 op = optimize_regclass(pRExC_state, cp_list,
19194 only_utf8_locale_list,
19195 upper_latin1_only_utf8_matches,
19196 has_runtime_dependency,
19198 &anyof_flags, &invert, &ret, flagp);
19199 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
19201 /* If optimized to something else, finish up and return */
19203 Set_Node_Offset_Length(REGNODE_p(ret), orig_parse - RExC_start,
19204 RExC_parse - orig_parse);;
19205 SvREFCNT_dec(cp_list);;
19206 SvREFCNT_dec(only_utf8_locale_list);
19207 SvREFCNT_dec(upper_latin1_only_utf8_matches);
19212 /* Here didn't optimize, or optimized to a specialized ANYOF node. If the
19213 * former, set the particular type */
19215 if (has_runtime_dependency & HAS_D_RUNTIME_DEPENDENCY) {
19226 ret = regnode_guts(pRExC_state, op, regarglen[op], "anyof");
19227 FILL_NODE(ret, op); /* We set the argument later */
19228 RExC_emit += 1 + regarglen[op];
19229 ANYOF_FLAGS(REGNODE_p(ret)) = anyof_flags;
19231 /* Here, <cp_list> contains all the code points we can determine at
19232 * compile time that match under all conditions. Go through it, and
19233 * for things that belong in the bitmap, put them there, and delete from
19234 * <cp_list>. While we are at it, see if everything above 255 is in the
19235 * list, and if so, set a flag to speed up execution */
19237 populate_ANYOF_from_invlist(REGNODE_p(ret), &cp_list);
19240 ANYOF_POSIXL_SET_TO_BITMAP(REGNODE_p(ret), posixl);
19244 ANYOF_FLAGS(REGNODE_p(ret)) |= ANYOF_INVERT;
19247 /* Here, the bitmap has been populated with all the Latin1 code points that
19248 * always match. Can now add to the overall list those that match only
19249 * when the target string is UTF-8 (<upper_latin1_only_utf8_matches>).
19251 if (upper_latin1_only_utf8_matches) {
19253 _invlist_union(cp_list,
19254 upper_latin1_only_utf8_matches,
19256 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
19259 cp_list = upper_latin1_only_utf8_matches;
19261 ANYOF_FLAGS(REGNODE_p(ret)) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
19264 set_ANYOF_arg(pRExC_state, REGNODE_p(ret), cp_list,
19265 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
19268 only_utf8_locale_list);
19270 SvREFCNT_dec(cp_list);;
19271 SvREFCNT_dec(only_utf8_locale_list);
19276 S_optimize_regclass(pTHX_
19277 RExC_state_t *pRExC_state,
19279 SV* only_utf8_locale_list,
19280 SV* upper_latin1_only_utf8_matches,
19281 const U32 has_runtime_dependency,
19285 regnode_offset * ret,
19289 /* This function exists just to make S_regclass() smaller. It extracts out
19290 * the code that looks for potential optimizations away from a full generic
19291 * ANYOF node. The parameter names are the same as the corresponding
19292 * variables in S_regclass.
19294 * It returns the new op (ANYOF if no optimization found) and sets *ret to
19295 * any created regnode. If the new op is sufficiently like plain ANYOF, it
19296 * leaves *ret unchanged for allocation in S_regclass.
19298 * Certain of the parameters may be updated as a result of the changes herein */
19300 U8 op = ANYOF; /* The returned node-type, initialized to the unoptimized
19303 PERL_UINT_FAST8_T i;
19304 UV partial_cp_count = 0;
19305 UV start[MAX_FOLD_FROMS+1] = { 0 }; /* +1 for the folded-to char */
19306 UV end[MAX_FOLD_FROMS+1] = { 0 };
19307 bool single_range = FALSE;
19309 PERL_ARGS_ASSERT_OPTIMIZE_REGCLASS;
19311 if (cp_list) { /* Count the code points in enough ranges that we would
19312 see all the ones possible in any fold in this version
19315 invlist_iterinit(cp_list);
19316 for (i = 0; i <= MAX_FOLD_FROMS; i++) {
19317 if (! invlist_iternext(cp_list, &start[i], &end[i])) {
19320 partial_cp_count += end[i] - start[i] + 1;
19324 single_range = TRUE;
19326 invlist_iterfinish(cp_list);
19329 /* If we know at compile time that this matches every possible code
19330 * point, any run-time dependencies don't matter */
19331 if (start[0] == 0 && end[0] == UV_MAX) {
19334 *ret = reganode(pRExC_state, op, 0);
19338 *ret = reg_node(pRExC_state, op);
19344 /* Similarly, for /l posix classes, if both a class and its
19345 * complement match, any run-time dependencies don't matter */
19348 for (namedclass = 0; namedclass < ANYOF_POSIXL_MAX;
19351 if ( POSIXL_TEST(posixl, namedclass) /* class */
19352 && POSIXL_TEST(posixl, namedclass + 1)) /* its complement */
19356 *ret = reganode(pRExC_state, op, 0);
19360 *ret = reg_node(pRExC_state, op);
19367 /* For well-behaved locales, some classes are subsets of others,
19368 * so complementing the subset and including the non-complemented
19369 * superset should match everything, like [\D[:alnum:]], and
19370 * [[:^alpha:][:alnum:]], but some implementations of locales are
19371 * buggy, and khw thinks its a bad idea to have optimization change
19372 * behavior, even if it avoids an OS bug in a given case */
19374 #define isSINGLE_BIT_SET(n) isPOWER_OF_2(n)
19376 /* If is a single posix /l class, can optimize to just that op.
19377 * Such a node will not match anything in the Latin1 range, as that
19378 * is not determinable until runtime, but will match whatever the
19379 * class does outside that range. (Note that some classes won't
19380 * match anything outside the range, like [:ascii:]) */
19381 if ( isSINGLE_BIT_SET(posixl)
19382 && (partial_cp_count == 0 || start[0] > 255))
19385 SV * class_above_latin1 = NULL;
19386 bool already_inverted;
19387 bool are_equivalent;
19389 /* Compute which bit is set, which is the same thing as, e.g.,
19390 * ANYOF_CNTRL. From
19391 * https://graphics.stanford.edu/~seander/bithacks.html#IntegerLogDeBruijn
19393 static const int MultiplyDeBruijnBitPosition2[32] =
19395 0, 1, 28, 2, 29, 14, 24, 3, 30, 22, 20, 15, 25, 17, 4, 8,
19396 31, 27, 13, 23, 21, 19, 16, 7, 26, 12, 18, 6, 11, 5, 10, 9
19399 namedclass = MultiplyDeBruijnBitPosition2[(posixl
19400 * 0x077CB531U) >> 27];
19401 classnum = namedclass_to_classnum(namedclass);
19403 /* The named classes are such that the inverted number is one
19404 * larger than the non-inverted one */
19405 already_inverted = namedclass
19406 - classnum_to_namedclass(classnum);
19408 /* Create an inversion list of the official property, inverted
19409 * if the constructed node list is inverted, and restricted to
19410 * only the above latin1 code points, which are the only ones
19411 * known at compile time */
19412 _invlist_intersection_maybe_complement_2nd(
19414 PL_XPosix_ptrs[classnum],
19416 &class_above_latin1);
19417 are_equivalent = _invlistEQ(class_above_latin1, cp_list,
19419 SvREFCNT_dec_NN(class_above_latin1);
19421 if (are_equivalent) {
19423 /* Resolve the run-time inversion flag with this possibly
19424 * inverted class */
19425 *invert = *invert ^ already_inverted;
19427 op = POSIXL + *invert * (NPOSIXL - POSIXL);
19428 *ret = reg_node(pRExC_state, op);
19429 FLAGS(REGNODE_p(*ret)) = classnum;
19435 /* khw can't think of any other possible transformation involving
19437 if (has_runtime_dependency & HAS_USER_DEFINED_PROPERTY) {
19441 if (! has_runtime_dependency) {
19443 /* If the list is empty, nothing matches. This happens, for
19444 * example, when a Unicode property that doesn't match anything is
19445 * the only element in the character class (perluniprops.pod notes
19446 * such properties). */
19447 if (partial_cp_count == 0) {
19450 *ret = reg_node(pRExC_state, op);
19454 *ret = reganode(pRExC_state, op, 0);
19460 /* If matches everything but \n */
19461 if ( start[0] == 0 && end[0] == '\n' - 1
19462 && start[1] == '\n' + 1 && end[1] == UV_MAX)
19464 assert (! *invert);
19466 *ret = reg_node(pRExC_state, op);
19472 /* Next see if can optimize classes that contain just a few code points
19473 * into an EXACTish node. The reason to do this is to let the
19474 * optimizer join this node with adjacent EXACTish ones, and ANYOF
19475 * nodes require conversion to code point from UTF-8.
19477 * An EXACTFish node can be generated even if not under /i, and vice
19478 * versa. But care must be taken. An EXACTFish node has to be such
19479 * that it only matches precisely the code points in the class, but we
19480 * want to generate the least restrictive one that does that, to
19481 * increase the odds of being able to join with an adjacent node. For
19482 * example, if the class contains [kK], we have to make it an EXACTFAA
19483 * node to prevent the KELVIN SIGN from matching. Whether we are under
19484 * /i or not is irrelevant in this case. Less obvious is the pattern
19485 * qr/[\x{02BC}]n/i. U+02BC is MODIFIER LETTER APOSTROPHE. That is
19486 * supposed to match the single character U+0149 LATIN SMALL LETTER N
19487 * PRECEDED BY APOSTROPHE. And so even though there is no simple fold
19488 * that includes \X{02BC}, there is a multi-char fold that does, and so
19489 * the node generated for it must be an EXACTFish one. On the other
19490 * hand qr/:/i should generate a plain EXACT node since the colon
19491 * participates in no fold whatsoever, and having it EXACT tells the
19492 * optimizer the target string cannot match unless it has a colon in
19498 /* Only try if there are no more code points in the class than
19499 * in the max possible fold */
19500 && inRANGE(partial_cp_count, 1, MAX_FOLD_FROMS + 1))
19502 if (partial_cp_count == 1 && ! upper_latin1_only_utf8_matches)
19504 /* We can always make a single code point class into an
19505 * EXACTish node. */
19509 /* Here is /l: Use EXACTL, except if there is a fold not
19510 * known until runtime so shows as only a single code point
19511 * here. For code points above 255, we know which can
19512 * cause problems by having a potential fold to the Latin1
19515 || ( start[0] > 255
19516 && ! is_PROBLEMATIC_LOCALE_FOLD_cp(start[0])))
19524 else if (! FOLD) { /* Not /l and not /i */
19525 op = (start[0] < 256) ? EXACT : EXACT_REQ8;
19527 else if (start[0] < 256) { /* /i, not /l, and the code point is
19530 /* Under /i, it gets a little tricky. A code point that
19531 * doesn't participate in a fold should be an EXACT node.
19532 * We know this one isn't the result of a simple fold, or
19533 * there'd be more than one code point in the list, but it
19534 * could be part of a multi- character fold. In that case
19535 * we better not create an EXACT node, as we would wrongly
19536 * be telling the optimizer that this code point must be in
19537 * the target string, and that is wrong. This is because
19538 * if the sequence around this code point forms a
19539 * multi-char fold, what needs to be in the string could be
19540 * the code point that folds to the sequence.
19542 * This handles the case of below-255 code points, as we
19543 * have an easy look up for those. The next clause handles
19544 * the above-256 one */
19545 op = IS_IN_SOME_FOLD_L1(start[0])
19549 else { /* /i, larger code point. Since we are under /i, and
19550 have just this code point, we know that it can't
19551 fold to something else, so PL_InMultiCharFold
19553 op = _invlist_contains_cp(PL_InMultiCharFold,
19561 else if ( ! (has_runtime_dependency & ~HAS_D_RUNTIME_DEPENDENCY)
19562 && _invlist_contains_cp(PL_in_some_fold, start[0]))
19564 /* Here, the only runtime dependency, if any, is from /d, and
19565 * the class matches more than one code point, and the lowest
19566 * code point participates in some fold. It might be that the
19567 * other code points are /i equivalent to this one, and hence
19568 * they would representable by an EXACTFish node. Above, we
19569 * eliminated classes that contain too many code points to be
19570 * EXACTFish, with the test for MAX_FOLD_FROMS
19572 * First, special case the ASCII fold pairs, like 'B' and 'b'.
19573 * We do this because we have EXACTFAA at our disposal for the
19575 if (partial_cp_count == 2 && isASCII(start[0])) {
19577 /* The only ASCII characters that participate in folds are
19579 assert(isALPHA(start[0]));
19580 if ( end[0] == start[0] /* First range is a single
19581 character, so 2nd exists */
19582 && isALPHA_FOLD_EQ(start[0], start[1]))
19585 /* Here, is part of an ASCII fold pair */
19587 if ( ASCII_FOLD_RESTRICTED
19588 || HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(start[0]))
19590 /* If the second clause just above was true, it
19591 * means we can't be under /i, or else the list
19592 * would have included more than this fold pair.
19593 * Therefore we have to exclude the possibility of
19594 * whatever else it is that folds to these, by
19595 * using EXACTFAA */
19598 else if (HAS_NONLATIN1_FOLD_CLOSURE(start[0])) {
19600 /* Here, there's no simple fold that start[0] is part
19601 * of, but there is a multi-character one. If we
19602 * are not under /i, we want to exclude that
19603 * possibility; if under /i, we want to include it
19605 op = (FOLD) ? EXACTFU : EXACTFAA;
19609 /* Here, the only possible fold start[0] particpates in
19610 * is with start[1]. /i or not isn't relevant */
19614 value = toFOLD(start[0]);
19617 else if ( ! upper_latin1_only_utf8_matches
19618 || ( _invlist_len(upper_latin1_only_utf8_matches)
19621 invlist_highest(upper_latin1_only_utf8_matches)]
19624 /* Here, the smallest character is non-ascii or there are
19625 * more than 2 code points matched by this node. Also, we
19626 * either don't have /d UTF-8 dependent matches, or if we
19627 * do, they look like they could be a single character that
19628 * is the fold of the lowest one in the always-match list.
19629 * This test quickly excludes most of the false positives
19630 * when there are /d UTF-8 depdendent matches. These are
19631 * like LATIN CAPITAL LETTER A WITH GRAVE matching LATIN
19632 * SMALL LETTER A WITH GRAVE iff the target string is
19633 * UTF-8. (We don't have to worry above about exceeding
19634 * the array bounds of PL_fold_latin1[] because any code
19635 * point in 'upper_latin1_only_utf8_matches' is below 256.)
19637 * EXACTFAA would apply only to pairs (hence exactly 2 code
19638 * points) in the ASCII range, so we can't use it here to
19639 * artificially restrict the fold domain, so we check if
19640 * the class does or does not match some EXACTFish node.
19641 * Further, if we aren't under /i, and the folded-to
19642 * character is part of a multi-character fold, we can't do
19643 * this optimization, as the sequence around it could be
19644 * that multi-character fold, and we don't here know the
19645 * context, so we have to assume it is that multi-char
19646 * fold, to prevent potential bugs.
19648 * To do the general case, we first find the fold of the
19649 * lowest code point (which may be higher than the lowest
19650 * one), then find everything that folds to it. (The data
19651 * structure we have only maps from the folded code points,
19652 * so we have to do the earlier step.) */
19655 U8 foldbuf[UTF8_MAXBYTES_CASE];
19656 UV folded = _to_uni_fold_flags(start[0],
19657 foldbuf, &foldlen, 0);
19659 const U32 * remaining_folds;
19660 Size_t folds_to_this_cp_count = _inverse_folds(
19664 Size_t folds_count = folds_to_this_cp_count + 1;
19665 SV * fold_list = _new_invlist(folds_count);
19668 /* If there are UTF-8 dependent matches, create a temporary
19669 * list of what this node matches, including them. */
19670 SV * all_cp_list = NULL;
19671 SV ** use_this_list = &cp_list;
19673 if (upper_latin1_only_utf8_matches) {
19674 all_cp_list = _new_invlist(0);
19675 use_this_list = &all_cp_list;
19676 _invlist_union(cp_list,
19677 upper_latin1_only_utf8_matches,
19681 /* Having gotten everything that participates in the fold
19682 * containing the lowest code point, we turn that into an
19683 * inversion list, making sure everything is included. */
19684 fold_list = add_cp_to_invlist(fold_list, start[0]);
19685 fold_list = add_cp_to_invlist(fold_list, folded);
19686 if (folds_to_this_cp_count > 0) {
19687 fold_list = add_cp_to_invlist(fold_list, first_fold);
19688 for (i = 0; i + 1 < folds_to_this_cp_count; i++) {
19689 fold_list = add_cp_to_invlist(fold_list,
19690 remaining_folds[i]);
19694 /* If the fold list is identical to what's in this ANYOF
19695 * node, the node can be represented by an EXACTFish one
19697 if (_invlistEQ(*use_this_list, fold_list,
19698 0 /* Don't complement */ )
19701 /* But, we have to be careful, as mentioned above.
19702 * Just the right sequence of characters could match
19703 * this if it is part of a multi-character fold. That
19704 * IS what we want if we are under /i. But it ISN'T
19705 * what we want if not under /i, as it could match when
19706 * it shouldn't. So, when we aren't under /i and this
19707 * character participates in a multi-char fold, we
19708 * don't optimize into an EXACTFish node. So, for each
19709 * case below we have to check if we are folding
19710 * and if not, if it is not part of a multi-char fold.
19712 if (start[0] > 255) { /* Highish code point */
19713 if (FOLD || ! _invlist_contains_cp(
19714 PL_InMultiCharFold, folded))
19718 : (ASCII_FOLD_RESTRICTED)
19723 } /* Below, the lowest code point < 256 */
19726 && DEPENDS_SEMANTICS)
19727 { /* An EXACTF node containing a single character
19728 's', can be an EXACTFU if it doesn't get
19729 joined with an adjacent 's' */
19730 op = EXACTFU_S_EDGE;
19734 || ! HAS_NONLATIN1_FOLD_CLOSURE(start[0]))
19736 if (upper_latin1_only_utf8_matches) {
19739 /* We can't use the fold, as that only matches
19743 else if ( UNLIKELY(start[0] == MICRO_SIGN)
19745 { /* EXACTFUP is a special node for this
19747 op = (ASCII_FOLD_RESTRICTED)
19750 value = MICRO_SIGN;
19752 else if ( ASCII_FOLD_RESTRICTED
19753 && ! isASCII(start[0]))
19754 { /* For ASCII under /iaa, we can use EXACTFU
19766 SvREFCNT_dec_NN(fold_list);
19767 SvREFCNT_dec(all_cp_list);
19774 /* Here, we have calculated what EXACTish node to use. Have to
19775 * convert to UTF-8 if not already there */
19778 SvREFCNT_dec(cp_list);;
19779 REQUIRE_UTF8(flagp);
19782 /* This is a kludge to the special casing issues with this
19783 * ligature under /aa. FB05 should fold to FB06, but the
19784 * call above to _to_uni_fold_flags() didn't find this, as
19785 * it didn't use the /aa restriction in order to not miss
19786 * other folds that would be affected. This is the only
19787 * instance likely to ever be a problem in all of Unicode.
19788 * So special case it. */
19789 if ( value == LATIN_SMALL_LIGATURE_LONG_S_T
19790 && ASCII_FOLD_RESTRICTED)
19792 value = LATIN_SMALL_LIGATURE_ST;
19796 len = (UTF) ? UVCHR_SKIP(value) : 1;
19798 *ret = regnode_guts(pRExC_state, op, len, "exact");
19799 FILL_NODE(*ret, op);
19800 RExC_emit += 1 + STR_SZ(len);
19801 setSTR_LEN(REGNODE_p(*ret), len);
19803 *STRINGs(REGNODE_p(*ret)) = (U8) value;
19806 uvchr_to_utf8((U8 *) STRINGs(REGNODE_p(*ret)), value);
19812 if (! has_runtime_dependency) {
19814 /* See if this can be turned into an ANYOFM node. Think about the
19815 * bit patterns in two different bytes. In some positions, the
19816 * bits in each will be 1; and in other positions both will be 0;
19817 * and in some positions the bit will be 1 in one byte, and 0 in
19818 * the other. Let 'n' be the number of positions where the bits
19819 * differ. We create a mask which has exactly 'n' 0 bits, each in
19820 * a position where the two bytes differ. Now take the set of all
19821 * bytes that when ANDed with the mask yield the same result. That
19822 * set has 2**n elements, and is representable by just two 8 bit
19823 * numbers: the result and the mask. Importantly, matching the set
19824 * can be vectorized by creating a word full of the result bytes,
19825 * and a word full of the mask bytes, yielding a significant speed
19826 * up. Here, see if this node matches such a set. As a concrete
19827 * example consider [01], and the byte representing '0' which is
19828 * 0x30 on ASCII machines. It has the bits 0011 0000. Take the
19829 * mask 1111 1110. If we AND 0x31 and 0x30 with that mask we get
19830 * 0x30. Any other bytes ANDed yield something else. So [01],
19831 * which is a common usage, is optimizable into ANYOFM, and can
19832 * benefit from the speed up. We can only do this on UTF-8
19833 * invariant bytes, because they have the same bit patterns under
19835 PERL_UINT_FAST8_T inverted = 0;
19837 const PERL_UINT_FAST8_T max_permissible = 0xFF;
19839 const PERL_UINT_FAST8_T max_permissible = 0x7F;
19841 /* If doesn't fit the criteria for ANYOFM, invert and try again.
19842 * If that works we will instead later generate an NANYOFM, and
19843 * invert back when through */
19844 if (invlist_highest(cp_list) > max_permissible) {
19845 _invlist_invert(cp_list);
19849 if (invlist_highest(cp_list) <= max_permissible) {
19850 UV this_start, this_end;
19851 UV lowest_cp = UV_MAX; /* init'ed to suppress compiler warn */
19852 U8 bits_differing = 0;
19853 Size_t full_cp_count = 0;
19854 bool first_time = TRUE;
19856 /* Go through the bytes and find the bit positions that differ
19858 invlist_iterinit(cp_list);
19859 while (invlist_iternext(cp_list, &this_start, &this_end)) {
19860 unsigned int i = this_start;
19863 if (! UVCHR_IS_INVARIANT(i)) {
19867 first_time = FALSE;
19868 lowest_cp = this_start;
19870 /* We have set up the code point to compare with.
19871 * Don't compare it with itself */
19875 /* Find the bit positions that differ from the lowest code
19876 * point in the node. Keep track of all such positions by
19878 for (; i <= this_end; i++) {
19879 if (! UVCHR_IS_INVARIANT(i)) {
19883 bits_differing |= i ^ lowest_cp;
19886 full_cp_count += this_end - this_start + 1;
19889 /* At the end of the loop, we count how many bits differ from
19890 * the bits in lowest code point, call the count 'd'. If the
19891 * set we found contains 2**d elements, it is the closure of
19892 * all code points that differ only in those bit positions. To
19893 * convince yourself of that, first note that the number in the
19894 * closure must be a power of 2, which we test for. The only
19895 * way we could have that count and it be some differing set,
19896 * is if we got some code points that don't differ from the
19897 * lowest code point in any position, but do differ from each
19898 * other in some other position. That means one code point has
19899 * a 1 in that position, and another has a 0. But that would
19900 * mean that one of them differs from the lowest code point in
19901 * that position, which possibility we've already excluded. */
19902 if ( (inverted || full_cp_count > 1)
19903 && full_cp_count == 1U << PL_bitcount[bits_differing])
19907 op = ANYOFM + inverted;;
19909 /* We need to make the bits that differ be 0's */
19910 ANYOFM_mask = ~ bits_differing; /* This goes into FLAGS */
19912 /* The argument is the lowest code point */
19913 *ret = reganode(pRExC_state, op, lowest_cp);
19914 FLAGS(REGNODE_p(*ret)) = ANYOFM_mask;
19918 invlist_iterfinish(cp_list);
19922 _invlist_invert(cp_list);
19929 /* XXX We could create an ANYOFR_LOW node here if we saved above if
19930 * all were invariants, it wasn't inverted, and there is a single
19931 * range. This would be faster than some of the posix nodes we
19932 * create below like /\d/a, but would be twice the size. Without
19933 * having actually measured the gain, khw doesn't think the
19934 * tradeoff is really worth it */
19937 if (! (*anyof_flags & ANYOF_LOCALE_FLAGS)) {
19938 PERL_UINT_FAST8_T type;
19939 SV * intersection = NULL;
19940 SV* d_invlist = NULL;
19942 /* See if this matches any of the POSIX classes. The POSIXA and
19943 * POSIXD ones are about the same speed as ANYOF ops, but take less
19944 * room; the ones that have above-Latin1 code point matches are
19945 * somewhat faster than ANYOF. */
19947 for (type = POSIXA; type >= POSIXD; type--) {
19950 if (type == POSIXL) { /* But not /l posix classes */
19954 for (posix_class = 0;
19955 posix_class <= _HIGHEST_REGCOMP_DOT_H_SYNC;
19958 SV** our_code_points = &cp_list;
19959 SV** official_code_points;
19962 if (type == POSIXA) {
19963 official_code_points = &PL_Posix_ptrs[posix_class];
19966 official_code_points = &PL_XPosix_ptrs[posix_class];
19969 /* Skip non-existent classes of this type. e.g. \v only
19970 * has an entry in PL_XPosix_ptrs */
19971 if (! *official_code_points) {
19975 /* Try both the regular class, and its inversion */
19976 for (try_inverted = 0; try_inverted < 2; try_inverted++) {
19977 bool this_inverted = *invert ^ try_inverted;
19979 if (type != POSIXD) {
19981 /* This class that isn't /d can't match if we have
19982 * /d dependencies */
19983 if (has_runtime_dependency
19984 & HAS_D_RUNTIME_DEPENDENCY)
19989 else /* is /d */ if (! this_inverted) {
19991 /* /d classes don't match anything non-ASCII below
19992 * 256 unconditionally (which cp_list contains) */
19993 _invlist_intersection(cp_list, PL_UpperLatin1,
19995 if (_invlist_len(intersection) != 0) {
19999 SvREFCNT_dec(d_invlist);
20000 d_invlist = invlist_clone(cp_list, NULL);
20002 /* But under UTF-8 it turns into using /u rules.
