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;
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 preceding 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 (OP(scan)==ACCEPT) {
6362 /* m{(*ACCEPT)x} does not have to start with 'x' */
6363 flags &= ~SCF_DO_STCLASS;
6365 data->flags |= SCF_SEEN_ACCEPT;
6371 else if (OP(scan) == COMMIT) {
6372 /* gh18770: m{abc(*COMMIT)xyz} must fail on "abc abcxyz", so we
6373 * must not end up with "abcxyz" as a fixed substring else we'll
6374 * skip straight to attempting to match at offset 4.
6376 if (flags & SCF_DO_SUBSTR) {
6377 scan_commit(pRExC_state, data, minlenp, is_inf);
6378 flags &= ~SCF_DO_SUBSTR;
6381 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
6383 if (flags & SCF_DO_SUBSTR) {
6384 scan_commit(pRExC_state, data, minlenp, is_inf);
6385 data->cur_is_floating = 1; /* float */
6387 is_inf = is_inf_internal = 1;
6388 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
6389 ssc_anything(data->start_class);
6390 flags &= ~SCF_DO_STCLASS;
6392 else if (OP(scan) == GPOS) {
6393 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
6394 !(delta || is_inf || (data && data->pos_delta)))
6396 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
6397 RExC_rx->intflags |= PREGf_ANCH_GPOS;
6398 if (RExC_rx->gofs < (STRLEN)min)
6399 RExC_rx->gofs = min;
6401 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
6405 #ifdef TRIE_STUDY_OPT
6406 #ifdef FULL_TRIE_STUDY
6407 else if (PL_regkind[OP(scan)] == TRIE) {
6408 /* NOTE - There is similar code to this block above for handling
6409 BRANCH nodes on the initial study. If you change stuff here
6411 regnode *trie_node= scan;
6412 regnode *tail= regnext(scan);
6413 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
6414 SSize_t max1 = 0, min1 = OPTIMIZE_INFTY;
6417 if (flags & SCF_DO_SUBSTR) { /* XXXX Add !SUSPEND? */
6418 /* Cannot merge strings after this. */
6419 scan_commit(pRExC_state, data, minlenp, is_inf);
6421 if (flags & SCF_DO_STCLASS)
6422 ssc_init_zero(pRExC_state, &accum);
6428 const regnode *nextbranch= NULL;
6431 for ( word=1 ; word <= trie->wordcount ; word++)
6433 SSize_t deltanext=0, minnext=0, f = 0, fake;
6434 regnode_ssc this_class;
6436 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
6438 data_fake.whilem_c = data->whilem_c;
6439 data_fake.last_closep = data->last_closep;
6442 data_fake.last_closep = &fake;
6443 data_fake.pos_delta = delta;
6444 if (flags & SCF_DO_STCLASS) {
6445 ssc_init(pRExC_state, &this_class);
6446 data_fake.start_class = &this_class;
6447 f = SCF_DO_STCLASS_AND;
6449 if (flags & SCF_WHILEM_VISITED_POS)
6450 f |= SCF_WHILEM_VISITED_POS;
6452 if (trie->jump[word]) {
6454 nextbranch = trie_node + trie->jump[0];
6455 scan= trie_node + trie->jump[word];
6456 /* We go from the jump point to the branch that follows
6457 it. Note this means we need the vestigal unused
6458 branches even though they arent otherwise used. */
6459 /* optimise study_chunk() for TRIE */
6460 minnext = study_chunk(pRExC_state, &scan, minlenp,
6461 &deltanext, (regnode *)nextbranch, &data_fake,
6462 stopparen, recursed_depth, NULL, f, depth+1,
6465 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
6466 nextbranch= regnext((regnode*)nextbranch);
6468 if (min1 > (SSize_t)(minnext + trie->minlen))
6469 min1 = minnext + trie->minlen;
6470 if (deltanext == OPTIMIZE_INFTY) {
6471 is_inf = is_inf_internal = 1;
6472 max1 = OPTIMIZE_INFTY;
6473 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
6474 max1 = minnext + deltanext + trie->maxlen;
6476 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
6478 if (data_fake.flags & SCF_SEEN_ACCEPT) {
6479 if ( stopmin > min + min1)
6480 stopmin = min + min1;
6481 flags &= ~SCF_DO_SUBSTR;
6483 data->flags |= SCF_SEEN_ACCEPT;
6486 if (data_fake.flags & SF_HAS_EVAL)
6487 data->flags |= SF_HAS_EVAL;
6488 data->whilem_c = data_fake.whilem_c;
6490 if (flags & SCF_DO_STCLASS)
6491 ssc_or(pRExC_state, &accum, (regnode_charclass *) &this_class);
6494 if (flags & SCF_DO_SUBSTR) {
6495 data->pos_min += min1;
6496 data->pos_delta += max1 - min1;
6497 if (max1 != min1 || is_inf)
6498 data->cur_is_floating = 1; /* float */
6501 if (delta != OPTIMIZE_INFTY) {
6502 if (OPTIMIZE_INFTY - (max1 - min1) >= delta)
6503 delta += max1 - min1;
6505 delta = OPTIMIZE_INFTY;
6507 if (flags & SCF_DO_STCLASS_OR) {
6508 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &accum);
6510 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6511 flags &= ~SCF_DO_STCLASS;
6514 else if (flags & SCF_DO_STCLASS_AND) {
6516 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
6517 flags &= ~SCF_DO_STCLASS;
6520 /* Switch to OR mode: cache the old value of
6521 * data->start_class */
6523 StructCopy(data->start_class, and_withp, regnode_ssc);
6524 flags &= ~SCF_DO_STCLASS_AND;
6525 StructCopy(&accum, data->start_class, regnode_ssc);
6526 flags |= SCF_DO_STCLASS_OR;
6533 else if (PL_regkind[OP(scan)] == TRIE) {
6534 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
6537 min += trie->minlen;
6538 delta += (trie->maxlen - trie->minlen);
6539 flags &= ~SCF_DO_STCLASS; /* xxx */
6540 if (flags & SCF_DO_SUBSTR) {
6541 /* Cannot expect anything... */
6542 scan_commit(pRExC_state, data, minlenp, is_inf);
6543 data->pos_min += trie->minlen;
6544 data->pos_delta += (trie->maxlen - trie->minlen);
6545 if (trie->maxlen != trie->minlen)
6546 data->cur_is_floating = 1; /* float */
6548 if (trie->jump) /* no more substrings -- for now /grr*/
6549 flags &= ~SCF_DO_SUBSTR;
6552 #endif /* old or new */
6553 #endif /* TRIE_STUDY_OPT */
6555 else if (OP(scan) == REGEX_SET) {
6556 Perl_croak(aTHX_ "panic: %s regnode should be resolved"
6557 " before optimization", PL_reg_name[REGEX_SET]);
6560 /* Else: zero-length, ignore. */
6561 scan = regnext(scan);
6566 /* we need to unwind recursion. */
6569 DEBUG_STUDYDATA("frame-end", data, depth, is_inf);
6570 DEBUG_PEEP("fend", scan, depth, flags);
6572 /* restore previous context */
6573 last = frame->last_regnode;
6574 scan = frame->next_regnode;
6575 stopparen = frame->stopparen;
6576 recursed_depth = frame->prev_recursed_depth;
6578 RExC_frame_last = frame->prev_frame;
6579 frame = frame->this_prev_frame;
6580 goto fake_study_recurse;
6584 DEBUG_STUDYDATA("pre-fin", data, depth, is_inf);
6587 *deltap = is_inf_internal ? OPTIMIZE_INFTY : delta;
6589 if (flags & SCF_DO_SUBSTR && is_inf)
6590 data->pos_delta = OPTIMIZE_INFTY - data->pos_min;
6591 if (is_par > (I32)U8_MAX)
6593 if (is_par && pars==1 && data) {
6594 data->flags |= SF_IN_PAR;
6595 data->flags &= ~SF_HAS_PAR;
6597 else if (pars && data) {
6598 data->flags |= SF_HAS_PAR;
6599 data->flags &= ~SF_IN_PAR;
6601 if (flags & SCF_DO_STCLASS_OR)
6602 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6603 if (flags & SCF_TRIE_RESTUDY)
6604 data->flags |= SCF_TRIE_RESTUDY;
6606 DEBUG_STUDYDATA("post-fin", data, depth, is_inf);
6608 final_minlen = min < stopmin
6611 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN)) {
6612 if (final_minlen > OPTIMIZE_INFTY - delta)
6613 RExC_maxlen = OPTIMIZE_INFTY;
6614 else if (RExC_maxlen < final_minlen + delta)
6615 RExC_maxlen = final_minlen + delta;
6617 return final_minlen;
6621 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
6623 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
6625 PERL_ARGS_ASSERT_ADD_DATA;
6627 Renewc(RExC_rxi->data,
6628 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
6629 char, struct reg_data);
6631 Renew(RExC_rxi->data->what, count + n, U8);
6633 Newx(RExC_rxi->data->what, n, U8);
6634 RExC_rxi->data->count = count + n;
6635 Copy(s, RExC_rxi->data->what + count, n, U8);
6639 /*XXX: todo make this not included in a non debugging perl, but appears to be
6640 * used anyway there, in 'use re' */
6641 #ifndef PERL_IN_XSUB_RE
6643 Perl_reginitcolors(pTHX)
6645 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
6647 char *t = savepv(s);
6651 t = strchr(t, '\t');
6657 PL_colors[i] = t = (char *)"";
6662 PL_colors[i++] = (char *)"";
6669 #ifdef TRIE_STUDY_OPT
6670 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
6673 (data.flags & SCF_TRIE_RESTUDY) \
6681 #define CHECK_RESTUDY_GOTO_butfirst
6685 * pregcomp - compile a regular expression into internal code
6687 * Decides which engine's compiler to call based on the hint currently in
6691 #ifndef PERL_IN_XSUB_RE
6693 /* return the currently in-scope regex engine (or the default if none) */
6695 regexp_engine const *
6696 Perl_current_re_engine(pTHX)
6698 if (IN_PERL_COMPILETIME) {
6699 HV * const table = GvHV(PL_hintgv);
6702 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
6703 return &PL_core_reg_engine;
6704 ptr = hv_fetchs(table, "regcomp", FALSE);
6705 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
6706 return &PL_core_reg_engine;
6707 return INT2PTR(regexp_engine*, SvIV(*ptr));
6711 if (!PL_curcop->cop_hints_hash)
6712 return &PL_core_reg_engine;
6713 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
6714 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
6715 return &PL_core_reg_engine;
6716 return INT2PTR(regexp_engine*, SvIV(ptr));
6722 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
6724 regexp_engine const *eng = current_re_engine();
6725 DECLARE_AND_GET_RE_DEBUG_FLAGS;
6727 PERL_ARGS_ASSERT_PREGCOMP;
6729 /* Dispatch a request to compile a regexp to correct regexp engine. */
6731 Perl_re_printf( aTHX_ "Using engine %" UVxf "\n",
6734 return CALLREGCOMP_ENG(eng, pattern, flags);
6738 /* public(ish) entry point for the perl core's own regex compiling code.
6739 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
6740 * pattern rather than a list of OPs, and uses the internal engine rather
6741 * than the current one */
6744 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
6746 SV *pat = pattern; /* defeat constness! */
6748 PERL_ARGS_ASSERT_RE_COMPILE;
6750 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
6751 #ifdef PERL_IN_XSUB_RE
6754 &PL_core_reg_engine,
6756 NULL, NULL, rx_flags, 0);
6760 S_free_codeblocks(pTHX_ struct reg_code_blocks *cbs)
6764 if (--cbs->refcnt > 0)
6766 for (n = 0; n < cbs->count; n++) {
6767 REGEXP *rx = cbs->cb[n].src_regex;
6769 cbs->cb[n].src_regex = NULL;
6770 SvREFCNT_dec_NN(rx);
6778 static struct reg_code_blocks *
6779 S_alloc_code_blocks(pTHX_ int ncode)
6781 struct reg_code_blocks *cbs;
6782 Newx(cbs, 1, struct reg_code_blocks);
6785 SAVEDESTRUCTOR_X(S_free_codeblocks, cbs);
6787 Newx(cbs->cb, ncode, struct reg_code_block);
6794 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
6795 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
6796 * point to the realloced string and length.
6798 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
6802 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
6803 char **pat_p, STRLEN *plen_p, int num_code_blocks)
6805 U8 *const src = (U8*)*pat_p;
6810 DECLARE_AND_GET_RE_DEBUG_FLAGS;
6812 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6813 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
6815 /* 1 for each byte + 1 for each byte that expands to two, + trailing NUL */
6816 Newx(dst, *plen_p + variant_under_utf8_count(src, src + *plen_p) + 1, U8);
6819 while (s < *plen_p) {
6820 append_utf8_from_native_byte(src[s], &d);
6822 if (n < num_code_blocks) {
6823 assert(pRExC_state->code_blocks);
6824 if (!do_end && pRExC_state->code_blocks->cb[n].start == s) {
6825 pRExC_state->code_blocks->cb[n].start = d - dst - 1;
6826 assert(*(d - 1) == '(');
6829 else if (do_end && pRExC_state->code_blocks->cb[n].end == s) {
6830 pRExC_state->code_blocks->cb[n].end = d - dst - 1;
6831 assert(*(d - 1) == ')');
6840 *pat_p = (char*) dst;
6842 RExC_orig_utf8 = RExC_utf8 = 1;
6847 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
6848 * while recording any code block indices, and handling overloading,
6849 * nested qr// objects etc. If pat is null, it will allocate a new
6850 * string, or just return the first arg, if there's only one.
6852 * Returns the malloced/updated pat.
6853 * patternp and pat_count is the array of SVs to be concatted;
6854 * oplist is the optional list of ops that generated the SVs;
6855 * recompile_p is a pointer to a boolean that will be set if
6856 * the regex will need to be recompiled.
6857 * delim, if non-null is an SV that will be inserted between each element
6861 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
6862 SV *pat, SV ** const patternp, int pat_count,
6863 OP *oplist, bool *recompile_p, SV *delim)
6867 bool use_delim = FALSE;
6868 bool alloced = FALSE;
6870 /* if we know we have at least two args, create an empty string,
6871 * then concatenate args to that. For no args, return an empty string */
6872 if (!pat && pat_count != 1) {
6878 for (svp = patternp; svp < patternp + pat_count; svp++) {
6881 STRLEN orig_patlen = 0;
6883 SV *msv = use_delim ? delim : *svp;
6884 if (!msv) msv = &PL_sv_undef;
6886 /* if we've got a delimiter, we go round the loop twice for each
6887 * svp slot (except the last), using the delimiter the second
6896 if (SvTYPE(msv) == SVt_PVAV) {
6897 /* we've encountered an interpolated array within
6898 * the pattern, e.g. /...@a..../. Expand the list of elements,
6899 * then recursively append elements.
6900 * The code in this block is based on S_pushav() */
6902 AV *const av = (AV*)msv;
6903 const SSize_t maxarg = AvFILL(av) + 1;
6907 assert(oplist->op_type == OP_PADAV
6908 || oplist->op_type == OP_RV2AV);
6909 oplist = OpSIBLING(oplist);
6912 if (SvRMAGICAL(av)) {
6915 Newx(array, maxarg, SV*);
6917 for (i=0; i < maxarg; i++) {
6918 SV ** const svp = av_fetch(av, i, FALSE);
6919 array[i] = svp ? *svp : &PL_sv_undef;
6923 array = AvARRAY(av);
6925 pat = S_concat_pat(aTHX_ pRExC_state, pat,
6926 array, maxarg, NULL, recompile_p,
6928 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
6934 /* we make the assumption here that each op in the list of
6935 * op_siblings maps to one SV pushed onto the stack,
6936 * except for code blocks, with have both an OP_NULL and
6938 * This allows us to match up the list of SVs against the
6939 * list of OPs to find the next code block.
6941 * Note that PUSHMARK PADSV PADSV ..
6943 * PADRANGE PADSV PADSV ..
6944 * so the alignment still works. */
6947 if (oplist->op_type == OP_NULL
6948 && (oplist->op_flags & OPf_SPECIAL))
6950 assert(n < pRExC_state->code_blocks->count);
6951 pRExC_state->code_blocks->cb[n].start = pat ? SvCUR(pat) : 0;
6952 pRExC_state->code_blocks->cb[n].block = oplist;
6953 pRExC_state->code_blocks->cb[n].src_regex = NULL;
6956 oplist = OpSIBLING(oplist); /* skip CONST */
6959 oplist = OpSIBLING(oplist);;
6962 /* apply magic and QR overloading to arg */
6965 if (SvROK(msv) && SvAMAGIC(msv)) {
6966 SV *sv = AMG_CALLunary(msv, regexp_amg);
6970 if (SvTYPE(sv) != SVt_REGEXP)
6971 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
6976 /* try concatenation overload ... */
6977 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
6978 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
6981 /* overloading involved: all bets are off over literal
6982 * code. Pretend we haven't seen it */
6984 pRExC_state->code_blocks->count -= n;
6988 /* ... or failing that, try "" overload */
6989 while (SvAMAGIC(msv)
6990 && (sv = AMG_CALLunary(msv, string_amg))
6994 && SvRV(msv) == SvRV(sv))
6999 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
7003 /* this is a partially unrolled
7004 * sv_catsv_nomg(pat, msv);
7005 * that allows us to adjust code block indices if
7008 char *dst = SvPV_force_nomg(pat, dlen);
7010 if (SvUTF8(msv) && !SvUTF8(pat)) {
7011 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
7012 sv_setpvn(pat, dst, dlen);
7015 sv_catsv_nomg(pat, msv);
7019 /* We have only one SV to process, but we need to verify
7020 * it is properly null terminated or we will fail asserts
7021 * later. In theory we probably shouldn't get such SV's,
7022 * but if we do we should handle it gracefully. */
7023 if ( SvTYPE(msv) != SVt_PV || (SvLEN(msv) > SvCUR(msv) && *(SvEND(msv)) == 0) || SvIsCOW_shared_hash(msv) ) {
7024 /* not a string, or a string with a trailing null */
7027 /* a string with no trailing null, we need to copy it
7028 * so it has a trailing null */
7029 pat = sv_2mortal(newSVsv(msv));
7034 pRExC_state->code_blocks->cb[n-1].end = SvCUR(pat)-1;
7037 /* extract any code blocks within any embedded qr//'s */
7038 if (rx && SvTYPE(rx) == SVt_REGEXP
7039 && RX_ENGINE((REGEXP*)rx)->op_comp)
7042 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
7043 if (ri->code_blocks && ri->code_blocks->count) {
7045 /* the presence of an embedded qr// with code means
7046 * we should always recompile: the text of the
7047 * qr// may not have changed, but it may be a
7048 * different closure than last time */
7050 if (pRExC_state->code_blocks) {
7051 int new_count = pRExC_state->code_blocks->count
7052 + ri->code_blocks->count;
7053 Renew(pRExC_state->code_blocks->cb,
7054 new_count, struct reg_code_block);
7055 pRExC_state->code_blocks->count = new_count;
7058 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_
7059 ri->code_blocks->count);
7061 for (i=0; i < ri->code_blocks->count; i++) {
7062 struct reg_code_block *src, *dst;
7063 STRLEN offset = orig_patlen
7064 + ReANY((REGEXP *)rx)->pre_prefix;
7065 assert(n < pRExC_state->code_blocks->count);
7066 src = &ri->code_blocks->cb[i];
7067 dst = &pRExC_state->code_blocks->cb[n];
7068 dst->start = src->start + offset;
7069 dst->end = src->end + offset;
7070 dst->block = src->block;
7071 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
7080 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
7089 /* see if there are any run-time code blocks in the pattern.
7090 * False positives are allowed */
7093 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
7094 char *pat, STRLEN plen)
7099 PERL_UNUSED_CONTEXT;
7101 for (s = 0; s < plen; s++) {
7102 if ( pRExC_state->code_blocks
7103 && n < pRExC_state->code_blocks->count
7104 && s == pRExC_state->code_blocks->cb[n].start)
7106 s = pRExC_state->code_blocks->cb[n].end;
7110 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
7112 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
7114 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
7121 /* Handle run-time code blocks. We will already have compiled any direct
7122 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
7123 * copy of it, but with any literal code blocks blanked out and
7124 * appropriate chars escaped; then feed it into
7126 * eval "qr'modified_pattern'"
7130 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
7134 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
7136 * After eval_sv()-ing that, grab any new code blocks from the returned qr
7137 * and merge them with any code blocks of the original regexp.
7139 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
7140 * instead, just save the qr and return FALSE; this tells our caller that
7141 * the original pattern needs upgrading to utf8.
7145 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
7146 char *pat, STRLEN plen)
7150 DECLARE_AND_GET_RE_DEBUG_FLAGS;
7152 if (pRExC_state->runtime_code_qr) {
7153 /* this is the second time we've been called; this should
7154 * only happen if the main pattern got upgraded to utf8
7155 * during compilation; re-use the qr we compiled first time
7156 * round (which should be utf8 too)
7158 qr = pRExC_state->runtime_code_qr;
7159 pRExC_state->runtime_code_qr = NULL;
7160 assert(RExC_utf8 && SvUTF8(qr));
7166 int newlen = plen + 7; /* allow for "qr''xx\0" extra chars */
7170 /* determine how many extra chars we need for ' and \ escaping */
7171 for (s = 0; s < plen; s++) {
7172 if (pat[s] == '\'' || pat[s] == '\\')
7176 Newx(newpat, newlen, char);
7178 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
7180 for (s = 0; s < plen; s++) {
7181 if ( pRExC_state->code_blocks
7182 && n < pRExC_state->code_blocks->count
7183 && s == pRExC_state->code_blocks->cb[n].start)
7185 /* blank out literal code block so that they aren't
7186 * recompiled: eg change from/to:
7196 assert(pat[s] == '(');
7197 assert(pat[s+1] == '?');
7201 while (s < pRExC_state->code_blocks->cb[n].end) {
7209 if (pat[s] == '\'' || pat[s] == '\\')
7214 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED) {
7216 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED_MORE) {
7222 Perl_re_printf( aTHX_
7223 "%sre-parsing pattern for runtime code:%s %s\n",
7224 PL_colors[4], PL_colors[5], newpat);
7227 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
7233 PUSHSTACKi(PERLSI_REQUIRE);
7234 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
7235 * parsing qr''; normally only q'' does this. It also alters
7237 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
7238 SvREFCNT_dec_NN(sv);
7243 SV * const errsv = ERRSV;
7244 if (SvTRUE_NN(errsv))
7245 /* use croak_sv ? */
7246 Perl_croak_nocontext("%" SVf, SVfARG(errsv));
7248 assert(SvROK(qr_ref));
7250 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
7251 /* the leaving below frees the tmp qr_ref.
7252 * Give qr a life of its own */
7260 if (!RExC_utf8 && SvUTF8(qr)) {
7261 /* first time through; the pattern got upgraded; save the
7262 * qr for the next time through */
7263 assert(!pRExC_state->runtime_code_qr);
7264 pRExC_state->runtime_code_qr = qr;
7269 /* extract any code blocks within the returned qr// */
7272 /* merge the main (r1) and run-time (r2) code blocks into one */
7274 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
7275 struct reg_code_block *new_block, *dst;
7276 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
7280 if (!r2->code_blocks || !r2->code_blocks->count) /* we guessed wrong */
7282 SvREFCNT_dec_NN(qr);
7286 if (!r1->code_blocks)
7287 r1->code_blocks = S_alloc_code_blocks(aTHX_ 0);
7289 r1c = r1->code_blocks->count;
7290 r2c = r2->code_blocks->count;
7292 Newx(new_block, r1c + r2c, struct reg_code_block);
7296 while (i1 < r1c || i2 < r2c) {
7297 struct reg_code_block *src;
7301 src = &r2->code_blocks->cb[i2++];
7305 src = &r1->code_blocks->cb[i1++];
7306 else if ( r1->code_blocks->cb[i1].start
7307 < r2->code_blocks->cb[i2].start)
7309 src = &r1->code_blocks->cb[i1++];
7310 assert(src->end < r2->code_blocks->cb[i2].start);
7313 assert( r1->code_blocks->cb[i1].start
7314 > r2->code_blocks->cb[i2].start);
7315 src = &r2->code_blocks->cb[i2++];
7317 assert(src->end < r1->code_blocks->cb[i1].start);
7320 assert(pat[src->start] == '(');
7321 assert(pat[src->end] == ')');
7322 dst->start = src->start;
7323 dst->end = src->end;
7324 dst->block = src->block;
7325 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
7329 r1->code_blocks->count += r2c;
7330 Safefree(r1->code_blocks->cb);
7331 r1->code_blocks->cb = new_block;
7334 SvREFCNT_dec_NN(qr);
7340 S_setup_longest(pTHX_ RExC_state_t *pRExC_state,
7341 struct reg_substr_datum *rsd,
7342 struct scan_data_substrs *sub,
7343 STRLEN longest_length)
7345 /* This is the common code for setting up the floating and fixed length
7346 * string data extracted from Perl_re_op_compile() below. Returns a boolean
7347 * as to whether succeeded or not */
7351 bool eol = cBOOL(sub->flags & SF_BEFORE_EOL);
7352 bool meol = cBOOL(sub->flags & SF_BEFORE_MEOL);
7354 if (! (longest_length
7355 || (eol /* Can't have SEOL and MULTI */
7356 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
7358 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
7359 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
7364 /* copy the information about the longest from the reg_scan_data
7365 over to the program. */
7366 if (SvUTF8(sub->str)) {
7368 rsd->utf8_substr = sub->str;
7370 rsd->substr = sub->str;
7371 rsd->utf8_substr = NULL;
7373 /* end_shift is how many chars that must be matched that
7374 follow this item. We calculate it ahead of time as once the
7375 lookbehind offset is added in we lose the ability to correctly
7377 ml = sub->minlenp ? *(sub->minlenp) : (SSize_t)longest_length;
7378 rsd->end_shift = ml - sub->min_offset
7380 /* XXX SvTAIL is always false here - did you mean FBMcf_TAIL
7382 + (SvTAIL(sub->str) != 0)
7386 t = (eol/* Can't have SEOL and MULTI */
7387 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
7388 fbm_compile(sub->str, t ? FBMcf_TAIL : 0);
7394 S_set_regex_pv(pTHX_ RExC_state_t *pRExC_state, REGEXP *Rx)
7396 /* Calculates and sets in the compiled pattern 'Rx' the string to compile,
7397 * properly wrapped with the right modifiers */
7399 bool has_p = ((RExC_rx->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
7400 bool has_charset = RExC_utf8 || (get_regex_charset(RExC_rx->extflags)
7401 != REGEX_DEPENDS_CHARSET);
7403 /* The caret is output if there are any defaults: if not all the STD
7404 * flags are set, or if no character set specifier is needed */
7406 (((RExC_rx->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
7408 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
7409 == REG_RUN_ON_COMMENT_SEEN);
7410 U8 reganch = (U8)((RExC_rx->extflags & RXf_PMf_STD_PMMOD)
7411 >> RXf_PMf_STD_PMMOD_SHIFT);
7412 const char *fptr = STD_PAT_MODS; /*"msixxn"*/
7414 STRLEN pat_len = RExC_precomp_end - RExC_precomp;
7416 /* We output all the necessary flags; we never output a minus, as all
7417 * those are defaults, so are
7418 * covered by the caret */
7419 const STRLEN wraplen = pat_len + has_p + has_runon
7420 + has_default /* If needs a caret */
7421 + PL_bitcount[reganch] /* 1 char for each set standard flag */
7423 /* If needs a character set specifier */
7424 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
7425 + (sizeof("(?:)") - 1);
7427 PERL_ARGS_ASSERT_SET_REGEX_PV;
7429 /* make sure PL_bitcount bounds not exceeded */
7430 STATIC_ASSERT_STMT(sizeof(STD_PAT_MODS) <= 8);
7432 p = sv_grow(MUTABLE_SV(Rx), wraplen + 1); /* +1 for the ending NUL */
7435 SvFLAGS(Rx) |= SVf_UTF8;
7438 /* If a default, cover it using the caret */
7440 *p++= DEFAULT_PAT_MOD;
7446 name = get_regex_charset_name(RExC_rx->extflags, &len);
7447 if (strEQ(name, DEPENDS_PAT_MODS)) { /* /d under UTF-8 => /u */
7449 name = UNICODE_PAT_MODS;
7450 len = sizeof(UNICODE_PAT_MODS) - 1;
7452 Copy(name, p, len, char);
7456 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
7459 while((ch = *fptr++)) {
7467 Copy(RExC_precomp, p, pat_len, char);
7468 assert ((RX_WRAPPED(Rx) - p) < 16);
7469 RExC_rx->pre_prefix = p - RX_WRAPPED(Rx);
7472 /* Adding a trailing \n causes this to compile properly:
7473 my $R = qr / A B C # D E/x; /($R)/
7474 Otherwise the parens are considered part of the comment */
7479 SvCUR_set(Rx, p - RX_WRAPPED(Rx));
7483 * Perl_re_op_compile - the perl internal RE engine's function to compile a
7484 * regular expression into internal code.
7485 * The pattern may be passed either as:
7486 * a list of SVs (patternp plus pat_count)
7487 * a list of OPs (expr)
7488 * If both are passed, the SV list is used, but the OP list indicates
7489 * which SVs are actually pre-compiled code blocks
7491 * The SVs in the list have magic and qr overloading applied to them (and
7492 * the list may be modified in-place with replacement SVs in the latter
7495 * If the pattern hasn't changed from old_re, then old_re will be
7498 * eng is the current engine. If that engine has an op_comp method, then
7499 * handle directly (i.e. we assume that op_comp was us); otherwise, just
7500 * do the initial concatenation of arguments and pass on to the external
7503 * If is_bare_re is not null, set it to a boolean indicating whether the
7504 * arg list reduced (after overloading) to a single bare regex which has
7505 * been returned (i.e. /$qr/).
7507 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
7509 * pm_flags contains the PMf_* flags, typically based on those from the
7510 * pm_flags field of the related PMOP. Currently we're only interested in
7511 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL, PMf_WILDCARD.
7513 * For many years this code had an initial sizing pass that calculated
7514 * (sometimes incorrectly, leading to security holes) the size needed for the
7515 * compiled pattern. That was changed by commit
7516 * 7c932d07cab18751bfc7515b4320436273a459e2 in 5.29, which reallocs the size, a
7517 * node at a time, as parsing goes along. Patches welcome to fix any obsolete
7518 * references to this sizing pass.
7520 * Now, an initial crude guess as to the size needed is made, based on the
7521 * length of the pattern. Patches welcome to improve that guess. That amount
7522 * of space is malloc'd and then immediately freed, and then clawed back node
7523 * by node. This design is to minimze, to the extent possible, memory churn
7524 * when doing the reallocs.
7526 * A separate parentheses counting pass may be needed in some cases.
7527 * (Previously the sizing pass did this.) Patches welcome to reduce the number
7530 * The existence of a sizing pass necessitated design decisions that are no
7531 * longer needed. There are potential areas of simplification.
7533 * Beware that the optimization-preparation code in here knows about some
7534 * of the structure of the compiled regexp. [I'll say.]
7538 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
7539 OP *expr, const regexp_engine* eng, REGEXP *old_re,
7540 bool *is_bare_re, const U32 orig_rx_flags, const U32 pm_flags)
7542 REGEXP *Rx; /* Capital 'R' means points to a REGEXP */
7550 SV** new_patternp = patternp;
7552 /* these are all flags - maybe they should be turned
7553 * into a single int with different bit masks */
7554 I32 sawlookahead = 0;
7559 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
7561 bool runtime_code = 0;
7563 RExC_state_t RExC_state;
7564 RExC_state_t * const pRExC_state = &RExC_state;
7565 #ifdef TRIE_STUDY_OPT
7567 RExC_state_t copyRExC_state;
7569 DECLARE_AND_GET_RE_DEBUG_FLAGS;
7571 PERL_ARGS_ASSERT_RE_OP_COMPILE;
7573 DEBUG_r(if (!PL_colorset) reginitcolors());
7576 pRExC_state->warn_text = NULL;
7577 pRExC_state->unlexed_names = NULL;
7578 pRExC_state->code_blocks = NULL;
7581 *is_bare_re = FALSE;
7583 if (expr && (expr->op_type == OP_LIST ||
7584 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
7585 /* allocate code_blocks if needed */
7589 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o))
7590 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
7591 ncode++; /* count of DO blocks */
7594 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_ ncode);
7598 /* compile-time pattern with just OP_CONSTs and DO blocks */
7603 /* find how many CONSTs there are */
7606 if (expr->op_type == OP_CONST)
7609 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
7610 if (o->op_type == OP_CONST)
7614 /* fake up an SV array */
7616 assert(!new_patternp);
7617 Newx(new_patternp, n, SV*);
7618 SAVEFREEPV(new_patternp);
7622 if (expr->op_type == OP_CONST)
7623 new_patternp[n] = cSVOPx_sv(expr);
7625 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
7626 if (o->op_type == OP_CONST)
7627 new_patternp[n++] = cSVOPo_sv;
7632 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7633 "Assembling pattern from %d elements%s\n", pat_count,
7634 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
7636 /* set expr to the first arg op */
7638 if (pRExC_state->code_blocks && pRExC_state->code_blocks->count
7639 && expr->op_type != OP_CONST)
7641 expr = cLISTOPx(expr)->op_first;
7642 assert( expr->op_type == OP_PUSHMARK
7643 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
7644 || expr->op_type == OP_PADRANGE);
7645 expr = OpSIBLING(expr);
7648 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
7649 expr, &recompile, NULL);
7651 /* handle bare (possibly after overloading) regex: foo =~ $re */
7656 if (SvTYPE(re) == SVt_REGEXP) {
7660 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7661 "Precompiled pattern%s\n",
7662 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
7668 exp = SvPV_nomg(pat, plen);
7670 if (!eng->op_comp) {
7671 if ((SvUTF8(pat) && IN_BYTES)
7672 || SvGMAGICAL(pat) || SvAMAGIC(pat))
7674 /* make a temporary copy; either to convert to bytes,
7675 * or to avoid repeating get-magic / overloaded stringify */
7676 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
7677 (IN_BYTES ? 0 : SvUTF8(pat)));
7679 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
7682 /* ignore the utf8ness if the pattern is 0 length */
7683 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
7684 RExC_uni_semantics = 0;
7685 RExC_contains_locale = 0;
7686 RExC_strict = cBOOL(pm_flags & RXf_PMf_STRICT);
7687 RExC_in_script_run = 0;
7688 RExC_study_started = 0;
7689 pRExC_state->runtime_code_qr = NULL;
7690 RExC_frame_head= NULL;
7691 RExC_frame_last= NULL;
7692 RExC_frame_count= 0;
7693 RExC_latest_warn_offset = 0;
7694 RExC_use_BRANCHJ = 0;
7695 RExC_warned_WARN_EXPERIMENTAL__VLB = 0;
7696 RExC_warned_WARN_EXPERIMENTAL__REGEX_SETS = 0;
7697 RExC_total_parens = 0;
7698 RExC_open_parens = NULL;
7699 RExC_close_parens = NULL;
7700 RExC_paren_names = NULL;
7702 RExC_seen_d_op = FALSE;
7704 RExC_paren_name_list = NULL;
7708 RExC_mysv1= sv_newmortal();
7709 RExC_mysv2= sv_newmortal();
7713 SV *dsv= sv_newmortal();
7714 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, PL_dump_re_max_len);
7715 Perl_re_printf( aTHX_ "%sCompiling REx%s %s\n",
7716 PL_colors[4], PL_colors[5], s);
7719 /* we jump here if we have to recompile, e.g., from upgrading the pattern
7722 if ((pm_flags & PMf_USE_RE_EVAL)
7723 /* this second condition covers the non-regex literal case,
7724 * i.e. $foo =~ '(?{})'. */
7725 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
7727 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
7730 /* return old regex if pattern hasn't changed */
7731 /* XXX: note in the below we have to check the flags as well as the
7734 * Things get a touch tricky as we have to compare the utf8 flag
7735 * independently from the compile flags. */
7739 && !!RX_UTF8(old_re) == !!RExC_utf8
7740 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
7741 && RX_PRECOMP(old_re)
7742 && RX_PRELEN(old_re) == plen
7743 && memEQ(RX_PRECOMP(old_re), exp, plen)
7744 && !runtime_code /* with runtime code, always recompile */ )
7747 SV *dsv= sv_newmortal();
7748 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, PL_dump_re_max_len);
7749 Perl_re_printf( aTHX_ "%sSkipping recompilation of unchanged REx%s %s\n",
7750 PL_colors[4], PL_colors[5], s);
7755 /* Allocate the pattern's SV */
7756 RExC_rx_sv = Rx = (REGEXP*) newSV_type(SVt_REGEXP);
7757 RExC_rx = ReANY(Rx);
7758 if ( RExC_rx == NULL )
7759 FAIL("Regexp out of space");
7761 rx_flags = orig_rx_flags;
7763 if ( toUSE_UNI_CHARSET_NOT_DEPENDS
7764 && initial_charset == REGEX_DEPENDS_CHARSET)
7767 /* Set to use unicode semantics if the pattern is in utf8 and has the
7768 * 'depends' charset specified, as it means unicode when utf8 */
7769 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
7770 RExC_uni_semantics = 1;
7773 RExC_pm_flags = pm_flags;
7776 assert(TAINTING_get || !TAINT_get);
7778 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
7780 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
7781 /* whoops, we have a non-utf8 pattern, whilst run-time code
7782 * got compiled as utf8. Try again with a utf8 pattern */
7783 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7784 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7788 assert(!pRExC_state->runtime_code_qr);
7794 RExC_in_lookaround = 0;
7795 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
7796 RExC_recode_x_to_native = 0;
7797 RExC_in_multi_char_class = 0;
7799 RExC_start = RExC_copy_start_in_constructed = RExC_copy_start_in_input = RExC_precomp = exp;
7800 RExC_precomp_end = RExC_end = exp + plen;
7802 RExC_whilem_seen = 0;
7804 RExC_recurse = NULL;
7805 RExC_study_chunk_recursed = NULL;
7806 RExC_study_chunk_recursed_bytes= 0;
7807 RExC_recurse_count = 0;
7808 RExC_sets_depth = 0;
7809 pRExC_state->code_index = 0;
7811 /* Initialize the string in the compiled pattern. This is so that there is
7812 * something to output if necessary */
7813 set_regex_pv(pRExC_state, Rx);
7816 Perl_re_printf( aTHX_
7817 "Starting parse and generation\n");
7819 RExC_lastparse=NULL;
7822 /* Allocate space and zero-initialize. Note, the two step process
7823 of zeroing when in debug mode, thus anything assigned has to
7824 happen after that */
7827 /* On the first pass of the parse, we guess how big this will be. Then
7828 * we grow in one operation to that amount and then give it back. As
7829 * we go along, we re-allocate what we need.
7831 * XXX Currently the guess is essentially that the pattern will be an
7832 * EXACT node with one byte input, one byte output. This is crude, and
7833 * better heuristics are welcome.
7835 * On any subsequent passes, we guess what we actually computed in the
7836 * latest earlier pass. Such a pass probably didn't complete so is
7837 * missing stuff. We could improve those guesses by knowing where the
7838 * parse stopped, and use the length so far plus apply the above
7839 * assumption to what's left. */
7840 RExC_size = STR_SZ(RExC_end - RExC_start);
7843 Newxc(RExC_rxi, sizeof(regexp_internal) + RExC_size, char, regexp_internal);
7844 if ( RExC_rxi == NULL )
7845 FAIL("Regexp out of space");
7847 Zero(RExC_rxi, sizeof(regexp_internal) + RExC_size, char);
7848 RXi_SET( RExC_rx, RExC_rxi );
7850 /* We start from 0 (over from 0 in the case this is a reparse. The first
7851 * node parsed will give back any excess memory we have allocated so far).
7855 /* non-zero initialization begins here */
7856 RExC_rx->engine= eng;
7857 RExC_rx->extflags = rx_flags;
7858 RXp_COMPFLAGS(RExC_rx) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
7860 if (pm_flags & PMf_IS_QR) {
7861 RExC_rxi->code_blocks = pRExC_state->code_blocks;
7862 if (RExC_rxi->code_blocks) {
7863 RExC_rxi->code_blocks->refcnt++;
7867 RExC_rx->intflags = 0;
7869 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
7872 /* This NUL is guaranteed because the pattern comes from an SV*, and the sv
7873 * code makes sure the final byte is an uncounted NUL. But should this
7874 * ever not be the case, lots of things could read beyond the end of the
7875 * buffer: loops like
7876 * while(isFOO(*RExC_parse)) RExC_parse++;
7877 * strchr(RExC_parse, "foo");
7878 * etc. So it is worth noting. */
7879 assert(*RExC_end == '\0');
7883 RExC_parens_buf_size = 0;
7884 RExC_emit_start = RExC_rxi->program;
7885 pRExC_state->code_index = 0;
7887 *((char*) RExC_emit_start) = (char) REG_MAGIC;
7891 if (reg(pRExC_state, 0, &flags, 1)) {
7893 /* Success!, But we may need to redo the parse knowing how many parens
7894 * there actually are */
7895 if (IN_PARENS_PASS) {
7896 flags |= RESTART_PARSE;
7899 /* We have that number in RExC_npar */
7900 RExC_total_parens = RExC_npar;
7902 else if (! MUST_RESTART(flags)) {
7904 Perl_croak(aTHX_ "panic: reg returned failure to re_op_compile, flags=%#" UVxf, (UV) flags);
7907 /* Here, we either have success, or we have to redo the parse for some reason */
7908 if (MUST_RESTART(flags)) {
7910 /* It's possible to write a regexp in ascii that represents Unicode
7911 codepoints outside of the byte range, such as via \x{100}. If we
7912 detect such a sequence we have to convert the entire pattern to utf8
7913 and then recompile, as our sizing calculation will have been based
7914 on 1 byte == 1 character, but we will need to use utf8 to encode
7915 at least some part of the pattern, and therefore must convert the whole
7918 if (flags & NEED_UTF8) {
7920 /* We have stored the offset of the final warning output so far.
7921 * That must be adjusted. Any variant characters between the start
7922 * of the pattern and this warning count for 2 bytes in the final,
7923 * so just add them again */
7924 if (UNLIKELY(RExC_latest_warn_offset > 0)) {
7925 RExC_latest_warn_offset +=
7926 variant_under_utf8_count((U8 *) exp, (U8 *) exp
7927 + RExC_latest_warn_offset);
7929 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7930 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7931 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "Need to redo parse after upgrade\n"));
7934 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "Need to redo parse\n"));
7937 if (ALL_PARENS_COUNTED) {
7938 /* Make enough room for all the known parens, and zero it */
7939 Renew(RExC_open_parens, RExC_total_parens, regnode_offset);
7940 Zero(RExC_open_parens, RExC_total_parens, regnode_offset);
7941 RExC_open_parens[0] = 1; /* +1 for REG_MAGIC */
7943 Renew(RExC_close_parens, RExC_total_parens, regnode_offset);
7944 Zero(RExC_close_parens, RExC_total_parens, regnode_offset);
7946 else { /* Parse did not complete. Reinitialize the parentheses
7948 RExC_total_parens = 0;
7949 if (RExC_open_parens) {
7950 Safefree(RExC_open_parens);
7951 RExC_open_parens = NULL;
7953 if (RExC_close_parens) {
7954 Safefree(RExC_close_parens);
7955 RExC_close_parens = NULL;
7959 /* Clean up what we did in this parse */
7960 SvREFCNT_dec_NN(RExC_rx_sv);
7965 /* Here, we have successfully parsed and generated the pattern's program
7966 * for the regex engine. We are ready to finish things up and look for
7969 /* Update the string to compile, with correct modifiers, etc */
7970 set_regex_pv(pRExC_state, Rx);
7972 RExC_rx->nparens = RExC_total_parens - 1;
7974 /* Uses the upper 4 bits of the FLAGS field, so keep within that size */
7975 if (RExC_whilem_seen > 15)
7976 RExC_whilem_seen = 15;
7979 Perl_re_printf( aTHX_
7980 "Required size %" IVdf " nodes\n", (IV)RExC_size);
7982 RExC_lastparse=NULL;
7985 #ifdef RE_TRACK_PATTERN_OFFSETS
7986 DEBUG_OFFSETS_r(Perl_re_printf( aTHX_
7987 "%s %" UVuf " bytes for offset annotations.\n",
7988 RExC_offsets ? "Got" : "Couldn't get",
7989 (UV)((RExC_offsets[0] * 2 + 1))));
7990 DEBUG_OFFSETS_r(if (RExC_offsets) {
7991 const STRLEN len = RExC_offsets[0];
7993 DECLARE_AND_GET_RE_DEBUG_FLAGS;
7994 Perl_re_printf( aTHX_
7995 "Offsets: [%" UVuf "]\n\t", (UV)RExC_offsets[0]);
7996 for (i = 1; i <= len; i++) {
7997 if (RExC_offsets[i*2-1] || RExC_offsets[i*2])
7998 Perl_re_printf( aTHX_ "%" UVuf ":%" UVuf "[%" UVuf "] ",
7999 (UV)i, (UV)RExC_offsets[i*2-1], (UV)RExC_offsets[i*2]);
8001 Perl_re_printf( aTHX_ "\n");
8005 SetProgLen(RExC_rxi,RExC_size);
8008 DEBUG_DUMP_PRE_OPTIMIZE_r({
8009 SV * const sv = sv_newmortal();
8010 RXi_GET_DECL(RExC_rx, ri);
8012 Perl_re_printf( aTHX_ "Program before optimization:\n");
8014 (void)dumpuntil(RExC_rx, ri->program, ri->program + 1, NULL, NULL,
8019 Perl_re_printf( aTHX_ "Starting post parse optimization\n");
8022 /* XXXX To minimize changes to RE engine we always allocate
8023 3-units-long substrs field. */
8024 Newx(RExC_rx->substrs, 1, struct reg_substr_data);
8025 if (RExC_recurse_count) {
8026 Newx(RExC_recurse, RExC_recurse_count, regnode *);
8027 SAVEFREEPV(RExC_recurse);
8030 if (RExC_seen & REG_RECURSE_SEEN) {
8031 /* Note, RExC_total_parens is 1 + the number of parens in a pattern.
8032 * So its 1 if there are no parens. */
8033 RExC_study_chunk_recursed_bytes= (RExC_total_parens >> 3) +
8034 ((RExC_total_parens & 0x07) != 0);
8035 Newx(RExC_study_chunk_recursed,
8036 RExC_study_chunk_recursed_bytes * RExC_total_parens, U8);
8037 SAVEFREEPV(RExC_study_chunk_recursed);
8041 RExC_rx->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
8043 RExC_study_chunk_recursed_count= 0;
8045 Zero(RExC_rx->substrs, 1, struct reg_substr_data);
8046 if (RExC_study_chunk_recursed) {
8047 Zero(RExC_study_chunk_recursed,
8048 RExC_study_chunk_recursed_bytes * RExC_total_parens, U8);
8052 #ifdef TRIE_STUDY_OPT
8054 StructCopy(&zero_scan_data, &data, scan_data_t);
8055 copyRExC_state = RExC_state;
8058 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "Restudying\n"));
8060 RExC_state = copyRExC_state;
8061 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
8062 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
8064 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
8065 StructCopy(&zero_scan_data, &data, scan_data_t);
8068 StructCopy(&zero_scan_data, &data, scan_data_t);
8071 /* Dig out information for optimizations. */
8072 RExC_rx->extflags = RExC_flags; /* was pm_op */
8073 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
8076 SvUTF8_on(Rx); /* Unicode in it? */
8077 RExC_rxi->regstclass = NULL;
8078 if (RExC_naughty >= TOO_NAUGHTY) /* Probably an expensive pattern. */
8079 RExC_rx->intflags |= PREGf_NAUGHTY;
8080 scan = RExC_rxi->program + 1; /* First BRANCH. */
8082 /* testing for BRANCH here tells us whether there is "must appear"
8083 data in the pattern. If there is then we can use it for optimisations */
8084 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
8087 STRLEN longest_length[2];
8088 regnode_ssc ch_class; /* pointed to by data */
8090 SSize_t last_close = 0; /* pointed to by data */
8091 regnode *first= scan;
8092 regnode *first_next= regnext(first);
8096 * Skip introductions and multiplicators >= 1
8097 * so that we can extract the 'meat' of the pattern that must
8098 * match in the large if() sequence following.
8099 * NOTE that EXACT is NOT covered here, as it is normally
8100 * picked up by the optimiser separately.
8102 * This is unfortunate as the optimiser isnt handling lookahead
8103 * properly currently.
8106 while ((OP(first) == OPEN && (sawopen = 1)) ||
8107 /* An OR of *one* alternative - should not happen now. */
8108 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
8109 /* for now we can't handle lookbehind IFMATCH*/
8110 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
8111 (OP(first) == PLUS) ||
8112 (OP(first) == MINMOD) ||
8113 /* An {n,m} with n>0 */
8114 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
8115 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
8118 * the only op that could be a regnode is PLUS, all the rest
8119 * will be regnode_1 or regnode_2.
8121 * (yves doesn't think this is true)
8123 if (OP(first) == PLUS)
8126 if (OP(first) == MINMOD)
8128 first += regarglen[OP(first)];
8130 first = NEXTOPER(first);
8131 first_next= regnext(first);
8134 /* Starting-point info. */
8136 DEBUG_PEEP("first:", first, 0, 0);
8137 /* Ignore EXACT as we deal with it later. */
8138 if (PL_regkind[OP(first)] == EXACT) {
8139 if (! isEXACTFish(OP(first))) {
8140 NOOP; /* Empty, get anchored substr later. */
8143 RExC_rxi->regstclass = first;
8146 else if (PL_regkind[OP(first)] == TRIE &&
8147 ((reg_trie_data *)RExC_rxi->data->data[ ARG(first) ])->minlen>0)
8149 /* this can happen only on restudy */
8150 RExC_rxi->regstclass = construct_ahocorasick_from_trie(pRExC_state, (regnode *)first, 0);
8153 else if (REGNODE_SIMPLE(OP(first)))
8154 RExC_rxi->regstclass = first;
8155 else if (PL_regkind[OP(first)] == BOUND ||
8156 PL_regkind[OP(first)] == NBOUND)
8157 RExC_rxi->regstclass = first;
8158 else if (PL_regkind[OP(first)] == BOL) {
8159 RExC_rx->intflags |= (OP(first) == MBOL
8162 first = NEXTOPER(first);
8165 else if (OP(first) == GPOS) {
8166 RExC_rx->intflags |= PREGf_ANCH_GPOS;
8167 first = NEXTOPER(first);
8170 else if ((!sawopen || !RExC_sawback) &&
8172 (OP(first) == STAR &&
8173 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
8174 !(RExC_rx->intflags & PREGf_ANCH) && !pRExC_state->code_blocks)
8176 /* turn .* into ^.* with an implied $*=1 */
8178 (OP(NEXTOPER(first)) == REG_ANY)
8181 RExC_rx->intflags |= (type | PREGf_IMPLICIT);
8182 first = NEXTOPER(first);
8185 if (sawplus && !sawminmod && !sawlookahead
8186 && (!sawopen || !RExC_sawback)
8187 && !pRExC_state->code_blocks) /* May examine pos and $& */
8188 /* x+ must match at the 1st pos of run of x's */
8189 RExC_rx->intflags |= PREGf_SKIP;
8191 /* Scan is after the zeroth branch, first is atomic matcher. */
8192 #ifdef TRIE_STUDY_OPT
8195 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
8196 (IV)(first - scan + 1))
8200 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
8201 (IV)(first - scan + 1))
8207 * If there's something expensive in the r.e., find the
8208 * longest literal string that must appear and make it the
8209 * regmust. Resolve ties in favor of later strings, since
8210 * the regstart check works with the beginning of the r.e.
8211 * and avoiding duplication strengthens checking. Not a
8212 * strong reason, but sufficient in the absence of others.
8213 * [Now we resolve ties in favor of the earlier string if
8214 * it happens that c_offset_min has been invalidated, since the
8215 * earlier string may buy us something the later one won't.]
8218 data.substrs[0].str = newSVpvs("");
8219 data.substrs[1].str = newSVpvs("");
8220 data.last_found = newSVpvs("");
8221 data.cur_is_floating = 0; /* initially any found substring is fixed */
8222 ENTER_with_name("study_chunk");
8223 SAVEFREESV(data.substrs[0].str);
8224 SAVEFREESV(data.substrs[1].str);
8225 SAVEFREESV(data.last_found);
8227 if (!RExC_rxi->regstclass) {
8228 ssc_init(pRExC_state, &ch_class);
8229 data.start_class = &ch_class;
8230 stclass_flag = SCF_DO_STCLASS_AND;
8231 } else /* XXXX Check for BOUND? */
8233 data.last_closep = &last_close;
8237 * MAIN ENTRY FOR study_chunk() FOR m/PATTERN/
8238 * (NO top level branches)
8240 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
8241 scan + RExC_size, /* Up to end */
8243 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
8244 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
8248 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
8251 if ( RExC_total_parens == 1 && !data.cur_is_floating
8252 && data.last_start_min == 0 && data.last_end > 0
8253 && !RExC_seen_zerolen
8254 && !(RExC_seen & REG_VERBARG_SEEN)
8255 && !(RExC_seen & REG_GPOS_SEEN)
8257 RExC_rx->extflags |= RXf_CHECK_ALL;
8259 scan_commit(pRExC_state, &data,&minlen, 0);
8262 /* XXX this is done in reverse order because that's the way the
8263 * code was before it was parameterised. Don't know whether it
8264 * actually needs doing in reverse order. DAPM */
8265 for (i = 1; i >= 0; i--) {
8266 longest_length[i] = CHR_SVLEN(data.substrs[i].str);
8269 && SvCUR(data.substrs[0].str) /* ok to leave SvCUR */
8270 && data.substrs[0].min_offset
8271 == data.substrs[1].min_offset
8272 && SvCUR(data.substrs[0].str)
8273 == SvCUR(data.substrs[1].str)
8275 && S_setup_longest (aTHX_ pRExC_state,
8276 &(RExC_rx->substrs->data[i]),
8280 RExC_rx->substrs->data[i].min_offset =
8281 data.substrs[i].min_offset - data.substrs[i].lookbehind;
8283 RExC_rx->substrs->data[i].max_offset = data.substrs[i].max_offset;
8284 /* Don't offset infinity */
8285 if (data.substrs[i].max_offset < OPTIMIZE_INFTY)
8286 RExC_rx->substrs->data[i].max_offset -= data.substrs[i].lookbehind;
8287 SvREFCNT_inc_simple_void_NN(data.substrs[i].str);
8290 RExC_rx->substrs->data[i].substr = NULL;
8291 RExC_rx->substrs->data[i].utf8_substr = NULL;
8292 longest_length[i] = 0;
8296 LEAVE_with_name("study_chunk");
8298 if (RExC_rxi->regstclass
8299 && (OP(RExC_rxi->regstclass) == REG_ANY || OP(RExC_rxi->regstclass) == SANY))
8300 RExC_rxi->regstclass = NULL;
8302 if ((!(RExC_rx->substrs->data[0].substr || RExC_rx->substrs->data[0].utf8_substr)
8303 || RExC_rx->substrs->data[0].min_offset)
8305 && ! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
8306 && is_ssc_worth_it(pRExC_state, data.start_class))
8308 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
8310 ssc_finalize(pRExC_state, data.start_class);
8312 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
8313 StructCopy(data.start_class,
8314 (regnode_ssc*)RExC_rxi->data->data[n],
8316 RExC_rxi->regstclass = (regnode*)RExC_rxi->data->data[n];
8317 RExC_rx->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
8318 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
8319 regprop(RExC_rx, sv, (regnode*)data.start_class, NULL, pRExC_state);
8320 Perl_re_printf( aTHX_
8321 "synthetic stclass \"%s\".\n",
8322 SvPVX_const(sv));});
8323 data.start_class = NULL;
8326 /* A temporary algorithm prefers floated substr to fixed one of
8327 * same length to dig more info. */
8328 i = (longest_length[0] <= longest_length[1]);
8329 RExC_rx->substrs->check_ix = i;
8330 RExC_rx->check_end_shift = RExC_rx->substrs->data[i].end_shift;
8331 RExC_rx->check_substr = RExC_rx->substrs->data[i].substr;
8332 RExC_rx->check_utf8 = RExC_rx->substrs->data[i].utf8_substr;
8333 RExC_rx->check_offset_min = RExC_rx->substrs->data[i].min_offset;
8334 RExC_rx->check_offset_max = RExC_rx->substrs->data[i].max_offset;
8335 if (!i && (RExC_rx->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS)))
8336 RExC_rx->intflags |= PREGf_NOSCAN;
8338 if ((RExC_rx->check_substr || RExC_rx->check_utf8) ) {
8339 RExC_rx->extflags |= RXf_USE_INTUIT;
8340 if (SvTAIL(RExC_rx->check_substr ? RExC_rx->check_substr : RExC_rx->check_utf8))
8341 RExC_rx->extflags |= RXf_INTUIT_TAIL;
8344 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
8345 if ( (STRLEN)minlen < longest_length[1] )
8346 minlen= longest_length[1];
8347 if ( (STRLEN)minlen < longest_length[0] )
8348 minlen= longest_length[0];
8352 /* Several toplevels. Best we can is to set minlen. */
8354 regnode_ssc ch_class;
8355 SSize_t last_close = 0;
8357 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "\nMulti Top Level\n"));
8359 scan = RExC_rxi->program + 1;
8360 ssc_init(pRExC_state, &ch_class);
8361 data.start_class = &ch_class;
8362 data.last_closep = &last_close;
8366 * MAIN ENTRY FOR study_chunk() FOR m/P1|P2|.../
8367 * (patterns WITH top level branches)
8369 minlen = study_chunk(pRExC_state,
8370 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
8371 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
8372 ? SCF_TRIE_DOING_RESTUDY
8376 CHECK_RESTUDY_GOTO_butfirst(NOOP);
8378 RExC_rx->check_substr = NULL;
8379 RExC_rx->check_utf8 = NULL;
8380 RExC_rx->substrs->data[0].substr = NULL;
8381 RExC_rx->substrs->data[0].utf8_substr = NULL;
8382 RExC_rx->substrs->data[1].substr = NULL;
8383 RExC_rx->substrs->data[1].utf8_substr = NULL;
8385 if (! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
8386 && is_ssc_worth_it(pRExC_state, data.start_class))
8388 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
8390 ssc_finalize(pRExC_state, data.start_class);
8392 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
8393 StructCopy(data.start_class,
8394 (regnode_ssc*)RExC_rxi->data->data[n],
8396 RExC_rxi->regstclass = (regnode*)RExC_rxi->data->data[n];
8397 RExC_rx->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
8398 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
8399 regprop(RExC_rx, sv, (regnode*)data.start_class, NULL, pRExC_state);
8400 Perl_re_printf( aTHX_
8401 "synthetic stclass \"%s\".\n",
8402 SvPVX_const(sv));});
8403 data.start_class = NULL;
8407 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
8408 RExC_rx->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
8409 RExC_rx->maxlen = REG_INFTY;
8412 RExC_rx->maxlen = RExC_maxlen;
8415 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
8416 the "real" pattern. */
8418 Perl_re_printf( aTHX_ "minlen: %" IVdf " RExC_rx->minlen:%" IVdf " maxlen:%" IVdf "\n",
8419 (IV)minlen, (IV)RExC_rx->minlen, (IV)RExC_maxlen);
8421 RExC_rx->minlenret = minlen;
8422 if (RExC_rx->minlen < minlen)
8423 RExC_rx->minlen = minlen;
8425 if (RExC_seen & REG_RECURSE_SEEN ) {
8426 RExC_rx->intflags |= PREGf_RECURSE_SEEN;
8427 Newx(RExC_rx->recurse_locinput, RExC_rx->nparens + 1, char *);
8429 if (RExC_seen & REG_GPOS_SEEN)
8430 RExC_rx->intflags |= PREGf_GPOS_SEEN;
8431 if (RExC_seen & REG_LOOKBEHIND_SEEN)
8432 RExC_rx->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
8434 if (pRExC_state->code_blocks)
8435 RExC_rx->extflags |= RXf_EVAL_SEEN;
8436 if (RExC_seen & REG_VERBARG_SEEN)
8438 RExC_rx->intflags |= PREGf_VERBARG_SEEN;
8439 RExC_rx->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
8441 if (RExC_seen & REG_CUTGROUP_SEEN)
8442 RExC_rx->intflags |= PREGf_CUTGROUP_SEEN;
8443 if (pm_flags & PMf_USE_RE_EVAL)
8444 RExC_rx->intflags |= PREGf_USE_RE_EVAL;
8445 if (RExC_paren_names)
8446 RXp_PAREN_NAMES(RExC_rx) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
8448 RXp_PAREN_NAMES(RExC_rx) = NULL;
8450 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
8451 * so it can be used in pp.c */
8452 if (RExC_rx->intflags & PREGf_ANCH)
8453 RExC_rx->extflags |= RXf_IS_ANCHORED;
8457 /* this is used to identify "special" patterns that might result
8458 * in Perl NOT calling the regex engine and instead doing the match "itself",
8459 * particularly special cases in split//. By having the regex compiler
8460 * do this pattern matching at a regop level (instead of by inspecting the pattern)
8461 * we avoid weird issues with equivalent patterns resulting in different behavior,
8462 * AND we allow non Perl engines to get the same optimizations by the setting the
8463 * flags appropriately - Yves */
8464 regnode *first = RExC_rxi->program + 1;
8466 regnode *next = NEXTOPER(first);
8467 /* It's safe to read through *next only if OP(first) is a regop of
8468 * the right type (not EXACT, for example).
8470 U8 nop = (fop == NOTHING || fop == MBOL || fop == SBOL || fop == PLUS)
8473 if (PL_regkind[fop] == NOTHING && nop == END)
8474 RExC_rx->extflags |= RXf_NULL;
8475 else if ((fop == MBOL || (fop == SBOL && !first->flags)) && nop == END)
8476 /* when fop is SBOL first->flags will be true only when it was
8477 * produced by parsing /\A/, and not when parsing /^/. This is
8478 * very important for the split code as there we want to
8479 * treat /^/ as /^/m, but we do not want to treat /\A/ as /^/m.
8480 * See rt #122761 for more details. -- Yves */
8481 RExC_rx->extflags |= RXf_START_ONLY;
8482 else if (fop == PLUS
8483 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
8484 && OP(regnext(first)) == END)
8485 RExC_rx->extflags |= RXf_WHITE;
8486 else if ( RExC_rx->extflags & RXf_SPLIT
8487 && (PL_regkind[fop] == EXACT && ! isEXACTFish(fop))
8488 && STR_LEN(first) == 1
8489 && *(STRING(first)) == ' '
8490 && OP(regnext(first)) == END )
8491 RExC_rx->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
8495 if (RExC_contains_locale) {
8496 RXp_EXTFLAGS(RExC_rx) |= RXf_TAINTED;
8500 if (RExC_paren_names) {
8501 RExC_rxi->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
8502 RExC_rxi->data->data[RExC_rxi->name_list_idx]
8503 = (void*)SvREFCNT_inc(RExC_paren_name_list);
8506 RExC_rxi->name_list_idx = 0;
8508 while ( RExC_recurse_count > 0 ) {
8509 const regnode *scan = RExC_recurse[ --RExC_recurse_count ];
8511 * This data structure is set up in study_chunk() and is used
8512 * to calculate the distance between a GOSUB regopcode and
8513 * the OPEN/CURLYM (CURLYM's are special and can act like OPEN's)
8516 * If for some reason someone writes code that optimises
8517 * away a GOSUB opcode then the assert should be changed to
8518 * an if(scan) to guard the ARG2L_SET() - Yves
8521 assert(scan && OP(scan) == GOSUB);
8522 ARG2L_SET( scan, RExC_open_parens[ARG(scan)] - REGNODE_OFFSET(scan));
8525 Newxz(RExC_rx->offs, RExC_total_parens, regexp_paren_pair);
8526 /* assume we don't need to swap parens around before we match */
8528 Perl_re_printf( aTHX_ "study_chunk_recursed_count: %lu\n",
8529 (unsigned long)RExC_study_chunk_recursed_count);
8533 Perl_re_printf( aTHX_ "Final program:\n");
8537 if (RExC_open_parens) {
8538 Safefree(RExC_open_parens);
8539 RExC_open_parens = NULL;
8541 if (RExC_close_parens) {
8542 Safefree(RExC_close_parens);
8543 RExC_close_parens = NULL;
8547 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
8548 * by setting the regexp SV to readonly-only instead. If the
8549 * pattern's been recompiled, the USEDness should remain. */
8550 if (old_re && SvREADONLY(old_re))
8558 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
8561 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
8563 PERL_UNUSED_ARG(value);
8565 if (flags & RXapif_FETCH) {
8566 return reg_named_buff_fetch(rx, key, flags);
8567 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
8568 Perl_croak_no_modify();
8570 } else if (flags & RXapif_EXISTS) {
8571 return reg_named_buff_exists(rx, key, flags)
8574 } else if (flags & RXapif_REGNAMES) {
8575 return reg_named_buff_all(rx, flags);
8576 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
8577 return reg_named_buff_scalar(rx, flags);
8579 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
8585 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
8588 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
8589 PERL_UNUSED_ARG(lastkey);
8591 if (flags & RXapif_FIRSTKEY)
8592 return reg_named_buff_firstkey(rx, flags);
8593 else if (flags & RXapif_NEXTKEY)
8594 return reg_named_buff_nextkey(rx, flags);
8596 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
8603 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
8607 struct regexp *const rx = ReANY(r);
8609 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
8611 if (rx && RXp_PAREN_NAMES(rx)) {
8612 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
8615 SV* sv_dat=HeVAL(he_str);
8616 I32 *nums=(I32*)SvPVX(sv_dat);
8617 AV * const retarray = (flags & RXapif_ALL) ? newAV() : NULL;
8618 for ( i=0; i<SvIVX(sv_dat); i++ ) {
8619 if ((I32)(rx->nparens) >= nums[i]
8620 && rx->offs[nums[i]].start != -1
8621 && rx->offs[nums[i]].end != -1)
8624 CALLREG_NUMBUF_FETCH(r, nums[i], ret);
8629 ret = newSVsv(&PL_sv_undef);
8632 av_push(retarray, ret);
8635 return newRV_noinc(MUTABLE_SV(retarray));
8642 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
8645 struct regexp *const rx = ReANY(r);
8647 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
8649 if (rx && RXp_PAREN_NAMES(rx)) {
8650 if (flags & RXapif_ALL) {
8651 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
8653 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
8655 SvREFCNT_dec_NN(sv);
8667 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
8669 struct regexp *const rx = ReANY(r);
8671 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
8673 if ( rx && RXp_PAREN_NAMES(rx) ) {
8674 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
8676 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
8683 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
8685 struct regexp *const rx = ReANY(r);
8686 DECLARE_AND_GET_RE_DEBUG_FLAGS;
8688 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
8690 if (rx && RXp_PAREN_NAMES(rx)) {
8691 HV *hv = RXp_PAREN_NAMES(rx);
8693 while ( (temphe = hv_iternext_flags(hv, 0)) ) {
8696 SV* sv_dat = HeVAL(temphe);
8697 I32 *nums = (I32*)SvPVX(sv_dat);
8698 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8699 if ((I32)(rx->lastparen) >= nums[i] &&
8700 rx->offs[nums[i]].start != -1 &&
8701 rx->offs[nums[i]].end != -1)
8707 if (parno || flags & RXapif_ALL) {
8708 return newSVhek(HeKEY_hek(temphe));
8716 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
8721 struct regexp *const rx = ReANY(r);
8723 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
8725 if (rx && RXp_PAREN_NAMES(rx)) {
8726 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
8727 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
8728 } else if (flags & RXapif_ONE) {
8729 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
8730 av = MUTABLE_AV(SvRV(ret));
8731 length = av_count(av);
8732 SvREFCNT_dec_NN(ret);
8733 return newSViv(length);
8735 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
8740 return &PL_sv_undef;
8744 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
8746 struct regexp *const rx = ReANY(r);
8749 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
8751 if (rx && RXp_PAREN_NAMES(rx)) {
8752 HV *hv= RXp_PAREN_NAMES(rx);
8754 (void)hv_iterinit(hv);
8755 while ( (temphe = hv_iternext_flags(hv, 0)) ) {
8758 SV* sv_dat = HeVAL(temphe);
8759 I32 *nums = (I32*)SvPVX(sv_dat);
8760 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8761 if ((I32)(rx->lastparen) >= nums[i] &&
8762 rx->offs[nums[i]].start != -1 &&
8763 rx->offs[nums[i]].end != -1)
8769 if (parno || flags & RXapif_ALL) {
8770 av_push(av, newSVhek(HeKEY_hek(temphe)));
8775 return newRV_noinc(MUTABLE_SV(av));
8779 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
8782 struct regexp *const rx = ReANY(r);
8788 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
8790 if ( n == RX_BUFF_IDX_CARET_PREMATCH
8791 || n == RX_BUFF_IDX_CARET_FULLMATCH
8792 || n == RX_BUFF_IDX_CARET_POSTMATCH
8795 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8797 /* on something like
8800 * the KEEPCOPY is set on the PMOP rather than the regex */
8801 if (PL_curpm && r == PM_GETRE(PL_curpm))
8802 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8811 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
8812 /* no need to distinguish between them any more */
8813 n = RX_BUFF_IDX_FULLMATCH;
8815 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
8816 && rx->offs[0].start != -1)
8818 /* $`, ${^PREMATCH} */
8819 i = rx->offs[0].start;
8823 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
8824 && rx->offs[0].end != -1)
8826 /* $', ${^POSTMATCH} */
8827 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
8828 i = rx->sublen + rx->suboffset - rx->offs[0].end;
8831 if (inRANGE(n, 0, (I32)rx->nparens) &&
8832 (s1 = rx->offs[n].start) != -1 &&
8833 (t1 = rx->offs[n].end) != -1)
8835 /* $&, ${^MATCH}, $1 ... */
8837 s = rx->subbeg + s1 - rx->suboffset;
8842 assert(s >= rx->subbeg);
8843 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
8845 #ifdef NO_TAINT_SUPPORT
8846 sv_setpvn(sv, s, i);
8848 const int oldtainted = TAINT_get;
8850 sv_setpvn(sv, s, i);
8851 TAINT_set(oldtainted);
8853 if (RXp_MATCH_UTF8(rx))
8858 if (RXp_MATCH_TAINTED(rx)) {
8859 if (SvTYPE(sv) >= SVt_PVMG) {
8860 MAGIC* const mg = SvMAGIC(sv);
8863 SvMAGIC_set(sv, mg->mg_moremagic);
8865 if ((mgt = SvMAGIC(sv))) {
8866 mg->mg_moremagic = mgt;
8867 SvMAGIC_set(sv, mg);
8884 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
8885 SV const * const value)
8887 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
8889 PERL_UNUSED_ARG(rx);
8890 PERL_UNUSED_ARG(paren);
8891 PERL_UNUSED_ARG(value);
8894 Perl_croak_no_modify();
8898 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
8901 struct regexp *const rx = ReANY(r);
8905 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
8907 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
8908 || paren == RX_BUFF_IDX_CARET_FULLMATCH
8909 || paren == RX_BUFF_IDX_CARET_POSTMATCH
8912 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8914 /* on something like
8917 * the KEEPCOPY is set on the PMOP rather than the regex */
8918 if (PL_curpm && r == PM_GETRE(PL_curpm))
8919 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8925 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
8927 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
8928 case RX_BUFF_IDX_PREMATCH: /* $` */
8929 if (rx->offs[0].start != -1) {
8930 i = rx->offs[0].start;
8939 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
8940 case RX_BUFF_IDX_POSTMATCH: /* $' */
8941 if (rx->offs[0].end != -1) {
8942 i = rx->sublen - rx->offs[0].end;
8944 s1 = rx->offs[0].end;
8951 default: /* $& / ${^MATCH}, $1, $2, ... */
8952 if (paren <= (I32)rx->nparens &&
8953 (s1 = rx->offs[paren].start) != -1 &&
8954 (t1 = rx->offs[paren].end) != -1)
8960 if (ckWARN(WARN_UNINITIALIZED))
8961 report_uninit((const SV *)sv);
8966 if (i > 0 && RXp_MATCH_UTF8(rx)) {
8967 const char * const s = rx->subbeg - rx->suboffset + s1;
8972 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
8979 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
8981 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
8982 PERL_UNUSED_ARG(rx);
8986 return newSVpvs("Regexp");
8989 /* Scans the name of a named buffer from the pattern.