20003 * Add the things it matches under these conditions
20004 * so that we check below that these are identical
20005 * to what the tested class should match */
20006 if (upper_latin1_only_utf8_matches) {
20009 upper_latin1_only_utf8_matches,
20012 our_code_points = &d_invlist;
20014 else { /* POSIXD, inverted. If this doesn't have this
20015 flag set, it isn't /d. */
20016 if (! (*anyof_flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
20020 our_code_points = &cp_list;
20023 /* Here, have weeded out some things. We want to see
20024 * if the list of characters this node contains
20025 * ('*our_code_points') precisely matches those of the
20026 * class we are currently checking against
20027 * ('*official_code_points'). */
20028 if (_invlistEQ(*our_code_points,
20029 *official_code_points,
20032 /* Here, they precisely match. Optimize this ANYOF
20033 * node into its equivalent POSIX one of the
20034 * correct type, possibly inverted */
20035 op = (try_inverted)
20036 ? type + NPOSIXA - POSIXA
20038 *ret = reg_node(pRExC_state, op);
20039 FLAGS(REGNODE_p(*ret)) = posix_class;
20040 SvREFCNT_dec(d_invlist);
20041 SvREFCNT_dec(intersection);
20047 SvREFCNT_dec(d_invlist);
20048 SvREFCNT_dec(intersection);
20051 /* If it is a single contiguous range, ANYOFR is an efficient regnode,
20052 * both in size and speed. Currently, a 20 bit range base (smallest
20053 * code point in the range), and a 12 bit maximum delta are packed into
20054 * a 32 bit word. This allows for using it on all of the Unicode code
20055 * points except for the highest plane, which is only for private use
20056 * code points. khw doubts that a bigger delta is likely in real world
20059 && ! has_runtime_dependency
20060 && *anyof_flags == 0
20061 && start[0] < (1 << ANYOFR_BASE_BITS)
20062 && end[0] - start[0]
20063 < ((1U << (sizeof(((struct regnode_1 *)NULL)->arg1)
20064 * CHARBITS - ANYOFR_BASE_BITS))))
20067 U8 low_utf8[UTF8_MAXBYTES+1];
20068 U8 high_utf8[UTF8_MAXBYTES+1];
20071 *ret = reganode(pRExC_state, op,
20072 (start[0] | (end[0] - start[0]) << ANYOFR_BASE_BITS));
20074 /* Place the lowest UTF-8 start byte in the flags field, so as to
20075 * allow efficient ruling out at run time of many possible inputs.
20077 (void) uvchr_to_utf8(low_utf8, start[0]);
20078 (void) uvchr_to_utf8(high_utf8, end[0]);
20080 /* If all code points share the same first byte, this can be an
20081 * ANYOFRb. Otherwise store the lowest UTF-8 start byte which can
20082 * quickly rule out many inputs at run-time without having to
20083 * compute the code point from UTF-8. For EBCDIC, we use I8, as
20084 * not doing that transformation would not rule out nearly so many
20086 if (low_utf8[0] == high_utf8[0]) {
20088 OP(REGNODE_p(*ret)) = op;
20089 ANYOF_FLAGS(REGNODE_p(*ret)) = low_utf8[0];
20092 ANYOF_FLAGS(REGNODE_p(*ret))
20093 = NATIVE_UTF8_TO_I8(low_utf8[0]);
20099 /* If didn't find an optimization and there is no need for a bitmap,
20100 * optimize to indicate that */
20101 if ( start[0] >= NUM_ANYOF_CODE_POINTS
20103 && ! upper_latin1_only_utf8_matches
20104 && *anyof_flags == 0)
20106 U8 low_utf8[UTF8_MAXBYTES+1];
20107 UV highest_cp = invlist_highest(cp_list);
20109 /* Currently the maximum allowed code point by the system is
20110 * IV_MAX. Higher ones are reserved for future internal use. This
20111 * particular regnode can be used for higher ones, but we can't
20112 * calculate the code point of those. IV_MAX suffices though, as
20113 * it will be a large first byte */
20114 Size_t low_len = uvchr_to_utf8(low_utf8, MIN(start[0], IV_MAX))
20117 /* We store the lowest possible first byte of the UTF-8
20118 * representation, using the flags field. This allows for quick
20119 * ruling out of some inputs without having to convert from UTF-8
20120 * to code point. For EBCDIC, we use I8, as not doing that
20121 * transformation would not rule out nearly so many things */
20122 *anyof_flags = NATIVE_UTF8_TO_I8(low_utf8[0]);
20126 /* If the first UTF-8 start byte for the highest code point in the
20127 * range is suitably small, we may be able to get an upper bound as
20129 if (highest_cp <= IV_MAX) {
20130 U8 high_utf8[UTF8_MAXBYTES+1];
20131 Size_t high_len = uvchr_to_utf8(high_utf8, highest_cp)
20134 /* If the lowest and highest are the same, we can get an exact
20135 * first byte instead of a just minimum or even a sequence of
20136 * exact leading bytes. We signal these with different
20138 if (low_utf8[0] == high_utf8[0]) {
20139 Size_t len = find_first_differing_byte_pos(low_utf8,
20141 MIN(low_len, high_len));
20145 /* No need to convert to I8 for EBCDIC as this is an
20147 *anyof_flags = low_utf8[0];
20152 *ret = regnode_guts(pRExC_state, op,
20153 regarglen[op] + STR_SZ(len),
20155 FILL_NODE(*ret, op);
20156 ((struct regnode_anyofhs *) REGNODE_p(*ret))->str_len
20158 Copy(low_utf8, /* Add the common bytes */
20159 ((struct regnode_anyofhs *) REGNODE_p(*ret))->string,
20161 RExC_emit += NODE_SZ_STR(REGNODE_p(*ret));
20162 set_ANYOF_arg(pRExC_state, REGNODE_p(*ret), cp_list,
20163 NULL, only_utf8_locale_list);
20167 else if (NATIVE_UTF8_TO_I8(high_utf8[0]) <= MAX_ANYOF_HRx_BYTE)
20170 /* Here, the high byte is not the same as the low, but is
20171 * small enough that its reasonable to have a loose upper
20172 * bound, which is packed in with the strict lower bound.
20173 * See comments at the definition of MAX_ANYOF_HRx_BYTE.
20174 * On EBCDIC platforms, I8 is used. On ASCII platforms I8
20175 * is the same thing as UTF-8 */
20178 U8 max_range_diff = MAX_ANYOF_HRx_BYTE - *anyof_flags;
20179 U8 range_diff = NATIVE_UTF8_TO_I8(high_utf8[0])
20182 if (range_diff <= max_range_diff / 8) {
20185 else if (range_diff <= max_range_diff / 4) {
20188 else if (range_diff <= max_range_diff / 2) {
20191 *anyof_flags = (*anyof_flags - 0xC0) << 2 | bits;
20200 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
20203 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
20204 regnode* const node,
20206 SV* const runtime_defns,
20207 SV* const only_utf8_locale_list)
20209 /* Sets the arg field of an ANYOF-type node 'node', using information about
20210 * the node passed-in. If there is nothing outside the node's bitmap, the
20211 * arg is set to ANYOF_ONLY_HAS_BITMAP. Otherwise, it sets the argument to
20212 * the count returned by add_data(), having allocated and stored an array,
20215 * av[0] stores the inversion list defining this class as far as known at
20216 * this time, or PL_sv_undef if nothing definite is now known.
20217 * av[1] stores the inversion list of code points that match only if the
20218 * current locale is UTF-8, or if none, PL_sv_undef if there is an
20219 * av[2], or no entry otherwise.
20220 * av[2] stores the list of user-defined properties whose subroutine
20221 * definitions aren't known at this time, or no entry if none. */
20225 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
20227 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
20228 assert(! (ANYOF_FLAGS(node)
20229 & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP));
20230 ARG_SET(node, ANYOF_ONLY_HAS_BITMAP);
20233 AV * const av = newAV();
20237 av_store(av, INVLIST_INDEX, SvREFCNT_inc_NN(cp_list));
20240 /* (Note that if any of this changes, the size calculations in
20241 * S_optimize_regclass() might need to be updated.) */
20243 if (only_utf8_locale_list) {
20244 av_store(av, ONLY_LOCALE_MATCHES_INDEX,
20245 SvREFCNT_inc_NN(only_utf8_locale_list));
20248 if (runtime_defns) {
20249 av_store(av, DEFERRED_USER_DEFINED_INDEX,
20250 SvREFCNT_inc_NN(runtime_defns));
20253 rv = newRV_noinc(MUTABLE_SV(av));
20254 n = add_data(pRExC_state, STR_WITH_LEN("s"));
20255 RExC_rxi->data->data[n] = (void*)rv;
20262 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
20263 Perl_get_regclass_nonbitmap_data(pTHX_ const regexp *prog, const regnode* node, bool doinit, SV** listsvp, SV** only_utf8_locale_ptr, SV** output_invlist)
20265 Perl_get_re_gclass_nonbitmap_data(pTHX_ const regexp *prog, const regnode* node, bool doinit, SV** listsvp, SV** only_utf8_locale_ptr, SV** output_invlist)
20269 /* For internal core use only.
20270 * Returns the inversion list for the input 'node' in the regex 'prog'.
20271 * If <doinit> is 'true', will attempt to create the inversion list if not
20273 * If <listsvp> is non-null, will return the printable contents of the
20274 * property definition. This can be used to get debugging information
20275 * even before the inversion list exists, by calling this function with
20276 * 'doinit' set to false, in which case the components that will be used
20277 * to eventually create the inversion list are returned (in a printable
20279 * If <only_utf8_locale_ptr> is not NULL, it is where this routine is to
20280 * store an inversion list of code points that should match only if the
20281 * execution-time locale is a UTF-8 one.
20282 * If <output_invlist> is not NULL, it is where this routine is to store an
20283 * inversion list of the code points that would be instead returned in
20284 * <listsvp> if this were NULL. Thus, what gets output in <listsvp>
20285 * when this parameter is used, is just the non-code point data that
20286 * will go into creating the inversion list. This currently should be just
20287 * user-defined properties whose definitions were not known at compile
20288 * time. Using this parameter allows for easier manipulation of the
20289 * inversion list's data by the caller. It is illegal to call this
20290 * function with this parameter set, but not <listsvp>
20292 * Tied intimately to how S_set_ANYOF_arg sets up the data structure. Note
20293 * that, in spite of this function's name, the inversion list it returns
20294 * may include the bitmap data as well */
20296 SV *si = NULL; /* Input initialization string */
20297 SV* invlist = NULL;
20299 RXi_GET_DECL(prog, progi);
20300 const struct reg_data * const data = prog ? progi->data : NULL;
20302 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
20303 PERL_ARGS_ASSERT_GET_REGCLASS_NONBITMAP_DATA;
20305 PERL_ARGS_ASSERT_GET_RE_GCLASS_NONBITMAP_DATA;
20307 assert(! output_invlist || listsvp);
20309 if (data && data->count) {
20310 const U32 n = ARG(node);
20312 if (data->what[n] == 's') {
20313 SV * const rv = MUTABLE_SV(data->data[n]);
20314 AV * const av = MUTABLE_AV(SvRV(rv));
20315 SV **const ary = AvARRAY(av);
20317 invlist = ary[INVLIST_INDEX];
20319 if (av_tindex_skip_len_mg(av) >= ONLY_LOCALE_MATCHES_INDEX) {
20320 *only_utf8_locale_ptr = ary[ONLY_LOCALE_MATCHES_INDEX];
20323 if (av_tindex_skip_len_mg(av) >= DEFERRED_USER_DEFINED_INDEX) {
20324 si = ary[DEFERRED_USER_DEFINED_INDEX];
20327 if (doinit && (si || invlist)) {
20330 SV * msg = newSVpvs_flags("", SVs_TEMP);
20332 SV * prop_definition = handle_user_defined_property(
20333 "", 0, FALSE, /* There is no \p{}, \P{} */
20334 SvPVX_const(si)[1] - '0', /* /i or not has been
20335 stored here for just
20337 TRUE, /* run time */
20338 FALSE, /* This call must find the defn */
20339 si, /* The property definition */
20342 0 /* base level call */
20346 assert(prop_definition == NULL);
20348 Perl_croak(aTHX_ "%" UTF8f,
20349 UTF8fARG(SvUTF8(msg), SvCUR(msg), SvPVX(msg)));
20353 _invlist_union(invlist, prop_definition, &invlist);
20354 SvREFCNT_dec_NN(prop_definition);
20357 invlist = prop_definition;
20360 STATIC_ASSERT_STMT(ONLY_LOCALE_MATCHES_INDEX == 1 + INVLIST_INDEX);
20361 STATIC_ASSERT_STMT(DEFERRED_USER_DEFINED_INDEX == 1 + ONLY_LOCALE_MATCHES_INDEX);
20363 ary[INVLIST_INDEX] = invlist;
20364 av_fill(av, (ary[ONLY_LOCALE_MATCHES_INDEX])
20365 ? ONLY_LOCALE_MATCHES_INDEX
20373 /* If requested, return a printable version of what this ANYOF node matches
20376 SV* matches_string = NULL;
20378 /* This function can be called at compile-time, before everything gets
20379 * resolved, in which case we return the currently best available
20380 * information, which is the string that will eventually be used to do
20381 * that resolving, 'si' */
20383 /* Here, we only have 'si' (and possibly some passed-in data in
20384 * 'invlist', which is handled below) If the caller only wants
20385 * 'si', use that. */
20386 if (! output_invlist) {
20387 matches_string = newSVsv(si);
20390 /* But if the caller wants an inversion list of the node, we
20391 * need to parse 'si' and place as much as possible in the
20392 * desired output inversion list, making 'matches_string' only
20393 * contain the currently unresolvable things */
20394 const char *si_string = SvPVX(si);
20395 STRLEN remaining = SvCUR(si);
20399 /* Ignore everything before and including the first new-line */
20400 si_string = (const char *) memchr(si_string, '\n', SvCUR(si));
20401 assert (si_string != NULL);
20403 remaining = SvPVX(si) + SvCUR(si) - si_string;
20405 while (remaining > 0) {
20407 /* The data consists of just strings defining user-defined
20408 * property names, but in prior incarnations, and perhaps
20409 * somehow from pluggable regex engines, it could still
20410 * hold hex code point definitions, all of which should be
20411 * legal (or it wouldn't have gotten this far). Each
20412 * component of a range would be separated by a tab, and
20413 * each range by a new-line. If these are found, instead
20414 * add them to the inversion list */
20415 I32 grok_flags = PERL_SCAN_SILENT_ILLDIGIT
20416 |PERL_SCAN_SILENT_NON_PORTABLE;
20417 STRLEN len = remaining;
20418 UV cp = grok_hex(si_string, &len, &grok_flags, NULL);
20420 /* If the hex decode routine found something, it should go
20421 * up to the next \n */
20422 if ( *(si_string + len) == '\n') {
20423 if (count) { /* 2nd code point on line */
20424 *output_invlist = _add_range_to_invlist(*output_invlist, prev_cp, cp);
20427 *output_invlist = add_cp_to_invlist(*output_invlist, cp);
20430 goto prepare_for_next_iteration;
20433 /* If the hex decode was instead for the lower range limit,
20434 * save it, and go parse the upper range limit */
20435 if (*(si_string + len) == '\t') {
20436 assert(count == 0);
20440 prepare_for_next_iteration:
20441 si_string += len + 1;
20442 remaining -= len + 1;
20446 /* Here, didn't find a legal hex number. Just add the text
20447 * from here up to the next \n, omitting any trailing
20451 len = strcspn(si_string,
20452 DEFERRED_COULD_BE_OFFICIAL_MARKERs "\n");
20454 if (matches_string) {
20455 sv_catpvn(matches_string, si_string, len);
20458 matches_string = newSVpvn(si_string, len);
20460 sv_catpvs(matches_string, " ");
20464 && UCHARAT(si_string)
20465 == DEFERRED_COULD_BE_OFFICIAL_MARKERc)
20470 if (remaining && UCHARAT(si_string) == '\n') {
20474 } /* end of loop through the text */
20476 assert(matches_string);
20477 if (SvCUR(matches_string)) { /* Get rid of trailing blank */
20478 SvCUR_set(matches_string, SvCUR(matches_string) - 1);
20480 } /* end of has an 'si' */
20483 /* Add the stuff that's already known */
20486 /* Again, if the caller doesn't want the output inversion list, put
20487 * everything in 'matches-string' */
20488 if (! output_invlist) {
20489 if ( ! matches_string) {
20490 matches_string = newSVpvs("\n");
20492 sv_catsv(matches_string, invlist_contents(invlist,
20493 TRUE /* traditional style */
20496 else if (! *output_invlist) {
20497 *output_invlist = invlist_clone(invlist, NULL);
20500 _invlist_union(*output_invlist, invlist, output_invlist);
20504 *listsvp = matches_string;
20510 /* reg_skipcomment()
20512 Absorbs an /x style # comment from the input stream,
20513 returning a pointer to the first character beyond the comment, or if the
20514 comment terminates the pattern without anything following it, this returns
20515 one past the final character of the pattern (in other words, RExC_end) and
20516 sets the REG_RUN_ON_COMMENT_SEEN flag.
20518 Note it's the callers responsibility to ensure that we are
20519 actually in /x mode
20523 PERL_STATIC_INLINE char*
20524 S_reg_skipcomment(RExC_state_t *pRExC_state, char* p)
20526 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
20530 while (p < RExC_end) {
20531 if (*(++p) == '\n') {
20536 /* we ran off the end of the pattern without ending the comment, so we have
20537 * to add an \n when wrapping */
20538 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
20543 S_skip_to_be_ignored_text(pTHX_ RExC_state_t *pRExC_state,
20545 const bool force_to_xmod
20548 /* If the text at the current parse position '*p' is a '(?#...)' comment,
20549 * or if we are under /x or 'force_to_xmod' is TRUE, and the text at '*p'
20550 * is /x whitespace, advance '*p' so that on exit it points to the first
20551 * byte past all such white space and comments */
20553 const bool use_xmod = force_to_xmod || (RExC_flags & RXf_PMf_EXTENDED);
20555 PERL_ARGS_ASSERT_SKIP_TO_BE_IGNORED_TEXT;
20557 assert( ! UTF || UTF8_IS_INVARIANT(**p) || UTF8_IS_START(**p));
20560 if (RExC_end - (*p) >= 3
20562 && *(*p + 1) == '?'
20563 && *(*p + 2) == '#')
20565 while (*(*p) != ')') {
20566 if ((*p) == RExC_end)
20567 FAIL("Sequence (?#... not terminated");
20575 const char * save_p = *p;
20576 while ((*p) < RExC_end) {
20578 if ((len = is_PATWS_safe((*p), RExC_end, UTF))) {
20581 else if (*(*p) == '#') {
20582 (*p) = reg_skipcomment(pRExC_state, (*p));
20588 if (*p != save_p) {
20601 Advances the parse position by one byte, unless that byte is the beginning
20602 of a '(?#...)' style comment, or is /x whitespace and /x is in effect. In
20603 those two cases, the parse position is advanced beyond all such comments and
20606 This is the UTF, (?#...), and /x friendly way of saying RExC_parse++.
20610 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
20612 PERL_ARGS_ASSERT_NEXTCHAR;
20614 if (RExC_parse < RExC_end) {
20616 || UTF8_IS_INVARIANT(*RExC_parse)
20617 || UTF8_IS_START(*RExC_parse));
20619 RExC_parse += (UTF)
20620 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
20623 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
20624 FALSE /* Don't force /x */ );
20629 S_change_engine_size(pTHX_ RExC_state_t *pRExC_state, const Ptrdiff_t size)
20631 /* 'size' is the delta number of smallest regnode equivalents to add or
20632 * subtract from the current memory allocated to the regex engine being
20635 PERL_ARGS_ASSERT_CHANGE_ENGINE_SIZE;
20640 sizeof(regexp_internal) + (RExC_size + 1) * sizeof(regnode),
20641 /* +1 for REG_MAGIC */
20644 if ( RExC_rxi == NULL )
20645 FAIL("Regexp out of space");
20646 RXi_SET(RExC_rx, RExC_rxi);
20648 RExC_emit_start = RExC_rxi->program;
20650 Zero(REGNODE_p(RExC_emit), size, regnode);
20653 #ifdef RE_TRACK_PATTERN_OFFSETS
20654 Renew(RExC_offsets, 2*RExC_size+1, U32);
20656 Zero(RExC_offsets + 2*(RExC_size - size) + 1, 2 * size, U32);
20658 RExC_offsets[0] = RExC_size;
20662 STATIC regnode_offset
20663 S_regnode_guts(pTHX_ RExC_state_t *pRExC_state, const U8 op, const STRLEN extra_size, const char* const name)
20665 /* Allocate a regnode for 'op', with 'extra_size' extra (smallest) regnode
20666 * equivalents space. It aligns and increments RExC_size
20668 * It returns the regnode's offset into the regex engine program */
20670 const regnode_offset ret = RExC_emit;
20672 DECLARE_AND_GET_RE_DEBUG_FLAGS;
20674 PERL_ARGS_ASSERT_REGNODE_GUTS;
20676 SIZE_ALIGN(RExC_size);
20677 change_engine_size(pRExC_state, (Ptrdiff_t) 1 + extra_size);
20678 NODE_ALIGN_FILL(REGNODE_p(ret));
20679 #ifndef RE_TRACK_PATTERN_OFFSETS
20680 PERL_UNUSED_ARG(name);
20681 PERL_UNUSED_ARG(op);
20683 assert(extra_size >= regarglen[op] || PL_regkind[op] == ANYOF);
20685 if (RExC_offsets) { /* MJD */
20687 ("%s:%d: (op %s) %s %" UVuf " (len %" UVuf ") (max %" UVuf ").\n",
20690 (UV)(RExC_emit) > RExC_offsets[0]
20691 ? "Overwriting end of array!\n" : "OK",
20693 (UV)(RExC_parse - RExC_start),
20694 (UV)RExC_offsets[0]));
20695 Set_Node_Offset(REGNODE_p(RExC_emit), RExC_parse + (op == END));
20702 - reg_node - emit a node
20704 STATIC regnode_offset /* Location. */
20705 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
20707 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reg_node");
20708 regnode_offset ptr = ret;
20710 PERL_ARGS_ASSERT_REG_NODE;
20712 assert(regarglen[op] == 0);
20714 FILL_ADVANCE_NODE(ptr, op);
20720 - reganode - emit a node with an argument
20722 STATIC regnode_offset /* Location. */
20723 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
20725 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reganode");
20726 regnode_offset ptr = ret;
20728 PERL_ARGS_ASSERT_REGANODE;
20730 /* ANYOF are special cased to allow non-length 1 args */
20731 assert(regarglen[op] == 1);
20733 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
20739 - regpnode - emit a temporary node with a SV* argument
20741 STATIC regnode_offset /* Location. */
20742 S_regpnode(pTHX_ RExC_state_t *pRExC_state, U8 op, SV * arg)
20744 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "regpnode");
20745 regnode_offset ptr = ret;
20747 PERL_ARGS_ASSERT_REGPNODE;
20749 FILL_ADVANCE_NODE_ARGp(ptr, op, arg);
20754 STATIC regnode_offset
20755 S_reg2Lanode(pTHX_ RExC_state_t *pRExC_state, const U8 op, const U32 arg1, const I32 arg2)
20757 /* emit a node with U32 and I32 arguments */
20759 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reg2Lanode");
20760 regnode_offset ptr = ret;
20762 PERL_ARGS_ASSERT_REG2LANODE;
20764 assert(regarglen[op] == 2);
20766 FILL_ADVANCE_NODE_2L_ARG(ptr, op, arg1, arg2);
20772 - reginsert - insert an operator in front of already-emitted operand
20774 * That means that on exit 'operand' is the offset of the newly inserted
20775 * operator, and the original operand has been relocated.
20777 * IMPORTANT NOTE - it is the *callers* responsibility to correctly
20778 * set up NEXT_OFF() of the inserted node if needed. Something like this:
20780 * reginsert(pRExC, OPFAIL, orig_emit, depth+1);
20781 * NEXT_OFF(orig_emit) = regarglen[OPFAIL] + NODE_STEP_REGNODE;
20783 * ALSO NOTE - FLAGS(newly-inserted-operator) will be set to 0 as well.
20786 S_reginsert(pTHX_ RExC_state_t *pRExC_state, const U8 op,
20787 const regnode_offset operand, const U32 depth)
20792 const int offset = regarglen[(U8)op];
20793 const int size = NODE_STEP_REGNODE + offset;
20794 DECLARE_AND_GET_RE_DEBUG_FLAGS;
20796 PERL_ARGS_ASSERT_REGINSERT;
20797 PERL_UNUSED_CONTEXT;
20798 PERL_UNUSED_ARG(depth);
20799 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
20800 DEBUG_PARSE_FMT("inst"," - %s", PL_reg_name[op]);
20801 assert(!RExC_study_started); /* I believe we should never use reginsert once we have started
20802 studying. If this is wrong then we need to adjust RExC_recurse
20803 below like we do with RExC_open_parens/RExC_close_parens. */
20804 change_engine_size(pRExC_state, (Ptrdiff_t) size);
20805 src = REGNODE_p(RExC_emit);
20807 dst = REGNODE_p(RExC_emit);
20809 /* If we are in a "count the parentheses" pass, the numbers are unreliable,
20810 * and [perl #133871] shows this can lead to problems, so skip this
20811 * realignment of parens until a later pass when they are reliable */
20812 if (! IN_PARENS_PASS && RExC_open_parens) {
20814 /*DEBUG_PARSE_FMT("inst"," - %" IVdf, (IV)RExC_npar);*/
20815 /* remember that RExC_npar is rex->nparens + 1,
20816 * iow it is 1 more than the number of parens seen in
20817 * the pattern so far. */
20818 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
20819 /* note, RExC_open_parens[0] is the start of the
20820 * regex, it can't move. RExC_close_parens[0] is the end
20821 * of the regex, it *can* move. */
20822 if ( paren && RExC_open_parens[paren] >= operand ) {
20823 /*DEBUG_PARSE_FMT("open"," - %d", size);*/
20824 RExC_open_parens[paren] += size;
20826 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
20828 if ( RExC_close_parens[paren] >= operand ) {
20829 /*DEBUG_PARSE_FMT("close"," - %d", size);*/
20830 RExC_close_parens[paren] += size;
20832 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
20837 RExC_end_op += size;
20839 while (src > REGNODE_p(operand)) {
20840 StructCopy(--src, --dst, regnode);
20841 #ifdef RE_TRACK_PATTERN_OFFSETS
20842 if (RExC_offsets) { /* MJD 20010112 */
20844 ("%s(%d): (op %s) %s copy %" UVuf " -> %" UVuf " (max %" UVuf ").\n",
20848 (UV)(REGNODE_OFFSET(dst)) > RExC_offsets[0]
20849 ? "Overwriting end of array!\n" : "OK",
20850 (UV)REGNODE_OFFSET(src),
20851 (UV)REGNODE_OFFSET(dst),
20852 (UV)RExC_offsets[0]));
20853 Set_Node_Offset_To_R(REGNODE_OFFSET(dst), Node_Offset(src));
20854 Set_Node_Length_To_R(REGNODE_OFFSET(dst), Node_Length(src));
20859 place = REGNODE_p(operand); /* Op node, where operand used to be. */
20860 #ifdef RE_TRACK_PATTERN_OFFSETS
20861 if (RExC_offsets) { /* MJD */
20863 ("%s(%d): (op %s) %s %" UVuf " <- %" UVuf " (max %" UVuf ").\n",
20867 (UV)REGNODE_OFFSET(place) > RExC_offsets[0]
20868 ? "Overwriting end of array!\n" : "OK",
20869 (UV)REGNODE_OFFSET(place),
20870 (UV)(RExC_parse - RExC_start),
20871 (UV)RExC_offsets[0]));
20872 Set_Node_Offset(place, RExC_parse);
20873 Set_Node_Length(place, 1);
20876 src = NEXTOPER(place);
20878 FILL_NODE(operand, op);
20880 /* Zero out any arguments in the new node */
20881 Zero(src, offset, regnode);
20885 - regtail - set the next-pointer at the end of a node chain of p to val. If
20886 that value won't fit in the space available, instead returns FALSE.
20887 (Except asserts if we can't fit in the largest space the regex
20888 engine is designed for.)
20889 - SEE ALSO: regtail_study
20892 S_regtail(pTHX_ RExC_state_t * pRExC_state,
20893 const regnode_offset p,
20894 const regnode_offset val,
20897 regnode_offset scan;
20898 DECLARE_AND_GET_RE_DEBUG_FLAGS;
20900 PERL_ARGS_ASSERT_REGTAIL;
20902 PERL_UNUSED_ARG(depth);
20905 /* The final node in the chain is the first one with a nonzero next pointer
20907 scan = (regnode_offset) p;
20909 regnode * const temp = regnext(REGNODE_p(scan));
20911 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
20912 regprop(RExC_rx, RExC_mysv, REGNODE_p(scan), NULL, pRExC_state);
20913 Perl_re_printf( aTHX_ "~ %s (%zu) %s %s\n",
20914 SvPV_nolen_const(RExC_mysv), scan,
20915 (temp == NULL ? "->" : ""),
20916 (temp == NULL ? PL_reg_name[OP(REGNODE_p(val))] : "")
20921 scan = REGNODE_OFFSET(temp);
20924 /* Populate this node's next pointer */
20925 assert(val >= scan);
20926 if (reg_off_by_arg[OP(REGNODE_p(scan))]) {
20927 assert((UV) (val - scan) <= U32_MAX);
20928 ARG_SET(REGNODE_p(scan), val - scan);
20931 if (val - scan > U16_MAX) {
20932 /* Populate this with something that won't loop and will likely
20933 * lead to a crash if the caller ignores the failure return, and
20934 * execution continues */
20935 NEXT_OFF(REGNODE_p(scan)) = U16_MAX;
20938 NEXT_OFF(REGNODE_p(scan)) = val - scan;
20946 - regtail_study - set the next-pointer at the end of a node chain of p to val.
20947 - Look for optimizable sequences at the same time.
20948 - currently only looks for EXACT chains.
20950 This is experimental code. The idea is to use this routine to perform
20951 in place optimizations on branches and groups as they are constructed,
20952 with the long term intention of removing optimization from study_chunk so
20953 that it is purely analytical.
20955 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
20956 to control which is which.
20958 This used to return a value that was ignored. It was a problem that it is
20959 #ifdef'd to be another function that didn't return a value. khw has changed it
20960 so both currently return a pass/fail return.
20963 /* TODO: All four parms should be const */
20966 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode_offset p,
20967 const regnode_offset val, U32 depth)
20969 regnode_offset scan;
20971 #ifdef EXPERIMENTAL_INPLACESCAN
20974 DECLARE_AND_GET_RE_DEBUG_FLAGS;
20976 PERL_ARGS_ASSERT_REGTAIL_STUDY;
20979 /* Find last node. */
20983 regnode * const temp = regnext(REGNODE_p(scan));
20984 #ifdef EXPERIMENTAL_INPLACESCAN
20985 if (PL_regkind[OP(REGNODE_p(scan))] == EXACT) {
20986 bool unfolded_multi_char; /* Unexamined in this routine */
20987 if (join_exact(pRExC_state, scan, &min,
20988 &unfolded_multi_char, 1, REGNODE_p(val), depth+1))
20989 return TRUE; /* Was return EXACT */
20993 if (PL_regkind[OP(REGNODE_p(scan))] == EXACT) {
20994 if (exact == PSEUDO )
20995 exact= OP(REGNODE_p(scan));
20996 else if (exact != OP(REGNODE_p(scan)) )
20999 else if (OP(REGNODE_p(scan)) != NOTHING) {
21004 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
21005 regprop(RExC_rx, RExC_mysv, REGNODE_p(scan), NULL, pRExC_state);
21006 Perl_re_printf( aTHX_ "~ %s (%zu) -> %s\n",
21007 SvPV_nolen_const(RExC_mysv),
21009 PL_reg_name[exact]);
21013 scan = REGNODE_OFFSET(temp);
21016 DEBUG_PARSE_MSG("");
21017 regprop(RExC_rx, RExC_mysv, REGNODE_p(val), NULL, pRExC_state);
21018 Perl_re_printf( aTHX_
21019 "~ attach to %s (%" IVdf ") offset to %" IVdf "\n",
21020 SvPV_nolen_const(RExC_mysv),
21025 if (reg_off_by_arg[OP(REGNODE_p(scan))]) {
21026 assert((UV) (val - scan) <= U32_MAX);
21027 ARG_SET(REGNODE_p(scan), val - scan);
21030 if (val - scan > U16_MAX) {
21031 /* Populate this with something that won't loop and will likely
21032 * lead to a crash if the caller ignores the failure return, and
21033 * execution continues */
21034 NEXT_OFF(REGNODE_p(scan)) = U16_MAX;
21037 NEXT_OFF(REGNODE_p(scan)) = val - scan;
21040 return TRUE; /* Was 'return exact' */
21045 S_get_ANYOFM_contents(pTHX_ const regnode * n) {
21047 /* Returns an inversion list of all the code points matched by the
21048 * ANYOFM/NANYOFM node 'n' */
21050 SV * cp_list = _new_invlist(-1);
21051 const U8 lowest = (U8) ARG(n);
21054 U8 needed = 1U << PL_bitcount[ (U8) ~ FLAGS(n)];
21056 PERL_ARGS_ASSERT_GET_ANYOFM_CONTENTS;
21058 /* Starting with the lowest code point, any code point that ANDed with the
21059 * mask yields the lowest code point is in the set */
21060 for (i = lowest; i <= 0xFF; i++) {
21061 if ((i & FLAGS(n)) == ARG(n)) {
21062 cp_list = add_cp_to_invlist(cp_list, i);
21065 /* We know how many code points (a power of two) that are in the
21066 * set. No use looking once we've got that number */
21067 if (count >= needed) break;
21071 if (OP(n) == NANYOFM) {
21072 _invlist_invert(cp_list);
21078 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
21083 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
21088 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
21090 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
21091 if (flags & (1<<bit)) {
21092 if (!set++ && lead)
21093 Perl_re_printf( aTHX_ "%s", lead);
21094 Perl_re_printf( aTHX_ "%s ", PL_reg_intflags_name[bit]);
21099 Perl_re_printf( aTHX_ "\n");
21101 Perl_re_printf( aTHX_ "%s[none-set]\n", lead);
21106 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
21112 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
21114 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
21115 if (flags & (1<<bit)) {
21116 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
21119 if (!set++ && lead)
21120 Perl_re_printf( aTHX_ "%s", lead);
21121 Perl_re_printf( aTHX_ "%s ", PL_reg_extflags_name[bit]);
21124 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
21125 if (!set++ && lead) {
21126 Perl_re_printf( aTHX_ "%s", lead);
21129 case REGEX_UNICODE_CHARSET:
21130 Perl_re_printf( aTHX_ "UNICODE");
21132 case REGEX_LOCALE_CHARSET:
21133 Perl_re_printf( aTHX_ "LOCALE");
21135 case REGEX_ASCII_RESTRICTED_CHARSET:
21136 Perl_re_printf( aTHX_ "ASCII-RESTRICTED");
21138 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
21139 Perl_re_printf( aTHX_ "ASCII-MORE_RESTRICTED");
21142 Perl_re_printf( aTHX_ "UNKNOWN CHARACTER SET");
21148 Perl_re_printf( aTHX_ "\n");
21150 Perl_re_printf( aTHX_ "%s[none-set]\n", lead);
21156 Perl_regdump(pTHX_ const regexp *r)
21160 SV * const sv = sv_newmortal();
21161 SV *dsv= sv_newmortal();
21162 RXi_GET_DECL(r, ri);
21163 DECLARE_AND_GET_RE_DEBUG_FLAGS;
21165 PERL_ARGS_ASSERT_REGDUMP;
21167 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
21169 /* Header fields of interest. */
21170 for (i = 0; i < 2; i++) {
21171 if (r->substrs->data[i].substr) {
21172 RE_PV_QUOTED_DECL(s, 0, dsv,
21173 SvPVX_const(r->substrs->data[i].substr),
21174 RE_SV_DUMPLEN(r->substrs->data[i].substr),
21175 PL_dump_re_max_len);
21176 Perl_re_printf( aTHX_
21177 "%s %s%s at %" IVdf "..%" UVuf " ",
21178 i ? "floating" : "anchored",
21180 RE_SV_TAIL(r->substrs->data[i].substr),
21181 (IV)r->substrs->data[i].min_offset,
21182 (UV)r->substrs->data[i].max_offset);
21184 else if (r->substrs->data[i].utf8_substr) {
21185 RE_PV_QUOTED_DECL(s, 1, dsv,
21186 SvPVX_const(r->substrs->data[i].utf8_substr),
21187 RE_SV_DUMPLEN(r->substrs->data[i].utf8_substr),
21189 Perl_re_printf( aTHX_
21190 "%s utf8 %s%s at %" IVdf "..%" UVuf " ",
21191 i ? "floating" : "anchored",
21193 RE_SV_TAIL(r->substrs->data[i].utf8_substr),
21194 (IV)r->substrs->data[i].min_offset,
21195 (UV)r->substrs->data[i].max_offset);
21199 if (r->check_substr || r->check_utf8)
21200 Perl_re_printf( aTHX_
21202 ( r->check_substr == r->substrs->data[1].substr
21203 && r->check_utf8 == r->substrs->data[1].utf8_substr
21204 ? "(checking floating" : "(checking anchored"));
21205 if (r->intflags & PREGf_NOSCAN)
21206 Perl_re_printf( aTHX_ " noscan");
21207 if (r->extflags & RXf_CHECK_ALL)
21208 Perl_re_printf( aTHX_ " isall");
21209 if (r->check_substr || r->check_utf8)
21210 Perl_re_printf( aTHX_ ") ");
21212 if (ri->regstclass) {
21213 regprop(r, sv, ri->regstclass, NULL, NULL);
21214 Perl_re_printf( aTHX_ "stclass %s ", SvPVX_const(sv));
21216 if (r->intflags & PREGf_ANCH) {
21217 Perl_re_printf( aTHX_ "anchored");
21218 if (r->intflags & PREGf_ANCH_MBOL)
21219 Perl_re_printf( aTHX_ "(MBOL)");
21220 if (r->intflags & PREGf_ANCH_SBOL)
21221 Perl_re_printf( aTHX_ "(SBOL)");
21222 if (r->intflags & PREGf_ANCH_GPOS)
21223 Perl_re_printf( aTHX_ "(GPOS)");
21224 Perl_re_printf( aTHX_ " ");
21226 if (r->intflags & PREGf_GPOS_SEEN)
21227 Perl_re_printf( aTHX_ "GPOS:%" UVuf " ", (UV)r->gofs);
21228 if (r->intflags & PREGf_SKIP)
21229 Perl_re_printf( aTHX_ "plus ");
21230 if (r->intflags & PREGf_IMPLICIT)
21231 Perl_re_printf( aTHX_ "implicit ");
21232 Perl_re_printf( aTHX_ "minlen %" IVdf " ", (IV)r->minlen);
21233 if (r->extflags & RXf_EVAL_SEEN)
21234 Perl_re_printf( aTHX_ "with eval ");
21235 Perl_re_printf( aTHX_ "\n");
21237 regdump_extflags("r->extflags: ", r->extflags);
21238 regdump_intflags("r->intflags: ", r->intflags);
21241 PERL_ARGS_ASSERT_REGDUMP;
21242 PERL_UNUSED_CONTEXT;
21243 PERL_UNUSED_ARG(r);
21244 #endif /* DEBUGGING */
21247 /* Should be synchronized with ANYOF_ #defines in regcomp.h */
21250 # if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 \
21251 || _CC_LOWER != 3 || _CC_UPPER != 4 || _CC_PUNCT != 5 \
21252 || _CC_PRINT != 6 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 \
21253 || _CC_CASED != 9 || _CC_SPACE != 10 || _CC_BLANK != 11 \
21254 || _CC_XDIGIT != 12 || _CC_CNTRL != 13 || _CC_ASCII != 14 \
21255 || _CC_VERTSPACE != 15
21256 # error Need to adjust order of anyofs[]
21258 static const char * const anyofs[] = {
21295 - regprop - printable representation of opcode, with run time support
21299 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo, const RExC_state_t *pRExC_state)
21303 RXi_GET_DECL(prog, progi);
21304 DECLARE_AND_GET_RE_DEBUG_FLAGS;
21306 PERL_ARGS_ASSERT_REGPROP;
21310 if (OP(o) > REGNODE_MAX) { /* regnode.type is unsigned */
21311 if (pRExC_state) { /* This gives more info, if we have it */
21312 FAIL3("panic: corrupted regexp opcode %d > %d",
21313 (int)OP(o), (int)REGNODE_MAX);
21316 Perl_croak(aTHX_ "panic: corrupted regexp opcode %d > %d",
21317 (int)OP(o), (int)REGNODE_MAX);
21320 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
21322 k = PL_regkind[OP(o)];
21325 sv_catpvs(sv, " ");
21326 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
21327 * is a crude hack but it may be the best for now since
21328 * we have no flag "this EXACTish node was UTF-8"
21330 pv_pretty(sv, STRING(o), STR_LEN(o), PL_dump_re_max_len,
21331 PL_colors[0], PL_colors[1],
21332 PERL_PV_ESCAPE_UNI_DETECT |
21333 PERL_PV_ESCAPE_NONASCII |
21334 PERL_PV_PRETTY_ELLIPSES |
21335 PERL_PV_PRETTY_LTGT |
21336 PERL_PV_PRETTY_NOCLEAR
21338 } else if (k == TRIE) {
21339 /* print the details of the trie in dumpuntil instead, as
21340 * progi->data isn't available here */
21341 const char op = OP(o);
21342 const U32 n = ARG(o);
21343 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
21344 (reg_ac_data *)progi->data->data[n] :
21346 const reg_trie_data * const trie
21347 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
21349 Perl_sv_catpvf(aTHX_ sv, "-%s", PL_reg_name[o->flags]);
21350 DEBUG_TRIE_COMPILE_r({
21352 sv_catpvs(sv, "(JUMP)");
21353 Perl_sv_catpvf(aTHX_ sv,
21354 "<S:%" UVuf "/%" IVdf " W:%" UVuf " L:%" UVuf "/%" UVuf " C:%" UVuf "/%" UVuf ">",
21355 (UV)trie->startstate,
21356 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
21357 (UV)trie->wordcount,
21360 (UV)TRIE_CHARCOUNT(trie),
21361 (UV)trie->uniquecharcount
21364 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
21365 sv_catpvs(sv, "[");
21366 (void) put_charclass_bitmap_innards(sv,
21367 ((IS_ANYOF_TRIE(op))
21369 : TRIE_BITMAP(trie)),
21376 sv_catpvs(sv, "]");
21378 } else if (k == CURLY) {
21379 U32 lo = ARG1(o), hi = ARG2(o);
21380 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
21381 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
21382 Perl_sv_catpvf(aTHX_ sv, "{%u,", (unsigned) lo);
21383 if (hi == REG_INFTY)
21384 sv_catpvs(sv, "INFTY");
21386 Perl_sv_catpvf(aTHX_ sv, "%u", (unsigned) hi);
21387 sv_catpvs(sv, "}");
21389 else if (k == WHILEM && o->flags) /* Ordinal/of */
21390 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
21391 else if (k == REF || k == OPEN || k == CLOSE
21392 || k == GROUPP || OP(o)==ACCEPT)
21394 AV *name_list= NULL;
21395 U32 parno= OP(o) == ACCEPT ? (U32)ARG2L(o) : ARG(o);
21396 Perl_sv_catpvf(aTHX_ sv, "%" UVuf, (UV)parno); /* Parenth number */
21397 if ( RXp_PAREN_NAMES(prog) ) {
21398 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
21399 } else if ( pRExC_state ) {
21400 name_list= RExC_paren_name_list;
21403 if ( k != REF || (OP(o) < REFN)) {
21404 SV **name= av_fetch(name_list, parno, 0 );
21406 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
21409 SV *sv_dat= MUTABLE_SV(progi->data->data[ parno ]);
21410 I32 *nums=(I32*)SvPVX(sv_dat);
21411 SV **name= av_fetch(name_list, nums[0], 0 );
21414 for ( n=0; n<SvIVX(sv_dat); n++ ) {
21415 Perl_sv_catpvf(aTHX_ sv, "%s%" IVdf,
21416 (n ? "," : ""), (IV)nums[n]);
21418 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
21422 if ( k == REF && reginfo) {
21423 U32 n = ARG(o); /* which paren pair */
21424 I32 ln = prog->offs[n].start;
21425 if (prog->lastparen < n || ln == -1 || prog->offs[n].end == -1)
21426 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
21427 else if (ln == prog->offs[n].end)
21428 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
21430 const char *s = reginfo->strbeg + ln;
21431 Perl_sv_catpvf(aTHX_ sv, ": ");
21432 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
21433 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
21436 } else if (k == GOSUB) {
21437 AV *name_list= NULL;
21438 if ( RXp_PAREN_NAMES(prog) ) {
21439 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
21440 } else if ( pRExC_state ) {
21441 name_list= RExC_paren_name_list;
21444 /* Paren and offset */
21445 Perl_sv_catpvf(aTHX_ sv, "%d[%+d:%d]", (int)ARG(o),(int)ARG2L(o),
21446 (int)((o + (int)ARG2L(o)) - progi->program) );
21448 SV **name= av_fetch(name_list, ARG(o), 0 );
21450 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
21453 else if (k == LOGICAL)
21454 /* 2: embedded, otherwise 1 */
21455 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
21456 else if (k == ANYOF || k == ANYOFR) {
21460 bool do_sep = FALSE; /* Do we need to separate various components of
21462 /* Set if there is still an unresolved user-defined property */
21463 SV *unresolved = NULL;
21465 /* Things that are ignored except when the runtime locale is UTF-8 */
21466 SV *only_utf8_locale_invlist = NULL;
21468 /* Code points that don't fit in the bitmap */
21469 SV *nonbitmap_invlist = NULL;
21471 /* And things that aren't in the bitmap, but are small enough to be */
21472 SV* bitmap_range_not_in_bitmap = NULL;
21476 if (inRANGE(OP(o), ANYOFH, ANYOFRb)) {
21482 flags = ANYOF_FLAGS(o);
21483 bitmap = ANYOF_BITMAP(o);
21487 if (OP(o) == ANYOFL || OP(o) == ANYOFPOSIXL) {
21488 if (ANYOFL_UTF8_LOCALE_REQD(flags)) {
21489 sv_catpvs(sv, "{utf8-locale-reqd}");
21491 if (flags & ANYOFL_FOLD) {
21492 sv_catpvs(sv, "{i}");
21496 inverted = flags & ANYOF_INVERT;
21498 /* If there is stuff outside the bitmap, get it */
21499 if (arg != ANYOF_ONLY_HAS_BITMAP) {
21500 if (inRANGE(OP(o), ANYOFR, ANYOFRb)) {
21501 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
21503 ANYOFRbase(o) + ANYOFRdelta(o));
21506 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
21507 (void) get_regclass_nonbitmap_data(prog, o, FALSE,
21509 &only_utf8_locale_invlist,
21510 &nonbitmap_invlist);
21512 (void) get_re_gclass_nonbitmap_data(prog, o, FALSE,
21514 &only_utf8_locale_invlist,
21515 &nonbitmap_invlist);
21519 /* The non-bitmap data may contain stuff that could fit in the
21520 * bitmap. This could come from a user-defined property being
21521 * finally resolved when this call was done; or much more likely
21522 * because there are matches that require UTF-8 to be valid, and so
21523 * aren't in the bitmap (or ANYOFR). This is teased apart later */
21524 _invlist_intersection(nonbitmap_invlist,
21526 &bitmap_range_not_in_bitmap);
21527 /* Leave just the things that don't fit into the bitmap */
21528 _invlist_subtract(nonbitmap_invlist,
21530 &nonbitmap_invlist);
21533 /* Obey this flag to add all above-the-bitmap code points */
21534 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
21535 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
21536 NUM_ANYOF_CODE_POINTS,
21540 /* Ready to start outputting. First, the initial left bracket */
21541 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
21543 /* ANYOFH by definition doesn't have anything that will fit inside the
21544 * bitmap; ANYOFR may or may not. */
21545 if ( ! inRANGE(OP(o), ANYOFH, ANYOFHr)
21546 && ( ! inRANGE(OP(o), ANYOFR, ANYOFRb)
21547 || ANYOFRbase(o) < NUM_ANYOF_CODE_POINTS))
21549 /* Then all the things that could fit in the bitmap */
21550 do_sep = put_charclass_bitmap_innards(sv,
21552 bitmap_range_not_in_bitmap,
21553 only_utf8_locale_invlist,
21557 /* Can't try inverting for a
21558 * better display if there
21559 * are things that haven't
21562 || inRANGE(OP(o), ANYOFR, ANYOFRb));
21563 SvREFCNT_dec(bitmap_range_not_in_bitmap);
21565 /* If there are user-defined properties which haven't been defined
21566 * yet, output them. If the result is not to be inverted, it is
21567 * clearest to output them in a separate [] from the bitmap range
21568 * stuff. If the result is to be complemented, we have to show
21569 * everything in one [], as the inversion applies to the whole
21570 * thing. Use {braces} to separate them from anything in the
21571 * bitmap and anything above the bitmap. */
21574 if (! do_sep) { /* If didn't output anything in the bitmap
21576 sv_catpvs(sv, "^");
21578 sv_catpvs(sv, "{");
21581 Perl_sv_catpvf(aTHX_ sv,"%s][%s", PL_colors[1],
21584 sv_catsv(sv, unresolved);
21586 sv_catpvs(sv, "}");
21588 do_sep = ! inverted;
21592 /* And, finally, add the above-the-bitmap stuff */
21593 if (nonbitmap_invlist && _invlist_len(nonbitmap_invlist)) {
21596 /* See if truncation size is overridden */
21597 const STRLEN dump_len = (PL_dump_re_max_len > 256)
21598 ? PL_dump_re_max_len
21601 /* This is output in a separate [] */
21603 Perl_sv_catpvf(aTHX_ sv,"%s][%s", PL_colors[1], PL_colors[0]);
21606 /* And, for easy of understanding, it is shown in the
21607 * uncomplemented form if possible. The one exception being if
21608 * there are unresolved items, where the inversion has to be
21609 * delayed until runtime */
21610 if (inverted && ! unresolved) {
21611 _invlist_invert(nonbitmap_invlist);
21612 _invlist_subtract(nonbitmap_invlist, PL_InBitmap, &nonbitmap_invlist);
21615 contents = invlist_contents(nonbitmap_invlist,
21616 FALSE /* output suitable for catsv */
21619 /* If the output is shorter than the permissible maximum, just do it. */
21620 if (SvCUR(contents) <= dump_len) {
21621 sv_catsv(sv, contents);
21624 const char * contents_string = SvPVX(contents);
21625 STRLEN i = dump_len;
21627 /* Otherwise, start at the permissible max and work back to the
21628 * first break possibility */
21629 while (i > 0 && contents_string[i] != ' ') {
21632 if (i == 0) { /* Fail-safe. Use the max if we couldn't
21633 find a legal break */
21637 sv_catpvn(sv, contents_string, i);
21638 sv_catpvs(sv, "...");
21641 SvREFCNT_dec_NN(contents);
21642 SvREFCNT_dec_NN(nonbitmap_invlist);
21645 /* And finally the matching, closing ']' */
21646 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
21648 if (OP(o) == ANYOFHs) {
21649 Perl_sv_catpvf(aTHX_ sv, " (Leading UTF-8 bytes=%s", _byte_dump_string((U8 *) ((struct regnode_anyofhs *) o)->string, FLAGS(o), 1));
21651 else if (inRANGE(OP(o), ANYOFH, ANYOFRb)) {
21652 U8 lowest = (OP(o) != ANYOFHr)
21654 : LOWEST_ANYOF_HRx_BYTE(FLAGS(o));
21655 U8 highest = (OP(o) == ANYOFHr)
21656 ? HIGHEST_ANYOF_HRx_BYTE(FLAGS(o))
21657 : (OP(o) == ANYOFH || OP(o) == ANYOFR)
21661 if (OP(o) != ANYOFR || ! isASCII(ANYOFRbase(o) + ANYOFRdelta(o)))
21664 Perl_sv_catpvf(aTHX_ sv, " (First UTF-8 byte=%02X", lowest);
21665 if (lowest != highest) {
21666 Perl_sv_catpvf(aTHX_ sv, "-%02X", highest);
21668 Perl_sv_catpvf(aTHX_ sv, ")");
21672 SvREFCNT_dec(unresolved);
21674 else if (k == ANYOFM) {
21675 SV * cp_list = get_ANYOFM_contents(o);
21677 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
21678 if (OP(o) == NANYOFM) {
21679 _invlist_invert(cp_list);
21682 put_charclass_bitmap_innards(sv, NULL, cp_list, NULL, NULL, 0, TRUE);
21683 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
21685 SvREFCNT_dec(cp_list);
21687 else if (k == POSIXD || k == NPOSIXD) {
21688 U8 index = FLAGS(o) * 2;
21689 if (index < C_ARRAY_LENGTH(anyofs)) {
21690 if (*anyofs[index] != '[') {
21691 sv_catpvs(sv, "[");
21693 sv_catpv(sv, anyofs[index]);
21694 if (*anyofs[index] != '[') {
21695 sv_catpvs(sv, "]");
21699 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
21702 else if (k == BOUND || k == NBOUND) {
21703 /* Must be synced with order of 'bound_type' in regcomp.h */
21704 const char * const bounds[] = {
21705 "", /* Traditional */
21711 assert(FLAGS(o) < C_ARRAY_LENGTH(bounds));
21712 sv_catpv(sv, bounds[FLAGS(o)]);
21714 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH)) {
21715 Perl_sv_catpvf(aTHX_ sv, "[%d", -(o->flags));
21717 Perl_sv_catpvf(aTHX_ sv, "..-%d", o->flags - o->next_off);
21719 Perl_sv_catpvf(aTHX_ sv, "]");
21721 else if (OP(o) == SBOL)
21722 Perl_sv_catpvf(aTHX_ sv, " /%s/", o->flags ? "\\A" : "^");
21724 /* add on the verb argument if there is one */
21725 if ( ( k == VERB || OP(o) == ACCEPT || OP(o) == OPFAIL ) && o->flags) {
21727 Perl_sv_catpvf(aTHX_ sv, ":%" SVf,
21728 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
21730 sv_catpvs(sv, ":NULL");
21733 PERL_UNUSED_CONTEXT;
21734 PERL_UNUSED_ARG(sv);
21735 PERL_UNUSED_ARG(o);
21736 PERL_UNUSED_ARG(prog);
21737 PERL_UNUSED_ARG(reginfo);
21738 PERL_UNUSED_ARG(pRExC_state);
21739 #endif /* DEBUGGING */
21745 Perl_re_intuit_string(pTHX_ REGEXP * const r)
21746 { /* Assume that RE_INTUIT is set */
21747 /* Returns an SV containing a string that must appear in the target for it
21748 * to match, or NULL if nothing is known that must match.