8990 * If flags is REG_RSN_RETURN_NULL returns null.
8991 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
8992 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
8993 * to the parsed name as looked up in the RExC_paren_names hash.
8994 * If there is an error throws a vFAIL().. type exception.
8997 #define REG_RSN_RETURN_NULL 0
8998 #define REG_RSN_RETURN_NAME 1
8999 #define REG_RSN_RETURN_DATA 2
9002 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
9004 char *name_start = RExC_parse;
9007 PERL_ARGS_ASSERT_REG_SCAN_NAME;
9009 assert (RExC_parse <= RExC_end);
9010 if (RExC_parse == RExC_end) NOOP;
9011 else if (isIDFIRST_lazy_if_safe(RExC_parse, RExC_end, UTF)) {
9012 /* Note that the code here assumes well-formed UTF-8. Skip IDFIRST by
9013 * using do...while */
9016 RExC_parse += UTF8SKIP(RExC_parse);
9017 } while ( RExC_parse < RExC_end
9018 && isWORDCHAR_utf8_safe((U8*)RExC_parse, (U8*) RExC_end));
9022 } while (RExC_parse < RExC_end && isWORDCHAR(*RExC_parse));
9024 RExC_parse++; /* so the <- from the vFAIL is after the offending
9026 vFAIL("Group name must start with a non-digit word character");
9028 sv_name = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
9029 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
9030 if ( flags == REG_RSN_RETURN_NAME)
9032 else if (flags==REG_RSN_RETURN_DATA) {
9035 if ( ! sv_name ) /* should not happen*/
9036 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
9037 if (RExC_paren_names)
9038 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
9040 sv_dat = HeVAL(he_str);
9041 if ( ! sv_dat ) { /* Didn't find group */
9043 /* It might be a forward reference; we can't fail until we
9044 * know, by completing the parse to get all the groups, and
9046 if (ALL_PARENS_COUNTED) {
9047 vFAIL("Reference to nonexistent named group");
9050 REQUIRE_PARENS_PASS;
9056 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
9057 (unsigned long) flags);
9060 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
9061 if (RExC_lastparse!=RExC_parse) { \
9062 Perl_re_printf( aTHX_ "%s", \
9063 Perl_pv_pretty(aTHX_ RExC_mysv1, RExC_parse, \
9064 RExC_end - RExC_parse, 16, \
9066 PERL_PV_ESCAPE_UNI_DETECT | \
9067 PERL_PV_PRETTY_ELLIPSES | \
9068 PERL_PV_PRETTY_LTGT | \
9069 PERL_PV_ESCAPE_RE | \
9070 PERL_PV_PRETTY_EXACTSIZE \
9074 Perl_re_printf( aTHX_ "%16s",""); \
9076 if (RExC_lastnum!=RExC_emit) \
9077 Perl_re_printf( aTHX_ "|%4zu", RExC_emit); \
9079 Perl_re_printf( aTHX_ "|%4s",""); \
9080 Perl_re_printf( aTHX_ "|%*s%-4s", \
9081 (int)((depth*2)), "", \
9084 RExC_lastnum=RExC_emit; \
9085 RExC_lastparse=RExC_parse; \
9090 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
9091 DEBUG_PARSE_MSG((funcname)); \
9092 Perl_re_printf( aTHX_ "%4s","\n"); \
9094 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({\
9095 DEBUG_PARSE_MSG((funcname)); \
9096 Perl_re_printf( aTHX_ fmt "\n",args); \
9099 /* This section of code defines the inversion list object and its methods. The
9100 * interfaces are highly subject to change, so as much as possible is static to
9101 * this file. An inversion list is here implemented as a malloc'd C UV array
9102 * as an SVt_INVLIST scalar.
9104 * An inversion list for Unicode is an array of code points, sorted by ordinal
9105 * number. Each element gives the code point that begins a range that extends
9106 * up-to but not including the code point given by the next element. The final
9107 * element gives the first code point of a range that extends to the platform's
9108 * infinity. The even-numbered elements (invlist[0], invlist[2], invlist[4],
9109 * ...) give ranges whose code points are all in the inversion list. We say
9110 * that those ranges are in the set. The odd-numbered elements give ranges
9111 * whose code points are not in the inversion list, and hence not in the set.
9112 * Thus, element [0] is the first code point in the list. Element [1]
9113 * is the first code point beyond that not in the list; and element [2] is the
9114 * first code point beyond that that is in the list. In other words, the first
9115 * range is invlist[0]..(invlist[1]-1), and all code points in that range are
9116 * in the inversion list. The second range is invlist[1]..(invlist[2]-1), and
9117 * all code points in that range are not in the inversion list. The third
9118 * range invlist[2]..(invlist[3]-1) gives code points that are in the inversion
9119 * list, and so forth. Thus every element whose index is divisible by two
9120 * gives the beginning of a range that is in the list, and every element whose
9121 * index is not divisible by two gives the beginning of a range not in the
9122 * list. If the final element's index is divisible by two, the inversion list
9123 * extends to the platform's infinity; otherwise the highest code point in the
9124 * inversion list is the contents of that element minus 1.
9126 * A range that contains just a single code point N will look like
9128 * invlist[i+1] == N+1
9130 * If N is UV_MAX (the highest representable code point on the machine), N+1 is
9131 * impossible to represent, so element [i+1] is omitted. The single element
9133 * invlist[0] == UV_MAX
9134 * contains just UV_MAX, but is interpreted as matching to infinity.
9136 * Taking the complement (inverting) an inversion list is quite simple, if the
9137 * first element is 0, remove it; otherwise add a 0 element at the beginning.
9138 * This implementation reserves an element at the beginning of each inversion
9139 * list to always contain 0; there is an additional flag in the header which
9140 * indicates if the list begins at the 0, or is offset to begin at the next
9141 * element. This means that the inversion list can be inverted without any
9142 * copying; just flip the flag.
9144 * More about inversion lists can be found in "Unicode Demystified"
9145 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
9147 * The inversion list data structure is currently implemented as an SV pointing
9148 * to an array of UVs that the SV thinks are bytes. This allows us to have an
9149 * array of UV whose memory management is automatically handled by the existing
9150 * facilities for SV's.
9152 * Some of the methods should always be private to the implementation, and some
9153 * should eventually be made public */
9155 /* The header definitions are in F<invlist_inline.h> */
9157 #ifndef PERL_IN_XSUB_RE
9159 PERL_STATIC_INLINE UV*
9160 S__invlist_array_init(SV* const invlist, const bool will_have_0)
9162 /* Returns a pointer to the first element in the inversion list's array.
9163 * This is called upon initialization of an inversion list. Where the
9164 * array begins depends on whether the list has the code point U+0000 in it
9165 * or not. The other parameter tells it whether the code that follows this
9166 * call is about to put a 0 in the inversion list or not. The first
9167 * element is either the element reserved for 0, if TRUE, or the element
9168 * after it, if FALSE */
9170 bool* offset = get_invlist_offset_addr(invlist);
9171 UV* zero_addr = (UV *) SvPVX(invlist);
9173 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
9176 assert(! _invlist_len(invlist));
9180 /* 1^1 = 0; 1^0 = 1 */
9181 *offset = 1 ^ will_have_0;
9182 return zero_addr + *offset;
9186 S_invlist_replace_list_destroys_src(pTHX_ SV * dest, SV * src)
9188 /* Replaces the inversion list in 'dest' with the one from 'src'. It
9189 * steals the list from 'src', so 'src' is made to have a NULL list. This
9190 * is similar to what SvSetMagicSV() would do, if it were implemented on
9191 * inversion lists, though this routine avoids a copy */
9193 const UV src_len = _invlist_len(src);
9194 const bool src_offset = *get_invlist_offset_addr(src);
9195 const STRLEN src_byte_len = SvLEN(src);
9196 char * array = SvPVX(src);
9198 const int oldtainted = TAINT_get;
9200 PERL_ARGS_ASSERT_INVLIST_REPLACE_LIST_DESTROYS_SRC;
9202 assert(is_invlist(src));
9203 assert(is_invlist(dest));
9204 assert(! invlist_is_iterating(src));
9205 assert(SvCUR(src) == 0 || SvCUR(src) < SvLEN(src));
9207 /* Make sure it ends in the right place with a NUL, as our inversion list
9208 * manipulations aren't careful to keep this true, but sv_usepvn_flags()
9210 array[src_byte_len - 1] = '\0';
9212 TAINT_NOT; /* Otherwise it breaks */
9213 sv_usepvn_flags(dest,
9217 /* This flag is documented to cause a copy to be avoided */
9218 SV_HAS_TRAILING_NUL);
9219 TAINT_set(oldtainted);
9224 /* Finish up copying over the other fields in an inversion list */
9225 *get_invlist_offset_addr(dest) = src_offset;
9226 invlist_set_len(dest, src_len, src_offset);
9227 *get_invlist_previous_index_addr(dest) = 0;
9228 invlist_iterfinish(dest);
9231 PERL_STATIC_INLINE IV*
9232 S_get_invlist_previous_index_addr(SV* invlist)
9234 /* Return the address of the IV that is reserved to hold the cached index
9236 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
9238 assert(is_invlist(invlist));
9240 return &(((XINVLIST*) SvANY(invlist))->prev_index);
9243 PERL_STATIC_INLINE IV
9244 S_invlist_previous_index(SV* const invlist)
9246 /* Returns cached index of previous search */
9248 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
9250 return *get_invlist_previous_index_addr(invlist);
9253 PERL_STATIC_INLINE void
9254 S_invlist_set_previous_index(SV* const invlist, const IV index)
9256 /* Caches <index> for later retrieval */
9258 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
9260 assert(index == 0 || index < (int) _invlist_len(invlist));
9262 *get_invlist_previous_index_addr(invlist) = index;
9265 PERL_STATIC_INLINE void
9266 S_invlist_trim(SV* invlist)
9268 /* Free the not currently-being-used space in an inversion list */
9270 /* But don't free up the space needed for the 0 UV that is always at the
9271 * beginning of the list, nor the trailing NUL */
9272 const UV min_size = TO_INTERNAL_SIZE(1) + 1;
9274 PERL_ARGS_ASSERT_INVLIST_TRIM;
9276 assert(is_invlist(invlist));
9278 SvPV_renew(invlist, MAX(min_size, SvCUR(invlist) + 1));
9281 PERL_STATIC_INLINE void
9282 S_invlist_clear(pTHX_ SV* invlist) /* Empty the inversion list */
9284 PERL_ARGS_ASSERT_INVLIST_CLEAR;
9286 assert(is_invlist(invlist));
9288 invlist_set_len(invlist, 0, 0);
9289 invlist_trim(invlist);
9292 #endif /* ifndef PERL_IN_XSUB_RE */
9294 PERL_STATIC_INLINE bool
9295 S_invlist_is_iterating(SV* const invlist)
9297 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
9299 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
9302 #ifndef PERL_IN_XSUB_RE
9304 PERL_STATIC_INLINE UV
9305 S_invlist_max(SV* const invlist)
9307 /* Returns the maximum number of elements storable in the inversion list's
9308 * array, without having to realloc() */
9310 PERL_ARGS_ASSERT_INVLIST_MAX;
9312 assert(is_invlist(invlist));
9314 /* Assumes worst case, in which the 0 element is not counted in the
9315 * inversion list, so subtracts 1 for that */
9316 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
9317 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
9318 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
9322 S_initialize_invlist_guts(pTHX_ SV* invlist, const Size_t initial_size)
9324 PERL_ARGS_ASSERT_INITIALIZE_INVLIST_GUTS;
9326 /* First 1 is in case the zero element isn't in the list; second 1 is for
9328 SvGROW(invlist, TO_INTERNAL_SIZE(initial_size + 1) + 1);
9329 invlist_set_len(invlist, 0, 0);
9331 /* Force iterinit() to be used to get iteration to work */
9332 invlist_iterfinish(invlist);
9334 *get_invlist_previous_index_addr(invlist) = 0;
9335 SvPOK_on(invlist); /* This allows B to extract the PV */
9339 Perl__new_invlist(pTHX_ IV initial_size)
9342 /* Return a pointer to a newly constructed inversion list, with enough
9343 * space to store 'initial_size' elements. If that number is negative, a
9344 * system default is used instead */
9348 if (initial_size < 0) {
9352 new_list = newSV_type(SVt_INVLIST);
9353 initialize_invlist_guts(new_list, initial_size);
9359 Perl__new_invlist_C_array(pTHX_ const UV* const list)
9361 /* Return a pointer to a newly constructed inversion list, initialized to
9362 * point to <list>, which has to be in the exact correct inversion list
9363 * form, including internal fields. Thus this is a dangerous routine that
9364 * should not be used in the wrong hands. The passed in 'list' contains
9365 * several header fields at the beginning that are not part of the
9366 * inversion list body proper */
9368 const STRLEN length = (STRLEN) list[0];
9369 const UV version_id = list[1];
9370 const bool offset = cBOOL(list[2]);
9371 #define HEADER_LENGTH 3
9372 /* If any of the above changes in any way, you must change HEADER_LENGTH
9373 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
9374 * perl -E 'say int(rand 2**31-1)'
9376 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
9377 data structure type, so that one being
9378 passed in can be validated to be an
9379 inversion list of the correct vintage.
9382 SV* invlist = newSV_type(SVt_INVLIST);
9384 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
9386 if (version_id != INVLIST_VERSION_ID) {
9387 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
9390 /* The generated array passed in includes header elements that aren't part
9391 * of the list proper, so start it just after them */
9392 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
9394 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
9395 shouldn't touch it */
9397 *(get_invlist_offset_addr(invlist)) = offset;
9399 /* The 'length' passed to us is the physical number of elements in the
9400 * inversion list. But if there is an offset the logical number is one
9402 invlist_set_len(invlist, length - offset, offset);
9404 invlist_set_previous_index(invlist, 0);
9406 /* Initialize the iteration pointer. */
9407 invlist_iterfinish(invlist);
9409 SvREADONLY_on(invlist);
9416 S__append_range_to_invlist(pTHX_ SV* const invlist,
9417 const UV start, const UV end)
9419 /* Subject to change or removal. Append the range from 'start' to 'end' at
9420 * the end of the inversion list. The range must be above any existing
9424 UV max = invlist_max(invlist);
9425 UV len = _invlist_len(invlist);
9428 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
9430 if (len == 0) { /* Empty lists must be initialized */
9431 offset = start != 0;
9432 array = _invlist_array_init(invlist, ! offset);
9435 /* Here, the existing list is non-empty. The current max entry in the
9436 * list is generally the first value not in the set, except when the
9437 * set extends to the end of permissible values, in which case it is
9438 * the first entry in that final set, and so this call is an attempt to
9439 * append out-of-order */
9441 UV final_element = len - 1;
9442 array = invlist_array(invlist);
9443 if ( array[final_element] > start
9444 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
9446 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",
9447 array[final_element], start,
9448 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
9451 /* Here, it is a legal append. If the new range begins 1 above the end
9452 * of the range below it, it is extending the range below it, so the
9453 * new first value not in the set is one greater than the newly
9454 * extended range. */
9455 offset = *get_invlist_offset_addr(invlist);
9456 if (array[final_element] == start) {
9457 if (end != UV_MAX) {
9458 array[final_element] = end + 1;
9461 /* But if the end is the maximum representable on the machine,
9462 * assume that infinity was actually what was meant. Just let
9463 * the range that this would extend to have no end */
9464 invlist_set_len(invlist, len - 1, offset);
9470 /* Here the new range doesn't extend any existing set. Add it */
9472 len += 2; /* Includes an element each for the start and end of range */
9474 /* If wll overflow the existing space, extend, which may cause the array to
9477 invlist_extend(invlist, len);
9479 /* Have to set len here to avoid assert failure in invlist_array() */
9480 invlist_set_len(invlist, len, offset);
9482 array = invlist_array(invlist);
9485 invlist_set_len(invlist, len, offset);
9488 /* The next item on the list starts the range, the one after that is
9489 * one past the new range. */
9490 array[len - 2] = start;
9491 if (end != UV_MAX) {
9492 array[len - 1] = end + 1;
9495 /* But if the end is the maximum representable on the machine, just let
9496 * the range have no end */
9497 invlist_set_len(invlist, len - 1, offset);
9502 Perl__invlist_search(SV* const invlist, const UV cp)
9504 /* Searches the inversion list for the entry that contains the input code
9505 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
9506 * return value is the index into the list's array of the range that
9507 * contains <cp>, that is, 'i' such that
9508 * array[i] <= cp < array[i+1]
9513 IV high = _invlist_len(invlist);
9514 const IV highest_element = high - 1;
9517 PERL_ARGS_ASSERT__INVLIST_SEARCH;
9519 /* If list is empty, return failure. */
9520 if (UNLIKELY(high == 0)) {
9524 /* (We can't get the array unless we know the list is non-empty) */
9525 array = invlist_array(invlist);
9527 mid = invlist_previous_index(invlist);
9529 if (UNLIKELY(mid > highest_element)) {
9530 mid = highest_element;
9533 /* <mid> contains the cache of the result of the previous call to this
9534 * function (0 the first time). See if this call is for the same result,
9535 * or if it is for mid-1. This is under the theory that calls to this
9536 * function will often be for related code points that are near each other.
9537 * And benchmarks show that caching gives better results. We also test
9538 * here if the code point is within the bounds of the list. These tests
9539 * replace others that would have had to be made anyway to make sure that
9540 * the array bounds were not exceeded, and these give us extra information
9541 * at the same time */
9542 if (cp >= array[mid]) {
9543 if (cp >= array[highest_element]) {
9544 return highest_element;
9547 /* Here, array[mid] <= cp < array[highest_element]. This means that
9548 * the final element is not the answer, so can exclude it; it also
9549 * means that <mid> is not the final element, so can refer to 'mid + 1'
9551 if (cp < array[mid + 1]) {
9557 else { /* cp < aray[mid] */
9558 if (cp < array[0]) { /* Fail if outside the array */
9562 if (cp >= array[mid - 1]) {
9567 /* Binary search. What we are looking for is <i> such that
9568 * array[i] <= cp < array[i+1]
9569 * The loop below converges on the i+1. Note that there may not be an
9570 * (i+1)th element in the array, and things work nonetheless */
9571 while (low < high) {
9572 mid = (low + high) / 2;
9573 assert(mid <= highest_element);
9574 if (array[mid] <= cp) { /* cp >= array[mid] */
9577 /* We could do this extra test to exit the loop early.
9578 if (cp < array[low]) {
9583 else { /* cp < array[mid] */
9590 invlist_set_previous_index(invlist, high);
9595 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9596 const bool complement_b, SV** output)
9598 /* Take the union of two inversion lists and point '*output' to it. On
9599 * input, '*output' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9600 * even 'a' or 'b'). If to an inversion list, the contents of the original
9601 * list will be replaced by the union. The first list, 'a', may be
9602 * NULL, in which case a copy of the second list is placed in '*output'.
9603 * If 'complement_b' is TRUE, the union is taken of the complement
9604 * (inversion) of 'b' instead of b itself.
9606 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9607 * Richard Gillam, published by Addison-Wesley, and explained at some
9608 * length there. The preface says to incorporate its examples into your
9609 * code at your own risk.
9611 * The algorithm is like a merge sort. */
9613 const UV* array_a; /* a's array */
9615 UV len_a; /* length of a's array */
9618 SV* u; /* the resulting union */
9622 UV i_a = 0; /* current index into a's array */
9626 /* running count, as explained in the algorithm source book; items are
9627 * stopped accumulating and are output when the count changes to/from 0.
9628 * The count is incremented when we start a range that's in an input's set,
9629 * and decremented when we start a range that's not in a set. So this
9630 * variable can be 0, 1, or 2. When it is 0 neither input is in their set,
9631 * and hence nothing goes into the union; 1, just one of the inputs is in
9632 * its set (and its current range gets added to the union); and 2 when both
9633 * inputs are in their sets. */
9636 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
9638 assert(*output == NULL || is_invlist(*output));
9640 len_b = _invlist_len(b);
9643 /* Here, 'b' is empty, hence it's complement is all possible code
9644 * points. So if the union includes the complement of 'b', it includes
9645 * everything, and we need not even look at 'a'. It's easiest to
9646 * create a new inversion list that matches everything. */
9648 SV* everything = _add_range_to_invlist(NULL, 0, UV_MAX);
9650 if (*output == NULL) { /* If the output didn't exist, just point it
9652 *output = everything;
9654 else { /* Otherwise, replace its contents with the new list */
9655 invlist_replace_list_destroys_src(*output, everything);
9656 SvREFCNT_dec_NN(everything);
9662 /* Here, we don't want the complement of 'b', and since 'b' is empty,
9663 * the union will come entirely from 'a'. If 'a' is NULL or empty, the
9664 * output will be empty */
9666 if (a == NULL || _invlist_len(a) == 0) {
9667 if (*output == NULL) {
9668 *output = _new_invlist(0);
9671 invlist_clear(*output);
9676 /* Here, 'a' is not empty, but 'b' is, so 'a' entirely determines the
9677 * union. We can just return a copy of 'a' if '*output' doesn't point
9678 * to an existing list */
9679 if (*output == NULL) {
9680 *output = invlist_clone(a, NULL);
9684 /* If the output is to overwrite 'a', we have a no-op, as it's
9690 /* Here, '*output' is to be overwritten by 'a' */
9691 u = invlist_clone(a, NULL);
9692 invlist_replace_list_destroys_src(*output, u);
9698 /* Here 'b' is not empty. See about 'a' */
9700 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
9702 /* Here, 'a' is empty (and b is not). That means the union will come
9703 * entirely from 'b'. If '*output' is NULL, we can directly return a
9704 * clone of 'b'. Otherwise, we replace the contents of '*output' with
9707 SV ** dest = (*output == NULL) ? output : &u;
9708 *dest = invlist_clone(b, NULL);
9710 _invlist_invert(*dest);
9714 invlist_replace_list_destroys_src(*output, u);
9721 /* Here both lists exist and are non-empty */
9722 array_a = invlist_array(a);
9723 array_b = invlist_array(b);
9725 /* If are to take the union of 'a' with the complement of b, set it
9726 * up so are looking at b's complement. */
9729 /* To complement, we invert: if the first element is 0, remove it. To
9730 * do this, we just pretend the array starts one later */
9731 if (array_b[0] == 0) {
9737 /* But if the first element is not zero, we pretend the list starts
9738 * at the 0 that is always stored immediately before the array. */
9744 /* Size the union for the worst case: that the sets are completely
9746 u = _new_invlist(len_a + len_b);
9748 /* Will contain U+0000 if either component does */
9749 array_u = _invlist_array_init(u, ( len_a > 0 && array_a[0] == 0)
9750 || (len_b > 0 && array_b[0] == 0));
9752 /* Go through each input list item by item, stopping when have exhausted
9754 while (i_a < len_a && i_b < len_b) {
9755 UV cp; /* The element to potentially add to the union's array */
9756 bool cp_in_set; /* is it in the input list's set or not */
9758 /* We need to take one or the other of the two inputs for the union.
9759 * Since we are merging two sorted lists, we take the smaller of the
9760 * next items. In case of a tie, we take first the one that is in its
9761 * set. If we first took the one not in its set, it would decrement
9762 * the count, possibly to 0 which would cause it to be output as ending
9763 * the range, and the next time through we would take the same number,
9764 * and output it again as beginning the next range. By doing it the
9765 * opposite way, there is no possibility that the count will be
9766 * momentarily decremented to 0, and thus the two adjoining ranges will
9767 * be seamlessly merged. (In a tie and both are in the set or both not
9768 * in the set, it doesn't matter which we take first.) */
9769 if ( array_a[i_a] < array_b[i_b]
9770 || ( array_a[i_a] == array_b[i_b]
9771 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9773 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9774 cp = array_a[i_a++];
9777 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9778 cp = array_b[i_b++];
9781 /* Here, have chosen which of the two inputs to look at. Only output
9782 * if the running count changes to/from 0, which marks the
9783 * beginning/end of a range that's in the set */
9786 array_u[i_u++] = cp;
9793 array_u[i_u++] = cp;
9799 /* The loop above increments the index into exactly one of the input lists
9800 * each iteration, and ends when either index gets to its list end. That
9801 * means the other index is lower than its end, and so something is
9802 * remaining in that one. We decrement 'count', as explained below, if
9803 * that list is in its set. (i_a and i_b each currently index the element
9804 * beyond the one we care about.) */
9805 if ( (i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9806 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9811 /* Above we decremented 'count' if the list that had unexamined elements in
9812 * it was in its set. This has made it so that 'count' being non-zero
9813 * means there isn't anything left to output; and 'count' equal to 0 means
9814 * that what is left to output is precisely that which is left in the
9815 * non-exhausted input list.
9817 * To see why, note first that the exhausted input obviously has nothing
9818 * left to add to the union. If it was in its set at its end, that means
9819 * the set extends from here to the platform's infinity, and hence so does
9820 * the union and the non-exhausted set is irrelevant. The exhausted set
9821 * also contributed 1 to 'count'. If 'count' was 2, it got decremented to
9822 * 1, but if it was 1, the non-exhausted set wasn't in its set, and so
9823 * 'count' remains at 1. This is consistent with the decremented 'count'
9824 * != 0 meaning there's nothing left to add to the union.
9826 * But if the exhausted input wasn't in its set, it contributed 0 to
9827 * 'count', and the rest of the union will be whatever the other input is.
9828 * If 'count' was 0, neither list was in its set, and 'count' remains 0;
9829 * otherwise it gets decremented to 0. This is consistent with 'count'
9830 * == 0 meaning the remainder of the union is whatever is left in the
9831 * non-exhausted list. */
9836 IV copy_count = len_a - i_a;
9837 if (copy_count > 0) { /* The non-exhausted input is 'a' */
9838 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
9840 else { /* The non-exhausted input is b */
9841 copy_count = len_b - i_b;
9842 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
9844 len_u = i_u + copy_count;
9847 /* Set the result to the final length, which can change the pointer to
9848 * array_u, so re-find it. (Note that it is unlikely that this will
9849 * change, as we are shrinking the space, not enlarging it) */
9850 if (len_u != _invlist_len(u)) {
9851 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
9853 array_u = invlist_array(u);
9856 if (*output == NULL) { /* Simply return the new inversion list */
9860 /* Otherwise, overwrite the inversion list that was in '*output'. We
9861 * could instead free '*output', and then set it to 'u', but experience
9862 * has shown [perl #127392] that if the input is a mortal, we can get a
9863 * huge build-up of these during regex compilation before they get
9865 invlist_replace_list_destroys_src(*output, u);
9873 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9874 const bool complement_b, SV** i)
9876 /* Take the intersection of two inversion lists and point '*i' to it. On
9877 * input, '*i' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9878 * even 'a' or 'b'). If to an inversion list, the contents of the original
9879 * list will be replaced by the intersection. The first list, 'a', may be
9880 * NULL, in which case '*i' will be an empty list. If 'complement_b' is
9881 * TRUE, the result will be the intersection of 'a' and the complement (or
9882 * inversion) of 'b' instead of 'b' directly.
9884 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9885 * Richard Gillam, published by Addison-Wesley, and explained at some
9886 * length there. The preface says to incorporate its examples into your
9887 * code at your own risk. In fact, it had bugs
9889 * The algorithm is like a merge sort, and is essentially the same as the
9893 const UV* array_a; /* a's array */
9895 UV len_a; /* length of a's array */
9898 SV* r; /* the resulting intersection */
9902 UV i_a = 0; /* current index into a's array */
9906 /* running count of how many of the two inputs are postitioned at ranges
9907 * that are in their sets. As explained in the algorithm source book,
9908 * items are stopped accumulating and are output when the count changes
9909 * to/from 2. The count is incremented when we start a range that's in an
9910 * input's set, and decremented when we start a range that's not in a set.
9911 * Only when it is 2 are we in the intersection. */
9914 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
9916 assert(*i == NULL || is_invlist(*i));
9918 /* Special case if either one is empty */
9919 len_a = (a == NULL) ? 0 : _invlist_len(a);
9920 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
9921 if (len_a != 0 && complement_b) {
9923 /* Here, 'a' is not empty, therefore from the enclosing 'if', 'b'
9924 * must be empty. Here, also we are using 'b's complement, which
9925 * hence must be every possible code point. Thus the intersection
9928 if (*i == a) { /* No-op */
9933 *i = invlist_clone(a, NULL);
9937 r = invlist_clone(a, NULL);
9938 invlist_replace_list_destroys_src(*i, r);
9943 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
9944 * intersection must be empty */
9946 *i = _new_invlist(0);
9954 /* Here both lists exist and are non-empty */
9955 array_a = invlist_array(a);
9956 array_b = invlist_array(b);
9958 /* If are to take the intersection of 'a' with the complement of b, set it
9959 * up so are looking at b's complement. */
9962 /* To complement, we invert: if the first element is 0, remove it. To
9963 * do this, we just pretend the array starts one later */
9964 if (array_b[0] == 0) {
9970 /* But if the first element is not zero, we pretend the list starts
9971 * at the 0 that is always stored immediately before the array. */
9977 /* Size the intersection for the worst case: that the intersection ends up
9978 * fragmenting everything to be completely disjoint */
9979 r= _new_invlist(len_a + len_b);
9981 /* Will contain U+0000 iff both components do */
9982 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
9983 && len_b > 0 && array_b[0] == 0);
9985 /* Go through each list item by item, stopping when have exhausted one of
9987 while (i_a < len_a && i_b < len_b) {
9988 UV cp; /* The element to potentially add to the intersection's
9990 bool cp_in_set; /* Is it in the input list's set or not */
9992 /* We need to take one or the other of the two inputs for the
9993 * intersection. Since we are merging two sorted lists, we take the
9994 * smaller of the next items. In case of a tie, we take first the one
9995 * that is not in its set (a difference from the union algorithm). If
9996 * we first took the one in its set, it would increment the count,
9997 * possibly to 2 which would cause it to be output as starting a range
9998 * in the intersection, and the next time through we would take that
9999 * same number, and output it again as ending the set. By doing the
10000 * opposite of this, there is no possibility that the count will be
10001 * momentarily incremented to 2. (In a tie and both are in the set or
10002 * both not in the set, it doesn't matter which we take first.) */
10003 if ( array_a[i_a] < array_b[i_b]
10004 || ( array_a[i_a] == array_b[i_b]
10005 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
10007 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
10008 cp = array_a[i_a++];
10011 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
10012 cp= array_b[i_b++];
10015 /* Here, have chosen which of the two inputs to look at. Only output
10016 * if the running count changes to/from 2, which marks the
10017 * beginning/end of a range that's in the intersection */
10021 array_r[i_r++] = cp;
10026 array_r[i_r++] = cp;
10033 /* The loop above increments the index into exactly one of the input lists
10034 * each iteration, and ends when either index gets to its list end. That
10035 * means the other index is lower than its end, and so something is
10036 * remaining in that one. We increment 'count', as explained below, if the
10037 * exhausted list was in its set. (i_a and i_b each currently index the
10038 * element beyond the one we care about.) */
10039 if ( (i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
10040 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
10045 /* Above we incremented 'count' if the exhausted list was in its set. This
10046 * has made it so that 'count' being below 2 means there is nothing left to
10047 * output; otheriwse what's left to add to the intersection is precisely
10048 * that which is left in the non-exhausted input list.
10050 * To see why, note first that the exhausted input obviously has nothing
10051 * left to affect the intersection. If it was in its set at its end, that
10052 * means the set extends from here to the platform's infinity, and hence
10053 * anything in the non-exhausted's list will be in the intersection, and
10054 * anything not in it won't be. Hence, the rest of the intersection is
10055 * precisely what's in the non-exhausted list The exhausted set also
10056 * contributed 1 to 'count', meaning 'count' was at least 1. Incrementing
10057 * it means 'count' is now at least 2. This is consistent with the
10058 * incremented 'count' being >= 2 means to add the non-exhausted list to
10059 * the intersection.
10061 * But if the exhausted input wasn't in its set, it contributed 0 to
10062 * 'count', and the intersection can't include anything further; the
10063 * non-exhausted set is irrelevant. 'count' was at most 1, and doesn't get
10064 * incremented. This is consistent with 'count' being < 2 meaning nothing
10065 * further to add to the intersection. */
10066 if (count < 2) { /* Nothing left to put in the intersection. */
10069 else { /* copy the non-exhausted list, unchanged. */
10070 IV copy_count = len_a - i_a;
10071 if (copy_count > 0) { /* a is the one with stuff left */
10072 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
10074 else { /* b is the one with stuff left */
10075 copy_count = len_b - i_b;
10076 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
10078 len_r = i_r + copy_count;
10081 /* Set the result to the final length, which can change the pointer to
10082 * array_r, so re-find it. (Note that it is unlikely that this will
10083 * change, as we are shrinking the space, not enlarging it) */
10084 if (len_r != _invlist_len(r)) {
10085 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
10087 array_r = invlist_array(r);
10090 if (*i == NULL) { /* Simply return the calculated intersection */
10093 else { /* Otherwise, replace the existing inversion list in '*i'. We could
10094 instead free '*i', and then set it to 'r', but experience has
10095 shown [perl #127392] that if the input is a mortal, we can get a
10096 huge build-up of these during regex compilation before they get
10099 invlist_replace_list_destroys_src(*i, r);
10104 SvREFCNT_dec_NN(r);
10111 Perl__add_range_to_invlist(pTHX_ SV* invlist, UV start, UV end)
10113 /* Add the range from 'start' to 'end' inclusive to the inversion list's
10114 * set. A pointer to the inversion list is returned. This may actually be
10115 * a new list, in which case the passed in one has been destroyed. The
10116 * passed-in inversion list can be NULL, in which case a new one is created
10117 * with just the one range in it. The new list is not necessarily
10118 * NUL-terminated. Space is not freed if the inversion list shrinks as a
10119 * result of this function. The gain would not be large, and in many
10120 * cases, this is called multiple times on a single inversion list, so
10121 * anything freed may almost immediately be needed again.
10123 * This used to mostly call the 'union' routine, but that is much more
10124 * heavyweight than really needed for a single range addition */
10126 UV* array; /* The array implementing the inversion list */
10127 UV len; /* How many elements in 'array' */
10128 SSize_t i_s; /* index into the invlist array where 'start'
10130 SSize_t i_e = 0; /* And the index where 'end' should go */
10131 UV cur_highest; /* The highest code point in the inversion list
10132 upon entry to this function */
10134 /* This range becomes the whole inversion list if none already existed */
10135 if (invlist == NULL) {
10136 invlist = _new_invlist(2);
10137 _append_range_to_invlist(invlist, start, end);
10141 /* Likewise, if the inversion list is currently empty */
10142 len = _invlist_len(invlist);
10144 _append_range_to_invlist(invlist, start, end);
10148 /* Starting here, we have to know the internals of the list */
10149 array = invlist_array(invlist);
10151 /* If the new range ends higher than the current highest ... */
10152 cur_highest = invlist_highest(invlist);
10153 if (end > cur_highest) {
10155 /* If the whole range is higher, we can just append it */
10156 if (start > cur_highest) {
10157 _append_range_to_invlist(invlist, start, end);
10161 /* Otherwise, add the portion that is higher ... */
10162 _append_range_to_invlist(invlist, cur_highest + 1, end);
10164 /* ... and continue on below to handle the rest. As a result of the
10165 * above append, we know that the index of the end of the range is the
10166 * final even numbered one of the array. Recall that the final element
10167 * always starts a range that extends to infinity. If that range is in
10168 * the set (meaning the set goes from here to infinity), it will be an
10169 * even index, but if it isn't in the set, it's odd, and the final
10170 * range in the set is one less, which is even. */
10171 if (end == UV_MAX) {
10179 /* We have dealt with appending, now see about prepending. If the new
10180 * range starts lower than the current lowest ... */
10181 if (start < array[0]) {
10183 /* Adding something which has 0 in it is somewhat tricky, and uncommon.
10184 * Let the union code handle it, rather than having to know the
10185 * trickiness in two code places. */
10186 if (UNLIKELY(start == 0)) {
10189 range_invlist = _new_invlist(2);
10190 _append_range_to_invlist(range_invlist, start, end);
10192 _invlist_union(invlist, range_invlist, &invlist);
10194 SvREFCNT_dec_NN(range_invlist);
10199 /* If the whole new range comes before the first entry, and doesn't
10200 * extend it, we have to insert it as an additional range */
10201 if (end < array[0] - 1) {
10203 goto splice_in_new_range;
10206 /* Here the new range adjoins the existing first range, extending it
10210 /* And continue on below to handle the rest. We know that the index of
10211 * the beginning of the range is the first one of the array */
10214 else { /* Not prepending any part of the new range to the existing list.
10215 * Find where in the list it should go. This finds i_s, such that:
10216 * invlist[i_s] <= start < array[i_s+1]
10218 i_s = _invlist_search(invlist, start);
10221 /* At this point, any extending before the beginning of the inversion list
10222 * and/or after the end has been done. This has made it so that, in the
10223 * code below, each endpoint of the new range is either in a range that is
10224 * in the set, or is in a gap between two ranges that are. This means we
10225 * don't have to worry about exceeding the array bounds.
10227 * Find where in the list the new range ends (but we can skip this if we
10228 * have already determined what it is, or if it will be the same as i_s,
10229 * which we already have computed) */
10231 i_e = (start == end)
10233 : _invlist_search(invlist, end);
10236 /* Here generally invlist[i_e] <= end < array[i_e+1]. But if invlist[i_e]
10237 * is a range that goes to infinity there is no element at invlist[i_e+1],
10238 * so only the first relation holds. */
10240 if ( ! ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
10242 /* Here, the ranges on either side of the beginning of the new range
10243 * are in the set, and this range starts in the gap between them.
10245 * The new range extends the range above it downwards if the new range
10246 * ends at or above that range's start */
10247 const bool extends_the_range_above = ( end == UV_MAX
10248 || end + 1 >= array[i_s+1]);
10250 /* The new range extends the range below it upwards if it begins just
10251 * after where that range ends */
10252 if (start == array[i_s]) {
10254 /* If the new range fills the entire gap between the other ranges,
10255 * they will get merged together. Other ranges may also get
10256 * merged, depending on how many of them the new range spans. In
10257 * the general case, we do the merge later, just once, after we
10258 * figure out how many to merge. But in the case where the new
10259 * range exactly spans just this one gap (possibly extending into
10260 * the one above), we do the merge here, and an early exit. This
10261 * is done here to avoid having to special case later. */
10262 if (i_e - i_s <= 1) {
10264 /* If i_e - i_s == 1, it means that the new range terminates
10265 * within the range above, and hence 'extends_the_range_above'
10266 * must be true. (If the range above it extends to infinity,
10267 * 'i_s+2' will be above the array's limit, but 'len-i_s-2'
10268 * will be 0, so no harm done.) */
10269 if (extends_the_range_above) {
10270 Move(array + i_s + 2, array + i_s, len - i_s - 2, UV);
10271 invlist_set_len(invlist,
10273 *(get_invlist_offset_addr(invlist)));
10277 /* Here, i_e must == i_s. We keep them in sync, as they apply
10278 * to the same range, and below we are about to decrement i_s
10283 /* Here, the new range is adjacent to the one below. (It may also
10284 * span beyond the range above, but that will get resolved later.)
10285 * Extend the range below to include this one. */
10286 array[i_s] = (end == UV_MAX) ? UV_MAX : end + 1;
10288 start = array[i_s];
10290 else if (extends_the_range_above) {
10292 /* Here the new range only extends the range above it, but not the
10293 * one below. It merges with the one above. Again, we keep i_e
10294 * and i_s in sync if they point to the same range */
10299 array[i_s] = start;
10303 /* Here, we've dealt with the new range start extending any adjoining
10306 * If the new range extends to infinity, it is now the final one,
10307 * regardless of what was there before */
10308 if (UNLIKELY(end == UV_MAX)) {
10309 invlist_set_len(invlist, i_s + 1, *(get_invlist_offset_addr(invlist)));
10313 /* If i_e started as == i_s, it has also been dealt with,
10314 * and been updated to the new i_s, which will fail the following if */
10315 if (! ELEMENT_RANGE_MATCHES_INVLIST(i_e)) {
10317 /* Here, the ranges on either side of the end of the new range are in
10318 * the set, and this range ends in the gap between them.
10320 * If this range is adjacent to (hence extends) the range above it, it
10321 * becomes part of that range; likewise if it extends the range below,
10322 * it becomes part of that range */
10323 if (end + 1 == array[i_e+1]) {
10325 array[i_e] = start;
10327 else if (start <= array[i_e]) {
10328 array[i_e] = end + 1;
10335 /* If the range fits entirely in an existing range (as possibly already
10336 * extended above), it doesn't add anything new */
10337 if (ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
10341 /* Here, no part of the range is in the list. Must add it. It will
10342 * occupy 2 more slots */
10343 splice_in_new_range:
10345 invlist_extend(invlist, len + 2);
10346 array = invlist_array(invlist);
10347 /* Move the rest of the array down two slots. Don't include any
10349 Move(array + i_e + 1, array + i_e + 3, len - i_e - 1, UV);
10351 /* Do the actual splice */
10352 array[i_e+1] = start;
10353 array[i_e+2] = end + 1;
10354 invlist_set_len(invlist, len + 2, *(get_invlist_offset_addr(invlist)));
10358 /* Here the new range crossed the boundaries of a pre-existing range. The
10359 * code above has adjusted things so that both ends are in ranges that are
10360 * in the set. This means everything in between must also be in the set.
10361 * Just squash things together */
10362 Move(array + i_e + 1, array + i_s + 1, len - i_e - 1, UV);
10363 invlist_set_len(invlist,
10365 *(get_invlist_offset_addr(invlist)));
10371 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
10372 UV** other_elements_ptr)
10374 /* Create and return an inversion list whose contents are to be populated
10375 * by the caller. The caller gives the number of elements (in 'size') and
10376 * the very first element ('element0'). This function will set
10377 * '*other_elements_ptr' to an array of UVs, where the remaining elements
10378 * are to be placed.
10380 * Obviously there is some trust involved that the caller will properly
10381 * fill in the other elements of the array.
10383 * (The first element needs to be passed in, as the underlying code does
10384 * things differently depending on whether it is zero or non-zero) */
10386 SV* invlist = _new_invlist(size);
10389 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
10391 invlist = add_cp_to_invlist(invlist, element0);
10392 offset = *get_invlist_offset_addr(invlist);
10394 invlist_set_len(invlist, size, offset);
10395 *other_elements_ptr = invlist_array(invlist) + 1;
10401 #ifndef PERL_IN_XSUB_RE
10403 Perl__invlist_invert(pTHX_ SV* const invlist)
10405 /* Complement the input inversion list. This adds a 0 if the list didn't
10406 * have a zero; removes it otherwise. As described above, the data
10407 * structure is set up so that this is very efficient */
10409 PERL_ARGS_ASSERT__INVLIST_INVERT;
10411 assert(! invlist_is_iterating(invlist));
10413 /* The inverse of matching nothing is matching everything */
10414 if (_invlist_len(invlist) == 0) {
10415 _append_range_to_invlist(invlist, 0, UV_MAX);
10419 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
10423 Perl_invlist_clone(pTHX_ SV* const invlist, SV* new_invlist)
10425 /* Return a new inversion list that is a copy of the input one, which is
10426 * unchanged. The new list will not be mortal even if the old one was. */
10428 const STRLEN nominal_length = _invlist_len(invlist);
10429 const STRLEN physical_length = SvCUR(invlist);
10430 const bool offset = *(get_invlist_offset_addr(invlist));
10432 PERL_ARGS_ASSERT_INVLIST_CLONE;
10434 if (new_invlist == NULL) {
10435 new_invlist = _new_invlist(nominal_length);
10438 sv_upgrade(new_invlist, SVt_INVLIST);
10439 initialize_invlist_guts(new_invlist, nominal_length);
10442 *(get_invlist_offset_addr(new_invlist)) = offset;
10443 invlist_set_len(new_invlist, nominal_length, offset);
10444 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
10446 return new_invlist;
10451 PERL_STATIC_INLINE UV
10452 S_invlist_lowest(SV* const invlist)
10454 /* Returns the lowest code point that matches an inversion list. This API
10455 * has an ambiguity, as it returns 0 under either the lowest is actually
10456 * 0, or if the list is empty. If this distinction matters to you, check
10457 * for emptiness before calling this function */
10459 UV len = _invlist_len(invlist);
10462 PERL_ARGS_ASSERT_INVLIST_LOWEST;
10468 array = invlist_array(invlist);
10474 S_invlist_contents(pTHX_ SV* const invlist, const bool traditional_style)
10476 /* Get the contents of an inversion list into a string SV so that they can
10477 * be printed out. If 'traditional_style' is TRUE, it uses the format
10478 * traditionally done for debug tracing; otherwise it uses a format
10479 * suitable for just copying to the output, with blanks between ranges and
10480 * a dash between range components */
10484 const char intra_range_delimiter = (traditional_style ? '\t' : '-');
10485 const char inter_range_delimiter = (traditional_style ? '\n' : ' ');
10487 if (traditional_style) {
10488 output = newSVpvs("\n");
10491 output = newSVpvs("");
10494 PERL_ARGS_ASSERT_INVLIST_CONTENTS;
10496 assert(! invlist_is_iterating(invlist));
10498 invlist_iterinit(invlist);
10499 while (invlist_iternext(invlist, &start, &end)) {
10500 if (end == UV_MAX) {
10501 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%cINFTY%c",
10502 start, intra_range_delimiter,
10503 inter_range_delimiter);
10505 else if (end != start) {
10506 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c%04" UVXf "%c",
10508 intra_range_delimiter,
10509 end, inter_range_delimiter);
10512 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c",
10513 start, inter_range_delimiter);
10517 if (SvCUR(output) && ! traditional_style) {/* Get rid of trailing blank */
10518 SvCUR_set(output, SvCUR(output) - 1);
10524 #ifndef PERL_IN_XSUB_RE
10526 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
10527 const char * const indent, SV* const invlist)
10529 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
10530 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
10531 * the string 'indent'. The output looks like this:
10532 [0] 0x000A .. 0x000D
10534 [4] 0x2028 .. 0x2029
10535 [6] 0x3104 .. INFTY
10536 * This means that the first range of code points matched by the list are
10537 * 0xA through 0xD; the second range contains only the single code point
10538 * 0x85, etc. An inversion list is an array of UVs. Two array elements
10539 * are used to define each range (except if the final range extends to
10540 * infinity, only a single element is needed). The array index of the
10541 * first element for the corresponding range is given in brackets. */
10546 PERL_ARGS_ASSERT__INVLIST_DUMP;
10548 if (invlist_is_iterating(invlist)) {
10549 Perl_dump_indent(aTHX_ level, file,
10550 "%sCan't dump inversion list because is in middle of iterating\n",
10555 invlist_iterinit(invlist);
10556 while (invlist_iternext(invlist, &start, &end)) {
10557 if (end == UV_MAX) {
10558 Perl_dump_indent(aTHX_ level, file,
10559 "%s[%" UVuf "] 0x%04" UVXf " .. INFTY\n",
10560 indent, (UV)count, start);
10562 else if (end != start) {
10563 Perl_dump_indent(aTHX_ level, file,
10564 "%s[%" UVuf "] 0x%04" UVXf " .. 0x%04" UVXf "\n",
10565 indent, (UV)count, start, end);
10568 Perl_dump_indent(aTHX_ level, file, "%s[%" UVuf "] 0x%04" UVXf "\n",
10569 indent, (UV)count, start);
10577 #if defined(PERL_ARGS_ASSERT__INVLISTEQ) && !defined(PERL_IN_XSUB_RE)
10579 Perl__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
10581 /* Return a boolean as to if the two passed in inversion lists are
10582 * identical. The final argument, if TRUE, says to take the complement of
10583 * the second inversion list before doing the comparison */
10585 const UV len_a = _invlist_len(a);
10586 UV len_b = _invlist_len(b);
10588 const UV* array_a = NULL;
10589 const UV* array_b = NULL;
10591 PERL_ARGS_ASSERT__INVLISTEQ;
10593 /* This code avoids accessing the arrays unless it knows the length is
10598 return ! complement_b;
10602 array_a = invlist_array(a);
10606 array_b = invlist_array(b);
10609 /* If are to compare 'a' with the complement of b, set it
10610 * up so are looking at b's complement. */
10611 if (complement_b) {
10613 /* The complement of nothing is everything, so <a> would have to have
10614 * just one element, starting at zero (ending at infinity) */
10616 return (len_a == 1 && array_a[0] == 0);
10618 if (array_b[0] == 0) {
10620 /* Otherwise, to complement, we invert. Here, the first element is
10621 * 0, just remove it. To do this, we just pretend the array starts
10629 /* But if the first element is not zero, we pretend the list starts
10630 * at the 0 that is always stored immediately before the array. */
10636 return len_a == len_b
10637 && memEQ(array_a, array_b, len_a * sizeof(array_a[0]));
10643 * As best we can, determine the characters that can match the start of
10644 * the given EXACTF-ish node. This is for use in creating ssc nodes, so there
10645 * can be false positive matches
10647 * Returns the invlist as a new SV*; it is the caller's responsibility to
10648 * call SvREFCNT_dec() when done with it.
10651 S_make_exactf_invlist(pTHX_ RExC_state_t *pRExC_state, regnode *node)
10653 const U8 * s = (U8*)STRING(node);
10654 SSize_t bytelen = STR_LEN(node);
10656 /* Start out big enough for 2 separate code points */
10657 SV* invlist = _new_invlist(4);
10659 PERL_ARGS_ASSERT_MAKE_EXACTF_INVLIST;
10664 /* We punt and assume can match anything if the node begins
10665 * with a multi-character fold. Things are complicated. For
10666 * example, /ffi/i could match any of:
10667 * "\N{LATIN SMALL LIGATURE FFI}"
10668 * "\N{LATIN SMALL LIGATURE FF}I"
10669 * "F\N{LATIN SMALL LIGATURE FI}"
10670 * plus several other things; and making sure we have all the
10671 * possibilities is hard. */
10672 if (is_MULTI_CHAR_FOLD_latin1_safe(s, s + bytelen)) {
10673 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10676 /* Any Latin1 range character can potentially match any
10677 * other depending on the locale, and in Turkic locales, U+130 and
10679 if (OP(node) == EXACTFL) {
10680 _invlist_union(invlist, PL_Latin1, &invlist);
10681 invlist = add_cp_to_invlist(invlist,
10682 LATIN_SMALL_LETTER_DOTLESS_I);
10683 invlist = add_cp_to_invlist(invlist,
10684 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
10687 /* But otherwise, it matches at least itself. We can
10688 * quickly tell if it has a distinct fold, and if so,
10689 * it matches that as well */
10690 invlist = add_cp_to_invlist(invlist, uc);
10691 if (IS_IN_SOME_FOLD_L1(uc))
10692 invlist = add_cp_to_invlist(invlist, PL_fold_latin1[uc]);
10695 /* Some characters match above-Latin1 ones under /i. This
10696 * is true of EXACTFL ones when the locale is UTF-8 */
10697 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(uc)
10698 && (! isASCII(uc) || ! inRANGE(OP(node), EXACTFAA,
10699 EXACTFAA_NO_TRIE)))
10701 add_above_Latin1_folds(pRExC_state, (U8) uc, &invlist);
10705 else { /* Pattern is UTF-8 */
10706 U8 folded[UTF8_MAX_FOLD_CHAR_EXPAND * UTF8_MAXBYTES_CASE + 1] = { '\0' };
10707 const U8* e = s + bytelen;
10710 fc = uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
10712 /* The only code points that aren't folded in a UTF EXACTFish
10713 * node are the problematic ones in EXACTFL nodes */
10714 if (OP(node) == EXACTFL && is_PROBLEMATIC_LOCALE_FOLDEDS_START_cp(uc)) {
10715 /* We need to check for the possibility that this EXACTFL
10716 * node begins with a multi-char fold. Therefore we fold
10717 * the first few characters of it so that we can make that
10723 for (i = 0; i < UTF8_MAX_FOLD_CHAR_EXPAND && s < e; i++) {
10725 *(d++) = (U8) toFOLD(*s);
10726 if (fc < 0) { /* Save the first fold */
10733 UV fold = toFOLD_utf8_safe(s, e, d, &len);
10734 if (fc < 0) { /* Save the first fold */
10742 /* And set up so the code below that looks in this folded
10743 * buffer instead of the node's string */
10748 /* When we reach here 's' points to the fold of the first
10749 * character(s) of the node; and 'e' points to far enough along
10750 * the folded string to be just past any possible multi-char
10753 * Like the non-UTF case above, we punt if the node begins with a
10754 * multi-char fold */
10756 if (is_MULTI_CHAR_FOLD_utf8_safe(s, e)) {
10757 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10759 else { /* Single char fold */
10762 const U32 * remaining_folds;
10763 Size_t folds_count;
10765 /* It matches itself */
10766 invlist = add_cp_to_invlist(invlist, fc);
10768 /* ... plus all the things that fold to it, which are found in
10769 * PL_utf8_foldclosures */
10770 folds_count = _inverse_folds(fc, &first_fold,
10772 for (k = 0; k < folds_count; k++) {
10773 UV c = (k == 0) ? first_fold : remaining_folds[k-1];
10775 /* /aa doesn't allow folds between ASCII and non- */
10776 if ( inRANGE(OP(node), EXACTFAA, EXACTFAA_NO_TRIE)
10777 && isASCII(c) != isASCII(fc))
10782 invlist = add_cp_to_invlist(invlist, c);
10785 if (OP(node) == EXACTFL) {
10787 /* If either [iI] are present in an EXACTFL node the above code
10788 * should have added its normal case pair, but under a Turkish
10789 * locale they could match instead the case pairs from it. Add
10790 * those as potential matches as well */
10791 if (isALPHA_FOLD_EQ(fc, 'I')) {
10792 invlist = add_cp_to_invlist(invlist,
10793 LATIN_SMALL_LETTER_DOTLESS_I);
10794 invlist = add_cp_to_invlist(invlist,
10795 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
10797 else if (fc == LATIN_SMALL_LETTER_DOTLESS_I) {
10798 invlist = add_cp_to_invlist(invlist, 'I');
10800 else if (fc == LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE) {
10801 invlist = add_cp_to_invlist(invlist, 'i');
10810 #undef HEADER_LENGTH
10811 #undef TO_INTERNAL_SIZE
10812 #undef FROM_INTERNAL_SIZE
10813 #undef INVLIST_VERSION_ID
10815 /* End of inversion list object */
10818 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
10820 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
10821 * constructs, and updates RExC_flags with them. On input, RExC_parse
10822 * should point to the first flag; it is updated on output to point to the
10823 * final ')' or ':'. There needs to be at least one flag, or this will
10826 /* for (?g), (?gc), and (?o) warnings; warning
10827 about (?c) will warn about (?g) -- japhy */
10829 #define WASTED_O 0x01
10830 #define WASTED_G 0x02
10831 #define WASTED_C 0x04
10832 #define WASTED_GC (WASTED_G|WASTED_C)
10833 I32 wastedflags = 0x00;
10834 U32 posflags = 0, negflags = 0;
10835 U32 *flagsp = &posflags;
10836 char has_charset_modifier = '\0';
10838 bool has_use_defaults = FALSE;
10839 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
10840 int x_mod_count = 0;
10842 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
10844 /* '^' as an initial flag sets certain defaults */
10845 if (UCHARAT(RExC_parse) == '^') {
10847 has_use_defaults = TRUE;
10848 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
10849 cs = (toUSE_UNI_CHARSET_NOT_DEPENDS)
10850 ? REGEX_UNICODE_CHARSET
10851 : REGEX_DEPENDS_CHARSET;
10852 set_regex_charset(&RExC_flags, cs);
10855 cs = get_regex_charset(RExC_flags);
10856 if ( cs == REGEX_DEPENDS_CHARSET
10857 && (toUSE_UNI_CHARSET_NOT_DEPENDS))
10859 cs = REGEX_UNICODE_CHARSET;
10863 while (RExC_parse < RExC_end) {
10864 /* && memCHRs("iogcmsx", *RExC_parse) */
10865 /* (?g), (?gc) and (?o) are useless here
10866 and must be globally applied -- japhy */
10867 if ((RExC_pm_flags & PMf_WILDCARD)) {
10868 if (flagsp == & negflags) {
10869 if (*RExC_parse == 'm') {
10871 /* diag_listed_as: Use of %s is not allowed in Unicode
10872 property wildcard subpatterns in regex; marked by <--
10874 vFAIL("Use of modifier '-m' is not allowed in Unicode"
10875 " property wildcard subpatterns");
10879 if (*RExC_parse == 's') {
10880 goto modifier_illegal_in_wildcard;
10885 switch (*RExC_parse) {
10887 /* Code for the imsxn flags */
10888 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp, x_mod_count);
10890 case LOCALE_PAT_MOD:
10891 if (has_charset_modifier) {
10892 goto excess_modifier;
10894 else if (flagsp == &negflags) {
10897 cs = REGEX_LOCALE_CHARSET;
10898 has_charset_modifier = LOCALE_PAT_MOD;
10900 case UNICODE_PAT_MOD:
10901 if (has_charset_modifier) {
10902 goto excess_modifier;
10904 else if (flagsp == &negflags) {
10907 cs = REGEX_UNICODE_CHARSET;
10908 has_charset_modifier = UNICODE_PAT_MOD;
10910 case ASCII_RESTRICT_PAT_MOD:
10911 if (flagsp == &negflags) {
10914 if (has_charset_modifier) {
10915 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
10916 goto excess_modifier;
10918 /* Doubled modifier implies more restricted */
10919 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
10922 cs = REGEX_ASCII_RESTRICTED_CHARSET;
10924 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
10926 case DEPENDS_PAT_MOD:
10927 if (has_use_defaults) {
10928 goto fail_modifiers;
10930 else if (flagsp == &negflags) {
10933 else if (has_charset_modifier) {
10934 goto excess_modifier;
10937 /* The dual charset means unicode semantics if the
10938 * pattern (or target, not known until runtime) are
10939 * utf8, or something in the pattern indicates unicode
10941 cs = (toUSE_UNI_CHARSET_NOT_DEPENDS)
10942 ? REGEX_UNICODE_CHARSET
10943 : REGEX_DEPENDS_CHARSET;
10944 has_charset_modifier = DEPENDS_PAT_MOD;
10948 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
10949 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
10951 else if (has_charset_modifier == *(RExC_parse - 1)) {
10952 vFAIL2("Regexp modifier \"%c\" may not appear twice",
10953 *(RExC_parse - 1));
10956 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
10958 NOT_REACHED; /*NOTREACHED*/
10961 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
10962 *(RExC_parse - 1));
10963 NOT_REACHED; /*NOTREACHED*/
10964 case GLOBAL_PAT_MOD: /* 'g' */
10965 if (RExC_pm_flags & PMf_WILDCARD) {
10966 goto modifier_illegal_in_wildcard;
10969 case ONCE_PAT_MOD: /* 'o' */
10970 if (ckWARN(WARN_REGEXP)) {
10971 const I32 wflagbit = *RExC_parse == 'o'
10974 if (! (wastedflags & wflagbit) ) {
10975 wastedflags |= wflagbit;
10976 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10979 "Useless (%s%c) - %suse /%c modifier",
10980 flagsp == &negflags ? "?-" : "?",
10982 flagsp == &negflags ? "don't " : "",
10989 case CONTINUE_PAT_MOD: /* 'c' */
10990 if (RExC_pm_flags & PMf_WILDCARD) {
10991 goto modifier_illegal_in_wildcard;
10993 if (ckWARN(WARN_REGEXP)) {
10994 if (! (wastedflags & WASTED_C) ) {
10995 wastedflags |= WASTED_GC;
10996 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10999 "Useless (%sc) - %suse /gc modifier",
11000 flagsp == &negflags ? "?-" : "?",
11001 flagsp == &negflags ? "don't " : ""
11006 case KEEPCOPY_PAT_MOD: /* 'p' */
11007 if (RExC_pm_flags & PMf_WILDCARD) {
11008 goto modifier_illegal_in_wildcard;
11010 if (flagsp == &negflags) {
11011 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
11013 *flagsp |= RXf_PMf_KEEPCOPY;
11017 /* A flag is a default iff it is following a minus, so
11018 * if there is a minus, it means will be trying to
11019 * re-specify a default which is an error */
11020 if (has_use_defaults || flagsp == &negflags) {
11021 goto fail_modifiers;
11023 flagsp = &negflags;
11024 wastedflags = 0; /* reset so (?g-c) warns twice */
11030 if ( (RExC_pm_flags & PMf_WILDCARD)
11031 && cs != REGEX_ASCII_MORE_RESTRICTED_CHARSET)
11034 /* diag_listed_as: Use of %s is not allowed in Unicode
11035 property wildcard subpatterns in regex; marked by <--
11037 vFAIL2("Use of modifier '%c' is not allowed in Unicode"
11038 " property wildcard subpatterns",
11039 has_charset_modifier);
11042 if ((posflags & (RXf_PMf_EXTENDED|RXf_PMf_EXTENDED_MORE)) == RXf_PMf_EXTENDED) {
11043 negflags |= RXf_PMf_EXTENDED_MORE;
11045 RExC_flags |= posflags;
11047 if (negflags & RXf_PMf_EXTENDED) {
11048 negflags |= RXf_PMf_EXTENDED_MORE;
11050 RExC_flags &= ~negflags;
11051 set_regex_charset(&RExC_flags, cs);
11056 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
11057 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11058 vFAIL2utf8f("Sequence (%" UTF8f "...) not recognized",
11059 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
11060 NOT_REACHED; /*NOTREACHED*/
11063 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11066 vFAIL("Sequence (?... not terminated");
11068 modifier_illegal_in_wildcard:
11070 /* diag_listed_as: Use of %s is not allowed in Unicode property wildcard
11071 subpatterns in regex; marked by <-- HERE in m/%s/ */
11072 vFAIL2("Use of modifier '%c' is not allowed in Unicode property wildcard"
11073 " subpatterns", *(RExC_parse - 1));
11077 - reg - regular expression, i.e. main body or parenthesized thing
11079 * Caller must absorb opening parenthesis.
11081 * Combining parenthesis handling with the base level of regular expression
11082 * is a trifle forced, but the need to tie the tails of the branches to what
11083 * follows makes it hard to avoid.