21750 * CAUTION: the SV can be freed during execution of the regex engine */
21752 struct regexp *const prog = ReANY(r);
21753 DECLARE_AND_GET_RE_DEBUG_FLAGS;
21755 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
21756 PERL_UNUSED_CONTEXT;
21760 if (prog->maxlen > 0) {
21761 const char * const s = SvPV_nolen_const(RX_UTF8(r)
21762 ? prog->check_utf8 : prog->check_substr);
21764 if (!PL_colorset) reginitcolors();
21765 Perl_re_printf( aTHX_
21766 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
21768 RX_UTF8(r) ? "utf8 " : "",
21769 PL_colors[5], PL_colors[0],
21772 (strlen(s) > PL_dump_re_max_len ? "..." : ""));
21776 /* use UTF8 check substring if regexp pattern itself is in UTF8 */
21777 return RX_UTF8(r) ? prog->check_utf8 : prog->check_substr;
21783 handles refcounting and freeing the perl core regexp structure. When
21784 it is necessary to actually free the structure the first thing it
21785 does is call the 'free' method of the regexp_engine associated to
21786 the regexp, allowing the handling of the void *pprivate; member
21787 first. (This routine is not overridable by extensions, which is why
21788 the extensions free is called first.)
21790 See regdupe and regdupe_internal if you change anything here.
21792 #ifndef PERL_IN_XSUB_RE
21794 Perl_pregfree(pTHX_ REGEXP *r)
21800 Perl_pregfree2(pTHX_ REGEXP *rx)
21802 struct regexp *const r = ReANY(rx);
21803 DECLARE_AND_GET_RE_DEBUG_FLAGS;
21805 PERL_ARGS_ASSERT_PREGFREE2;
21810 if (r->mother_re) {
21811 ReREFCNT_dec(r->mother_re);
21813 CALLREGFREE_PVT(rx); /* free the private data */
21814 SvREFCNT_dec(RXp_PAREN_NAMES(r));
21818 for (i = 0; i < 2; i++) {
21819 SvREFCNT_dec(r->substrs->data[i].substr);
21820 SvREFCNT_dec(r->substrs->data[i].utf8_substr);
21822 Safefree(r->substrs);
21824 RX_MATCH_COPY_FREE(rx);
21825 #ifdef PERL_ANY_COW
21826 SvREFCNT_dec(r->saved_copy);
21829 SvREFCNT_dec(r->qr_anoncv);
21830 if (r->recurse_locinput)
21831 Safefree(r->recurse_locinput);
21837 Copy ssv to dsv, both of which should of type SVt_REGEXP or SVt_PVLV,
21838 except that dsv will be created if NULL.
21840 This function is used in two main ways. First to implement
21841 $r = qr/....; $s = $$r;
21843 Secondly, it is used as a hacky workaround to the structural issue of
21845 being stored in the regexp structure which is in turn stored in
21846 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
21847 could be PL_curpm in multiple contexts, and could require multiple
21848 result sets being associated with the pattern simultaneously, such
21849 as when doing a recursive match with (??{$qr})
21851 The solution is to make a lightweight copy of the regexp structure
21852 when a qr// is returned from the code executed by (??{$qr}) this
21853 lightweight copy doesn't actually own any of its data except for
21854 the starp/end and the actual regexp structure itself.
21860 Perl_reg_temp_copy(pTHX_ REGEXP *dsv, REGEXP *ssv)
21862 struct regexp *drx;
21863 struct regexp *const srx = ReANY(ssv);
21864 const bool islv = dsv && SvTYPE(dsv) == SVt_PVLV;
21866 PERL_ARGS_ASSERT_REG_TEMP_COPY;
21869 dsv = (REGEXP*) newSV_type(SVt_REGEXP);
21871 assert(SvTYPE(dsv) == SVt_REGEXP || (SvTYPE(dsv) == SVt_PVLV));
21873 /* our only valid caller, sv_setsv_flags(), should have done
21874 * a SV_CHECK_THINKFIRST_COW_DROP() by now */
21875 assert(!SvOOK(dsv));
21876 assert(!SvIsCOW(dsv));
21877 assert(!SvROK(dsv));
21879 if (SvPVX_const(dsv)) {
21881 Safefree(SvPVX(dsv));
21886 SvOK_off((SV *)dsv);
21889 /* For PVLVs, the head (sv_any) points to an XPVLV, while
21890 * the LV's xpvlenu_rx will point to a regexp body, which
21891 * we allocate here */
21892 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
21893 assert(!SvPVX(dsv));
21894 ((XPV*)SvANY(dsv))->xpv_len_u.xpvlenu_rx = temp->sv_any;
21895 temp->sv_any = NULL;
21896 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
21897 SvREFCNT_dec_NN(temp);
21898 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
21899 ing below will not set it. */
21900 SvCUR_set(dsv, SvCUR(ssv));
21903 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
21904 sv_force_normal(sv) is called. */
21908 SvFLAGS(dsv) |= SvFLAGS(ssv) & (SVf_POK|SVp_POK|SVf_UTF8);
21909 SvPV_set(dsv, RX_WRAPPED(ssv));
21910 /* We share the same string buffer as the original regexp, on which we
21911 hold a reference count, incremented when mother_re is set below.
21912 The string pointer is copied here, being part of the regexp struct.
21914 memcpy(&(drx->xpv_cur), &(srx->xpv_cur),
21915 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
21919 const I32 npar = srx->nparens+1;
21920 Newx(drx->offs, npar, regexp_paren_pair);
21921 Copy(srx->offs, drx->offs, npar, regexp_paren_pair);
21923 if (srx->substrs) {
21925 Newx(drx->substrs, 1, struct reg_substr_data);
21926 StructCopy(srx->substrs, drx->substrs, struct reg_substr_data);
21928 for (i = 0; i < 2; i++) {
21929 SvREFCNT_inc_void(drx->substrs->data[i].substr);
21930 SvREFCNT_inc_void(drx->substrs->data[i].utf8_substr);
21933 /* check_substr and check_utf8, if non-NULL, point to either their
21934 anchored or float namesakes, and don't hold a second reference. */
21936 RX_MATCH_COPIED_off(dsv);
21937 #ifdef PERL_ANY_COW
21938 drx->saved_copy = NULL;
21940 drx->mother_re = ReREFCNT_inc(srx->mother_re ? srx->mother_re : ssv);
21941 SvREFCNT_inc_void(drx->qr_anoncv);
21942 if (srx->recurse_locinput)
21943 Newx(drx->recurse_locinput, srx->nparens + 1, char *);
21950 /* regfree_internal()
21952 Free the private data in a regexp. This is overloadable by
21953 extensions. Perl takes care of the regexp structure in pregfree(),
21954 this covers the *pprivate pointer which technically perl doesn't
21955 know about, however of course we have to handle the
21956 regexp_internal structure when no extension is in use.
21958 Note this is called before freeing anything in the regexp
21963 Perl_regfree_internal(pTHX_ REGEXP * const rx)
21965 struct regexp *const r = ReANY(rx);
21966 RXi_GET_DECL(r, ri);
21967 DECLARE_AND_GET_RE_DEBUG_FLAGS;
21969 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
21979 SV *dsv= sv_newmortal();
21980 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
21981 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), PL_dump_re_max_len);
21982 Perl_re_printf( aTHX_ "%sFreeing REx:%s %s\n",
21983 PL_colors[4], PL_colors[5], s);
21987 #ifdef RE_TRACK_PATTERN_OFFSETS
21989 Safefree(ri->u.offsets); /* 20010421 MJD */
21991 if (ri->code_blocks)
21992 S_free_codeblocks(aTHX_ ri->code_blocks);
21995 int n = ri->data->count;
21998 /* If you add a ->what type here, update the comment in regcomp.h */
21999 switch (ri->data->what[n]) {
22005 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
22008 Safefree(ri->data->data[n]);
22014 { /* Aho Corasick add-on structure for a trie node.
22015 Used in stclass optimization only */
22017 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
22019 refcount = --aho->refcount;
22022 PerlMemShared_free(aho->states);
22023 PerlMemShared_free(aho->fail);
22024 /* do this last!!!! */
22025 PerlMemShared_free(ri->data->data[n]);
22026 /* we should only ever get called once, so
22027 * assert as much, and also guard the free
22028 * which /might/ happen twice. At the least
22029 * it will make code anlyzers happy and it
22030 * doesn't cost much. - Yves */
22031 assert(ri->regstclass);
22032 if (ri->regstclass) {
22033 PerlMemShared_free(ri->regstclass);
22034 ri->regstclass = 0;
22041 /* trie structure. */
22043 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
22045 refcount = --trie->refcount;
22048 PerlMemShared_free(trie->charmap);
22049 PerlMemShared_free(trie->states);
22050 PerlMemShared_free(trie->trans);
22052 PerlMemShared_free(trie->bitmap);
22054 PerlMemShared_free(trie->jump);
22055 PerlMemShared_free(trie->wordinfo);
22056 /* do this last!!!! */
22057 PerlMemShared_free(ri->data->data[n]);
22062 Perl_croak(aTHX_ "panic: regfree data code '%c'",
22063 ri->data->what[n]);
22066 Safefree(ri->data->what);
22067 Safefree(ri->data);
22073 #define av_dup_inc(s, t) MUTABLE_AV(sv_dup_inc((const SV *)s, t))
22074 #define hv_dup_inc(s, t) MUTABLE_HV(sv_dup_inc((const SV *)s, t))
22075 #define SAVEPVN(p, n) ((p) ? savepvn(p, n) : NULL)
22078 =for apidoc re_dup_guts
22079 Duplicate a regexp.
22081 This routine is expected to clone a given regexp structure. It is only
22082 compiled under USE_ITHREADS.
22084 After all of the core data stored in struct regexp is duplicated
22085 the C<regexp_engine.dupe> method is used to copy any private data
22086 stored in the *pprivate pointer. This allows extensions to handle
22087 any duplication they need to do.
22091 See pregfree() and regfree_internal() if you change anything here.
22093 #if defined(USE_ITHREADS)
22094 #ifndef PERL_IN_XSUB_RE
22096 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
22099 const struct regexp *r = ReANY(sstr);
22100 struct regexp *ret = ReANY(dstr);
22102 PERL_ARGS_ASSERT_RE_DUP_GUTS;
22104 npar = r->nparens+1;
22105 Newx(ret->offs, npar, regexp_paren_pair);
22106 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
22108 if (ret->substrs) {
22109 /* Do it this way to avoid reading from *r after the StructCopy().
22110 That way, if any of the sv_dup_inc()s dislodge *r from the L1
22111 cache, it doesn't matter. */
22113 const bool anchored = r->check_substr
22114 ? r->check_substr == r->substrs->data[0].substr
22115 : r->check_utf8 == r->substrs->data[0].utf8_substr;
22116 Newx(ret->substrs, 1, struct reg_substr_data);
22117 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
22119 for (i = 0; i < 2; i++) {
22120 ret->substrs->data[i].substr =
22121 sv_dup_inc(ret->substrs->data[i].substr, param);
22122 ret->substrs->data[i].utf8_substr =
22123 sv_dup_inc(ret->substrs->data[i].utf8_substr, param);
22126 /* check_substr and check_utf8, if non-NULL, point to either their
22127 anchored or float namesakes, and don't hold a second reference. */
22129 if (ret->check_substr) {
22131 assert(r->check_utf8 == r->substrs->data[0].utf8_substr);
22133 ret->check_substr = ret->substrs->data[0].substr;
22134 ret->check_utf8 = ret->substrs->data[0].utf8_substr;
22136 assert(r->check_substr == r->substrs->data[1].substr);
22137 assert(r->check_utf8 == r->substrs->data[1].utf8_substr);
22139 ret->check_substr = ret->substrs->data[1].substr;
22140 ret->check_utf8 = ret->substrs->data[1].utf8_substr;
22142 } else if (ret->check_utf8) {
22144 ret->check_utf8 = ret->substrs->data[0].utf8_substr;
22146 ret->check_utf8 = ret->substrs->data[1].utf8_substr;
22151 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
22152 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
22153 if (r->recurse_locinput)
22154 Newx(ret->recurse_locinput, r->nparens + 1, char *);
22157 RXi_SET(ret, CALLREGDUPE_PVT(dstr, param));
22159 if (RX_MATCH_COPIED(dstr))
22160 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
22162 ret->subbeg = NULL;
22163 #ifdef PERL_ANY_COW
22164 ret->saved_copy = NULL;
22167 /* Whether mother_re be set or no, we need to copy the string. We
22168 cannot refrain from copying it when the storage points directly to
22169 our mother regexp, because that's
22170 1: a buffer in a different thread
22171 2: something we no longer hold a reference on
22172 so we need to copy it locally. */
22173 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED_const(sstr), SvCUR(sstr)+1);
22174 /* set malloced length to a non-zero value so it will be freed
22175 * (otherwise in combination with SVf_FAKE it looks like an alien
22176 * buffer). It doesn't have to be the actual malloced size, since it
22177 * should never be grown */
22178 SvLEN_set(dstr, SvCUR(sstr)+1);
22179 ret->mother_re = NULL;
22181 #endif /* PERL_IN_XSUB_RE */
22186 This is the internal complement to regdupe() which is used to copy
22187 the structure pointed to by the *pprivate pointer in the regexp.
22188 This is the core version of the extension overridable cloning hook.
22189 The regexp structure being duplicated will be copied by perl prior
22190 to this and will be provided as the regexp *r argument, however
22191 with the /old/ structures pprivate pointer value. Thus this routine
22192 may override any copying normally done by perl.
22194 It returns a pointer to the new regexp_internal structure.
22198 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
22200 struct regexp *const r = ReANY(rx);
22201 regexp_internal *reti;
22203 RXi_GET_DECL(r, ri);
22205 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
22209 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
22210 char, regexp_internal);
22211 Copy(ri->program, reti->program, len+1, regnode);
22214 if (ri->code_blocks) {
22216 Newx(reti->code_blocks, 1, struct reg_code_blocks);
22217 Newx(reti->code_blocks->cb, ri->code_blocks->count,
22218 struct reg_code_block);
22219 Copy(ri->code_blocks->cb, reti->code_blocks->cb,
22220 ri->code_blocks->count, struct reg_code_block);
22221 for (n = 0; n < ri->code_blocks->count; n++)
22222 reti->code_blocks->cb[n].src_regex = (REGEXP*)
22223 sv_dup_inc((SV*)(ri->code_blocks->cb[n].src_regex), param);
22224 reti->code_blocks->count = ri->code_blocks->count;
22225 reti->code_blocks->refcnt = 1;
22228 reti->code_blocks = NULL;
22230 reti->regstclass = NULL;
22233 struct reg_data *d;
22234 const int count = ri->data->count;
22237 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
22238 char, struct reg_data);
22239 Newx(d->what, count, U8);
22242 for (i = 0; i < count; i++) {
22243 d->what[i] = ri->data->what[i];
22244 switch (d->what[i]) {
22245 /* see also regcomp.h and regfree_internal() */
22246 case 'a': /* actually an AV, but the dup function is identical.
22247 values seem to be "plain sv's" generally. */
22248 case 'r': /* a compiled regex (but still just another SV) */
22249 case 's': /* an RV (currently only used for an RV to an AV by the ANYOF code)
22250 this use case should go away, the code could have used
22251 'a' instead - see S_set_ANYOF_arg() for array contents. */
22252 case 'S': /* actually an SV, but the dup function is identical. */
22253 case 'u': /* actually an HV, but the dup function is identical.
22254 values are "plain sv's" */
22255 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
22258 /* Synthetic Start Class - "Fake" charclass we generate to optimize
22259 * patterns which could start with several different things. Pre-TRIE
22260 * this was more important than it is now, however this still helps
22261 * in some places, for instance /x?a+/ might produce a SSC equivalent
22262 * to [xa]. This is used by Perl_re_intuit_start() and S_find_byclass()
22265 /* This is cheating. */
22266 Newx(d->data[i], 1, regnode_ssc);
22267 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
22268 reti->regstclass = (regnode*)d->data[i];
22271 /* AHO-CORASICK fail table */
22272 /* Trie stclasses are readonly and can thus be shared
22273 * without duplication. We free the stclass in pregfree
22274 * when the corresponding reg_ac_data struct is freed.
22276 reti->regstclass= ri->regstclass;
22279 /* TRIE transition table */
22281 ((reg_trie_data*)ri->data->data[i])->refcount++;
22284 case 'l': /* (?{...}) or (??{ ... }) code (cb->block) */
22285 case 'L': /* same when RExC_pm_flags & PMf_HAS_CV and code
22286 is not from another regexp */
22287 d->data[i] = ri->data->data[i];
22290 Perl_croak(aTHX_ "panic: re_dup_guts unknown data code '%c'",
22291 ri->data->what[i]);
22300 reti->name_list_idx = ri->name_list_idx;
22302 #ifdef RE_TRACK_PATTERN_OFFSETS
22303 if (ri->u.offsets) {
22304 Newx(reti->u.offsets, 2*len+1, U32);
22305 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
22308 SetProgLen(reti, len);
22311 return (void*)reti;
22314 #endif /* USE_ITHREADS */
22316 #ifndef PERL_IN_XSUB_RE
22319 - regnext - dig the "next" pointer out of a node
22322 Perl_regnext(pTHX_ regnode *p)
22329 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
22330 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
22331 (int)OP(p), (int)REGNODE_MAX);
22334 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
22344 S_re_croak(pTHX_ bool utf8, const char* pat,...)
22347 STRLEN len = strlen(pat);
22350 const char *message;
22352 PERL_ARGS_ASSERT_RE_CROAK;
22356 Copy(pat, buf, len , char);
22358 buf[len + 1] = '\0';
22359 va_start(args, pat);
22360 msv = vmess(buf, &args);
22362 message = SvPV_const(msv, len);
22365 Copy(message, buf, len , char);
22366 /* len-1 to avoid \n */
22367 Perl_croak(aTHX_ "%" UTF8f, UTF8fARG(utf8, len-1, buf));
22370 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
22372 #ifndef PERL_IN_XSUB_RE
22374 Perl_save_re_context(pTHX)
22379 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
22382 const REGEXP * const rx = PM_GETRE(PL_curpm);
22384 nparens = RX_NPARENS(rx);
22387 /* RT #124109. This is a complete hack; in the SWASHNEW case we know
22388 * that PL_curpm will be null, but that utf8.pm and the modules it
22389 * loads will only use $1..$3.
22390 * The t/porting/re_context.t test file checks this assumption.
22395 for (i = 1; i <= nparens; i++) {
22396 char digits[TYPE_CHARS(long)];
22397 const STRLEN len = my_snprintf(digits, sizeof(digits),
22399 GV *const *const gvp
22400 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
22403 GV * const gv = *gvp;
22404 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
22414 S_put_code_point(pTHX_ SV *sv, UV c)
22416 PERL_ARGS_ASSERT_PUT_CODE_POINT;
22419 Perl_sv_catpvf(aTHX_ sv, "\\x{%04" UVXf "}", c);
22421 else if (isPRINT(c)) {
22422 const char string = (char) c;
22424 /* We use {phrase} as metanotation in the class, so also escape literal
22426 if (isBACKSLASHED_PUNCT(c) || c == '{' || c == '}')
22427 sv_catpvs(sv, "\\");
22428 sv_catpvn(sv, &string, 1);
22430 else if (isMNEMONIC_CNTRL(c)) {
22431 Perl_sv_catpvf(aTHX_ sv, "%s", cntrl_to_mnemonic((U8) c));
22434 Perl_sv_catpvf(aTHX_ sv, "\\x%02X", (U8) c);
22439 S_put_range(pTHX_ SV *sv, UV start, const UV end, const bool allow_literals)
22441 /* Appends to 'sv' a displayable version of the range of code points from
22442 * 'start' to 'end'. Mnemonics (like '\r') are used for the few controls
22443 * that have them, when they occur at the beginning or end of the range.
22444 * It uses hex to output the remaining code points, unless 'allow_literals'
22445 * is true, in which case the printable ASCII ones are output as-is (though
22446 * some of these will be escaped by put_code_point()).