11085 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
11087 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
11089 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
11092 STATIC regnode_offset
11093 S_handle_named_backref(pTHX_ RExC_state_t *pRExC_state,
11095 char * parse_start,
11099 regnode_offset ret;
11100 char* name_start = RExC_parse;
11102 SV *sv_dat = reg_scan_name(pRExC_state, REG_RSN_RETURN_DATA);
11103 DECLARE_AND_GET_RE_DEBUG_FLAGS;
11105 PERL_ARGS_ASSERT_HANDLE_NAMED_BACKREF;
11107 if (RExC_parse != name_start && ch == '}') {
11108 while (isBLANK(*RExC_parse)) {
11112 if (RExC_parse == name_start || *RExC_parse != ch) {
11113 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
11114 vFAIL2("Sequence %.3s... not terminated", parse_start);
11118 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11119 RExC_rxi->data->data[num]=(void*)sv_dat;
11120 SvREFCNT_inc_simple_void_NN(sv_dat);
11123 ret = reganode(pRExC_state,
11126 : (ASCII_FOLD_RESTRICTED)
11128 : (AT_LEAST_UNI_SEMANTICS)
11134 *flagp |= HASWIDTH;
11136 Set_Node_Offset(REGNODE_p(ret), parse_start+1);
11137 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
11139 nextchar(pRExC_state);
11143 /* On success, returns the offset at which any next node should be placed into
11144 * the regex engine program being compiled.
11146 * Returns 0 otherwise, with *flagp set to indicate why:
11147 * TRYAGAIN at the end of (?) that only sets flags.
11148 * RESTART_PARSE if the parse needs to be restarted, or'd with
11149 * NEED_UTF8 if the pattern needs to be upgraded to UTF-8.
11150 * Otherwise would only return 0 if regbranch() returns 0, which cannot
11152 STATIC regnode_offset
11153 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp, U32 depth)
11154 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
11155 * 2 is like 1, but indicates that nextchar() has been called to advance
11156 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
11157 * this flag alerts us to the need to check for that */
11159 regnode_offset ret = 0; /* Will be the head of the group. */
11161 regnode_offset lastbr;
11162 regnode_offset ender = 0;
11165 U32 oregflags = RExC_flags;
11166 bool have_branch = 0;
11168 I32 freeze_paren = 0;
11169 I32 after_freeze = 0;
11170 I32 num; /* numeric backreferences */
11171 SV * max_open; /* Max number of unclosed parens */
11172 I32 was_in_lookaround = RExC_in_lookaround;
11174 char * parse_start = RExC_parse; /* MJD */
11175 char * const oregcomp_parse = RExC_parse;
11177 DECLARE_AND_GET_RE_DEBUG_FLAGS;
11179 PERL_ARGS_ASSERT_REG;
11180 DEBUG_PARSE("reg ");
11182 max_open = get_sv(RE_COMPILE_RECURSION_LIMIT, GV_ADD);
11184 if (!SvIOK(max_open)) {
11185 sv_setiv(max_open, RE_COMPILE_RECURSION_INIT);
11187 if (depth > 4 * (UV) SvIV(max_open)) { /* We increase depth by 4 for each
11189 vFAIL("Too many nested open parens");
11192 *flagp = 0; /* Initialize. */
11194 /* Having this true makes it feasible to have a lot fewer tests for the
11195 * parse pointer being in scope. For example, we can write
11196 * while(isFOO(*RExC_parse)) RExC_parse++;
11198 * while(RExC_parse < RExC_end && isFOO(*RExC_parse)) RExC_parse++;
11200 assert(*RExC_end == '\0');
11202 /* Make an OPEN node, if parenthesized. */
11205 /* Under /x, space and comments can be gobbled up between the '(' and
11206 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
11207 * intervening space, as the sequence is a token, and a token should be
11209 bool has_intervening_patws = (paren == 2)
11210 && *(RExC_parse - 1) != '(';
11212 if (RExC_parse >= RExC_end) {
11213 vFAIL("Unmatched (");
11216 if (paren == 'r') { /* Atomic script run */
11220 else if ( *RExC_parse == '*') { /* (*VERB:ARG), (*construct:...) */
11221 char *start_verb = RExC_parse + 1;
11223 char *start_arg = NULL;
11224 unsigned char op = 0;
11225 int arg_required = 0;
11226 int internal_argval = -1; /* if >-1 we are not allowed an argument*/
11227 bool has_upper = FALSE;
11229 if (has_intervening_patws) {
11230 RExC_parse++; /* past the '*' */
11232 /* For strict backwards compatibility, don't change the message
11233 * now that we also have lowercase operands */
11234 if (isUPPER(*RExC_parse)) {
11235 vFAIL("In '(*VERB...)', the '(' and '*' must be adjacent");
11238 vFAIL("In '(*...)', the '(' and '*' must be adjacent");
11241 while (RExC_parse < RExC_end && *RExC_parse != ')' ) {
11242 if ( *RExC_parse == ':' ) {
11243 start_arg = RExC_parse + 1;
11247 if (isUPPER(*RExC_parse)) {
11253 RExC_parse += UTF8SKIP(RExC_parse);
11256 verb_len = RExC_parse - start_verb;
11258 if (RExC_parse >= RExC_end) {
11259 goto unterminated_verb_pattern;
11262 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11263 while ( RExC_parse < RExC_end && *RExC_parse != ')' ) {
11264 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11266 if ( RExC_parse >= RExC_end || *RExC_parse != ')' ) {
11267 unterminated_verb_pattern:
11269 vFAIL("Unterminated verb pattern argument");
11272 vFAIL("Unterminated '(*...' argument");
11276 if ( RExC_parse >= RExC_end || *RExC_parse != ')' ) {
11278 vFAIL("Unterminated verb pattern");
11281 vFAIL("Unterminated '(*...' construct");
11286 /* Here, we know that RExC_parse < RExC_end */
11288 switch ( *start_verb ) {
11289 case 'A': /* (*ACCEPT) */
11290 if ( memEQs(start_verb, verb_len,"ACCEPT") ) {
11292 internal_argval = RExC_nestroot;
11295 case 'C': /* (*COMMIT) */
11296 if ( memEQs(start_verb, verb_len,"COMMIT") )
11299 case 'F': /* (*FAIL) */
11300 if ( verb_len==1 || memEQs(start_verb, verb_len,"FAIL") ) {
11304 case ':': /* (*:NAME) */
11305 case 'M': /* (*MARK:NAME) */
11306 if ( verb_len==0 || memEQs(start_verb, verb_len,"MARK") ) {
11311 case 'P': /* (*PRUNE) */
11312 if ( memEQs(start_verb, verb_len,"PRUNE") )
11315 case 'S': /* (*SKIP) */
11316 if ( memEQs(start_verb, verb_len,"SKIP") )
11319 case 'T': /* (*THEN) */
11320 /* [19:06] <TimToady> :: is then */
11321 if ( memEQs(start_verb, verb_len,"THEN") ) {
11323 RExC_seen |= REG_CUTGROUP_SEEN;
11327 if ( memEQs(start_verb, verb_len, "asr")
11328 || memEQs(start_verb, verb_len, "atomic_script_run"))
11330 paren = 'r'; /* Mnemonic: recursed run */
11333 else if (memEQs(start_verb, verb_len, "atomic")) {
11334 paren = 't'; /* AtOMIC */
11335 goto alpha_assertions;
11339 if ( memEQs(start_verb, verb_len, "plb")
11340 || memEQs(start_verb, verb_len, "positive_lookbehind"))
11343 goto lookbehind_alpha_assertions;
11345 else if ( memEQs(start_verb, verb_len, "pla")
11346 || memEQs(start_verb, verb_len, "positive_lookahead"))
11349 goto alpha_assertions;
11353 if ( memEQs(start_verb, verb_len, "nlb")
11354 || memEQs(start_verb, verb_len, "negative_lookbehind"))
11357 goto lookbehind_alpha_assertions;
11359 else if ( memEQs(start_verb, verb_len, "nla")
11360 || memEQs(start_verb, verb_len, "negative_lookahead"))
11363 goto alpha_assertions;
11367 if ( memEQs(start_verb, verb_len, "sr")
11368 || memEQs(start_verb, verb_len, "script_run"))
11370 regnode_offset atomic;
11376 /* This indicates Unicode rules. */
11377 REQUIRE_UNI_RULES(flagp, 0);
11383 RExC_parse = start_arg;
11385 if (RExC_in_script_run) {
11387 /* Nested script runs are treated as no-ops, because
11388 * if the nested one fails, the outer one must as
11389 * well. It could fail sooner, and avoid (??{} with
11390 * side effects, but that is explicitly documented as
11391 * undefined behavior. */
11395 if (paren == 's') {
11400 /* But, the atomic part of a nested atomic script run
11401 * isn't a no-op, but can be treated just like a '(?>'
11407 if (paren == 's') {
11408 /* Here, we're starting a new regular script run */
11409 ret = reg_node(pRExC_state, SROPEN);
11410 RExC_in_script_run = 1;
11415 /* Here, we are starting an atomic script run. This is
11416 * handled by recursing to deal with the atomic portion
11417 * separately, enclosed in SROPEN ... SRCLOSE nodes */
11419 ret = reg_node(pRExC_state, SROPEN);
11421 RExC_in_script_run = 1;
11423 atomic = reg(pRExC_state, 'r', &flags, depth);
11424 if (flags & (RESTART_PARSE|NEED_UTF8)) {
11425 *flagp = flags & (RESTART_PARSE|NEED_UTF8);
11429 if (! REGTAIL(pRExC_state, ret, atomic)) {
11430 REQUIRE_BRANCHJ(flagp, 0);
11433 if (! REGTAIL(pRExC_state, atomic, reg_node(pRExC_state,
11436 REQUIRE_BRANCHJ(flagp, 0);
11439 RExC_in_script_run = 0;
11445 lookbehind_alpha_assertions:
11446 RExC_seen |= REG_LOOKBEHIND_SEEN;
11451 RExC_in_lookaround++;
11452 RExC_seen_zerolen++;
11458 /* An empty negative lookahead assertion simply is failure */
11459 if (paren == 'A' && RExC_parse == start_arg) {
11460 ret=reganode(pRExC_state, OPFAIL, 0);
11461 nextchar(pRExC_state);
11465 RExC_parse = start_arg;
11470 "'(*%" UTF8f "' requires a terminating ':'",
11471 UTF8fARG(UTF, verb_len, start_verb));
11472 NOT_REACHED; /*NOTREACHED*/
11474 } /* End of switch */
11477 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
11479 if (has_upper || verb_len == 0) {
11481 "Unknown verb pattern '%" UTF8f "'",
11482 UTF8fARG(UTF, verb_len, start_verb));
11486 "Unknown '(*...)' construct '%" UTF8f "'",
11487 UTF8fARG(UTF, verb_len, start_verb));
11490 if ( RExC_parse == start_arg ) {
11493 if ( arg_required && !start_arg ) {
11494 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
11495 (int) verb_len, start_verb);
11497 if (internal_argval == -1) {
11498 ret = reganode(pRExC_state, op, 0);
11500 ret = reg2Lanode(pRExC_state, op, 0, internal_argval);
11502 RExC_seen |= REG_VERBARG_SEEN;
11504 SV *sv = newSVpvn( start_arg,
11505 RExC_parse - start_arg);
11506 ARG(REGNODE_p(ret)) = add_data( pRExC_state,
11507 STR_WITH_LEN("S"));
11508 RExC_rxi->data->data[ARG(REGNODE_p(ret))]=(void*)sv;
11509 FLAGS(REGNODE_p(ret)) = 1;
11511 FLAGS(REGNODE_p(ret)) = 0;
11513 if ( internal_argval != -1 )
11514 ARG2L_SET(REGNODE_p(ret), internal_argval);
11515 nextchar(pRExC_state);
11518 else if (*RExC_parse == '?') { /* (?...) */
11519 bool is_logical = 0;
11520 const char * const seqstart = RExC_parse;
11521 const char * endptr;
11522 const char non_existent_group_msg[]
11523 = "Reference to nonexistent group";
11524 const char impossible_group[] = "Invalid reference to group";
11526 if (has_intervening_patws) {
11528 vFAIL("In '(?...)', the '(' and '?' must be adjacent");
11531 RExC_parse++; /* past the '?' */
11532 paren = *RExC_parse; /* might be a trailing NUL, if not
11534 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11535 if (RExC_parse > RExC_end) {
11538 ret = 0; /* For look-ahead/behind. */
11541 case 'P': /* (?P...) variants for those used to PCRE/Python */
11542 paren = *RExC_parse;
11543 if ( paren == '<') { /* (?P<...>) named capture */
11545 if (RExC_parse >= RExC_end) {
11546 vFAIL("Sequence (?P<... not terminated");
11548 goto named_capture;
11550 else if (paren == '>') { /* (?P>name) named recursion */
11552 if (RExC_parse >= RExC_end) {
11553 vFAIL("Sequence (?P>... not terminated");
11555 goto named_recursion;
11557 else if (paren == '=') { /* (?P=...) named backref */
11559 return handle_named_backref(pRExC_state, flagp,
11562 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
11563 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11564 vFAIL3("Sequence (%.*s...) not recognized",
11565 (int) (RExC_parse - seqstart), seqstart);
11566 NOT_REACHED; /*NOTREACHED*/
11567 case '<': /* (?<...) */
11568 /* If you want to support (?<*...), first reconcile with GH #17363 */
11569 if (*RExC_parse == '!')
11571 else if (*RExC_parse != '=')
11578 case '\'': /* (?'...') */
11579 name_start = RExC_parse;
11580 svname = reg_scan_name(pRExC_state, REG_RSN_RETURN_NAME);
11581 if ( RExC_parse == name_start
11582 || RExC_parse >= RExC_end
11583 || *RExC_parse != paren)
11585 vFAIL2("Sequence (?%c... not terminated",
11586 paren=='>' ? '<' : (char) paren);
11591 if (!svname) /* shouldn't happen */
11593 "panic: reg_scan_name returned NULL");
11594 if (!RExC_paren_names) {
11595 RExC_paren_names= newHV();
11596 sv_2mortal(MUTABLE_SV(RExC_paren_names));
11598 RExC_paren_name_list= newAV();
11599 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
11602 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
11604 sv_dat = HeVAL(he_str);
11606 /* croak baby croak */
11608 "panic: paren_name hash element allocation failed");
11609 } else if ( SvPOK(sv_dat) ) {
11610 /* (?|...) can mean we have dupes so scan to check
11611 its already been stored. Maybe a flag indicating
11612 we are inside such a construct would be useful,
11613 but the arrays are likely to be quite small, so
11614 for now we punt -- dmq */
11615 IV count = SvIV(sv_dat);
11616 I32 *pv = (I32*)SvPVX(sv_dat);
11618 for ( i = 0 ; i < count ; i++ ) {
11619 if ( pv[i] == RExC_npar ) {
11625 pv = (I32*)SvGROW(sv_dat,
11626 SvCUR(sv_dat) + sizeof(I32)+1);
11627 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
11628 pv[count] = RExC_npar;
11629 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
11632 (void)SvUPGRADE(sv_dat, SVt_PVNV);
11633 sv_setpvn(sv_dat, (char *)&(RExC_npar),
11636 SvIV_set(sv_dat, 1);
11639 /* Yes this does cause a memory leak in debugging Perls
11641 if (!av_store(RExC_paren_name_list,
11642 RExC_npar, SvREFCNT_inc_NN(svname)))
11643 SvREFCNT_dec_NN(svname);
11646 /*sv_dump(sv_dat);*/
11648 nextchar(pRExC_state);
11650 goto capturing_parens;
11653 RExC_seen |= REG_LOOKBEHIND_SEEN;
11654 RExC_in_lookaround++;
11656 if (RExC_parse >= RExC_end) {
11657 vFAIL("Sequence (?... not terminated");
11659 RExC_seen_zerolen++;
11661 case '=': /* (?=...) */
11662 RExC_seen_zerolen++;
11663 RExC_in_lookaround++;
11665 case '!': /* (?!...) */
11666 RExC_seen_zerolen++;
11667 /* check if we're really just a "FAIL" assertion */
11668 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
11669 FALSE /* Don't force to /x */ );
11670 if (*RExC_parse == ')') {
11671 ret=reganode(pRExC_state, OPFAIL, 0);
11672 nextchar(pRExC_state);
11675 RExC_in_lookaround++;
11677 case '|': /* (?|...) */
11678 /* branch reset, behave like a (?:...) except that
11679 buffers in alternations share the same numbers */
11681 after_freeze = freeze_paren = RExC_npar;
11683 /* XXX This construct currently requires an extra pass.
11684 * Investigation would be required to see if that could be
11686 REQUIRE_PARENS_PASS;
11688 case ':': /* (?:...) */
11689 case '>': /* (?>...) */
11691 case '$': /* (?$...) */
11692 case '@': /* (?@...) */
11693 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
11695 case '0' : /* (?0) */
11696 case 'R' : /* (?R) */
11697 if (RExC_parse == RExC_end || *RExC_parse != ')')
11698 FAIL("Sequence (?R) not terminated");
11700 RExC_seen |= REG_RECURSE_SEEN;
11702 /* XXX These constructs currently require an extra pass.
11703 * It probably could be changed */
11704 REQUIRE_PARENS_PASS;
11706 *flagp |= POSTPONED;
11707 goto gen_recurse_regop;
11709 /* named and numeric backreferences */
11710 case '&': /* (?&NAME) */
11711 parse_start = RExC_parse - 1;
11714 SV *sv_dat = reg_scan_name(pRExC_state,
11715 REG_RSN_RETURN_DATA);
11716 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
11718 if (RExC_parse >= RExC_end || *RExC_parse != ')')
11719 vFAIL("Sequence (?&... not terminated");
11720 goto gen_recurse_regop;
11723 if (! inRANGE(RExC_parse[0], '1', '9')) {
11725 vFAIL("Illegal pattern");
11727 goto parse_recursion;
11729 case '-': /* (?-1) */
11730 if (! inRANGE(RExC_parse[0], '1', '9')) {
11731 RExC_parse--; /* rewind to let it be handled later */
11735 case '1': case '2': case '3': case '4': /* (?1) */
11736 case '5': case '6': case '7': case '8': case '9':
11737 RExC_parse = (char *) seqstart + 1; /* Point to the digit */
11740 bool is_neg = FALSE;
11742 parse_start = RExC_parse - 1; /* MJD */
11743 if (*RExC_parse == '-') {
11748 if (grok_atoUV(RExC_parse, &unum, &endptr)
11752 RExC_parse = (char*)endptr;
11754 else { /* Overflow, or something like that. Position
11755 beyond all digits for the message */
11756 while (RExC_parse < RExC_end && isDIGIT(*RExC_parse)) {
11759 vFAIL(impossible_group);
11762 /* -num is always representable on 1 and 2's complement
11767 if (*RExC_parse!=')')
11768 vFAIL("Expecting close bracket");
11771 if (paren == '-' || paren == '+') {
11773 /* Don't overflow */
11774 if (UNLIKELY(I32_MAX - RExC_npar < num)) {
11776 vFAIL(impossible_group);
11780 Diagram of capture buffer numbering.
11781 Top line is the normal capture buffer numbers
11782 Bottom line is the negative indexing as from
11786 /(a(x)y)(a(b(c(?+2)d)e)f)(g(h))/
11787 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
11788 - 5 4 3 2 1 X Y x x
11790 Resolve to absolute group. Recall that RExC_npar is +1 of
11791 the actual parenthesis group number. For lookahead, we
11792 have to compensate for that. Using the above example, when
11793 we get to Y in the parse, num is 2 and RExC_npar is 6. We
11794 want 7 for +2, and 4 for -2.
11796 if ( paren == '+' ) {
11802 if (paren == '-' && num < 1) {
11804 vFAIL(non_existent_group_msg);
11808 if (num >= RExC_npar) {
11810 /* It might be a forward reference; we can't fail until we
11811 * know, by completing the parse to get all the groups, and
11812 * then reparsing */
11813 if (ALL_PARENS_COUNTED) {
11814 if (num >= RExC_total_parens) {
11816 vFAIL(non_existent_group_msg);
11820 REQUIRE_PARENS_PASS;
11824 /* We keep track how many GOSUB items we have produced.
11825 To start off the ARG2L() of the GOSUB holds its "id",
11826 which is used later in conjunction with RExC_recurse
11827 to calculate the offset we need to jump for the GOSUB,
11828 which it will store in the final representation.
11829 We have to defer the actual calculation until much later
11830 as the regop may move.
11832 ret = reg2Lanode(pRExC_state, GOSUB, num, RExC_recurse_count);
11833 RExC_recurse_count++;
11834 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11835 "%*s%*s Recurse #%" UVuf " to %" IVdf "\n",
11836 22, "| |", (int)(depth * 2 + 1), "",
11837 (UV)ARG(REGNODE_p(ret)),
11838 (IV)ARG2L(REGNODE_p(ret))));
11839 RExC_seen |= REG_RECURSE_SEEN;
11841 Set_Node_Length(REGNODE_p(ret),
11842 1 + regarglen[OP(REGNODE_p(ret))]); /* MJD */
11843 Set_Node_Offset(REGNODE_p(ret), parse_start); /* MJD */
11845 *flagp |= POSTPONED;
11846 assert(*RExC_parse == ')');
11847 nextchar(pRExC_state);
11852 case '?': /* (??...) */
11854 if (*RExC_parse != '{') {
11855 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
11856 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11858 "Sequence (%" UTF8f "...) not recognized",
11859 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
11860 NOT_REACHED; /*NOTREACHED*/
11862 *flagp |= POSTPONED;
11866 case '{': /* (?{...}) */
11869 struct reg_code_block *cb;
11872 RExC_seen_zerolen++;
11874 if ( !pRExC_state->code_blocks
11875 || pRExC_state->code_index
11876 >= pRExC_state->code_blocks->count
11877 || pRExC_state->code_blocks->cb[pRExC_state->code_index].start
11878 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
11881 if (RExC_pm_flags & PMf_USE_RE_EVAL)
11882 FAIL("panic: Sequence (?{...}): no code block found\n");
11883 FAIL("Eval-group not allowed at runtime, use re 'eval'");
11885 /* this is a pre-compiled code block (?{...}) */
11886 cb = &pRExC_state->code_blocks->cb[pRExC_state->code_index];
11887 RExC_parse = RExC_start + cb->end;
11889 if (cb->src_regex) {
11890 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
11891 RExC_rxi->data->data[n] =
11892 (void*)SvREFCNT_inc((SV*)cb->src_regex);
11893 RExC_rxi->data->data[n+1] = (void*)o;
11896 n = add_data(pRExC_state,
11897 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
11898 RExC_rxi->data->data[n] = (void*)o;
11900 pRExC_state->code_index++;
11901 nextchar(pRExC_state);
11904 regnode_offset eval;
11905 ret = reg_node(pRExC_state, LOGICAL);
11907 eval = reg2Lanode(pRExC_state, EVAL,
11910 /* for later propagation into (??{})
11912 RExC_flags & RXf_PMf_COMPILETIME
11914 FLAGS(REGNODE_p(ret)) = 2;
11915 if (! REGTAIL(pRExC_state, ret, eval)) {
11916 REQUIRE_BRANCHJ(flagp, 0);
11918 /* deal with the length of this later - MJD */
11921 ret = reg2Lanode(pRExC_state, EVAL, n, 0);
11922 Set_Node_Length(REGNODE_p(ret), RExC_parse - parse_start + 1);
11923 Set_Node_Offset(REGNODE_p(ret), parse_start);
11926 case '(': /* (?(?{...})...) and (?(?=...)...) */
11929 const int DEFINE_len = sizeof("DEFINE") - 1;
11930 if ( RExC_parse < RExC_end - 1
11931 && ( ( RExC_parse[0] == '?' /* (?(?...)) */
11932 && ( RExC_parse[1] == '='
11933 || RExC_parse[1] == '!'
11934 || RExC_parse[1] == '<'
11935 || RExC_parse[1] == '{'))
11936 || ( RExC_parse[0] == '*' /* (?(*...)) */
11937 && ( memBEGINs(RExC_parse + 1,
11938 (Size_t) (RExC_end - (RExC_parse + 1)),
11940 || memBEGINs(RExC_parse + 1,
11941 (Size_t) (RExC_end - (RExC_parse + 1)),
11943 || memBEGINs(RExC_parse + 1,
11944 (Size_t) (RExC_end - (RExC_parse + 1)),
11946 || memBEGINs(RExC_parse + 1,
11947 (Size_t) (RExC_end - (RExC_parse + 1)),
11949 || memBEGINs(RExC_parse + 1,
11950 (Size_t) (RExC_end - (RExC_parse + 1)),
11951 "positive_lookahead:")
11952 || memBEGINs(RExC_parse + 1,
11953 (Size_t) (RExC_end - (RExC_parse + 1)),
11954 "positive_lookbehind:")
11955 || memBEGINs(RExC_parse + 1,
11956 (Size_t) (RExC_end - (RExC_parse + 1)),
11957 "negative_lookahead:")
11958 || memBEGINs(RExC_parse + 1,
11959 (Size_t) (RExC_end - (RExC_parse + 1)),
11960 "negative_lookbehind:"))))
11961 ) { /* Lookahead or eval. */
11963 regnode_offset tail;
11965 ret = reg_node(pRExC_state, LOGICAL);
11966 FLAGS(REGNODE_p(ret)) = 1;
11968 tail = reg(pRExC_state, 1, &flag, depth+1);
11969 RETURN_FAIL_ON_RESTART(flag, flagp);
11970 if (! REGTAIL(pRExC_state, ret, tail)) {
11971 REQUIRE_BRANCHJ(flagp, 0);
11975 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
11976 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
11978 char ch = RExC_parse[0] == '<' ? '>' : '\'';
11979 char *name_start= RExC_parse++;
11981 SV *sv_dat=reg_scan_name(pRExC_state, REG_RSN_RETURN_DATA);
11982 if ( RExC_parse == name_start
11983 || RExC_parse >= RExC_end
11984 || *RExC_parse != ch)
11986 vFAIL2("Sequence (?(%c... not terminated",
11987 (ch == '>' ? '<' : ch));
11991 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11992 RExC_rxi->data->data[num]=(void*)sv_dat;
11993 SvREFCNT_inc_simple_void_NN(sv_dat);
11995 ret = reganode(pRExC_state, GROUPPN, num);
11996 goto insert_if_check_paren;
11998 else if (memBEGINs(RExC_parse,
11999 (STRLEN) (RExC_end - RExC_parse),
12002 ret = reganode(pRExC_state, DEFINEP, 0);
12003 RExC_parse += DEFINE_len;
12005 goto insert_if_check_paren;
12007 else if (RExC_parse[0] == 'R') {
12009 /* parno == 0 => /(?(R)YES|NO)/ "in any form of recursion OR eval"
12010 * parno == 1 => /(?(R0)YES|NO)/ "in GOSUB (?0) / (?R)"
12011 * parno == 2 => /(?(R1)YES|NO)/ "in GOSUB (?1) (parno-1)"
12014 if (RExC_parse[0] == '0') {
12018 else if (inRANGE(RExC_parse[0], '1', '9')) {
12021 if (grok_atoUV(RExC_parse, &uv, &endptr)
12024 parno = (I32)uv + 1;
12025 RExC_parse = (char*)endptr;
12027 /* else "Switch condition not recognized" below */
12028 } else if (RExC_parse[0] == '&') {
12031 sv_dat = reg_scan_name(pRExC_state,
12032 REG_RSN_RETURN_DATA);
12034 parno = 1 + *((I32 *)SvPVX(sv_dat));
12036 ret = reganode(pRExC_state, INSUBP, parno);
12037 goto insert_if_check_paren;
12039 else if (inRANGE(RExC_parse[0], '1', '9')) {
12044 if (grok_atoUV(RExC_parse, &uv, &endptr)
12048 RExC_parse = (char*)endptr;
12051 vFAIL("panic: grok_atoUV returned FALSE");
12053 ret = reganode(pRExC_state, GROUPP, parno);
12055 insert_if_check_paren:
12056 if (UCHARAT(RExC_parse) != ')') {
12058 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
12060 vFAIL("Switch condition not recognized");
12062 nextchar(pRExC_state);
12064 if (! REGTAIL(pRExC_state, ret, reganode(pRExC_state,
12067 REQUIRE_BRANCHJ(flagp, 0);
12069 br = regbranch(pRExC_state, &flags, 1, depth+1);
12071 RETURN_FAIL_ON_RESTART(flags,flagp);
12072 FAIL2("panic: regbranch returned failure, flags=%#" UVxf,
12075 if (! REGTAIL(pRExC_state, br, reganode(pRExC_state,
12078 REQUIRE_BRANCHJ(flagp, 0);
12080 c = UCHARAT(RExC_parse);
12081 nextchar(pRExC_state);
12082 if (flags&HASWIDTH)
12083 *flagp |= HASWIDTH;
12086 vFAIL("(?(DEFINE)....) does not allow branches");
12088 /* Fake one for optimizer. */
12089 lastbr = reganode(pRExC_state, IFTHEN, 0);
12091 if (!regbranch(pRExC_state, &flags, 1, depth+1)) {
12092 RETURN_FAIL_ON_RESTART(flags, flagp);
12093 FAIL2("panic: regbranch returned failure, flags=%#" UVxf,
12096 if (! REGTAIL(pRExC_state, ret, lastbr)) {
12097 REQUIRE_BRANCHJ(flagp, 0);
12099 if (flags&HASWIDTH)
12100 *flagp |= HASWIDTH;
12101 c = UCHARAT(RExC_parse);
12102 nextchar(pRExC_state);
12107 if (RExC_parse >= RExC_end)
12108 vFAIL("Switch (?(condition)... not terminated");
12110 vFAIL("Switch (?(condition)... contains too many branches");
12112 ender = reg_node(pRExC_state, TAIL);
12113 if (! REGTAIL(pRExC_state, br, ender)) {
12114 REQUIRE_BRANCHJ(flagp, 0);
12117 if (! REGTAIL(pRExC_state, lastbr, ender)) {
12118 REQUIRE_BRANCHJ(flagp, 0);
12120 if (! REGTAIL(pRExC_state,
12123 NEXTOPER(REGNODE_p(lastbr)))),
12126 REQUIRE_BRANCHJ(flagp, 0);
12130 if (! REGTAIL(pRExC_state, ret, ender)) {
12131 REQUIRE_BRANCHJ(flagp, 0);
12133 #if 0 /* Removing this doesn't cause failures in the test suite -- khw */
12134 RExC_size++; /* XXX WHY do we need this?!!
12135 For large programs it seems to be required
12136 but I can't figure out why. -- dmq*/
12141 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
12143 vFAIL("Unknown switch condition (?(...))");
12145 case '[': /* (?[ ... ]) */
12146 return handle_regex_sets(pRExC_state, NULL, flagp, depth+1,
12148 case 0: /* A NUL */
12149 RExC_parse--; /* for vFAIL to print correctly */
12150 vFAIL("Sequence (? incomplete");
12154 if (RExC_strict) { /* [perl #132851] */
12155 ckWARNreg(RExC_parse, "Empty (?) without any modifiers");
12158 case '*': /* If you want to support (?*...), first reconcile with GH #17363 */
12160 default: /* e.g., (?i) */
12161 RExC_parse = (char *) seqstart + 1;
12163 parse_lparen_question_flags(pRExC_state);
12164 if (UCHARAT(RExC_parse) != ':') {
12165 if (RExC_parse < RExC_end)
12166 nextchar(pRExC_state);
12171 nextchar(pRExC_state);
12176 else if (!(RExC_flags & RXf_PMf_NOCAPTURE)) { /* (...) */
12180 if (! ALL_PARENS_COUNTED) {
12181 /* If we are in our first pass through (and maybe only pass),
12182 * we need to allocate memory for the capturing parentheses
12186 if (!RExC_parens_buf_size) {
12187 /* first guess at number of parens we might encounter */
12188 RExC_parens_buf_size = 10;
12190 /* setup RExC_open_parens, which holds the address of each
12191 * OPEN tag, and to make things simpler for the 0 index the
12192 * start of the program - this is used later for offsets */
12193 Newxz(RExC_open_parens, RExC_parens_buf_size,
12195 RExC_open_parens[0] = 1; /* +1 for REG_MAGIC */
12197 /* setup RExC_close_parens, which holds the address of each
12198 * CLOSE tag, and to make things simpler for the 0 index
12199 * the end of the program - this is used later for offsets
12201 Newxz(RExC_close_parens, RExC_parens_buf_size,
12203 /* we dont know where end op starts yet, so we dont need to
12204 * set RExC_close_parens[0] like we do RExC_open_parens[0]
12207 else if (RExC_npar > RExC_parens_buf_size) {
12208 I32 old_size = RExC_parens_buf_size;
12210 RExC_parens_buf_size *= 2;
12212 Renew(RExC_open_parens, RExC_parens_buf_size,
12214 Zero(RExC_open_parens + old_size,
12215 RExC_parens_buf_size - old_size, regnode_offset);
12217 Renew(RExC_close_parens, RExC_parens_buf_size,
12219 Zero(RExC_close_parens + old_size,
12220 RExC_parens_buf_size - old_size, regnode_offset);
12224 ret = reganode(pRExC_state, OPEN, parno);
12225 if (!RExC_nestroot)
12226 RExC_nestroot = parno;
12227 if (RExC_open_parens && !RExC_open_parens[parno])
12229 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12230 "%*s%*s Setting open paren #%" IVdf " to %zu\n",
12231 22, "| |", (int)(depth * 2 + 1), "",
12233 RExC_open_parens[parno]= ret;
12236 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
12237 Set_Node_Offset(REGNODE_p(ret), RExC_parse); /* MJD */
12240 /* with RXf_PMf_NOCAPTURE treat (...) as (?:...) */
12249 /* Pick up the branches, linking them together. */
12250 parse_start = RExC_parse; /* MJD */
12251 br = regbranch(pRExC_state, &flags, 1, depth+1);
12253 /* branch_len = (paren != 0); */
12256 RETURN_FAIL_ON_RESTART(flags, flagp);
12257 FAIL2("panic: regbranch returned failure, flags=%#" UVxf, (UV) flags);
12259 if (*RExC_parse == '|') {
12260 if (RExC_use_BRANCHJ) {
12261 reginsert(pRExC_state, BRANCHJ, br, depth+1);
12264 reginsert(pRExC_state, BRANCH, br, depth+1);
12265 Set_Node_Length(REGNODE_p(br), paren != 0);
12266 Set_Node_Offset_To_R(br, parse_start-RExC_start);
12270 else if (paren == ':') {
12271 *flagp |= flags&SIMPLE;
12273 if (is_open) { /* Starts with OPEN. */
12274 if (! REGTAIL(pRExC_state, ret, br)) { /* OPEN -> first. */
12275 REQUIRE_BRANCHJ(flagp, 0);
12278 else if (paren != '?') /* Not Conditional */
12280 *flagp |= flags & (HASWIDTH | POSTPONED);
12282 while (*RExC_parse == '|') {
12283 if (RExC_use_BRANCHJ) {
12286 ender = reganode(pRExC_state, LONGJMP, 0);
12288 /* Append to the previous. */
12289 shut_gcc_up = REGTAIL(pRExC_state,
12290 REGNODE_OFFSET(NEXTOPER(NEXTOPER(REGNODE_p(lastbr)))),
12292 PERL_UNUSED_VAR(shut_gcc_up);
12294 nextchar(pRExC_state);
12295 if (freeze_paren) {
12296 if (RExC_npar > after_freeze)
12297 after_freeze = RExC_npar;
12298 RExC_npar = freeze_paren;
12300 br = regbranch(pRExC_state, &flags, 0, depth+1);
12303 RETURN_FAIL_ON_RESTART(flags, flagp);
12304 FAIL2("panic: regbranch returned failure, flags=%#" UVxf, (UV) flags);
12306 if (! REGTAIL(pRExC_state, lastbr, br)) { /* BRANCH -> BRANCH. */
12307 REQUIRE_BRANCHJ(flagp, 0);
12310 *flagp |= flags & (HASWIDTH | POSTPONED);
12313 if (have_branch || paren != ':') {
12316 /* Make a closing node, and hook it on the end. */
12319 ender = reg_node(pRExC_state, TAIL);
12322 ender = reganode(pRExC_state, CLOSE, parno);
12323 if ( RExC_close_parens ) {
12324 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12325 "%*s%*s Setting close paren #%" IVdf " to %zu\n",
12326 22, "| |", (int)(depth * 2 + 1), "",
12327 (IV)parno, ender));
12328 RExC_close_parens[parno]= ender;
12329 if (RExC_nestroot == parno)
12332 Set_Node_Offset(REGNODE_p(ender), RExC_parse+1); /* MJD */
12333 Set_Node_Length(REGNODE_p(ender), 1); /* MJD */
12336 ender = reg_node(pRExC_state, SRCLOSE);
12337 RExC_in_script_run = 0;
12347 *flagp &= ~HASWIDTH;
12349 case 't': /* aTomic */
12351 ender = reg_node(pRExC_state, SUCCEED);
12354 ender = reg_node(pRExC_state, END);
12355 assert(!RExC_end_op); /* there can only be one! */
12356 RExC_end_op = REGNODE_p(ender);
12357 if (RExC_close_parens) {
12358 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12359 "%*s%*s Setting close paren #0 (END) to %zu\n",
12360 22, "| |", (int)(depth * 2 + 1), "",
12363 RExC_close_parens[0]= ender;
12368 DEBUG_PARSE_MSG("lsbr");
12369 regprop(RExC_rx, RExC_mysv1, REGNODE_p(lastbr), NULL, pRExC_state);
12370 regprop(RExC_rx, RExC_mysv2, REGNODE_p(ender), NULL, pRExC_state);
12371 Perl_re_printf( aTHX_ "~ tying lastbr %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
12372 SvPV_nolen_const(RExC_mysv1),
12374 SvPV_nolen_const(RExC_mysv2),
12376 (IV)(ender - lastbr)
12379 if (! REGTAIL(pRExC_state, lastbr, ender)) {
12380 REQUIRE_BRANCHJ(flagp, 0);
12384 char is_nothing= 1;
12386 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
12388 /* Hook the tails of the branches to the closing node. */
12389 for (br = REGNODE_p(ret); br; br = regnext(br)) {
12390 const U8 op = PL_regkind[OP(br)];
12391 if (op == BRANCH) {
12392 if (! REGTAIL_STUDY(pRExC_state,
12393 REGNODE_OFFSET(NEXTOPER(br)),
12396 REQUIRE_BRANCHJ(flagp, 0);
12398 if ( OP(NEXTOPER(br)) != NOTHING
12399 || regnext(NEXTOPER(br)) != REGNODE_p(ender))
12402 else if (op == BRANCHJ) {
12403 bool shut_gcc_up = REGTAIL_STUDY(pRExC_state,
12404 REGNODE_OFFSET(NEXTOPER(NEXTOPER(br))),
12406 PERL_UNUSED_VAR(shut_gcc_up);
12407 /* for now we always disable this optimisation * /
12408 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
12409 || regnext(NEXTOPER(NEXTOPER(br))) != REGNODE_p(ender))
12415 regnode * ret_as_regnode = REGNODE_p(ret);
12416 br= PL_regkind[OP(ret_as_regnode)] != BRANCH
12417 ? regnext(ret_as_regnode)
12420 DEBUG_PARSE_MSG("NADA");
12421 regprop(RExC_rx, RExC_mysv1, ret_as_regnode,
12422 NULL, pRExC_state);
12423 regprop(RExC_rx, RExC_mysv2, REGNODE_p(ender),
12424 NULL, pRExC_state);
12425 Perl_re_printf( aTHX_ "~ converting ret %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
12426 SvPV_nolen_const(RExC_mysv1),
12427 (IV)REG_NODE_NUM(ret_as_regnode),
12428 SvPV_nolen_const(RExC_mysv2),
12434 if (OP(REGNODE_p(ender)) == TAIL) {
12436 RExC_emit= REGNODE_OFFSET(br) + 1;
12439 for ( opt= br + 1; opt < REGNODE_p(ender) ; opt++ )
12440 OP(opt)= OPTIMIZED;
12441 NEXT_OFF(br)= REGNODE_p(ender) - br;
12449 /* Even/odd or x=don't care: 010101x10x */
12450 static const char parens[] = "=!aA<,>Bbt";
12451 /* flag below is set to 0 up through 'A'; 1 for larger */
12453 if (paren && (p = strchr(parens, paren))) {
12454 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
12455 int flag = (p - parens) > 3;
12457 if (paren == '>' || paren == 't') {
12458 node = SUSPEND, flag = 0;
12461 reginsert(pRExC_state, node, ret, depth+1);
12462 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
12463 Set_Node_Offset(REGNODE_p(ret), parse_start + 1);
12464 FLAGS(REGNODE_p(ret)) = flag;
12465 if (! REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL)))
12467 REQUIRE_BRANCHJ(flagp, 0);
12472 /* Check for proper termination. */
12474 /* restore original flags, but keep (?p) and, if we've encountered
12475 * something in the parse that changes /d rules into /u, keep the /u */
12476 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
12477 if (DEPENDS_SEMANTICS && toUSE_UNI_CHARSET_NOT_DEPENDS) {
12478 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
12480 if (RExC_parse >= RExC_end || UCHARAT(RExC_parse) != ')') {
12481 RExC_parse = oregcomp_parse;
12482 vFAIL("Unmatched (");
12484 nextchar(pRExC_state);
12486 else if (!paren && RExC_parse < RExC_end) {
12487 if (*RExC_parse == ')') {
12489 vFAIL("Unmatched )");
12492 FAIL("Junk on end of regexp"); /* "Can't happen". */
12493 NOT_REACHED; /* NOTREACHED */
12496 if (after_freeze > RExC_npar)
12497 RExC_npar = after_freeze;
12499 RExC_in_lookaround = was_in_lookaround;
12505 - regbranch - one alternative of an | operator
12507 * Implements the concatenation operator.
12509 * On success, returns the offset at which any next node should be placed into
12510 * the regex engine program being compiled.
12512 * Returns 0 otherwise, setting flagp to RESTART_PARSE if the parse needs
12513 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to
12516 STATIC regnode_offset
12517 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
12519 regnode_offset ret;
12520 regnode_offset chain = 0;
12521 regnode_offset latest;
12522 I32 flags = 0, c = 0;
12523 DECLARE_AND_GET_RE_DEBUG_FLAGS;
12525 PERL_ARGS_ASSERT_REGBRANCH;
12527 DEBUG_PARSE("brnc");
12532 if (RExC_use_BRANCHJ)
12533 ret = reganode(pRExC_state, BRANCHJ, 0);
12535 ret = reg_node(pRExC_state, BRANCH);
12536 Set_Node_Length(REGNODE_p(ret), 1);
12540 *flagp = 0; /* Initialize. */
12542 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
12543 FALSE /* Don't force to /x */ );
12544 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
12545 flags &= ~TRYAGAIN;
12546 latest = regpiece(pRExC_state, &flags, depth+1);
12548 if (flags & TRYAGAIN)
12550 RETURN_FAIL_ON_RESTART(flags, flagp);
12551 FAIL2("panic: regpiece returned failure, flags=%#" UVxf, (UV) flags);
12555 *flagp |= flags&(HASWIDTH|POSTPONED);
12557 /* FIXME adding one for every branch after the first is probably
12558 * excessive now we have TRIE support. (hv) */
12560 if (! REGTAIL(pRExC_state, chain, latest)) {
12561 /* XXX We could just redo this branch, but figuring out what
12562 * bookkeeping needs to be reset is a pain, and it's likely
12563 * that other branches that goto END will also be too large */
12564 REQUIRE_BRANCHJ(flagp, 0);
12570 if (chain == 0) { /* Loop ran zero times. */
12571 chain = reg_node(pRExC_state, NOTHING);
12576 *flagp |= flags&SIMPLE;
12588 #ifndef PERL_IN_XSUB_RE
12590 Perl_regcurly(const char *s, const char *e, const char * result[5])
12592 /* This function matches a {m,n} quantifier. When called with a NULL final
12593 * argument, it simply parses the input from 's' up through 'e-1', and
12594 * returns a boolean as to whether or not this input is syntactically a
12595 * {m,n} quantifier.
12597 * When called with a non-NULL final parameter, and when the function
12598 * returns TRUE, it additionally stores information into the array
12599 * specified by that parameter about what it found in the parse. The
12600 * parameter must be a pointer into a 5 element array of 'const char *'
12601 * elements. The returned information is as follows:
12602 * result[RBRACE] points to the closing brace
12603 * result[MIN_S] points to the first byte of the lower bound
12604 * result[MIN_E] points to one beyond the final byte of the lower bound
12605 * result[MAX_S] points to the first byte of the upper bound
12606 * result[MAX_E] points to one beyond the final byte of the upper bound
12608 * If the quantifier is of the form {m,} (meaning an infinite upper
12609 * bound), result[MAX_E] is set to result[MAX_S]; what they actually point
12610 * to is irrelevant, just that it's the same place
12612 * If instead the quantifier is of the form {m} there is actually only
12613 * one bound, and both the upper and lower result[] elements are set to
12616 * This function checks only for syntactic validity; it leaves checking for
12617 * semantic validity and raising any diagnostics to the caller. This
12618 * function is called in multiple places to check for syntax, but only from
12619 * one for semantics. It makes it as simple as possible for the
12620 * syntax-only callers, while furnishing just enough information for the
12624 const char * min_start = NULL;
12625 const char * max_start = NULL;
12626 const char * min_end = NULL;
12627 const char * max_end = NULL;
12629 bool has_comma = FALSE;
12631 PERL_ARGS_ASSERT_REGCURLY;
12633 if (s >= e || *s++ != '{')
12636 while (s < e && isBLANK(*s)) {
12644 } while (s < e && isDIGIT(*s));
12648 while (s < e && isBLANK(*s)) {
12656 while (s < e && isBLANK(*s)) {
12664 } while (s < e && isDIGIT(*s));
12669 while (s < e && isBLANK(*s)) {
12672 /* Need at least one number */
12673 if (s >= e || *s != '}' || (! min_start && ! max_end)) {
12679 result[RBRACE] = s;
12681 result[MIN_S] = min_start;
12682 result[MIN_E] = min_end;
12685 result[MAX_S] = max_start;
12686 result[MAX_E] = max_end;
12689 /* Having no value after the comma is signalled by setting
12690 * start and end to the same value. What that value is isn't
12691 * relevant; NULL is chosen simply because it will fail if the
12692 * caller mistakenly uses it */
12693 result[MAX_S] = result[MAX_E] = NULL;
12696 else { /* No comma means lower and upper bounds are the same */
12697 result[MAX_S] = min_start;
12698 result[MAX_E] = min_end;
12707 S_get_quantifier_value(pTHX_ RExC_state_t *pRExC_state,
12708 const char * start, const char * end)
12710 /* This is a helper function for regpiece() to compute, given the
12711 * quantifier {m,n}, the value of either m or n, based on the starting
12712 * position 'start' in the string, through the byte 'end-1', returning it
12713 * if valid, and failing appropriately if not. It knows the restrictions
12714 * imposed on quantifier values */
12717 STATIC_ASSERT_DECL(REG_INFTY <= U32_MAX);
12719 PERL_ARGS_ASSERT_GET_QUANTIFIER_VALUE;
12721 if (grok_atoUV(start, &uv, &end)) {
12722 if (uv < REG_INFTY) { /* A valid, small-enough number */
12726 else if (*start == '0') { /* grok_atoUV() fails for only two reasons:
12727 leading zeros or overflow */
12728 RExC_parse = (char * ) end;
12730 /* Perhaps too generic a msg for what is only failure from having
12731 * leading zeros, but this is how it's always behaved. */
12732 vFAIL("Invalid quantifier in {,}");
12733 NOT_REACHED; /*NOTREACHED*/
12736 /* Here, found a quantifier, but was too large; either it overflowed or was
12737 * too big a legal number */
12738 RExC_parse = (char * ) end;
12739 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
12741 NOT_REACHED; /*NOTREACHED*/
12742 return U32_MAX; /* Perhaps some compilers will be expecting a return */
12746 - regpiece - something followed by possible quantifier * + ? {n,m}
12748 * Note that the branching code sequences used for ? and the general cases
12749 * of * and + are somewhat optimized: they use the same NOTHING node as
12750 * both the endmarker for their branch list and the body of the last branch.
12751 * It might seem that this node could be dispensed with entirely, but the
12752 * endmarker role is not redundant.
12754 * On success, returns the offset at which any next node should be placed into
12755 * the regex engine program being compiled.
12757 * Returns 0 otherwise, with *flagp set to indicate why:
12758 * TRYAGAIN if regatom() returns 0 with TRYAGAIN.
12759 * RESTART_PARSE if the parse needs to be restarted, or'd with
12760 * NEED_UTF8 if the pattern needs to be upgraded to UTF-8.
12762 STATIC regnode_offset
12763 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
12765 regnode_offset ret;
12768 const char * const origparse = RExC_parse;
12770 I32 max = REG_INFTY;
12771 #ifdef RE_TRACK_PATTERN_OFFSETS
12775 /* Save the original in case we change the emitted regop to a FAIL. */
12776 const regnode_offset orig_emit = RExC_emit;
12778 DECLARE_AND_GET_RE_DEBUG_FLAGS;
12780 PERL_ARGS_ASSERT_REGPIECE;
12782 DEBUG_PARSE("piec");
12784 ret = regatom(pRExC_state, &flags, depth+1);
12786 RETURN_FAIL_ON_RESTART_OR_FLAGS(flags, flagp, TRYAGAIN);
12787 FAIL2("panic: regatom returned failure, flags=%#" UVxf, (UV) flags);
12790 #ifdef RE_TRACK_PATTERN_OFFSETS
12791 parse_start = RExC_parse;
12796 const char * regcurly_return[5];
12799 nextchar(pRExC_state);
12804 nextchar(pRExC_state);
12809 nextchar(pRExC_state);
12813 case '{': /* A '{' may or may not indicate a quantifier; call regcurly()
12814 to determine which */
12815 if (regcurly(RExC_parse, RExC_end, regcurly_return)) {
12816 const char * min_start = regcurly_return[MIN_S];
12817 const char * min_end = regcurly_return[MIN_E];
12818 const char * max_start = regcurly_return[MAX_S];
12819 const char * max_end = regcurly_return[MAX_E];
12822 min = get_quantifier_value(pRExC_state, min_start, min_end);
12828 if (max_start == max_end) { /* Was of the form {m,} */
12831 else if (max_start == min_start) { /* Was of the form {m} */
12834 else { /* Was of the form {m,n} */
12835 assert(max_end >= max_start);
12837 max = get_quantifier_value(pRExC_state, max_start, max_end);
12840 RExC_parse = (char *) regcurly_return[RBRACE];
12841 nextchar(pRExC_state);
12843 if (max < min) { /* If can't match, warn and optimize to fail
12845 reginsert(pRExC_state, OPFAIL, orig_emit, depth+1);
12846 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
12847 NEXT_OFF(REGNODE_p(orig_emit)) =
12848 regarglen[OPFAIL] + NODE_STEP_REGNODE;
12851 else if (min == max && *RExC_parse == '?') {
12852 ckWARN2reg(RExC_parse + 1,
12853 "Useless use of greediness modifier '%c'",
12858 } /* End of is {m,n} */
12860 /* Here was a '{', but what followed it didn't form a quantifier. */
12866 NOT_REACHED; /*NOTREACHED*/
12869 /* Here we have a quantifier, and have calculated 'min' and 'max'.
12871 * Check and possibly adjust a zero width operand */
12872 if (! (flags & (HASWIDTH|POSTPONED))) {
12873 if (max > REG_INFTY/3) {
12874 if (origparse[0] == '\\' && origparse[1] == 'K') {
12876 "%" UTF8f " is forbidden - matches null string"
12878 UTF8fARG(UTF, (RExC_parse >= origparse
12879 ? RExC_parse - origparse
12883 ckWARN2reg(RExC_parse,
12884 "%" UTF8f " matches null string many times",
12885 UTF8fARG(UTF, (RExC_parse >= origparse
12886 ? RExC_parse - origparse
12892 /* There's no point in trying to match something 0 length more than
12893 * once except for extra side effects, which we don't have here since
12903 /* If this is a code block pass it up */
12904 *flagp |= (flags & POSTPONED);
12907 *flagp |= (flags & HASWIDTH);
12908 if (max == REG_INFTY)
12909 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
12912 /* 'SIMPLE' operands don't require full generality */
12913 if ((flags&SIMPLE)) {
12914 if (max == REG_INFTY) {
12916 if (UNLIKELY(RExC_pm_flags & PMf_WILDCARD)) {
12917 goto min0_maxINF_wildcard_forbidden;
12920 reginsert(pRExC_state, STAR, ret, depth+1);
12924 else if (min == 1) {
12925 reginsert(pRExC_state, PLUS, ret, depth+1);
12931 /* Here, SIMPLE, but not the '*' and '+' special cases */
12933 MARK_NAUGHTY_EXP(2, 2);
12934 reginsert(pRExC_state, CURLY, ret, depth+1);
12935 Set_Node_Offset(REGNODE_p(ret), parse_start+1); /* MJD */
12936 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
12938 else { /* not SIMPLE */
12939 const regnode_offset w = reg_node(pRExC_state, WHILEM);
12941 FLAGS(REGNODE_p(w)) = 0;
12942 if (! REGTAIL(pRExC_state, ret, w)) {
12943 REQUIRE_BRANCHJ(flagp, 0);
12945 if (RExC_use_BRANCHJ) {
12946 reginsert(pRExC_state, LONGJMP, ret, depth+1);
12947 reginsert(pRExC_state, NOTHING, ret, depth+1);
12948 NEXT_OFF(REGNODE_p(ret)) = 3; /* Go over LONGJMP. */
12950 reginsert(pRExC_state, CURLYX, ret, depth+1);
12952 Set_Node_Offset(REGNODE_p(ret), parse_start+1);
12953 Set_Node_Length(REGNODE_p(ret),
12954 op == '{' ? (RExC_parse - parse_start) : 1);
12956 if (RExC_use_BRANCHJ)
12957 NEXT_OFF(REGNODE_p(ret)) = 3; /* Go over NOTHING to
12959 if (! REGTAIL(pRExC_state, ret, reg_node(pRExC_state,
12962 REQUIRE_BRANCHJ(flagp, 0);
12964 RExC_whilem_seen++;
12965 MARK_NAUGHTY_EXP(1, 4); /* compound interest */
12968 /* Finish up the CURLY/CURLYX case */
12969 FLAGS(REGNODE_p(ret)) = 0;
12971 ARG1_SET(REGNODE_p(ret), (U16)min);
12972 ARG2_SET(REGNODE_p(ret), (U16)max);
12976 /* Process any greediness modifiers */
12977 if (*RExC_parse == '?') {
12978 nextchar(pRExC_state);
12979 reginsert(pRExC_state, MINMOD, ret, depth+1);
12980 if (! REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE)) {
12981 REQUIRE_BRANCHJ(flagp, 0);
12984 else if (*RExC_parse == '+') {
12985 regnode_offset ender;
12986 nextchar(pRExC_state);
12987 ender = reg_node(pRExC_state, SUCCEED);
12988 if (! REGTAIL(pRExC_state, ret, ender)) {
12989 REQUIRE_BRANCHJ(flagp, 0);
12991 reginsert(pRExC_state, SUSPEND, ret, depth+1);
12992 ender = reg_node(pRExC_state, TAIL);
12993 if (! REGTAIL(pRExC_state, ret, ender)) {
12994 REQUIRE_BRANCHJ(flagp, 0);
12998 /* Forbid extra quantifiers */
12999 if (isQUANTIFIER(RExC_parse, RExC_end)) {
13001 vFAIL("Nested quantifiers");
13006 min0_maxINF_wildcard_forbidden:
13008 /* Here we are in a wildcard match, and the minimum match length is 0, and
13009 * the max could be infinity. This is currently forbidden. The only
13010 * reason is to make it harder to write patterns that take a long long time
13011 * to halt, and because the use of this construct isn't necessary in
13012 * matching Unicode property values */
13014 /* diag_listed_as: Use of %s is not allowed in Unicode property wildcard
13015 subpatterns in regex; marked by <-- HERE in m/%s/
13017 vFAIL("Use of quantifier '*' is not allowed in Unicode property wildcard"
13020 /* Note, don't need to worry about the input being '{0,}', as a '}' isn't
13021 * legal at all in wildcards, so can't get this far */
13023 NOT_REACHED; /*NOTREACHED*/
13027 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state,
13028 regnode_offset * node_p,
13036 /* This routine teases apart the various meanings of \N and returns
13037 * accordingly. The input parameters constrain which meaning(s) is/are valid
13038 * in the current context.
13040 * Exactly one of <node_p> and <code_point_p> must be non-NULL.
13042 * If <code_point_p> is not NULL, the context is expecting the result to be a
13043 * single code point. If this \N instance turns out to a single code point,
13044 * the function returns TRUE and sets *code_point_p to that code point.
13046 * If <node_p> is not NULL, the context is expecting the result to be one of
13047 * the things representable by a regnode. If this \N instance turns out to be
13048 * one such, the function generates the regnode, returns TRUE and sets *node_p
13049 * to point to the offset of that regnode into the regex engine program being
13052 * If this instance of \N isn't legal in any context, this function will
13053 * generate a fatal error and not return.
13055 * On input, RExC_parse should point to the first char following the \N at the
13056 * time of the call. On successful return, RExC_parse will have been updated
13057 * to point to just after the sequence identified by this routine. Also
13058 * *flagp has been updated as needed.
13060 * When there is some problem with the current context and this \N instance,
13061 * the function returns FALSE, without advancing RExC_parse, nor setting
13062 * *node_p, nor *code_point_p, nor *flagp.
13064 * If <cp_count> is not NULL, the caller wants to know the length (in code
13065 * points) that this \N sequence matches. This is set, and the input is
13066 * parsed for errors, even if the function returns FALSE, as detailed below.
13068 * There are 6 possibilities here, as detailed in the next 6 paragraphs.
13070 * Probably the most common case is for the \N to specify a single code point.
13071 * *cp_count will be set to 1, and *code_point_p will be set to that code
13074 * Another possibility is for the input to be an empty \N{}. This is no
13075 * longer accepted, and will generate a fatal error.
13077 * Another possibility is for a custom charnames handler to be in effect which
13078 * translates the input name to an empty string. *cp_count will be set to 0.
13079 * *node_p will be set to a generated NOTHING node.
13081 * Still another possibility is for the \N to mean [^\n]. *cp_count will be
13082 * set to 0. *node_p will be set to a generated REG_ANY node.
13084 * The fifth possibility is that \N resolves to a sequence of more than one
13085 * code points. *cp_count will be set to the number of code points in the
13086 * sequence. *node_p will be set to a generated node returned by this
13087 * function calling S_reg().
13089 * The sixth and final possibility is that it is premature to be calling this
13090 * function; the parse needs to be restarted. This can happen when this
13091 * changes from /d to /u rules, or when the pattern needs to be upgraded to
13092 * UTF-8. The latter occurs only when the fifth possibility would otherwise
13093 * be in effect, and is because one of those code points requires the pattern
13094 * to be recompiled as UTF-8. The function returns FALSE, and sets the
13095 * RESTART_PARSE and NEED_UTF8 flags in *flagp, as appropriate. When this
13096 * happens, the caller needs to desist from continuing parsing, and return
13097 * this information to its caller. This is not set for when there is only one
13098 * code point, as this can be called as part of an ANYOF node, and they can
13099 * store above-Latin1 code points without the pattern having to be in UTF-8.
13101 * For non-single-quoted regexes, the tokenizer has resolved character and
13102 * sequence names inside \N{...} into their Unicode values, normalizing the
13103 * result into what we should see here: '\N{U+c1.c2...}', where c1... are the
13104 * hex-represented code points in the sequence. This is done there because
13105 * the names can vary based on what charnames pragma is in scope at the time,
13106 * so we need a way to take a snapshot of what they resolve to at the time of
13107 * the original parse. [perl #56444].
13109 * That parsing is skipped for single-quoted regexes, so here we may get
13110 * '\N{NAME}', which is parsed now. If the single-quoted regex is something
13111 * like '\N{U+41}', that code point is Unicode, and has to be translated into
13112 * the native character set for non-ASCII platforms. The other possibilities
13113 * are already native, so no translation is done. */
13115 char * endbrace; /* points to '}' following the name */
13116 char * e; /* points to final non-blank before endbrace */
13117 char* p = RExC_parse; /* Temporary */
13119 SV * substitute_parse = NULL;
13124 DECLARE_AND_GET_RE_DEBUG_FLAGS;
13126 PERL_ARGS_ASSERT_GROK_BSLASH_N;
13128 assert(cBOOL(node_p) ^ cBOOL(code_point_p)); /* Exactly one should be set */
13129 assert(! (node_p && cp_count)); /* At most 1 should be set */
13131 if (cp_count) { /* Initialize return for the most common case */
13135 /* The [^\n] meaning of \N ignores spaces and comments under the /x
13136 * modifier. The other meanings do not (except blanks adjacent to and
13137 * within the braces), so use a temporary until we find out which we are
13138 * being called with */
13139 skip_to_be_ignored_text(pRExC_state, &p,
13140 FALSE /* Don't force to /x */ );
13142 /* Disambiguate between \N meaning a named character versus \N meaning
13143 * [^\n]. The latter is assumed when the {...} following the \N is a legal
13144 * quantifier, or if there is no '{' at all */
13145 if (*p != '{' || regcurly(p, RExC_end, NULL)) {
13155 *node_p = reg_node(pRExC_state, REG_ANY);
13156 *flagp |= HASWIDTH|SIMPLE;
13158 Set_Node_Length(REGNODE_p(*(node_p)), 1); /* MJD */
13162 /* The test above made sure that the next real character is a '{', but
13163 * under the /x modifier, it could be separated by space (or a comment and
13164 * \n) and this is not allowed (for consistency with \x{...} and the
13165 * tokenizer handling of \N{NAME}). */
13166 if (*RExC_parse != '{') {
13167 vFAIL("Missing braces on \\N{}");
13170 RExC_parse++; /* Skip past the '{' */
13172 endbrace = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
13173 if (! endbrace) { /* no trailing brace */
13174 vFAIL2("Missing right brace on \\%c{}", 'N');
13177 /* Here, we have decided it should be a named character or sequence. These
13178 * imply Unicode semantics */
13179 REQUIRE_UNI_RULES(flagp, FALSE);
13181 /* \N{_} is what toke.c returns to us to indicate a name that evaluates to
13182 * nothing at all (not allowed under strict) */
13183 if (endbrace - RExC_parse == 1 && *RExC_parse == '_') {
13184 RExC_parse = endbrace;
13186 RExC_parse++; /* Position after the "}" */
13187 vFAIL("Zero length \\N{}");
13193 nextchar(pRExC_state);
13198 *node_p = reg_node(pRExC_state, NOTHING);
13202 while (isBLANK(*RExC_parse)) {
13207 while (RExC_parse < e && isBLANK(*(e-1))) {
13211 if (e - RExC_parse < 2 || ! strBEGINs(RExC_parse, "U+")) {
13213 /* Here, the name isn't of the form U+.... This can happen if the
13214 * pattern is single-quoted, so didn't get evaluated in toke.c. Now
13215 * is the time to find out what the name means */
13217 const STRLEN name_len = e - RExC_parse;
13218 SV * value_sv; /* What does this name evaluate to */
13220 const U8 * value; /* string of name's value */
13221 STRLEN value_len; /* and its length */
13223 /* RExC_unlexed_names is a hash of names that weren't evaluated by
13224 * toke.c, and their values. Make sure is initialized */
13225 if (! RExC_unlexed_names) {
13226 RExC_unlexed_names = newHV();
13229 /* If we have already seen this name in this pattern, use that. This
13230 * allows us to only call the charnames handler once per name per
13231 * pattern. A broken or malicious handler could return something
13232 * different each time, which could cause the results to vary depending
13233 * on if something gets added or subtracted from the pattern that
13234 * causes the number of passes to change, for example */
13235 if ((value_svp = hv_fetch(RExC_unlexed_names, RExC_parse,
13238 value_sv = *value_svp;
13240 else { /* Otherwise we have to go out and get the name */
13241 const char * error_msg = NULL;
13242 value_sv = get_and_check_backslash_N_name(RExC_parse, e,
13246 RExC_parse = endbrace;
13250 /* If no error message, should have gotten a valid return */
13253 /* Save the name's meaning for later use */
13254 if (! hv_store(RExC_unlexed_names, RExC_parse, name_len,
13257 Perl_croak(aTHX_ "panic: hv_store() unexpectedly failed");
13261 /* Here, we have the value the name evaluates to in 'value_sv' */
13262 value = (U8 *) SvPV(value_sv, value_len);
13264 /* See if the result is one code point vs 0 or multiple */
13265 if (inRANGE(value_len, 1, ((UV) SvUTF8(value_sv)
13269 /* Here, exactly one code point. If that isn't what is wanted,
13271 if (! code_point_p) {
13276 /* Convert from string to numeric code point */
13277 *code_point_p = (SvUTF8(value_sv))
13278 ? valid_utf8_to_uvchr(value, NULL)
13281 /* Have parsed this entire single code point \N{...}. *cp_count
13282 * has already been set to 1, so don't do it again. */
13283 RExC_parse = endbrace;
13284 nextchar(pRExC_state);
13286 } /* End of is a single code point */
13288 /* Count the code points, if caller desires. The API says to do this
13289 * even if we will later return FALSE */
13293 *cp_count = (SvUTF8(value_sv))
13294 ? utf8_length(value, value + value_len)
13298 /* Fail if caller doesn't want to handle a multi-code-point sequence.
13299 * But don't back the pointer up if the caller wants to know how many
13300 * code points there are (they need to handle it themselves in this
13309 /* Convert this to a sub-pattern of the form "(?: ... )", and then call
13310 * reg recursively to parse it. That way, it retains its atomicness,
13311 * while not having to worry about any special handling that some code
13312 * points may have. */
13314 substitute_parse = newSVpvs("?:");
13315 sv_catsv(substitute_parse, value_sv);
13316 sv_catpv(substitute_parse, ")");
13318 /* The value should already be native, so no need to convert on EBCDIC
13320 assert(! RExC_recode_x_to_native);
13323 else { /* \N{U+...} */
13324 Size_t count = 0; /* code point count kept internally */
13326 /* We can get to here when the input is \N{U+...} or when toke.c has
13327 * converted a name to the \N{U+...} form. This include changing a
13328 * name that evaluates to multiple code points to \N{U+c1.c2.c3 ...} */
13330 RExC_parse += 2; /* Skip past the 'U+' */
13332 /* Code points are separated by dots. The '}' terminates the whole
13335 do { /* Loop until the ending brace */
13336 I32 flags = PERL_SCAN_SILENT_OVERFLOW
13337 | PERL_SCAN_SILENT_ILLDIGIT
13338 | PERL_SCAN_NOTIFY_ILLDIGIT
13339 | PERL_SCAN_ALLOW_MEDIAL_UNDERSCORES
13340 | PERL_SCAN_DISALLOW_PREFIX;
13341 STRLEN len = e - RExC_parse;
13343 char * start_digit = RExC_parse;
13344 UV cp = grok_hex(RExC_parse, &len, &flags, &overflow_value);
13349 vFAIL("Invalid hexadecimal number in \\N{U+...}");
13354 if (cp > MAX_LEGAL_CP) {
13355 vFAIL(form_cp_too_large_msg(16, start_digit, len, 0));
13358 if (RExC_parse >= e) { /* Got to the closing '}' */
13363 /* Here, is a single code point; fail if doesn't want that */
13364 if (! code_point_p) {
13369 /* A single code point is easy to handle; just return it */
13370 *code_point_p = UNI_TO_NATIVE(cp);
13371 RExC_parse = endbrace;
13372 nextchar(pRExC_state);
13376 /* Here, the parse stopped bfore the ending brace. This is legal
13377 * only if that character is a dot separating code points, like a
13378 * multiple character sequence (of the form "\N{U+c1.c2. ... }".