22448 * NOTE: This is designed only for printing ranges of code points that fit
22449 * inside an ANYOF bitmap. Higher code points are simply suppressed
22452 const unsigned int min_range_count = 3;
22454 assert(start <= end);
22456 PERL_ARGS_ASSERT_PUT_RANGE;
22458 while (start <= end) {
22460 const char * format;
22462 if ( end - start < min_range_count
22463 && (end - start <= 2 || (isPRINT_A(start) && isPRINT_A(end))))
22465 /* Output a range of 1 or 2 chars individually, or longer ranges
22466 * when printable */
22467 for (; start <= end; start++) {
22468 put_code_point(sv, start);
22473 /* If permitted by the input options, and there is a possibility that
22474 * this range contains a printable literal, look to see if there is
22476 if (allow_literals && start <= MAX_PRINT_A) {
22478 /* If the character at the beginning of the range isn't an ASCII
22479 * printable, effectively split the range into two parts:
22480 * 1) the portion before the first such printable,
22482 * and output them separately. */
22483 if (! isPRINT_A(start)) {
22484 UV temp_end = start + 1;
22486 /* There is no point looking beyond the final possible
22487 * printable, in MAX_PRINT_A */
22488 UV max = MIN(end, MAX_PRINT_A);
22490 while (temp_end <= max && ! isPRINT_A(temp_end)) {
22494 /* Here, temp_end points to one beyond the first printable if
22495 * found, or to one beyond 'max' if not. If none found, make
22496 * sure that we use the entire range */
22497 if (temp_end > MAX_PRINT_A) {
22498 temp_end = end + 1;
22501 /* Output the first part of the split range: the part that
22502 * doesn't have printables, with the parameter set to not look
22503 * for literals (otherwise we would infinitely recurse) */
22504 put_range(sv, start, temp_end - 1, FALSE);
22506 /* The 2nd part of the range (if any) starts here. */
22509 /* We do a continue, instead of dropping down, because even if
22510 * the 2nd part is non-empty, it could be so short that we want
22511 * to output it as individual characters, as tested for at the
22512 * top of this loop. */
22516 /* Here, 'start' is a printable ASCII. If it is an alphanumeric,
22517 * output a sub-range of just the digits or letters, then process
22518 * the remaining portion as usual. */
22519 if (isALPHANUMERIC_A(start)) {
22520 UV mask = (isDIGIT_A(start))
22525 UV temp_end = start + 1;
22527 /* Find the end of the sub-range that includes just the
22528 * characters in the same class as the first character in it */
22529 while (temp_end <= end && _generic_isCC_A(temp_end, mask)) {
22534 /* For short ranges, don't duplicate the code above to output
22535 * them; just call recursively */
22536 if (temp_end - start < min_range_count) {
22537 put_range(sv, start, temp_end, FALSE);
22539 else { /* Output as a range */
22540 put_code_point(sv, start);
22541 sv_catpvs(sv, "-");
22542 put_code_point(sv, temp_end);
22544 start = temp_end + 1;
22548 /* We output any other printables as individual characters */
22549 if (isPUNCT_A(start) || isSPACE_A(start)) {
22550 while (start <= end && (isPUNCT_A(start)
22551 || isSPACE_A(start)))
22553 put_code_point(sv, start);
22558 } /* End of looking for literals */
22560 /* Here is not to output as a literal. Some control characters have
22561 * mnemonic names. Split off any of those at the beginning and end of
22562 * the range to print mnemonically. It isn't possible for many of
22563 * these to be in a row, so this won't overwhelm with output */
22565 && (isMNEMONIC_CNTRL(start) || isMNEMONIC_CNTRL(end)))
22567 while (isMNEMONIC_CNTRL(start) && start <= end) {
22568 put_code_point(sv, start);
22572 /* If this didn't take care of the whole range ... */
22573 if (start <= end) {
22575 /* Look backwards from the end to find the final non-mnemonic
22578 while (isMNEMONIC_CNTRL(temp_end)) {
22582 /* And separately output the interior range that doesn't start
22583 * or end with mnemonics */
22584 put_range(sv, start, temp_end, FALSE);
22586 /* Then output the mnemonic trailing controls */
22587 start = temp_end + 1;
22588 while (start <= end) {
22589 put_code_point(sv, start);
22596 /* As a final resort, output the range or subrange as hex. */
22598 if (start >= NUM_ANYOF_CODE_POINTS) {
22601 else { /* Have to split range at the bitmap boundary */
22602 this_end = (end < NUM_ANYOF_CODE_POINTS)
22604 : NUM_ANYOF_CODE_POINTS - 1;
22606 #if NUM_ANYOF_CODE_POINTS > 256
22607 format = (this_end < 256)
22608 ? "\\x%02" UVXf "-\\x%02" UVXf
22609 : "\\x{%04" UVXf "}-\\x{%04" UVXf "}";
22611 format = "\\x%02" UVXf "-\\x%02" UVXf;
22613 GCC_DIAG_IGNORE_STMT(-Wformat-nonliteral);
22614 Perl_sv_catpvf(aTHX_ sv, format, start, this_end);
22615 GCC_DIAG_RESTORE_STMT;
22621 S_put_charclass_bitmap_innards_invlist(pTHX_ SV *sv, SV* invlist)
22623 /* Concatenate onto the PV in 'sv' a displayable form of the inversion list
22627 bool allow_literals = TRUE;
22629 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_INVLIST;
22631 /* Generally, it is more readable if printable characters are output as
22632 * literals, but if a range (nearly) spans all of them, it's best to output
22633 * it as a single range. This code will use a single range if all but 2
22634 * ASCII printables are in it */
22635 invlist_iterinit(invlist);
22636 while (invlist_iternext(invlist, &start, &end)) {
22638 /* If the range starts beyond the final printable, it doesn't have any
22640 if (start > MAX_PRINT_A) {
22644 /* In both ASCII and EBCDIC, a SPACE is the lowest printable. To span
22645 * all but two, the range must start and end no later than 2 from
22647 if (start < ' ' + 2 && end > MAX_PRINT_A - 2) {
22648 if (end > MAX_PRINT_A) {
22654 if (end - start >= MAX_PRINT_A - ' ' - 2) {
22655 allow_literals = FALSE;
22660 invlist_iterfinish(invlist);
22662 /* Here we have figured things out. Output each range */
22663 invlist_iterinit(invlist);
22664 while (invlist_iternext(invlist, &start, &end)) {
22665 if (start >= NUM_ANYOF_CODE_POINTS) {
22668 put_range(sv, start, end, allow_literals);
22670 invlist_iterfinish(invlist);
22676 S_put_charclass_bitmap_innards_common(pTHX_
22677 SV* invlist, /* The bitmap */
22678 SV* posixes, /* Under /l, things like [:word:], \S */
22679 SV* only_utf8, /* Under /d, matches iff the target is UTF-8 */
22680 SV* not_utf8, /* /d, matches iff the target isn't UTF-8 */
22681 SV* only_utf8_locale, /* Under /l, matches if the locale is UTF-8 */
22682 const bool invert /* Is the result to be inverted? */
22685 /* Create and return an SV containing a displayable version of the bitmap
22686 * and associated information determined by the input parameters. If the
22687 * output would have been only the inversion indicator '^', NULL is instead
22692 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_COMMON;
22695 output = newSVpvs("^");
22698 output = newSVpvs("");
22701 /* First, the code points in the bitmap that are unconditionally there */
22702 put_charclass_bitmap_innards_invlist(output, invlist);
22704 /* Traditionally, these have been placed after the main code points */
22706 sv_catsv(output, posixes);
22709 if (only_utf8 && _invlist_len(only_utf8)) {
22710 Perl_sv_catpvf(aTHX_ output, "%s{utf8}%s", PL_colors[1], PL_colors[0]);
22711 put_charclass_bitmap_innards_invlist(output, only_utf8);
22714 if (not_utf8 && _invlist_len(not_utf8)) {
22715 Perl_sv_catpvf(aTHX_ output, "%s{not utf8}%s", PL_colors[1], PL_colors[0]);
22716 put_charclass_bitmap_innards_invlist(output, not_utf8);
22719 if (only_utf8_locale && _invlist_len(only_utf8_locale)) {
22720 Perl_sv_catpvf(aTHX_ output, "%s{utf8 locale}%s", PL_colors[1], PL_colors[0]);
22721 put_charclass_bitmap_innards_invlist(output, only_utf8_locale);
22723 /* This is the only list in this routine that can legally contain code
22724 * points outside the bitmap range. The call just above to
22725 * 'put_charclass_bitmap_innards_invlist' will simply suppress them, so
22726 * output them here. There's about a half-dozen possible, and none in
22727 * contiguous ranges longer than 2 */
22728 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
22730 SV* above_bitmap = NULL;
22732 _invlist_subtract(only_utf8_locale, PL_InBitmap, &above_bitmap);
22734 invlist_iterinit(above_bitmap);
22735 while (invlist_iternext(above_bitmap, &start, &end)) {
22738 for (i = start; i <= end; i++) {
22739 put_code_point(output, i);
22742 invlist_iterfinish(above_bitmap);
22743 SvREFCNT_dec_NN(above_bitmap);
22747 if (invert && SvCUR(output) == 1) {
22755 S_put_charclass_bitmap_innards(pTHX_ SV *sv,
22757 SV *nonbitmap_invlist,
22758 SV *only_utf8_locale_invlist,
22759 const regnode * const node,
22761 const bool force_as_is_display)
22763 /* Appends to 'sv' a displayable version of the innards of the bracketed
22764 * character class defined by the other arguments:
22765 * 'bitmap' points to the bitmap, or NULL if to ignore that.
22766 * 'nonbitmap_invlist' is an inversion list of the code points that are in
22767 * the bitmap range, but for some reason aren't in the bitmap; NULL if
22768 * none. The reasons for this could be that they require some
22769 * condition such as the target string being or not being in UTF-8
22770 * (under /d), or because they came from a user-defined property that
22771 * was not resolved at the time of the regex compilation (under /u)
22772 * 'only_utf8_locale_invlist' is an inversion list of the code points that
22773 * are valid only if the runtime locale is a UTF-8 one; NULL if none
22774 * 'node' is the regex pattern ANYOF node. It is needed only when the
22775 * above two parameters are not null, and is passed so that this
22776 * routine can tease apart the various reasons for them.
22777 * 'flags' is the flags field of 'node'
22778 * 'force_as_is_display' is TRUE if this routine should definitely NOT try
22779 * to invert things to see if that leads to a cleaner display. If
22780 * FALSE, this routine is free to use its judgment about doing this.
22782 * It returns TRUE if there was actually something output. (It may be that
22783 * the bitmap, etc is empty.)
22785 * When called for outputting the bitmap of a non-ANYOF node, just pass the
22786 * bitmap, with the succeeding parameters set to NULL, and the final one to
22790 /* In general, it tries to display the 'cleanest' representation of the
22791 * innards, choosing whether to display them inverted or not, regardless of
22792 * whether the class itself is to be inverted. However, there are some
22793 * cases where it can't try inverting, as what actually matches isn't known
22794 * until runtime, and hence the inversion isn't either. */
22796 bool inverting_allowed = ! force_as_is_display;
22799 STRLEN orig_sv_cur = SvCUR(sv);
22801 SV* invlist; /* Inversion list we accumulate of code points that
22802 are unconditionally matched */
22803 SV* only_utf8 = NULL; /* Under /d, list of matches iff the target is
22805 SV* not_utf8 = NULL; /* /d, list of matches iff the target isn't UTF-8
22807 SV* posixes = NULL; /* Under /l, string of things like [:word:], \D */
22808 SV* only_utf8_locale = NULL; /* Under /l, list of matches if the locale
22811 SV* as_is_display; /* The output string when we take the inputs
22813 SV* inverted_display; /* The output string when we invert the inputs */
22815 bool invert = cBOOL(flags & ANYOF_INVERT); /* Is the input to be inverted
22817 /* We are biased in favor of displaying things without them being inverted,
22818 * as that is generally easier to understand */
22819 const int bias = 5;
22821 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS;
22823 /* Start off with whatever code points are passed in. (We clone, so we
22824 * don't change the caller's list) */
22825 if (nonbitmap_invlist) {
22826 assert(invlist_highest(nonbitmap_invlist) < NUM_ANYOF_CODE_POINTS);
22827 invlist = invlist_clone(nonbitmap_invlist, NULL);
22829 else { /* Worst case size is every other code point is matched */
22830 invlist = _new_invlist(NUM_ANYOF_CODE_POINTS / 2);
22834 if (OP(node) == ANYOFD) {
22836 /* This flag indicates that the code points below 0x100 in the
22837 * nonbitmap list are precisely the ones that match only when the
22838 * target is UTF-8 (they should all be non-ASCII). */
22839 if (flags & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)
22841 _invlist_intersection(invlist, PL_UpperLatin1, &only_utf8);
22842 _invlist_subtract(invlist, only_utf8, &invlist);
22845 /* And this flag for matching all non-ASCII 0xFF and below */
22846 if (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)
22848 not_utf8 = invlist_clone(PL_UpperLatin1, NULL);
22851 else if (OP(node) == ANYOFL || OP(node) == ANYOFPOSIXL) {
22853 /* If either of these flags are set, what matches isn't
22854 * determinable except during execution, so don't know enough here
22856 if (flags & (ANYOFL_FOLD|ANYOF_MATCHES_POSIXL)) {
22857 inverting_allowed = FALSE;
22860 /* What the posix classes match also varies at runtime, so these
22861 * will be output symbolically. */
22862 if (ANYOF_POSIXL_TEST_ANY_SET(node)) {
22865 posixes = newSVpvs("");
22866 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
22867 if (ANYOF_POSIXL_TEST(node, i)) {
22868 sv_catpv(posixes, anyofs[i]);
22875 /* Accumulate the bit map into the unconditional match list */
22877 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
22878 if (BITMAP_TEST(bitmap, i)) {
22881 i < NUM_ANYOF_CODE_POINTS && BITMAP_TEST(bitmap, i);
22884 invlist = _add_range_to_invlist(invlist, start, i-1);
22889 /* Make sure that the conditional match lists don't have anything in them
22890 * that match unconditionally; otherwise the output is quite confusing.
22891 * This could happen if the code that populates these misses some
22894 _invlist_subtract(only_utf8, invlist, &only_utf8);
22897 _invlist_subtract(not_utf8, invlist, ¬_utf8);
22900 if (only_utf8_locale_invlist) {
22902 /* Since this list is passed in, we have to make a copy before
22904 only_utf8_locale = invlist_clone(only_utf8_locale_invlist, NULL);
22906 _invlist_subtract(only_utf8_locale, invlist, &only_utf8_locale);
22908 /* And, it can get really weird for us to try outputting an inverted
22909 * form of this list when it has things above the bitmap, so don't even
22911 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
22912 inverting_allowed = FALSE;
22916 /* Calculate what the output would be if we take the input as-is */
22917 as_is_display = put_charclass_bitmap_innards_common(invlist,
22924 /* If have to take the output as-is, just do that */
22925 if (! inverting_allowed) {
22926 if (as_is_display) {
22927 sv_catsv(sv, as_is_display);
22928 SvREFCNT_dec_NN(as_is_display);
22931 else { /* But otherwise, create the output again on the inverted input, and
22932 use whichever version is shorter */
22934 int inverted_bias, as_is_bias;
22936 /* We will apply our bias to whichever of the results doesn't have
22946 inverted_bias = bias;
22949 /* Now invert each of the lists that contribute to the output,
22950 * excluding from the result things outside the possible range */
22952 /* For the unconditional inversion list, we have to add in all the
22953 * conditional code points, so that when inverted, they will be gone
22955 _invlist_union(only_utf8, invlist, &invlist);
22956 _invlist_union(not_utf8, invlist, &invlist);
22957 _invlist_union(only_utf8_locale, invlist, &invlist);
22958 _invlist_invert(invlist);
22959 _invlist_intersection(invlist, PL_InBitmap, &invlist);
22962 _invlist_invert(only_utf8);
22963 _invlist_intersection(only_utf8, PL_UpperLatin1, &only_utf8);
22965 else if (not_utf8) {
22967 /* If a code point matches iff the target string is not in UTF-8,
22968 * then complementing the result has it not match iff not in UTF-8,
22969 * which is the same thing as matching iff it is UTF-8. */
22970 only_utf8 = not_utf8;
22974 if (only_utf8_locale) {
22975 _invlist_invert(only_utf8_locale);
22976 _invlist_intersection(only_utf8_locale,
22978 &only_utf8_locale);
22981 inverted_display = put_charclass_bitmap_innards_common(
22986 only_utf8_locale, invert);
22988 /* Use the shortest representation, taking into account our bias
22989 * against showing it inverted */
22990 if ( inverted_display
22991 && ( ! as_is_display
22992 || ( SvCUR(inverted_display) + inverted_bias
22993 < SvCUR(as_is_display) + as_is_bias)))
22995 sv_catsv(sv, inverted_display);
22997 else if (as_is_display) {
22998 sv_catsv(sv, as_is_display);
23001 SvREFCNT_dec(as_is_display);
23002 SvREFCNT_dec(inverted_display);
23005 SvREFCNT_dec_NN(invlist);
23006 SvREFCNT_dec(only_utf8);
23007 SvREFCNT_dec(not_utf8);
23008 SvREFCNT_dec(posixes);
23009 SvREFCNT_dec(only_utf8_locale);
23011 return SvCUR(sv) > orig_sv_cur;
23014 #define CLEAR_OPTSTART \
23015 if (optstart) STMT_START { \
23016 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ \
23017 " (%" IVdf " nodes)\n", (IV)(node - optstart))); \
23021 #define DUMPUNTIL(b,e) \
23023 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
23025 STATIC const regnode *
23026 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
23027 const regnode *last, const regnode *plast,
23028 SV* sv, I32 indent, U32 depth)
23030 U8 op = PSEUDO; /* Arbitrary non-END op. */
23031 const regnode *next;
23032 const regnode *optstart= NULL;
23034 RXi_GET_DECL(r, ri);
23035 DECLARE_AND_GET_RE_DEBUG_FLAGS;
23037 PERL_ARGS_ASSERT_DUMPUNTIL;
23039 #ifdef DEBUG_DUMPUNTIL
23040 Perl_re_printf( aTHX_ "--- %d : %d - %d - %d\n", indent, node-start,
23041 last ? last-start : 0, plast ? plast-start : 0);
23044 if (plast && plast < last)
23047 while (PL_regkind[op] != END && (!last || node < last)) {
23049 /* While that wasn't END last time... */
23052 if (op == CLOSE || op == SRCLOSE || op == WHILEM)
23054 next = regnext((regnode *)node);
23057 if (OP(node) == OPTIMIZED) {
23058 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
23065 regprop(r, sv, node, NULL, NULL);
23066 Perl_re_printf( aTHX_ "%4" IVdf ":%*s%s", (IV)(node - start),
23067 (int)(2*indent + 1), "", SvPVX_const(sv));
23069 if (OP(node) != OPTIMIZED) {
23070 if (next == NULL) /* Next ptr. */
23071 Perl_re_printf( aTHX_ " (0)");
23072 else if (PL_regkind[(U8)op] == BRANCH
23073 && PL_regkind[OP(next)] != BRANCH )
23074 Perl_re_printf( aTHX_ " (FAIL)");
23076 Perl_re_printf( aTHX_ " (%" IVdf ")", (IV)(next - start));
23077 Perl_re_printf( aTHX_ "\n");
23081 if (PL_regkind[(U8)op] == BRANCHJ) {
23084 const regnode *nnode = (OP(next) == LONGJMP
23085 ? regnext((regnode *)next)
23087 if (last && nnode > last)
23089 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
23092 else if (PL_regkind[(U8)op] == BRANCH) {
23094 DUMPUNTIL(NEXTOPER(node), next);
23096 else if ( PL_regkind[(U8)op] == TRIE ) {
23097 const regnode *this_trie = node;
23098 const char op = OP(node);
23099 const U32 n = ARG(node);
23100 const reg_ac_data * const ac = op>=AHOCORASICK ?
23101 (reg_ac_data *)ri->data->data[n] :
23103 const reg_trie_data * const trie =
23104 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
23106 AV *const trie_words
23107 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
23109 const regnode *nextbranch= NULL;
23112 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
23113 SV ** const elem_ptr = av_fetch(trie_words, word_idx, 0);
23115 Perl_re_indentf( aTHX_ "%s ",
23118 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
23119 SvCUR(*elem_ptr), PL_dump_re_max_len,
23120 PL_colors[0], PL_colors[1],
23122 ? PERL_PV_ESCAPE_UNI
23124 | PERL_PV_PRETTY_ELLIPSES
23125 | PERL_PV_PRETTY_LTGT
23130 U16 dist= trie->jump[word_idx+1];
23131 Perl_re_printf( aTHX_ "(%" UVuf ")\n",
23132 (UV)((dist ? this_trie + dist : next) - start));
23135 nextbranch= this_trie + trie->jump[0];
23136 DUMPUNTIL(this_trie + dist, nextbranch);
23138 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
23139 nextbranch= regnext((regnode *)nextbranch);
23141 Perl_re_printf( aTHX_ "\n");
23144 if (last && next > last)
23149 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
23150 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
23151 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
23153 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
23155 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
23157 else if ( op == PLUS || op == STAR) {
23158 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
23160 else if (PL_regkind[(U8)op] == EXACT || op == ANYOFHs) {
23161 /* Literal string, where present. */
23162 node += NODE_SZ_STR(node) - 1;
23163 node = NEXTOPER(node);
23166 node = NEXTOPER(node);
23167 node += regarglen[(U8)op];
23169 if (op == CURLYX || op == OPEN || op == SROPEN)
23173 #ifdef DEBUG_DUMPUNTIL
23174 Perl_re_printf( aTHX_ "--- %d\n", (int)indent);
23179 #endif /* DEBUGGING */
23181 #ifndef PERL_IN_XSUB_RE
23183 # include "uni_keywords.h"
23186 Perl_init_uniprops(pTHX)
23190 char * dump_len_string;
23192 dump_len_string = PerlEnv_getenv("PERL_DUMP_RE_MAX_LEN");
23193 if ( ! dump_len_string
23194 || ! grok_atoUV(dump_len_string, (UV *)&PL_dump_re_max_len, NULL))
23196 PL_dump_re_max_len = 60; /* A reasonable default */
23200 PL_user_def_props = newHV();
23202 # ifdef USE_ITHREADS
23204 HvSHAREKEYS_off(PL_user_def_props);
23205 PL_user_def_props_aTHX = aTHX;
23209 /* Set up the inversion list interpreter-level variables */
23211 PL_XPosix_ptrs[_CC_ASCII] = _new_invlist_C_array(uni_prop_ptrs[UNI_ASCII]);
23212 PL_XPosix_ptrs[_CC_ALPHANUMERIC] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXALNUM]);
23213 PL_XPosix_ptrs[_CC_ALPHA] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXALPHA]);
23214 PL_XPosix_ptrs[_CC_BLANK] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXBLANK]);
23215 PL_XPosix_ptrs[_CC_CASED] = _new_invlist_C_array(uni_prop_ptrs[UNI_CASED]);
23216 PL_XPosix_ptrs[_CC_CNTRL] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXCNTRL]);
23217 PL_XPosix_ptrs[_CC_DIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXDIGIT]);
23218 PL_XPosix_ptrs[_CC_GRAPH] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXGRAPH]);
23219 PL_XPosix_ptrs[_CC_LOWER] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXLOWER]);
23220 PL_XPosix_ptrs[_CC_PRINT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXPRINT]);
23221 PL_XPosix_ptrs[_CC_PUNCT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXPUNCT]);
23222 PL_XPosix_ptrs[_CC_SPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXSPACE]);
23223 PL_XPosix_ptrs[_CC_UPPER] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXUPPER]);
23224 PL_XPosix_ptrs[_CC_VERTSPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_VERTSPACE]);
23225 PL_XPosix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXWORD]);
23226 PL_XPosix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXXDIGIT]);
23228 PL_Posix_ptrs[_CC_ASCII] = _new_invlist_C_array(uni_prop_ptrs[UNI_ASCII]);
23229 PL_Posix_ptrs[_CC_ALPHANUMERIC] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXALNUM]);
23230 PL_Posix_ptrs[_CC_ALPHA] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXALPHA]);
23231 PL_Posix_ptrs[_CC_BLANK] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXBLANK]);
23232 PL_Posix_ptrs[_CC_CASED] = PL_Posix_ptrs[_CC_ALPHA];
23233 PL_Posix_ptrs[_CC_CNTRL] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXCNTRL]);
23234 PL_Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXDIGIT]);
23235 PL_Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXGRAPH]);
23236 PL_Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXLOWER]);
23237 PL_Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXPRINT]);
23238 PL_Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXPUNCT]);
23239 PL_Posix_ptrs[_CC_SPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXSPACE]);
23240 PL_Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXUPPER]);
23241 PL_Posix_ptrs[_CC_VERTSPACE] = NULL;
23242 PL_Posix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXWORD]);
23243 PL_Posix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXXDIGIT]);
23245 PL_GCB_invlist = _new_invlist_C_array(_Perl_GCB_invlist);
23246 PL_SB_invlist = _new_invlist_C_array(_Perl_SB_invlist);
23247 PL_WB_invlist = _new_invlist_C_array(_Perl_WB_invlist);
23248 PL_LB_invlist = _new_invlist_C_array(_Perl_LB_invlist);
23249 PL_SCX_invlist = _new_invlist_C_array(_Perl_SCX_invlist);
23251 PL_InBitmap = _new_invlist_C_array(InBitmap_invlist);
23252 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
23253 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
23254 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
23256 PL_Assigned_invlist = _new_invlist_C_array(uni_prop_ptrs[UNI_ASSIGNED]);
23258 PL_utf8_perl_idstart = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_IDSTART]);
23259 PL_utf8_perl_idcont = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_IDCONT]);
23261 PL_utf8_charname_begin = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_CHARNAME_BEGIN]);
23262 PL_utf8_charname_continue = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_CHARNAME_CONTINUE]);
23264 PL_in_some_fold = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_ANY_FOLDS]);
23265 PL_HasMultiCharFold = _new_invlist_C_array(uni_prop_ptrs[
23266 UNI__PERL_FOLDS_TO_MULTI_CHAR]);
23267 PL_InMultiCharFold = _new_invlist_C_array(uni_prop_ptrs[
23268 UNI__PERL_IS_IN_MULTI_CHAR_FOLD]);
23269 PL_utf8_toupper = _new_invlist_C_array(Uppercase_Mapping_invlist);
23270 PL_utf8_tolower = _new_invlist_C_array(Lowercase_Mapping_invlist);
23271 PL_utf8_totitle = _new_invlist_C_array(Titlecase_Mapping_invlist);
23272 PL_utf8_tofold = _new_invlist_C_array(Case_Folding_invlist);
23273 PL_utf8_tosimplefold = _new_invlist_C_array(Simple_Case_Folding_invlist);
23274 PL_utf8_foldclosures = _new_invlist_C_array(_Perl_IVCF_invlist);
23275 PL_utf8_mark = _new_invlist_C_array(uni_prop_ptrs[UNI_M]);
23276 PL_CCC_non0_non230 = _new_invlist_C_array(_Perl_CCC_non0_non230_invlist);
23277 PL_Private_Use = _new_invlist_C_array(uni_prop_ptrs[UNI_CO]);
23280 /* The below are used only by deprecated functions. They could be removed */
23281 PL_utf8_xidcont = _new_invlist_C_array(uni_prop_ptrs[UNI_XIDC]);
23282 PL_utf8_idcont = _new_invlist_C_array(uni_prop_ptrs[UNI_IDC]);
23283 PL_utf8_xidstart = _new_invlist_C_array(uni_prop_ptrs[UNI_XIDS]);
23287 /* These four functions are compiled only in regcomp.c, where they have access
23288 * to the data they return. They are a way for re_comp.c to get access to that
23289 * data without having to compile the whole data structures. */
23292 Perl_do_uniprop_match(const char * const key, const U16 key_len)
23294 PERL_ARGS_ASSERT_DO_UNIPROP_MATCH;
23296 return match_uniprop((U8 *) key, key_len);
23300 Perl_get_prop_definition(pTHX_ const int table_index)
23302 PERL_ARGS_ASSERT_GET_PROP_DEFINITION;
23304 /* Create and return the inversion list */
23305 return _new_invlist_C_array(uni_prop_ptrs[table_index]);
23308 const char * const *
23309 Perl_get_prop_values(const int table_index)
23311 PERL_ARGS_ASSERT_GET_PROP_VALUES;
23313 return UNI_prop_value_ptrs[table_index];
23317 Perl_get_deprecated_property_msg(const Size_t warning_offset)
23319 PERL_ARGS_ASSERT_GET_DEPRECATED_PROPERTY_MSG;
23321 return deprecated_property_msgs[warning_offset];
23326 This code was mainly added for backcompat to give a warning for non-portable
23327 code points in user-defined properties. But experiments showed that the
23328 warning in earlier perls were only omitted on overflow, which should be an
23329 error, so there really isnt a backcompat issue, and actually adding the
23330 warning when none was present before might cause breakage, for little gain. So
23331 khw left this code in, but not enabled. Tests were never added.
23334 Ei |const char *|get_extended_utf8_msg|const UV cp
23336 PERL_STATIC_INLINE const char *
23337 S_get_extended_utf8_msg(pTHX_ const UV cp)
23339 U8 dummy[UTF8_MAXBYTES + 1];
23343 uvchr_to_utf8_flags_msgs(dummy, cp, UNICODE_WARN_PERL_EXTENDED,
23346 msg = hv_fetchs(msgs, "text", 0);
23349 (void) sv_2mortal((SV *) msgs);
23351 return SvPVX(*msg);
23355 #endif /* end of ! PERL_IN_XSUB_RE */
23358 S_compile_wildcard(pTHX_ const char * subpattern, const STRLEN len,
23359 const bool ignore_case)
23361 /* Pretends that the input subpattern is qr/subpattern/aam, compiling it
23362 * possibly with /i if the 'ignore_case' parameter is true. Use /aa
23363 * because nothing outside of ASCII will match. Use /m because the input
23364 * string may be a bunch of lines strung together.
23366 * Also sets up the debugging info */
23368 U32 flags = PMf_MULTILINE|PMf_WILDCARD;
23370 SV * subpattern_sv = sv_2mortal(newSVpvn(subpattern, len));
23371 REGEXP * subpattern_re;
23372 DECLARE_AND_GET_RE_DEBUG_FLAGS;
23374 PERL_ARGS_ASSERT_COMPILE_WILDCARD;
23379 set_regex_charset(&flags, REGEX_ASCII_MORE_RESTRICTED_CHARSET);
23381 /* Like in op.c, we copy the compile time pm flags to the rx ones */
23382 rx_flags = flags & RXf_PMf_COMPILETIME;
23384 #ifndef PERL_IN_XSUB_RE
23385 /* Use the core engine if this file is regcomp.c. That means no
23386 * 'use re "Debug ..." is in effect, so the core engine is sufficient */
23387 subpattern_re = Perl_re_op_compile(aTHX_ &subpattern_sv, 1, NULL,
23388 &PL_core_reg_engine,
23392 if (isDEBUG_WILDCARD) {
23393 /* Use the special debugging engine if this file is re_comp.c and wants
23394 * to output the wildcard matching. This uses whatever
23395 * 'use re "Debug ..." is in effect */
23396 subpattern_re = Perl_re_op_compile(aTHX_ &subpattern_sv, 1, NULL,
23402 /* Use the special wildcard engine if this file is re_comp.c and
23403 * doesn't want to output the wildcard matching. This uses whatever
23404 * 'use re "Debug ..." is in effect for compilation, but this engine
23405 * structure has been set up so that it uses the core engine for
23406 * execution, so no execution debugging as a result of re.pm will be
23408 subpattern_re = Perl_re_op_compile(aTHX_ &subpattern_sv, 1, NULL,
23412 /* XXX The above has the effect that any user-supplied regex engine
23413 * won't be called for matching wildcards. That might be good, or bad.
23414 * It could be changed in several ways. The reason it is done the
23415 * current way is to avoid having to save and restore
23416 * ^{^RE_DEBUG_FLAGS} around the execution. save_scalar() perhaps
23417 * could be used. Another suggestion is to keep the authoritative
23418 * value of the debug flags in a thread-local variable and add set/get
23419 * magic to ${^RE_DEBUG_FLAGS} to keep the C level variable up to date.
23420 * Still another is to pass a flag, say in the engine's intflags that
23421 * would be checked each time before doing the debug output */
23425 assert(subpattern_re); /* Should have died if didn't compile successfully */
23426 return subpattern_re;
23430 S_execute_wildcard(pTHX_ REGEXP * const prog, char* stringarg, char *strend,
23431 char *strbeg, SSize_t minend, SV *screamer, U32 nosave)
23434 DECLARE_AND_GET_RE_DEBUG_FLAGS;
23436 PERL_ARGS_ASSERT_EXECUTE_WILDCARD;
23440 /* The compilation has set things up so that if the program doesn't want to
23441 * see the wildcard matching procedure, it will get the core execution
23442 * engine, which is subject only to -Dr. So we have to turn that off
23443 * around this procedure */
23444 if (! isDEBUG_WILDCARD) {
23445 /* Note! Casts away 'volatile' */
23447 PL_debug &= ~ DEBUG_r_FLAG;
23450 result = CALLREGEXEC(prog, stringarg, strend, strbeg, minend, screamer,
23458 S_handle_user_defined_property(pTHX_
23460 /* Parses the contents of a user-defined property definition; returning the
23461 * expanded definition if possible. If so, the return is an inversion
23464 * If there are subroutines that are part of the expansion and which aren't
23465 * known at the time of the call to this function, this returns what
23466 * parse_uniprop_string() returned for the first one encountered.
23468 * If an error was found, NULL is returned, and 'msg' gets a suitable
23469 * message appended to it. (Appending allows the back trace of how we got
23470 * to the faulty definition to be displayed through nested calls of
23471 * user-defined subs.)
23473 * The caller IS responsible for freeing any returned SV.
23475 * The syntax of the contents is pretty much described in perlunicode.pod,
23476 * but we also allow comments on each line */
23478 const char * name, /* Name of property */
23479 const STRLEN name_len, /* The name's length in bytes */
23480 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
23481 const bool to_fold, /* ? Is this under /i */
23482 const bool runtime, /* ? Are we in compile- or run-time */
23483 const bool deferrable, /* Is it ok for this property's full definition
23484 to be deferred until later? */
23485 SV* contents, /* The property's definition */
23486 bool *user_defined_ptr, /* This will be set TRUE as we wouldn't be
23487 getting called unless this is thought to be
23488 a user-defined property */
23489 SV * msg, /* Any error or warning msg(s) are appended to
23491 const STRLEN level) /* Recursion level of this call */
23494 const char * string = SvPV_const(contents, len);
23495 const char * const e = string + len;
23496 const bool is_contents_utf8 = cBOOL(SvUTF8(contents));
23497 const STRLEN msgs_length_on_entry = SvCUR(msg);
23499 const char * s0 = string; /* Points to first byte in the current line
23500 being parsed in 'string' */
23501 const char overflow_msg[] = "Code point too large in \"";
23502 SV* running_definition = NULL;
23504 PERL_ARGS_ASSERT_HANDLE_USER_DEFINED_PROPERTY;
23506 *user_defined_ptr = TRUE;
23508 /* Look at each line */
23510 const char * s; /* Current byte */
23511 char op = '+'; /* Default operation is 'union' */
23512 IV min = 0; /* range begin code point */
23513 IV max = -1; /* and range end */
23514 SV* this_definition;
23516 /* Skip comment lines */
23518 s0 = strchr(s0, '\n');
23526 /* For backcompat, allow an empty first line */
23532 /* First character in the line may optionally be the operation */
23541 /* If the line is one or two hex digits separated by blank space, its
23542 * a range; otherwise it is either another user-defined property or an
23547 if (! isXDIGIT(*s)) {
23548 goto check_if_property;
23551 do { /* Each new hex digit will add 4 bits. */
23552 if (min > ( (IV) MAX_LEGAL_CP >> 4)) {
23553 s = strchr(s, '\n');
23557 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23558 sv_catpv(msg, overflow_msg);
23559 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
23560 UTF8fARG(is_contents_utf8, s - s0, s0));
23561 sv_catpvs(msg, "\"");
23562 goto return_failure;
23565 /* Accumulate this digit into the value */
23566 min = (min << 4) + READ_XDIGIT(s);
23567 } while (isXDIGIT(*s));
23569 while (isBLANK(*s)) { s++; }
23571 /* We allow comments at the end of the line */
23573 s = strchr(s, '\n');
23579 else if (s < e && *s != '\n') {
23580 if (! isXDIGIT(*s)) {
23581 goto check_if_property;
23584 /* Look for the high point of the range */
23587 if (max > ( (IV) MAX_LEGAL_CP >> 4)) {
23588 s = strchr(s, '\n');
23592 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23593 sv_catpv(msg, overflow_msg);
23594 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
23595 UTF8fARG(is_contents_utf8, s - s0, s0));
23596 sv_catpvs(msg, "\"");
23597 goto return_failure;
23600 max = (max << 4) + READ_XDIGIT(s);
23601 } while (isXDIGIT(*s));
23603 while (isBLANK(*s)) { s++; }
23606 s = strchr(s, '\n');
23611 else if (s < e && *s != '\n') {
23612 goto check_if_property;
23616 if (max == -1) { /* The line only had one entry */
23619 else if (max < min) {
23620 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23621 sv_catpvs(msg, "Illegal range in \"");
23622 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
23623 UTF8fARG(is_contents_utf8, s - s0, s0));
23624 sv_catpvs(msg, "\"");
23625 goto return_failure;
23628 # if 0 /* See explanation at definition above of get_extended_utf8_msg() */
23630 if ( UNICODE_IS_PERL_EXTENDED(min)
23631 || UNICODE_IS_PERL_EXTENDED(max))
23633 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23635 /* If both code points are non-portable, warn only on the lower
23637 sv_catpv(msg, get_extended_utf8_msg(
23638 (UNICODE_IS_PERL_EXTENDED(min))
23640 sv_catpvs(msg, " in \"");
23641 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
23642 UTF8fARG(is_contents_utf8, s - s0, s0));
23643 sv_catpvs(msg, "\"");
23648 /* Here, this line contains a legal range */
23649 this_definition = sv_2mortal(_new_invlist(2));
23650 this_definition = _add_range_to_invlist(this_definition, min, max);
23655 /* Here it isn't a legal range line. See if it is a legal property
23656 * line. First find the end of the meat of the line */
23657 s = strpbrk(s, "#\n");
23662 /* Ignore trailing blanks in keeping with the requirements of
23663 * parse_uniprop_string() */
23665 while (s > s0 && isBLANK_A(*s)) {
23670 this_definition = parse_uniprop_string(s0, s - s0,
23671 is_utf8, to_fold, runtime,
23674 user_defined_ptr, msg,
23676 ? level /* Don't increase level
23677 if input is empty */
23680 if (this_definition == NULL) {
23681 goto return_failure; /* 'msg' should have had the reason
23682 appended to it by the above call */
23685 if (! is_invlist(this_definition)) { /* Unknown at this time */
23686 return newSVsv(this_definition);
23690 s = strchr(s, '\n');
23700 _invlist_union(running_definition, this_definition,
23701 &running_definition);
23704 _invlist_subtract(running_definition, this_definition,
23705 &running_definition);
23708 _invlist_intersection(running_definition, this_definition,
23709 &running_definition);
23712 _invlist_union_complement_2nd(running_definition,
23713 this_definition, &running_definition);
23716 Perl_croak(aTHX_ "panic: %s: %d: Unexpected operation %d",
23717 __FILE__, __LINE__, op);
23721 /* Position past the '\n' */
23723 } /* End of loop through the lines of 'contents' */
23725 /* Here, we processed all the lines in 'contents' without error. If we
23726 * didn't add any warnings, simply return success */
23727 if (msgs_length_on_entry == SvCUR(msg)) {
23729 /* If the expansion was empty, the answer isn't nothing: its an empty
23730 * inversion list */
23731 if (running_definition == NULL) {
23732 running_definition = _new_invlist(1);
23735 return running_definition;
23738 /* Otherwise, add some explanatory text, but we will return success */
23742 running_definition = NULL;
23746 if (name_len > 0) {
23747 sv_catpvs(msg, " in expansion of ");
23748 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8, name_len, name));
23751 return running_definition;
23754 /* As explained below, certain operations need to take place in the first
23755 * thread created. These macros switch contexts */
23756 # ifdef USE_ITHREADS
23757 # define DECLARATION_FOR_GLOBAL_CONTEXT \
23758 PerlInterpreter * save_aTHX = aTHX;
23759 # define SWITCH_TO_GLOBAL_CONTEXT \
23760 PERL_SET_CONTEXT((aTHX = PL_user_def_props_aTHX))
23761 # define RESTORE_CONTEXT PERL_SET_CONTEXT((aTHX = save_aTHX));
23762 # define CUR_CONTEXT aTHX
23763 # define ORIGINAL_CONTEXT save_aTHX
23765 # define DECLARATION_FOR_GLOBAL_CONTEXT dNOOP
23766 # define SWITCH_TO_GLOBAL_CONTEXT NOOP
23767 # define RESTORE_CONTEXT NOOP
23768 # define CUR_CONTEXT NULL
23769 # define ORIGINAL_CONTEXT NULL
23773 S_delete_recursion_entry(pTHX_ void *key)
23775 /* Deletes the entry used to detect recursion when expanding user-defined
23776 * properties. This is a function so it can be set up to be called even if
23777 * the program unexpectedly quits */
23779 SV ** current_entry;
23780 const STRLEN key_len = strlen((const char *) key);
23781 DECLARATION_FOR_GLOBAL_CONTEXT;
23783 SWITCH_TO_GLOBAL_CONTEXT;
23785 /* If the entry is one of these types, it is a permanent entry, and not the
23786 * one used to detect recursions. This function should delete only the
23787 * recursion entry */
23788 current_entry = hv_fetch(PL_user_def_props, (const char *) key, key_len, 0);
23790 && ! is_invlist(*current_entry)
23791 && ! SvPOK(*current_entry))
23793 (void) hv_delete(PL_user_def_props, (const char *) key, key_len,
23801 S_get_fq_name(pTHX_
23802 const char * const name, /* The first non-blank in the \p{}, \P{} */
23803 const Size_t name_len, /* Its length in bytes, not including any trailing space */
23804 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
23805 const bool has_colon_colon
23808 /* Returns a mortal SV containing the fully qualified version of the input
23813 fq_name = newSVpvs_flags("", SVs_TEMP);
23815 /* Use the current package if it wasn't included in our input */
23816 if (! has_colon_colon) {
23817 const HV * pkg = (IN_PERL_COMPILETIME)
23819 : CopSTASH(PL_curcop);
23820 const char* pkgname = HvNAME(pkg);
23822 Perl_sv_catpvf(aTHX_ fq_name, "%" UTF8f,
23823 UTF8fARG(is_utf8, strlen(pkgname), pkgname));
23824 sv_catpvs(fq_name, "::");
23827 Perl_sv_catpvf(aTHX_ fq_name, "%" UTF8f,
23828 UTF8fARG(is_utf8, name_len, name));
23833 S_parse_uniprop_string(pTHX_
23835 /* Parse the interior of a \p{}, \P{}. Returns its definition if knowable
23836 * now. If so, the return is an inversion list.
23838 * If the property is user-defined, it is a subroutine, which in turn
23839 * may call other subroutines. This function will call the whole nest of
23840 * them to get the definition they return; if some aren't known at the time
23841 * of the call to this function, the fully qualified name of the highest
23842 * level sub is returned. It is an error to call this function at runtime
23843 * without every sub defined.
23845 * If an error was found, NULL is returned, and 'msg' gets a suitable
23846 * message appended to it. (Appending allows the back trace of how we got
23847 * to the faulty definition to be displayed through nested calls of
23848 * user-defined subs.)
23850 * The caller should NOT try to free any returned inversion list.
23852 * Other parameters will be set on return as described below */
23854 const char * const name, /* The first non-blank in the \p{}, \P{} */
23855 Size_t name_len, /* Its length in bytes, not including any
23857 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
23858 const bool to_fold, /* ? Is this under /i */
23859 const bool runtime, /* TRUE if this is being called at run time */
23860 const bool deferrable, /* TRUE if it's ok for the definition to not be
23861 known at this call */
23862 AV ** strings, /* To return string property values, like named
23864 bool *user_defined_ptr, /* Upon return from this function it will be
23865 set to TRUE if any component is a
23866 user-defined property */
23867 SV * msg, /* Any error or warning msg(s) are appended to
23869 const STRLEN level) /* Recursion level of this call */
23871 char* lookup_name; /* normalized name for lookup in our tables */
23872 unsigned lookup_len; /* Its length */
23873 enum { Not_Strict = 0, /* Some properties have stricter name */
23874 Strict, /* normalization rules, which we decide */
23875 As_Is /* upon based on parsing */
23876 } stricter = Not_Strict;
23878 /* nv= or numeric_value=, or possibly one of the cjk numeric properties
23879 * (though it requires extra effort to download them from Unicode and
23880 * compile perl to know about them) */
23881 bool is_nv_type = FALSE;
23883 unsigned int i, j = 0;
23884 int equals_pos = -1; /* Where the '=' is found, or negative if none */
23885 int slash_pos = -1; /* Where the '/' is found, or negative if none */
23886 int table_index = 0; /* The entry number for this property in the table
23887 of all Unicode property names */
23888 bool starts_with_Is = FALSE; /* ? Does the name start with 'Is' */
23889 Size_t lookup_offset = 0; /* Used to ignore the first few characters of
23890 the normalized name in certain situations */
23891 Size_t non_pkg_begin = 0; /* Offset of first byte in 'name' that isn't
23892 part of a package name */
23893 Size_t lun_non_pkg_begin = 0; /* Similarly for 'lookup_name' */
23894 bool could_be_user_defined = TRUE; /* ? Could this be a user-defined
23895 property rather than a Unicode
23897 SV * prop_definition = NULL; /* The returned definition of 'name' or NULL
23898 if an error. If it is an inversion list,
23899 it is the definition. Otherwise it is a
23900 string containing the fully qualified sub
23902 SV * fq_name = NULL; /* For user-defined properties, the fully
23904 bool invert_return = FALSE; /* ? Do we need to complement the result before
23906 bool stripped_utf8_pkg = FALSE; /* Set TRUE if the input includes an
23907 explicit utf8:: package that we strip
23909 /* The expansion of properties that could be either user-defined or
23910 * official unicode ones is deferred until runtime, including a marker for
23911 * those that might be in the latter category. This boolean indicates if
23912 * we've seen that marker. If not, what we're parsing can't be such an
23913 * official Unicode property whose expansion was deferred */
23914 bool could_be_deferred_official = FALSE;
23916 PERL_ARGS_ASSERT_PARSE_UNIPROP_STRING;
23918 /* The input will be normalized into 'lookup_name' */
23919 Newx(lookup_name, name_len, char);
23920 SAVEFREEPV(lookup_name);
23922 /* Parse the input. */
23923 for (i = 0; i < name_len; i++) {
23924 char cur = name[i];
23926 /* Most of the characters in the input will be of this ilk, being parts
23928 if (isIDCONT_A(cur)) {
23930 /* Case differences are ignored. Our lookup routine assumes
23931 * everything is lowercase, so normalize to that */
23932 if (isUPPER_A(cur)) {
23933 lookup_name[j++] = toLOWER_A(cur);
23937 if (cur == '_') { /* Don't include these in the normalized name */
23941 lookup_name[j++] = cur;
23943 /* The first character in a user-defined name must be of this type.
23945 if (i - non_pkg_begin == 0 && ! isIDFIRST_A(cur)) {
23946 could_be_user_defined = FALSE;
23952 /* Here, the character is not something typically in a name, But these
23953 * two types of characters (and the '_' above) can be freely ignored in
23954 * most situations. Later it may turn out we shouldn't have ignored
23955 * them, and we have to reparse, but we don't have enough information
23956 * yet to make that decision */
23957 if (cur == '-' || isSPACE_A(cur)) {
23958 could_be_user_defined = FALSE;
23962 /* An equals sign or single colon mark the end of the first part of
23963 * the property name */
23965 || (cur == ':' && (i >= name_len - 1 || name[i+1] != ':')))
23967 lookup_name[j++] = '='; /* Treat the colon as an '=' */
23968 equals_pos = j; /* Note where it occurred in the input */
23969 could_be_user_defined = FALSE;
23973 /* If this looks like it is a marker we inserted at compile time,
23974 * set a flag and otherwise ignore it. If it isn't in the final
23975 * position, keep it as it would have been user input. */
23976 if ( UNLIKELY(cur == DEFERRED_COULD_BE_OFFICIAL_MARKERc)
23978 && could_be_user_defined
23979 && i == name_len - 1)
23982 could_be_deferred_official = TRUE;
23986 /* Otherwise, this character is part of the name. */
23987 lookup_name[j++] = cur;
23989 /* Here it isn't a single colon, so if it is a colon, it must be a
23993 /* A double colon should be a package qualifier. We note its
23994 * position and continue. Note that one could have
23995 * pkg1::pkg2::...::foo
23996 * so that the position at the end of the loop will be just after
23997 * the final qualifier */
24000 non_pkg_begin = i + 1;
24001 lookup_name[j++] = ':';
24002 lun_non_pkg_begin = j;
24004 else { /* Only word chars (and '::') can be in a user-defined name */
24005 could_be_user_defined = FALSE;
24007 } /* End of parsing through the lhs of the property name (or all of it if
24010 # define STRLENs(s) (sizeof("" s "") - 1)
24012 /* If there is a single package name 'utf8::', it is ambiguous. It could
24013 * be for a user-defined property, or it could be a Unicode property, as
24014 * all of them are considered to be for that package. For the purposes of
24015 * parsing the rest of the property, strip it off */
24016 if (non_pkg_begin == STRLENs("utf8::") && memBEGINPs(name, name_len, "utf8::")) {
24017 lookup_name += STRLENs("utf8::");
24018 j -= STRLENs("utf8::");
24019 equals_pos -= STRLENs("utf8::");
24020 stripped_utf8_pkg = TRUE;
24023 /* Here, we are either done with the whole property name, if it was simple;
24024 * or are positioned just after the '=' if it is compound. */
24026 if (equals_pos >= 0) {
24027 assert(stricter == Not_Strict); /* We shouldn't have set this yet */
24029 /* Space immediately after the '=' is ignored */
24031 for (; i < name_len; i++) {
24032 if (! isSPACE_A(name[i])) {
24037 /* Most punctuation after the equals indicates a subpattern, like
24039 if ( isPUNCT_A(name[i])
24044 /* A backslash means the real delimitter is the next character,
24045 * but it must be punctuation */
24046 && (name[i] != '\\' || (i < name_len && isPUNCT_A(name[i+1]))))
24048 bool special_property = memEQs(lookup_name, j - 1, "name")
24049 || memEQs(lookup_name, j - 1, "na");
24050 if (! special_property) {
24051 /* Find the property. The table includes the equals sign, so
24052 * we use 'j' as-is */
24053 table_index = do_uniprop_match(lookup_name, j);
24055 if (special_property || table_index) {
24056 REGEXP * subpattern_re;
24057 char open = name[i++];
24059 const char * pos_in_brackets;
24060 const char * const * prop_values;
24063 /* Backslash => delimitter is the character following. We
24064 * already checked that it is punctuation */
24065 if (open == '\\') {
24070 /* This data structure is constructed so that the matching
24071 * closing bracket is 3 past its matching opening. The second
24072 * set of closing is so that if the opening is something like
24073 * ']', the closing will be that as well. Something similar is
24074 * done in toke.c */
24075 pos_in_brackets = memCHRs("([<)]>)]>", open);
24076 close = (pos_in_brackets) ? pos_in_brackets[3] : open;
24079 || name[name_len-1] != close
24080 || (escaped && name[name_len-2] != '\\')
24081 /* Also make sure that there are enough characters.