13379 * So the next character must be a dot (and the one after that
13380 * can't be the ending brace, or we'd have something like
13383 if (*RExC_parse != '.' || RExC_parse + 1 >= e) {
13384 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
13385 ? UTF8SKIP(RExC_parse)
13387 RExC_parse = MIN(e, RExC_parse);/* Guard against malformed utf8
13392 /* Here, looks like its really a multiple character sequence. Fail
13393 * if that's not what the caller wants. But continue with counting
13394 * and error checking if they still want a count */
13395 if (! node_p && ! cp_count) {
13399 /* What is done here is to convert this to a sub-pattern of the
13400 * form \x{char1}\x{char2}... and then call reg recursively to
13401 * parse it (enclosing in "(?: ... )" ). That way, it retains its
13402 * atomicness, while not having to worry about special handling
13403 * that some code points may have. We don't create a subpattern,
13404 * but go through the motions of code point counting and error
13405 * checking, if the caller doesn't want a node returned. */
13407 if (node_p && ! substitute_parse) {
13408 substitute_parse = newSVpvs("?:");
13414 /* Convert to notation the rest of the code understands */
13415 sv_catpvs(substitute_parse, "\\x{");
13416 sv_catpvn(substitute_parse, start_digit,
13417 RExC_parse - start_digit);
13418 sv_catpvs(substitute_parse, "}");
13421 /* Move to after the dot (or ending brace the final time through.)
13426 } while (RExC_parse < e);
13428 if (! node_p) { /* Doesn't want the node */
13435 sv_catpvs(substitute_parse, ")");
13437 /* The values are Unicode, and therefore have to be converted to native
13438 * on a non-Unicode (meaning non-ASCII) platform. */
13439 SET_recode_x_to_native(1);
13442 /* Here, we have the string the name evaluates to, ready to be parsed,
13443 * stored in 'substitute_parse' as a series of valid "\x{...}\x{...}"
13444 * constructs. This can be called from within a substitute parse already.
13445 * The error reporting mechanism doesn't work for 2 levels of this, but the
13446 * code above has validated this new construct, so there should be no
13447 * errors generated by the below. And this isn't an exact copy, so the
13448 * mechanism to seamlessly deal with this won't work, so turn off warnings
13450 save_start = RExC_start;
13451 orig_end = RExC_end;
13453 RExC_parse = RExC_start = SvPVX(substitute_parse);
13454 RExC_end = RExC_parse + SvCUR(substitute_parse);
13455 TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE;
13457 *node_p = reg(pRExC_state, 1, &flags, depth+1);
13459 /* Restore the saved values */
13461 RExC_start = save_start;
13462 RExC_parse = endbrace;
13463 RExC_end = orig_end;
13464 SET_recode_x_to_native(0);
13466 SvREFCNT_dec_NN(substitute_parse);
13469 RETURN_FAIL_ON_RESTART(flags, flagp);
13470 FAIL2("panic: reg returned failure to grok_bslash_N, flags=%#" UVxf,
13473 *flagp |= flags&(HASWIDTH|SIMPLE|POSTPONED);
13475 nextchar(pRExC_state);
13482 S_compute_EXACTish(RExC_state_t *pRExC_state)
13486 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
13494 op = get_regex_charset(RExC_flags);
13495 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
13496 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
13497 been, so there is no hole */
13500 return op + EXACTF;
13503 /* Parse backref decimal value, unless it's too big to sensibly be a backref,
13504 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
13507 S_backref_value(char *p, char *e)
13509 const char* endptr = e;
13511 if (grok_atoUV(p, &val, &endptr) && val <= I32_MAX)
13518 - regatom - the lowest level
13520 Try to identify anything special at the start of the current parse position.
13521 If there is, then handle it as required. This may involve generating a
13522 single regop, such as for an assertion; or it may involve recursing, such as
13523 to handle a () structure.
13525 If the string doesn't start with something special then we gobble up
13526 as much literal text as we can. If we encounter a quantifier, we have to
13527 back off the final literal character, as that quantifier applies to just it
13528 and not to the whole string of literals.
13530 Once we have been able to handle whatever type of thing started the
13531 sequence, we return the offset into the regex engine program being compiled
13532 at which any next regnode should be placed.
13534 Returns 0, setting *flagp to TRYAGAIN if reg() returns 0 with TRYAGAIN.
13535 Returns 0, setting *flagp to RESTART_PARSE if the parse needs to be
13536 restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
13537 Otherwise does not return 0.
13539 Note: we have to be careful with escapes, as they can be both literal
13540 and special, and in the case of \10 and friends, context determines which.
13542 A summary of the code structure is:
13544 switch (first_byte) {
13545 cases for each special:
13546 handle this special;
13549 switch (2nd byte) {
13550 cases for each unambiguous special:
13551 handle this special;
13553 cases for each ambigous special/literal:
13555 if (special) handle here
13557 default: // unambiguously literal:
13560 default: // is a literal char
13563 create EXACTish node for literal;
13564 while (more input and node isn't full) {
13565 switch (input_byte) {
13566 cases for each special;
13567 make sure parse pointer is set so that the next call to
13568 regatom will see this special first
13569 goto loopdone; // EXACTish node terminated by prev. char
13571 append char to EXACTISH node;
13573 get next input byte;
13577 return the generated node;
13579 Specifically there are two separate switches for handling
13580 escape sequences, with the one for handling literal escapes requiring
13581 a dummy entry for all of the special escapes that are actually handled
13586 STATIC regnode_offset
13587 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
13589 regnode_offset ret = 0;
13595 DECLARE_AND_GET_RE_DEBUG_FLAGS;
13597 *flagp = 0; /* Initialize. */
13599 DEBUG_PARSE("atom");
13601 PERL_ARGS_ASSERT_REGATOM;
13604 parse_start = RExC_parse;
13605 assert(RExC_parse < RExC_end);
13606 switch ((U8)*RExC_parse) {
13608 RExC_seen_zerolen++;
13609 nextchar(pRExC_state);
13610 if (RExC_flags & RXf_PMf_MULTILINE)
13611 ret = reg_node(pRExC_state, MBOL);
13613 ret = reg_node(pRExC_state, SBOL);
13614 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13617 nextchar(pRExC_state);
13619 RExC_seen_zerolen++;
13620 if (RExC_flags & RXf_PMf_MULTILINE)
13621 ret = reg_node(pRExC_state, MEOL);
13623 ret = reg_node(pRExC_state, SEOL);
13624 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13627 nextchar(pRExC_state);
13628 if (RExC_flags & RXf_PMf_SINGLELINE)
13629 ret = reg_node(pRExC_state, SANY);
13631 ret = reg_node(pRExC_state, REG_ANY);
13632 *flagp |= HASWIDTH|SIMPLE;
13634 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13638 char * const oregcomp_parse = ++RExC_parse;
13639 ret = regclass(pRExC_state, flagp, depth+1,
13640 FALSE, /* means parse the whole char class */
13641 TRUE, /* allow multi-char folds */
13642 FALSE, /* don't silence non-portable warnings. */
13643 (bool) RExC_strict,
13644 TRUE, /* Allow an optimized regnode result */
13647 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13648 FAIL2("panic: regclass returned failure to regatom, flags=%#" UVxf,
13651 if (*RExC_parse != ']') {
13652 RExC_parse = oregcomp_parse;
13653 vFAIL("Unmatched [");
13655 nextchar(pRExC_state);
13656 Set_Node_Length(REGNODE_p(ret), RExC_parse - oregcomp_parse + 1); /* MJD */
13660 nextchar(pRExC_state);
13661 ret = reg(pRExC_state, 2, &flags, depth+1);
13663 if (flags & TRYAGAIN) {
13664 if (RExC_parse >= RExC_end) {
13665 /* Make parent create an empty node if needed. */
13666 *flagp |= TRYAGAIN;
13671 RETURN_FAIL_ON_RESTART(flags, flagp);
13672 FAIL2("panic: reg returned failure to regatom, flags=%#" UVxf,
13675 *flagp |= flags&(HASWIDTH|SIMPLE|POSTPONED);
13679 if (flags & TRYAGAIN) {
13680 *flagp |= TRYAGAIN;
13683 vFAIL("Internal urp");
13684 /* Supposed to be caught earlier. */
13690 vFAIL("Quantifier follows nothing");
13695 This switch handles escape sequences that resolve to some kind
13696 of special regop and not to literal text. Escape sequences that
13697 resolve to literal text are handled below in the switch marked
13700 Every entry in this switch *must* have a corresponding entry
13701 in the literal escape switch. However, the opposite is not
13702 required, as the default for this switch is to jump to the
13703 literal text handling code.
13706 switch ((U8)*RExC_parse) {
13707 /* Special Escapes */
13709 RExC_seen_zerolen++;
13710 /* Under wildcards, this is changed to match \n; should be
13711 * invisible to the user, as they have to compile under /m */
13712 if (RExC_pm_flags & PMf_WILDCARD) {
13713 ret = reg_node(pRExC_state, MBOL);
13716 ret = reg_node(pRExC_state, SBOL);
13717 /* SBOL is shared with /^/ so we set the flags so we can tell
13718 * /\A/ from /^/ in split. */
13719 FLAGS(REGNODE_p(ret)) = 1;
13721 goto finish_meta_pat;
13723 if (RExC_pm_flags & PMf_WILDCARD) {
13725 /* diag_listed_as: Use of %s is not allowed in Unicode property
13726 wildcard subpatterns in regex; marked by <-- HERE in m/%s/
13728 vFAIL("Use of '\\G' is not allowed in Unicode property"
13729 " wildcard subpatterns");
13731 ret = reg_node(pRExC_state, GPOS);
13732 RExC_seen |= REG_GPOS_SEEN;
13733 goto finish_meta_pat;
13735 if (!RExC_in_lookaround) {
13736 RExC_seen_zerolen++;
13737 ret = reg_node(pRExC_state, KEEPS);
13738 /* XXX:dmq : disabling in-place substitution seems to
13739 * be necessary here to avoid cases of memory corruption, as
13740 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
13742 RExC_seen |= REG_LOOKBEHIND_SEEN;
13743 goto finish_meta_pat;
13746 ++RExC_parse; /* advance past the 'K' */
13747 vFAIL("\\K not permitted in lookahead/lookbehind");
13750 if (RExC_pm_flags & PMf_WILDCARD) {
13751 /* See comment under \A above */
13752 ret = reg_node(pRExC_state, MEOL);
13755 ret = reg_node(pRExC_state, SEOL);
13757 RExC_seen_zerolen++; /* Do not optimize RE away */
13758 goto finish_meta_pat;
13760 if (RExC_pm_flags & PMf_WILDCARD) {
13761 /* See comment under \A above */
13762 ret = reg_node(pRExC_state, MEOL);
13765 ret = reg_node(pRExC_state, EOS);
13767 RExC_seen_zerolen++; /* Do not optimize RE away */
13768 goto finish_meta_pat;
13770 vFAIL("\\C no longer supported");
13772 ret = reg_node(pRExC_state, CLUMP);
13773 *flagp |= HASWIDTH;
13774 goto finish_meta_pat;
13782 regex_charset charset = get_regex_charset(RExC_flags);
13784 RExC_seen_zerolen++;
13785 RExC_seen |= REG_LOOKBEHIND_SEEN;
13786 op = BOUND + charset;
13788 if (RExC_parse >= RExC_end || *(RExC_parse + 1) != '{') {
13789 flags = TRADITIONAL_BOUND;
13790 if (op > BOUNDA) { /* /aa is same as /a */
13796 char name = *RExC_parse;
13797 char * endbrace = (char *) memchr(RExC_parse, '}',
13798 RExC_end - RExC_parse);
13799 char * e = endbrace;
13804 vFAIL2("Missing right brace on \\%c{}", name);
13807 while (isBLANK(*RExC_parse)) {
13811 while (RExC_parse < e && isBLANK(*(e - 1))) {
13815 if (e == RExC_parse) {
13816 RExC_parse = endbrace + 1; /* After the '}' */
13817 vFAIL2("Empty \\%c{}", name);
13820 length = e - RExC_parse;
13822 switch (*RExC_parse) {
13825 && (memNEs(RExC_parse + 1, length - 1, "cb")))
13827 goto bad_bound_type;
13832 if (length != 2 || *(RExC_parse + 1) != 'b') {
13833 goto bad_bound_type;
13838 if (length != 2 || *(RExC_parse + 1) != 'b') {
13839 goto bad_bound_type;
13844 if (length != 2 || *(RExC_parse + 1) != 'b') {
13845 goto bad_bound_type;
13853 "'%" UTF8f "' is an unknown bound type",
13854 UTF8fARG(UTF, length, e - length));
13855 NOT_REACHED; /*NOTREACHED*/
13857 RExC_parse = endbrace;
13858 REQUIRE_UNI_RULES(flagp, 0);
13863 else if (op >= BOUNDA) { /* /aa is same as /a */
13867 /* Don't have to worry about UTF-8, in this message because
13868 * to get here the contents of the \b must be ASCII */
13869 ckWARN4reg(RExC_parse + 1, /* Include the '}' in msg */
13870 "Using /u for '%.*s' instead of /%s",
13872 endbrace - length + 1,
13873 (charset == REGEX_ASCII_RESTRICTED_CHARSET)
13874 ? ASCII_RESTRICT_PAT_MODS
13875 : ASCII_MORE_RESTRICT_PAT_MODS);
13880 RExC_seen_d_op = TRUE;
13882 else if (op == BOUNDL) {
13883 RExC_contains_locale = 1;
13887 op += NBOUND - BOUND;
13890 ret = reg_node(pRExC_state, op);
13891 FLAGS(REGNODE_p(ret)) = flags;
13893 goto finish_meta_pat;
13897 ret = reg_node(pRExC_state, LNBREAK);
13898 *flagp |= HASWIDTH|SIMPLE;
13899 goto finish_meta_pat;
13913 /* These all have the same meaning inside [brackets], and it knows
13914 * how to do the best optimizations for them. So, pretend we found
13915 * these within brackets, and let it do the work */
13918 ret = regclass(pRExC_state, flagp, depth+1,
13919 TRUE, /* means just parse this element */
13920 FALSE, /* don't allow multi-char folds */
13921 FALSE, /* don't silence non-portable warnings. It
13922 would be a bug if these returned
13924 (bool) RExC_strict,
13925 TRUE, /* Allow an optimized regnode result */
13927 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13928 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
13929 * multi-char folds are allowed. */
13931 FAIL2("panic: regclass returned failure to regatom, flags=%#" UVxf,
13934 RExC_parse--; /* regclass() leaves this one too far ahead */
13937 /* The escapes above that don't take a parameter can't be
13938 * followed by a '{'. But 'pX', 'p{foo}' and
13939 * correspondingly 'P' can be */
13940 if ( RExC_parse - parse_start == 1
13941 && UCHARAT(RExC_parse + 1) == '{'
13942 && UNLIKELY(! regcurly(RExC_parse + 1, RExC_end, NULL)))
13945 vFAIL("Unescaped left brace in regex is illegal here");
13947 Set_Node_Offset(REGNODE_p(ret), parse_start);
13948 Set_Node_Length(REGNODE_p(ret), RExC_parse - parse_start + 1); /* MJD */
13949 nextchar(pRExC_state);
13952 /* Handle \N, \N{} and \N{NAMED SEQUENCE} (the latter meaning the
13953 * \N{...} evaluates to a sequence of more than one code points).
13954 * The function call below returns a regnode, which is our result.
13955 * The parameters cause it to fail if the \N{} evaluates to a
13956 * single code point; we handle those like any other literal. The
13957 * reason that the multicharacter case is handled here and not as
13958 * part of the EXACtish code is because of quantifiers. In
13959 * /\N{BLAH}+/, the '+' applies to the whole thing, and doing it
13960 * this way makes that Just Happen. dmq.
13961 * join_exact() will join this up with adjacent EXACTish nodes
13962 * later on, if appropriate. */
13964 if (grok_bslash_N(pRExC_state,
13965 &ret, /* Want a regnode returned */
13966 NULL, /* Fail if evaluates to a single code
13968 NULL, /* Don't need a count of how many code
13977 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13979 /* Here, evaluates to a single code point. Go get that */
13980 RExC_parse = parse_start;
13983 case 'k': /* Handle \k<NAME> and \k'NAME' and \k{NAME} */
13984 parse_named_seq: /* Also handle non-numeric \g{...} */
13987 if ( RExC_parse >= RExC_end - 1
13988 || (( ch = RExC_parse[1]) != '<'
13993 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
13994 vFAIL2("Sequence %.2s... not terminated", parse_start);
13998 while (isBLANK(*RExC_parse)) {
14002 ret = handle_named_backref(pRExC_state,
14014 case '1': case '2': case '3': case '4':
14015 case '5': case '6': case '7': case '8': case '9':
14018 char * endbrace = NULL;
14019 char * s = RExC_parse;
14020 char * e = RExC_end;
14027 endbrace = (char *) memchr(s, '}', RExC_end - s);
14030 /* Missing '}'. Position after the number to give
14031 * a better indication to the user of where the
14038 /* If it looks to be a name and not a number, go
14039 * handle it there */
14040 if (! isDIGIT(*s)) {
14041 goto parse_named_seq;
14046 } while isDIGIT(*s);
14049 vFAIL("Unterminated \\g{...} pattern");
14052 s++; /* Past the '{' */
14054 while (isBLANK(*s)) {
14058 /* Ignore trailing blanks */
14060 while (s < e && isBLANK(*(e - 1))) {
14065 /* Here, have isolated the meat of the construct from any
14066 * surrounding braces */
14073 if (endbrace && !isDIGIT(*s)) {
14074 goto parse_named_seq;
14078 num = S_backref_value(RExC_parse, RExC_end);
14080 vFAIL("Reference to invalid group 0");
14081 else if (num == I32_MAX) {
14082 if (isDIGIT(*RExC_parse))
14083 vFAIL("Reference to nonexistent group");
14085 vFAIL("Unterminated \\g... pattern");
14089 num = RExC_npar - num;
14091 vFAIL("Reference to nonexistent or unclosed group");
14095 num = S_backref_value(RExC_parse, RExC_end);
14096 /* bare \NNN might be backref or octal - if it is larger
14097 * than or equal RExC_npar then it is assumed to be an
14098 * octal escape. Note RExC_npar is +1 from the actual
14099 * number of parens. */
14100 /* Note we do NOT check if num == I32_MAX here, as that is
14101 * handled by the RExC_npar check */
14103 if ( /* any numeric escape < 10 is always a backref */
14105 /* any numeric escape < RExC_npar is a backref */
14106 && num >= RExC_npar
14107 /* cannot be an octal escape if it starts with [89]
14109 && ! inRANGE(*RExC_parse, '8', '9')
14111 /* Probably not meant to be a backref, instead likely
14112 * to be an octal character escape, e.g. \35 or \777.
14113 * The above logic should make it obvious why using
14114 * octal escapes in patterns is problematic. - Yves */
14115 RExC_parse = parse_start;
14120 /* At this point RExC_parse points at a numeric escape like
14121 * \12 or \88 or the digits in \g{34} or \g34 or something
14122 * similar, which we should NOT treat as an octal escape. It
14123 * may or may not be a valid backref escape. For instance
14124 * \88888888 is unlikely to be a valid backref.
14126 * We've already figured out what value the digits represent.
14127 * Now, move the parse to beyond them. */
14129 RExC_parse = endbrace + 1;
14131 else while (isDIGIT(*RExC_parse)) {
14135 if (num >= (I32)RExC_npar) {
14137 /* It might be a forward reference; we can't fail until we
14138 * know, by completing the parse to get all the groups, and
14139 * then reparsing */
14140 if (ALL_PARENS_COUNTED) {
14141 if (num >= RExC_total_parens) {
14142 vFAIL("Reference to nonexistent group");
14146 REQUIRE_PARENS_PASS;
14150 ret = reganode(pRExC_state,
14153 : (ASCII_FOLD_RESTRICTED)
14155 : (AT_LEAST_UNI_SEMANTICS)
14161 if (OP(REGNODE_p(ret)) == REFF) {
14162 RExC_seen_d_op = TRUE;
14164 *flagp |= HASWIDTH;
14166 /* override incorrect value set in reganode MJD */
14167 Set_Node_Offset(REGNODE_p(ret), parse_start);
14168 Set_Node_Cur_Length(REGNODE_p(ret), parse_start-1);
14169 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14170 FALSE /* Don't force to /x */ );
14174 if (RExC_parse >= RExC_end)
14175 FAIL("Trailing \\");
14178 /* Do not generate "unrecognized" warnings here, we fall
14179 back into the quick-grab loop below */
14180 RExC_parse = parse_start;
14182 } /* end of switch on a \foo sequence */
14187 /* '#' comments should have been spaced over before this function was
14189 assert((RExC_flags & RXf_PMf_EXTENDED) == 0);
14191 if (RExC_flags & RXf_PMf_EXTENDED) {
14192 RExC_parse = reg_skipcomment( pRExC_state, RExC_parse );
14193 if (RExC_parse < RExC_end)
14203 /* Here, we have determined that the next thing is probably a
14204 * literal character. RExC_parse points to the first byte of its
14205 * definition. (It still may be an escape sequence that evaluates
14206 * to a single character) */
14211 char *s, *old_s = NULL, *old_old_s = NULL;
14213 U32 max_string_len = 255;
14215 /* We may have to reparse the node, artificially stopping filling
14216 * it early, based on info gleaned in the first parse. This
14217 * variable gives where we stop. Make it above the normal stopping
14218 * place first time through; otherwise it would stop too early */
14219 U32 upper_fill = max_string_len + 1;
14221 /* We start out as an EXACT node, even if under /i, until we find a
14222 * character which is in a fold. The algorithm now segregates into
14223 * separate nodes, characters that fold from those that don't under
14224 * /i. (This hopefully will create nodes that are fixed strings
14225 * even under /i, giving the optimizer something to grab on to.)
14226 * So, if a node has something in it and the next character is in
14227 * the opposite category, that node is closed up, and the function
14228 * returns. Then regatom is called again, and a new node is
14229 * created for the new category. */
14230 U8 node_type = EXACT;
14232 /* Assume the node will be fully used; the excess is given back at
14233 * the end. Under /i, we may need to temporarily add the fold of
14234 * an extra character or two at the end to check for splitting
14235 * multi-char folds, so allocate extra space for that. We can't
14236 * make any other length assumptions, as a byte input sequence
14237 * could shrink down. */
14238 Ptrdiff_t current_string_nodes = STR_SZ(max_string_len
14242 ? UTF8_MAXBYTES_CASE
14243 /* Max non-UTF-8 expansion is 2 */ : 2)));
14245 bool next_is_quantifier;
14246 char * oldp = NULL;
14248 /* We can convert EXACTF nodes to EXACTFU if they contain only
14249 * characters that match identically regardless of the target
14250 * string's UTF8ness. The reason to do this is that EXACTF is not
14251 * trie-able, EXACTFU is, and EXACTFU requires fewer operations at
14254 * Similarly, we can convert EXACTFL nodes to EXACTFLU8 if they
14255 * contain only above-Latin1 characters (hence must be in UTF8),
14256 * which don't participate in folds with Latin1-range characters,
14257 * as the latter's folds aren't known until runtime. */
14258 bool maybe_exactfu = FOLD && (DEPENDS_SEMANTICS || LOC);
14260 /* Single-character EXACTish nodes are almost always SIMPLE. This
14261 * allows us to override this as encountered */
14262 U8 maybe_SIMPLE = SIMPLE;
14264 /* Does this node contain something that can't match unless the
14265 * target string is (also) in UTF-8 */
14266 bool requires_utf8_target = FALSE;
14268 /* The sequence 'ss' is problematic in non-UTF-8 patterns. */
14269 bool has_ss = FALSE;
14271 /* So is the MICRO SIGN */
14272 bool has_micro_sign = FALSE;
14274 /* Set when we fill up the current node and there is still more
14275 * text to process */
14278 /* Allocate an EXACT node. The node_type may change below to
14279 * another EXACTish node, but since the size of the node doesn't
14280 * change, it works */
14281 ret = regnode_guts(pRExC_state, node_type, current_string_nodes,
14283 FILL_NODE(ret, node_type);
14286 s = STRING(REGNODE_p(ret));
14297 maybe_exactfu = FOLD && (DEPENDS_SEMANTICS || LOC);
14298 maybe_SIMPLE = SIMPLE;
14299 requires_utf8_target = FALSE;
14301 has_micro_sign = FALSE;
14305 /* This breaks under rare circumstances. If folding, we do not
14306 * want to split a node at a character that is a non-final in a
14307 * multi-char fold, as an input string could just happen to want to
14308 * match across the node boundary. The code at the end of the loop
14309 * looks for this, and backs off until it finds not such a
14310 * character, but it is possible (though extremely, extremely
14311 * unlikely) for all characters in the node to be non-final fold
14312 * ones, in which case we just leave the node fully filled, and
14313 * hope that it doesn't match the string in just the wrong place */
14315 assert( ! UTF /* Is at the beginning of a character */
14316 || UTF8_IS_INVARIANT(UCHARAT(RExC_parse))
14317 || UTF8_IS_START(UCHARAT(RExC_parse)));
14319 overflowed = FALSE;
14321 /* Here, we have a literal character. Find the maximal string of
14322 * them in the input that we can fit into a single EXACTish node.
14323 * We quit at the first non-literal or when the node gets full, or
14324 * under /i the categorization of folding/non-folding character
14326 while (p < RExC_end && len < upper_fill) {
14328 /* In most cases each iteration adds one byte to the output.
14329 * The exceptions override this */
14330 Size_t added_len = 1;
14336 /* White space has already been ignored */
14337 assert( (RExC_flags & RXf_PMf_EXTENDED) == 0
14338 || ! is_PATWS_safe((p), RExC_end, UTF));
14341 const char* message;
14354 /* Literal Escapes Switch
14356 This switch is meant to handle escape sequences that
14357 resolve to a literal character.
14359 Every escape sequence that represents something
14360 else, like an assertion or a char class, is handled
14361 in the switch marked 'Special Escapes' above in this
14362 routine, but also has an entry here as anything that
14363 isn't explicitly mentioned here will be treated as
14364 an unescaped equivalent literal.
14367 switch ((U8)*++p) {
14369 /* These are all the special escapes. */
14370 case 'A': /* Start assertion */
14371 case 'b': case 'B': /* Word-boundary assertion*/
14372 case 'C': /* Single char !DANGEROUS! */
14373 case 'd': case 'D': /* digit class */
14374 case 'g': case 'G': /* generic-backref, pos assertion */
14375 case 'h': case 'H': /* HORIZWS */
14376 case 'k': case 'K': /* named backref, keep marker */
14377 case 'p': case 'P': /* Unicode property */
14378 case 'R': /* LNBREAK */
14379 case 's': case 'S': /* space class */
14380 case 'v': case 'V': /* VERTWS */
14381 case 'w': case 'W': /* word class */
14382 case 'X': /* eXtended Unicode "combining
14383 character sequence" */
14384 case 'z': case 'Z': /* End of line/string assertion */
14388 /* Anything after here is an escape that resolves to a
14389 literal. (Except digits, which may or may not)
14395 case 'N': /* Handle a single-code point named character. */
14396 RExC_parse = p + 1;
14397 if (! grok_bslash_N(pRExC_state,
14398 NULL, /* Fail if evaluates to
14399 anything other than a
14400 single code point */
14401 &ender, /* The returned single code
14403 NULL, /* Don't need a count of
14404 how many code points */
14409 if (*flagp & NEED_UTF8)
14410 FAIL("panic: grok_bslash_N set NEED_UTF8");
14411 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
14413 /* Here, it wasn't a single code point. Go close
14414 * up this EXACTish node. The switch() prior to
14415 * this switch handles the other cases */
14416 RExC_parse = p = oldp;
14420 RExC_parse = parse_start;
14422 /* The \N{} means the pattern, if previously /d,
14423 * becomes /u. That means it can't be an EXACTF node,
14424 * but an EXACTFU */
14425 if (node_type == EXACTF) {
14426 node_type = EXACTFU;
14428 /* If the node already contains something that
14429 * differs between EXACTF and EXACTFU, reparse it
14431 if (! maybe_exactfu) {
14452 ender = ESC_NATIVE;
14460 if (! grok_bslash_o(&p,
14465 (bool) RExC_strict,
14466 FALSE, /* No illegal cp's */
14469 RExC_parse = p; /* going to die anyway; point to
14470 exact spot of failure */
14474 if (message && TO_OUTPUT_WARNINGS(p)) {
14475 warn_non_literal_string(p, packed_warn, message);
14479 if (! grok_bslash_x(&p,
14484 (bool) RExC_strict,
14485 FALSE, /* No illegal cp's */
14488 RExC_parse = p; /* going to die anyway; point
14489 to exact spot of failure */
14493 if (message && TO_OUTPUT_WARNINGS(p)) {
14494 warn_non_literal_string(p, packed_warn, message);
14498 if (ender < 0x100) {
14499 if (RExC_recode_x_to_native) {
14500 ender = LATIN1_TO_NATIVE(ender);
14507 if (! grok_bslash_c(*p, &grok_c_char,
14508 &message, &packed_warn))
14510 /* going to die anyway; point to exact spot of
14512 RExC_parse = p + ((UTF)
14513 ? UTF8_SAFE_SKIP(p, RExC_end)
14518 ender = grok_c_char;
14520 if (message && TO_OUTPUT_WARNINGS(p)) {
14521 warn_non_literal_string(p, packed_warn, message);
14525 case '8': case '9': /* must be a backreference */
14527 /* we have an escape like \8 which cannot be an octal escape
14528 * so we exit the loop, and let the outer loop handle this
14529 * escape which may or may not be a legitimate backref. */
14531 case '1': case '2': case '3':case '4':
14532 case '5': case '6': case '7':
14534 /* When we parse backslash escapes there is ambiguity
14535 * between backreferences and octal escapes. Any escape
14536 * from \1 - \9 is a backreference, any multi-digit
14537 * escape which does not start with 0 and which when
14538 * evaluated as decimal could refer to an already
14539 * parsed capture buffer is a back reference. Anything
14542 * Note this implies that \118 could be interpreted as
14543 * 118 OR as "\11" . "8" depending on whether there
14544 * were 118 capture buffers defined already in the
14547 /* NOTE, RExC_npar is 1 more than the actual number of
14548 * parens we have seen so far, hence the "<" as opposed
14550 if ( !isDIGIT(p[1]) || S_backref_value(p, RExC_end) < RExC_npar)
14551 { /* Not to be treated as an octal constant, go
14559 I32 flags = PERL_SCAN_SILENT_ILLDIGIT
14560 | PERL_SCAN_NOTIFY_ILLDIGIT;
14562 ender = grok_oct(p, &numlen, &flags, NULL);
14564 if ( (flags & PERL_SCAN_NOTIFY_ILLDIGIT)
14565 && isDIGIT(*p) /* like \08, \178 */
14566 && ckWARN(WARN_REGEXP))
14568 reg_warn_non_literal_string(
14570 form_alien_digit_msg(8, numlen, p,
14571 RExC_end, UTF, FALSE));
14577 FAIL("Trailing \\");
14580 if (isALPHANUMERIC(*p)) {
14581 /* An alpha followed by '{' is going to fail next
14582 * iteration, so don't output this warning in that
14584 if (! isALPHA(*p) || *(p + 1) != '{') {
14585 ckWARN2reg(p + 1, "Unrecognized escape \\%.1s"
14586 " passed through", p);
14589 goto normal_default;
14590 } /* End of switch on '\' */
14593 /* Trying to gain new uses for '{' without breaking too
14594 * much existing code is hard. The solution currently
14596 * 1) If there is no ambiguity that a '{' should always
14597 * be taken literally, at the start of a construct, we
14599 * 2) If the literal '{' conflicts with our desired use
14600 * of it as a metacharacter, we die. The deprecation
14601 * cycles for this have come and gone.
14602 * 3) If there is ambiguity, we raise a simple warning.
14603 * This could happen, for example, if the user
14604 * intended it to introduce a quantifier, but slightly
14605 * misspelled the quantifier. Without this warning,
14606 * the quantifier would silently be taken as a literal
14607 * string of characters instead of a meta construct */
14608 if (len || (p > RExC_start && isALPHA_A(*(p - 1)))) {
14610 || ( p > parse_start + 1
14611 && isALPHA_A(*(p - 1))
14612 && *(p - 2) == '\\'))
14614 RExC_parse = p + 1;
14615 vFAIL("Unescaped left brace in regex is "
14618 ckWARNreg(p + 1, "Unescaped left brace in regex is"
14619 " passed through");
14621 goto normal_default;
14624 if (p > RExC_parse && RExC_strict) {
14625 ckWARN2reg(p + 1, "Unescaped literal '%c'", *p);
14628 default: /* A literal character */
14630 if (! UTF8_IS_INVARIANT(*p) && UTF) {
14632 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
14633 &numlen, UTF8_ALLOW_DEFAULT);
14639 } /* End of switch on the literal */
14641 /* Here, have looked at the literal character, and <ender>
14642 * contains its ordinal; <p> points to the character after it.
14646 REQUIRE_UTF8(flagp);
14647 if ( UNICODE_IS_PERL_EXTENDED(ender)
14648 && TO_OUTPUT_WARNINGS(p))
14650 ckWARN2_non_literal_string(p,
14651 packWARN(WARN_PORTABLE),
14652 PL_extended_cp_format,
14657 /* We need to check if the next non-ignored thing is a
14658 * quantifier. Move <p> to after anything that should be
14659 * ignored, which, as a side effect, positions <p> for the next
14660 * loop iteration */
14661 skip_to_be_ignored_text(pRExC_state, &p,
14662 FALSE /* Don't force to /x */ );
14664 /* If the next thing is a quantifier, it applies to this
14665 * character only, which means that this character has to be in
14666 * its own node and can't just be appended to the string in an
14667 * existing node, so if there are already other characters in
14668 * the node, close the node with just them, and set up to do
14669 * this character again next time through, when it will be the
14670 * only thing in its new node */
14672 next_is_quantifier = LIKELY(p < RExC_end)
14673 && UNLIKELY(isQUANTIFIER(p, RExC_end));
14675 if (next_is_quantifier && LIKELY(len)) {
14680 /* Ready to add 'ender' to the node */
14682 if (! FOLD) { /* The simple case, just append the literal */
14685 /* Don't output if it would overflow */
14686 if (UNLIKELY(len > max_string_len - ((UTF)
14687 ? UVCHR_SKIP(ender)
14694 if (UVCHR_IS_INVARIANT(ender) || ! UTF) {
14695 *(s++) = (char) ender;
14698 U8 * new_s = uvchr_to_utf8((U8*)s, ender);
14699 added_len = (char *) new_s - s;
14700 s = (char *) new_s;
14703 requires_utf8_target = TRUE;
14707 else if (LOC && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)) {
14709 /* Here are folding under /l, and the code point is
14710 * problematic. If this is the first character in the
14711 * node, change the node type to folding. Otherwise, if
14712 * this is the first problematic character, close up the
14713 * existing node, so can start a new node with this one */
14715 node_type = EXACTFL;
14716 RExC_contains_locale = 1;
14718 else if (node_type == EXACT) {
14723 /* This problematic code point means we can't simplify
14725 maybe_exactfu = FALSE;
14727 /* Although these two characters have folds that are
14728 * locale-problematic, they also have folds to above Latin1
14729 * that aren't a problem. Doing these now helps at
14731 if (UNLIKELY( ender == GREEK_CAPITAL_LETTER_MU
14732 || ender == LATIN_CAPITAL_LETTER_SHARP_S))
14737 /* Here, we are adding a problematic fold character.
14738 * "Problematic" in this context means that its fold isn't
14739 * known until runtime. (The non-problematic code points
14740 * are the above-Latin1 ones that fold to also all
14741 * above-Latin1. Their folds don't vary no matter what the
14742 * locale is.) But here we have characters whose fold
14743 * depends on the locale. We just add in the unfolded
14744 * character, and wait until runtime to fold it */
14745 goto not_fold_common;
14747 else /* regular fold; see if actually is in a fold */
14748 if ( (ender < 256 && ! IS_IN_SOME_FOLD_L1(ender))
14750 && ! _invlist_contains_cp(PL_in_some_fold, ender)))
14752 /* Here, folding, but the character isn't in a fold.
14754 * Start a new node if previous characters in the node were
14756 if (len && node_type != EXACT) {
14761 /* Here, continuing a node with non-folded characters. Add
14763 goto not_fold_common;
14765 else { /* Here, does participate in some fold */
14767 /* If this is the first character in the node, change its
14768 * type to folding. Otherwise, if this is the first
14769 * folding character in the node, close up the existing
14770 * node, so can start a new node with this one. */
14772 node_type = compute_EXACTish(pRExC_state);
14774 else if (node_type == EXACT) {
14779 if (UTF) { /* Alway use the folded value for UTF-8
14781 if (UVCHR_IS_INVARIANT(ender)) {
14782 if (UNLIKELY(len + 1 > max_string_len)) {
14787 *(s)++ = (U8) toFOLD(ender);
14793 folded = _to_uni_fold_flags(
14795 (U8 *) s, /* We have allocated extra space
14796 in 's' so can't run off the
14800 | (( ASCII_FOLD_RESTRICTED
14801 || node_type == EXACTFL)
14802 ? FOLD_FLAGS_NOMIX_ASCII
14804 if (UNLIKELY(len + added_len > max_string_len)) {
14812 && LIKELY(folded != GREEK_SMALL_LETTER_MU))
14814 /* U+B5 folds to the MU, so its possible for a
14815 * non-UTF-8 target to match it */
14816 requires_utf8_target = TRUE;
14820 else { /* Here is non-UTF8. */
14822 /* The fold will be one or (rarely) two characters.
14823 * Check that there's room for at least a single one
14824 * before setting any flags, etc. Because otherwise an
14825 * overflowing character could cause a flag to be set
14826 * even though it doesn't end up in this node. (For
14827 * the two character fold, we check again, before
14828 * setting any flags) */
14829 if (UNLIKELY(len + 1 > max_string_len)) {
14834 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
14835 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
14836 || UNICODE_DOT_DOT_VERSION > 0)
14838 /* On non-ancient Unicodes, check for the only possible
14839 * multi-char fold */
14840 if (UNLIKELY(ender == LATIN_SMALL_LETTER_SHARP_S)) {
14842 /* This potential multi-char fold means the node
14843 * can't be simple (because it could match more
14844 * than a single char). And in some cases it will
14845 * match 'ss', so set that flag */
14849 /* It can't change to be an EXACTFU (unless already
14850 * is one). We fold it iff under /u rules. */
14851 if (node_type != EXACTFU) {
14852 maybe_exactfu = FALSE;
14855 if (UNLIKELY(len + 2 > max_string_len)) {
14864 goto done_with_this_char;
14867 else if ( UNLIKELY(isALPHA_FOLD_EQ(ender, 's'))
14869 && UNLIKELY(isALPHA_FOLD_EQ(*(s-1), 's')))
14871 /* Also, the sequence 'ss' is special when not
14872 * under /u. If the target string is UTF-8, it
14873 * should match SHARP S; otherwise it won't. So,
14874 * here we have to exclude the possibility of this
14875 * node moving to /u.*/
14877 maybe_exactfu = FALSE;
14880 /* Here, the fold will be a single character */
14882 if (UNLIKELY(ender == MICRO_SIGN)) {
14883 has_micro_sign = TRUE;
14885 else if (PL_fold[ender] != PL_fold_latin1[ender]) {
14887 /* If the character's fold differs between /d and
14888 * /u, this can't change to be an EXACTFU node */
14889 maybe_exactfu = FALSE;
14892 *(s++) = (DEPENDS_SEMANTICS)
14893 ? (char) toFOLD(ender)
14895 /* Under /u, the fold of any character in
14896 * the 0-255 range happens to be its
14897 * lowercase equivalent, except for LATIN
14898 * SMALL LETTER SHARP S, which was handled
14899 * above, and the MICRO SIGN, whose fold
14900 * requires UTF-8 to represent. */
14901 : (char) toLOWER_L1(ender);
14903 } /* End of adding current character to the node */
14905 done_with_this_char:
14909 if (next_is_quantifier) {
14911 /* Here, the next input is a quantifier, and to get here,
14912 * the current character is the only one in the node. */
14916 } /* End of loop through literal characters */
14918 /* Here we have either exhausted the input or run out of room in
14919 * the node. If the former, we are done. (If we encountered a
14920 * character that can't be in the node, transfer is made directly
14921 * to <loopdone>, and so we wouldn't have fallen off the end of the
14923 if (LIKELY(! overflowed)) {
14927 /* Here we have run out of room. We can grow plain EXACT and
14928 * LEXACT nodes. If the pattern is gigantic enough, though,
14929 * eventually we'll have to artificially chunk the pattern into
14930 * multiple nodes. */
14931 if (! LOC && (node_type == EXACT || node_type == LEXACT)) {
14932 Size_t overhead = 1 + regarglen[OP(REGNODE_p(ret))];
14933 Size_t overhead_expansion = 0;
14935 Size_t max_nodes_for_string;
14939 /* Here we couldn't fit the final character in the current
14940 * node, so it will have to be reparsed, no matter what else we
14944 /* If would have overflowed a regular EXACT node, switch
14945 * instead to an LEXACT. The code below is structured so that
14946 * the actual growing code is common to changing from an EXACT
14947 * or just increasing the LEXACT size. This means that we have
14948 * to save the string in the EXACT case before growing, and
14949 * then copy it afterwards to its new location */
14950 if (node_type == EXACT) {
14951 overhead_expansion = regarglen[LEXACT] - regarglen[EXACT];
14952 RExC_emit += overhead_expansion;
14953 Copy(s0, temp, len, char);
14956 /* Ready to grow. If it was a plain EXACT, the string was
14957 * saved, and the first few bytes of it overwritten by adding
14958 * an argument field. We assume, as we do elsewhere in this
14959 * file, that one byte of remaining input will translate into
14960 * one byte of output, and if that's too small, we grow again,
14961 * if too large the excess memory is freed at the end */
14963 max_nodes_for_string = U16_MAX - overhead - overhead_expansion;
14964 achievable = MIN(max_nodes_for_string,
14965 current_string_nodes + STR_SZ(RExC_end - p));
14966 delta = achievable - current_string_nodes;
14968 /* If there is just no more room, go finish up this chunk of
14974 change_engine_size(pRExC_state, delta + overhead_expansion);
14975 current_string_nodes += delta;
14977 = sizeof(struct regnode) * current_string_nodes;
14978 upper_fill = max_string_len + 1;
14980 /* If the length was small, we know this was originally an
14981 * EXACT node now converted to LEXACT, and the string has to be
14982 * restored. Otherwise the string was untouched. 260 is just
14983 * a number safely above 255 so don't have to worry about
14984 * getting it precise */
14986 node_type = LEXACT;
14987 FILL_NODE(ret, node_type);
14988 s0 = STRING(REGNODE_p(ret));
14989 Copy(temp, s0, len, char);
14993 goto continue_parse;
14996 bool splittable = FALSE;
14997 bool backed_up = FALSE;
14998 char * e; /* should this be U8? */
14999 char * s_start; /* should this be U8? */
15001 /* Here is /i. Running out of room creates a problem if we are
15002 * folding, and the split happens in the middle of a
15003 * multi-character fold, as a match that should have occurred,
15004 * won't, due to the way nodes are matched, and our artificial
15005 * boundary. So back off until we aren't splitting such a
15006 * fold. If there is no such place to back off to, we end up
15007 * taking the entire node as-is. This can happen if the node
15008 * consists entirely of 'f' or entirely of 's' characters (or
15009 * things that fold to them) as 'ff' and 'ss' are
15010 * multi-character folds.
15012 * The Unicode standard says that multi character folds consist
15013 * of either two or three characters. That means we would be
15014 * splitting one if the final character in the node is at the
15015 * beginning of either type, or is the second of a three
15019 * ender is the code point of the character that won't fit
15021 * s points to just beyond the final byte in the node.
15022 * It's where we would place ender if there were
15023 * room, and where in fact we do place ender's fold
15024 * in the code below, as we've over-allocated space
15025 * for s0 (hence s) to allow for this
15026 * e starts at 's' and advances as we append things.
15027 * old_s is the same as 's'. (If ender had fit, 's' would
15028 * have been advanced to beyond it).
15029 * old_old_s points to the beginning byte of the final
15030 * character in the node
15031 * p points to the beginning byte in the input of the
15032 * character beyond 'ender'.
15033 * oldp points to the beginning byte in the input of
15036 * In the case of /il, we haven't folded anything that could be
15037 * affected by the locale. That means only above-Latin1
15038 * characters that fold to other above-latin1 characters get
15039 * folded at compile time. To check where a good place to
15040 * split nodes is, everything in it will have to be folded.
15041 * The boolean 'maybe_exactfu' keeps track in /il if there are
15042 * any unfolded characters in the node. */
15043 bool need_to_fold_loc = LOC && ! maybe_exactfu;
15045 /* If we do need to fold the node, we need a place to store the
15046 * folded copy, and a way to map back to the unfolded original
15048 char * locfold_buf = NULL;
15049 Size_t * loc_correspondence = NULL;
15051 if (! need_to_fold_loc) { /* The normal case. Just
15052 initialize to the actual node */
15055 s = old_old_s; /* Point to the beginning of the final char
15056 that fits in the node */
15060 /* Here, we have filled a /il node, and there are unfolded
15061 * characters in it. If the runtime locale turns out to be
15062 * UTF-8, there are possible multi-character folds, just
15063 * like when not under /l. The node hence can't terminate
15064 * in the middle of such a fold. To determine this, we
15065 * have to create a folded copy of this node. That means
15066 * reparsing the node, folding everything assuming a UTF-8
15067 * locale. (If at runtime it isn't such a locale, the
15068 * actions here wouldn't have been necessary, but we have
15069 * to assume the worst case.) If we find we need to back
15070 * off the folded string, we do so, and then map that
15071 * position back to the original unfolded node, which then
15072 * gets output, truncated at that spot */
15074 char * redo_p = RExC_parse;
15078 /* Allow enough space assuming a single byte input folds to
15079 * a single byte output, plus assume that the two unparsed
15080 * characters (that we may need) fold to the largest number
15081 * of bytes possible, plus extra for one more worst case
15082 * scenario. In the loop below, if we start eating into
15083 * that final spare space, we enlarge this initial space */
15084 Size_t size = max_string_len + (3 * UTF8_MAXBYTES_CASE) + 1;
15086 Newxz(locfold_buf, size, char);
15087 Newxz(loc_correspondence, size, Size_t);
15089 /* Redo this node's parse, folding into 'locfold_buf' */
15090 redo_p = RExC_parse;
15091 old_redo_e = redo_e = locfold_buf;
15092 while (redo_p <= oldp) {
15094 old_redo_e = redo_e;
15095 loc_correspondence[redo_e - locfold_buf]
15096 = redo_p - RExC_parse;
15101 (void) _to_utf8_fold_flags((U8 *) redo_p,
15106 redo_e += added_len;
15107 redo_p += UTF8SKIP(redo_p);
15111 /* Note that if this code is run on some ancient
15112 * Unicode versions, SHARP S doesn't fold to 'ss',
15113 * but rather than clutter the code with #ifdef's,
15114 * as is done above, we ignore that possibility.
15115 * This is ok because this code doesn't affect what
15116 * gets matched, but merely where the node gets
15118 if (UCHARAT(redo_p) != LATIN_SMALL_LETTER_SHARP_S) {
15119 *redo_e++ = toLOWER_L1(UCHARAT(redo_p));
15129 /* If we're getting so close to the end that a
15130 * worst-case fold in the next character would cause us
15131 * to overflow, increase, assuming one byte output byte
15132 * per one byte input one, plus room for another worst
15134 if ( redo_p <= oldp
15135 && redo_e > locfold_buf + size
15136 - (UTF8_MAXBYTES_CASE + 1))
15138 Size_t new_size = size
15140 + UTF8_MAXBYTES_CASE + 1;
15141 Ptrdiff_t e_offset = redo_e - locfold_buf;
15143 Renew(locfold_buf, new_size, char);
15144 Renew(loc_correspondence, new_size, Size_t);
15147 redo_e = locfold_buf + e_offset;
15151 /* Set so that things are in terms of the folded, temporary
15154 s_start = locfold_buf;
15159 /* Here, we have 's', 's_start' and 'e' set up to point to the
15160 * input that goes into the node, folded.
15162 * If the final character of the node and the fold of ender
15163 * form the first two characters of a three character fold, we
15164 * need to peek ahead at the next (unparsed) character in the
15165 * input to determine if the three actually do form such a
15166 * fold. Just looking at that character is not generally
15167 * sufficient, as it could be, for example, an escape sequence
15168 * that evaluates to something else, and it needs to be folded.
15170 * khw originally thought to just go through the parse loop one
15171 * extra time, but that doesn't work easily as that iteration
15172 * could cause things to think that the parse is over and to
15173 * goto loopdone. The character could be a '$' for example, or
15174 * the character beyond could be a quantifier, and other
15175 * glitches as well.
15177 * The solution used here for peeking ahead is to look at that
15178 * next character. If it isn't ASCII punctuation, then it will
15179 * be something that would continue on in an EXACTish node if
15180 * there were space. We append the fold of it to s, having
15181 * reserved enough room in s0 for the purpose. If we can't
15182 * reasonably peek ahead, we instead assume the worst case:
15183 * that it is something that would form the completion of a
15186 * If we can't split between s and ender, we work backwards
15187 * character-by-character down to s0. At each current point
15188 * see if we are at the beginning of a multi-char fold. If so,
15189 * that means we would be splitting the fold across nodes, and
15190 * so we back up one and try again.
15192 * If we're not at the beginning, we still could be at the
15193 * final two characters of a (rare) three character fold. We
15194 * check if the sequence starting at the character before the
15195 * current position (and including the current and next
15196 * characters) is a three character fold. If not, the node can
15197 * be split here. If it is, we have to backup two characters
15200 * Otherwise, the node can be split at the current position.
15202 * The same logic is used for UTF-8 patterns and not */
15206 /* Append the fold of ender */
15207 (void) _to_uni_fold_flags(
15211 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
15212 ? FOLD_FLAGS_NOMIX_ASCII
15216 /* 's' and the character folded to by ender may be the
15217 * first two of a three-character fold, in which case the
15218 * node should not be split here. That may mean examining
15219 * the so-far unparsed character starting at 'p'. But if
15220 * ender folded to more than one character, we already have
15221 * three characters to look at. Also, we first check if
15222 * the sequence consisting of s and the next character form
15223 * the first two of some three character fold. If not,
15224 * there's no need to peek ahead. */
15225 if ( added_len <= UTF8SKIP(e - added_len)
15226 && UNLIKELY(is_THREE_CHAR_FOLD_HEAD_utf8_safe(s, e)))
15228 /* Here, the two do form the beginning of a potential
15229 * three character fold. The unexamined character may
15230 * or may not complete it. Peek at it. It might be
15231 * something that ends the node or an escape sequence,
15232 * in which case we don't know without a lot of work
15233 * what it evaluates to, so we have to assume the worst
15234 * case: that it does complete the fold, and so we
15235 * can't split here. All such instances will have
15236 * that character be an ASCII punctuation character,
15237 * like a backslash. So, for that case, backup one and
15238 * drop down to try at that position */
15240 s = (char *) utf8_hop_back((U8 *) s, -1,
15245 /* Here, since it's not punctuation, it must be a
15246 * real character, and we can append its fold to
15247 * 'e' (having deliberately reserved enough space
15248 * for this eventuality) and drop down to check if
15249 * the three actually do form a folded sequence */
15250 (void) _to_utf8_fold_flags(
15251 (U8 *) p, (U8 *) RExC_end,
15254 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
15255 ? FOLD_FLAGS_NOMIX_ASCII
15261 /* Here, we either have three characters available in
15262 * sequence starting at 's', or we have two characters and
15263 * know that the following one can't possibly be part of a
15264 * three character fold. We go through the node backwards
15265 * until we find a place where we can split it without
15266 * breaking apart a multi-character fold. At any given
15267 * point we have to worry about if such a fold begins at
15268 * the current 's', and also if a three-character fold
15269 * begins at s-1, (containing s and s+1). Splitting in
15270 * either case would break apart a fold */
15272 char *prev_s = (char *) utf8_hop_back((U8 *) s, -1,
15275 /* If is a multi-char fold, can't split here. Backup
15276 * one char and try again */
15277 if (UNLIKELY(is_MULTI_CHAR_FOLD_utf8_safe(s, e))) {
15283 /* If the two characters beginning at 's' are part of a
15284 * three character fold starting at the character
15285 * before s, we can't split either before or after s.
15286 * Backup two chars and try again */
15287 if ( LIKELY(s > s_start)
15288 && UNLIKELY(is_THREE_CHAR_FOLD_utf8_safe(prev_s, e)))
15291 s = (char *) utf8_hop_back((U8 *) s, -1, (U8 *) s_start);
15296 /* Here there's no multi-char fold between s and the
15297 * next character following it. We can split */
15301 } while (s > s_start); /* End of loops backing up through the node */
15303 /* Here we either couldn't find a place to split the node,
15304 * or else we broke out of the loop setting 'splittable' to
15305 * true. In the latter case, the place to split is between
15306 * the first and second characters in the sequence starting
15312 else { /* Pattern not UTF-8 */
15313 if ( ender != LATIN_SMALL_LETTER_SHARP_S
15314 || ASCII_FOLD_RESTRICTED)
15316 assert( toLOWER_L1(ender) < 256 );
15317 *e++ = (char)(toLOWER_L1(ender)); /* should e and the cast be U8? */
15325 && UNLIKELY(is_THREE_CHAR_FOLD_HEAD_latin1_safe(s, e)))
15332 if ( UCHARAT(p) != LATIN_SMALL_LETTER_SHARP_S
15333 || ASCII_FOLD_RESTRICTED)
15335 assert( toLOWER_L1(ender) < 256 );
15336 *e++ = (char)(toLOWER_L1(ender)); /* should e and the cast be U8? */
15346 if (UNLIKELY(is_MULTI_CHAR_FOLD_latin1_safe(s, e))) {
15352 if ( LIKELY(s > s_start)
15353 && UNLIKELY(is_THREE_CHAR_FOLD_latin1_safe(s - 1, e)))
15363 } while (s > s_start);
15370 /* Here, we are done backing up. If we didn't backup at all
15371 * (the likely case), just proceed */
15374 /* If we did find a place to split, reparse the entire node
15375 * stopping where we have calculated. */
15378 /* If we created a temporary folded string under /l, we
15379 * have to map that back to the original */
15380 if (need_to_fold_loc) {
15381 upper_fill = loc_correspondence[s - s_start];
15382 if (upper_fill == 0) {
15383 FAIL2("panic: loc_correspondence[%d] is 0",
15384 (int) (s - s_start));
15386 Safefree(locfold_buf);
15387 Safefree(loc_correspondence);
15390 upper_fill = s - s0;
15395 /* Here the node consists entirely of non-final multi-char
15396 * folds. (Likely it is all 'f's or all 's's.) There's no
15397 * decent place to split it, so give up and just take the
15402 if (need_to_fold_loc) {
15403 Safefree(locfold_buf);
15404 Safefree(loc_correspondence);
15406 } /* End of verifying node ends with an appropriate char */
15408 /* We need to start the next node at the character that didn't fit
15412 loopdone: /* Jumped to when encounters something that shouldn't be
15415 /* Free up any over-allocated space; cast is to silence bogus
15416 * warning in MS VC */
15417 change_engine_size(pRExC_state,
15418 - (Ptrdiff_t) (current_string_nodes - STR_SZ(len)));
15420 /* I (khw) don't know if you can get here with zero length, but the
15421 * old code handled this situation by creating a zero-length EXACT
15422 * node. Might as well be NOTHING instead */
15424 OP(REGNODE_p(ret)) = NOTHING;
15428 /* If the node type is EXACT here, check to see if it
15429 * should be EXACTL, or EXACT_REQ8. */
15430 if (node_type == EXACT) {
15432 node_type = EXACTL;
15434 else if (requires_utf8_target) {
15435 node_type = EXACT_REQ8;
15438 else if (node_type == LEXACT) {
15439 if (requires_utf8_target) {
15440 node_type = LEXACT_REQ8;
15444 if ( UNLIKELY(has_micro_sign || has_ss)
15445 && (node_type == EXACTFU || ( node_type == EXACTF
15446 && maybe_exactfu)))
15447 { /* These two conditions are problematic in non-UTF-8
15450 node_type = EXACTFUP;
15452 else if (node_type == EXACTFL) {
15454 /* 'maybe_exactfu' is deliberately set above to
15455 * indicate this node type, where all code points in it
15457 if (maybe_exactfu) {
15458 node_type = EXACTFLU8;
15461 _invlist_contains_cp(PL_HasMultiCharFold, ender)))
15463 /* A character that folds to more than one will
15464 * match multiple characters, so can't be SIMPLE.
15465 * We don't have to worry about this with EXACTFLU8
15466 * nodes just above, as they have already been
15467 * folded (since the fold doesn't vary at run
15468 * time). Here, if the final character in the node
15469 * folds to multiple, it can't be simple. (This
15470 * only has an effect if the node has only a single
15471 * character, hence the final one, as elsewhere we
15472 * turn off simple for nodes whose length > 1 */
15476 else if (node_type == EXACTF) { /* Means is /di */
15478 /* This intermediate variable is needed solely because
15479 * the asserts in the macro where used exceed Win32's
15480 * literal string capacity */
15481 char first_char = * STRING(REGNODE_p(ret));
15483 /* If 'maybe_exactfu' is clear, then we need to stay
15484 * /di. If it is set, it means there are no code
15485 * points that match differently depending on UTF8ness
15486 * of the target string, so it can become an EXACTFU
15488 if (! maybe_exactfu) {
15489 RExC_seen_d_op = TRUE;
15491 else if ( isALPHA_FOLD_EQ(first_char, 's')
15492 || isALPHA_FOLD_EQ(ender, 's'))
15494 /* But, if the node begins or ends in an 's' we
15495 * have to defer changing it into an EXACTFU, as
15496 * the node could later get joined with another one
15497 * that ends or begins with 's' creating an 'ss'
15498 * sequence which would then wrongly match the
15499 * sharp s without the target being UTF-8. We
15500 * create a special node that we resolve later when
15501 * we join nodes together */
15503 node_type = EXACTFU_S_EDGE;
15506 node_type = EXACTFU;
15510 if (requires_utf8_target && node_type == EXACTFU) {
15511 node_type = EXACTFU_REQ8;
15515 OP(REGNODE_p(ret)) = node_type;
15516 setSTR_LEN(REGNODE_p(ret), len);
15517 RExC_emit += STR_SZ(len);
15519 /* If the node isn't a single character, it can't be SIMPLE */
15520 if (len > (Size_t) ((UTF) ? UTF8SKIP(STRING(REGNODE_p(ret))) : 1)) {
15524 *flagp |= HASWIDTH | maybe_SIMPLE;
15527 Set_Node_Length(REGNODE_p(ret), p - parse_start - 1);
15531 /* len is STRLEN which is unsigned, need to copy to signed */
15534 vFAIL("Internal disaster");
15537 } /* End of label 'defchar:' */
15539 } /* End of giant switch on input character */
15541 /* Position parse to next real character */
15542 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
15543 FALSE /* Don't force to /x */ );
15544 if ( *RExC_parse == '{'
15545 && OP(REGNODE_p(ret)) != SBOL && ! regcurly(RExC_parse, RExC_end, NULL))
15549 vFAIL("Unescaped left brace in regex is illegal here");
15551 ckWARNreg(RExC_parse + 1, "Unescaped left brace in regex is"
15552 " passed through");
15560 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
15562 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
15563 * sets up the bitmap and any flags, removing those code points from the
15564 * inversion list, setting it to NULL should it become completely empty */
15567 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
15568 assert(PL_regkind[OP(node)] == ANYOF);
15570 /* There is no bitmap for this node type */
15571 if (inRANGE(OP(node), ANYOFH, ANYOFRb)) {
15575 ANYOF_BITMAP_ZERO(node);
15576 if (*invlist_ptr) {
15578 /* This gets set if we actually need to modify things */
15579 bool change_invlist = FALSE;
15583 /* Start looking through *invlist_ptr */
15584 invlist_iterinit(*invlist_ptr);
15585 while (invlist_iternext(*invlist_ptr, &start, &end)) {
15589 if (end == UV_MAX && start <= NUM_ANYOF_CODE_POINTS) {
15590 ANYOF_FLAGS(node) |= ANYOF_MATCHES_ALL_ABOVE_BITMAP;
15593 /* Quit if are above what we should change */
15594 if (start >= NUM_ANYOF_CODE_POINTS) {
15598 change_invlist = TRUE;
15600 /* Set all the bits in the range, up to the max that we are doing */
15601 high = (end < NUM_ANYOF_CODE_POINTS - 1)
15603 : NUM_ANYOF_CODE_POINTS - 1;
15604 for (i = start; i <= (int) high; i++) {
15605 ANYOF_BITMAP_SET(node, i);
15608 invlist_iterfinish(*invlist_ptr);
15610 /* Done with loop; remove any code points that are in the bitmap from
15611 * *invlist_ptr; similarly for code points above the bitmap if we have
15612 * a flag to match all of them anyways */
15613 if (change_invlist) {
15614 _invlist_subtract(*invlist_ptr, PL_InBitmap, invlist_ptr);
15616 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
15617 _invlist_intersection(*invlist_ptr, PL_InBitmap, invlist_ptr);
15620 /* If have completely emptied it, remove it completely */
15621 if (_invlist_len(*invlist_ptr) == 0) {
15622 SvREFCNT_dec_NN(*invlist_ptr);
15623 *invlist_ptr = NULL;
15628 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
15629 Character classes ([:foo:]) can also be negated ([:^foo:]).
15630 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
15631 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
15632 but trigger failures because they are currently unimplemented. */
15634 #define POSIXCC_DONE(c) ((c) == ':')
15635 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
15636 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
15637 #define MAYBE_POSIXCC(c) (POSIXCC(c) || (c) == '^' || (c) == ';')
15639 #define WARNING_PREFIX "Assuming NOT a POSIX class since "
15640 #define NO_BLANKS_POSIX_WARNING "no blanks are allowed in one"
15641 #define SEMI_COLON_POSIX_WARNING "a semi-colon was found instead of a colon"
15643 #define NOT_MEANT_TO_BE_A_POSIX_CLASS (OOB_NAMEDCLASS - 1)
15645 /* 'posix_warnings' and 'warn_text' are names of variables in the following
15647 #define ADD_POSIX_WARNING(p, text) STMT_START { \
15648 if (posix_warnings) { \
15649 if (! RExC_warn_text ) RExC_warn_text = \
15650 (AV *) sv_2mortal((SV *) newAV()); \
15651 av_push(RExC_warn_text, Perl_newSVpvf(aTHX_ \
15655 REPORT_LOCATION_ARGS(p))); \
15658 #define CLEAR_POSIX_WARNINGS() \
15660 if (posix_warnings && RExC_warn_text) \
15661 av_clear(RExC_warn_text); \
15664 #define CLEAR_POSIX_WARNINGS_AND_RETURN(ret) \
15666 CLEAR_POSIX_WARNINGS(); \
15671 S_handle_possible_posix(pTHX_ RExC_state_t *pRExC_state,
15673 const char * const s, /* Where the putative posix class begins.
15674 Normally, this is one past the '['. This
15675 parameter exists so it can be somewhere
15676 besides RExC_parse. */
15677 char ** updated_parse_ptr, /* Where to set the updated parse pointer, or
15679 AV ** posix_warnings, /* Where to place any generated warnings, or
15681 const bool check_only /* Don't die if error */
15684 /* This parses what the caller thinks may be one of the three POSIX
15686 * 1) a character class, like [:blank:]
15687 * 2) a collating symbol, like [. .]
15688 * 3) an equivalence class, like [= =]
15689 * In the latter two cases, it croaks if it finds a syntactically legal
15690 * one, as these are not handled by Perl.
15692 * The main purpose is to look for a POSIX character class. It returns:
15693 * a) the class number
15694 * if it is a completely syntactically and semantically legal class.
15695 * 'updated_parse_ptr', if not NULL, is set to point to just after the
15696 * closing ']' of the class
15697 * b) OOB_NAMEDCLASS
15698 * if it appears that one of the three POSIX constructs was meant, but
15699 * its specification was somehow defective. 'updated_parse_ptr', if
15700 * not NULL, is set to point to the character just after the end
15701 * character of the class. See below for handling of warnings.
15702 * c) NOT_MEANT_TO_BE_A_POSIX_CLASS
15703 * if it doesn't appear that a POSIX construct was intended.
15704 * 'updated_parse_ptr' is not changed. No warnings nor errors are
15707 * In b) there may be errors or warnings generated. If 'check_only' is
15708 * TRUE, then any errors are discarded. Warnings are returned to the
15709 * caller via an AV* created into '*posix_warnings' if it is not NULL. If
15710 * instead it is NULL, warnings are suppressed.
15712 * The reason for this function, and its complexity is that a bracketed
15713 * character class can contain just about anything. But it's easy to
15714 * mistype the very specific posix class syntax but yielding a valid
15715 * regular bracketed class, so it silently gets compiled into something
15716 * quite unintended.
15718 * The solution adopted here maintains backward compatibility except that
15719 * it adds a warning if it looks like a posix class was intended but
15720 * improperly specified. The warning is not raised unless what is input
15721 * very closely resembles one of the 14 legal posix classes. To do this,
15722 * it uses fuzzy parsing. It calculates how many single-character edits it
15723 * would take to transform what was input into a legal posix class. Only
15724 * if that number is quite small does it think that the intention was a
15725 * posix class. Obviously these are heuristics, and there will be cases
15726 * where it errs on one side or another, and they can be tweaked as
15727 * experience informs.
15729 * The syntax for a legal posix class is:
15731 * qr/(?xa: \[ : \^? [[:lower:]]{4,6} : \] )/
15733 * What this routine considers syntactically to be an intended posix class
15734 * is this (the comments indicate some restrictions that the pattern
15737 * qr/(?x: \[? # The left bracket, possibly
15739 * \h* # possibly followed by blanks
15740 * (?: \^ \h* )? # possibly a misplaced caret
15741 * [:;]? # The opening class character,
15742 * # possibly omitted. A typo
15743 * # semi-colon can also be used.
15745 * \^? # possibly a correctly placed
15746 * # caret, but not if there was also
15747 * # a misplaced one
15749 * .{3,15} # The class name. If there are
15750 * # deviations from the legal syntax,
15751 * # its edit distance must be close
15752 * # to a real class name in order
15753 * # for it to be considered to be
15754 * # an intended posix class.
15756 * [[:punct:]]? # The closing class character,
15757 * # possibly omitted. If not a colon
15758 * # nor semi colon, the class name
15759 * # must be even closer to a valid
15762 * \]? # The right bracket, possibly
15766 * In the above, \h must be ASCII-only.
15768 * These are heuristics, and can be tweaked as field experience dictates.
15769 * There will be cases when someone didn't intend to specify a posix class
15770 * that this warns as being so. The goal is to minimize these, while
15771 * maximizing the catching of things intended to be a posix class that
15772 * aren't parsed as such.
15776 const char * const e = RExC_end;
15777 unsigned complement = 0; /* If to complement the class */
15778 bool found_problem = FALSE; /* Assume OK until proven otherwise */
15779 bool has_opening_bracket = FALSE;
15780 bool has_opening_colon = FALSE;
15781 int class_number = OOB_NAMEDCLASS; /* Out-of-bounds until find
15783 const char * possible_end = NULL; /* used for a 2nd parse pass */
15784 const char* name_start; /* ptr to class name first char */
15786 /* If the number of single-character typos the input name is away from a
15787 * legal name is no more than this number, it is considered to have meant
15788 * the legal name */
15789 int max_distance = 2;
15791 /* to store the name. The size determines the maximum length before we
15792 * decide that no posix class was intended. Should be at least
15793 * sizeof("alphanumeric") */
15795 STATIC_ASSERT_DECL(C_ARRAY_LENGTH(input_text) >= sizeof "alphanumeric");
15797 PERL_ARGS_ASSERT_HANDLE_POSSIBLE_POSIX;
15799 CLEAR_POSIX_WARNINGS();
15802 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
15805 if (*(p - 1) != '[') {
15806 ADD_POSIX_WARNING(p, "it doesn't start with a '['");
15807 found_problem = TRUE;
15810 has_opening_bracket = TRUE;
15813 /* They could be confused and think you can put spaces between the
15816 found_problem = TRUE;
15820 } while (p < e && isBLANK(*p));
15822 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15825 /* For [. .] and [= =]. These are quite different internally from [: :],
15826 * so they are handled separately. */
15827 if (POSIXCC_NOTYET(*p) && p < e - 3) /* 1 for the close, and 1 for the ']'
15828 and 1 for at least one char in it
15831 const char open_char = *p;
15832 const char * temp_ptr = p + 1;
15834 /* These two constructs are not handled by perl, and if we find a
15835 * syntactically valid one, we croak. khw, who wrote this code, finds
15836 * this explanation of them very unclear:
15837 * http://pubs.opengroup.org/onlinepubs/009696899/basedefs/xbd_chap09.html
15838 * And searching the rest of the internet wasn't very helpful either.
15839 * It looks like just about any byte can be in these constructs,
15840 * depending on the locale. But unless the pattern is being compiled
15841 * under /l, which is very rare, Perl runs under the C or POSIX locale.
15842 * In that case, it looks like [= =] isn't allowed at all, and that
15843 * [. .] could be any single code point, but for longer strings the
15844 * constituent characters would have to be the ASCII alphabetics plus
15845 * the minus-hyphen. Any sensible locale definition would limit itself
15846 * to these. And any portable one definitely should. Trying to parse
15847 * the general case is a nightmare (see [perl #127604]). So, this code
15848 * looks only for interiors of these constructs that match:
15850 * Using \w relaxes the apparent rules a little, without adding much
15851 * danger of mistaking something else for one of these constructs.
15853 * [. .] in some implementations described on the internet is usable to
15854 * escape a character that otherwise is special in bracketed character
15855 * classes. For example [.].] means a literal right bracket instead of
15856 * the ending of the class
15858 * [= =] can legitimately contain a [. .] construct, but we don't
15859 * handle this case, as that [. .] construct will later get parsed
15860 * itself and croak then. And [= =] is checked for even when not under
15861 * /l, as Perl has long done so.
15863 * The code below relies on there being a trailing NUL, so it doesn't
15864 * have to keep checking if the parse ptr < e.
15866 if (temp_ptr[1] == open_char) {
15869 else while ( temp_ptr < e
15870 && (isWORDCHAR(*temp_ptr) || *temp_ptr == '-'))
15875 if (*temp_ptr == open_char) {
15877 if (*temp_ptr == ']') {
15879 if (! found_problem && ! check_only) {
15880 RExC_parse = (char *) temp_ptr;
15881 vFAIL3("POSIX syntax [%c %c] is reserved for future "
15882 "extensions", open_char, open_char);
15885 /* Here, the syntax wasn't completely valid, or else the call
15886 * is to check-only */
15887 if (updated_parse_ptr) {
15888 *updated_parse_ptr = (char *) temp_ptr;
15891 CLEAR_POSIX_WARNINGS_AND_RETURN(OOB_NAMEDCLASS);
15895 /* If we find something that started out to look like one of these
15896 * constructs, but isn't, we continue below so that it can be checked
15897 * for being a class name with a typo of '.' or '=' instead of a colon.
15901 /* Here, we think there is a possibility that a [: :] class was meant, and
15902 * we have the first real character. It could be they think the '^' comes
15905 found_problem = TRUE;
15906 ADD_POSIX_WARNING(p + 1, "the '^' must come after the colon");
15911 found_problem = TRUE;
15915 } while (p < e && isBLANK(*p));
15917 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15921 /* But the first character should be a colon, which they could have easily
15922 * mistyped on a qwerty keyboard as a semi-colon (and which may be hard to
15923 * distinguish from a colon, so treat that as a colon). */
15926 has_opening_colon = TRUE;
15928 else if (*p == ';') {
15929 found_problem = TRUE;
15931 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
15932 has_opening_colon = TRUE;
15935 found_problem = TRUE;
15936 ADD_POSIX_WARNING(p, "there must be a starting ':'");
15938 /* Consider an initial punctuation (not one of the recognized ones) to
15939 * be a left terminator */
15940 if (*p != '^' && *p != ']' && isPUNCT(*p)) {
15945 /* They may think that you can put spaces between the components */
15947 found_problem = TRUE;
15951 } while (p < e && isBLANK(*p));
15953 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15958 /* We consider something like [^:^alnum:]] to not have been intended to
15959 * be a posix class, but XXX maybe we should */
15961 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15968 /* Again, they may think that you can put spaces between the components */
15970 found_problem = TRUE;
15974 } while (p < e && isBLANK(*p));
15976 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15981 /* XXX This ']' may be a typo, and something else was meant. But
15982 * treating it as such creates enough complications, that that
15983 * possibility isn't currently considered here. So we assume that the
15984 * ']' is what is intended, and if we've already found an initial '[',
15985 * this leaves this construct looking like [:] or [:^], which almost
15986 * certainly weren't intended to be posix classes */
15987 if (has_opening_bracket) {
15988 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15991 /* But this function can be called when we parse the colon for
15992 * something like qr/[alpha:]]/, so we back up to look for the
15997 found_problem = TRUE;
15998 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
16000 else if (*p != ':') {
16002 /* XXX We are currently very restrictive here, so this code doesn't
16003 * consider the possibility that, say, /[alpha.]]/ was intended to
16004 * be a posix class. */
16005 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
16008 /* Here we have something like 'foo:]'. There was no initial colon,
16009 * and we back up over 'foo. XXX Unlike the going forward case, we
16010 * don't handle typos of non-word chars in the middle */
16011 has_opening_colon = FALSE;
16014 while (p > RExC_start && isWORDCHAR(*p)) {
16019 /* Here, we have positioned ourselves to where we think the first
16020 * character in the potential class is */
16023 /* Now the interior really starts. There are certain key characters that
16024 * can end the interior, or these could just be typos. To catch both
16025 * cases, we may have to do two passes. In the first pass, we keep on
16026 * going unless we come to a sequence that matches
16027 * qr/ [[:punct:]] [[:blank:]]* \] /xa
16028 * This means it takes a sequence to end the pass, so two typos in a row if
16029 * that wasn't what was intended. If the class is perfectly formed, just
16030 * this one pass is needed. We also stop if there are too many characters
16031 * being accumulated, but this number is deliberately set higher than any
16032 * real class. It is set high enough so that someone who thinks that
16033 * 'alphanumeric' is a correct name would get warned that it wasn't.
16034 * While doing the pass, we keep track of where the key characters were in
16035 * it. If we don't find an end to the class, and one of the key characters
16036 * was found, we redo the pass, but stop when we get to that character.
16037 * Thus the key character was considered a typo in the first pass, but a
16038 * terminator in the second. If two key characters are found, we stop at
16039 * the second one in the first pass. Again this can miss two typos, but
16040 * catches a single one
16042 * In the first pass, 'possible_end' starts as NULL, and then gets set to
16043 * point to the first key character. For the second pass, it starts as -1.
16049 bool has_blank = FALSE;
16050 bool has_upper = FALSE;
16051 bool has_terminating_colon = FALSE;
16052 bool has_terminating_bracket = FALSE;
16053 bool has_semi_colon = FALSE;
16054 unsigned int name_len = 0;
16055 int punct_count = 0;
16059 /* Squeeze out blanks when looking up the class name below */
16060 if (isBLANK(*p) ) {
16062 found_problem = TRUE;
16067 /* The name will end with a punctuation */
16069 const char * peek = p + 1;
16071 /* Treat any non-']' punctuation followed by a ']' (possibly
16072 * with intervening blanks) as trying to terminate the class.
16073 * ']]' is very likely to mean a class was intended (but
16074 * missing the colon), but the warning message that gets
16075 * generated shows the error position better if we exit the
16076 * loop at the bottom (eventually), so skip it here. */
16078 if (peek < e && isBLANK(*peek)) {
16080 found_problem = TRUE;
16083 } while (peek < e && isBLANK(*peek));
16086 if (peek < e && *peek == ']') {
16087 has_terminating_bracket = TRUE;
16089 has_terminating_colon = TRUE;
16091 else if (*p == ';') {
16092 has_semi_colon = TRUE;
16093 has_terminating_colon = TRUE;
16096 found_problem = TRUE;
16103 /* Here we have punctuation we thought didn't end the class.
16104 * Keep track of the position of the key characters that are
16105 * more likely to have been class-enders */
16106 if (*p == ']' || *p == '[' || *p == ':' || *p == ';') {
16108 /* Allow just one such possible class-ender not actually
16109 * ending the class. */
16110 if (possible_end) {
16116 /* If we have too many punctuation characters, no use in
16118 if (++punct_count > max_distance) {
16122 /* Treat the punctuation as a typo. */
16123 input_text[name_len++] = *p;
16126 else if (isUPPER(*p)) { /* Use lowercase for lookup */
16127 input_text[name_len++] = toLOWER(*p);
16129 found_problem = TRUE;
16131 } else if (! UTF || UTF8_IS_INVARIANT(*p)) {
16132 input_text[name_len++] = *p;
16136 input_text[name_len++] = utf8_to_uvchr_buf((U8 *) p, e, NULL);
16140 /* The declaration of 'input_text' is how long we allow a potential
16141 * class name to be, before saying they didn't mean a class name at
16143 if (name_len >= C_ARRAY_LENGTH(input_text)) {
16148 /* We get to here when the possible class name hasn't been properly
16149 * terminated before:
16150 * 1) we ran off the end of the pattern; or
16151 * 2) found two characters, each of which might have been intended to
16152 * be the name's terminator
16153 * 3) found so many punctuation characters in the purported name,
16154 * that the edit distance to a valid one is exceeded
16155 * 4) we decided it was more characters than anyone could have
16156 * intended to be one. */
16158 found_problem = TRUE;
16160 /* In the final two cases, we know that looking up what we've
16161 * accumulated won't lead to a match, even a fuzzy one. */
16162 if ( name_len >= C_ARRAY_LENGTH(input_text)
16163 || punct_count > max_distance)
16165 /* If there was an intermediate key character that could have been
16166 * an intended end, redo the parse, but stop there */
16167 if (possible_end && possible_end != (char *) -1) {
16168 possible_end = (char *) -1; /* Special signal value to say
16169 we've done a first pass */
16174 /* Otherwise, it can't have meant to have been a class */
16175 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
16178 /* If we ran off the end, and the final character was a punctuation
16179 * one, back up one, to look at that final one just below. Later, we
16180 * will restore the parse pointer if appropriate */
16181 if (name_len && p == e && isPUNCT(*(p-1))) {
16186 if (p < e && isPUNCT(*p)) {
16188 has_terminating_bracket = TRUE;
16190 /* If this is a 2nd ']', and the first one is just below this
16191 * one, consider that to be the real terminator. This gives a
16192 * uniform and better positioning for the warning message */
16194 && possible_end != (char *) -1
16195 && *possible_end == ']'
16196 && name_len && input_text[name_len - 1] == ']')
16201 /* And this is actually equivalent to having done the 2nd
16202 * pass now, so set it to not try again */
16203 possible_end = (char *) -1;
16208 has_terminating_colon = TRUE;
16210 else if (*p == ';') {
16211 has_semi_colon = TRUE;
16212 has_terminating_colon = TRUE;
16220 /* Here, we have a class name to look up. We can short circuit the
16221 * stuff below for short names that can't possibly be meant to be a
16222 * class name. (We can do this on the first pass, as any second pass
16223 * will yield an even shorter name) */
16224 if (name_len < 3) {
16225 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
16228 /* Find which class it is. Initially switch on the length of the name.
16230 switch (name_len) {
16232 if (memEQs(name_start, 4, "word")) {
16233 /* this is not POSIX, this is the Perl \w */
16234 class_number = ANYOF_WORDCHAR;
16238 /* Names all of length 5: alnum alpha ascii blank cntrl digit
16239 * graph lower print punct space upper
16240 * Offset 4 gives the best switch position. */
16241 switch (name_start[4]) {
16243 if (memBEGINs(name_start, 5, "alph")) /* alpha */
16244 class_number = ANYOF_ALPHA;
16247 if (memBEGINs(name_start, 5, "spac")) /* space */
16248 class_number = ANYOF_SPACE;
16251 if (memBEGINs(name_start, 5, "grap")) /* graph */
16252 class_number = ANYOF_GRAPH;
16255 if (memBEGINs(name_start, 5, "asci")) /* ascii */
16256 class_number = ANYOF_ASCII;
16259 if (memBEGINs(name_start, 5, "blan")) /* blank */
16260 class_number = ANYOF_BLANK;
16263 if (memBEGINs(name_start, 5, "cntr")) /* cntrl */
16264 class_number = ANYOF_CNTRL;
16267 if (memBEGINs(name_start, 5, "alnu")) /* alnum */
16268 class_number = ANYOF_ALPHANUMERIC;
16271 if (memBEGINs(name_start, 5, "lowe")) /* lower */
16272 class_number = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
16273 else if (memBEGINs(name_start, 5, "uppe")) /* upper */
16274 class_number = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
16277 if (memBEGINs(name_start, 5, "digi")) /* digit */
16278 class_number = ANYOF_DIGIT;
16279 else if (memBEGINs(name_start, 5, "prin")) /* print */
16280 class_number = ANYOF_PRINT;
16281 else if (memBEGINs(name_start, 5, "punc")) /* punct */
16282 class_number = ANYOF_PUNCT;
16287 if (memEQs(name_start, 6, "xdigit"))
16288 class_number = ANYOF_XDIGIT;
16292 /* If the name exactly matches a posix class name the class number will
16293 * here be set to it, and the input almost certainly was meant to be a
16294 * posix class, so we can skip further checking. If instead the syntax
16295 * is exactly correct, but the name isn't one of the legal ones, we
16296 * will return that as an error below. But if neither of these apply,
16297 * it could be that no posix class was intended at all, or that one
16298 * was, but there was a typo. We tease these apart by doing fuzzy
16299 * matching on the name */
16300 if (class_number == OOB_NAMEDCLASS && found_problem) {
16301 const UV posix_names[][6] = {
16302 { 'a', 'l', 'n', 'u', 'm' },
16303 { 'a', 'l', 'p', 'h', 'a' },
16304 { 'a', 's', 'c', 'i', 'i' },
16305 { 'b', 'l', 'a', 'n', 'k' },
16306 { 'c', 'n', 't', 'r', 'l' },
16307 { 'd', 'i', 'g', 'i', 't' },
16308 { 'g', 'r', 'a', 'p', 'h' },
16309 { 'l', 'o', 'w', 'e', 'r' },
16310 { 'p', 'r', 'i', 'n', 't' },
16311 { 'p', 'u', 'n', 'c', 't' },
16312 { 's', 'p', 'a', 'c', 'e' },
16313 { 'u', 'p', 'p', 'e', 'r' },
16314 { 'w', 'o', 'r', 'd' },
16315 { 'x', 'd', 'i', 'g', 'i', 't' }
16317 /* The names of the above all have added NULs to make them the same
16318 * size, so we need to also have the real lengths */
16319 const UV posix_name_lengths[] = {
16320 sizeof("alnum") - 1,
16321 sizeof("alpha") - 1,
16322 sizeof("ascii") - 1,
16323 sizeof("blank") - 1,
16324 sizeof("cntrl") - 1,
16325 sizeof("digit") - 1,
16326 sizeof("graph") - 1,
16327 sizeof("lower") - 1,
16328 sizeof("print") - 1,
16329 sizeof("punct") - 1,
16330 sizeof("space") - 1,
16331 sizeof("upper") - 1,
16332 sizeof("word") - 1,
16333 sizeof("xdigit")- 1
16336 int temp_max = max_distance; /* Use a temporary, so if we
16337 reparse, we haven't changed the
16340 /* Use a smaller max edit distance if we are missing one of the
16342 if ( has_opening_bracket + has_opening_colon < 2
16343 || has_terminating_bracket + has_terminating_colon < 2)
16348 /* See if the input name is close to a legal one */
16349 for (i = 0; i < C_ARRAY_LENGTH(posix_names); i++) {
16351 /* Short circuit call if the lengths are too far apart to be
16353 if (abs( (int) (name_len - posix_name_lengths[i]))
16359 if (edit_distance(input_text,
16362 posix_name_lengths[i],
16366 { /* If it is close, it probably was intended to be a class */
16367 goto probably_meant_to_be;
16371 /* Here the input name is not close enough to a valid class name
16372 * for us to consider it to be intended to be a posix class. If
16373 * we haven't already done so, and the parse found a character that
16374 * could have been terminators for the name, but which we absorbed
16375 * as typos during the first pass, repeat the parse, signalling it
16376 * to stop at that character */
16377 if (possible_end && possible_end != (char *) -1) {
16378 possible_end = (char *) -1;
16383 /* Here neither pass found a close-enough class name */
16384 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
16387 probably_meant_to_be:
16389 /* Here we think that a posix specification was intended. Update any
16391 if (updated_parse_ptr) {
16392 *updated_parse_ptr = (char *) p;
16395 /* If a posix class name was intended but incorrectly specified, we
16396 * output or return the warnings */
16397 if (found_problem) {
16399 /* We set flags for these issues in the parse loop above instead of
16400 * adding them to the list of warnings, because we can parse it
16401 * twice, and we only want one warning instance */
16403 ADD_POSIX_WARNING(p, "the name must be all lowercase letters");
16406 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
16408 if (has_semi_colon) {
16409 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
16411 else if (! has_terminating_colon) {
16412 ADD_POSIX_WARNING(p, "there is no terminating ':'");
16414 if (! has_terminating_bracket) {
16415 ADD_POSIX_WARNING(p, "there is no terminating ']'");
16418 if ( posix_warnings
16420 && av_count(RExC_warn_text) > 0)
16422 *posix_warnings = RExC_warn_text;
16425 else if (class_number != OOB_NAMEDCLASS) {
16426 /* If it is a known class, return the class. The class number
16427 * #defines are structured so each complement is +1 to the normal
16429 CLEAR_POSIX_WARNINGS_AND_RETURN(class_number + complement);
16431 else if (! check_only) {
16433 /* Here, it is an unrecognized class. This is an error (unless the
16434 * call is to check only, which we've already handled above) */
16435 const char * const complement_string = (complement)
16438 RExC_parse = (char *) p;
16439 vFAIL3utf8f("POSIX class [:%s%" UTF8f ":] unknown",
16441 UTF8fARG(UTF, RExC_parse - name_start - 2, name_start));
16445 return OOB_NAMEDCLASS;
16447 #undef ADD_POSIX_WARNING
16449 STATIC unsigned int
16450 S_regex_set_precedence(const U8 my_operator) {
16452 /* Returns the precedence in the (?[...]) construct of the input operator,
16453 * specified by its character representation. The precedence follows
16454 * general Perl rules, but it extends this so that ')' and ']' have (low)
16455 * precedence even though they aren't really operators */
16457 switch (my_operator) {
16473 NOT_REACHED; /* NOTREACHED */
16474 return 0; /* Silence compiler warning */
16477 STATIC regnode_offset
16478 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
16479 I32 *flagp, U32 depth,
16480 char * const oregcomp_parse)
16482 /* Handle the (?[...]) construct to do set operations */
16484 U8 curchar; /* Current character being parsed */
16485 UV start, end; /* End points of code point ranges */
16486 SV* final = NULL; /* The end result inversion list */
16487 SV* result_string; /* 'final' stringified */
16488 AV* stack; /* stack of operators and operands not yet
16490 AV* fence_stack = NULL; /* A stack containing the positions in
16491 'stack' of where the undealt-with left
16492 parens would be if they were actually
16494 /* The 'volatile' is a workaround for an optimiser bug
16495 * in Solaris Studio 12.3. See RT #127455 */
16496 volatile IV fence = 0; /* Position of where most recent undealt-
16497 with left paren in stack is; -1 if none.
16499 STRLEN len; /* Temporary */
16500 regnode_offset node; /* Temporary, and final regnode returned by
16502 const bool save_fold = FOLD; /* Temporary */
16503 char *save_end, *save_parse; /* Temporaries */
16504 const bool in_locale = LOC; /* we turn off /l during processing */
16506 DECLARE_AND_GET_RE_DEBUG_FLAGS;
16508 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
16509 PERL_UNUSED_ARG(oregcomp_parse); /* Only for Set_Node_Length */
16511 DEBUG_PARSE("xcls");
16514 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
16517 /* The use of this operator implies /u. This is required so that the
16518 * compile time values are valid in all runtime cases */
16519 REQUIRE_UNI_RULES(flagp, 0);
16521 ckWARNexperimental(RExC_parse,
16522 WARN_EXPERIMENTAL__REGEX_SETS,
16523 "The regex_sets feature is experimental");
16525 /* Everything in this construct is a metacharacter. Operands begin with
16526 * either a '\' (for an escape sequence), or a '[' for a bracketed
16527 * character class. Any other character should be an operator, or
16528 * parenthesis for grouping. Both types of operands are handled by calling
16529 * regclass() to parse them. It is called with a parameter to indicate to
16530 * return the computed inversion list. The parsing here is implemented via
16531 * a stack. Each entry on the stack is a single character representing one
16532 * of the operators; or else a pointer to an operand inversion list. */
16534 #define IS_OPERATOR(a) SvIOK(a)
16535 #define IS_OPERAND(a) (! IS_OPERATOR(a))
16537 /* The stack is kept in Łukasiewicz order. (That's pronounced similar
16538 * to luke-a-shave-itch (or -itz), but people who didn't want to bother
16539 * with pronouncing it called it Reverse Polish instead, but now that YOU
16540 * know how to pronounce it you can use the correct term, thus giving due
16541 * credit to the person who invented it, and impressing your geek friends.
16542 * Wikipedia says that the pronounciation of "Ł" has been changing so that
16543 * it is now more like an English initial W (as in wonk) than an L.)
16545 * This means that, for example, 'a | b & c' is stored on the stack as
16553 * where the numbers in brackets give the stack [array] element number.
16554 * In this implementation, parentheses are not stored on the stack.
16555 * Instead a '(' creates a "fence" so that the part of the stack below the
16556 * fence is invisible except to the corresponding ')' (this allows us to
16557 * replace testing for parens, by using instead subtraction of the fence
16558 * position). As new operands are processed they are pushed onto the stack
16559 * (except as noted in the next paragraph). New operators of higher
16560 * precedence than the current final one are inserted on the stack before
16561 * the lhs operand (so that when the rhs is pushed next, everything will be
16562 * in the correct positions shown above. When an operator of equal or
16563 * lower precedence is encountered in parsing, all the stacked operations
16564 * of equal or higher precedence are evaluated, leaving the result as the
16565 * top entry on the stack. This makes higher precedence operations
16566 * evaluate before lower precedence ones, and causes operations of equal
16567 * precedence to left associate.
16569 * The only unary operator '!' is immediately pushed onto the stack when
16570 * encountered. When an operand is encountered, if the top of the stack is
16571 * a '!", the complement is immediately performed, and the '!' popped. The
16572 * resulting value is treated as a new operand, and the logic in the
16573 * previous paragraph is executed. Thus in the expression
16575 * the stack looks like
16581 * as 'b' gets parsed, the latter gets evaluated to '!b', and the stack
16588 * A ')' is treated as an operator with lower precedence than all the
16589 * aforementioned ones, which causes all operations on the stack above the
16590 * corresponding '(' to be evaluated down to a single resultant operand.
16591 * Then the fence for the '(' is removed, and the operand goes through the
16592 * algorithm above, without the fence.
16594 * A separate stack is kept of the fence positions, so that the position of
16595 * the latest so-far unbalanced '(' is at the top of it.
16597 * The ']' ending the construct is treated as the lowest operator of all,
16598 * so that everything gets evaluated down to a single operand, which is the
16601 sv_2mortal((SV *)(stack = newAV()));
16602 sv_2mortal((SV *)(fence_stack = newAV()));
16604 while (RExC_parse < RExC_end) {
16605 I32 top_index; /* Index of top-most element in 'stack' */
16606 SV** top_ptr; /* Pointer to top 'stack' element */
16607 SV* current = NULL; /* To contain the current inversion list
16609 SV* only_to_avoid_leaks;
16611 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
16612 TRUE /* Force /x */ );
16613 if (RExC_parse >= RExC_end) { /* Fail */
16617 curchar = UCHARAT(RExC_parse);
16621 #ifdef ENABLE_REGEX_SETS_DEBUGGING
16622 /* Enable with -Accflags=-DENABLE_REGEX_SETS_DEBUGGING */
16623 DEBUG_U(dump_regex_sets_structures(pRExC_state,
16624 stack, fence, fence_stack));
16627 top_index = av_tindex_skip_len_mg(stack);
16630 SV** stacked_ptr; /* Ptr to something already on 'stack' */
16631 char stacked_operator; /* The topmost operator on the 'stack'. */
16632 SV* lhs; /* Operand to the left of the operator */
16633 SV* rhs; /* Operand to the right of the operator */
16634 SV* fence_ptr; /* Pointer to top element of the fence
16638 if ( RExC_parse < RExC_end - 2
16639 && UCHARAT(RExC_parse + 1) == '?'
16640 && UCHARAT(RExC_parse + 2) == '^')
16642 const regnode_offset orig_emit = RExC_emit;
16643 SV * resultant_invlist;
16645 /* If is a '(?^', could be an embedded '(?^flags:(?[...])'.
16646 * This happens when we have some thing like
16648 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
16650 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
16652 * Here we would be handling the interpolated
16653 * '$thai_or_lao'. We handle this by a recursive call to
16654 * reg which returns the inversion list the
16655 * interpolated expression evaluates to. Actually, the
16656 * return is a special regnode containing a pointer to that
16657 * inversion list. If the return isn't that regnode alone,
16658 * we know that this wasn't such an interpolation, which is
16659 * an error: we need to get a single inversion list back
16660 * from the recursion */
16665 node = reg(pRExC_state, 2, flagp, depth+1);
16666 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16668 if ( OP(REGNODE_p(node)) != REGEX_SET
16669 /* If more than a single node returned, the nested
16670 * parens evaluated to more than just a (?[...]),
16671 * which isn't legal */
16672 || RExC_emit != orig_emit
16673 + NODE_STEP_REGNODE
16674 + regarglen[REGEX_SET])
16676 vFAIL("Expecting interpolated extended charclass");
16678 resultant_invlist = (SV *) ARGp(REGNODE_p(node));
16679 current = invlist_clone(resultant_invlist, NULL);
16680 SvREFCNT_dec(resultant_invlist);
16683 RExC_emit = orig_emit;
16684 goto handle_operand;
16687 /* A regular '('. Look behind for illegal syntax */
16688 if (top_index - fence >= 0) {
16689 /* If the top entry on the stack is an operator, it had
16690 * better be a '!', otherwise the entry below the top
16691 * operand should be an operator */
16692 if ( ! (top_ptr = av_fetch(stack, top_index, FALSE))
16693 || (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) != '!')
16694 || ( IS_OPERAND(*top_ptr)
16695 && ( top_index - fence < 1
16696 || ! (stacked_ptr = av_fetch(stack,
16699 || ! IS_OPERATOR(*stacked_ptr))))
16702 vFAIL("Unexpected '(' with no preceding operator");
16706 /* Stack the position of this undealt-with left paren */
16707 av_push(fence_stack, newSViv(fence));
16708 fence = top_index + 1;
16712 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
16713 * multi-char folds are allowed. */
16714 if (!regclass(pRExC_state, flagp, depth+1,
16715 TRUE, /* means parse just the next thing */
16716 FALSE, /* don't allow multi-char folds */
16717 FALSE, /* don't silence non-portable warnings. */
16719 FALSE, /* Require return to be an ANYOF */
16722 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16723 goto regclass_failed;
16728 /* regclass() will return with parsing just the \ sequence,
16729 * leaving the parse pointer at the next thing to parse */
16731 goto handle_operand;
16733 case '[': /* Is a bracketed character class */
16735 /* See if this is a [:posix:] class. */
16736 bool is_posix_class = (OOB_NAMEDCLASS
16737 < handle_possible_posix(pRExC_state,
16741 TRUE /* checking only */));
16742 /* If it is a posix class, leave the parse pointer at the '['
16743 * to fool regclass() into thinking it is part of a
16744 * '[[:posix:]]'. */
16745 if (! is_posix_class) {
16749 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
16750 * multi-char folds are allowed. */
16751 if (!regclass(pRExC_state, flagp, depth+1,
16752 is_posix_class, /* parse the whole char
16753 class only if not a
16755 FALSE, /* don't allow multi-char folds */
16756 TRUE, /* silence non-portable warnings. */
16758 FALSE, /* Require return to be an ANYOF */
16761 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16762 goto regclass_failed;
16767 /* function call leaves parse pointing to the ']', except if we
16769 if (is_posix_class) {
16773 goto handle_operand;
16777 if (top_index >= 1) {
16778 goto join_operators;
16781 /* Only a single operand on the stack: are done */
16785 if (av_tindex_skip_len_mg(fence_stack) < 0) {
16786 if (UCHARAT(RExC_parse - 1) == ']') {
16790 vFAIL("Unexpected ')'");
16793 /* If nothing after the fence, is missing an operand */
16794 if (top_index - fence < 0) {
16798 /* If at least two things on the stack, treat this as an
16800 if (top_index - fence >= 1) {
16801 goto join_operators;
16804 /* Here only a single thing on the fenced stack, and there is a
16805 * fence. Get rid of it */
16806 fence_ptr = av_pop(fence_stack);
16808 fence = SvIV(fence_ptr);
16809 SvREFCNT_dec_NN(fence_ptr);
16816 /* Having gotten rid of the fence, we pop the operand at the
16817 * stack top and process it as a newly encountered operand */
16818 current = av_pop(stack);
16819 if (IS_OPERAND(current)) {
16820 goto handle_operand;
16832 /* These binary operators should have a left operand already
16834 if ( top_index - fence < 0
16835 || top_index - fence == 1
16836 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
16837 || ! IS_OPERAND(*top_ptr))
16839 goto unexpected_binary;
16842 /* If only the one operand is on the part of the stack visible
16843 * to us, we just place this operator in the proper position */
16844 if (top_index - fence < 2) {
16846 /* Place the operator before the operand */
16848 SV* lhs = av_pop(stack);
16849 av_push(stack, newSVuv(curchar));
16850 av_push(stack, lhs);
16854 /* But if there is something else on the stack, we need to
16855 * process it before this new operator if and only if the
16856 * stacked operation has equal or higher precedence than the
16861 /* The operator on the stack is supposed to be below both its
16863 if ( ! (stacked_ptr = av_fetch(stack, top_index - 2, FALSE))
16864 || IS_OPERAND(*stacked_ptr))
16866 /* But if not, it's legal and indicates we are completely
16867 * done if and only if we're currently processing a ']',
16868 * which should be the final thing in the expression */
16869 if (curchar == ']') {
16875 vFAIL2("Unexpected binary operator '%c' with no "
16876 "preceding operand", curchar);
16878 stacked_operator = (char) SvUV(*stacked_ptr);
16880 if (regex_set_precedence(curchar)
16881 > regex_set_precedence(stacked_operator))
16883 /* Here, the new operator has higher precedence than the
16884 * stacked one. This means we need to add the new one to
16885 * the stack to await its rhs operand (and maybe more
16886 * stuff). We put it before the lhs operand, leaving
16887 * untouched the stacked operator and everything below it
16889 lhs = av_pop(stack);
16890 assert(IS_OPERAND(lhs));
16892 av_push(stack, newSVuv(curchar));
16893 av_push(stack, lhs);
16897 /* Here, the new operator has equal or lower precedence than
16898 * what's already there. This means the operation already
16899 * there should be performed now, before the new one. */
16901 rhs = av_pop(stack);
16902 if (! IS_OPERAND(rhs)) {
16904 /* This can happen when a ! is not followed by an operand,
16905 * like in /(?[\t &!])/ */
16909 lhs = av_pop(stack);
16911 if (! IS_OPERAND(lhs)) {
16913 /* This can happen when there is an empty (), like in
16914 * /(?[[0]+()+])/ */
16918 switch (stacked_operator) {
16920 _invlist_intersection(lhs, rhs, &rhs);
16925 _invlist_union(lhs, rhs, &rhs);
16929 _invlist_subtract(lhs, rhs, &rhs);
16932 case '^': /* The union minus the intersection */
16937 _invlist_union(lhs, rhs, &u);
16938 _invlist_intersection(lhs, rhs, &i);
16939 _invlist_subtract(u, i, &rhs);
16940 SvREFCNT_dec_NN(i);
16941 SvREFCNT_dec_NN(u);
16947 /* Here, the higher precedence operation has been done, and the
16948 * result is in 'rhs'. We overwrite the stacked operator with
16949 * the result. Then we redo this code to either push the new
16950 * operator onto the stack or perform any higher precedence
16951 * stacked operation */
16952 only_to_avoid_leaks = av_pop(stack);
16953 SvREFCNT_dec(only_to_avoid_leaks);
16954 av_push(stack, rhs);
16957 case '!': /* Highest priority, right associative */
16959 /* If what's already at the top of the stack is another '!",
16960 * they just cancel each other out */
16961 if ( (top_ptr = av_fetch(stack, top_index, FALSE))
16962 && (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) == '!'))
16964 only_to_avoid_leaks = av_pop(stack);
16965 SvREFCNT_dec(only_to_avoid_leaks);
16967 else { /* Otherwise, since it's right associative, just push
16969 av_push(stack, newSVuv(curchar));
16974 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16975 if (RExC_parse >= RExC_end) {
16978 vFAIL("Unexpected character");
16982 /* Here 'current' is the operand. If something is already on the
16983 * stack, we have to check if it is a !. But first, the code above
16984 * may have altered the stack in the time since we earlier set
16987 top_index = av_tindex_skip_len_mg(stack);
16988 if (top_index - fence >= 0) {
16989 /* If the top entry on the stack is an operator, it had better
16990 * be a '!', otherwise the entry below the top operand should
16991 * be an operator */
16992 top_ptr = av_fetch(stack, top_index, FALSE);
16994 if (IS_OPERATOR(*top_ptr)) {
16996 /* The only permissible operator at the top of the stack is
16997 * '!', which is applied immediately to this operand. */
16998 curchar = (char) SvUV(*top_ptr);
16999 if (curchar != '!') {
17000 SvREFCNT_dec(current);
17001 vFAIL2("Unexpected binary operator '%c' with no "
17002 "preceding operand", curchar);
17005 _invlist_invert(current);
17007 only_to_avoid_leaks = av_pop(stack);
17008 SvREFCNT_dec(only_to_avoid_leaks);
17010 /* And we redo with the inverted operand. This allows
17011 * handling multiple ! in a row */
17012 goto handle_operand;
17014 /* Single operand is ok only for the non-binary ')'
17016 else if ((top_index - fence == 0 && curchar != ')')
17017 || (top_index - fence > 0
17018 && (! (stacked_ptr = av_fetch(stack,
17021 || IS_OPERAND(*stacked_ptr))))
17023 SvREFCNT_dec(current);
17024 vFAIL("Operand with no preceding operator");
17028 /* Here there was nothing on the stack or the top element was
17029 * another operand. Just add this new one */
17030 av_push(stack, current);
17032 } /* End of switch on next parse token */
17034 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
17035 } /* End of loop parsing through the construct */
17037 vFAIL("Syntax error in (?[...])");
17041 if (RExC_parse >= RExC_end || RExC_parse[1] != ')') {
17042 if (RExC_parse < RExC_end) {
17046 vFAIL("Unexpected ']' with no following ')' in (?[...");
17049 if (av_tindex_skip_len_mg(fence_stack) >= 0) {
17050 vFAIL("Unmatched (");
17053 if (av_tindex_skip_len_mg(stack) < 0 /* Was empty */
17054 || ((final = av_pop(stack)) == NULL)
17055 || ! IS_OPERAND(final)
17056 || ! is_invlist(final)
17057 || av_tindex_skip_len_mg(stack) >= 0) /* More left on stack */
17060 SvREFCNT_dec(final);
17061 vFAIL("Incomplete expression within '(?[ ])'");
17064 /* Here, 'final' is the resultant inversion list from evaluating the
17065 * expression. Return it if so requested */
17066 if (return_invlist) {
17067 *return_invlist = final;
17071 if (RExC_sets_depth) { /* If within a recursive call, return in a special
17074 node = regpnode(pRExC_state, REGEX_SET, final);
17078 /* Otherwise generate a resultant node, based on 'final'. regclass()
17079 * is expecting a string of ranges and individual code points */
17080 invlist_iterinit(final);
17081 result_string = newSVpvs("");
17082 while (invlist_iternext(final, &start, &end)) {
17083 if (start == end) {
17084 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}", start);
17087 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}-\\x{%"
17088 UVXf "}", start, end);
17092 /* About to generate an ANYOF (or similar) node from the inversion list
17093 * we have calculated */
17094 save_parse = RExC_parse;
17095 RExC_parse = SvPV(result_string, len);
17096 save_end = RExC_end;
17097 RExC_end = RExC_parse + len;
17098 TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE;
17100 /* We turn off folding around the call, as the class we have
17101 * constructed already has all folding taken into consideration, and we
17102 * don't want regclass() to add to that */
17103 RExC_flags &= ~RXf_PMf_FOLD;
17104 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if multi-char
17105 * folds are allowed. */
17106 node = regclass(pRExC_state, flagp, depth+1,
17107 FALSE, /* means parse the whole char class */
17108 FALSE, /* don't allow multi-char folds */
17109 TRUE, /* silence non-portable warnings. The above may
17110 very well have generated non-portable code
17111 points, but they're valid on this machine */
17112 FALSE, /* similarly, no need for strict */
17114 /* We can optimize into something besides an ANYOF,
17115 * except under /l, which needs to be ANYOF because of
17116 * runtime checks for locale sanity, etc */
17122 RExC_parse = save_parse + 1;
17123 RExC_end = save_end;
17124 SvREFCNT_dec_NN(final);
17125 SvREFCNT_dec_NN(result_string);
17128 RExC_flags |= RXf_PMf_FOLD;
17132 RETURN_FAIL_ON_RESTART(*flagp, flagp);
17133 goto regclass_failed;
17136 /* Fix up the node type if we are in locale. (We have pretended we are
17137 * under /u for the purposes of regclass(), as this construct will only
17138 * work under UTF-8 locales. But now we change the opcode to be ANYOFL
17139 * (so as to cause any warnings about bad locales to be output in
17140 * regexec.c), and add the flag that indicates to check if not in a
17141 * UTF-8 locale. The reason we above forbid optimization into
17142 * something other than an ANYOF node is simply to minimize the number
17143 * of code changes in regexec.c. Otherwise we would have to create new
17144 * EXACTish node types and deal with them. This decision could be
17145 * revisited should this construct become popular.
17147 * (One might think we could look at the resulting ANYOF node and
17148 * suppress the flag if everything is above 255, as those would be
17149 * UTF-8 only, but this isn't true, as the components that led to that
17150 * result could have been locale-affected, and just happen to cancel
17151 * each other out under UTF-8 locales.) */
17153 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
17155 assert(OP(REGNODE_p(node)) == ANYOF);
17157 OP(REGNODE_p(node)) = ANYOFL;
17158 ANYOF_FLAGS(REGNODE_p(node))
17159 |= ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
17163 nextchar(pRExC_state);
17164 Set_Node_Length(REGNODE_p(node), RExC_parse - oregcomp_parse + 1); /* MJD */
17168 FAIL2("panic: regclass returned failure to handle_sets, " "flags=%#" UVxf,
17172 #ifdef ENABLE_REGEX_SETS_DEBUGGING
17175 S_dump_regex_sets_structures(pTHX_ RExC_state_t *pRExC_state,
17176 AV * stack, const IV fence, AV * fence_stack)
17177 { /* Dumps the stacks in handle_regex_sets() */
17179 const SSize_t stack_top = av_tindex_skip_len_mg(stack);
17180 const SSize_t fence_stack_top = av_tindex_skip_len_mg(fence_stack);
17183 PERL_ARGS_ASSERT_DUMP_REGEX_SETS_STRUCTURES;
17185 PerlIO_printf(Perl_debug_log, "\nParse position is:%s\n", RExC_parse);
17187 if (stack_top < 0) {
17188 PerlIO_printf(Perl_debug_log, "Nothing on stack\n");
17191 PerlIO_printf(Perl_debug_log, "Stack: (fence=%d)\n", (int) fence);
17192 for (i = stack_top; i >= 0; i--) {
17193 SV ** element_ptr = av_fetch(stack, i, FALSE);
17194 if (! element_ptr) {
17197 if (IS_OPERATOR(*element_ptr)) {
17198 PerlIO_printf(Perl_debug_log, "[%d]: %c\n",
17199 (int) i, (int) SvIV(*element_ptr));
17202 PerlIO_printf(Perl_debug_log, "[%d] ", (int) i);
17203 sv_dump(*element_ptr);
17208 if (fence_stack_top < 0) {
17209 PerlIO_printf(Perl_debug_log, "Nothing on fence_stack\n");
17212 PerlIO_printf(Perl_debug_log, "Fence_stack: \n");
17213 for (i = fence_stack_top; i >= 0; i--) {
17214 SV ** element_ptr = av_fetch(fence_stack, i, FALSE);
17215 if (! element_ptr) {
17218 PerlIO_printf(Perl_debug_log, "[%d]: %d\n",
17219 (int) i, (int) SvIV(*element_ptr));
17230 S_add_above_Latin1_folds(pTHX_ RExC_state_t *pRExC_state, const U8 cp, SV** invlist)
17232 /* This adds the Latin1/above-Latin1 folding rules.
17234 * This should be called only for a Latin1-range code points, cp, which is
17235 * known to be involved in a simple fold with other code points above
17236 * Latin1. It would give false results if /aa has been specified.
17237 * Multi-char folds are outside the scope of this, and must be handled
17240 PERL_ARGS_ASSERT_ADD_ABOVE_LATIN1_FOLDS;
17242 assert(HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(cp));
17244 /* The rules that are valid for all Unicode versions are hard-coded in */
17249 add_cp_to_invlist(*invlist, KELVIN_SIGN);
17253 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_LONG_S);
17256 *invlist = add_cp_to_invlist(*invlist, GREEK_CAPITAL_LETTER_MU);
17257 *invlist = add_cp_to_invlist(*invlist, GREEK_SMALL_LETTER_MU);
17259 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
17260 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
17261 *invlist = add_cp_to_invlist(*invlist, ANGSTROM_SIGN);
17263 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
17264 *invlist = add_cp_to_invlist(*invlist,
17265 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
17268 default: /* Other code points are checked against the data for the
17269 current Unicode version */
17271 Size_t folds_count;
17273 const U32 * remaining_folds;
17277 folded_cp = toFOLD(cp);
17280 U8 dummy_fold[UTF8_MAXBYTES_CASE+1];
17282 folded_cp = _to_fold_latin1(cp, dummy_fold, &dummy_len, 0);
17285 if (folded_cp > 255) {
17286 *invlist = add_cp_to_invlist(*invlist, folded_cp);
17289 folds_count = _inverse_folds(folded_cp, &first_fold,
17291 if (folds_count == 0) {
17293 /* Use deprecated warning to increase the chances of this being
17295 ckWARN2reg_d(RExC_parse,
17296 "Perl folding rules are not up-to-date for 0x%02X;"
17297 " please use the perlbug utility to report;", cp);
17302 if (first_fold > 255) {
17303 *invlist = add_cp_to_invlist(*invlist, first_fold);
17305 for (i = 0; i < folds_count - 1; i++) {
17306 if (remaining_folds[i] > 255) {
17307 *invlist = add_cp_to_invlist(*invlist,
17308 remaining_folds[i]);
17318 S_output_posix_warnings(pTHX_ RExC_state_t *pRExC_state, AV* posix_warnings)
17320 /* Output the elements of the array given by '*posix_warnings' as REGEXP
17324 const bool first_is_fatal = ckDEAD(packWARN(WARN_REGEXP));
17326 PERL_ARGS_ASSERT_OUTPUT_POSIX_WARNINGS;
17328 if (! TO_OUTPUT_WARNINGS(RExC_parse)) {
17329 CLEAR_POSIX_WARNINGS();
17333 while ((msg = av_shift(posix_warnings)) != &PL_sv_undef) {
17334 if (first_is_fatal) { /* Avoid leaking this */
17335 av_undef(posix_warnings); /* This isn't necessary if the
17336 array is mortal, but is a
17338 (void) sv_2mortal(msg);
17341 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s", SvPVX(msg));
17342 SvREFCNT_dec_NN(msg);
17345 UPDATE_WARNINGS_LOC(RExC_parse);
17348 PERL_STATIC_INLINE Size_t
17349 S_find_first_differing_byte_pos(const U8 * s1, const U8 * s2, const Size_t max)
17351 const U8 * const start = s1;
17352 const U8 * const send = start + max;
17354 PERL_ARGS_ASSERT_FIND_FIRST_DIFFERING_BYTE_POS;
17356 while (s1 < send && *s1 == *s2) {
17365 S_add_multi_match(pTHX_ AV* multi_char_matches, SV* multi_string, const STRLEN cp_count)
17367 /* This adds the string scalar <multi_string> to the array
17368 * <multi_char_matches>. <multi_string> is known to have exactly
17369 * <cp_count> code points in it. This is used when constructing a
17370 * bracketed character class and we find something that needs to match more
17371 * than a single character.
17373 * <multi_char_matches> is actually an array of arrays. Each top-level
17374 * element is an array that contains all the strings known so far that are
17375 * the same length. And that length (in number of code points) is the same
17376 * as the index of the top-level array. Hence, the [2] element is an
17377 * array, each element thereof is a string containing TWO code points;
17378 * while element [3] is for strings of THREE characters, and so on. Since
17379 * this is for multi-char strings there can never be a [0] nor [1] element.
17381 * When we rewrite the character class below, we will do so such that the
17382 * longest strings are written first, so that it prefers the longest
17383 * matching strings first. This is done even if it turns out that any
17384 * quantifier is non-greedy, out of this programmer's (khw) laziness. Tom
17385 * Christiansen has agreed that this is ok. This makes the test for the
17386 * ligature 'ffi' come before the test for 'ff', for example */
17389 AV** this_array_ptr;
17391 PERL_ARGS_ASSERT_ADD_MULTI_MATCH;
17393 if (! multi_char_matches) {
17394 multi_char_matches = newAV();
17397 if (av_exists(multi_char_matches, cp_count)) {
17398 this_array_ptr = (AV**) av_fetch(multi_char_matches, cp_count, FALSE);
17399 this_array = *this_array_ptr;
17402 this_array = newAV();
17403 av_store(multi_char_matches, cp_count,
17406 av_push(this_array, multi_string);
17408 return multi_char_matches;
17411 /* The names of properties whose definitions are not known at compile time are
17412 * stored in this SV, after a constant heading. So if the length has been
17413 * changed since initialization, then there is a run-time definition. */
17414 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
17415 (SvCUR(listsv) != initial_listsv_len)
17417 /* There is a restricted set of white space characters that are legal when
17418 * ignoring white space in a bracketed character class. This generates the
17419 * code to skip them.
17421 * There is a line below that uses the same white space criteria but is outside
17422 * this macro. Both here and there must use the same definition */
17423 #define SKIP_BRACKETED_WHITE_SPACE(do_skip, p, stop_p) \
17426 while (p < stop_p && isBLANK_A(UCHARAT(p))) \
17433 STATIC regnode_offset
17434 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
17435 const bool stop_at_1, /* Just parse the next thing, don't
17436 look for a full character class */
17437 bool allow_mutiple_chars,
17438 const bool silence_non_portable, /* Don't output warnings
17442 bool optimizable, /* ? Allow a non-ANYOF return
17444 SV** ret_invlist /* Return an inversion list, not a node */
17447 /* parse a bracketed class specification. Most of these will produce an
17448 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
17449 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
17450 * under /i with multi-character folds: it will be rewritten following the
17451 * paradigm of this example, where the <multi-fold>s are characters which
17452 * fold to multiple character sequences:
17453 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
17454 * gets effectively rewritten as:
17455 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
17456 * reg() gets called (recursively) on the rewritten version, and this
17457 * function will return what it constructs. (Actually the <multi-fold>s
17458 * aren't physically removed from the [abcdefghi], it's just that they are
17459 * ignored in the recursion by means of a flag:
17460 * <RExC_in_multi_char_class>.)
17462 * ANYOF nodes contain a bit map for the first NUM_ANYOF_CODE_POINTS
17463 * characters, with the corresponding bit set if that character is in the
17464 * list. For characters above this, an inversion list is used. There
17465 * are extra bits for \w, etc. in locale ANYOFs, as what these match is not
17466 * determinable at compile time
17468 * On success, returns the offset at which any next node should be placed
17469 * into the regex engine program being compiled.
17471 * Returns 0 otherwise, setting flagp to RESTART_PARSE if the parse needs
17472 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to
17476 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
17478 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
17479 regnode_offset ret = -1; /* Initialized to an illegal value */
17481 int namedclass = OOB_NAMEDCLASS;
17482 char *rangebegin = NULL;
17483 SV *listsv = NULL; /* List of \p{user-defined} whose definitions
17484 aren't available at the time this was called */
17485 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
17486 than just initialized. */
17487 SV* properties = NULL; /* Code points that match \p{} \P{} */
17488 SV* posixes = NULL; /* Code points that match classes like [:word:],
17489 extended beyond the Latin1 range. These have to
17490 be kept separate from other code points for much
17491 of this function because their handling is
17492 different under /i, and for most classes under
17494 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
17495 separate for a while from the non-complemented
17496 versions because of complications with /d
17498 SV* simple_posixes = NULL; /* But under some conditions, the classes can be
17499 treated more simply than the general case,
17500 leading to less compilation and execution
17502 UV element_count = 0; /* Number of distinct elements in the class.
17503 Optimizations may be possible if this is tiny */
17504 AV * multi_char_matches = NULL; /* Code points that fold to more than one
17505 character; used under /i */
17507 char * stop_ptr = RExC_end; /* where to stop parsing */
17509 /* ignore unescaped whitespace? */
17510 const bool skip_white = cBOOL( ret_invlist
17511 || (RExC_flags & RXf_PMf_EXTENDED_MORE));
17513 /* inversion list of code points this node matches only when the target
17514 * string is in UTF-8. These are all non-ASCII, < 256. (Because is under
17516 SV* upper_latin1_only_utf8_matches = NULL;
17518 /* Inversion list of code points this node matches regardless of things
17519 * like locale, folding, utf8ness of the target string */
17520 SV* cp_list = NULL;
17522 /* Like cp_list, but code points on this list need to be checked for things
17523 * that fold to/from them under /i */
17524 SV* cp_foldable_list = NULL;
17526 /* Like cp_list, but code points on this list are valid only when the
17527 * runtime locale is UTF-8 */
17528 SV* only_utf8_locale_list = NULL;
17530 /* In a range, if one of the endpoints is non-character-set portable,
17531 * meaning that it hard-codes a code point that may mean a different
17532 * charactger in ASCII vs. EBCDIC, as opposed to, say, a literal 'A' or a
17533 * mnemonic '\t' which each mean the same character no matter which
17534 * character set the platform is on. */
17535 unsigned int non_portable_endpoint = 0;
17537 /* Is the range unicode? which means on a platform that isn't 1-1 native
17538 * to Unicode (i.e. non-ASCII), each code point in it should be considered
17539 * to be a Unicode value. */
17540 bool unicode_range = FALSE;
17541 bool invert = FALSE; /* Is this class to be complemented */
17543 bool warn_super = ALWAYS_WARN_SUPER;
17545 const char * orig_parse = RExC_parse;
17547 /* This variable is used to mark where the end in the input is of something
17548 * that looks like a POSIX construct but isn't. During the parse, when
17549 * something looks like it could be such a construct is encountered, it is
17550 * checked for being one, but not if we've already checked this area of the
17551 * input. Only after this position is reached do we check again */
17552 char *not_posix_region_end = RExC_parse - 1;
17554 AV* posix_warnings = NULL;
17555 const bool do_posix_warnings = ckWARN(WARN_REGEXP);
17556 U8 op = ANYOF; /* The returned node-type, initialized to the expected
17558 U8 anyof_flags = 0; /* flag bits if the node is an ANYOF-type */
17559 U32 posixl = 0; /* bit field of posix classes matched under /l */
17562 /* Flags as to what things aren't knowable until runtime. (Note that these are
17563 * mutually exclusive.) */
17564 #define HAS_USER_DEFINED_PROPERTY 0x01 /* /u any user-defined properties that
17565 haven't been defined as of yet */
17566 #define HAS_D_RUNTIME_DEPENDENCY 0x02 /* /d if the target being matched is
17568 #define HAS_L_RUNTIME_DEPENDENCY 0x04 /* /l what the posix classes match and
17569 what gets folded */
17570 U32 has_runtime_dependency = 0; /* OR of the above flags */
17572 DECLARE_AND_GET_RE_DEBUG_FLAGS;
17574 PERL_ARGS_ASSERT_REGCLASS;
17576 PERL_UNUSED_ARG(depth);
17579 assert(! (ret_invlist && allow_mutiple_chars));
17581 /* If wants an inversion list returned, we can't optimize to something
17584 optimizable = FALSE;
17587 DEBUG_PARSE("clas");
17589 #if UNICODE_MAJOR_VERSION < 3 /* no multifolds in early Unicode */ \
17590 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0 \
17591 && UNICODE_DOT_DOT_VERSION == 0)
17592 allow_mutiple_chars = FALSE;
17595 /* We include the /i status at the beginning of this so that we can
17596 * know it at runtime */
17597 listsv = sv_2mortal(Perl_newSVpvf(aTHX_ "#%d\n", cBOOL(FOLD)));
17598 initial_listsv_len = SvCUR(listsv);
17599 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
17601 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse, RExC_end);
17603 assert(RExC_parse <= RExC_end);
17605 if (UCHARAT(RExC_parse) == '^') { /* Complement the class */
17608 allow_mutiple_chars = FALSE;
17610 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse, RExC_end);
17613 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
17614 if (! ret_invlist && MAYBE_POSIXCC(UCHARAT(RExC_parse))) {
17615 int maybe_class = handle_possible_posix(pRExC_state,
17617 ¬_posix_region_end,
17619 TRUE /* checking only */);
17620 if (maybe_class >= OOB_NAMEDCLASS && do_posix_warnings) {
17621 ckWARN4reg(not_posix_region_end,
17622 "POSIX syntax [%c %c] belongs inside character classes%s",
17623 *RExC_parse, *RExC_parse,
17624 (maybe_class == OOB_NAMEDCLASS)
17625 ? ((POSIXCC_NOTYET(*RExC_parse))
17626 ? " (but this one isn't implemented)"
17627 : " (but this one isn't fully valid)")
17633 /* If the caller wants us to just parse a single element, accomplish this
17634 * by faking the loop ending condition */
17635 if (stop_at_1 && RExC_end > RExC_parse) {
17636 stop_ptr = RExC_parse + 1;
17639 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
17640 if (UCHARAT(RExC_parse) == ']')
17641 goto charclassloop;
17645 if ( posix_warnings
17646 && av_tindex_skip_len_mg(posix_warnings) >= 0
17647 && RExC_parse > not_posix_region_end)
17649 /* Warnings about posix class issues are considered tentative until
17650 * we are far enough along in the parse that we can no longer
17651 * change our mind, at which point we output them. This is done
17652 * each time through the loop so that a later class won't zap them
17653 * before they have been dealt with. */
17654 output_posix_warnings(pRExC_state, posix_warnings);
17657 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse, RExC_end);
17659 if (RExC_parse >= stop_ptr) {
17663 if (UCHARAT(RExC_parse) == ']') {
17669 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
17670 save_value = value;
17671 save_prevvalue = prevvalue;
17674 rangebegin = RExC_parse;
17676 non_portable_endpoint = 0;
17678 if (UTF && ! UTF8_IS_INVARIANT(* RExC_parse)) {
17679 value = utf8n_to_uvchr((U8*)RExC_parse,
17680 RExC_end - RExC_parse,
17681 &numlen, UTF8_ALLOW_DEFAULT);
17682 RExC_parse += numlen;
17685 value = UCHARAT(RExC_parse++);
17687 if (value == '[') {
17688 char * posix_class_end;
17689 namedclass = handle_possible_posix(pRExC_state,
17692 do_posix_warnings ? &posix_warnings : NULL,
17693 FALSE /* die if error */);
17694 if (namedclass > OOB_NAMEDCLASS) {
17696 /* If there was an earlier attempt to parse this particular
17697 * posix class, and it failed, it was a false alarm, as this
17698 * successful one proves */
17699 if ( posix_warnings
17700 && av_tindex_skip_len_mg(posix_warnings) >= 0
17701 && not_posix_region_end >= RExC_parse
17702 && not_posix_region_end <= posix_class_end)
17704 av_undef(posix_warnings);
17707 RExC_parse = posix_class_end;
17709 else if (namedclass == OOB_NAMEDCLASS) {
17710 not_posix_region_end = posix_class_end;
17713 namedclass = OOB_NAMEDCLASS;
17716 else if ( RExC_parse - 1 > not_posix_region_end
17717 && MAYBE_POSIXCC(value))
17719 (void) handle_possible_posix(
17721 RExC_parse - 1, /* -1 because parse has already been
17723 ¬_posix_region_end,
17724 do_posix_warnings ? &posix_warnings : NULL,
17725 TRUE /* checking only */);
17727 else if ( strict && ! skip_white
17728 && ( _generic_isCC(value, _CC_VERTSPACE)
17729 || is_VERTWS_cp_high(value)))
17731 vFAIL("Literal vertical space in [] is illegal except under /x");
17733 else if (value == '\\') {
17734 /* Is a backslash; get the code point of the char after it */
17736 if (RExC_parse >= RExC_end) {
17737 vFAIL("Unmatched [");
17740 if (UTF && ! UTF8_IS_INVARIANT(UCHARAT(RExC_parse))) {
17741 value = utf8n_to_uvchr((U8*)RExC_parse,
17742 RExC_end - RExC_parse,
17743 &numlen, UTF8_ALLOW_DEFAULT);
17744 RExC_parse += numlen;
17747 value = UCHARAT(RExC_parse++);
17749 /* Some compilers cannot handle switching on 64-bit integer
17750 * values, therefore value cannot be an UV. Yes, this will
17751 * be a problem later if we want switch on Unicode.
17752 * A similar issue a little bit later when switching on
17753 * namedclass. --jhi */
17755 /* If the \ is escaping white space when white space is being
17756 * skipped, it means that that white space is wanted literally, and
17757 * is already in 'value'. Otherwise, need to translate the escape
17758 * into what it signifies. */
17759 if (! skip_white || ! isBLANK_A(value)) switch ((I32)value) {
17760 const char * message;
17764 case 'w': namedclass = ANYOF_WORDCHAR; break;
17765 case 'W': namedclass = ANYOF_NWORDCHAR; break;
17766 case 's': namedclass = ANYOF_SPACE; break;
17767 case 'S': namedclass = ANYOF_NSPACE; break;
17768 case 'd': namedclass = ANYOF_DIGIT; break;
17769 case 'D': namedclass = ANYOF_NDIGIT; break;
17770 case 'v': namedclass = ANYOF_VERTWS; break;
17771 case 'V': namedclass = ANYOF_NVERTWS; break;
17772 case 'h': namedclass = ANYOF_HORIZWS; break;
17773 case 'H': namedclass = ANYOF_NHORIZWS; break;
17774 case 'N': /* Handle \N{NAME} in class */
17776 const char * const backslash_N_beg = RExC_parse - 2;
17779 if (! grok_bslash_N(pRExC_state,
17780 NULL, /* No regnode */
17781 &value, /* Yes single value */
17782 &cp_count, /* Multiple code pt count */
17788 if (*flagp & NEED_UTF8)
17789 FAIL("panic: grok_bslash_N set NEED_UTF8");
17791 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
17793 if (cp_count < 0) {
17794 vFAIL("\\N in a character class must be a named character: \\N{...}");
17796 else if (cp_count == 0) {
17797 ckWARNreg(RExC_parse,
17798 "Ignoring zero length \\N{} in character class");
17800 else { /* cp_count > 1 */
17801 assert(cp_count > 1);
17802 if (! RExC_in_multi_char_class) {
17803 if ( ! allow_mutiple_chars
17806 || *RExC_parse == '-')
17810 vFAIL("\\N{} here is restricted to one character");
17812 ckWARNreg(RExC_parse, "Using just the first character returned by \\N{} in character class");
17813 break; /* <value> contains the first code
17814 point. Drop out of the switch to
17818 SV * multi_char_N = newSVpvn(backslash_N_beg,
17819 RExC_parse - backslash_N_beg);
17821 = add_multi_match(multi_char_matches,
17826 } /* End of cp_count != 1 */
17828 /* This element should not be processed further in this
17831 value = save_value;
17832 prevvalue = save_prevvalue;
17833 continue; /* Back to top of loop to get next char */
17836 /* Here, is a single code point, and <value> contains it */
17837 unicode_range = TRUE; /* \N{} are Unicode */
17845 if (RExC_pm_flags & PMf_WILDCARD) {
17847 /* diag_listed_as: Use of %s is not allowed in Unicode
17848 property wildcard subpatterns in regex; marked by <--
17850 vFAIL3("Use of '\\%c%c' is not allowed in Unicode property"
17851 " wildcard subpatterns", (char) value, *(RExC_parse - 1));
17854 /* \p means they want Unicode semantics */
17855 REQUIRE_UNI_RULES(flagp, 0);
17857 if (RExC_parse >= RExC_end)
17858 vFAIL2("Empty \\%c", (U8)value);
17859 if (*RExC_parse == '{') {
17860 const U8 c = (U8)value;
17861 e = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
17864 vFAIL2("Missing right brace on \\%c{}", c);
17869 /* White space is allowed adjacent to the braces and after
17870 * any '^', even when not under /x */
17871 while (isSPACE(*RExC_parse)) {
17875 if (UCHARAT(RExC_parse) == '^') {
17877 /* toggle. (The rhs xor gets the single bit that
17878 * differs between P and p; the other xor inverts just
17880 value ^= 'P' ^ 'p';
17883 while (isSPACE(*RExC_parse)) {
17888 if (e == RExC_parse)
17889 vFAIL2("Empty \\%c{}", c);
17891 n = e - RExC_parse;
17892 while (isSPACE(*(RExC_parse + n - 1)))
17895 } /* The \p isn't immediately followed by a '{' */
17896 else if (! isALPHA(*RExC_parse)) {
17897 RExC_parse += (UTF)
17898 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
17900 vFAIL2("Character following \\%c must be '{' or a "
17901 "single-character Unicode property name",
17909 char* name = RExC_parse;
17911 /* Any message returned about expanding the definition */
17912 SV* msg = newSVpvs_flags("", SVs_TEMP);
17914 /* If set TRUE, the property is user-defined as opposed to
17915 * official Unicode */
17916 bool user_defined = FALSE;
17917 AV * strings = NULL;
17919 SV * prop_definition = parse_uniprop_string(
17920 name, n, UTF, FOLD,
17921 FALSE, /* This is compile-time */
17923 /* We can't defer this defn when
17924 * the full result is required in
17926 ! cBOOL(ret_invlist),
17933 if (SvCUR(msg)) { /* Assumes any error causes a msg */
17934 assert(prop_definition == NULL);
17935 RExC_parse = e + 1;
17936 if (SvUTF8(msg)) { /* msg being UTF-8 makes the whole
17937 thing so, or else the display is
17941 /* diag_listed_as: Can't find Unicode property definition "%s" in regex; marked by <-- HERE in m/%s/ */
17942 vFAIL2utf8f("%" UTF8f, UTF8fARG(SvUTF8(msg),
17943 SvCUR(msg), SvPVX(msg)));
17946 assert(prop_definition || strings);
17950 if (! prop_definition) {
17951 RExC_parse = e + 1;
17952 vFAIL("Unicode string properties are not implemented in (?[...])");
17956 "Using just the single character results"
17957 " returned by \\p{} in (?[...])");
17960 else if (! RExC_in_multi_char_class) {
17961 if (invert ^ (value == 'P')) {
17962 RExC_parse = e + 1;
17963 vFAIL("Inverting a character class which contains"
17964 " a multi-character sequence is illegal");
17967 /* For each multi-character string ... */
17968 while (av_count(strings) > 0) {
17969 /* ... Each entry is itself an array of code
17971 AV * this_string = (AV *) av_shift( strings);
17972 STRLEN cp_count = av_count(this_string);
17973 SV * final = newSV(cp_count * 4);
17976 /* Create another string of sequences of \x{...} */
17977 while (av_count(this_string) > 0) {
17978 SV * character = av_shift(this_string);
17979 UV cp = SvUV(character);
17982 REQUIRE_UTF8(flagp);
17984 Perl_sv_catpvf(aTHX_ final, "\\x{%" UVXf "}",
17986 SvREFCNT_dec_NN(character);
17988 SvREFCNT_dec_NN(this_string);
17990 /* And add that to the list of such things */
17992 = add_multi_match(multi_char_matches,
17997 SvREFCNT_dec_NN(strings);
18000 if (! prop_definition) { /* If we got only a string,
18001 this iteration didn't really
18002 find a character */
18005 else if (! is_invlist(prop_definition)) {
18007 /* Here, the definition isn't known, so we have gotten
18008 * returned a string that will be evaluated if and when
18009 * encountered at runtime. We add it to the list of
18010 * such properties, along with whether it should be
18011 * complemented or not */
18012 if (value == 'P') {
18013 sv_catpvs(listsv, "!");
18016 sv_catpvs(listsv, "+");
18018 sv_catsv(listsv, prop_definition);
18020 has_runtime_dependency |= HAS_USER_DEFINED_PROPERTY;
18022 /* We don't know yet what this matches, so have to flag
18024 anyof_flags |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
18027 assert (prop_definition && is_invlist(prop_definition));
18029 /* Here we do have the complete property definition
18031 * Temporary workaround for [perl #133136]. For this
18032 * precise input that is in the .t that is failing,
18033 * load utf8.pm, which is what the test wants, so that
18034 * that .t passes */
18035 if ( memEQs(RExC_start, e + 1 - RExC_start,
18037 && ! hv_common(GvHVn(PL_incgv),
18039 "utf8.pm", sizeof("utf8.pm") - 1,
18040 0, HV_FETCH_ISEXISTS, NULL, 0))
18042 require_pv("utf8.pm");
18045 if (! user_defined &&
18046 /* We warn on matching an above-Unicode code point
18047 * if the match would return true, except don't
18048 * warn for \p{All}, which has exactly one element
18050 (_invlist_contains_cp(prop_definition, 0x110000)
18051 && (! (_invlist_len(prop_definition) == 1
18052 && *invlist_array(prop_definition) == 0))))
18057 /* Invert if asking for the complement */
18058 if (value == 'P') {
18059 _invlist_union_complement_2nd(properties,
18064 _invlist_union(properties, prop_definition, &properties);
18069 RExC_parse = e + 1;
18070 namedclass = ANYOF_UNIPROP; /* no official name, but it's
18074 case 'n': value = '\n'; break;
18075 case 'r': value = '\r'; break;
18076 case 't': value = '\t'; break;
18077 case 'f': value = '\f'; break;
18078 case 'b': value = '\b'; break;
18079 case 'e': value = ESC_NATIVE; break;
18080 case 'a': value = '\a'; break;
18082 RExC_parse--; /* function expects to be pointed at the 'o' */
18083 if (! grok_bslash_o(&RExC_parse,
18089 cBOOL(range), /* MAX_UV allowed for range
18095 else if (message && TO_OUTPUT_WARNINGS(RExC_parse)) {
18096 warn_non_literal_string(RExC_parse, packed_warn, message);
18100 non_portable_endpoint++;
18104 RExC_parse--; /* function expects to be pointed at the 'x' */
18105 if (! grok_bslash_x(&RExC_parse,
18111 cBOOL(range), /* MAX_UV allowed for range
18117 else if (message && TO_OUTPUT_WARNINGS(RExC_parse)) {
18118 warn_non_literal_string(RExC_parse, packed_warn, message);
18122 non_portable_endpoint++;
18126 if (! grok_bslash_c(*RExC_parse, &grok_c_char, &message,
18129 /* going to die anyway; point to exact spot of
18131 RExC_parse += (UTF)
18132 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
18137 value = grok_c_char;
18139 if (message && TO_OUTPUT_WARNINGS(RExC_parse)) {
18140 warn_non_literal_string(RExC_parse, packed_warn, message);
18143 non_portable_endpoint++;
18145 case '0': case '1': case '2': case '3': case '4':
18146 case '5': case '6': case '7':
18148 /* Take 1-3 octal digits */
18149 I32 flags = PERL_SCAN_SILENT_ILLDIGIT
18150 | PERL_SCAN_NOTIFY_ILLDIGIT;
18151 numlen = (strict) ? 4 : 3;
18152 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
18153 RExC_parse += numlen;
18156 RExC_parse += (UTF)
18157 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
18159 vFAIL("Need exactly 3 octal digits");
18161 else if ( (flags & PERL_SCAN_NOTIFY_ILLDIGIT)
18162 && RExC_parse < RExC_end
18163 && isDIGIT(*RExC_parse)
18164 && ckWARN(WARN_REGEXP))
18166 reg_warn_non_literal_string(
18168 form_alien_digit_msg(8, numlen, RExC_parse,
18169 RExC_end, UTF, FALSE));
18173 non_portable_endpoint++;
18178 /* Allow \_ to not give an error */
18179 if (isWORDCHAR(value) && value != '_') {
18181 vFAIL2("Unrecognized escape \\%c in character class",
18185 ckWARN2reg(RExC_parse,
18186 "Unrecognized escape \\%c in character class passed through",
18191 } /* End of switch on char following backslash */
18192 } /* end of handling backslash escape sequences */
18194 /* Here, we have the current token in 'value' */
18196 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
18199 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
18200 * literal, as is the character that began the false range, i.e.
18201 * the 'a' in the examples */
18203 const int w = (RExC_parse >= rangebegin)
18204 ? RExC_parse - rangebegin
18208 "False [] range \"%" UTF8f "\"",
18209 UTF8fARG(UTF, w, rangebegin));
18212 ckWARN2reg(RExC_parse,
18213 "False [] range \"%" UTF8f "\"",
18214 UTF8fARG(UTF, w, rangebegin));
18215 cp_list = add_cp_to_invlist(cp_list, '-');
18216 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
18220 range = 0; /* this was not a true range */
18221 element_count += 2; /* So counts for three values */
18224 classnum = namedclass_to_classnum(namedclass);
18226 if (LOC && namedclass < ANYOF_POSIXL_MAX
18227 #ifndef HAS_ISASCII
18228 && classnum != _CC_ASCII
18231 SV* scratch_list = NULL;
18233 /* What the Posix classes (like \w, [:space:]) match isn't
18234 * generally knowable under locale until actual match time. A
18235 * special node is used for these which has extra space for a
18236 * bitmap, with a bit reserved for each named class that is to
18237 * be matched against. (This isn't needed for \p{} and
18238 * pseudo-classes, as they are not affected by locale, and
18239 * hence are dealt with separately.) However, if a named class
18240 * and its complement are both present, then it matches
18241 * everything, and there is no runtime dependency. Odd numbers
18242 * are the complements of the next lower number, so xor works.
18243 * (Note that something like [\w\D] should match everything,
18244 * because \d should be a proper subset of \w. But rather than
18245 * trust that the locale is well behaved, we leave this to
18246 * runtime to sort out) */
18247 if (POSIXL_TEST(posixl, namedclass ^ 1)) {
18248 cp_list = _add_range_to_invlist(cp_list, 0, UV_MAX);
18249 POSIXL_ZERO(posixl);
18250 has_runtime_dependency &= ~HAS_L_RUNTIME_DEPENDENCY;
18251 anyof_flags &= ~ANYOF_MATCHES_POSIXL;
18252 continue; /* We could ignore the rest of the class, but
18253 best to parse it for any errors */
18255 else { /* Here, isn't the complement of any already parsed
18257 POSIXL_SET(posixl, namedclass);
18258 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
18259 anyof_flags |= ANYOF_MATCHES_POSIXL;
18261 /* The above-Latin1 characters are not subject to locale
18262 * rules. Just add them to the unconditionally-matched
18265 /* Get the list of the above-Latin1 code points this
18267 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
18268 PL_XPosix_ptrs[classnum],
18270 /* Odd numbers are complements,
18271 * like NDIGIT, NASCII, ... */
18272 namedclass % 2 != 0,
18274 /* Checking if 'cp_list' is NULL first saves an extra
18275 * clone. Its reference count will be decremented at the
18276 * next union, etc, or if this is the only instance, at the
18277 * end of the routine */
18279 cp_list = scratch_list;
18282 _invlist_union(cp_list, scratch_list, &cp_list);
18283 SvREFCNT_dec_NN(scratch_list);
18285 continue; /* Go get next character */
18290 /* Here, is not /l, or is a POSIX class for which /l doesn't
18291 * matter (or is a Unicode property, which is skipped here). */
18292 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
18293 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
18295 /* Here, should be \h, \H, \v, or \V. None of /d, /i
18296 * nor /l make a difference in what these match,
18297 * therefore we just add what they match to cp_list. */
18298 if (classnum != _CC_VERTSPACE) {
18299 assert( namedclass == ANYOF_HORIZWS
18300 || namedclass == ANYOF_NHORIZWS);
18302 /* It turns out that \h is just a synonym for
18304 classnum = _CC_BLANK;
18307 _invlist_union_maybe_complement_2nd(
18309 PL_XPosix_ptrs[classnum],
18310 namedclass % 2 != 0, /* Complement if odd
18311 (NHORIZWS, NVERTWS)
18316 else if ( AT_LEAST_UNI_SEMANTICS
18317 || classnum == _CC_ASCII
18318 || (DEPENDS_SEMANTICS && ( classnum == _CC_DIGIT
18319 || classnum == _CC_XDIGIT)))
18321 /* We usually have to worry about /d affecting what POSIX
18322 * classes match, with special code needed because we won't
18323 * know until runtime what all matches. But there is no
18324 * extra work needed under /u and /a; and [:ascii:] is
18325 * unaffected by /d; and :digit: and :xdigit: don't have
18326 * runtime differences under /d. So we can special case
18327 * these, and avoid some extra work below, and at runtime.
18329 _invlist_union_maybe_complement_2nd(
18331 ((AT_LEAST_ASCII_RESTRICTED)
18332 ? PL_Posix_ptrs[classnum]
18333 : PL_XPosix_ptrs[classnum]),
18334 namedclass % 2 != 0,
18337 else { /* Garden variety class. If is NUPPER, NALPHA, ...
18338 complement and use nposixes */
18339 SV** posixes_ptr = namedclass % 2 == 0
18342 _invlist_union_maybe_complement_2nd(
18344 PL_XPosix_ptrs[classnum],
18345 namedclass % 2 != 0,
18349 } /* end of namedclass \blah */
18351 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse, RExC_end);
18353 /* If 'range' is set, 'value' is the ending of a range--check its
18354 * validity. (If value isn't a single code point in the case of a
18355 * range, we should have figured that out above in the code that
18356 * catches false ranges). Later, we will handle each individual code
18357 * point in the range. If 'range' isn't set, this could be the
18358 * beginning of a range, so check for that by looking ahead to see if
18359 * the next real character to be processed is the range indicator--the
18364 /* For unicode ranges, we have to test that the Unicode as opposed
18365 * to the native values are not decreasing. (Above 255, there is
18366 * no difference between native and Unicode) */
18367 if (unicode_range && prevvalue < 255 && value < 255) {
18368 if (NATIVE_TO_LATIN1(prevvalue) > NATIVE_TO_LATIN1(value)) {
18369 goto backwards_range;
18374 if (prevvalue > value) /* b-a */ {
18379 w = RExC_parse - rangebegin;
18381 "Invalid [] range \"%" UTF8f "\"",
18382 UTF8fARG(UTF, w, rangebegin));
18383 NOT_REACHED; /* NOTREACHED */
18387 prevvalue = value; /* save the beginning of the potential range */
18388 if (! stop_at_1 /* Can't be a range if parsing just one thing */
18389 && *RExC_parse == '-')
18391 char* next_char_ptr = RExC_parse + 1;
18393 /* Get the next real char after the '-' */
18394 SKIP_BRACKETED_WHITE_SPACE(skip_white, next_char_ptr, RExC_end);
18396 /* If the '-' is at the end of the class (just before the ']',
18397 * it is a literal minus; otherwise it is a range */
18398 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
18399 RExC_parse = next_char_ptr;
18401 /* a bad range like \w-, [:word:]- ? */
18402 if (namedclass > OOB_NAMEDCLASS) {
18403 if (strict || ckWARN(WARN_REGEXP)) {
18404 const int w = RExC_parse >= rangebegin
18405 ? RExC_parse - rangebegin
18408 vFAIL4("False [] range \"%*.*s\"",
18413 "False [] range \"%*.*s\"",
18417 cp_list = add_cp_to_invlist(cp_list, '-');
18420 range = 1; /* yeah, it's a range! */
18421 continue; /* but do it the next time */
18426 if (namedclass > OOB_NAMEDCLASS) {
18430 /* Here, we have a single value this time through the loop, and
18431 * <prevvalue> is the beginning of the range, if any; or <value> if
18434 /* non-Latin1 code point implies unicode semantics. */
18436 if (value > MAX_LEGAL_CP && ( value != UV_MAX
18437 || prevvalue > MAX_LEGAL_CP))
18439 vFAIL(form_cp_too_large_msg(16, NULL, 0, value));
18441 REQUIRE_UNI_RULES(flagp, 0);
18442 if ( ! silence_non_portable
18443 && UNICODE_IS_PERL_EXTENDED(value)
18444 && TO_OUTPUT_WARNINGS(RExC_parse))
18446 ckWARN2_non_literal_string(RExC_parse,
18447 packWARN(WARN_PORTABLE),
18448 PL_extended_cp_format,
18453 /* Ready to process either the single value, or the completed range.
18454 * For single-valued non-inverted ranges, we consider the possibility
18455 * of multi-char folds. (We made a conscious decision to not do this
18456 * for the other cases because it can often lead to non-intuitive
18457 * results. For example, you have the peculiar case that:
18458 * "s s" =~ /^[^\xDF]+$/i => Y
18459 * "ss" =~ /^[^\xDF]+$/i => N
18461 * See [perl #89750] */
18462 if (FOLD && allow_mutiple_chars && value == prevvalue) {
18463 if ( value == LATIN_SMALL_LETTER_SHARP_S
18464 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
18467 /* Here <value> is indeed a multi-char fold. Get what it is */
18469 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
18472 UV folded = _to_uni_fold_flags(
18476 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
18477 ? FOLD_FLAGS_NOMIX_ASCII
18481 /* Here, <folded> should be the first character of the
18482 * multi-char fold of <value>, with <foldbuf> containing the
18483 * whole thing. But, if this fold is not allowed (because of
18484 * the flags), <fold> will be the same as <value>, and should
18485 * be processed like any other character, so skip the special
18487 if (folded != value) {
18489 /* Skip if we are recursed, currently parsing the class
18490 * again. Otherwise add this character to the list of
18491 * multi-char folds. */
18492 if (! RExC_in_multi_char_class) {
18493 STRLEN cp_count = utf8_length(foldbuf,
18494 foldbuf + foldlen);
18495 SV* multi_fold = sv_2mortal(newSVpvs(""));
18497 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%" UVXf "}", value);
18500 = add_multi_match(multi_char_matches,
18506 /* This element should not be processed further in this
18509 value = save_value;
18510 prevvalue = save_prevvalue;
18516 if (strict && ckWARN(WARN_REGEXP)) {
18519 /* If the range starts above 255, everything is portable and
18520 * likely to be so for any forseeable character set, so don't
18522 if (unicode_range && non_portable_endpoint && prevvalue < 256) {
18523 vWARN(RExC_parse, "Both or neither range ends should be Unicode");
18525 else if (prevvalue != value) {
18527 /* Under strict, ranges that stop and/or end in an ASCII
18528 * printable should have each end point be a portable value
18529 * for it (preferably like 'A', but we don't warn if it is
18530 * a (portable) Unicode name or code point), and the range
18531 * must be all digits or all letters of the same case.
18532 * Otherwise, the range is non-portable and unclear as to
18533 * what it contains */
18534 if ( (isPRINT_A(prevvalue) || isPRINT_A(value))
18535 && ( non_portable_endpoint
18536 || ! ( (isDIGIT_A(prevvalue) && isDIGIT_A(value))
18537 || (isLOWER_A(prevvalue) && isLOWER_A(value))
18538 || (isUPPER_A(prevvalue) && isUPPER_A(value))
18540 vWARN(RExC_parse, "Ranges of ASCII printables should"
18541 " be some subset of \"0-9\","
18542 " \"A-Z\", or \"a-z\"");
18544 else if (prevvalue >= FIRST_NON_ASCII_DECIMAL_DIGIT) {
18545 SSize_t index_start;
18546 SSize_t index_final;
18548 /* But the nature of Unicode and languages mean we
18549 * can't do the same checks for above-ASCII ranges,
18550 * except in the case of digit ones. These should
18551 * contain only digits from the same group of 10. The
18552 * ASCII case is handled just above. Hence here, the
18553 * range could be a range of digits. First some
18554 * unlikely special cases. Grandfather in that a range
18555 * ending in 19DA (NEW TAI LUE THAM DIGIT ONE) is bad
18556 * if its starting value is one of the 10 digits prior
18557 * to it. This is because it is an alternate way of
18558 * writing 19D1, and some people may expect it to be in
18559 * that group. But it is bad, because it won't give
18560 * the expected results. In Unicode 5.2 it was
18561 * considered to be in that group (of 11, hence), but
18562 * this was fixed in the next version */
18564 if (UNLIKELY(value == 0x19DA && prevvalue >= 0x19D0)) {
18565 goto warn_bad_digit_range;
18567 else if (UNLIKELY( prevvalue >= 0x1D7CE
18568 && value <= 0x1D7FF))
18570 /* This is the only other case currently in Unicode
18571 * where the algorithm below fails. The code
18572 * points just above are the end points of a single
18573 * range containing only decimal digits. It is 5
18574 * different series of 0-9. All other ranges of
18575 * digits currently in Unicode are just a single
18576 * series. (And mktables will notify us if a later
18577 * Unicode version breaks this.)
18579 * If the range being checked is at most 9 long,
18580 * and the digit values represented are in
18581 * numerical order, they are from the same series.
18583 if ( value - prevvalue > 9
18584 || ((( value - 0x1D7CE) % 10)
18585 <= (prevvalue - 0x1D7CE) % 10))
18587 goto warn_bad_digit_range;
18592 /* For all other ranges of digits in Unicode, the
18593 * algorithm is just to check if both end points
18594 * are in the same series, which is the same range.
18596 index_start = _invlist_search(
18597 PL_XPosix_ptrs[_CC_DIGIT],
18600 /* Warn if the range starts and ends with a digit,
18601 * and they are not in the same group of 10. */
18602 if ( index_start >= 0
18603 && ELEMENT_RANGE_MATCHES_INVLIST(index_start)
18605 _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
18606 value)) != index_start
18607 && index_final >= 0
18608 && ELEMENT_RANGE_MATCHES_INVLIST(index_final))
18610 warn_bad_digit_range:
18611 vWARN(RExC_parse, "Ranges of digits should be"
18612 " from the same group of"
18619 if ((! range || prevvalue == value) && non_portable_endpoint) {
18620 if (isPRINT_A(value)) {
18623 if (isBACKSLASHED_PUNCT(value)) {
18624 literal[d++] = '\\';
18626 literal[d++] = (char) value;
18627 literal[d++] = '\0';
18630 "\"%.*s\" is more clearly written simply as \"%s\"",
18631 (int) (RExC_parse - rangebegin),
18636 else if (isMNEMONIC_CNTRL(value)) {
18638 "\"%.*s\" is more clearly written simply as \"%s\"",
18639 (int) (RExC_parse - rangebegin),
18641 cntrl_to_mnemonic((U8) value)
18647 /* Deal with this element of the class */
18650 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
18653 /* On non-ASCII platforms, for ranges that span all of 0..255, and ones
18654 * that don't require special handling, we can just add the range like
18655 * we do for ASCII platforms */
18656 if ((UNLIKELY(prevvalue == 0) && value >= 255)
18657 || ! (prevvalue < 256
18659 || (! non_portable_endpoint
18660 && ((isLOWER_A(prevvalue) && isLOWER_A(value))
18661 || (isUPPER_A(prevvalue)
18662 && isUPPER_A(value)))))))
18664 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
18668 /* Here, requires special handling. This can be because it is a
18669 * range whose code points are considered to be Unicode, and so
18670 * must be individually translated into native, or because its a
18671 * subrange of 'A-Z' or 'a-z' which each aren't contiguous in
18672 * EBCDIC, but we have defined them to include only the "expected"
18673 * upper or lower case ASCII alphabetics. Subranges above 255 are
18674 * the same in native and Unicode, so can be added as a range */
18675 U8 start = NATIVE_TO_LATIN1(prevvalue);
18677 U8 end = (value < 256) ? NATIVE_TO_LATIN1(value) : 255;
18678 for (j = start; j <= end; j++) {
18679 cp_foldable_list = add_cp_to_invlist(cp_foldable_list, LATIN1_TO_NATIVE(j));
18682 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
18688 range = 0; /* this range (if it was one) is done now */
18689 } /* End of loop through all the text within the brackets */
18691 if ( posix_warnings && av_tindex_skip_len_mg(posix_warnings) >= 0) {
18692 output_posix_warnings(pRExC_state, posix_warnings);
18695 /* If anything in the class expands to more than one character, we have to
18696 * deal with them by building up a substitute parse string, and recursively
18697 * calling reg() on it, instead of proceeding */
18698 if (multi_char_matches) {
18699 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
18702 char *save_end = RExC_end;
18703 char *save_parse = RExC_parse;
18704 char *save_start = RExC_start;
18705 Size_t constructed_prefix_len = 0; /* This gives the length of the
18706 constructed portion of the
18707 substitute parse. */
18708 bool first_time = TRUE; /* First multi-char occurrence doesn't get
18713 /* Only one level of recursion allowed */
18714 assert(RExC_copy_start_in_constructed == RExC_precomp);
18716 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
18717 because too confusing */
18719 sv_catpvs(substitute_parse, "(?:");
18723 /* Look at the longest strings first */
18724 for (cp_count = av_tindex_skip_len_mg(multi_char_matches);
18729 if (av_exists(multi_char_matches, cp_count)) {
18730 AV** this_array_ptr;
18733 this_array_ptr = (AV**) av_fetch(multi_char_matches,
18735 while ((this_sequence = av_pop(*this_array_ptr)) !=
18738 if (! first_time) {
18739 sv_catpvs(substitute_parse, "|");
18741 first_time = FALSE;
18743 sv_catpv(substitute_parse, SvPVX(this_sequence));
18748 /* If the character class contains anything else besides these
18749 * multi-character strings, have to include it in recursive parsing */
18750 if (element_count) {
18751 bool has_l_bracket = orig_parse > RExC_start && *(orig_parse - 1) == '[';
18753 sv_catpvs(substitute_parse, "|");
18754 if (has_l_bracket) { /* Add an [ if the original had one */
18755 sv_catpvs(substitute_parse, "[");
18757 constructed_prefix_len = SvCUR(substitute_parse);
18758 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
18760 /* Put in a closing ']' to match any opening one, but not if going
18761 * off the end, as otherwise we are adding something that really
18763 if (has_l_bracket && RExC_parse < RExC_end) {
18764 sv_catpvs(substitute_parse, "]");
18768 sv_catpvs(substitute_parse, ")");
18771 /* This is a way to get the parse to skip forward a whole named
18772 * sequence instead of matching the 2nd character when it fails the
18774 sv_catpvs(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
18778 /* Set up the data structure so that any errors will be properly
18779 * reported. See the comments at the definition of
18780 * REPORT_LOCATION_ARGS for details */
18781 RExC_copy_start_in_input = (char *) orig_parse;
18782 RExC_start = RExC_parse = SvPV(substitute_parse, len);
18783 RExC_copy_start_in_constructed = RExC_start + constructed_prefix_len;
18784 RExC_end = RExC_parse + len;
18785 RExC_in_multi_char_class = 1;
18787 ret = reg(pRExC_state, 1, ®_flags, depth+1);
18789 *flagp |= reg_flags & (HASWIDTH|SIMPLE|POSTPONED|RESTART_PARSE|NEED_UTF8);
18791 /* And restore so can parse the rest of the pattern */
18792 RExC_parse = save_parse;
18793 RExC_start = RExC_copy_start_in_constructed = RExC_copy_start_in_input = save_start;
18794 RExC_end = save_end;
18795 RExC_in_multi_char_class = 0;
18796 SvREFCNT_dec_NN(multi_char_matches);
18797 SvREFCNT_dec(properties);
18798 SvREFCNT_dec(cp_list);
18799 SvREFCNT_dec(simple_posixes);
18800 SvREFCNT_dec(posixes);
18801 SvREFCNT_dec(nposixes);
18802 SvREFCNT_dec(cp_foldable_list);
18806 /* If folding, we calculate all characters that could fold to or from the
18807 * ones already on the list */
18808 if (cp_foldable_list) {
18810 UV start, end; /* End points of code point ranges */
18812 SV* fold_intersection = NULL;
18815 /* Our calculated list will be for Unicode rules. For locale
18816 * matching, we have to keep a separate list that is consulted at
18817 * runtime only when the locale indicates Unicode rules (and we
18818 * don't include potential matches in the ASCII/Latin1 range, as
18819 * any code point could fold to any other, based on the run-time
18820 * locale). For non-locale, we just use the general list */
18822 use_list = &only_utf8_locale_list;
18825 use_list = &cp_list;
18828 /* Only the characters in this class that participate in folds need
18829 * be checked. Get the intersection of this class and all the
18830 * possible characters that are foldable. This can quickly narrow
18831 * down a large class */
18832 _invlist_intersection(PL_in_some_fold, cp_foldable_list,
18833 &fold_intersection);
18835 /* Now look at the foldable characters in this class individually */
18836 invlist_iterinit(fold_intersection);
18837 while (invlist_iternext(fold_intersection, &start, &end)) {
18841 /* Look at every character in the range */
18842 for (j = start; j <= end; j++) {
18843 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
18846 Size_t folds_count;
18848 const U32 * remaining_folds;
18852 /* Under /l, we don't know what code points below 256
18853 * fold to, except we do know the MICRO SIGN folds to
18854 * an above-255 character if the locale is UTF-8, so we
18855 * add it to the special list (in *use_list) Otherwise
18856 * we know now what things can match, though some folds
18857 * are valid under /d only if the target is UTF-8.
18858 * Those go in a separate list */
18859 if ( IS_IN_SOME_FOLD_L1(j)
18860 && ! (LOC && j != MICRO_SIGN))
18863 /* ASCII is always matched; non-ASCII is matched
18864 * only under Unicode rules (which could happen
18865 * under /l if the locale is a UTF-8 one */
18866 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
18867 *use_list = add_cp_to_invlist(*use_list,
18868 PL_fold_latin1[j]);
18870 else if (j != PL_fold_latin1[j]) {
18871 upper_latin1_only_utf8_matches
18872 = add_cp_to_invlist(
18873 upper_latin1_only_utf8_matches,
18874 PL_fold_latin1[j]);
18878 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(j)
18879 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
18881 add_above_Latin1_folds(pRExC_state,
18888 /* Here is an above Latin1 character. We don't have the
18889 * rules hard-coded for it. First, get its fold. This is
18890 * the simple fold, as the multi-character folds have been
18891 * handled earlier and separated out */
18892 folded = _to_uni_fold_flags(j, foldbuf, &foldlen,
18893 (ASCII_FOLD_RESTRICTED)
18894 ? FOLD_FLAGS_NOMIX_ASCII
18897 /* Single character fold of above Latin1. Add everything
18898 * in its fold closure to the list that this node should
18900 folds_count = _inverse_folds(folded, &first_fold,
18902 for (k = 0; k <= folds_count; k++) {
18903 UV c = (k == 0) /* First time through use itself */
18905 : (k == 1) /* 2nd time use, the first fold */
18908 /* Then the remaining ones */
18909 : remaining_folds[k-2];
18911 /* /aa doesn't allow folds between ASCII and non- */
18912 if (( ASCII_FOLD_RESTRICTED
18913 && (isASCII(c) != isASCII(j))))
18918 /* Folds under /l which cross the 255/256 boundary are
18919 * added to a separate list. (These are valid only
18920 * when the locale is UTF-8.) */
18921 if (c < 256 && LOC) {
18922 *use_list = add_cp_to_invlist(*use_list, c);
18926 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
18928 cp_list = add_cp_to_invlist(cp_list, c);
18931 /* Similarly folds involving non-ascii Latin1
18932 * characters under /d are added to their list */
18933 upper_latin1_only_utf8_matches
18934 = add_cp_to_invlist(
18935 upper_latin1_only_utf8_matches,
18941 SvREFCNT_dec_NN(fold_intersection);
18944 /* Now that we have finished adding all the folds, there is no reason
18945 * to keep the foldable list separate */
18946 _invlist_union(cp_list, cp_foldable_list, &cp_list);
18947 SvREFCNT_dec_NN(cp_foldable_list);
18950 /* And combine the result (if any) with any inversion lists from posix
18951 * classes. The lists are kept separate up to now because we don't want to
18952 * fold the classes */
18953 if (simple_posixes) { /* These are the classes known to be unaffected by
18956 _invlist_union(cp_list, simple_posixes, &cp_list);
18957 SvREFCNT_dec_NN(simple_posixes);
18960 cp_list = simple_posixes;
18963 if (posixes || nposixes) {
18964 if (! DEPENDS_SEMANTICS) {
18966 /* For everything but /d, we can just add the current 'posixes' and
18967 * 'nposixes' to the main list */
18970 _invlist_union(cp_list, posixes, &cp_list);
18971 SvREFCNT_dec_NN(posixes);
18979 _invlist_union(cp_list, nposixes, &cp_list);
18980 SvREFCNT_dec_NN(nposixes);
18983 cp_list = nposixes;
18988 /* Under /d, things like \w match upper Latin1 characters only if
18989 * the target string is in UTF-8. But things like \W match all the
18990 * upper Latin1 characters if the target string is not in UTF-8.
18992 * Handle the case with something like \W separately */
18994 SV* only_non_utf8_list = invlist_clone(PL_UpperLatin1, NULL);
18996 /* A complemented posix class matches all upper Latin1
18997 * characters if not in UTF-8. And it matches just certain
18998 * ones when in UTF-8. That means those certain ones are
18999 * matched regardless, so can just be added to the
19000 * unconditional list */
19002 _invlist_union(cp_list, nposixes, &cp_list);
19003 SvREFCNT_dec_NN(nposixes);
19007 cp_list = nposixes;
19010 /* Likewise for 'posixes' */
19011 _invlist_union(posixes, cp_list, &cp_list);
19012 SvREFCNT_dec(posixes);
19014 /* Likewise for anything else in the range that matched only
19016 if (upper_latin1_only_utf8_matches) {
19017 _invlist_union(cp_list,
19018 upper_latin1_only_utf8_matches,
19020 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
19021 upper_latin1_only_utf8_matches = NULL;
19024 /* If we don't match all the upper Latin1 characters regardless
19025 * of UTF-8ness, we have to set a flag to match the rest when
19027 _invlist_subtract(only_non_utf8_list, cp_list,
19028 &only_non_utf8_list);
19029 if (_invlist_len(only_non_utf8_list) != 0) {
19030 anyof_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
19032 SvREFCNT_dec_NN(only_non_utf8_list);
19035 /* Here there were no complemented posix classes. That means
19036 * the upper Latin1 characters in 'posixes' match only when the
19037 * target string is in UTF-8. So we have to add them to the
19038 * list of those types of code points, while adding the
19039 * remainder to the unconditional list.
19041 * First calculate what they are */
19042 SV* nonascii_but_latin1_properties = NULL;
19043 _invlist_intersection(posixes, PL_UpperLatin1,
19044 &nonascii_but_latin1_properties);
19046 /* And add them to the final list of such characters. */
19047 _invlist_union(upper_latin1_only_utf8_matches,
19048 nonascii_but_latin1_properties,
19049 &upper_latin1_only_utf8_matches);
19051 /* Remove them from what now becomes the unconditional list */
19052 _invlist_subtract(posixes, nonascii_but_latin1_properties,
19055 /* And add those unconditional ones to the final list */
19057 _invlist_union(cp_list, posixes, &cp_list);
19058 SvREFCNT_dec_NN(posixes);
19065 SvREFCNT_dec(nonascii_but_latin1_properties);
19067 /* Get rid of any characters from the conditional list that we
19068 * now know are matched unconditionally, which may make that
19070 _invlist_subtract(upper_latin1_only_utf8_matches,
19072 &upper_latin1_only_utf8_matches);
19073 if (_invlist_len(upper_latin1_only_utf8_matches) == 0) {
19074 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
19075 upper_latin1_only_utf8_matches = NULL;
19081 /* And combine the result (if any) with any inversion list from properties.
19082 * The lists are kept separate up to now so that we can distinguish the two
19083 * in regards to matching above-Unicode. A run-time warning is generated
19084 * if a Unicode property is matched against a non-Unicode code point. But,
19085 * we allow user-defined properties to match anything, without any warning,
19086 * and we also suppress the warning if there is a portion of the character
19087 * class that isn't a Unicode property, and which matches above Unicode, \W
19088 * or [\x{110000}] for example.
19089 * (Note that in this case, unlike the Posix one above, there is no
19090 * <upper_latin1_only_utf8_matches>, because having a Unicode property
19091 * forces Unicode semantics */
19095 /* If it matters to the final outcome, see if a non-property
19096 * component of the class matches above Unicode. If so, the
19097 * warning gets suppressed. This is true even if just a single
19098 * such code point is specified, as, though not strictly correct if
19099 * another such code point is matched against, the fact that they
19100 * are using above-Unicode code points indicates they should know
19101 * the issues involved */
19103 warn_super = ! (invert
19104 ^ (UNICODE_IS_SUPER(invlist_highest(cp_list))));
19107 _invlist_union(properties, cp_list, &cp_list);
19108 SvREFCNT_dec_NN(properties);
19111 cp_list = properties;
19116 |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
19118 /* Because an ANYOF node is the only one that warns, this node
19119 * can't be optimized into something else */
19120 optimizable = FALSE;
19124 /* Here, we have calculated what code points should be in the character
19127 * Now we can see about various optimizations. Fold calculation (which we
19128 * did above) needs to take place before inversion. Otherwise /[^k]/i
19129 * would invert to include K, which under /i would match k, which it
19130 * shouldn't. Therefore we can't invert folded locale now, as it won't be
19131 * folded until runtime */
19133 /* If we didn't do folding, it's because some information isn't available
19134 * until runtime; set the run-time fold flag for these We know to set the
19135 * flag if we have a non-NULL list for UTF-8 locales, or the class matches
19136 * at least one 0-255 range code point */
19139 /* Some things on the list might be unconditionally included because of
19140 * other components. Remove them, and clean up the list if it goes to
19142 if (only_utf8_locale_list && cp_list) {
19143 _invlist_subtract(only_utf8_locale_list, cp_list,
19144 &only_utf8_locale_list);
19146 if (_invlist_len(only_utf8_locale_list) == 0) {
19147 SvREFCNT_dec_NN(only_utf8_locale_list);
19148 only_utf8_locale_list = NULL;
19151 if ( only_utf8_locale_list
19152 || (cp_list && ( _invlist_contains_cp(cp_list, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE)
19153 || _invlist_contains_cp(cp_list, LATIN_SMALL_LETTER_DOTLESS_I))))
19155 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
19158 | ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
19160 else if (cp_list && invlist_lowest(cp_list) < 256) {
19161 /* If nothing is below 256, has no locale dependency; otherwise it
19163 anyof_flags |= ANYOFL_FOLD;
19164 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
19167 else if ( DEPENDS_SEMANTICS
19168 && ( upper_latin1_only_utf8_matches
19169 || (anyof_flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)))
19171 RExC_seen_d_op = TRUE;
19172 has_runtime_dependency |= HAS_D_RUNTIME_DEPENDENCY;
19175 /* Optimize inverted patterns (e.g. [^a-z]) when everything is known at
19179 && ! has_runtime_dependency)
19181 _invlist_invert(cp_list);
19183 /* Clear the invert flag since have just done it here */
19187 /* All possible optimizations below still have these characteristics.
19188 * (Multi-char folds aren't SIMPLE, but they don't get this far in this
19190 *flagp |= HASWIDTH|SIMPLE;
19193 *ret_invlist = cp_list;
19195 return (cp_list) ? RExC_emit : 0;
19198 if (anyof_flags & ANYOF_LOCALE_FLAGS) {
19199 RExC_contains_locale = 1;
19204 /* Some character classes are equivalent to other nodes. Such nodes
19205 * take up less room, and some nodes require fewer operations to
19206 * execute, than ANYOF nodes. EXACTish nodes may be joinable with
19207 * adjacent nodes to improve efficiency. */
19208 op = optimize_regclass(pRExC_state, cp_list,
19209 only_utf8_locale_list,
19210 upper_latin1_only_utf8_matches,
19211 has_runtime_dependency,
19213 &anyof_flags, &invert, &ret, flagp);
19214 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
19216 /* If optimized to something else and emitted, clean up and return */
19218 Set_Node_Offset_Length(REGNODE_p(ret), orig_parse - RExC_start,
19219 RExC_parse - orig_parse);;
19220 SvREFCNT_dec(cp_list);;
19221 SvREFCNT_dec(only_utf8_locale_list);
19222 SvREFCNT_dec(upper_latin1_only_utf8_matches);
19227 /* Here are going to emit an ANYOF; set the particular type */
19229 if (has_runtime_dependency & HAS_D_RUNTIME_DEPENDENCY) {
19240 ret = regnode_guts(pRExC_state, op, regarglen[op], "anyof");
19241 FILL_NODE(ret, op); /* We set the argument later */
19242 RExC_emit += 1 + regarglen[op];
19243 ANYOF_FLAGS(REGNODE_p(ret)) = anyof_flags;
19245 /* Here, <cp_list> contains all the code points we can determine at
19246 * compile time that match under all conditions. Go through it, and
19247 * for things that belong in the bitmap, put them there, and delete from
19248 * <cp_list>. While we are at it, see if everything above 255 is in the
19249 * list, and if so, set a flag to speed up execution */
19251 populate_ANYOF_from_invlist(REGNODE_p(ret), &cp_list);
19254 ANYOF_POSIXL_SET_TO_BITMAP(REGNODE_p(ret), posixl);
19258 ANYOF_FLAGS(REGNODE_p(ret)) |= ANYOF_INVERT;
19261 /* Here, the bitmap has been populated with all the Latin1 code points that
19262 * always match. Can now add to the overall list those that match only
19263 * when the target string is UTF-8 (<upper_latin1_only_utf8_matches>).
19265 if (upper_latin1_only_utf8_matches) {
19267 _invlist_union(cp_list,
19268 upper_latin1_only_utf8_matches,
19270 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
19273 cp_list = upper_latin1_only_utf8_matches;
19275 ANYOF_FLAGS(REGNODE_p(ret)) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
19278 set_ANYOF_arg(pRExC_state, REGNODE_p(ret), cp_list,
19279 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
19282 only_utf8_locale_list);
19284 SvREFCNT_dec(cp_list);;
19285 SvREFCNT_dec(only_utf8_locale_list);
19290 S_optimize_regclass(pTHX_
19291 RExC_state_t *pRExC_state,
19293 SV* only_utf8_locale_list,
19294 SV* upper_latin1_only_utf8_matches,
19295 const U32 has_runtime_dependency,
19299 regnode_offset * ret,
19303 /* This function exists just to make S_regclass() smaller. It extracts out
19304 * the code that looks for potential optimizations away from a full generic
19305 * ANYOF node. The parameter names are the same as the corresponding
19306 * variables in S_regclass.
19308 * It returns the new op (ANYOF if no optimization found) and sets *ret to
19309 * any created regnode. If the new op is sufficiently like plain ANYOF, it
19310 * leaves *ret unchanged for allocation in S_regclass.
19312 * Certain of the parameters may be updated as a result of the changes
19315 U8 op = ANYOF; /* The returned node-type, initialized to the unoptimized
19318 PERL_UINT_FAST8_T i;
19319 UV partial_cp_count = 0;
19320 UV start[MAX_FOLD_FROMS+1] = { 0 }; /* +1 for the folded-to char */
19321 UV end[MAX_FOLD_FROMS+1] = { 0 };
19322 bool single_range = FALSE;
19324 PERL_ARGS_ASSERT_OPTIMIZE_REGCLASS;
19326 if (cp_list) { /* Count the code points in enough ranges that we would see
19327 all the ones possible in any fold in this version of
19330 invlist_iterinit(cp_list);
19331 for (i = 0; i <= MAX_FOLD_FROMS; i++) {
19332 if (! invlist_iternext(cp_list, &start[i], &end[i])) {
19335 partial_cp_count += end[i] - start[i] + 1;
19339 single_range = TRUE;
19341 invlist_iterfinish(cp_list);
19344 /* If we know at compile time that this matches every possible code point,
19345 * any run-time dependencies don't matter */
19346 if (start[0] == 0 && end[0] == UV_MAX) {
19349 *ret = reganode(pRExC_state, op, 0);
19353 *ret = reg_node(pRExC_state, op);
19359 /* Similarly, for /l posix classes, if both a class and its complement
19360 * match, any run-time dependencies don't matter */
19363 for (namedclass = 0; namedclass < ANYOF_POSIXL_MAX; namedclass += 2) {
19364 if ( POSIXL_TEST(posixl, namedclass) /* class */
19365 && POSIXL_TEST(posixl, namedclass + 1)) /* its complement */
19369 *ret = reganode(pRExC_state, op, 0);
19373 *ret = reg_node(pRExC_state, op);
19380 /* For well-behaved locales, some classes are subsets of others, so
19381 * complementing the subset and including the non-complemented superset
19382 * should match everything, like [\D[:alnum:]], and
19383 * [[:^alpha:][:alnum:]], but some implementations of locales are
19384 * buggy, and khw thinks its a bad idea to have optimization change
19385 * behavior, even if it avoids an OS bug in a given case */
19387 #define isSINGLE_BIT_SET(n) isPOWER_OF_2(n)
19389 /* If is a single posix /l class, can optimize to just that op. Such a
19390 * node will not match anything in the Latin1 range, as that is not
19391 * determinable until runtime, but will match whatever the class does
19392 * outside that range. (Note that some classes won't match anything
19393 * outside the range, like [:ascii:]) */
19394 if ( isSINGLE_BIT_SET(posixl)
19395 && (partial_cp_count == 0 || start[0] > 255))
19398 SV * class_above_latin1 = NULL;
19399 bool already_inverted;
19400 bool are_equivalent;
19403 namedclass = single_1bit_pos32(posixl);
19404 classnum = namedclass_to_classnum(namedclass);
19406 /* The named classes are such that the inverted number is one
19407 * larger than the non-inverted one */
19408 already_inverted = namedclass - classnum_to_namedclass(classnum);
19410 /* Create an inversion list of the official property, inverted if
19411 * the constructed node list is inverted, and restricted to only
19412 * the above latin1 code points, which are the only ones known at
19414 _invlist_intersection_maybe_complement_2nd(
19416 PL_XPosix_ptrs[classnum],
19418 &class_above_latin1);
19419 are_equivalent = _invlistEQ(class_above_latin1, cp_list, FALSE);
19420 SvREFCNT_dec_NN(class_above_latin1);
19422 if (are_equivalent) {
19424 /* Resolve the run-time inversion flag with this possibly
19425 * inverted class */
19426 *invert = *invert ^ already_inverted;
19428 op = POSIXL + *invert * (NPOSIXL - POSIXL);
19429 *ret = reg_node(pRExC_state, op);
19430 FLAGS(REGNODE_p(*ret)) = classnum;
19436 /* khw can't think of any other possible transformation involving these. */
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 example,
19444 * when a Unicode property that doesn't match anything is the only
19445 * element in the character class (perluniprops.pod notes such
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 optimizer
19474 * join this node with adjacent EXACTish ones, and ANYOF nodes require
19475 * runtime conversion to code point from UTF-8, which we'd like to avoid.
19477 * An EXACTFish node can be generated even if not under /i, and vice versa.
19478 * But care must be taken. An EXACTFish node has to be such that it only
19479 * matches precisely the code points in the class, but we want to generate
19480 * the least restrictive one that does that, to increase the odds of being
19481 * able to join with an adjacent node. For example, if the class contains
19482 * [kK], we have to make it an EXACTFAA node to prevent the KELVIN SIGN
19483 * from matching. Whether we are under /i or not is irrelevant in this
19484 * case. Less obvious is the pattern qr/[\x{02BC}]n/i. U+02BC is MODIFIER
19485 * LETTER APOSTROPHE. That is supposed to match the single character U+0149
19486 * LATIN SMALL LETTER N PRECEDED BY APOSTROPHE. And so even though there
19487 * is no simple fold that includes \X{02BC}, there is a multi-char fold
19488 * that does, and so the node generated for it must be an EXACTFish one.
19489 * On the other hand qr/:/i should generate a plain EXACT node since the
19490 * colon participates in no fold whatsoever, and having it be EXACT tells
19491 * the optimizer the target string cannot match unless it has a colon in
19496 /* Only try if there are no more code points in the class than in
19497 * the max possible fold */
19498 && inRANGE(partial_cp_count, 1, MAX_FOLD_FROMS + 1))
19500 /* We can always make a single code point class into an EXACTish node.
19502 if (partial_cp_count == 1 && ! upper_latin1_only_utf8_matches) {
19505 /* Here is /l: Use EXACTL, except if there is a fold not known
19506 * until runtime so shows as only a single code point here.
19507 * For code points above 255, we know which can cause problems
19508 * by having a potential fold to the Latin1 range. */
19510 || ( start[0] > 255
19511 && ! is_PROBLEMATIC_LOCALE_FOLD_cp(start[0])))
19519 else if (! FOLD) { /* Not /l and not /i */
19520 op = (start[0] < 256) ? EXACT : EXACT_REQ8;
19522 else if (start[0] < 256) { /* /i, not /l, and the code point is
19525 /* Under /i, it gets a little tricky. A code point that
19526 * doesn't participate in a fold should be an EXACT node. We
19527 * know this one isn't the result of a simple fold, or there'd
19528 * be more than one code point in the list, but it could be
19529 * part of a multi-character fold. In that case we better not
19530 * create an EXACT node, as we would wrongly be telling the
19531 * optimizer that this code point must be in the target string,
19532 * and that is wrong. This is because if the sequence around
19533 * this code point forms a multi-char fold, what needs to be in
19534 * the string could be the code point that folds to the
19537 * This handles the case of below-255 code points, as we have
19538 * an easy look up for those. The next clause handles the
19540 op = IS_IN_SOME_FOLD_L1(start[0])
19544 else { /* /i, larger code point. Since we are under /i, and have
19545 just this code point, we know that it can't fold to
19546 something else, so PL_InMultiCharFold applies to it */
19547 op = (_invlist_contains_cp(PL_InMultiCharFold, start[0]))
19554 else if ( ! (has_runtime_dependency & ~HAS_D_RUNTIME_DEPENDENCY)
19555 && _invlist_contains_cp(PL_in_some_fold, start[0]))
19557 /* Here, the only runtime dependency, if any, is from /d, and the
19558 * class matches more than one code point, and the lowest code
19559 * point participates in some fold. It might be that the other
19560 * code points are /i equivalent to this one, and hence they would
19561 * be representable by an EXACTFish node. Above, we eliminated
19562 * classes that contain too many code points to be EXACTFish, with
19563 * the test for MAX_FOLD_FROMS
19565 * First, special case the ASCII fold pairs, like 'B' and 'b'. We
19566 * do this because we have EXACTFAA at our disposal for the ASCII
19568 if (partial_cp_count == 2 && isASCII(start[0])) {
19570 /* The only ASCII characters that participate in folds are
19572 assert(isALPHA(start[0]));
19573 if ( end[0] == start[0] /* First range is a single
19574 character, so 2nd exists */
19575 && isALPHA_FOLD_EQ(start[0], start[1]))
19577 /* Here, is part of an ASCII fold pair */
19579 if ( ASCII_FOLD_RESTRICTED
19580 || HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(start[0]))
19582 /* If the second clause just above was true, it means
19583 * we can't be under /i, or else the list would have
19584 * included more than this fold pair. Therefore we
19585 * have to exclude the possibility of whatever else it
19586 * is that folds to these, by using EXACTFAA */
19589 else if (HAS_NONLATIN1_FOLD_CLOSURE(start[0])) {
19591 /* Here, there's no simple fold that start[0] is part
19592 * of, but there is a multi-character one. If we are
19593 * not under /i, we want to exclude that possibility;
19594 * if under /i, we want to include it */
19595 op = (FOLD) ? EXACTFU : EXACTFAA;
19599 /* Here, the only possible fold start[0] particpates in
19600 * is with start[1]. /i or not isn't relevant */
19604 value = toFOLD(start[0]);
19607 else if ( ! upper_latin1_only_utf8_matches
19608 || ( _invlist_len(upper_latin1_only_utf8_matches) == 2
19610 invlist_highest(upper_latin1_only_utf8_matches)]
19613 /* Here, the smallest character is non-ascii or there are more
19614 * than 2 code points matched by this node. Also, we either
19615 * don't have /d UTF-8 dependent matches, or if we do, they
19616 * look like they could be a single character that is the fold
19617 * of the lowest one is in the always-match list. This test
19618 * quickly excludes most of the false positives when there are
19619 * /d UTF-8 depdendent matches. These are like LATIN CAPITAL
19620 * LETTER A WITH GRAVE matching LATIN SMALL LETTER A WITH GRAVE
19621 * iff the target string is UTF-8. (We don't have to worry
19622 * above about exceeding the array bounds of PL_fold_latin1[]
19623 * because any code point in 'upper_latin1_only_utf8_matches'
19626 * EXACTFAA would apply only to pairs (hence exactly 2 code
19627 * points) in the ASCII range, so we can't use it here to
19628 * artificially restrict the fold domain, so we check if the
19629 * class does or does not match some EXACTFish node. Further,
19630 * if we aren't under /i, and and the folded-to character is
19631 * part of a multi-character fold, we can't do this
19632 * optimization, as the sequence around it could be that
19633 * multi-character fold, and we don't here know the context, so
19634 * we have to assume it is that multi-char fold, to prevent
19637 * To do the general case, we first find the fold of the lowest
19638 * code point (which may be higher than that lowest unfolded
19639 * one), then find everything that folds to it. (The data
19640 * structure we have only maps from the folded code points, so
19641 * we have to do the earlier step.) */
19644 U8 foldbuf[UTF8_MAXBYTES_CASE];
19645 UV folded = _to_uni_fold_flags(start[0], foldbuf, &foldlen, 0);
19647 const U32 * remaining_folds;
19648 Size_t folds_to_this_cp_count = _inverse_folds(
19652 Size_t folds_count = folds_to_this_cp_count + 1;
19653 SV * fold_list = _new_invlist(folds_count);
19656 /* If there are UTF-8 dependent matches, create a temporary
19657 * list of what this node matches, including them. */
19658 SV * all_cp_list = NULL;
19659 SV ** use_this_list = &cp_list;
19661 if (upper_latin1_only_utf8_matches) {
19662 all_cp_list = _new_invlist(0);
19663 use_this_list = &all_cp_list;
19664 _invlist_union(cp_list,
19665 upper_latin1_only_utf8_matches,
19669 /* Having gotten everything that participates in the fold
19670 * containing the lowest code point, we turn that into an
19671 * inversion list, making sure everything is included. */
19672 fold_list = add_cp_to_invlist(fold_list, start[0]);
19673 fold_list = add_cp_to_invlist(fold_list, folded);
19674 if (folds_to_this_cp_count > 0) {
19675 fold_list = add_cp_to_invlist(fold_list, first_fold);
19676 for (i = 0; i + 1 < folds_to_this_cp_count; i++) {
19677 fold_list = add_cp_to_invlist(fold_list,
19678 remaining_folds[i]);
19682 /* If the fold list is identical to what's in this ANYOF node,
19683 * the node can be represented by an EXACTFish one instead */
19684 if (_invlistEQ(*use_this_list, fold_list,
19685 0 /* Don't complement */ )
19688 /* But, we have to be careful, as mentioned above. Just
19689 * the right sequence of characters could match this if it
19690 * is part of a multi-character fold. That IS what we want
19691 * if we are under /i. But it ISN'T what we want if not
19692 * under /i, as it could match when it shouldn't. So, when
19693 * we aren't under /i and this character participates in a
19694 * multi-char fold, we don't optimize into an EXACTFish
19695 * node. So, for each case below we have to check if we
19696 * are folding, and if not, if it is not part of a
19697 * multi-char fold. */
19698 if (start[0] > 255) { /* Highish code point */
19699 if (FOLD || ! _invlist_contains_cp(
19700 PL_InMultiCharFold, folded))
19704 : (ASCII_FOLD_RESTRICTED)
19709 } /* Below, the lowest code point < 256 */
19712 && DEPENDS_SEMANTICS)
19713 { /* An EXACTF node containing a single character 's',
19714 can be an EXACTFU if it doesn't get joined with an
19716 op = EXACTFU_S_EDGE;
19720 || ! HAS_NONLATIN1_FOLD_CLOSURE(start[0]))
19722 if (upper_latin1_only_utf8_matches) {
19725 /* We can't use the fold, as that only matches
19729 else if ( UNLIKELY(start[0] == MICRO_SIGN)
19731 { /* EXACTFUP is a special node for this character */
19732 op = (ASCII_FOLD_RESTRICTED)
19735 value = MICRO_SIGN;
19737 else if ( ASCII_FOLD_RESTRICTED
19738 && ! isASCII(start[0]))
19739 { /* For ASCII under /iaa, we can use EXACTFU below
19751 SvREFCNT_dec_NN(fold_list);
19752 SvREFCNT_dec(all_cp_list);
19759 /* Here, we have calculated what EXACTish node to use. Have to
19760 * convert to UTF-8 if not already there */
19763 SvREFCNT_dec(cp_list);;
19764 REQUIRE_UTF8(flagp);
19767 /* This is a kludge to the special casing issues with this
19768 * ligature under /aa. FB05 should fold to FB06, but the call
19769 * above to _to_uni_fold_flags() didn't find this, as it didn't
19770 * use the /aa restriction in order to not miss other folds
19771 * that would be affected. This is the only instance likely to
19772 * ever be a problem in all of Unicode. So special case it. */
19773 if ( value == LATIN_SMALL_LIGATURE_LONG_S_T
19774 && ASCII_FOLD_RESTRICTED)
19776 value = LATIN_SMALL_LIGATURE_ST;
19780 len = (UTF) ? UVCHR_SKIP(value) : 1;
19782 *ret = regnode_guts(pRExC_state, op, len, "exact");
19783 FILL_NODE(*ret, op);
19784 RExC_emit += 1 + STR_SZ(len);
19785 setSTR_LEN(REGNODE_p(*ret), len);
19787 *STRINGs(REGNODE_p(*ret)) = (U8) value;
19790 uvchr_to_utf8((U8 *) STRINGs(REGNODE_p(*ret)), value);
19796 if (! has_runtime_dependency) {
19798 /* See if this can be turned into an ANYOFM node. Think about the bit
19799 * patterns in two different bytes. In some positions, the bits in
19800 * each will be 1; and in other positions both will be 0; and in some
19801 * positions the bit will be 1 in one byte, and 0 in the other. Let
19802 * 'n' be the number of positions where the bits differ. We create a
19803 * mask which has exactly 'n' 0 bits, each in a position where the two
19804 * bytes differ. Now take the set of all bytes that when ANDed with
19805 * the mask yield the same result. That set has 2**n elements, and is
19806 * representable by just two 8 bit numbers: the result and the mask.
19807 * Importantly, matching the set can be vectorized by creating a word
19808 * full of the result bytes, and a word full of the mask bytes,
19809 * yielding a significant speed up. Here, see if this node matches
19810 * such a set. As a concrete example consider [01], and the byte
19811 * representing '0' which is 0x30 on ASCII machines. It has the bits
19812 * 0011 0000. Take the mask 1111 1110. If we AND 0x31 and 0x30 with
19813 * that mask we get 0x30. Any other bytes ANDed yield something else.
19814 * So [01], which is a common usage, is optimizable into ANYOFM, and
19815 * can benefit from the speed up. We can only do this on UTF-8
19816 * invariant bytes, because they have the same bit patterns under UTF-8
19818 PERL_UINT_FAST8_T inverted = 0;
19820 /* Highest possible UTF-8 invariant is 7F on ASCII platforms; FF on
19822 const PERL_UINT_FAST8_T max_permissible
19823 = nBIT_UMAX(7 + ONE_IF_EBCDIC_ZERO_IF_NOT);
19825 /* If doesn't fit the criteria for ANYOFM, invert and try again. If
19826 * that works we will instead later generate an NANYOFM, and invert
19827 * back when through */
19828 if (invlist_highest(cp_list) > max_permissible) {
19829 _invlist_invert(cp_list);
19833 if (invlist_highest(cp_list) <= max_permissible) {
19834 UV this_start, this_end;
19835 UV lowest_cp = UV_MAX; /* init'ed to suppress compiler warn */
19836 U8 bits_differing = 0;
19837 Size_t full_cp_count = 0;
19838 bool first_time = TRUE;
19840 /* Go through the bytes and find the bit positions that differ */
19841 invlist_iterinit(cp_list);
19842 while (invlist_iternext(cp_list, &this_start, &this_end)) {
19843 unsigned int i = this_start;
19846 if (! UVCHR_IS_INVARIANT(i)) {
19850 first_time = FALSE;
19851 lowest_cp = this_start;
19853 /* We have set up the code point to compare with. Don't
19854 * compare it with itself */
19858 /* Find the bit positions that differ from the lowest code
19859 * point in the node. Keep track of all such positions by
19861 for (; i <= this_end; i++) {
19862 if (! UVCHR_IS_INVARIANT(i)) {
19866 bits_differing |= i ^ lowest_cp;
19869 full_cp_count += this_end - this_start + 1;
19872 /* At the end of the loop, we count how many bits differ from the
19873 * bits in lowest code point, call the count 'd'. If the set we
19874 * found contains 2**d elements, it is the closure of all code
19875 * points that differ only in those bit positions. To convince
19876 * yourself of that, first note that the number in the closure must
19877 * be a power of 2, which we test for. The only way we could have
19878 * that count and it be some differing set, is if we got some code
19879 * points that don't differ from the lowest code point in any
19880 * position, but do differ from each other in some other position.
19881 * That means one code point has a 1 in that position, and another
19882 * has a 0. But that would mean that one of them differs from the
19883 * lowest code point in that position, which possibility we've
19884 * already excluded. */
19885 if ( (inverted || full_cp_count > 1)
19886 && full_cp_count == 1U << PL_bitcount[bits_differing])
19890 op = ANYOFM + inverted;;
19892 /* We need to make the bits that differ be 0's */
19893 ANYOFM_mask = ~ bits_differing; /* This goes into FLAGS */
19895 /* The argument is the lowest code point */
19896 *ret = reganode(pRExC_state, op, lowest_cp);
19897 FLAGS(REGNODE_p(*ret)) = ANYOFM_mask;
19901 invlist_iterfinish(cp_list);
19905 _invlist_invert(cp_list);
19912 /* XXX We could create an ANYOFR_LOW node here if we saved above if all
19913 * were invariants, it wasn't inverted, and there is a single range.
19914 * This would be faster than some of the posix nodes we create below
19915 * like /\d/a, but would be twice the size. Without having actually
19916 * measured the gain, khw doesn't think the tradeoff is really worth it
19920 if (! (*anyof_flags & ANYOF_LOCALE_FLAGS)) {
19921 PERL_UINT_FAST8_T type;
19922 SV * intersection = NULL;
19923 SV* d_invlist = NULL;
19925 /* See if this matches any of the POSIX classes. The POSIXA and POSIXD
19926 * ones are about the same speed as ANYOF ops, but take less room; the
19927 * ones that have above-Latin1 code point matches are somewhat faster
19930 for (type = POSIXA; type >= POSIXD; type--) {
19933 if (type == POSIXL) { /* But not /l posix classes */
19937 for (posix_class = 0;
19938 posix_class <= _HIGHEST_REGCOMP_DOT_H_SYNC;
19941 SV** our_code_points = &cp_list;
19942 SV** official_code_points;
19945 if (type == POSIXA) {
19946 official_code_points = &PL_Posix_ptrs[posix_class];
19949 official_code_points = &PL_XPosix_ptrs[posix_class];
19952 /* Skip non-existent classes of this type. e.g. \v only has an
19953 * entry in PL_XPosix_ptrs */
19954 if (! *official_code_points) {
19958 /* Try both the regular class, and its inversion */
19959 for (try_inverted = 0; try_inverted < 2; try_inverted++) {
19960 bool this_inverted = *invert ^ try_inverted;
19962 if (type != POSIXD) {
19964 /* This class that isn't /d can't match if we have /d
19966 if (has_runtime_dependency
19967 & HAS_D_RUNTIME_DEPENDENCY)
19972 else /* is /d */ if (! this_inverted) {
19974 /* /d classes don't match anything non-ASCII below 256
19975 * unconditionally (which cp_list contains) */
19976 _invlist_intersection(cp_list, PL_UpperLatin1,
19978 if (_invlist_len(intersection) != 0) {
19982 SvREFCNT_dec(d_invlist);
19983 d_invlist = invlist_clone(cp_list, NULL);
19985 /* But under UTF-8 it turns into using /u rules. Add
19986 * the things it matches under these conditions so that
19987 * we check below that these are identical to what the
19988 * tested class should match */
19989 if (upper_latin1_only_utf8_matches) {
19992 upper_latin1_only_utf8_matches,
19995 our_code_points = &d_invlist;
19997 else { /* POSIXD, inverted. If this doesn't have this
19998 flag set, it isn't /d. */
19999 if (! (*anyof_flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
20003 our_code_points = &cp_list;
20006 /* Here, have weeded out some things. We want to see if
20007 * the list of characters this node contains
20008 * ('*our_code_points') precisely matches those of the
20009 * class we are currently checking against
20010 * ('*official_code_points'). */
20011 if (_invlistEQ(*our_code_points,
20012 *official_code_points,
20015 /* Here, they precisely match. Optimize this ANYOF
20016 * node into its equivalent POSIX one of the correct
20017 * type, possibly inverted */
20018 op = (try_inverted)
20019 ? type + NPOSIXA - POSIXA
20021 *ret = reg_node(pRExC_state, op);
20022 FLAGS(REGNODE_p(*ret)) = posix_class;
20023 SvREFCNT_dec(d_invlist);
20024 SvREFCNT_dec(intersection);
20030 SvREFCNT_dec(d_invlist);
20031 SvREFCNT_dec(intersection);
20034 /* If it is a single contiguous range, ANYOFR is an efficient regnode, both
20035 * in size and speed. Currently, a 20 bit range base (smallest code point
20036 * in the range), and a 12 bit maximum delta are packed into a 32 bit word.
20037 * This allows for using it on all of the Unicode code points except for
20038 * the highest plane, which is only for private use code points. khw
20039 * doubts that a bigger delta is likely in real world applications */
20041 && ! has_runtime_dependency
20042 && *anyof_flags == 0
20043 && start[0] < (1 << ANYOFR_BASE_BITS)
20044 && end[0] - start[0]
20045 < ((1U << (sizeof(((struct regnode_1 *)NULL)->arg1)
20046 * CHARBITS - ANYOFR_BASE_BITS))))
20049 U8 low_utf8[UTF8_MAXBYTES+1];
20050 U8 high_utf8[UTF8_MAXBYTES+1];
20053 *ret = reganode(pRExC_state, op,
20054 (start[0] | (end[0] - start[0]) << ANYOFR_BASE_BITS));
20056 /* Place the lowest UTF-8 start byte in the flags field, so as to allow
20057 * efficient ruling out at run time of many possible inputs. */
20058 (void) uvchr_to_utf8(low_utf8, start[0]);
20059 (void) uvchr_to_utf8(high_utf8, end[0]);
20061 /* If all code points share the same first byte, this can be an
20062 * ANYOFRb. Otherwise store the lowest UTF-8 start byte which can
20063 * quickly rule out many inputs at run-time without having to compute
20064 * the code point from UTF-8. For EBCDIC, we use I8, as not doing that
20065 * transformation would not rule out nearly so many things */
20066 if (low_utf8[0] == high_utf8[0]) {
20068 OP(REGNODE_p(*ret)) = op;
20069 ANYOF_FLAGS(REGNODE_p(*ret)) = low_utf8[0];
20072 ANYOF_FLAGS(REGNODE_p(*ret)) = NATIVE_UTF8_TO_I8(low_utf8[0]);
20078 /* If didn't find an optimization and there is no need for a bitmap,
20079 * optimize to indicate that */
20080 if ( start[0] >= NUM_ANYOF_CODE_POINTS
20082 && ! upper_latin1_only_utf8_matches
20083 && *anyof_flags == 0)
20085 U8 low_utf8[UTF8_MAXBYTES+1];
20086 UV highest_cp = invlist_highest(cp_list);
20088 /* Currently the maximum allowed code point by the system is IV_MAX.
20089 * Higher ones are reserved for future internal use. This particular
20090 * regnode can be used for higher ones, but we can't calculate the code
20091 * point of those. IV_MAX suffices though, as it will be a large first
20093 Size_t low_len = uvchr_to_utf8(low_utf8, MIN(start[0], IV_MAX))
20096 /* We store the lowest possible first byte of the UTF-8 representation,
20097 * using the flags field. This allows for quick ruling out of some
20098 * inputs without having to convert from UTF-8 to code point. For
20099 * EBCDIC, we use I8, as not doing that transformation would not rule
20100 * out nearly so many things */
20101 *anyof_flags = NATIVE_UTF8_TO_I8(low_utf8[0]);
20105 /* If the first UTF-8 start byte for the highest code point in the
20106 * range is suitably small, we may be able to get an upper bound as
20108 if (highest_cp <= IV_MAX) {
20109 U8 high_utf8[UTF8_MAXBYTES+1];
20110 Size_t high_len = uvchr_to_utf8(high_utf8, highest_cp) - high_utf8;
20112 /* If the lowest and highest are the same, we can get an exact
20113 * first byte instead of a just minimum or even a sequence of exact
20114 * leading bytes. We signal these with different regnodes */
20115 if (low_utf8[0] == high_utf8[0]) {
20116 Size_t len = find_first_differing_byte_pos(low_utf8,
20118 MIN(low_len, high_len));
20122 /* No need to convert to I8 for EBCDIC as this is an exact
20124 *anyof_flags = low_utf8[0];
20129 *ret = regnode_guts(pRExC_state, op,
20130 regarglen[op] + STR_SZ(len),
20132 FILL_NODE(*ret, op);
20133 ((struct regnode_anyofhs *) REGNODE_p(*ret))->str_len
20135 Copy(low_utf8, /* Add the common bytes */
20136 ((struct regnode_anyofhs *) REGNODE_p(*ret))->string,
20138 RExC_emit += NODE_SZ_STR(REGNODE_p(*ret));
20139 set_ANYOF_arg(pRExC_state, REGNODE_p(*ret), cp_list,
20140 NULL, only_utf8_locale_list);
20144 else if (NATIVE_UTF8_TO_I8(high_utf8[0]) <= MAX_ANYOF_HRx_BYTE) {
20146 /* Here, the high byte is not the same as the low, but is small
20147 * enough that its reasonable to have a loose upper bound,
20148 * which is packed in with the strict lower bound. See
20149 * comments at the definition of MAX_ANYOF_HRx_BYTE. On EBCDIC
20150 * platforms, I8 is used. On ASCII platforms I8 is the same
20151 * thing as UTF-8 */
20154 U8 max_range_diff = MAX_ANYOF_HRx_BYTE - *anyof_flags;
20155 U8 range_diff = NATIVE_UTF8_TO_I8(high_utf8[0])
20158 if (range_diff <= max_range_diff / 8) {
20161 else if (range_diff <= max_range_diff / 4) {
20164 else if (range_diff <= max_range_diff / 2) {
20167 *anyof_flags = (*anyof_flags - 0xC0) << 2 | bits;
20176 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
20179 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
20180 regnode* const node,
20182 SV* const runtime_defns,
20183 SV* const only_utf8_locale_list)
20185 /* Sets the arg field of an ANYOF-type node 'node', using information about
20186 * the node passed-in. If there is nothing outside the node's bitmap, the
20187 * arg is set to ANYOF_ONLY_HAS_BITMAP. Otherwise, it sets the argument to
20188 * the count returned by add_data(), having allocated and stored an array,
20191 * av[0] stores the inversion list defining this class as far as known at
20192 * this time, or PL_sv_undef if nothing definite is now known.
20193 * av[1] stores the inversion list of code points that match only if the
20194 * current locale is UTF-8, or if none, PL_sv_undef if there is an
20195 * av[2], or no entry otherwise.
20196 * av[2] stores the list of user-defined properties whose subroutine
20197 * definitions aren't known at this time, or no entry if none. */
20201 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
20203 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
20204 assert(! (ANYOF_FLAGS(node)
20205 & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP));
20206 ARG_SET(node, ANYOF_ONLY_HAS_BITMAP);
20209 AV * const av = newAV();
20213 av_store(av, INVLIST_INDEX, SvREFCNT_inc_NN(cp_list));
20216 /* (Note that if any of this changes, the size calculations in
20217 * S_optimize_regclass() might need to be updated.) */
20219 if (only_utf8_locale_list) {
20220 av_store(av, ONLY_LOCALE_MATCHES_INDEX,
20221 SvREFCNT_inc_NN(only_utf8_locale_list));
20224 if (runtime_defns) {
20225 av_store(av, DEFERRED_USER_DEFINED_INDEX,
20226 SvREFCNT_inc_NN(runtime_defns));
20229 rv = newRV_noinc(MUTABLE_SV(av));
20230 n = add_data(pRExC_state, STR_WITH_LEN("s"));
20231 RExC_rxi->data->data[n] = (void*)rv;
20238 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
20239 Perl_get_regclass_nonbitmap_data(pTHX_ const regexp *prog, const regnode* node, bool doinit, SV** listsvp, SV** only_utf8_locale_ptr, SV** output_invlist)
20241 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)
20245 /* For internal core use only.
20246 * Returns the inversion list for the input 'node' in the regex 'prog'.
20247 * If <doinit> is 'true', will attempt to create the inversion list if not
20249 * If <listsvp> is non-null, will return the printable contents of the
20250 * property definition. This can be used to get debugging information
20251 * even before the inversion list exists, by calling this function with
20252 * 'doinit' set to false, in which case the components that will be used
20253 * to eventually create the inversion list are returned (in a printable
20255 * If <only_utf8_locale_ptr> is not NULL, it is where this routine is to
20256 * store an inversion list of code points that should match only if the
20257 * execution-time locale is a UTF-8 one.
20258 * If <output_invlist> is not NULL, it is where this routine is to store an
20259 * inversion list of the code points that would be instead returned in
20260 * <listsvp> if this were NULL. Thus, what gets output in <listsvp>
20261 * when this parameter is used, is just the non-code point data that
20262 * will go into creating the inversion list. This currently should be just
20263 * user-defined properties whose definitions were not known at compile
20264 * time. Using this parameter allows for easier manipulation of the
20265 * inversion list's data by the caller. It is illegal to call this
20266 * function with this parameter set, but not <listsvp>
20268 * Tied intimately to how S_set_ANYOF_arg sets up the data structure. Note
20269 * that, in spite of this function's name, the inversion list it returns
20270 * may include the bitmap data as well */
20272 SV *si = NULL; /* Input initialization string */
20273 SV* invlist = NULL;
20275 RXi_GET_DECL(prog, progi);
20276 const struct reg_data * const data = prog ? progi->data : NULL;
20278 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
20279 PERL_ARGS_ASSERT_GET_REGCLASS_NONBITMAP_DATA;
20281 PERL_ARGS_ASSERT_GET_RE_GCLASS_NONBITMAP_DATA;
20283 assert(! output_invlist || listsvp);
20285 if (data && data->count) {
20286 const U32 n = ARG(node);
20288 if (data->what[n] == 's') {
20289 SV * const rv = MUTABLE_SV(data->data[n]);
20290 AV * const av = MUTABLE_AV(SvRV(rv));
20291 SV **const ary = AvARRAY(av);
20293 invlist = ary[INVLIST_INDEX];
20295 if (av_tindex_skip_len_mg(av) >= ONLY_LOCALE_MATCHES_INDEX) {
20296 *only_utf8_locale_ptr = ary[ONLY_LOCALE_MATCHES_INDEX];
20299 if (av_tindex_skip_len_mg(av) >= DEFERRED_USER_DEFINED_INDEX) {
20300 si = ary[DEFERRED_USER_DEFINED_INDEX];
20303 if (doinit && (si || invlist)) {
20306 SV * msg = newSVpvs_flags("", SVs_TEMP);
20308 SV * prop_definition = handle_user_defined_property(
20309 "", 0, FALSE, /* There is no \p{}, \P{} */
20310 SvPVX_const(si)[1] - '0', /* /i or not has been
20311 stored here for just
20313 TRUE, /* run time */
20314 FALSE, /* This call must find the defn */
20315 si, /* The property definition */
20318 0 /* base level call */
20322 assert(prop_definition == NULL);
20324 Perl_croak(aTHX_ "%" UTF8f,
20325 UTF8fARG(SvUTF8(msg), SvCUR(msg), SvPVX(msg)));
20329 _invlist_union(invlist, prop_definition, &invlist);
20330 SvREFCNT_dec_NN(prop_definition);
20333 invlist = prop_definition;
20336 STATIC_ASSERT_STMT(ONLY_LOCALE_MATCHES_INDEX == 1 + INVLIST_INDEX);
20337 STATIC_ASSERT_STMT(DEFERRED_USER_DEFINED_INDEX == 1 + ONLY_LOCALE_MATCHES_INDEX);
20339 ary[INVLIST_INDEX] = invlist;
20340 av_fill(av, (ary[ONLY_LOCALE_MATCHES_INDEX])
20341 ? ONLY_LOCALE_MATCHES_INDEX
20349 /* If requested, return a printable version of what this ANYOF node matches
20352 SV* matches_string = NULL;
20354 /* This function can be called at compile-time, before everything gets
20355 * resolved, in which case we return the currently best available
20356 * information, which is the string that will eventually be used to do
20357 * that resolving, 'si' */
20359 /* Here, we only have 'si' (and possibly some passed-in data in
20360 * 'invlist', which is handled below) If the caller only wants
20361 * 'si', use that. */
20362 if (! output_invlist) {
20363 matches_string = newSVsv(si);
20366 /* But if the caller wants an inversion list of the node, we
20367 * need to parse 'si' and place as much as possible in the
20368 * desired output inversion list, making 'matches_string' only
20369 * contain the currently unresolvable things */
20370 const char *si_string = SvPVX(si);
20371 STRLEN remaining = SvCUR(si);
20375 /* Ignore everything before and including the first new-line */
20376 si_string = (const char *) memchr(si_string, '\n', SvCUR(si));
20377 assert (si_string != NULL);
20379 remaining = SvPVX(si) + SvCUR(si) - si_string;
20381 while (remaining > 0) {
20383 /* The data consists of just strings defining user-defined
20384 * property names, but in prior incarnations, and perhaps
20385 * somehow from pluggable regex engines, it could still
20386 * hold hex code point definitions, all of which should be
20387 * legal (or it wouldn't have gotten this far). Each
20388 * component of a range would be separated by a tab, and
20389 * each range by a new-line. If these are found, instead
20390 * add them to the inversion list */
20391 I32 grok_flags = PERL_SCAN_SILENT_ILLDIGIT
20392 |PERL_SCAN_SILENT_NON_PORTABLE;
20393 STRLEN len = remaining;
20394 UV cp = grok_hex(si_string, &len, &grok_flags, NULL);
20396 /* If the hex decode routine found something, it should go
20397 * up to the next \n */
20398 if ( *(si_string + len) == '\n') {
20399 if (count) { /* 2nd code point on line */
20400 *output_invlist = _add_range_to_invlist(*output_invlist, prev_cp, cp);
20403 *output_invlist = add_cp_to_invlist(*output_invlist, cp);
20406 goto prepare_for_next_iteration;
20409 /* If the hex decode was instead for the lower range limit,
20410 * save it, and go parse the upper range limit */
20411 if (*(si_string + len) == '\t') {
20412 assert(count == 0);
20416 prepare_for_next_iteration:
20417 si_string += len + 1;
20418 remaining -= len + 1;
20422 /* Here, didn't find a legal hex number. Just add the text
20423 * from here up to the next \n, omitting any trailing
20427 len = strcspn(si_string,
20428 DEFERRED_COULD_BE_OFFICIAL_MARKERs "\n");
20430 if (matches_string) {
20431 sv_catpvn(matches_string, si_string, len);
20434 matches_string = newSVpvn(si_string, len);
20436 sv_catpvs(matches_string, " ");
20440 && UCHARAT(si_string)
20441 == DEFERRED_COULD_BE_OFFICIAL_MARKERc)
20446 if (remaining && UCHARAT(si_string) == '\n') {
20450 } /* end of loop through the text */
20452 assert(matches_string);
20453 if (SvCUR(matches_string)) { /* Get rid of trailing blank */
20454 SvCUR_set(matches_string, SvCUR(matches_string) - 1);
20456 } /* end of has an 'si' */
20459 /* Add the stuff that's already known */
20462 /* Again, if the caller doesn't want the output inversion list, put
20463 * everything in 'matches-string' */
20464 if (! output_invlist) {
20465 if ( ! matches_string) {
20466 matches_string = newSVpvs("\n");
20468 sv_catsv(matches_string, invlist_contents(invlist,
20469 TRUE /* traditional style */
20472 else if (! *output_invlist) {
20473 *output_invlist = invlist_clone(invlist, NULL);
20476 _invlist_union(*output_invlist, invlist, output_invlist);
20480 *listsvp = matches_string;
20486 /* reg_skipcomment()
20488 Absorbs an /x style # comment from the input stream,
20489 returning a pointer to the first character beyond the comment, or if the
20490 comment terminates the pattern without anything following it, this returns
20491 one past the final character of the pattern (in other words, RExC_end) and
20492 sets the REG_RUN_ON_COMMENT_SEEN flag.
20494 Note it's the callers responsibility to ensure that we are
20495 actually in /x mode
20499 PERL_STATIC_INLINE char*
20500 S_reg_skipcomment(RExC_state_t *pRExC_state, char* p)
20502 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
20506 while (p < RExC_end) {
20507 if (*(++p) == '\n') {
20512 /* we ran off the end of the pattern without ending the comment, so we have
20513 * to add an \n when wrapping */
20514 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
20519 S_skip_to_be_ignored_text(pTHX_ RExC_state_t *pRExC_state,
20521 const bool force_to_xmod
20524 /* If the text at the current parse position '*p' is a '(?#...)' comment,
20525 * or if we are under /x or 'force_to_xmod' is TRUE, and the text at '*p'
20526 * is /x whitespace, advance '*p' so that on exit it points to the first
20527 * byte past all such white space and comments */
20529 const bool use_xmod = force_to_xmod || (RExC_flags & RXf_PMf_EXTENDED);
20531 PERL_ARGS_ASSERT_SKIP_TO_BE_IGNORED_TEXT;
20533 assert( ! UTF || UTF8_IS_INVARIANT(**p) || UTF8_IS_START(**p));
20536 if (RExC_end - (*p) >= 3
20538 && *(*p + 1) == '?'
20539 && *(*p + 2) == '#')
20541 while (*(*p) != ')') {
20542 if ((*p) == RExC_end)
20543 FAIL("Sequence (?#... not terminated");
20551 const char * save_p = *p;
20552 while ((*p) < RExC_end) {
20554 if ((len = is_PATWS_safe((*p), RExC_end, UTF))) {
20557 else if (*(*p) == '#') {
20558 (*p) = reg_skipcomment(pRExC_state, (*p));
20564 if (*p != save_p) {
20577 Advances the parse position by one byte, unless that byte is the beginning
20578 of a '(?#...)' style comment, or is /x whitespace and /x is in effect. In
20579 those two cases, the parse position is advanced beyond all such comments and
20582 This is the UTF, (?#...), and /x friendly way of saying RExC_parse++.
20586 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
20588 PERL_ARGS_ASSERT_NEXTCHAR;
20590 if (RExC_parse < RExC_end) {
20592 || UTF8_IS_INVARIANT(*RExC_parse)
20593 || UTF8_IS_START(*RExC_parse));
20595 RExC_parse += (UTF)
20596 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
20599 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
20600 FALSE /* Don't force /x */ );
20605 S_change_engine_size(pTHX_ RExC_state_t *pRExC_state, const Ptrdiff_t size)
20607 /* 'size' is the delta number of smallest regnode equivalents to add or
20608 * subtract from the current memory allocated to the regex engine being
20611 PERL_ARGS_ASSERT_CHANGE_ENGINE_SIZE;
20616 sizeof(regexp_internal) + (RExC_size + 1) * sizeof(regnode),
20617 /* +1 for REG_MAGIC */
20620 if ( RExC_rxi == NULL )
20621 FAIL("Regexp out of space");
20622 RXi_SET(RExC_rx, RExC_rxi);
20624 RExC_emit_start = RExC_rxi->program;
20626 Zero(REGNODE_p(RExC_emit), size, regnode);
20629 #ifdef RE_TRACK_PATTERN_OFFSETS
20630 Renew(RExC_offsets, 2*RExC_size+1, U32);
20632 Zero(RExC_offsets + 2*(RExC_size - size) + 1, 2 * size, U32);
20634 RExC_offsets[0] = RExC_size;
20638 STATIC regnode_offset
20639 S_regnode_guts(pTHX_ RExC_state_t *pRExC_state, const U8 op, const STRLEN extra_size, const char* const name)
20641 /* Allocate a regnode for 'op', with 'extra_size' extra (smallest) regnode
20642 * equivalents space. It aligns and increments RExC_size
20644 * It returns the regnode's offset into the regex engine program */
20646 const regnode_offset ret = RExC_emit;
20648 DECLARE_AND_GET_RE_DEBUG_FLAGS;
20650 PERL_ARGS_ASSERT_REGNODE_GUTS;
20652 SIZE_ALIGN(RExC_size);
20653 change_engine_size(pRExC_state, (Ptrdiff_t) 1 + extra_size);
20654 NODE_ALIGN_FILL(REGNODE_p(ret));
20655 #ifndef RE_TRACK_PATTERN_OFFSETS
20656 PERL_UNUSED_ARG(name);
20657 PERL_UNUSED_ARG(op);
20659 assert(extra_size >= regarglen[op] || PL_regkind[op] == ANYOF);
20661 if (RExC_offsets) { /* MJD */
20663 ("%s:%d: (op %s) %s %" UVuf " (len %" UVuf ") (max %" UVuf ").\n",
20666 (UV)(RExC_emit) > RExC_offsets[0]
20667 ? "Overwriting end of array!\n" : "OK",
20669 (UV)(RExC_parse - RExC_start),
20670 (UV)RExC_offsets[0]));
20671 Set_Node_Offset(REGNODE_p(RExC_emit), RExC_parse + (op == END));
20678 - reg_node - emit a node
20680 STATIC regnode_offset /* Location. */
20681 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
20683 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reg_node");
20684 regnode_offset ptr = ret;
20686 PERL_ARGS_ASSERT_REG_NODE;
20688 assert(regarglen[op] == 0);
20690 FILL_ADVANCE_NODE(ptr, op);
20696 - reganode - emit a node with an argument
20698 STATIC regnode_offset /* Location. */
20699 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
20701 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reganode");
20702 regnode_offset ptr = ret;
20704 PERL_ARGS_ASSERT_REGANODE;
20706 /* ANYOF are special cased to allow non-length 1 args */
20707 assert(regarglen[op] == 1);
20709 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
20715 - regpnode - emit a temporary node with a SV* argument
20717 STATIC regnode_offset /* Location. */
20718 S_regpnode(pTHX_ RExC_state_t *pRExC_state, U8 op, SV * arg)
20720 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "regpnode");
20721 regnode_offset ptr = ret;
20723 PERL_ARGS_ASSERT_REGPNODE;
20725 FILL_ADVANCE_NODE_ARGp(ptr, op, arg);
20730 STATIC regnode_offset
20731 S_reg2Lanode(pTHX_ RExC_state_t *pRExC_state, const U8 op, const U32 arg1, const I32 arg2)
20733 /* emit a node with U32 and I32 arguments */
20735 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reg2Lanode");
20736 regnode_offset ptr = ret;
20738 PERL_ARGS_ASSERT_REG2LANODE;
20740 assert(regarglen[op] == 2);
20742 FILL_ADVANCE_NODE_2L_ARG(ptr, op, arg1, arg2);
20748 - reginsert - insert an operator in front of already-emitted operand
20750 * That means that on exit 'operand' is the offset of the newly inserted
20751 * operator, and the original operand has been relocated.
20753 * IMPORTANT NOTE - it is the *callers* responsibility to correctly
20754 * set up NEXT_OFF() of the inserted node if needed. Something like this:
20756 * reginsert(pRExC, OPFAIL, orig_emit, depth+1);
20757 * NEXT_OFF(orig_emit) = regarglen[OPFAIL] + NODE_STEP_REGNODE;
20759 * ALSO NOTE - FLAGS(newly-inserted-operator) will be set to 0 as well.
20762 S_reginsert(pTHX_ RExC_state_t *pRExC_state, const U8 op,
20763 const regnode_offset operand, const U32 depth)
20768 const int offset = regarglen[(U8)op];
20769 const int size = NODE_STEP_REGNODE + offset;
20770 DECLARE_AND_GET_RE_DEBUG_FLAGS;
20772 PERL_ARGS_ASSERT_REGINSERT;
20773 PERL_UNUSED_CONTEXT;
20774 PERL_UNUSED_ARG(depth);
20775 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
20776 DEBUG_PARSE_FMT("inst"," - %s", PL_reg_name[op]);
20777 assert(!RExC_study_started); /* I believe we should never use reginsert once we have started
20778 studying. If this is wrong then we need to adjust RExC_recurse
20779 below like we do with RExC_open_parens/RExC_close_parens. */
20780 change_engine_size(pRExC_state, (Ptrdiff_t) size);
20781 src = REGNODE_p(RExC_emit);
20783 dst = REGNODE_p(RExC_emit);
20785 /* If we are in a "count the parentheses" pass, the numbers are unreliable,
20786 * and [perl #133871] shows this can lead to problems, so skip this
20787 * realignment of parens until a later pass when they are reliable */
20788 if (! IN_PARENS_PASS && RExC_open_parens) {
20790 /*DEBUG_PARSE_FMT("inst"," - %" IVdf, (IV)RExC_npar);*/
20791 /* remember that RExC_npar is rex->nparens + 1,
20792 * iow it is 1 more than the number of parens seen in
20793 * the pattern so far. */
20794 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
20795 /* note, RExC_open_parens[0] is the start of the
20796 * regex, it can't move. RExC_close_parens[0] is the end
20797 * of the regex, it *can* move. */
20798 if ( paren && RExC_open_parens[paren] >= operand ) {
20799 /*DEBUG_PARSE_FMT("open"," - %d", size);*/
20800 RExC_open_parens[paren] += size;
20802 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
20804 if ( RExC_close_parens[paren] >= operand ) {
20805 /*DEBUG_PARSE_FMT("close"," - %d", size);*/
20806 RExC_close_parens[paren] += size;
20808 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
20813 RExC_end_op += size;
20815 while (src > REGNODE_p(operand)) {
20816 StructCopy(--src, --dst, regnode);
20817 #ifdef RE_TRACK_PATTERN_OFFSETS
20818 if (RExC_offsets) { /* MJD 20010112 */
20820 ("%s(%d): (op %s) %s copy %" UVuf " -> %" UVuf " (max %" UVuf ").\n",
20824 (UV)(REGNODE_OFFSET(dst)) > RExC_offsets[0]
20825 ? "Overwriting end of array!\n" : "OK",
20826 (UV)REGNODE_OFFSET(src),
20827 (UV)REGNODE_OFFSET(dst),
20828 (UV)RExC_offsets[0]));
20829 Set_Node_Offset_To_R(REGNODE_OFFSET(dst), Node_Offset(src));
20830 Set_Node_Length_To_R(REGNODE_OFFSET(dst), Node_Length(src));
20835 place = REGNODE_p(operand); /* Op node, where operand used to be. */
20836 #ifdef RE_TRACK_PATTERN_OFFSETS
20837 if (RExC_offsets) { /* MJD */
20839 ("%s(%d): (op %s) %s %" UVuf " <- %" UVuf " (max %" UVuf ").\n",
20843 (UV)REGNODE_OFFSET(place) > RExC_offsets[0]
20844 ? "Overwriting end of array!\n" : "OK",
20845 (UV)REGNODE_OFFSET(place),
20846 (UV)(RExC_parse - RExC_start),
20847 (UV)RExC_offsets[0]));
20848 Set_Node_Offset(place, RExC_parse);
20849 Set_Node_Length(place, 1);
20852 src = NEXTOPER(place);
20854 FILL_NODE(operand, op);
20856 /* Zero out any arguments in the new node */
20857 Zero(src, offset, regnode);
20861 - regtail - set the next-pointer at the end of a node chain of p to val. If
20862 that value won't fit in the space available, instead returns FALSE.
20863 (Except asserts if we can't fit in the largest space the regex
20864 engine is designed for.)
20865 - SEE ALSO: regtail_study
20868 S_regtail(pTHX_ RExC_state_t * pRExC_state,
20869 const regnode_offset p,
20870 const regnode_offset val,
20873 regnode_offset scan;
20874 DECLARE_AND_GET_RE_DEBUG_FLAGS;
20876 PERL_ARGS_ASSERT_REGTAIL;
20878 PERL_UNUSED_ARG(depth);
20881 /* The final node in the chain is the first one with a nonzero next pointer
20883 scan = (regnode_offset) p;
20885 regnode * const temp = regnext(REGNODE_p(scan));
20887 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
20888 regprop(RExC_rx, RExC_mysv, REGNODE_p(scan), NULL, pRExC_state);
20889 Perl_re_printf( aTHX_ "~ %s (%zu) %s %s\n",
20890 SvPV_nolen_const(RExC_mysv), scan,
20891 (temp == NULL ? "->" : ""),
20892 (temp == NULL ? PL_reg_name[OP(REGNODE_p(val))] : "")
20897 scan = REGNODE_OFFSET(temp);
20900 /* Populate this node's next pointer */
20901 assert(val >= scan);
20902 if (reg_off_by_arg[OP(REGNODE_p(scan))]) {
20903 assert((UV) (val - scan) <= U32_MAX);
20904 ARG_SET(REGNODE_p(scan), val - scan);
20907 if (val - scan > U16_MAX) {
20908 /* Populate this with something that won't loop and will likely
20909 * lead to a crash if the caller ignores the failure return, and
20910 * execution continues */
20911 NEXT_OFF(REGNODE_p(scan)) = U16_MAX;
20914 NEXT_OFF(REGNODE_p(scan)) = val - scan;
20922 - regtail_study - set the next-pointer at the end of a node chain of p to val.
20923 - Look for optimizable sequences at the same time.
20924 - currently only looks for EXACT chains.
20926 This is experimental code. The idea is to use this routine to perform
20927 in place optimizations on branches and groups as they are constructed,
20928 with the long term intention of removing optimization from study_chunk so
20929 that it is purely analytical.
20931 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
20932 to control which is which.
20934 This used to return a value that was ignored. It was a problem that it is
20935 #ifdef'd to be another function that didn't return a value. khw has changed it
20936 so both currently return a pass/fail return.
20939 /* TODO: All four parms should be const */
20942 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode_offset p,
20943 const regnode_offset val, U32 depth)
20945 regnode_offset scan;
20947 #ifdef EXPERIMENTAL_INPLACESCAN
20950 DECLARE_AND_GET_RE_DEBUG_FLAGS;
20952 PERL_ARGS_ASSERT_REGTAIL_STUDY;
20955 /* Find last node. */
20959 regnode * const temp = regnext(REGNODE_p(scan));
20960 #ifdef EXPERIMENTAL_INPLACESCAN
20961 if (PL_regkind[OP(REGNODE_p(scan))] == EXACT) {
20962 bool unfolded_multi_char; /* Unexamined in this routine */
20963 if (join_exact(pRExC_state, scan, &min,
20964 &unfolded_multi_char, 1, REGNODE_p(val), depth+1))
20965 return TRUE; /* Was return EXACT */
20969 if (PL_regkind[OP(REGNODE_p(scan))] == EXACT) {
20970 if (exact == PSEUDO )
20971 exact= OP(REGNODE_p(scan));
20972 else if (exact != OP(REGNODE_p(scan)) )
20975 else if (OP(REGNODE_p(scan)) != NOTHING) {
20980 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
20981 regprop(RExC_rx, RExC_mysv, REGNODE_p(scan), NULL, pRExC_state);
20982 Perl_re_printf( aTHX_ "~ %s (%zu) -> %s\n",
20983 SvPV_nolen_const(RExC_mysv),
20985 PL_reg_name[exact]);
20989 scan = REGNODE_OFFSET(temp);
20992 DEBUG_PARSE_MSG("");
20993 regprop(RExC_rx, RExC_mysv, REGNODE_p(val), NULL, pRExC_state);
20994 Perl_re_printf( aTHX_
20995 "~ attach to %s (%" IVdf ") offset to %" IVdf "\n",
20996 SvPV_nolen_const(RExC_mysv),
21001 if (reg_off_by_arg[OP(REGNODE_p(scan))]) {
21002 assert((UV) (val - scan) <= U32_MAX);
21003 ARG_SET(REGNODE_p(scan), val - scan);
21006 if (val - scan > U16_MAX) {
21007 /* Populate this with something that won't loop and will likely
21008 * lead to a crash if the caller ignores the failure return, and
21009 * execution continues */
21010 NEXT_OFF(REGNODE_p(scan)) = U16_MAX;
21013 NEXT_OFF(REGNODE_p(scan)) = val - scan;
21016 return TRUE; /* Was 'return exact' */
21021 S_get_ANYOFM_contents(pTHX_ const regnode * n) {
21023 /* Returns an inversion list of all the code points matched by the
21024 * ANYOFM/NANYOFM node 'n' */
21026 SV * cp_list = _new_invlist(-1);
21027 const U8 lowest = (U8) ARG(n);
21030 U8 needed = 1U << PL_bitcount[ (U8) ~ FLAGS(n)];
21032 PERL_ARGS_ASSERT_GET_ANYOFM_CONTENTS;
21034 /* Starting with the lowest code point, any code point that ANDed with the
21035 * mask yields the lowest code point is in the set */
21036 for (i = lowest; i <= 0xFF; i++) {
21037 if ((i & FLAGS(n)) == ARG(n)) {
21038 cp_list = add_cp_to_invlist(cp_list, i);
21041 /* We know how many code points (a power of two) that are in the
21042 * set. No use looking once we've got that number */
21043 if (count >= needed) break;
21047 if (OP(n) == NANYOFM) {
21048 _invlist_invert(cp_list);
21054 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
21059 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
21064 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
21066 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
21067 if (flags & (1<<bit)) {
21068 if (!set++ && lead)
21069 Perl_re_printf( aTHX_ "%s", lead);
21070 Perl_re_printf( aTHX_ "%s ", PL_reg_intflags_name[bit]);
21075 Perl_re_printf( aTHX_ "\n");
21077 Perl_re_printf( aTHX_ "%s[none-set]\n", lead);
21082 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
21088 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
21090 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
21091 if (flags & (1<<bit)) {
21092 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
21095 if (!set++ && lead)
21096 Perl_re_printf( aTHX_ "%s", lead);
21097 Perl_re_printf( aTHX_ "%s ", PL_reg_extflags_name[bit]);
21100 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
21101 if (!set++ && lead) {
21102 Perl_re_printf( aTHX_ "%s", lead);
21105 case REGEX_UNICODE_CHARSET:
21106 Perl_re_printf( aTHX_ "UNICODE");
21108 case REGEX_LOCALE_CHARSET:
21109 Perl_re_printf( aTHX_ "LOCALE");
21111 case REGEX_ASCII_RESTRICTED_CHARSET:
21112 Perl_re_printf( aTHX_ "ASCII-RESTRICTED");
21114 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
21115 Perl_re_printf( aTHX_ "ASCII-MORE_RESTRICTED");
21118 Perl_re_printf( aTHX_ "UNKNOWN CHARACTER SET");
21124 Perl_re_printf( aTHX_ "\n");
21126 Perl_re_printf( aTHX_ "%s[none-set]\n", lead);
21132 Perl_regdump(pTHX_ const regexp *r)
21136 SV * const sv = sv_newmortal();
21137 SV *dsv= sv_newmortal();
21138 RXi_GET_DECL(r, ri);
21139 DECLARE_AND_GET_RE_DEBUG_FLAGS;
21141 PERL_ARGS_ASSERT_REGDUMP;
21143 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
21145 /* Header fields of interest. */
21146 for (i = 0; i < 2; i++) {
21147 if (r->substrs->data[i].substr) {
21148 RE_PV_QUOTED_DECL(s, 0, dsv,
21149 SvPVX_const(r->substrs->data[i].substr),
21150 RE_SV_DUMPLEN(r->substrs->data[i].substr),
21151 PL_dump_re_max_len);
21152 Perl_re_printf( aTHX_
21153 "%s %s%s at %" IVdf "..%" UVuf " ",
21154 i ? "floating" : "anchored",
21156 RE_SV_TAIL(r->substrs->data[i].substr),
21157 (IV)r->substrs->data[i].min_offset,
21158 (UV)r->substrs->data[i].max_offset);
21160 else if (r->substrs->data[i].utf8_substr) {
21161 RE_PV_QUOTED_DECL(s, 1, dsv,
21162 SvPVX_const(r->substrs->data[i].utf8_substr),
21163 RE_SV_DUMPLEN(r->substrs->data[i].utf8_substr),
21165 Perl_re_printf( aTHX_
21166 "%s utf8 %s%s at %" IVdf "..%" UVuf " ",
21167 i ? "floating" : "anchored",
21169 RE_SV_TAIL(r->substrs->data[i].utf8_substr),
21170 (IV)r->substrs->data[i].min_offset,
21171 (UV)r->substrs->data[i].max_offset);
21175 if (r->check_substr || r->check_utf8)
21176 Perl_re_printf( aTHX_
21178 ( r->check_substr == r->substrs->data[1].substr
21179 && r->check_utf8 == r->substrs->data[1].utf8_substr
21180 ? "(checking floating" : "(checking anchored"));
21181 if (r->intflags & PREGf_NOSCAN)
21182 Perl_re_printf( aTHX_ " noscan");
21183 if (r->extflags & RXf_CHECK_ALL)
21184 Perl_re_printf( aTHX_ " isall");
21185 if (r->check_substr || r->check_utf8)
21186 Perl_re_printf( aTHX_ ") ");
21188 if (ri->regstclass) {
21189 regprop(r, sv, ri->regstclass, NULL, NULL);
21190 Perl_re_printf( aTHX_ "stclass %s ", SvPVX_const(sv));
21192 if (r->intflags & PREGf_ANCH) {
21193 Perl_re_printf( aTHX_ "anchored");
21194 if (r->intflags & PREGf_ANCH_MBOL)
21195 Perl_re_printf( aTHX_ "(MBOL)");
21196 if (r->intflags & PREGf_ANCH_SBOL)
21197 Perl_re_printf( aTHX_ "(SBOL)");
21198 if (r->intflags & PREGf_ANCH_GPOS)
21199 Perl_re_printf( aTHX_ "(GPOS)");
21200 Perl_re_printf( aTHX_ " ");
21202 if (r->intflags & PREGf_GPOS_SEEN)
21203 Perl_re_printf( aTHX_ "GPOS:%" UVuf " ", (UV)r->gofs);
21204 if (r->intflags & PREGf_SKIP)
21205 Perl_re_printf( aTHX_ "plus ");
21206 if (r->intflags & PREGf_IMPLICIT)
21207 Perl_re_printf( aTHX_ "implicit ");
21208 Perl_re_printf( aTHX_ "minlen %" IVdf " ", (IV)r->minlen);
21209 if (r->extflags & RXf_EVAL_SEEN)
21210 Perl_re_printf( aTHX_ "with eval ");
21211 Perl_re_printf( aTHX_ "\n");
21213 regdump_extflags("r->extflags: ", r->extflags);
21214 regdump_intflags("r->intflags: ", r->intflags);
21217 PERL_ARGS_ASSERT_REGDUMP;
21218 PERL_UNUSED_CONTEXT;
21219 PERL_UNUSED_ARG(r);
21220 #endif /* DEBUGGING */
21223 /* Should be synchronized with ANYOF_ #defines in regcomp.h */
21226 # if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 \
21227 || _CC_LOWER != 3 || _CC_UPPER != 4 || _CC_PUNCT != 5 \
21228 || _CC_PRINT != 6 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 \
21229 || _CC_CASED != 9 || _CC_SPACE != 10 || _CC_BLANK != 11 \
21230 || _CC_XDIGIT != 12 || _CC_CNTRL != 13 || _CC_ASCII != 14 \
21231 || _CC_VERTSPACE != 15
21232 # error Need to adjust order of anyofs[]
21234 static const char * const anyofs[] = {
21271 - regprop - printable representation of opcode, with run time support
21275 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo, const RExC_state_t *pRExC_state)
21279 RXi_GET_DECL(prog, progi);
21280 DECLARE_AND_GET_RE_DEBUG_FLAGS;
21282 PERL_ARGS_ASSERT_REGPROP;
21286 if (OP(o) > REGNODE_MAX) { /* regnode.type is unsigned */
21287 if (pRExC_state) { /* This gives more info, if we have it */
21288 FAIL3("panic: corrupted regexp opcode %d > %d",
21289 (int)OP(o), (int)REGNODE_MAX);
21292 Perl_croak(aTHX_ "panic: corrupted regexp opcode %d > %d",
21293 (int)OP(o), (int)REGNODE_MAX);
21296 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
21298 k = PL_regkind[OP(o)];
21301 sv_catpvs(sv, " ");
21302 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
21303 * is a crude hack but it may be the best for now since
21304 * we have no flag "this EXACTish node was UTF-8"
21306 pv_pretty(sv, STRING(o), STR_LEN(o), PL_dump_re_max_len,
21307 PL_colors[0], PL_colors[1],
21308 PERL_PV_ESCAPE_UNI_DETECT |
21309 PERL_PV_ESCAPE_NONASCII |
21310 PERL_PV_PRETTY_ELLIPSES |
21311 PERL_PV_PRETTY_LTGT |
21312 PERL_PV_PRETTY_NOCLEAR
21314 } else if (k == TRIE) {
21315 /* print the details of the trie in dumpuntil instead, as
21316 * progi->data isn't available here */
21317 const char op = OP(o);
21318 const U32 n = ARG(o);
21319 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
21320 (reg_ac_data *)progi->data->data[n] :
21322 const reg_trie_data * const trie
21323 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
21325 Perl_sv_catpvf(aTHX_ sv, "-%s", PL_reg_name[o->flags]);
21326 DEBUG_TRIE_COMPILE_r({
21328 sv_catpvs(sv, "(JUMP)");
21329 Perl_sv_catpvf(aTHX_ sv,
21330 "<S:%" UVuf "/%" IVdf " W:%" UVuf " L:%" UVuf "/%" UVuf " C:%" UVuf "/%" UVuf ">",
21331 (UV)trie->startstate,
21332 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
21333 (UV)trie->wordcount,
21336 (UV)TRIE_CHARCOUNT(trie),
21337 (UV)trie->uniquecharcount
21340 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
21341 sv_catpvs(sv, "[");
21342 (void) put_charclass_bitmap_innards(sv,
21343 ((IS_ANYOF_TRIE(op))
21345 : TRIE_BITMAP(trie)),
21352 sv_catpvs(sv, "]");
21354 } else if (k == CURLY) {
21355 U32 lo = ARG1(o), hi = ARG2(o);
21356 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
21357 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
21358 Perl_sv_catpvf(aTHX_ sv, "{%u,", (unsigned) lo);
21359 if (hi == REG_INFTY)
21360 sv_catpvs(sv, "INFTY");
21362 Perl_sv_catpvf(aTHX_ sv, "%u", (unsigned) hi);
21363 sv_catpvs(sv, "}");
21365 else if (k == WHILEM && o->flags) /* Ordinal/of */
21366 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
21367 else if (k == REF || k == OPEN || k == CLOSE
21368 || k == GROUPP || OP(o)==ACCEPT)
21370 AV *name_list= NULL;
21371 U32 parno= OP(o) == ACCEPT ? (U32)ARG2L(o) : ARG(o);
21372 Perl_sv_catpvf(aTHX_ sv, "%" UVuf, (UV)parno); /* Parenth number */
21373 if ( RXp_PAREN_NAMES(prog) ) {
21374 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
21375 } else if ( pRExC_state ) {
21376 name_list= RExC_paren_name_list;
21379 if ( k != REF || (OP(o) < REFN)) {
21380 SV **name= av_fetch(name_list, parno, 0 );
21382 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
21385 SV *sv_dat= MUTABLE_SV(progi->data->data[ parno ]);
21386 I32 *nums=(I32*)SvPVX(sv_dat);
21387 SV **name= av_fetch(name_list, nums[0], 0 );
21390 for ( n=0; n<SvIVX(sv_dat); n++ ) {
21391 Perl_sv_catpvf(aTHX_ sv, "%s%" IVdf,
21392 (n ? "," : ""), (IV)nums[n]);
21394 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
21398 if ( k == REF && reginfo) {
21399 U32 n = ARG(o); /* which paren pair */
21400 I32 ln = prog->offs[n].start;
21401 if (prog->lastparen < n || ln == -1 || prog->offs[n].end == -1)
21402 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
21403 else if (ln == prog->offs[n].end)
21404 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
21406 const char *s = reginfo->strbeg + ln;
21407 Perl_sv_catpvf(aTHX_ sv, ": ");
21408 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
21409 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
21412 } else if (k == GOSUB) {
21413 AV *name_list= NULL;
21414 if ( RXp_PAREN_NAMES(prog) ) {
21415 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
21416 } else if ( pRExC_state ) {
21417 name_list= RExC_paren_name_list;
21420 /* Paren and offset */
21421 Perl_sv_catpvf(aTHX_ sv, "%d[%+d:%d]", (int)ARG(o),(int)ARG2L(o),
21422 (int)((o + (int)ARG2L(o)) - progi->program) );
21424 SV **name= av_fetch(name_list, ARG(o), 0 );
21426 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
21429 else if (k == LOGICAL)
21430 /* 2: embedded, otherwise 1 */
21431 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
21432 else if (k == ANYOF || k == ANYOFR) {
21436 bool do_sep = FALSE; /* Do we need to separate various components of
21438 /* Set if there is still an unresolved user-defined property */
21439 SV *unresolved = NULL;
21441 /* Things that are ignored except when the runtime locale is UTF-8 */
21442 SV *only_utf8_locale_invlist = NULL;
21444 /* Code points that don't fit in the bitmap */
21445 SV *nonbitmap_invlist = NULL;
21447 /* And things that aren't in the bitmap, but are small enough to be */
21448 SV* bitmap_range_not_in_bitmap = NULL;
21452 if (inRANGE(OP(o), ANYOFH, ANYOFRb)) {
21458 flags = ANYOF_FLAGS(o);
21459 bitmap = ANYOF_BITMAP(o);
21463 if (OP(o) == ANYOFL || OP(o) == ANYOFPOSIXL) {
21464 if (ANYOFL_UTF8_LOCALE_REQD(flags)) {
21465 sv_catpvs(sv, "{utf8-locale-reqd}");
21467 if (flags & ANYOFL_FOLD) {
21468 sv_catpvs(sv, "{i}");
21472 inverted = flags & ANYOF_INVERT;
21474 /* If there is stuff outside the bitmap, get it */
21475 if (arg != ANYOF_ONLY_HAS_BITMAP) {
21476 if (inRANGE(OP(o), ANYOFR, ANYOFRb)) {
21477 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
21479 ANYOFRbase(o) + ANYOFRdelta(o));
21482 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
21483 (void) get_regclass_nonbitmap_data(prog, o, FALSE,
21485 &only_utf8_locale_invlist,
21486 &nonbitmap_invlist);
21488 (void) get_re_gclass_nonbitmap_data(prog, o, FALSE,
21490 &only_utf8_locale_invlist,
21491 &nonbitmap_invlist);
21495 /* The non-bitmap data may contain stuff that could fit in the
21496 * bitmap. This could come from a user-defined property being
21497 * finally resolved when this call was done; or much more likely
21498 * because there are matches that require UTF-8 to be valid, and so
21499 * aren't in the bitmap (or ANYOFR). This is teased apart later */
21500 _invlist_intersection(nonbitmap_invlist,
21502 &bitmap_range_not_in_bitmap);
21503 /* Leave just the things that don't fit into the bitmap */
21504 _invlist_subtract(nonbitmap_invlist,
21506 &nonbitmap_invlist);
21509 /* Obey this flag to add all above-the-bitmap code points */
21510 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
21511 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
21512 NUM_ANYOF_CODE_POINTS,
21516 /* Ready to start outputting. First, the initial left bracket */
21517 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
21519 /* ANYOFH by definition doesn't have anything that will fit inside the
21520 * bitmap; ANYOFR may or may not. */
21521 if ( ! inRANGE(OP(o), ANYOFH, ANYOFHr)
21522 && ( ! inRANGE(OP(o), ANYOFR, ANYOFRb)
21523 || ANYOFRbase(o) < NUM_ANYOF_CODE_POINTS))
21525 /* Then all the things that could fit in the bitmap */
21526 do_sep = put_charclass_bitmap_innards(sv,
21528 bitmap_range_not_in_bitmap,
21529 only_utf8_locale_invlist,
21533 /* Can't try inverting for a
21534 * better display if there
21535 * are things that haven't
21538 || inRANGE(OP(o), ANYOFR, ANYOFRb));
21539 SvREFCNT_dec(bitmap_range_not_in_bitmap);
21541 /* If there are user-defined properties which haven't been defined
21542 * yet, output them. If the result is not to be inverted, it is
21543 * clearest to output them in a separate [] from the bitmap range
21544 * stuff. If the result is to be complemented, we have to show
21545 * everything in one [], as the inversion applies to the whole
21546 * thing. Use {braces} to separate them from anything in the
21547 * bitmap and anything above the bitmap. */
21550 if (! do_sep) { /* If didn't output anything in the bitmap
21552 sv_catpvs(sv, "^");
21554 sv_catpvs(sv, "{");
21557 Perl_sv_catpvf(aTHX_ sv,"%s][%s", PL_colors[1],
21560 sv_catsv(sv, unresolved);
21562 sv_catpvs(sv, "}");
21564 do_sep = ! inverted;
21568 /* And, finally, add the above-the-bitmap stuff */
21569 if (nonbitmap_invlist && _invlist_len(nonbitmap_invlist)) {
21572 /* See if truncation size is overridden */
21573 const STRLEN dump_len = (PL_dump_re_max_len > 256)
21574 ? PL_dump_re_max_len
21577 /* This is output in a separate [] */
21579 Perl_sv_catpvf(aTHX_ sv,"%s][%s", PL_colors[1], PL_colors[0]);
21582 /* And, for easy of understanding, it is shown in the
21583 * uncomplemented form if possible. The one exception being if
21584 * there are unresolved items, where the inversion has to be
21585 * delayed until runtime */
21586 if (inverted && ! unresolved) {
21587 _invlist_invert(nonbitmap_invlist);
21588 _invlist_subtract(nonbitmap_invlist, PL_InBitmap, &nonbitmap_invlist);
21591 contents = invlist_contents(nonbitmap_invlist,
21592 FALSE /* output suitable for catsv */
21595 /* If the output is shorter than the permissible maximum, just do it. */
21596 if (SvCUR(contents) <= dump_len) {
21597 sv_catsv(sv, contents);
21600 const char * contents_string = SvPVX(contents);
21601 STRLEN i = dump_len;
21603 /* Otherwise, start at the permissible max and work back to the
21604 * first break possibility */
21605 while (i > 0 && contents_string[i] != ' ') {
21608 if (i == 0) { /* Fail-safe. Use the max if we couldn't
21609 find a legal break */
21613 sv_catpvn(sv, contents_string, i);
21614 sv_catpvs(sv, "...");
21617 SvREFCNT_dec_NN(contents);
21618 SvREFCNT_dec_NN(nonbitmap_invlist);
21621 /* And finally the matching, closing ']' */
21622 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
21624 if (OP(o) == ANYOFHs) {
21625 Perl_sv_catpvf(aTHX_ sv, " (Leading UTF-8 bytes=%s", _byte_dump_string((U8 *) ((struct regnode_anyofhs *) o)->string, FLAGS(o), 1));
21627 else if (inRANGE(OP(o), ANYOFH, ANYOFRb)) {
21628 U8 lowest = (OP(o) != ANYOFHr)
21630 : LOWEST_ANYOF_HRx_BYTE(FLAGS(o));
21631 U8 highest = (OP(o) == ANYOFHr)
21632 ? HIGHEST_ANYOF_HRx_BYTE(FLAGS(o))
21633 : (OP(o) == ANYOFH || OP(o) == ANYOFR)
21637 if (OP(o) != ANYOFR || ! isASCII(ANYOFRbase(o) + ANYOFRdelta(o)))
21640 Perl_sv_catpvf(aTHX_ sv, " (First UTF-8 byte=%02X", lowest);
21641 if (lowest != highest) {
21642 Perl_sv_catpvf(aTHX_ sv, "-%02X", highest);
21644 Perl_sv_catpvf(aTHX_ sv, ")");
21648 SvREFCNT_dec(unresolved);
21650 else if (k == ANYOFM) {
21651 SV * cp_list = get_ANYOFM_contents(o);
21653 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
21654 if (OP(o) == NANYOFM) {
21655 _invlist_invert(cp_list);
21658 put_charclass_bitmap_innards(sv, NULL, cp_list, NULL, NULL, 0, TRUE);
21659 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
21661 SvREFCNT_dec(cp_list);
21663 else if (k == POSIXD || k == NPOSIXD) {
21664 U8 index = FLAGS(o) * 2;
21665 if (index < C_ARRAY_LENGTH(anyofs)) {
21666 if (*anyofs[index] != '[') {
21667 sv_catpvs(sv, "[");
21669 sv_catpv(sv, anyofs[index]);
21670 if (*anyofs[index] != '[') {
21671 sv_catpvs(sv, "]");
21675 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
21678 else if (k == BOUND || k == NBOUND) {
21679 /* Must be synced with order of 'bound_type' in regcomp.h */
21680 const char * const bounds[] = {
21681 "", /* Traditional */
21687 assert(FLAGS(o) < C_ARRAY_LENGTH(bounds));
21688 sv_catpv(sv, bounds[FLAGS(o)]);
21690 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH)) {
21691 Perl_sv_catpvf(aTHX_ sv, "[%d", -(o->flags));
21693 Perl_sv_catpvf(aTHX_ sv, "..-%d", o->flags - o->next_off);
21695 Perl_sv_catpvf(aTHX_ sv, "]");
21697 else if (OP(o) == SBOL)
21698 Perl_sv_catpvf(aTHX_ sv, " /%s/", o->flags ? "\\A" : "^");
21700 /* add on the verb argument if there is one */
21701 if ( ( k == VERB || OP(o) == ACCEPT || OP(o) == OPFAIL ) && o->flags) {
21703 Perl_sv_catpvf(aTHX_ sv, ":%" SVf,
21704 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
21706 sv_catpvs(sv, ":NULL");
21709 PERL_UNUSED_CONTEXT;
21710 PERL_UNUSED_ARG(sv);
21711 PERL_UNUSED_ARG(o);
21712 PERL_UNUSED_ARG(prog);
21713 PERL_UNUSED_ARG(reginfo);
21714 PERL_UNUSED_ARG(pRExC_state);
21715 #endif /* DEBUGGING */
21721 Perl_re_intuit_string(pTHX_ REGEXP * const r)
21722 { /* Assume that RE_INTUIT is set */
21723 /* Returns an SV containing a string that must appear in the target for it
21724 * to match, or NULL if nothing is known that must match.
21726 * CAUTION: the SV can be freed during execution of the regex engine */
21728 struct regexp *const prog = ReANY(r);
21729 DECLARE_AND_GET_RE_DEBUG_FLAGS;
21731 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
21732 PERL_UNUSED_CONTEXT;
21736 if (prog->maxlen > 0) {
21737 const char * const s = SvPV_nolen_const(RX_UTF8(r)
21738 ? prog->check_utf8 : prog->check_substr);
21740 if (!PL_colorset) reginitcolors();
21741 Perl_re_printf( aTHX_
21742 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
21744 RX_UTF8(r) ? "utf8 " : "",
21745 PL_colors[5], PL_colors[0],
21748 (strlen(s) > PL_dump_re_max_len ? "..." : ""));
21752 /* use UTF8 check substring if regexp pattern itself is in UTF8 */
21753 return RX_UTF8(r) ? prog->check_utf8 : prog->check_substr;
21759 handles refcounting and freeing the perl core regexp structure. When
21760 it is necessary to actually free the structure the first thing it
21761 does is call the 'free' method of the regexp_engine associated to
21762 the regexp, allowing the handling of the void *pprivate; member
21763 first. (This routine is not overridable by extensions, which is why
21764 the extensions free is called first.)
21766 See regdupe and regdupe_internal if you change anything here.
21768 #ifndef PERL_IN_XSUB_RE
21770 Perl_pregfree(pTHX_ REGEXP *r)
21776 Perl_pregfree2(pTHX_ REGEXP *rx)
21778 struct regexp *const r = ReANY(rx);
21779 DECLARE_AND_GET_RE_DEBUG_FLAGS;
21781 PERL_ARGS_ASSERT_PREGFREE2;
21786 if (r->mother_re) {
21787 ReREFCNT_dec(r->mother_re);
21789 CALLREGFREE_PVT(rx); /* free the private data */
21790 SvREFCNT_dec(RXp_PAREN_NAMES(r));
21794 for (i = 0; i < 2; i++) {
21795 SvREFCNT_dec(r->substrs->data[i].substr);
21796 SvREFCNT_dec(r->substrs->data[i].utf8_substr);
21798 Safefree(r->substrs);
21800 RX_MATCH_COPY_FREE(rx);
21801 #ifdef PERL_ANY_COW
21802 SvREFCNT_dec(r->saved_copy);
21805 SvREFCNT_dec(r->qr_anoncv);
21806 if (r->recurse_locinput)
21807 Safefree(r->recurse_locinput);
21813 Copy ssv to dsv, both of which should of type SVt_REGEXP or SVt_PVLV,
21814 except that dsv will be created if NULL.
21816 This function is used in two main ways. First to implement
21817 $r = qr/....; $s = $$r;
21819 Secondly, it is used as a hacky workaround to the structural issue of
21821 being stored in the regexp structure which is in turn stored in
21822 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
21823 could be PL_curpm in multiple contexts, and could require multiple
21824 result sets being associated with the pattern simultaneously, such
21825 as when doing a recursive match with (??{$qr})
21827 The solution is to make a lightweight copy of the regexp structure
21828 when a qr// is returned from the code executed by (??{$qr}) this
21829 lightweight copy doesn't actually own any of its data except for
21830 the starp/end and the actual regexp structure itself.
21836 Perl_reg_temp_copy(pTHX_ REGEXP *dsv, REGEXP *ssv)
21838 struct regexp *drx;
21839 struct regexp *const srx = ReANY(ssv);
21840 const bool islv = dsv && SvTYPE(dsv) == SVt_PVLV;
21842 PERL_ARGS_ASSERT_REG_TEMP_COPY;
21845 dsv = (REGEXP*) newSV_type(SVt_REGEXP);
21847 assert(SvTYPE(dsv) == SVt_REGEXP || (SvTYPE(dsv) == SVt_PVLV));
21849 /* our only valid caller, sv_setsv_flags(), should have done
21850 * a SV_CHECK_THINKFIRST_COW_DROP() by now */
21851 assert(!SvOOK(dsv));
21852 assert(!SvIsCOW(dsv));
21853 assert(!SvROK(dsv));
21855 if (SvPVX_const(dsv)) {
21857 Safefree(SvPVX(dsv));
21862 SvOK_off((SV *)dsv);
21865 /* For PVLVs, the head (sv_any) points to an XPVLV, while
21866 * the LV's xpvlenu_rx will point to a regexp body, which
21867 * we allocate here */
21868 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
21869 assert(!SvPVX(dsv));
21870 ((XPV*)SvANY(dsv))->xpv_len_u.xpvlenu_rx = temp->sv_any;
21871 temp->sv_any = NULL;
21872 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
21873 SvREFCNT_dec_NN(temp);
21874 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
21875 ing below will not set it. */
21876 SvCUR_set(dsv, SvCUR(ssv));
21879 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
21880 sv_force_normal(sv) is called. */
21884 SvFLAGS(dsv) |= SvFLAGS(ssv) & (SVf_POK|SVp_POK|SVf_UTF8);
21885 SvPV_set(dsv, RX_WRAPPED(ssv));
21886 /* We share the same string buffer as the original regexp, on which we
21887 hold a reference count, incremented when mother_re is set below.
21888 The string pointer is copied here, being part of the regexp struct.
21890 memcpy(&(drx->xpv_cur), &(srx->xpv_cur),
21891 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
21895 const I32 npar = srx->nparens+1;
21896 Newx(drx->offs, npar, regexp_paren_pair);
21897 Copy(srx->offs, drx->offs, npar, regexp_paren_pair);
21899 if (srx->substrs) {
21901 Newx(drx->substrs, 1, struct reg_substr_data);
21902 StructCopy(srx->substrs, drx->substrs, struct reg_substr_data);
21904 for (i = 0; i < 2; i++) {
21905 SvREFCNT_inc_void(drx->substrs->data[i].substr);
21906 SvREFCNT_inc_void(drx->substrs->data[i].utf8_substr);
21909 /* check_substr and check_utf8, if non-NULL, point to either their
21910 anchored or float namesakes, and don't hold a second reference. */
21912 RX_MATCH_COPIED_off(dsv);
21913 #ifdef PERL_ANY_COW
21914 drx->saved_copy = NULL;
21916 drx->mother_re = ReREFCNT_inc(srx->mother_re ? srx->mother_re : ssv);
21917 SvREFCNT_inc_void(drx->qr_anoncv);
21918 if (srx->recurse_locinput)
21919 Newx(drx->recurse_locinput, srx->nparens + 1, char *);
21926 /* regfree_internal()
21928 Free the private data in a regexp. This is overloadable by
21929 extensions. Perl takes care of the regexp structure in pregfree(),
21930 this covers the *pprivate pointer which technically perl doesn't
21931 know about, however of course we have to handle the
21932 regexp_internal structure when no extension is in use.
21934 Note this is called before freeing anything in the regexp
21939 Perl_regfree_internal(pTHX_ REGEXP * const rx)
21941 struct regexp *const r = ReANY(rx);
21942 RXi_GET_DECL(r, ri);
21943 DECLARE_AND_GET_RE_DEBUG_FLAGS;
21945 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
21955 SV *dsv= sv_newmortal();
21956 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
21957 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), PL_dump_re_max_len);
21958 Perl_re_printf( aTHX_ "%sFreeing REx:%s %s\n",
21959 PL_colors[4], PL_colors[5], s);
21963 #ifdef RE_TRACK_PATTERN_OFFSETS
21965 Safefree(ri->u.offsets); /* 20010421 MJD */
21967 if (ri->code_blocks)
21968 S_free_codeblocks(aTHX_ ri->code_blocks);
21971 int n = ri->data->count;
21974 /* If you add a ->what type here, update the comment in regcomp.h */
21975 switch (ri->data->what[n]) {
21981 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
21984 Safefree(ri->data->data[n]);
21990 { /* Aho Corasick add-on structure for a trie node.
21991 Used in stclass optimization only */
21993 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
21995 refcount = --aho->refcount;
21998 PerlMemShared_free(aho->states);
21999 PerlMemShared_free(aho->fail);
22000 /* do this last!!!! */
22001 PerlMemShared_free(ri->data->data[n]);
22002 /* we should only ever get called once, so
22003 * assert as much, and also guard the free
22004 * which /might/ happen twice. At the least
22005 * it will make code anlyzers happy and it
22006 * doesn't cost much. - Yves */
22007 assert(ri->regstclass);
22008 if (ri->regstclass) {
22009 PerlMemShared_free(ri->regstclass);
22010 ri->regstclass = 0;
22017 /* trie structure. */
22019 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
22021 refcount = --trie->refcount;
22024 PerlMemShared_free(trie->charmap);
22025 PerlMemShared_free(trie->states);
22026 PerlMemShared_free(trie->trans);
22028 PerlMemShared_free(trie->bitmap);
22030 PerlMemShared_free(trie->jump);
22031 PerlMemShared_free(trie->wordinfo);
22032 /* do this last!!!! */
22033 PerlMemShared_free(ri->data->data[n]);
22038 Perl_croak(aTHX_ "panic: regfree data code '%c'",
22039 ri->data->what[n]);
22042 Safefree(ri->data->what);
22043 Safefree(ri->data);
22049 #define av_dup_inc(s, t) MUTABLE_AV(sv_dup_inc((const SV *)s, t))
22050 #define hv_dup_inc(s, t) MUTABLE_HV(sv_dup_inc((const SV *)s, t))
22051 #define SAVEPVN(p, n) ((p) ? savepvn(p, n) : NULL)
22054 =for apidoc re_dup_guts
22055 Duplicate a regexp.
22057 This routine is expected to clone a given regexp structure. It is only
22058 compiled under USE_ITHREADS.
22060 After all of the core data stored in struct regexp is duplicated
22061 the C<regexp_engine.dupe> method is used to copy any private data
22062 stored in the *pprivate pointer. This allows extensions to handle
22063 any duplication they need to do.
22067 See pregfree() and regfree_internal() if you change anything here.
22069 #if defined(USE_ITHREADS)
22070 #ifndef PERL_IN_XSUB_RE
22072 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
22075 const struct regexp *r = ReANY(sstr);
22076 struct regexp *ret = ReANY(dstr);
22078 PERL_ARGS_ASSERT_RE_DUP_GUTS;
22080 npar = r->nparens+1;
22081 Newx(ret->offs, npar, regexp_paren_pair);
22082 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
22084 if (ret->substrs) {
22085 /* Do it this way to avoid reading from *r after the StructCopy().
22086 That way, if any of the sv_dup_inc()s dislodge *r from the L1
22087 cache, it doesn't matter. */
22089 const bool anchored = r->check_substr
22090 ? r->check_substr == r->substrs->data[0].substr
22091 : r->check_utf8 == r->substrs->data[0].utf8_substr;
22092 Newx(ret->substrs, 1, struct reg_substr_data);
22093 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
22095 for (i = 0; i < 2; i++) {
22096 ret->substrs->data[i].substr =
22097 sv_dup_inc(ret->substrs->data[i].substr, param);
22098 ret->substrs->data[i].utf8_substr =
22099 sv_dup_inc(ret->substrs->data[i].utf8_substr, param);
22102 /* check_substr and check_utf8, if non-NULL, point to either their
22103 anchored or float namesakes, and don't hold a second reference. */
22105 if (ret->check_substr) {
22107 assert(r->check_utf8 == r->substrs->data[0].utf8_substr);
22109 ret->check_substr = ret->substrs->data[0].substr;
22110 ret->check_utf8 = ret->substrs->data[0].utf8_substr;
22112 assert(r->check_substr == r->substrs->data[1].substr);
22113 assert(r->check_utf8 == r->substrs->data[1].utf8_substr);
22115 ret->check_substr = ret->substrs->data[1].substr;
22116 ret->check_utf8 = ret->substrs->data[1].utf8_substr;
22118 } else if (ret->check_utf8) {
22120 ret->check_utf8 = ret->substrs->data[0].utf8_substr;
22122 ret->check_utf8 = ret->substrs->data[1].utf8_substr;
22127 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
22128 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
22129 if (r->recurse_locinput)
22130 Newx(ret->recurse_locinput, r->nparens + 1, char *);
22133 RXi_SET(ret, CALLREGDUPE_PVT(dstr, param));
22135 if (RX_MATCH_COPIED(dstr))
22136 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
22138 ret->subbeg = NULL;
22139 #ifdef PERL_ANY_COW
22140 ret->saved_copy = NULL;
22143 /* Whether mother_re be set or no, we need to copy the string. We
22144 cannot refrain from copying it when the storage points directly to
22145 our mother regexp, because that's
22146 1: a buffer in a different thread
22147 2: something we no longer hold a reference on
22148 so we need to copy it locally. */
22149 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED_const(sstr), SvCUR(sstr)+1);
22150 /* set malloced length to a non-zero value so it will be freed
22151 * (otherwise in combination with SVf_FAKE it looks like an alien
22152 * buffer). It doesn't have to be the actual malloced size, since it
22153 * should never be grown */
22154 SvLEN_set(dstr, SvCUR(sstr)+1);
22155 ret->mother_re = NULL;
22157 #endif /* PERL_IN_XSUB_RE */
22162 This is the internal complement to regdupe() which is used to copy
22163 the structure pointed to by the *pprivate pointer in the regexp.
22164 This is the core version of the extension overridable cloning hook.
22165 The regexp structure being duplicated will be copied by perl prior
22166 to this and will be provided as the regexp *r argument, however
22167 with the /old/ structures pprivate pointer value. Thus this routine
22168 may override any copying normally done by perl.
22170 It returns a pointer to the new regexp_internal structure.
22174 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
22176 struct regexp *const r = ReANY(rx);
22177 regexp_internal *reti;
22179 RXi_GET_DECL(r, ri);
22181 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
22185 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
22186 char, regexp_internal);
22187 Copy(ri->program, reti->program, len+1, regnode);
22190 if (ri->code_blocks) {
22192 Newx(reti->code_blocks, 1, struct reg_code_blocks);
22193 Newx(reti->code_blocks->cb, ri->code_blocks->count,
22194 struct reg_code_block);
22195 Copy(ri->code_blocks->cb, reti->code_blocks->cb,
22196 ri->code_blocks->count, struct reg_code_block);
22197 for (n = 0; n < ri->code_blocks->count; n++)
22198 reti->code_blocks->cb[n].src_regex = (REGEXP*)
22199 sv_dup_inc((SV*)(ri->code_blocks->cb[n].src_regex), param);
22200 reti->code_blocks->count = ri->code_blocks->count;
22201 reti->code_blocks->refcnt = 1;
22204 reti->code_blocks = NULL;
22206 reti->regstclass = NULL;
22209 struct reg_data *d;
22210 const int count = ri->data->count;
22213 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
22214 char, struct reg_data);
22215 Newx(d->what, count, U8);
22218 for (i = 0; i < count; i++) {
22219 d->what[i] = ri->data->what[i];
22220 switch (d->what[i]) {
22221 /* see also regcomp.h and regfree_internal() */
22222 case 'a': /* actually an AV, but the dup function is identical.
22223 values seem to be "plain sv's" generally. */
22224 case 'r': /* a compiled regex (but still just another SV) */
22225 case 's': /* an RV (currently only used for an RV to an AV by the ANYOF code)
22226 this use case should go away, the code could have used
22227 'a' instead - see S_set_ANYOF_arg() for array contents. */
22228 case 'S': /* actually an SV, but the dup function is identical. */
22229 case 'u': /* actually an HV, but the dup function is identical.
22230 values are "plain sv's" */
22231 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
22234 /* Synthetic Start Class - "Fake" charclass we generate to optimize
22235 * patterns which could start with several different things. Pre-TRIE
22236 * this was more important than it is now, however this still helps
22237 * in some places, for instance /x?a+/ might produce a SSC equivalent
22238 * to [xa]. This is used by Perl_re_intuit_start() and S_find_byclass()
22241 /* This is cheating. */
22242 Newx(d->data[i], 1, regnode_ssc);
22243 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
22244 reti->regstclass = (regnode*)d->data[i];
22247 /* AHO-CORASICK fail table */
22248 /* Trie stclasses are readonly and can thus be shared
22249 * without duplication. We free the stclass in pregfree
22250 * when the corresponding reg_ac_data struct is freed.
22252 reti->regstclass= ri->regstclass;
22255 /* TRIE transition table */
22257 ((reg_trie_data*)ri->data->data[i])->refcount++;
22260 case 'l': /* (?{...}) or (??{ ... }) code (cb->block) */
22261 case 'L': /* same when RExC_pm_flags & PMf_HAS_CV and code
22262 is not from another regexp */
22263 d->data[i] = ri->data->data[i];
22266 Perl_croak(aTHX_ "panic: re_dup_guts unknown data code '%c'",
22267 ri->data->what[i]);
22276 reti->name_list_idx = ri->name_list_idx;
22278 #ifdef RE_TRACK_PATTERN_OFFSETS
22279 if (ri->u.offsets) {
22280 Newx(reti->u.offsets, 2*len+1, U32);
22281 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
22284 SetProgLen(reti, len);
22287 return (void*)reti;
22290 #endif /* USE_ITHREADS */
22292 #ifndef PERL_IN_XSUB_RE
22295 - regnext - dig the "next" pointer out of a node
22298 Perl_regnext(pTHX_ regnode *p)
22305 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
22306 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
22307 (int)OP(p), (int)REGNODE_MAX);
22310 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
22320 S_re_croak(pTHX_ bool utf8, const char* pat,...)
22323 STRLEN len = strlen(pat);
22326 const char *message;
22328 PERL_ARGS_ASSERT_RE_CROAK;
22332 Copy(pat, buf, len , char);
22334 buf[len + 1] = '\0';
22335 va_start(args, pat);
22336 msv = vmess(buf, &args);
22338 message = SvPV_const(msv, len);
22341 Copy(message, buf, len , char);
22342 /* len-1 to avoid \n */
22343 Perl_croak(aTHX_ "%" UTF8f, UTF8fARG(utf8, len-1, buf));
22346 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
22348 #ifndef PERL_IN_XSUB_RE
22350 Perl_save_re_context(pTHX)
22355 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
22358 const REGEXP * const rx = PM_GETRE(PL_curpm);
22360 nparens = RX_NPARENS(rx);
22363 /* RT #124109. This is a complete hack; in the SWASHNEW case we know
22364 * that PL_curpm will be null, but that utf8.pm and the modules it
22365 * loads will only use $1..$3.
22366 * The t/porting/re_context.t test file checks this assumption.
22371 for (i = 1; i <= nparens; i++) {
22372 char digits[TYPE_CHARS(long)];
22373 const STRLEN len = my_snprintf(digits, sizeof(digits),
22375 GV *const *const gvp
22376 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
22379 GV * const gv = *gvp;
22380 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
22390 S_put_code_point(pTHX_ SV *sv, UV c)
22392 PERL_ARGS_ASSERT_PUT_CODE_POINT;
22395 Perl_sv_catpvf(aTHX_ sv, "\\x{%04" UVXf "}", c);
22397 else if (isPRINT(c)) {
22398 const char string = (char) c;
22400 /* We use {phrase} as metanotation in the class, so also escape literal
22402 if (isBACKSLASHED_PUNCT(c) || c == '{' || c == '}')
22403 sv_catpvs(sv, "\\");
22404 sv_catpvn(sv, &string, 1);
22406 else if (isMNEMONIC_CNTRL(c)) {
22407 Perl_sv_catpvf(aTHX_ sv, "%s", cntrl_to_mnemonic((U8) c));
22410 Perl_sv_catpvf(aTHX_ sv, "\\x%02X", (U8) c);
22415 S_put_range(pTHX_ SV *sv, UV start, const UV end, const bool allow_literals)
22417 /* Appends to 'sv' a displayable version of the range of code points from
22418 * 'start' to 'end'. Mnemonics (like '\r') are used for the few controls
22419 * that have them, when they occur at the beginning or end of the range.
22420 * It uses hex to output the remaining code points, unless 'allow_literals'
22421 * is true, in which case the printable ASCII ones are output as-is (though
22422 * some of these will be escaped by put_code_point()).
22424 * NOTE: This is designed only for printing ranges of code points that fit
22425 * inside an ANYOF bitmap. Higher code points are simply suppressed
22428 const unsigned int min_range_count = 3;
22430 assert(start <= end);
22432 PERL_ARGS_ASSERT_PUT_RANGE;
22434 while (start <= end) {
22436 const char * format;
22438 if ( end - start < min_range_count
22439 && (end - start <= 2 || (isPRINT_A(start) && isPRINT_A(end))))
22441 /* Output a range of 1 or 2 chars individually, or longer ranges
22442 * when printable */
22443 for (; start <= end; start++) {
22444 put_code_point(sv, start);
22449 /* If permitted by the input options, and there is a possibility that
22450 * this range contains a printable literal, look to see if there is
22452 if (allow_literals && start <= MAX_PRINT_A) {
22454 /* If the character at the beginning of the range isn't an ASCII
22455 * printable, effectively split the range into two parts:
22456 * 1) the portion before the first such printable,
22458 * and output them separately. */
22459 if (! isPRINT_A(start)) {
22460 UV temp_end = start + 1;
22462 /* There is no point looking beyond the final possible
22463 * printable, in MAX_PRINT_A */
22464 UV max = MIN(end, MAX_PRINT_A);
22466 while (temp_end <= max && ! isPRINT_A(temp_end)) {
22470 /* Here, temp_end points to one beyond the first printable if
22471 * found, or to one beyond 'max' if not. If none found, make
22472 * sure that we use the entire range */
22473 if (temp_end > MAX_PRINT_A) {
22474 temp_end = end + 1;
22477 /* Output the first part of the split range: the part that
22478 * doesn't have printables, with the parameter set to not look
22479 * for literals (otherwise we would infinitely recurse) */
22480 put_range(sv, start, temp_end - 1, FALSE);
22482 /* The 2nd part of the range (if any) starts here. */
22485 /* We do a continue, instead of dropping down, because even if
22486 * the 2nd part is non-empty, it could be so short that we want
22487 * to output it as individual characters, as tested for at the
22488 * top of this loop. */
22492 /* Here, 'start' is a printable ASCII. If it is an alphanumeric,
22493 * output a sub-range of just the digits or letters, then process
22494 * the remaining portion as usual. */
22495 if (isALPHANUMERIC_A(start)) {
22496 UV mask = (isDIGIT_A(start))
22501 UV temp_end = start + 1;
22503 /* Find the end of the sub-range that includes just the
22504 * characters in the same class as the first character in it */
22505 while (temp_end <= end && _generic_isCC_A(temp_end, mask)) {
22510 /* For short ranges, don't duplicate the code above to output
22511 * them; just call recursively */
22512 if (temp_end - start < min_range_count) {
22513 put_range(sv, start, temp_end, FALSE);
22515 else { /* Output as a range */
22516 put_code_point(sv, start);
22517 sv_catpvs(sv, "-");
22518 put_code_point(sv, temp_end);
22520 start = temp_end + 1;
22524 /* We output any other printables as individual characters */
22525 if (isPUNCT_A(start) || isSPACE_A(start)) {
22526 while (start <= end && (isPUNCT_A(start)
22527 || isSPACE_A(start)))
22529 put_code_point(sv, start);
22534 } /* End of looking for literals */
22536 /* Here is not to output as a literal. Some control characters have
22537 * mnemonic names. Split off any of those at the beginning and end of
22538 * the range to print mnemonically. It isn't possible for many of
22539 * these to be in a row, so this won't overwhelm with output */
22541 && (isMNEMONIC_CNTRL(start) || isMNEMONIC_CNTRL(end)))
22543 while (isMNEMONIC_CNTRL(start) && start <= end) {
22544 put_code_point(sv, start);
22548 /* If this didn't take care of the whole range ... */
22549 if (start <= end) {
22551 /* Look backwards from the end to find the final non-mnemonic
22554 while (isMNEMONIC_CNTRL(temp_end)) {
22558 /* And separately output the interior range that doesn't start
22559 * or end with mnemonics */
22560 put_range(sv, start, temp_end, FALSE);
22562 /* Then output the mnemonic trailing controls */
22563 start = temp_end + 1;
22564 while (start <= end) {
22565 put_code_point(sv, start);
22572 /* As a final resort, output the range or subrange as hex. */
22574 if (start >= NUM_ANYOF_CODE_POINTS) {
22577 else { /* Have to split range at the bitmap boundary */
22578 this_end = (end < NUM_ANYOF_CODE_POINTS)
22580 : NUM_ANYOF_CODE_POINTS - 1;
22582 #if NUM_ANYOF_CODE_POINTS > 256
22583 format = (this_end < 256)
22584 ? "\\x%02" UVXf "-\\x%02" UVXf
22585 : "\\x{%04" UVXf "}-\\x{%04" UVXf "}";
22587 format = "\\x%02" UVXf "-\\x%02" UVXf;
22589 GCC_DIAG_IGNORE_STMT(-Wformat-nonliteral);
22590 Perl_sv_catpvf(aTHX_ sv, format, start, this_end);
22591 GCC_DIAG_RESTORE_STMT;
22597 S_put_charclass_bitmap_innards_invlist(pTHX_ SV *sv, SV* invlist)
22599 /* Concatenate onto the PV in 'sv' a displayable form of the inversion list
22603 bool allow_literals = TRUE;
22605 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_INVLIST;
22607 /* Generally, it is more readable if printable characters are output as
22608 * literals, but if a range (nearly) spans all of them, it's best to output
22609 * it as a single range. This code will use a single range if all but 2
22610 * ASCII printables are in it */
22611 invlist_iterinit(invlist);
22612 while (invlist_iternext(invlist, &start, &end)) {
22614 /* If the range starts beyond the final printable, it doesn't have any
22616 if (start > MAX_PRINT_A) {
22620 /* In both ASCII and EBCDIC, a SPACE is the lowest printable. To span
22621 * all but two, the range must start and end no later than 2 from
22623 if (start < ' ' + 2 && end > MAX_PRINT_A - 2) {
22624 if (end > MAX_PRINT_A) {
22630 if (end - start >= MAX_PRINT_A - ' ' - 2) {
22631 allow_literals = FALSE;
22636 invlist_iterfinish(invlist);
22638 /* Here we have figured things out. Output each range */
22639 invlist_iterinit(invlist);
22640 while (invlist_iternext(invlist, &start, &end)) {
22641 if (start >= NUM_ANYOF_CODE_POINTS) {
22644 put_range(sv, start, end, allow_literals);
22646 invlist_iterfinish(invlist);
22652 S_put_charclass_bitmap_innards_common(pTHX_
22653 SV* invlist, /* The bitmap */
22654 SV* posixes, /* Under /l, things like [:word:], \S */
22655 SV* only_utf8, /* Under /d, matches iff the target is UTF-8 */
22656 SV* not_utf8, /* /d, matches iff the target isn't UTF-8 */
22657 SV* only_utf8_locale, /* Under /l, matches if the locale is UTF-8 */
22658 const bool invert /* Is the result to be inverted? */
22661 /* Create and return an SV containing a displayable version of the bitmap
22662 * and associated information determined by the input parameters. If the
22663 * output would have been only the inversion indicator '^', NULL is instead
22668 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_COMMON;
22671 output = newSVpvs("^");
22674 output = newSVpvs("");
22677 /* First, the code points in the bitmap that are unconditionally there */
22678 put_charclass_bitmap_innards_invlist(output, invlist);
22680 /* Traditionally, these have been placed after the main code points */
22682 sv_catsv(output, posixes);
22685 if (only_utf8 && _invlist_len(only_utf8)) {
22686 Perl_sv_catpvf(aTHX_ output, "%s{utf8}%s", PL_colors[1], PL_colors[0]);
22687 put_charclass_bitmap_innards_invlist(output, only_utf8);
22690 if (not_utf8 && _invlist_len(not_utf8)) {
22691 Perl_sv_catpvf(aTHX_ output, "%s{not utf8}%s", PL_colors[1], PL_colors[0]);
22692 put_charclass_bitmap_innards_invlist(output, not_utf8);
22695 if (only_utf8_locale && _invlist_len(only_utf8_locale)) {
22696 Perl_sv_catpvf(aTHX_ output, "%s{utf8 locale}%s", PL_colors[1], PL_colors[0]);
22697 put_charclass_bitmap_innards_invlist(output, only_utf8_locale);
22699 /* This is the only list in this routine that can legally contain code
22700 * points outside the bitmap range. The call just above to
22701 * 'put_charclass_bitmap_innards_invlist' will simply suppress them, so
22702 * output them here. There's about a half-dozen possible, and none in
22703 * contiguous ranges longer than 2 */
22704 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
22706 SV* above_bitmap = NULL;
22708 _invlist_subtract(only_utf8_locale, PL_InBitmap, &above_bitmap);
22710 invlist_iterinit(above_bitmap);
22711 while (invlist_iternext(above_bitmap, &start, &end)) {
22714 for (i = start; i <= end; i++) {
22715 put_code_point(output, i);
22718 invlist_iterfinish(above_bitmap);
22719 SvREFCNT_dec_NN(above_bitmap);
22723 if (invert && SvCUR(output) == 1) {
22731 S_put_charclass_bitmap_innards(pTHX_ SV *sv,
22733 SV *nonbitmap_invlist,
22734 SV *only_utf8_locale_invlist,
22735 const regnode * const node,
22737 const bool force_as_is_display)
22739 /* Appends to 'sv' a displayable version of the innards of the bracketed
22740 * character class defined by the other arguments:
22741 * 'bitmap' points to the bitmap, or NULL if to ignore that.
22742 * 'nonbitmap_invlist' is an inversion list of the code points that are in
22743 * the bitmap range, but for some reason aren't in the bitmap; NULL if
22744 * none. The reasons for this could be that they require some
22745 * condition such as the target string being or not being in UTF-8
22746 * (under /d), or because they came from a user-defined property that
22747 * was not resolved at the time of the regex compilation (under /u)
22748 * 'only_utf8_locale_invlist' is an inversion list of the code points that
22749 * are valid only if the runtime locale is a UTF-8 one; NULL if none
22750 * 'node' is the regex pattern ANYOF node. It is needed only when the
22751 * above two parameters are not null, and is passed so that this
22752 * routine can tease apart the various reasons for them.
22753 * 'flags' is the flags field of 'node'
22754 * 'force_as_is_display' is TRUE if this routine should definitely NOT try
22755 * to invert things to see if that leads to a cleaner display. If
22756 * FALSE, this routine is free to use its judgment about doing this.
22758 * It returns TRUE if there was actually something output. (It may be that
22759 * the bitmap, etc is empty.)
22761 * When called for outputting the bitmap of a non-ANYOF node, just pass the
22762 * bitmap, with the succeeding parameters set to NULL, and the final one to
22766 /* In general, it tries to display the 'cleanest' representation of the
22767 * innards, choosing whether to display them inverted or not, regardless of
22768 * whether the class itself is to be inverted. However, there are some
22769 * cases where it can't try inverting, as what actually matches isn't known
22770 * until runtime, and hence the inversion isn't either. */
22772 bool inverting_allowed = ! force_as_is_display;
22775 STRLEN orig_sv_cur = SvCUR(sv);
22777 SV* invlist; /* Inversion list we accumulate of code points that
22778 are unconditionally matched */
22779 SV* only_utf8 = NULL; /* Under /d, list of matches iff the target is
22781 SV* not_utf8 = NULL; /* /d, list of matches iff the target isn't UTF-8
22783 SV* posixes = NULL; /* Under /l, string of things like [:word:], \D */
22784 SV* only_utf8_locale = NULL; /* Under /l, list of matches if the locale
22787 SV* as_is_display; /* The output string when we take the inputs
22789 SV* inverted_display; /* The output string when we invert the inputs */
22791 bool invert = cBOOL(flags & ANYOF_INVERT); /* Is the input to be inverted
22793 /* We are biased in favor of displaying things without them being inverted,
22794 * as that is generally easier to understand */
22795 const int bias = 5;
22797 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS;
22799 /* Start off with whatever code points are passed in. (We clone, so we
22800 * don't change the caller's list) */
22801 if (nonbitmap_invlist) {
22802 assert(invlist_highest(nonbitmap_invlist) < NUM_ANYOF_CODE_POINTS);
22803 invlist = invlist_clone(nonbitmap_invlist, NULL);
22805 else { /* Worst case size is every other code point is matched */
22806 invlist = _new_invlist(NUM_ANYOF_CODE_POINTS / 2);
22810 if (OP(node) == ANYOFD) {
22812 /* This flag indicates that the code points below 0x100 in the
22813 * nonbitmap list are precisely the ones that match only when the
22814 * target is UTF-8 (they should all be non-ASCII). */
22815 if (flags & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)
22817 _invlist_intersection(invlist, PL_UpperLatin1, &only_utf8);
22818 _invlist_subtract(invlist, only_utf8, &invlist);
22821 /* And this flag for matching all non-ASCII 0xFF and below */
22822 if (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)
22824 not_utf8 = invlist_clone(PL_UpperLatin1, NULL);
22827 else if (OP(node) == ANYOFL || OP(node) == ANYOFPOSIXL) {
22829 /* If either of these flags are set, what matches isn't
22830 * determinable except during execution, so don't know enough here
22832 if (flags & (ANYOFL_FOLD|ANYOF_MATCHES_POSIXL)) {
22833 inverting_allowed = FALSE;
22836 /* What the posix classes match also varies at runtime, so these
22837 * will be output symbolically. */
22838 if (ANYOF_POSIXL_TEST_ANY_SET(node)) {
22841 posixes = newSVpvs("");
22842 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
22843 if (ANYOF_POSIXL_TEST(node, i)) {
22844 sv_catpv(posixes, anyofs[i]);
22851 /* Accumulate the bit map into the unconditional match list */
22853 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
22854 if (BITMAP_TEST(bitmap, i)) {
22857 i < NUM_ANYOF_CODE_POINTS && BITMAP_TEST(bitmap, i);
22860 invlist = _add_range_to_invlist(invlist, start, i-1);
22865 /* Make sure that the conditional match lists don't have anything in them
22866 * that match unconditionally; otherwise the output is quite confusing.
22867 * This could happen if the code that populates these misses some
22870 _invlist_subtract(only_utf8, invlist, &only_utf8);
22873 _invlist_subtract(not_utf8, invlist, ¬_utf8);
22876 if (only_utf8_locale_invlist) {
22878 /* Since this list is passed in, we have to make a copy before
22880 only_utf8_locale = invlist_clone(only_utf8_locale_invlist, NULL);
22882 _invlist_subtract(only_utf8_locale, invlist, &only_utf8_locale);
22884 /* And, it can get really weird for us to try outputting an inverted
22885 * form of this list when it has things above the bitmap, so don't even
22887 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
22888 inverting_allowed = FALSE;
22892 /* Calculate what the output would be if we take the input as-is */
22893 as_is_display = put_charclass_bitmap_innards_common(invlist,
22900 /* If have to take the output as-is, just do that */
22901 if (! inverting_allowed) {
22902 if (as_is_display) {
22903 sv_catsv(sv, as_is_display);
22904 SvREFCNT_dec_NN(as_is_display);
22907 else { /* But otherwise, create the output again on the inverted input, and
22908 use whichever version is shorter */
22910 int inverted_bias, as_is_bias;
22912 /* We will apply our bias to whichever of the results doesn't have
22922 inverted_bias = bias;
22925 /* Now invert each of the lists that contribute to the output,
22926 * excluding from the result things outside the possible range */
22928 /* For the unconditional inversion list, we have to add in all the
22929 * conditional code points, so that when inverted, they will be gone
22931 _invlist_union(only_utf8, invlist, &invlist);
22932 _invlist_union(not_utf8, invlist, &invlist);
22933 _invlist_union(only_utf8_locale, invlist, &invlist);
22934 _invlist_invert(invlist);
22935 _invlist_intersection(invlist, PL_InBitmap, &invlist);
22938 _invlist_invert(only_utf8);
22939 _invlist_intersection(only_utf8, PL_UpperLatin1, &only_utf8);
22941 else if (not_utf8) {
22943 /* If a code point matches iff the target string is not in UTF-8,
22944 * then complementing the result has it not match iff not in UTF-8,
22945 * which is the same thing as matching iff it is UTF-8. */
22946 only_utf8 = not_utf8;
22950 if (only_utf8_locale) {
22951 _invlist_invert(only_utf8_locale);
22952 _invlist_intersection(only_utf8_locale,
22954 &only_utf8_locale);
22957 inverted_display = put_charclass_bitmap_innards_common(
22962 only_utf8_locale, invert);
22964 /* Use the shortest representation, taking into account our bias
22965 * against showing it inverted */
22966 if ( inverted_display
22967 && ( ! as_is_display
22968 || ( SvCUR(inverted_display) + inverted_bias
22969 < SvCUR(as_is_display) + as_is_bias)))
22971 sv_catsv(sv, inverted_display);
22973 else if (as_is_display) {
22974 sv_catsv(sv, as_is_display);
22977 SvREFCNT_dec(as_is_display);
22978 SvREFCNT_dec(inverted_display);
22981 SvREFCNT_dec_NN(invlist);
22982 SvREFCNT_dec(only_utf8);
22983 SvREFCNT_dec(not_utf8);
22984 SvREFCNT_dec(posixes);
22985 SvREFCNT_dec(only_utf8_locale);
22987 return SvCUR(sv) > orig_sv_cur;
22990 #define CLEAR_OPTSTART \
22991 if (optstart) STMT_START { \
22992 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ \
22993 " (%" IVdf " nodes)\n", (IV)(node - optstart))); \
22997 #define DUMPUNTIL(b,e) \
22999 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
23001 STATIC const regnode *
23002 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
23003 const regnode *last, const regnode *plast,
23004 SV* sv, I32 indent, U32 depth)
23006 U8 op = PSEUDO; /* Arbitrary non-END op. */
23007 const regnode *next;
23008 const regnode *optstart= NULL;
23010 RXi_GET_DECL(r, ri);
23011 DECLARE_AND_GET_RE_DEBUG_FLAGS;
23013 PERL_ARGS_ASSERT_DUMPUNTIL;
23015 #ifdef DEBUG_DUMPUNTIL
23016 Perl_re_printf( aTHX_ "--- %d : %d - %d - %d\n", indent, node-start,
23017 last ? last-start : 0, plast ? plast-start : 0);
23020 if (plast && plast < last)
23023 while (PL_regkind[op] != END && (!last || node < last)) {
23025 /* While that wasn't END last time... */
23028 if (op == CLOSE || op == SRCLOSE || op == WHILEM)
23030 next = regnext((regnode *)node);
23033 if (OP(node) == OPTIMIZED) {
23034 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
23041 regprop(r, sv, node, NULL, NULL);
23042 Perl_re_printf( aTHX_ "%4" IVdf ":%*s%s", (IV)(node - start),
23043 (int)(2*indent + 1), "", SvPVX_const(sv));
23045 if (OP(node) != OPTIMIZED) {
23046 if (next == NULL) /* Next ptr. */
23047 Perl_re_printf( aTHX_ " (0)");
23048 else if (PL_regkind[(U8)op] == BRANCH
23049 && PL_regkind[OP(next)] != BRANCH )
23050 Perl_re_printf( aTHX_ " (FAIL)");
23052 Perl_re_printf( aTHX_ " (%" IVdf ")", (IV)(next - start));
23053 Perl_re_printf( aTHX_ "\n");
23057 if (PL_regkind[(U8)op] == BRANCHJ) {
23060 const regnode *nnode = (OP(next) == LONGJMP
23061 ? regnext((regnode *)next)
23063 if (last && nnode > last)
23065 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
23068 else if (PL_regkind[(U8)op] == BRANCH) {
23070 DUMPUNTIL(NEXTOPER(node), next);
23072 else if ( PL_regkind[(U8)op] == TRIE ) {
23073 const regnode *this_trie = node;
23074 const char op = OP(node);
23075 const U32 n = ARG(node);
23076 const reg_ac_data * const ac = op>=AHOCORASICK ?
23077 (reg_ac_data *)ri->data->data[n] :
23079 const reg_trie_data * const trie =
23080 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
23082 AV *const trie_words
23083 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
23085 const regnode *nextbranch= NULL;
23088 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
23089 SV ** const elem_ptr = av_fetch(trie_words, word_idx, 0);
23091 Perl_re_indentf( aTHX_ "%s ",
23094 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
23095 SvCUR(*elem_ptr), PL_dump_re_max_len,
23096 PL_colors[0], PL_colors[1],
23098 ? PERL_PV_ESCAPE_UNI
23100 | PERL_PV_PRETTY_ELLIPSES
23101 | PERL_PV_PRETTY_LTGT
23106 U16 dist= trie->jump[word_idx+1];
23107 Perl_re_printf( aTHX_ "(%" UVuf ")\n",
23108 (UV)((dist ? this_trie + dist : next) - start));
23111 nextbranch= this_trie + trie->jump[0];
23112 DUMPUNTIL(this_trie + dist, nextbranch);
23114 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
23115 nextbranch= regnext((regnode *)nextbranch);
23117 Perl_re_printf( aTHX_ "\n");
23120 if (last && next > last)
23125 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
23126 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
23127 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
23129 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
23131 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
23133 else if ( op == PLUS || op == STAR) {
23134 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
23136 else if (PL_regkind[(U8)op] == EXACT || op == ANYOFHs) {
23137 /* Literal string, where present. */
23138 node += NODE_SZ_STR(node) - 1;
23139 node = NEXTOPER(node);
23142 node = NEXTOPER(node);
23143 node += regarglen[(U8)op];
23145 if (op == CURLYX || op == OPEN || op == SROPEN)
23149 #ifdef DEBUG_DUMPUNTIL
23150 Perl_re_printf( aTHX_ "--- %d\n", (int)indent);
23155 #endif /* DEBUGGING */
23157 #ifndef PERL_IN_XSUB_RE
23159 # include "uni_keywords.h"
23162 Perl_init_uniprops(pTHX)
23166 char * dump_len_string;
23168 dump_len_string = PerlEnv_getenv("PERL_DUMP_RE_MAX_LEN");
23169 if ( ! dump_len_string
23170 || ! grok_atoUV(dump_len_string, (UV *)&PL_dump_re_max_len, NULL))
23172 PL_dump_re_max_len = 60; /* A reasonable default */
23176 PL_user_def_props = newHV();
23178 # ifdef USE_ITHREADS
23180 HvSHAREKEYS_off(PL_user_def_props);
23181 PL_user_def_props_aTHX = aTHX;
23185 /* Set up the inversion list interpreter-level variables */
23187 PL_XPosix_ptrs[_CC_ASCII] = _new_invlist_C_array(uni_prop_ptrs[UNI_ASCII]);
23188 PL_XPosix_ptrs[_CC_ALPHANUMERIC] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXALNUM]);
23189 PL_XPosix_ptrs[_CC_ALPHA] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXALPHA]);
23190 PL_XPosix_ptrs[_CC_BLANK] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXBLANK]);
23191 PL_XPosix_ptrs[_CC_CASED] = _new_invlist_C_array(uni_prop_ptrs[UNI_CASED]);
23192 PL_XPosix_ptrs[_CC_CNTRL] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXCNTRL]);
23193 PL_XPosix_ptrs[_CC_DIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXDIGIT]);
23194 PL_XPosix_ptrs[_CC_GRAPH] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXGRAPH]);
23195 PL_XPosix_ptrs[_CC_LOWER] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXLOWER]);
23196 PL_XPosix_ptrs[_CC_PRINT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXPRINT]);
23197 PL_XPosix_ptrs[_CC_PUNCT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXPUNCT]);
23198 PL_XPosix_ptrs[_CC_SPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXSPACE]);
23199 PL_XPosix_ptrs[_CC_UPPER] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXUPPER]);
23200 PL_XPosix_ptrs[_CC_VERTSPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_VERTSPACE]);
23201 PL_XPosix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXWORD]);
23202 PL_XPosix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXXDIGIT]);
23204 PL_Posix_ptrs[_CC_ASCII] = _new_invlist_C_array(uni_prop_ptrs[UNI_ASCII]);
23205 PL_Posix_ptrs[_CC_ALPHANUMERIC] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXALNUM]);
23206 PL_Posix_ptrs[_CC_ALPHA] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXALPHA]);
23207 PL_Posix_ptrs[_CC_BLANK] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXBLANK]);
23208 PL_Posix_ptrs[_CC_CASED] = PL_Posix_ptrs[_CC_ALPHA];
23209 PL_Posix_ptrs[_CC_CNTRL] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXCNTRL]);
23210 PL_Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXDIGIT]);
23211 PL_Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXGRAPH]);
23212 PL_Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXLOWER]);
23213 PL_Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXPRINT]);
23214 PL_Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXPUNCT]);
23215 PL_Posix_ptrs[_CC_SPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXSPACE]);
23216 PL_Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXUPPER]);
23217 PL_Posix_ptrs[_CC_VERTSPACE] = NULL;
23218 PL_Posix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXWORD]);
23219 PL_Posix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXXDIGIT]);
23221 PL_GCB_invlist = _new_invlist_C_array(_Perl_GCB_invlist);
23222 PL_SB_invlist = _new_invlist_C_array(_Perl_SB_invlist);
23223 PL_WB_invlist = _new_invlist_C_array(_Perl_WB_invlist);
23224 PL_LB_invlist = _new_invlist_C_array(_Perl_LB_invlist);
23225 PL_SCX_invlist = _new_invlist_C_array(_Perl_SCX_invlist);
23227 PL_InBitmap = _new_invlist_C_array(InBitmap_invlist);
23228 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
23229 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
23230 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
23232 PL_Assigned_invlist = _new_invlist_C_array(uni_prop_ptrs[UNI_ASSIGNED]);
23234 PL_utf8_perl_idstart = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_IDSTART]);
23235 PL_utf8_perl_idcont = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_IDCONT]);
23237 PL_utf8_charname_begin = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_CHARNAME_BEGIN]);
23238 PL_utf8_charname_continue = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_CHARNAME_CONTINUE]);
23240 PL_in_some_fold = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_ANY_FOLDS]);
23241 PL_HasMultiCharFold = _new_invlist_C_array(uni_prop_ptrs[
23242 UNI__PERL_FOLDS_TO_MULTI_CHAR]);
23243 PL_InMultiCharFold = _new_invlist_C_array(uni_prop_ptrs[
23244 UNI__PERL_IS_IN_MULTI_CHAR_FOLD]);
23245 PL_utf8_toupper = _new_invlist_C_array(Uppercase_Mapping_invlist);
23246 PL_utf8_tolower = _new_invlist_C_array(Lowercase_Mapping_invlist);
23247 PL_utf8_totitle = _new_invlist_C_array(Titlecase_Mapping_invlist);
23248 PL_utf8_tofold = _new_invlist_C_array(Case_Folding_invlist);
23249 PL_utf8_tosimplefold = _new_invlist_C_array(Simple_Case_Folding_invlist);
23250 PL_utf8_foldclosures = _new_invlist_C_array(_Perl_IVCF_invlist);
23251 PL_utf8_mark = _new_invlist_C_array(uni_prop_ptrs[UNI_M]);
23252 PL_CCC_non0_non230 = _new_invlist_C_array(_Perl_CCC_non0_non230_invlist);
23253 PL_Private_Use = _new_invlist_C_array(uni_prop_ptrs[UNI_CO]);
23256 /* The below are used only by deprecated functions. They could be removed */
23257 PL_utf8_xidcont = _new_invlist_C_array(uni_prop_ptrs[UNI_XIDC]);
23258 PL_utf8_idcont = _new_invlist_C_array(uni_prop_ptrs[UNI_IDC]);
23259 PL_utf8_xidstart = _new_invlist_C_array(uni_prop_ptrs[UNI_XIDS]);
23263 /* These four functions are compiled only in regcomp.c, where they have access
23264 * to the data they return. They are a way for re_comp.c to get access to that
23265 * data without having to compile the whole data structures. */
23268 Perl_do_uniprop_match(const char * const key, const U16 key_len)
23270 PERL_ARGS_ASSERT_DO_UNIPROP_MATCH;
23272 return match_uniprop((U8 *) key, key_len);
23276 Perl_get_prop_definition(pTHX_ const int table_index)
23278 PERL_ARGS_ASSERT_GET_PROP_DEFINITION;
23280 /* Create and return the inversion list */
23281 return _new_invlist_C_array(uni_prop_ptrs[table_index]);
23284 const char * const *
23285 Perl_get_prop_values(const int table_index)
23287 PERL_ARGS_ASSERT_GET_PROP_VALUES;
23289 return UNI_prop_value_ptrs[table_index];
23293 Perl_get_deprecated_property_msg(const Size_t warning_offset)
23295 PERL_ARGS_ASSERT_GET_DEPRECATED_PROPERTY_MSG;
23297 return deprecated_property_msgs[warning_offset];
23302 This code was mainly added for backcompat to give a warning for non-portable
23303 code points in user-defined properties. But experiments showed that the
23304 warning in earlier perls were only omitted on overflow, which should be an
23305 error, so there really isnt a backcompat issue, and actually adding the
23306 warning when none was present before might cause breakage, for little gain. So
23307 khw left this code in, but not enabled. Tests were never added.
23310 Ei |const char *|get_extended_utf8_msg|const UV cp
23312 PERL_STATIC_INLINE const char *
23313 S_get_extended_utf8_msg(pTHX_ const UV cp)
23315 U8 dummy[UTF8_MAXBYTES + 1];
23319 uvchr_to_utf8_flags_msgs(dummy, cp, UNICODE_WARN_PERL_EXTENDED,
23322 msg = hv_fetchs(msgs, "text", 0);
23325 (void) sv_2mortal((SV *) msgs);
23327 return SvPVX(*msg);
23331 #endif /* end of ! PERL_IN_XSUB_RE */
23334 S_compile_wildcard(pTHX_ const char * subpattern, const STRLEN len,
23335 const bool ignore_case)
23337 /* Pretends that the input subpattern is qr/subpattern/aam, compiling it
23338 * possibly with /i if the 'ignore_case' parameter is true. Use /aa
23339 * because nothing outside of ASCII will match. Use /m because the input
23340 * string may be a bunch of lines strung together.
23342 * Also sets up the debugging info */
23344 U32 flags = PMf_MULTILINE|PMf_WILDCARD;
23346 SV * subpattern_sv = sv_2mortal(newSVpvn(subpattern, len));
23347 REGEXP * subpattern_re;
23348 DECLARE_AND_GET_RE_DEBUG_FLAGS;
23350 PERL_ARGS_ASSERT_COMPILE_WILDCARD;
23355 set_regex_charset(&flags, REGEX_ASCII_MORE_RESTRICTED_CHARSET);
23357 /* Like in op.c, we copy the compile time pm flags to the rx ones */
23358 rx_flags = flags & RXf_PMf_COMPILETIME;
23360 #ifndef PERL_IN_XSUB_RE
23361 /* Use the core engine if this file is regcomp.c. That means no
23362 * 'use re "Debug ..." is in effect, so the core engine is sufficient */
23363 subpattern_re = Perl_re_op_compile(aTHX_ &subpattern_sv, 1, NULL,
23364 &PL_core_reg_engine,
23368 if (isDEBUG_WILDCARD) {
23369 /* Use the special debugging engine if this file is re_comp.c and wants
23370 * to output the wildcard matching. This uses whatever
23371 * 'use re "Debug ..." is in effect */
23372 subpattern_re = Perl_re_op_compile(aTHX_ &subpattern_sv, 1, NULL,
23378 /* Use the special wildcard engine if this file is re_comp.c and
23379 * doesn't want to output the wildcard matching. This uses whatever
23380 * 'use re "Debug ..." is in effect for compilation, but this engine
23381 * structure has been set up so that it uses the core engine for
23382 * execution, so no execution debugging as a result of re.pm will be
23384 subpattern_re = Perl_re_op_compile(aTHX_ &subpattern_sv, 1, NULL,
23388 /* XXX The above has the effect that any user-supplied regex engine
23389 * won't be called for matching wildcards. That might be good, or bad.
23390 * It could be changed in several ways. The reason it is done the
23391 * current way is to avoid having to save and restore
23392 * ^{^RE_DEBUG_FLAGS} around the execution. save_scalar() perhaps
23393 * could be used. Another suggestion is to keep the authoritative
23394 * value of the debug flags in a thread-local variable and add set/get
23395 * magic to ${^RE_DEBUG_FLAGS} to keep the C level variable up to date.
23396 * Still another is to pass a flag, say in the engine's intflags that
23397 * would be checked each time before doing the debug output */
23401 assert(subpattern_re); /* Should have died if didn't compile successfully */
23402 return subpattern_re;
23406 S_execute_wildcard(pTHX_ REGEXP * const prog, char* stringarg, char *strend,
23407 char *strbeg, SSize_t minend, SV *screamer, U32 nosave)
23410 DECLARE_AND_GET_RE_DEBUG_FLAGS;
23412 PERL_ARGS_ASSERT_EXECUTE_WILDCARD;
23416 /* The compilation has set things up so that if the program doesn't want to
23417 * see the wildcard matching procedure, it will get the core execution
23418 * engine, which is subject only to -Dr. So we have to turn that off
23419 * around this procedure */
23420 if (! isDEBUG_WILDCARD) {
23421 /* Note! Casts away 'volatile' */
23423 PL_debug &= ~ DEBUG_r_FLAG;
23426 result = CALLREGEXEC(prog, stringarg, strend, strbeg, minend, screamer,
23434 S_handle_user_defined_property(pTHX_
23436 /* Parses the contents of a user-defined property definition; returning the
23437 * expanded definition if possible. If so, the return is an inversion
23440 * If there are subroutines that are part of the expansion and which aren't
23441 * known at the time of the call to this function, this returns what
23442 * parse_uniprop_string() returned for the first one encountered.
23444 * If an error was found, NULL is returned, and 'msg' gets a suitable
23445 * message appended to it. (Appending allows the back trace of how we got
23446 * to the faulty definition to be displayed through nested calls of
23447 * user-defined subs.)
23449 * The caller IS responsible for freeing any returned SV.
23451 * The syntax of the contents is pretty much described in perlunicode.pod,
23452 * but we also allow comments on each line */
23454 const char * name, /* Name of property */
23455 const STRLEN name_len, /* The name's length in bytes */
23456 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
23457 const bool to_fold, /* ? Is this under /i */
23458 const bool runtime, /* ? Are we in compile- or run-time */
23459 const bool deferrable, /* Is it ok for this property's full definition
23460 to be deferred until later? */
23461 SV* contents, /* The property's definition */
23462 bool *user_defined_ptr, /* This will be set TRUE as we wouldn't be
23463 getting called unless this is thought to be
23464 a user-defined property */
23465 SV * msg, /* Any error or warning msg(s) are appended to
23467 const STRLEN level) /* Recursion level of this call */
23470 const char * string = SvPV_const(contents, len);
23471 const char * const e = string + len;
23472 const bool is_contents_utf8 = cBOOL(SvUTF8(contents));
23473 const STRLEN msgs_length_on_entry = SvCUR(msg);
23475 const char * s0 = string; /* Points to first byte in the current line
23476 being parsed in 'string' */
23477 const char overflow_msg[] = "Code point too large in \"";
23478 SV* running_definition = NULL;
23480 PERL_ARGS_ASSERT_HANDLE_USER_DEFINED_PROPERTY;
23482 *user_defined_ptr = TRUE;
23484 /* Look at each line */
23486 const char * s; /* Current byte */
23487 char op = '+'; /* Default operation is 'union' */
23488 IV min = 0; /* range begin code point */
23489 IV max = -1; /* and range end */
23490 SV* this_definition;
23492 /* Skip comment lines */
23494 s0 = strchr(s0, '\n');
23502 /* For backcompat, allow an empty first line */
23508 /* First character in the line may optionally be the operation */
23517 /* If the line is one or two hex digits separated by blank space, its
23518 * a range; otherwise it is either another user-defined property or an
23523 if (! isXDIGIT(*s)) {
23524 goto check_if_property;
23527 do { /* Each new hex digit will add 4 bits. */
23528 if (min > ( (IV) MAX_LEGAL_CP >> 4)) {
23529 s = strchr(s, '\n');
23533 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23534 sv_catpv(msg, overflow_msg);
23535 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
23536 UTF8fARG(is_contents_utf8, s - s0, s0));
23537 sv_catpvs(msg, "\"");
23538 goto return_failure;
23541 /* Accumulate this digit into the value */
23542 min = (min << 4) + READ_XDIGIT(s);
23543 } while (isXDIGIT(*s));
23545 while (isBLANK(*s)) { s++; }
23547 /* We allow comments at the end of the line */
23549 s = strchr(s, '\n');
23555 else if (s < e && *s != '\n') {
23556 if (! isXDIGIT(*s)) {
23557 goto check_if_property;
23560 /* Look for the high point of the range */
23563 if (max > ( (IV) MAX_LEGAL_CP >> 4)) {
23564 s = strchr(s, '\n');
23568 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23569 sv_catpv(msg, overflow_msg);
23570 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
23571 UTF8fARG(is_contents_utf8, s - s0, s0));
23572 sv_catpvs(msg, "\"");
23573 goto return_failure;
23576 max = (max << 4) + READ_XDIGIT(s);
23577 } while (isXDIGIT(*s));
23579 while (isBLANK(*s)) { s++; }
23582 s = strchr(s, '\n');
23587 else if (s < e && *s != '\n') {
23588 goto check_if_property;
23592 if (max == -1) { /* The line only had one entry */
23595 else if (max < min) {
23596 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23597 sv_catpvs(msg, "Illegal range in \"");
23598 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
23599 UTF8fARG(is_contents_utf8, s - s0, s0));
23600 sv_catpvs(msg, "\"");
23601 goto return_failure;
23604 # if 0 /* See explanation at definition above of get_extended_utf8_msg() */
23606 if ( UNICODE_IS_PERL_EXTENDED(min)
23607 || UNICODE_IS_PERL_EXTENDED(max))
23609 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23611 /* If both code points are non-portable, warn only on the lower
23613 sv_catpv(msg, get_extended_utf8_msg(
23614 (UNICODE_IS_PERL_EXTENDED(min))
23616 sv_catpvs(msg, " in \"");
23617 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
23618 UTF8fARG(is_contents_utf8, s - s0, s0));
23619 sv_catpvs(msg, "\"");
23624 /* Here, this line contains a legal range */
23625 this_definition = sv_2mortal(_new_invlist(2));
23626 this_definition = _add_range_to_invlist(this_definition, min, max);
23631 /* Here it isn't a legal range line. See if it is a legal property
23632 * line. First find the end of the meat of the line */
23633 s = strpbrk(s, "#\n");
23638 /* Ignore trailing blanks in keeping with the requirements of
23639 * parse_uniprop_string() */
23641 while (s > s0 && isBLANK_A(*s)) {
23646 this_definition = parse_uniprop_string(s0, s - s0,
23647 is_utf8, to_fold, runtime,
23650 user_defined_ptr, msg,
23652 ? level /* Don't increase level
23653 if input is empty */
23656 if (this_definition == NULL) {
23657 goto return_failure; /* 'msg' should have had the reason
23658 appended to it by the above call */
23661 if (! is_invlist(this_definition)) { /* Unknown at this time */
23662 return newSVsv(this_definition);
23666 s = strchr(s, '\n');
23676 _invlist_union(running_definition, this_definition,
23677 &running_definition);
23680 _invlist_subtract(running_definition, this_definition,
23681 &running_definition);
23684 _invlist_intersection(running_definition, this_definition,
23685 &running_definition);
23688 _invlist_union_complement_2nd(running_definition,
23689 this_definition, &running_definition);
23692 Perl_croak(aTHX_ "panic: %s: %d: Unexpected operation %d",
23693 __FILE__, __LINE__, op);
23697 /* Position past the '\n' */
23699 } /* End of loop through the lines of 'contents' */
23701 /* Here, we processed all the lines in 'contents' without error. If we
23702 * didn't add any warnings, simply return success */
23703 if (msgs_length_on_entry == SvCUR(msg)) {
23705 /* If the expansion was empty, the answer isn't nothing: its an empty
23706 * inversion list */
23707 if (running_definition == NULL) {
23708 running_definition = _new_invlist(1);
23711 return running_definition;
23714 /* Otherwise, add some explanatory text, but we will return success */
23718 running_definition = NULL;
23722 if (name_len > 0) {
23723 sv_catpvs(msg, " in expansion of ");
23724 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8, name_len, name));
23727 return running_definition;
23730 /* As explained below, certain operations need to take place in the first
23731 * thread created. These macros switch contexts */
23732 # ifdef USE_ITHREADS
23733 # define DECLARATION_FOR_GLOBAL_CONTEXT \
23734 PerlInterpreter * save_aTHX = aTHX;
23735 # define SWITCH_TO_GLOBAL_CONTEXT \
23736 PERL_SET_CONTEXT((aTHX = PL_user_def_props_aTHX))
23737 # define RESTORE_CONTEXT PERL_SET_CONTEXT((aTHX = save_aTHX));
23738 # define CUR_CONTEXT aTHX
23739 # define ORIGINAL_CONTEXT save_aTHX
23741 # define DECLARATION_FOR_GLOBAL_CONTEXT dNOOP
23742 # define SWITCH_TO_GLOBAL_CONTEXT NOOP
23743 # define RESTORE_CONTEXT NOOP
23744 # define CUR_CONTEXT NULL
23745 # define ORIGINAL_CONTEXT NULL
23749 S_delete_recursion_entry(pTHX_ void *key)
23751 /* Deletes the entry used to detect recursion when expanding user-defined
23752 * properties. This is a function so it can be set up to be called even if
23753 * the program unexpectedly quits */
23755 SV ** current_entry;
23756 const STRLEN key_len = strlen((const char *) key);
23757 DECLARATION_FOR_GLOBAL_CONTEXT;
23759 SWITCH_TO_GLOBAL_CONTEXT;
23761 /* If the entry is one of these types, it is a permanent entry, and not the
23762 * one used to detect recursions. This function should delete only the
23763 * recursion entry */
23764 current_entry = hv_fetch(PL_user_def_props, (const char *) key, key_len, 0);
23766 && ! is_invlist(*current_entry)
23767 && ! SvPOK(*current_entry))
23769 (void) hv_delete(PL_user_def_props, (const char *) key, key_len,
23777 S_get_fq_name(pTHX_
23778 const char * const name, /* The first non-blank in the \p{}, \P{} */
23779 const Size_t name_len, /* Its length in bytes, not including any trailing space */
23780 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
23781 const bool has_colon_colon
23784 /* Returns a mortal SV containing the fully qualified version of the input
23789 fq_name = newSVpvs_flags("", SVs_TEMP);
23791 /* Use the current package if it wasn't included in our input */
23792 if (! has_colon_colon) {
23793 const HV * pkg = (IN_PERL_COMPILETIME)
23795 : CopSTASH(PL_curcop);
23796 const char* pkgname = HvNAME(pkg);
23798 Perl_sv_catpvf(aTHX_ fq_name, "%" UTF8f,
23799 UTF8fARG(is_utf8, strlen(pkgname), pkgname));
23800 sv_catpvs(fq_name, "::");
23803 Perl_sv_catpvf(aTHX_ fq_name, "%" UTF8f,
23804 UTF8fARG(is_utf8, name_len, name));
23809 S_parse_uniprop_string(pTHX_
23811 /* Parse the interior of a \p{}, \P{}. Returns its definition if knowable
23812 * now. If so, the return is an inversion list.
23814 * If the property is user-defined, it is a subroutine, which in turn
23815 * may call other subroutines. This function will call the whole nest of
23816 * them to get the definition they return; if some aren't known at the time
23817 * of the call to this function, the fully qualified name of the highest
23818 * level sub is returned. It is an error to call this function at runtime
23819 * without every sub defined.
23821 * If an error was found, NULL is returned, and 'msg' gets a suitable
23822 * message appended to it. (Appending allows the back trace of how we got
23823 * to the faulty definition to be displayed through nested calls of
23824 * user-defined subs.)
23826 * The caller should NOT try to free any returned inversion list.
23828 * Other parameters will be set on return as described below */
23830 const char * const name, /* The first non-blank in the \p{}, \P{} */
23831 Size_t name_len, /* Its length in bytes, not including any
23833 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
23834 const bool to_fold, /* ? Is this under /i */
23835 const bool runtime, /* TRUE if this is being called at run time */
23836 const bool deferrable, /* TRUE if it's ok for the definition to not be
23837 known at this call */
23838 AV ** strings, /* To return string property values, like named
23840 bool *user_defined_ptr, /* Upon return from this function it will be
23841 set to TRUE if any component is a
23842 user-defined property */
23843 SV * msg, /* Any error or warning msg(s) are appended to
23845 const STRLEN level) /* Recursion level of this call */
23847 char* lookup_name; /* normalized name for lookup in our tables */
23848 unsigned lookup_len; /* Its length */
23849 enum { Not_Strict = 0, /* Some properties have stricter name */
23850 Strict, /* normalization rules, which we decide */
23851 As_Is /* upon based on parsing */
23852 } stricter = Not_Strict;
23854 /* nv= or numeric_value=, or possibly one of the cjk numeric properties
23855 * (though it requires extra effort to download them from Unicode and
23856 * compile perl to know about them) */
23857 bool is_nv_type = FALSE;
23859 unsigned int i, j = 0;
23860 int equals_pos = -1; /* Where the '=' is found, or negative if none */
23861 int slash_pos = -1; /* Where the '/' is found, or negative if none */
23862 int table_index = 0; /* The entry number for this property in the table
23863 of all Unicode property names */
23864 bool starts_with_Is = FALSE; /* ? Does the name start with 'Is' */
23865 Size_t lookup_offset = 0; /* Used to ignore the first few characters of
23866 the normalized name in certain situations */
23867 Size_t non_pkg_begin = 0; /* Offset of first byte in 'name' that isn't
23868 part of a package name */
23869 Size_t lun_non_pkg_begin = 0; /* Similarly for 'lookup_name' */
23870 bool could_be_user_defined = TRUE; /* ? Could this be a user-defined
23871 property rather than a Unicode
23873 SV * prop_definition = NULL; /* The returned definition of 'name' or NULL
23874 if an error. If it is an inversion list,
23875 it is the definition. Otherwise it is a
23876 string containing the fully qualified sub
23878 SV * fq_name = NULL; /* For user-defined properties, the fully
23880 bool invert_return = FALSE; /* ? Do we need to complement the result before
23882 bool stripped_utf8_pkg = FALSE; /* Set TRUE if the input includes an
23883 explicit utf8:: package that we strip
23885 /* The expansion of properties that could be either user-defined or
23886 * official unicode ones is deferred until runtime, including a marker for
23887 * those that might be in the latter category. This boolean indicates if
23888 * we've seen that marker. If not, what we're parsing can't be such an
23889 * official Unicode property whose expansion was deferred */
23890 bool could_be_deferred_official = FALSE;
23892 PERL_ARGS_ASSERT_PARSE_UNIPROP_STRING;
23894 /* The input will be normalized into 'lookup_name' */
23895 Newx(lookup_name, name_len, char);
23896 SAVEFREEPV(lookup_name);
23898 /* Parse the input. */
23899 for (i = 0; i < name_len; i++) {
23900 char cur = name[i];
23902 /* Most of the characters in the input will be of this ilk, being parts
23904 if (isIDCONT_A(cur)) {
23906 /* Case differences are ignored. Our lookup routine assumes
23907 * everything is lowercase, so normalize to that */
23908 if (isUPPER_A(cur)) {
23909 lookup_name[j++] = toLOWER_A(cur);
23913 if (cur == '_') { /* Don't include these in the normalized name */
23917 lookup_name[j++] = cur;
23919 /* The first character in a user-defined name must be of this type.
23921 if (i - non_pkg_begin == 0 && ! isIDFIRST_A(cur)) {
23922 could_be_user_defined = FALSE;
23928 /* Here, the character is not something typically in a name, But these
23929 * two types of characters (and the '_' above) can be freely ignored in
23930 * most situations. Later it may turn out we shouldn't have ignored
23931 * them, and we have to reparse, but we don't have enough information
23932 * yet to make that decision */
23933 if (cur == '-' || isSPACE_A(cur)) {
23934 could_be_user_defined = FALSE;
23938 /* An equals sign or single colon mark the end of the first part of
23939 * the property name */
23941 || (cur == ':' && (i >= name_len - 1 || name[i+1] != ':')))
23943 lookup_name[j++] = '='; /* Treat the colon as an '=' */
23944 equals_pos = j; /* Note where it occurred in the input */
23945 could_be_user_defined = FALSE;
23949 /* If this looks like it is a marker we inserted at compile time,
23950 * set a flag and otherwise ignore it. If it isn't in the final
23951 * position, keep it as it would have been user input. */
23952 if ( UNLIKELY(cur == DEFERRED_COULD_BE_OFFICIAL_MARKERc)
23954 && could_be_user_defined
23955 && i == name_len - 1)
23958 could_be_deferred_official = TRUE;
23962 /* Otherwise, this character is part of the name. */
23963 lookup_name[j++] = cur;
23965 /* Here it isn't a single colon, so if it is a colon, it must be a
23969 /* A double colon should be a package qualifier. We note its
23970 * position and continue. Note that one could have
23971 * pkg1::pkg2::...::foo
23972 * so that the position at the end of the loop will be just after
23973 * the final qualifier */
23976 non_pkg_begin = i + 1;
23977 lookup_name[j++] = ':';
23978 lun_non_pkg_begin = j;
23980 else { /* Only word chars (and '::') can be in a user-defined name */
23981 could_be_user_defined = FALSE;
23983 } /* End of parsing through the lhs of the property name (or all of it if
23986 # define STRLENs(s) (sizeof("" s "") - 1)
23988 /* If there is a single package name 'utf8::', it is ambiguous. It could
23989 * be for a user-defined property, or it could be a Unicode property, as
23990 * all of them are considered to be for that package. For the purposes of
23991 * parsing the rest of the property, strip it off */
23992 if (non_pkg_begin == STRLENs("utf8::") && memBEGINPs(name, name_len, "utf8::")) {
23993 lookup_name += STRLENs("utf8::");
23994 j -= STRLENs("utf8::");
23995 equals_pos -= STRLENs("utf8::");
23996 stripped_utf8_pkg = TRUE;
23999 /* Here, we are either done with the whole property name, if it was simple;
24000 * or are positioned just after the '=' if it is compound. */
24002 if (equals_pos >= 0) {
24003 assert(stricter == Not_Strict); /* We shouldn't have set this yet */
24005 /* Space immediately after the '=' is ignored */
24007 for (; i < name_len; i++) {
24008 if (! isSPACE_A(name[i])) {
24013 /* Most punctuation after the equals indicates a subpattern, like
24015 if ( isPUNCT_A(name[i])
24020 /* A backslash means the real delimitter is the next character,
24021 * but it must be punctuation */
24022 && (name[i] != '\\' || (i < name_len && isPUNCT_A(name[i+1]))))
24024 bool special_property = memEQs(lookup_name, j - 1, "name")
24025 || memEQs(lookup_name, j - 1, "na");
24026 if (! special_property) {
24027 /* Find the property. The table includes the equals sign, so
24028 * we use 'j' as-is */
24029 table_index = do_uniprop_match(lookup_name, j);
24031 if (special_property || table_index) {
24032 REGEXP * subpattern_re;
24033 char open = name[i++];
24035 const char * pos_in_brackets;
24036 const char * const * prop_values;
24039 /* Backslash => delimitter is the character following. We
24040 * already checked that it is punctuation */
24041 if (open == '\\') {
24046 /* This data structure is constructed so that the matching
24047 * closing bracket is 3 past its matching opening. The second
24048 * set of closing is so that if the opening is something like
24049 * ']', the closing will be that as well. Something similar is
24050 * done in toke.c */
24051 pos_in_brackets = memCHRs("([<)]>)]>", open);
24052 close = (pos_in_brackets) ? pos_in_brackets[3] : open;
24055 || name[name_len-1] != close
24056 || (escaped && name[name_len-2] != '\\')
24057 /* Also make sure that there are enough characters.
24058 * e.g., '\\\' would show up incorrectly as legal even
24059 * though it is too short */
24060 || (SSize_t) (name_len - i - 1 - escaped) < 0)
24062 sv_catpvs(msg, "Unicode property wildcard not terminated");
24063 goto append_name_to_msg;
24066 Perl_ck_warner_d(aTHX_
24067 packWARN(WARN_EXPERIMENTAL__UNIPROP_WILDCARDS),
24068 "The Unicode property wildcards feature is experimental");
24070 if (special_property) {
24071 const char * error_msg;
24072 const char * revised_name = name + i;
24073 Size_t revised_name_len = name_len - (i + 1 + escaped);
24075 /* Currently, the only 'special_property' is name, which we
24076 * lookup in _charnames.pm */
24078 if (! load_charnames(newSVpvs("placeholder"),
24079 revised_name, revised_name_len,
24082 sv_catpv(msg, error_msg);
24083 goto append_name_to_msg;
24086 /* Farm this out to a function just to make the current
24087 * function less unwieldy */
24088 if (handle_names_wildcard(revised_name, revised_name_len,
24092 return prop_definition;
24098 prop_values = get_prop_values(table_index);
24100 /* Now create and compile the wildcard subpattern. Use /i
24101 * because the property values are supposed to match with case
24103 subpattern_re = compile_wildcard(name + i,
24104 name_len - i - 1 - escaped,
24108 /* For each legal property value, see if the supplied pattern
24110 while (*prop_values) {
24111 const char * const entry = *prop_values;
24112 const Size_t len = strlen(entry);
24113 SV* entry_sv = newSVpvn_flags(entry, len, SVs_TEMP);
24115 if (execute_wildcard(subpattern_re,
24117 (char *) entry + len,
24121 { /* Here, matched. Add to the returned list */
24122 Size_t total_len = j + len;
24123 SV * sub_invlist = NULL;
24124 char * this_string;
24126 /* We know this is a legal \p{property=value}. Call
24127 * the function to return the list of code points that
24129 Newxz(this_string, total_len + 1, char);
24130 Copy(lookup_name, this_string, j, char);
24131 my_strlcat(this_string, entry, total_len + 1);
24132 SAVEFREEPV(this_string);
24133 sub_invlist = parse_uniprop_string(this_string,
24143 _invlist_union(prop_definition, sub_invlist,
24147 prop_values++; /* Next iteration, look at next propvalue */
24148 } /* End of looking through property values; (the data
24149 structure is terminated by a NULL ptr) */
24151 SvREFCNT_dec_NN(subpattern_re);
24153 if (prop_definition) {
24154 return prop_definition;
24157 sv_catpvs(msg, "No Unicode property value wildcard matches:");
24158 goto append_name_to_msg;
24161 /* Here's how khw thinks we should proceed to handle the properties
24162 * not yet done: Bidi Mirroring Glyph can map to ""
24163 Bidi Paired Bracket can map to ""
24164 Case Folding (both full and simple)
24165 Shouldn't /i be good enough for Full
24166 Decomposition Mapping
24167 Equivalent Unified Ideograph can map to ""
24168 Lowercase Mapping (both full and simple)
24169 NFKC Case Fold can map to ""
24170 Titlecase Mapping (both full and simple)
24171 Uppercase Mapping (both full and simple)
24172 * Handle these the same way Name is done, using say, _wild.pm, but
24173 * having both loose and full, like in charclass_invlists.h.
24174 * Perhaps move block and script to that as they are somewhat large
24175 * in charclass_invlists.h.
24176 * For properties where the default is the code point itself, such
24177 * as any of the case changing mappings, the string would otherwise
24178 * consist of all Unicode code points in UTF-8 strung together.
24179 * This would be impractical. So instead, examine their compiled
24180 * pattern, looking at the ssc. If none, reject the pattern as an
24181 * error. Otherwise run the pattern against every code point in
24182 * the ssc. The ssc is kind of like tr18's 3.9 Possible Match Sets
24183 * And it might be good to create an API to return the ssc.
24184 * Or handle them like the algorithmic names are done
24186 } /* End of is a wildcard subppattern */
24188 /* \p{name=...} is handled specially. Instead of using the normal
24189 * mechanism involving charclass_invlists.h, it uses _charnames.pm
24190 * which has the necessary (huge) data accessible to it, and which
24191 * doesn't get loaded unless necessary. The legal syntax for names is
24192 * somewhat different than other properties due both to the vagaries of
24193 * a few outlier official names, and the fact that only a few ASCII
24194 * characters are permitted in them */
24195 if ( memEQs(lookup_name, j - 1, "name")
24196 || memEQs(lookup_name, j - 1, "na"))
24201 const char * error_msg;
24203 SV * character_name;
24204 STRLEN character_len;
24209 /* Since the RHS (after skipping initial space) is passed unchanged
24210 * to charnames, and there are different criteria for what are
24211 * legal characters in the name, just parse it here. A character
24212 * name must begin with an ASCII alphabetic */
24213 if (! isALPHA(name[i])) {
24216 lookup_name[j++] = name[i];
24218 for (++i; i < name_len; i++) {
24219 /* Official names can only be in the ASCII range, and only
24220 * certain characters */
24221 if (! isASCII(name[i]) || ! isCHARNAME_CONT(name[i])) {
24224 lookup_name[j++] = name[i];
24227 /* Finished parsing, save the name into an SV */
24228 character_name = newSVpvn(lookup_name + equals_pos, j - equals_pos);
24230 /* Make sure _charnames is loaded. (The parameters give context
24231 * for any errors generated */
24232 table = load_charnames(character_name, name, name_len, &error_msg);
24233 if (table == NULL) {
24234 sv_catpv(msg, error_msg);
24235 goto append_name_to_msg;
24238 lookup_loose = get_cvs("_charnames::_loose_regcomp_lookup", 0);
24239 if (! lookup_loose) {
24241 "panic: Can't find '_charnames::_loose_regcomp_lookup");
24244 PUSHSTACKi(PERLSI_REGCOMP);
24250 XPUSHs(character_name);
24252 call_sv(MUTABLE_SV(lookup_loose), G_SCALAR);
24257 SvREFCNT_inc_simple_void_NN(character);
24264 if (! SvOK(character)) {
24268 cp = valid_utf8_to_uvchr((U8 *) SvPVX(character), &character_len);
24269 if (character_len == SvCUR(character)) {
24270 prop_definition = add_cp_to_invlist(NULL, cp);
24275 /* First of the remaining characters in the string. */
24276 char * remaining = SvPVX(character) + character_len;
24278 if (strings == NULL) {
24279 goto failed; /* XXX Perhaps a specific msg instead, like
24280 'not available here' */
24283 if (*strings == NULL) {
24284 *strings = newAV();
24287 this_string = newAV();
24288 av_push(this_string, newSVuv(cp));
24291 cp = valid_utf8_to_uvchr((U8 *) remaining, &character_len);
24292 av_push(this_string, newSVuv(cp));
24293 remaining += character_len;
24294 } while (remaining < SvEND(character));
24296 av_push(*strings, (SV *) this_string);
24299 return prop_definition;
24302 /* Certain properties whose values are numeric need special handling.
24303 * They may optionally be prefixed by 'is'. Ignore that prefix for the
24304 * purposes of checking if this is one of those properties */
24305 if (memBEGINPs(lookup_name, j, "is")) {
24309 /* Then check if it is one of these specially-handled properties. The
24310 * possibilities are hard-coded because easier this way, and the list
24311 * is unlikely to change.
24313 * All numeric value type properties are of this ilk, and are also
24314 * special in a different way later on. So find those first. There
24315 * are several numeric value type properties in the Unihan DB (which is
24316 * unlikely to be compiled with perl, but we handle it here in case it
24317 * does get compiled). They all end with 'numeric'. The interiors
24318 * aren't checked for the precise property. This would stop working if
24319 * a cjk property were to be created that ended with 'numeric' and
24320 * wasn't a numeric type */
24321 is_nv_type = memEQs(lookup_name + lookup_offset,
24322 j - 1 - lookup_offset, "numericvalue")
24323 || memEQs(lookup_name + lookup_offset,
24324 j - 1 - lookup_offset, "nv")
24325 || ( memENDPs(lookup_name + lookup_offset,
24326 j - 1 - lookup_offset, "numeric")
24327 && ( memBEGINPs(lookup_name + lookup_offset,
24328 j - 1 - lookup_offset, "cjk")
24329 || memBEGINPs(lookup_name + lookup_offset,
24330 j - 1 - lookup_offset, "k")));
24332 || memEQs(lookup_name + lookup_offset,
24333 j - 1 - lookup_offset, "canonicalcombiningclass")
24334 || memEQs(lookup_name + lookup_offset,
24335 j - 1 - lookup_offset, "ccc")
24336 || memEQs(lookup_name + lookup_offset,
24337 j - 1 - lookup_offset, "age")
24338 || memEQs(lookup_name + lookup_offset,
24339 j - 1 - lookup_offset, "in")
24340 || memEQs(lookup_name + lookup_offset,
24341 j - 1 - lookup_offset, "presentin"))
24345 /* Since the stuff after the '=' is a number, we can't throw away
24346 * '-' willy-nilly, as those could be a minus sign. Other stricter
24347 * rules also apply. However, these properties all can have the
24348 * rhs not be a number, in which case they contain at least one
24349 * alphabetic. In those cases, the stricter rules don't apply.
24350 * But the numeric type properties can have the alphas [Ee] to
24351 * signify an exponent, and it is still a number with stricter
24352 * rules. So look for an alpha that signifies not-strict */
24354 for (k = i; k < name_len; k++) {
24355 if ( isALPHA_A(name[k])
24356 && (! is_nv_type || ! isALPHA_FOLD_EQ(name[k], 'E')))
24358 stricter = Not_Strict;
24366 /* A number may have a leading '+' or '-'. The latter is retained
24368 if (name[i] == '+') {
24371 else if (name[i] == '-') {
24372 lookup_name[j++] = '-';
24376 /* Skip leading zeros including single underscores separating the
24377 * zeros, or between the final leading zero and the first other
24379 for (; i < name_len - 1; i++) {
24380 if ( name[i] != '0'
24381 && (name[i] != '_' || ! isDIGIT_A(name[i+1])))
24388 else { /* No '=' */
24390 /* Only a few properties without an '=' should be parsed with stricter
24391 * rules. The list is unlikely to change. */
24392 if ( memBEGINPs(lookup_name, j, "perl")
24393 && memNEs(lookup_name + 4, j - 4, "space")
24394 && memNEs(lookup_name + 4, j - 4, "word"))
24398 /* We set the inputs back to 0 and the code below will reparse,
24404 /* Here, we have either finished the property, or are positioned to parse
24405 * the remainder, and we know if stricter rules apply. Finish out, if not
24407 for (; i < name_len; i++) {
24408 char cur = name[i];
24410 /* In all instances, case differences are ignored, and we normalize to
24412 if (isUPPER_A(cur)) {
24413 lookup_name[j++] = toLOWER(cur);
24417 /* An underscore is skipped, but not under strict rules unless it
24418 * separates two digits */
24421 && ( i == 0 || (int) i == equals_pos || i == name_len- 1
24422 || ! isDIGIT_A(name[i-1]) || ! isDIGIT_A(name[i+1])))
24424 lookup_name[j++] = '_';
24429 /* Hyphens are skipped except under strict */
24430 if (cur == '-' && ! stricter) {
24434 /* XXX Bug in documentation. It says white space skipped adjacent to
24435 * non-word char. Maybe we should, but shouldn't skip it next to a dot
24437 if (isSPACE_A(cur) && ! stricter) {
24441 lookup_name[j++] = cur;
24443 /* Unless this is a non-trailing slash, we are done with it */
24444 if (i >= name_len - 1 || cur != '/') {
24450 /* A slash in the 'numeric value' property indicates that what follows
24451 * is a denominator. It can have a leading '+' and '0's that should be
24452 * skipped. But we have never allowed a negative denominator, so treat
24453 * a minus like every other character. (No need to rule out a second
24454 * '/', as that won't match anything anyway */
24457 if (i < name_len && name[i] == '+') {
24461 /* Skip leading zeros including underscores separating digits */
24462 for (; i < name_len - 1; i++) {
24463 if ( name[i] != '0'
24464 && (name[i] != '_' || ! isDIGIT_A(name[i+1])))
24470 /* Store the first real character in the denominator */
24471 if (i < name_len) {
24472 lookup_name[j++] = name[i];
24477 /* Here are completely done parsing the input 'name', and 'lookup_name'
24478 * contains a copy, normalized.
24480 * This special case is grandfathered in: 'L_' and 'GC=L_' are accepted and
24481 * different from without the underscores. */
24482 if ( ( UNLIKELY(memEQs(lookup_name, j, "l"))
24483 || UNLIKELY(memEQs(lookup_name, j, "gc=l")))
24484 && UNLIKELY(name[name_len-1] == '_'))
24486 lookup_name[j++] = '&';
24489 /* If the original input began with 'In' or 'Is', it could be a subroutine
24490 * call to a user-defined property instead of a Unicode property name. */
24491 if ( name_len - non_pkg_begin > 2
24492 && name[non_pkg_begin+0] == 'I'
24493 && (name[non_pkg_begin+1] == 'n' || name[non_pkg_begin+1] == 's'))
24495 /* Names that start with In have different characterstics than those
24496 * that start with Is */
24497 if (name[non_pkg_begin+1] == 's') {
24498 starts_with_Is = TRUE;
24502 could_be_user_defined = FALSE;
24505 if (could_be_user_defined) {
24508 /* If the user defined property returns the empty string, it could
24509 * easily be because the pattern is being compiled before the data it
24510 * actually needs to compile is available. This could be argued to be
24511 * a bug in the perl code, but this is a change of behavior for Perl,
24512 * so we handle it. This means that intentionally returning nothing
24513 * will not be resolved until runtime */
24514 bool empty_return = FALSE;
24516 /* Here, the name could be for a user defined property, which are
24517 * implemented as subs. */
24518 user_sub = get_cvn_flags(name, name_len, 0);
24521 /* Here, the property name could be a user-defined one, but there
24522 * is no subroutine to handle it (as of now). Defer handling it
24523 * until runtime. Otherwise, a block defined by Unicode in a later
24524 * release would get the synonym InFoo added for it, and existing
24525 * code that used that name would suddenly break if it referred to
24526 * the property before the sub was declared. See [perl #134146] */
24528 goto definition_deferred;
24531 /* Here, we are at runtime, and didn't find the user property. It
24532 * could be an official property, but only if no package was
24533 * specified, or just the utf8:: package. */
24534 if (could_be_deferred_official) {
24535 lookup_name += lun_non_pkg_begin;
24536 j -= lun_non_pkg_begin;
24538 else if (! stripped_utf8_pkg) {
24539 goto unknown_user_defined;
24542 /* Drop down to look up in the official properties */
24545 const char insecure[] = "Insecure user-defined property";
24547 /* Here, there is a sub by the correct name. Normally we call it
24548 * to get the property definition */
24550 SV * user_sub_sv = MUTABLE_SV(user_sub);
24551 SV * error; /* Any error returned by calling 'user_sub' */
24552 SV * key; /* The key into the hash of user defined sub names
24555 SV ** saved_user_prop_ptr; /* Hash entry for this property */
24557 /* How many times to retry when another thread is in the middle of
24558 * expanding the same definition we want */
24559 PERL_INT_FAST8_T retry_countdown = 10;
24561 DECLARATION_FOR_GLOBAL_CONTEXT;
24563 /* If we get here, we know this property is user-defined */
24564 *user_defined_ptr = TRUE;
24566 /* We refuse to call a potentially tainted subroutine; returning an
24569 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24570 sv_catpvn(msg, insecure, sizeof(insecure) - 1);
24571 goto append_name_to_msg;
24574 /* In principal, we only call each subroutine property definition
24575 * once during the life of the program. This guarantees that the
24576 * property definition never changes. The results of the single
24577 * sub call are stored in a hash, which is used instead for future
24578 * references to this property. The property definition is thus
24579 * immutable. But, to allow the user to have a /i-dependent
24580 * definition, we call the sub once for non-/i, and once for /i,
24581 * should the need arise, passing the /i status as a parameter.
24583 * We start by constructing the hash key name, consisting of the
24584 * fully qualified subroutine name, preceded