24082 * e.g., '\\\' would show up incorrectly as legal even
24083 * though it is too short */
24084 || (SSize_t) (name_len - i - 1 - escaped) < 0)
24086 sv_catpvs(msg, "Unicode property wildcard not terminated");
24087 goto append_name_to_msg;
24090 Perl_ck_warner_d(aTHX_
24091 packWARN(WARN_EXPERIMENTAL__UNIPROP_WILDCARDS),
24092 "The Unicode property wildcards feature is experimental");
24094 if (special_property) {
24095 const char * error_msg;
24096 const char * revised_name = name + i;
24097 Size_t revised_name_len = name_len - (i + 1 + escaped);
24099 /* Currently, the only 'special_property' is name, which we
24100 * lookup in _charnames.pm */
24102 if (! load_charnames(newSVpvs("placeholder"),
24103 revised_name, revised_name_len,
24106 sv_catpv(msg, error_msg);
24107 goto append_name_to_msg;
24110 /* Farm this out to a function just to make the current
24111 * function less unwieldy */
24112 if (handle_names_wildcard(revised_name, revised_name_len,
24116 return prop_definition;
24122 prop_values = get_prop_values(table_index);
24124 /* Now create and compile the wildcard subpattern. Use /i
24125 * because the property values are supposed to match with case
24127 subpattern_re = compile_wildcard(name + i,
24128 name_len - i - 1 - escaped,
24132 /* For each legal property value, see if the supplied pattern
24134 while (*prop_values) {
24135 const char * const entry = *prop_values;
24136 const Size_t len = strlen(entry);
24137 SV* entry_sv = newSVpvn_flags(entry, len, SVs_TEMP);
24139 if (execute_wildcard(subpattern_re,
24141 (char *) entry + len,
24145 { /* Here, matched. Add to the returned list */
24146 Size_t total_len = j + len;
24147 SV * sub_invlist = NULL;
24148 char * this_string;
24150 /* We know this is a legal \p{property=value}. Call
24151 * the function to return the list of code points that
24153 Newxz(this_string, total_len + 1, char);
24154 Copy(lookup_name, this_string, j, char);
24155 my_strlcat(this_string, entry, total_len + 1);
24156 SAVEFREEPV(this_string);
24157 sub_invlist = parse_uniprop_string(this_string,
24167 _invlist_union(prop_definition, sub_invlist,
24171 prop_values++; /* Next iteration, look at next propvalue */
24172 } /* End of looking through property values; (the data
24173 structure is terminated by a NULL ptr) */
24175 SvREFCNT_dec_NN(subpattern_re);
24177 if (prop_definition) {
24178 return prop_definition;
24181 sv_catpvs(msg, "No Unicode property value wildcard matches:");
24182 goto append_name_to_msg;
24185 /* Here's how khw thinks we should proceed to handle the properties
24186 * not yet done: Bidi Mirroring Glyph can map to ""
24187 Bidi Paired Bracket can map to ""
24188 Case Folding (both full and simple)
24189 Shouldn't /i be good enough for Full
24190 Decomposition Mapping
24191 Equivalent Unified Ideograph can map to ""
24192 Lowercase Mapping (both full and simple)
24193 NFKC Case Fold can map to ""
24194 Titlecase Mapping (both full and simple)
24195 Uppercase Mapping (both full and simple)
24196 * Handle these the same way Name is done, using say, _wild.pm, but
24197 * having both loose and full, like in charclass_invlists.h.
24198 * Perhaps move block and script to that as they are somewhat large
24199 * in charclass_invlists.h.
24200 * For properties where the default is the code point itself, such
24201 * as any of the case changing mappings, the string would otherwise
24202 * consist of all Unicode code points in UTF-8 strung together.
24203 * This would be impractical. So instead, examine their compiled
24204 * pattern, looking at the ssc. If none, reject the pattern as an
24205 * error. Otherwise run the pattern against every code point in
24206 * the ssc. The ssc is kind of like tr18's 3.9 Possible Match Sets
24207 * And it might be good to create an API to return the ssc.
24208 * Or handle them like the algorithmic names are done
24210 } /* End of is a wildcard subppattern */
24212 /* \p{name=...} is handled specially. Instead of using the normal
24213 * mechanism involving charclass_invlists.h, it uses _charnames.pm
24214 * which has the necessary (huge) data accessible to it, and which
24215 * doesn't get loaded unless necessary. The legal syntax for names is
24216 * somewhat different than other properties due both to the vagaries of
24217 * a few outlier official names, and the fact that only a few ASCII
24218 * characters are permitted in them */
24219 if ( memEQs(lookup_name, j - 1, "name")
24220 || memEQs(lookup_name, j - 1, "na"))
24225 const char * error_msg;
24227 SV * character_name;
24228 STRLEN character_len;
24233 /* Since the RHS (after skipping initial space) is passed unchanged
24234 * to charnames, and there are different criteria for what are
24235 * legal characters in the name, just parse it here. A character
24236 * name must begin with an ASCII alphabetic */
24237 if (! isALPHA(name[i])) {
24240 lookup_name[j++] = name[i];
24242 for (++i; i < name_len; i++) {
24243 /* Official names can only be in the ASCII range, and only
24244 * certain characters */
24245 if (! isASCII(name[i]) || ! isCHARNAME_CONT(name[i])) {
24248 lookup_name[j++] = name[i];
24251 /* Finished parsing, save the name into an SV */
24252 character_name = newSVpvn(lookup_name + equals_pos, j - equals_pos);
24254 /* Make sure _charnames is loaded. (The parameters give context
24255 * for any errors generated */
24256 table = load_charnames(character_name, name, name_len, &error_msg);
24257 if (table == NULL) {
24258 sv_catpv(msg, error_msg);
24259 goto append_name_to_msg;
24262 lookup_loose = get_cvs("_charnames::_loose_regcomp_lookup", 0);
24263 if (! lookup_loose) {
24265 "panic: Can't find '_charnames::_loose_regcomp_lookup");
24268 PUSHSTACKi(PERLSI_REGCOMP);
24274 XPUSHs(character_name);
24276 call_sv(MUTABLE_SV(lookup_loose), G_SCALAR);
24281 SvREFCNT_inc_simple_void_NN(character);
24288 if (! SvOK(character)) {
24292 cp = valid_utf8_to_uvchr((U8 *) SvPVX(character), &character_len);
24293 if (character_len == SvCUR(character)) {
24294 prop_definition = add_cp_to_invlist(NULL, cp);
24299 /* First of the remaining characters in the string. */
24300 char * remaining = SvPVX(character) + character_len;
24302 if (strings == NULL) {
24303 goto failed; /* XXX Perhaps a specific msg instead, like
24304 'not available here' */
24307 if (*strings == NULL) {
24308 *strings = newAV();
24311 this_string = newAV();
24312 av_push(this_string, newSVuv(cp));
24315 cp = valid_utf8_to_uvchr((U8 *) remaining, &character_len);
24316 av_push(this_string, newSVuv(cp));
24317 remaining += character_len;
24318 } while (remaining < SvEND(character));
24320 av_push(*strings, (SV *) this_string);
24323 return prop_definition;
24326 /* Certain properties whose values are numeric need special handling.
24327 * They may optionally be prefixed by 'is'. Ignore that prefix for the
24328 * purposes of checking if this is one of those properties */
24329 if (memBEGINPs(lookup_name, j, "is")) {
24333 /* Then check if it is one of these specially-handled properties. The
24334 * possibilities are hard-coded because easier this way, and the list
24335 * is unlikely to change.
24337 * All numeric value type properties are of this ilk, and are also
24338 * special in a different way later on. So find those first. There
24339 * are several numeric value type properties in the Unihan DB (which is
24340 * unlikely to be compiled with perl, but we handle it here in case it
24341 * does get compiled). They all end with 'numeric'. The interiors
24342 * aren't checked for the precise property. This would stop working if
24343 * a cjk property were to be created that ended with 'numeric' and
24344 * wasn't a numeric type */
24345 is_nv_type = memEQs(lookup_name + lookup_offset,
24346 j - 1 - lookup_offset, "numericvalue")
24347 || memEQs(lookup_name + lookup_offset,
24348 j - 1 - lookup_offset, "nv")
24349 || ( memENDPs(lookup_name + lookup_offset,
24350 j - 1 - lookup_offset, "numeric")
24351 && ( memBEGINPs(lookup_name + lookup_offset,
24352 j - 1 - lookup_offset, "cjk")
24353 || memBEGINPs(lookup_name + lookup_offset,
24354 j - 1 - lookup_offset, "k")));
24356 || memEQs(lookup_name + lookup_offset,
24357 j - 1 - lookup_offset, "canonicalcombiningclass")
24358 || memEQs(lookup_name + lookup_offset,
24359 j - 1 - lookup_offset, "ccc")
24360 || memEQs(lookup_name + lookup_offset,
24361 j - 1 - lookup_offset, "age")
24362 || memEQs(lookup_name + lookup_offset,
24363 j - 1 - lookup_offset, "in")
24364 || memEQs(lookup_name + lookup_offset,
24365 j - 1 - lookup_offset, "presentin"))
24369 /* Since the stuff after the '=' is a number, we can't throw away
24370 * '-' willy-nilly, as those could be a minus sign. Other stricter
24371 * rules also apply. However, these properties all can have the
24372 * rhs not be a number, in which case they contain at least one
24373 * alphabetic. In those cases, the stricter rules don't apply.
24374 * But the numeric type properties can have the alphas [Ee] to
24375 * signify an exponent, and it is still a number with stricter
24376 * rules. So look for an alpha that signifies not-strict */
24378 for (k = i; k < name_len; k++) {
24379 if ( isALPHA_A(name[k])
24380 && (! is_nv_type || ! isALPHA_FOLD_EQ(name[k], 'E')))
24382 stricter = Not_Strict;
24390 /* A number may have a leading '+' or '-'. The latter is retained
24392 if (name[i] == '+') {
24395 else if (name[i] == '-') {
24396 lookup_name[j++] = '-';
24400 /* Skip leading zeros including single underscores separating the
24401 * zeros, or between the final leading zero and the first other
24403 for (; i < name_len - 1; i++) {
24404 if ( name[i] != '0'
24405 && (name[i] != '_' || ! isDIGIT_A(name[i+1])))
24412 else { /* No '=' */
24414 /* Only a few properties without an '=' should be parsed with stricter
24415 * rules. The list is unlikely to change. */
24416 if ( memBEGINPs(lookup_name, j, "perl")
24417 && memNEs(lookup_name + 4, j - 4, "space")
24418 && memNEs(lookup_name + 4, j - 4, "word"))
24422 /* We set the inputs back to 0 and the code below will reparse,
24428 /* Here, we have either finished the property, or are positioned to parse
24429 * the remainder, and we know if stricter rules apply. Finish out, if not
24431 for (; i < name_len; i++) {
24432 char cur = name[i];
24434 /* In all instances, case differences are ignored, and we normalize to
24436 if (isUPPER_A(cur)) {
24437 lookup_name[j++] = toLOWER(cur);
24441 /* An underscore is skipped, but not under strict rules unless it
24442 * separates two digits */
24445 && ( i == 0 || (int) i == equals_pos || i == name_len- 1
24446 || ! isDIGIT_A(name[i-1]) || ! isDIGIT_A(name[i+1])))
24448 lookup_name[j++] = '_';
24453 /* Hyphens are skipped except under strict */
24454 if (cur == '-' && ! stricter) {
24458 /* XXX Bug in documentation. It says white space skipped adjacent to
24459 * non-word char. Maybe we should, but shouldn't skip it next to a dot
24461 if (isSPACE_A(cur) && ! stricter) {
24465 lookup_name[j++] = cur;
24467 /* Unless this is a non-trailing slash, we are done with it */
24468 if (i >= name_len - 1 || cur != '/') {
24474 /* A slash in the 'numeric value' property indicates that what follows
24475 * is a denominator. It can have a leading '+' and '0's that should be
24476 * skipped. But we have never allowed a negative denominator, so treat
24477 * a minus like every other character. (No need to rule out a second
24478 * '/', as that won't match anything anyway */
24481 if (i < name_len && name[i] == '+') {
24485 /* Skip leading zeros including underscores separating digits */
24486 for (; i < name_len - 1; i++) {
24487 if ( name[i] != '0'
24488 && (name[i] != '_' || ! isDIGIT_A(name[i+1])))
24494 /* Store the first real character in the denominator */
24495 if (i < name_len) {
24496 lookup_name[j++] = name[i];
24501 /* Here are completely done parsing the input 'name', and 'lookup_name'
24502 * contains a copy, normalized.
24504 * This special case is grandfathered in: 'L_' and 'GC=L_' are accepted and
24505 * different from without the underscores. */
24506 if ( ( UNLIKELY(memEQs(lookup_name, j, "l"))
24507 || UNLIKELY(memEQs(lookup_name, j, "gc=l")))
24508 && UNLIKELY(name[name_len-1] == '_'))
24510 lookup_name[j++] = '&';
24513 /* If the original input began with 'In' or 'Is', it could be a subroutine
24514 * call to a user-defined property instead of a Unicode property name. */
24515 if ( name_len - non_pkg_begin > 2
24516 && name[non_pkg_begin+0] == 'I'
24517 && (name[non_pkg_begin+1] == 'n' || name[non_pkg_begin+1] == 's'))
24519 /* Names that start with In have different characterstics than those
24520 * that start with Is */
24521 if (name[non_pkg_begin+1] == 's') {
24522 starts_with_Is = TRUE;
24526 could_be_user_defined = FALSE;
24529 if (could_be_user_defined) {
24532 /* If the user defined property returns the empty string, it could
24533 * easily be because the pattern is being compiled before the data it
24534 * actually needs to compile is available. This could be argued to be
24535 * a bug in the perl code, but this is a change of behavior for Perl,
24536 * so we handle it. This means that intentionally returning nothing
24537 * will not be resolved until runtime */
24538 bool empty_return = FALSE;
24540 /* Here, the name could be for a user defined property, which are
24541 * implemented as subs. */
24542 user_sub = get_cvn_flags(name, name_len, 0);
24545 /* Here, the property name could be a user-defined one, but there
24546 * is no subroutine to handle it (as of now). Defer handling it
24547 * until runtime. Otherwise, a block defined by Unicode in a later
24548 * release would get the synonym InFoo added for it, and existing
24549 * code that used that name would suddenly break if it referred to
24550 * the property before the sub was declared. See [perl #134146] */
24552 goto definition_deferred;
24555 /* Here, we are at runtime, and didn't find the user property. It
24556 * could be an official property, but only if no package was
24557 * specified, or just the utf8:: package. */
24558 if (could_be_deferred_official) {
24559 lookup_name += lun_non_pkg_begin;
24560 j -= lun_non_pkg_begin;
24562 else if (! stripped_utf8_pkg) {
24563 goto unknown_user_defined;
24566 /* Drop down to look up in the official properties */
24569 const char insecure[] = "Insecure user-defined property";
24571 /* Here, there is a sub by the correct name. Normally we call it
24572 * to get the property definition */
24574 SV * user_sub_sv = MUTABLE_SV(user_sub);
24575 SV * error; /* Any error returned by calling 'user_sub' */
24576 SV * key; /* The key into the hash of user defined sub names
24579 SV ** saved_user_prop_ptr; /* Hash entry for this property */
24581 /* How many times to retry when another thread is in the middle of
24582 * expanding the same definition we want */
24583 PERL_INT_FAST8_T retry_countdown = 10;
24585 DECLARATION_FOR_GLOBAL_CONTEXT;
24587 /* If we get here, we know this property is user-defined */
24588 *user_defined_ptr = TRUE;
24590 /* We refuse to call a potentially tainted subroutine; returning an
24593 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24594 sv_catpvn(msg, insecure, sizeof(insecure) - 1);
24595 goto append_name_to_msg;
24598 /* In principal, we only call each subroutine property definition
24599 * once during the life of the program. This guarantees that the
24600 * property definition never changes. The results of the single
24601 * sub call are stored in a hash, which is used instead for future
24602 * references to this property. The property definition is thus
24603 * immutable. But, to allow the user to have a /i-dependent
24604 * definition, we call the sub once for non-/i, and once for /i,
24605 * should the need arise, passing the /i status as a parameter.
24607 * We start by constructing the hash key name, consisting of the
24608 * fully qualified subroutine name, preceded by the /i status, so
24609 * that there is a key for /i and a different key for non-/i */
24610 key = newSVpvn(((to_fold) ? "1" : "0"), 1);
24611 fq_name = S_get_fq_name(aTHX_ name, name_len, is_utf8,
24612 non_pkg_begin != 0);
24613 sv_catsv(key, fq_name);
24616 /* We only call the sub once throughout the life of the program
24617 * (with the /i, non-/i exception noted above). That means the
24618 * hash must be global and accessible to all threads. It is
24619 * created at program start-up, before any threads are created, so
24620 * is accessible to all children. But this creates some
24623 * 1) The keys can't be shared, or else problems arise; sharing is
24624 * turned off at hash creation time
24625 * 2) All SVs in it are there for the remainder of the life of the
24626 * program, and must be created in the same interpreter context
24627 * as the hash, or else they will be freed from the wrong pool
24628 * at global destruction time. This is handled by switching to
24629 * the hash's context to create each SV going into it, and then
24630 * immediately switching back
24631 * 3) All accesses to the hash must be controlled by a mutex, to
24632 * prevent two threads from getting an unstable state should
24633 * they simultaneously be accessing it. The code below is
24634 * crafted so that the mutex is locked whenever there is an
24635 * access and unlocked only when the next stable state is
24638 * The hash stores either the definition of the property if it was
24639 * valid, or, if invalid, the error message that was raised. We
24640 * use the type of SV to distinguish.
24642 * There's also the need to guard against the definition expansion
24643 * from infinitely recursing. This is handled by storing the aTHX
24644 * of the expanding thread during the expansion. Again the SV type
24645 * is used to distinguish this from the other two cases. If we
24646 * come to here and the hash entry for this property is our aTHX,
24647 * it means we have recursed, and the code assumes that we would
24648 * infinitely recurse, so instead stops and raises an error.
24649 * (Any recursion has always been treated as infinite recursion in
24652 * If instead, the entry is for a different aTHX, it means that
24653 * that thread has gotten here first, and hasn't finished expanding
24654 * the definition yet. We just have to wait until it is done. We
24655 * sleep and retry a few times, returning an error if the other
24656 * thread doesn't complete. */
24659 USER_PROP_MUTEX_LOCK;
24661 /* If we have an entry for this key, the subroutine has already
24662 * been called once with this /i status. */
24663 saved_user_prop_ptr = hv_fetch(PL_user_def_props,
24664 SvPVX(key), SvCUR(key), 0);
24665 if (saved_user_prop_ptr) {
24667 /* If the saved result is an inversion list, it is the valid
24668 * definition of this property */
24669 if (is_invlist(*saved_user_prop_ptr)) {
24670 prop_definition = *saved_user_prop_ptr;
24672 /* The SV in the hash won't be removed until global
24673 * destruction, so it is stable and we can unlock */
24674 USER_PROP_MUTEX_UNLOCK;
24676 /* The caller shouldn't try to free this SV */
24677 return prop_definition;
24680 /* Otherwise, if it is a string, it is the error message
24681 * that was returned when we first tried to evaluate this
24682 * property. Fail, and append the message */
24683 if (SvPOK(*saved_user_prop_ptr)) {
24684 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24685 sv_catsv(msg, *saved_user_prop_ptr);
24687 /* The SV in the hash won't be removed until global
24688 * destruction, so it is stable and we can unlock */
24689 USER_PROP_MUTEX_UNLOCK;
24694 assert(SvIOK(*saved_user_prop_ptr));
24696 /* Here, we have an unstable entry in the hash. Either another
24697 * thread is in the middle of expanding the property's
24698 * definition, or we are ourselves recursing. We use the aTHX
24699 * in it to distinguish */
24700 if (SvIV(*saved_user_prop_ptr) != PTR2IV(CUR_CONTEXT)) {
24702 /* Here, it's another thread doing the expanding. We've
24703 * looked as much as we are going to at the contents of the
24704 * hash entry. It's safe to unlock. */
24705 USER_PROP_MUTEX_UNLOCK;
24707 /* Retry a few times */
24708 if (retry_countdown-- > 0) {
24713 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24714 sv_catpvs(msg, "Timeout waiting for another thread to "
24716 goto append_name_to_msg;
24719 /* Here, we are recursing; don't dig any deeper */
24720 USER_PROP_MUTEX_UNLOCK;
24722 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24724 "Infinite recursion in user-defined property");
24725 goto append_name_to_msg;
24728 /* Here, this thread has exclusive control, and there is no entry
24729 * for this property in the hash. So we have the go ahead to
24730 * expand the definition ourselves. */
24732 PUSHSTACKi(PERLSI_REGCOMP);
24735 /* Create a temporary placeholder in the hash to detect recursion
24737 SWITCH_TO_GLOBAL_CONTEXT;
24738 placeholder= newSVuv(PTR2IV(ORIGINAL_CONTEXT));
24739 (void) hv_store_ent(PL_user_def_props, key, placeholder, 0);
24742 /* Now that we have a placeholder, we can let other threads
24744 USER_PROP_MUTEX_UNLOCK;
24746 /* Make sure the placeholder always gets destroyed */
24747 SAVEDESTRUCTOR_X(S_delete_recursion_entry, SvPVX(key));
24752 /* Call the user's function, with the /i status as a parameter.
24753 * Note that we have gone to a lot of trouble to keep this call
24754 * from being within the locked mutex region. */
24755 XPUSHs(boolSV(to_fold));
24758 /* The following block was taken from swash_init(). Presumably
24759 * they apply to here as well, though we no longer use a swash --
24763 /* We might get here via a subroutine signature which uses a utf8
24764 * parameter name, at which point PL_subname will have been set
24765 * but not yet used. */
24766 save_item(PL_subname);
24768 /* G_SCALAR guarantees a single return value */
24769 (void) call_sv(user_sub_sv, G_EVAL|G_SCALAR);
24774 if (TAINT_get || SvTRUE(error)) {
24775 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24776 if (SvTRUE(error)) {
24777 sv_catpvs(msg, "Error \"");
24778 sv_catsv(msg, error);
24779 sv_catpvs(msg, "\"");
24782 if (SvTRUE(error)) sv_catpvs(msg, "; ");
24783 sv_catpvn(msg, insecure, sizeof(insecure) - 1);
24786 if (name_len > 0) {
24787 sv_catpvs(msg, " in expansion of ");
24788 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8,
24794 prop_definition = NULL;
24797 SV * contents = POPs;
24799 /* The contents is supposed to be the expansion of the property
24800 * definition. If the definition is deferrable, and we got an
24801 * empty string back, set a flag to later defer it (after clean
24804 && (! SvPOK(contents) || SvCUR(contents) == 0))
24806 empty_return = TRUE;
24808 else { /* Otherwise, call a function to check for valid syntax,
24811 prop_definition = handle_user_defined_property(
24813 is_utf8, to_fold, runtime,
24815 contents, user_defined_ptr,
24821 /* Here, we have the results of the expansion. Delete the
24822 * placeholder, and if the definition is now known, replace it with
24823 * that definition. We need exclusive access to the hash, and we
24824 * can't let anyone else in, between when we delete the placeholder
24825 * and add the permanent entry */
24826 USER_PROP_MUTEX_LOCK;
24828 S_delete_recursion_entry(aTHX_ SvPVX(key));
24830 if ( ! empty_return
24831 && (! prop_definition || is_invlist(prop_definition)))
24833 /* If we got success we use the inversion list defining the
24834 * property; otherwise use the error message */
24835 SWITCH_TO_GLOBAL_CONTEXT;
24836 (void) hv_store_ent(PL_user_def_props,
24839 ? newSVsv(prop_definition)
24845 /* All done, and the hash now has a permanent entry for this
24846 * property. Give up exclusive control */
24847 USER_PROP_MUTEX_UNLOCK;
24853 if (empty_return) {
24854 goto definition_deferred;
24857 if (prop_definition) {
24859 /* If the definition is for something not known at this time,
24860 * we toss it, and go return the main property name, as that's
24861 * the one the user will be aware of */
24862 if (! is_invlist(prop_definition)) {
24863 SvREFCNT_dec_NN(prop_definition);
24864 goto definition_deferred;
24867 sv_2mortal(prop_definition);
24871 return prop_definition;
24873 } /* End of calling the subroutine for the user-defined property */
24874 } /* End of it could be a user-defined property */
24876 /* Here it wasn't a user-defined property that is known at this time. See
24877 * if it is a Unicode property */
24879 lookup_len = j; /* This is a more mnemonic name than 'j' */
24881 /* Get the index into our pointer table of the inversion list corresponding
24882 * to the property */
24883 table_index = do_uniprop_match(lookup_name, lookup_len);
24885 /* If it didn't find the property ... */
24886 if (table_index == 0) {
24888 /* Try again stripping off any initial 'Is'. This is because we
24889 * promise that an initial Is is optional. The same isn't true of
24890 * names that start with 'In'. Those can match only blocks, and the
24891 * lookup table already has those accounted for. The lookup table also
24892 * has already accounted for Perl extensions (without and = sign)
24893 * starting with 'i's'. */
24894 if (starts_with_Is && equals_pos >= 0) {
24900 table_index = do_uniprop_match(lookup_name, lookup_len);
24903 if (table_index == 0) {
24906 /* Here, we didn't find it. If not a numeric type property, and
24907 * can't be a user-defined one, it isn't a legal property */
24908 if (! is_nv_type) {
24909 if (! could_be_user_defined) {
24913 /* Here, the property name is legal as a user-defined one. At
24914 * compile time, it might just be that the subroutine for that
24915 * property hasn't been encountered yet, but at runtime, it's
24916 * an error to try to use an undefined one */
24917 if (! deferrable) {
24918 goto unknown_user_defined;;
24921 goto definition_deferred;
24922 } /* End of isn't a numeric type property */
24924 /* The numeric type properties need more work to decide. What we
24925 * do is make sure we have the number in canonical form and look
24928 if (slash_pos < 0) { /* No slash */
24930 /* When it isn't a rational, take the input, convert it to a
24931 * NV, then create a canonical string representation of that
24935 SSize_t value_len = lookup_len - equals_pos;
24937 /* Get the value */
24938 if ( value_len <= 0
24939 || my_atof3(lookup_name + equals_pos, &value,
24941 != lookup_name + lookup_len)
24946 /* If the value is an integer, the canonical value is integral
24948 if (Perl_ceil(value) == value) {
24949 canonical = Perl_form(aTHX_ "%.*s%.0" NVff,
24950 equals_pos, lookup_name, value);
24952 else { /* Otherwise, it is %e with a known precision */
24955 canonical = Perl_form(aTHX_ "%.*s%.*" NVef,
24956 equals_pos, lookup_name,
24957 PL_E_FORMAT_PRECISION, value);
24959 /* The exponent generated is expecting two digits, whereas
24960 * %e on some systems will generate three. Remove leading
24961 * zeros in excess of 2 from the exponent. We start
24962 * looking for them after the '=' */
24963 exp_ptr = strchr(canonical + equals_pos, 'e');
24965 char * cur_ptr = exp_ptr + 2; /* past the 'e[+-]' */
24966 SSize_t excess_exponent_len = strlen(cur_ptr) - 2;
24968 assert(*(cur_ptr - 1) == '-' || *(cur_ptr - 1) == '+');
24970 if (excess_exponent_len > 0) {
24971 SSize_t leading_zeros = strspn(cur_ptr, "0");
24972 SSize_t excess_leading_zeros
24973 = MIN(leading_zeros, excess_exponent_len);
24974 if (excess_leading_zeros > 0) {
24975 Move(cur_ptr + excess_leading_zeros,
24977 strlen(cur_ptr) - excess_leading_zeros
24978 + 1, /* Copy the NUL as well */
24985 else { /* Has a slash. Create a rational in canonical form */
24986 UV numerator, denominator, gcd, trial;
24987 const char * end_ptr;
24988 const char * sign = "";
24990 /* We can't just find the numerator, denominator, and do the
24991 * division, then use the method above, because that is
24992 * inexact. And the input could be a rational that is within
24993 * epsilon (given our precision) of a valid rational, and would
24994 * then incorrectly compare valid.
24996 * We're only interested in the part after the '=' */
24997 const char * this_lookup_name = lookup_name + equals_pos;
24998 lookup_len -= equals_pos;
24999 slash_pos -= equals_pos;
25001 /* Handle any leading minus */
25002 if (this_lookup_name[0] == '-') {
25004 this_lookup_name++;
25009 /* Convert the numerator to numeric */
25010 end_ptr = this_lookup_name + slash_pos;
25011 if (! grok_atoUV(this_lookup_name, &numerator, &end_ptr)) {
25015 /* It better have included all characters before the slash */
25016 if (*end_ptr != '/') {
25020 /* Set to look at just the denominator */
25021 this_lookup_name += slash_pos;
25022 lookup_len -= slash_pos;
25023 end_ptr = this_lookup_name + lookup_len;
25025 /* Convert the denominator to numeric */
25026 if (! grok_atoUV(this_lookup_name, &denominator, &end_ptr)) {
25030 /* It better be the rest of the characters, and don't divide by
25032 if ( end_ptr != this_lookup_name + lookup_len
25033 || denominator == 0)
25038 /* Get the greatest common denominator using
25039 http://en.wikipedia.org/wiki/Euclidean_algorithm */
25041 trial = denominator;
25042 while (trial != 0) {
25044 trial = gcd % trial;
25048 /* If already in lowest possible terms, we have already tried
25049 * looking this up */
25054 /* Reduce the rational, which should put it in canonical form
25057 denominator /= gcd;
25059 canonical = Perl_form(aTHX_ "%.*s%s%" UVuf "/%" UVuf,
25060 equals_pos, lookup_name, sign, numerator, denominator);
25063 /* Here, we have the number in canonical form. Try that */
25064 table_index = do_uniprop_match(canonical, strlen(canonical));
25065 if (table_index == 0) {
25068 } /* End of still didn't find the property in our table */
25069 } /* End of didn't find the property in our table */
25071 /* Here, we have a non-zero return, which is an index into a table of ptrs.
25072 * A negative return signifies that the real index is the absolute value,
25073 * but the result needs to be inverted */
25074 if (table_index < 0) {
25075 invert_return = TRUE;
25076 table_index = -table_index;
25079 /* Out-of band indices indicate a deprecated property. The proper index is
25080 * modulo it with the table size. And dividing by the table size yields
25081 * an offset into a table constructed by regen/mk_invlists.pl to contain
25082 * the corresponding warning message */
25083 if (table_index > MAX_UNI_KEYWORD_INDEX) {
25084 Size_t warning_offset = table_index / MAX_UNI_KEYWORD_INDEX;
25085 table_index %= MAX_UNI_KEYWORD_INDEX;
25086 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED),
25087 "Use of '%.*s' in \\p{} or \\P{} is deprecated because: %s",
25088 (int) name_len, name,
25089 get_deprecated_property_msg(warning_offset));
25092 /* In a few properties, a different property is used under /i. These are
25093 * unlikely to change, so are hard-coded here. */
25095 if ( table_index == UNI_XPOSIXUPPER
25096 || table_index == UNI_XPOSIXLOWER
25097 || table_index == UNI_TITLE)
25099 table_index = UNI_CASED;
25101 else if ( table_index == UNI_UPPERCASELETTER
25102 || table_index == UNI_LOWERCASELETTER
25103 # ifdef UNI_TITLECASELETTER /* Missing from early Unicodes */
25104 || table_index == UNI_TITLECASELETTER
25107 table_index = UNI_CASEDLETTER;
25109 else if ( table_index == UNI_POSIXUPPER
25110 || table_index == UNI_POSIXLOWER)
25112 table_index = UNI_POSIXALPHA;
25116 /* Create and return the inversion list */
25117 prop_definition = get_prop_definition(table_index);
25118 sv_2mortal(prop_definition);
25120 /* See if there is a private use override to add to this definition */
25122 COPHH * hinthash = (IN_PERL_COMPILETIME)
25123 ? CopHINTHASH_get(&PL_compiling)
25124 : CopHINTHASH_get(PL_curcop);
25125 SV * pu_overrides = cophh_fetch_pv(hinthash, "private_use", 0, 0);
25127 if (UNLIKELY(pu_overrides && SvPOK(pu_overrides))) {
25129 /* See if there is an element in the hints hash for this table */
25130 SV * pu_lookup = Perl_newSVpvf(aTHX_ "%d=", table_index);
25131 const char * pos = strstr(SvPVX(pu_overrides), SvPVX(pu_lookup));
25135 SV * pu_definition;
25137 SV * expanded_prop_definition =
25138 sv_2mortal(invlist_clone(prop_definition, NULL));
25140 /* If so, it's definition is the string from here to the next
25141 * \a character. And its format is the same as a user-defined
25143 pos += SvCUR(pu_lookup);
25144 pu_definition = newSVpvn(pos, strchr(pos, '\a') - pos);
25145 pu_invlist = handle_user_defined_property(lookup_name,
25148 0, /* Not folded */
25156 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
25157 sv_catpvs(msg, "Insecure private-use override");
25158 goto append_name_to_msg;
25161 /* For now, as a safety measure, make sure that it doesn't
25162 * override non-private use code points */
25163 _invlist_intersection(pu_invlist, PL_Private_Use, &pu_invlist);
25165 /* Add it to the list to be returned */
25166 _invlist_union(prop_definition, pu_invlist,
25167 &expanded_prop_definition);
25168 prop_definition = expanded_prop_definition;
25169 Perl_ck_warner_d(aTHX_ packWARN(WARN_EXPERIMENTAL__PRIVATE_USE), "The private_use feature is experimental");
25174 if (invert_return) {
25175 _invlist_invert(prop_definition);
25177 return prop_definition;
25179 unknown_user_defined:
25180 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
25181 sv_catpvs(msg, "Unknown user-defined property name");
25182 goto append_name_to_msg;
25185 if (non_pkg_begin != 0) {
25186 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
25187 sv_catpvs(msg, "Illegal user-defined property name");
25190 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
25191 sv_catpvs(msg, "Can't find Unicode property definition");
25195 append_name_to_msg:
25197 const char * prefix = (runtime && level == 0) ? " \\p{" : " \"";
25198 const char * suffix = (runtime && level == 0) ? "}" : "\"";
25200 sv_catpv(msg, prefix);
25201 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8, name_len, name));
25202 sv_catpv(msg, suffix);
25207 definition_deferred:
25210 bool is_qualified = non_pkg_begin != 0; /* If has "::" */
25212 /* Here it could yet to be defined, so defer evaluation of this until
25213 * its needed at runtime. We need the fully qualified property name to
25214 * avoid ambiguity */
25216 fq_name = S_get_fq_name(aTHX_ name, name_len, is_utf8,
25220 /* If it didn't come with a package, or the package is utf8::, this
25221 * actually could be an official Unicode property whose inclusion we
25222 * are deferring until runtime to make sure that it isn't overridden by
25223 * a user-defined property of the same name (which we haven't
25224 * encountered yet). Add a marker to indicate this possibility, for
25225 * use at such time when we first need the definition during pattern
25226 * matching execution */
25227 if (! is_qualified || memBEGINPs(name, non_pkg_begin, "utf8::")) {
25228 sv_catpvs(fq_name, DEFERRED_COULD_BE_OFFICIAL_MARKERs);
25231 /* We also need a trailing newline */
25232 sv_catpvs(fq_name, "\n");
25234 *user_defined_ptr = TRUE;
25240 S_handle_names_wildcard(pTHX_ const char * wname, /* wildcard name to match */
25241 const STRLEN wname_len, /* Its length */
25242 SV ** prop_definition,
25245 /* Deal with Name property wildcard subpatterns; returns TRUE if there were
25246 * any matches, adding them to prop_definition */
25250 CV * get_names_info; /* entry to charnames.pm to get info we need */
25251 SV * names_string; /* Contains all character names, except algo */
25252 SV * algorithmic_names; /* Contains info about algorithmically
25253 generated character names */
25254 REGEXP * subpattern_re; /* The user's pattern to match with */
25255 struct regexp * prog; /* The compiled pattern */
25256 char * all_names_start; /* lib/unicore/Name.pl string of every
25257 (non-algorithmic) character name */
25258 char * cur_pos; /* We match, effectively using /gc; this is
25259 where we are now */
25260 bool found_matches = FALSE; /* Did any name match so far? */
25261 SV * empty; /* For matching zero length names */
25262 SV * must_sv; /* Contains the substring, if any, that must be
25263 in a name for the subpattern to match */
25264 const char * must; /* The PV of 'must' */
25265 STRLEN must_len; /* And its length */
25266 SV * syllable_name = NULL; /* For Hangul syllables */
25267 const char hangul_prefix[] = "HANGUL SYLLABLE ";
25268 const STRLEN hangul_prefix_len = sizeof(hangul_prefix) - 1;
25270 /* By inspection, there are a maximum of 7 bytes in the suffix of a hangul
25271 * syllable name, and these are immutable and guaranteed by the Unicode
25272 * standard to never be extended */
25273 const STRLEN syl_max_len = hangul_prefix_len + 7;
25277 PERL_ARGS_ASSERT_HANDLE_NAMES_WILDCARD;
25279 /* Make sure _charnames is loaded. (The parameters give context
25280 * for any errors generated */
25281 get_names_info = get_cv("_charnames::_get_names_info", 0);
25282 if (! get_names_info) {
25283 Perl_croak(aTHX_ "panic: Can't find '_charnames::_get_names_info");
25286 /* Get the charnames data */
25287 PUSHSTACKi(PERLSI_REGCOMP);
25295 /* Special _charnames entry point that returns the info this routine
25297 call_sv(MUTABLE_SV(get_names_info), G_ARRAY);
25301 /* Data structure for names which end in their very own code points */
25302 algorithmic_names = POPs;
25303 SvREFCNT_inc_simple_void_NN(algorithmic_names);
25305 /* The lib/unicore/Name.pl string */
25306 names_string = POPs;
25307 SvREFCNT_inc_simple_void_NN(names_string);
25314 if ( ! SvROK(names_string)
25315 || ! SvROK(algorithmic_names))
25316 { /* Perhaps should panic instead XXX */
25317 SvREFCNT_dec(names_string);
25318 SvREFCNT_dec(algorithmic_names);
25322 names_string = sv_2mortal(SvRV(names_string));
25323 all_names_start = SvPVX(names_string);
25324 cur_pos = all_names_start;
25326 algorithmic_names= sv_2mortal(SvRV(algorithmic_names));
25328 /* Compile the subpattern consisting of the name being looked for */
25329 subpattern_re = compile_wildcard(wname, wname_len, FALSE /* /-i */ );
25331 must_sv = re_intuit_string(subpattern_re);
25333 /* regexec.c can free the re_intuit_string() return. GH #17734 */
25334 must_sv = sv_2mortal(newSVsv(must_sv));
25335 must = SvPV(must_sv, must_len);
25342 /* (Note: 'must' could contain a NUL. And yet we use strspn() below on it.
25343 * This works because the NUL causes the function to return early, thus
25344 * showing that there are characters in it other than the acceptable ones,
25345 * which is our desired result.) */
25347 prog = ReANY(subpattern_re);
25349 /* If only nothing is matched, skip to where empty names are looked for */
25350 if (prog->maxlen == 0) {
25354 /* And match against the string of all names /gc. Don't even try if it
25355 * must match a character not found in any name. */
25356 if (strspn(must, "\n -0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ()") == must_len)
25358 while (execute_wildcard(subpattern_re,
25360 SvEND(names_string),
25361 all_names_start, 0,
25364 { /* Here, matched. */
25366 /* Note the string entries look like
25367 * 00001\nSTART OF HEADING\n\n
25368 * so we could match anywhere in that string. We have to rule out
25369 * matching a code point line */
25370 char * this_name_start = all_names_start
25371 + RX_OFFS(subpattern_re)->start;
25372 char * this_name_end = all_names_start
25373 + RX_OFFS(subpattern_re)->end;
25376 UV cp = 0; /* Silences some compilers */
25377 AV * this_string = NULL;
25378 bool is_multi = FALSE;
25380 /* If matched nothing, advance to next possible match */
25381 if (this_name_start == this_name_end) {
25382 cur_pos = (char *) memchr(this_name_end + 1, '\n',
25383 SvEND(names_string) - this_name_end);
25384 if (cur_pos == NULL) {
25389 /* Position the next match to start beyond the current returned
25391 cur_pos = (char *) memchr(this_name_end, '\n',
25392 SvEND(names_string) - this_name_end);
25395 /* Back up to the \n just before the beginning of the character. */
25396 cp_end = (char *) my_memrchr(all_names_start,
25398 this_name_start - all_names_start);
25400 /* If we didn't find a \n, it means it matched somewhere in the
25401 * initial '00000' in the string, so isn't a real match */
25402 if (cp_end == NULL) {
25406 this_name_start = cp_end + 1; /* The name starts just after */
25407 cp_end--; /* the \n, and the code point */
25408 /* ends just before it */
25410 /* All code points are 5 digits long */
25411 cp_start = cp_end - 4;
25413 /* This shouldn't happen, as we found a \n, and the first \n is
25414 * further along than what we subtracted */
25415 assert(cp_start >= all_names_start);
25417 if (cp_start == all_names_start) {
25418 *prop_definition = add_cp_to_invlist(*prop_definition, 0);
25422 /* If the character is a blank, we either have a named sequence, or
25423 * something is wrong */
25424 if (*(cp_start - 1) == ' ') {
25425 cp_start = (char *) my_memrchr(all_names_start,
25427 cp_start - all_names_start);
25431 assert(cp_start != NULL && cp_start >= all_names_start + 2);
25433 /* Except for the first line in the string, the sequence before the
25434 * code point is \n\n. If that isn't the case here, we didn't
25435 * match the name of a character. (We could have matched a named
25436 * sequence, not currently handled */
25437 if (*(cp_start - 1) != '\n' || *(cp_start - 2) != '\n') {
25441 /* We matched! Add this to the list */
25442 found_matches = TRUE;
25444 /* Loop through all the code points in the sequence */
25445 while (cp_start < cp_end) {
25447 /* Calculate this code point from its 5 digits */
25448 cp = (XDIGIT_VALUE(cp_start[0]) << 16)
25449 + (XDIGIT_VALUE(cp_start[1]) << 12)
25450 + (XDIGIT_VALUE(cp_start[2]) << 8)
25451 + (XDIGIT_VALUE(cp_start[3]) << 4)
25452 + XDIGIT_VALUE(cp_start[4]);
25454 cp_start += 6; /* Go past any blank */
25456 if (cp_start < cp_end || is_multi) {
25457 if (this_string == NULL) {
25458 this_string = newAV();
25462 av_push(this_string, newSVuv(cp));
25466 if (is_multi) { /* Was more than one code point */
25467 if (*strings == NULL) {
25468 *strings = newAV();
25471 av_push(*strings, (SV *) this_string);
25473 else { /* Only a single code point */
25474 *prop_definition = add_cp_to_invlist(*prop_definition, cp);
25476 } /* End of loop through the non-algorithmic names string */
25479 /* There are also character names not in 'names_string'. These are
25480 * algorithmically generatable. Try this pattern on each possible one.
25481 * (khw originally planned to leave this out given the large number of
25482 * matches attempted; but the speed turned out to be quite acceptable
25484 * There are plenty of opportunities to optimize to skip many of the tests.
25485 * beyond the rudimentary ones already here */
25487 /* First see if the subpattern matches any of the algorithmic generatable
25488 * Hangul syllable names.
25490 * We know none of these syllable names will match if the input pattern
25491 * requires more bytes than any syllable has, or if the input pattern only
25492 * matches an empty name, or if the pattern has something it must match and
25493 * one of the characters in that isn't in any Hangul syllable. */
25494 if ( prog->minlen <= (SSize_t) syl_max_len
25495 && prog->maxlen > 0
25496 && (strspn(must, "\n ABCDEGHIJKLMNOPRSTUWY") == must_len))
25498 /* These constants, names, values, and algorithm are adapted from the
25499 * Unicode standard, version 5.1, section 3.12, and should never
25501 const char * JamoL[] = {
25502 "G", "GG", "N", "D", "DD", "R", "M", "B", "BB",
25503 "S", "SS", "", "J", "JJ", "C", "K", "T", "P", "H"
25505 const int LCount = C_ARRAY_LENGTH(JamoL);
25507 const char * JamoV[] = {
25508 "A", "AE", "YA", "YAE", "EO", "E", "YEO", "YE", "O", "WA",
25509 "WAE", "OE", "YO", "U", "WEO", "WE", "WI", "YU", "EU", "YI",
25512 const int VCount = C_ARRAY_LENGTH(JamoV);
25514 const char * JamoT[] = {
25515 "", "G", "GG", "GS", "N", "NJ", "NH", "D", "L",
25516 "LG", "LM", "LB", "LS", "LT", "LP", "LH", "M", "B",
25517 "BS", "S", "SS", "NG", "J", "C", "K", "T", "P", "H"
25519 const int TCount = C_ARRAY_LENGTH(JamoT);
25523 /* This is the initial Hangul syllable code point; each time through the
25524 * inner loop, it maps to the next higher code point. For more info,
25525 * see the Hangul syllable section of the Unicode standard. */
25528 syllable_name = sv_2mortal(newSV(syl_max_len));
25529 sv_setpvn(syllable_name, hangul_prefix, hangul_prefix_len);
25531 for (L = 0; L < LCount; L++) {
25532 for (V = 0; V < VCount; V++) {
25533 for (T = 0; T < TCount; T++) {
25535 /* Truncate back to the prefix, which is unvarying */
25536 SvCUR_set(syllable_name, hangul_prefix_len);
25538 sv_catpv(syllable_name, JamoL[L]);
25539 sv_catpv(syllable_name, JamoV[V]);
25540 sv_catpv(syllable_name, JamoT[T]);
25542 if (execute_wildcard(subpattern_re,
25543 SvPVX(syllable_name),
25544 SvEND(syllable_name),
25545 SvPVX(syllable_name), 0,
25549 *prop_definition = add_cp_to_invlist(*prop_definition,
25551 found_matches = TRUE;
25560 /* The rest of the algorithmically generatable names are of the form
25561 * "PREFIX-code_point". The prefixes and the code point limits of each
25562 * were returned to us in the array 'algorithmic_names' from data in
25563 * lib/unicore/Name.pm. 'code_point' in the name is expressed in hex. */
25564 for (i = 0; i <= av_top_index((AV *) algorithmic_names); i++) {
25567 /* Each element of the array is a hash, giving the details for the
25568 * series of names it covers. There is the base name of the characters
25569 * in the series, and the low and high code points in the series. And,
25570 * for optimization purposes a string containing all the legal
25571 * characters that could possibly be in a name in this series. */
25572 HV * this_series = (HV *) SvRV(* av_fetch((AV *) algorithmic_names, i, 0));
25573 SV * prefix = * hv_fetchs(this_series, "name", 0);
25574 IV low = SvIV(* hv_fetchs(this_series, "low", 0));
25575 IV high = SvIV(* hv_fetchs(this_series, "high", 0));
25576 char * legal = SvPVX(* hv_fetchs(this_series, "legal", 0));
25578 /* Pre-allocate an SV with enough space */
25579 SV * algo_name = sv_2mortal(Perl_newSVpvf(aTHX_ "%s-0000",
25581 if (high >= 0x10000) {
25582 sv_catpvs(algo_name, "0");
25585 /* This series can be skipped entirely if the pattern requires
25586 * something longer than any name in the series, or can only match an
25587 * empty name, or contains a character not found in any name in the
25589 if ( prog->minlen <= (SSize_t) SvCUR(algo_name)
25590 && prog->maxlen > 0
25591 && (strspn(must, legal) == must_len))
25593 for (j = low; j <= high; j++) { /* For each code point in the series */
25595 /* Get its name, and see if it matches the subpattern */
25596 Perl_sv_setpvf(aTHX_ algo_name, "%s-%X", SvPVX(prefix),
25599 if (execute_wildcard(subpattern_re,
25602 SvPVX(algo_name), 0,
25606 *prop_definition = add_cp_to_invlist(*prop_definition, j);
25607 found_matches = TRUE;
25614 /* Finally, see if the subpattern matches an empty string */
25615 empty = newSVpvs("");
25616 if (execute_wildcard(subpattern_re,
25623 /* Many code points have empty names. Currently these are the \p{GC=C}
25624 * ones, minus CC and CF */
25626 SV * empty_names_ref = get_prop_definition(UNI_C);
25627 SV * empty_names = invlist_clone(empty_names_ref, NULL);
25629 SV * subtract = get_prop_definition(UNI_CC);
25631 _invlist_subtract(empty_names, subtract, &empty_names);
25632 SvREFCNT_dec_NN(empty_names_ref);
25633 SvREFCNT_dec_NN(subtract);
25635 subtract = get_prop_definition(UNI_CF);
25636 _invlist_subtract(empty_names, subtract, &empty_names);
25637 SvREFCNT_dec_NN(subtract);
25639 _invlist_union(*prop_definition, empty_names, prop_definition);
25640 found_matches = TRUE;
25641 SvREFCNT_dec_NN(empty_names);
25643 SvREFCNT_dec_NN(empty);
25646 /* If we ever were to accept aliases for, say private use names, we would
25647 * need to do something fancier to find empty names. The code below works
25648 * (at the time it was written), and is slower than the above */
25649 const char empties_pat[] = "^.";
25650 if (strNE(name, empties_pat)) {
25651 SV * empty = newSVpvs("");
25652 if (execute_wildcard(subpattern_re,
25659 SV * empties = NULL;
25661 (void) handle_names_wildcard(empties_pat, strlen(empties_pat), &empties);
25663 _invlist_union_complement_2nd(*prop_definition, empties, prop_definition);
25664 SvREFCNT_dec_NN(empties);
25666 found_matches = TRUE;
25668 SvREFCNT_dec_NN(empty);
25672 SvREFCNT_dec_NN(subpattern_re);
25673 return found_matches;
25677 * ex: set ts=8 sts=4 sw=4 et: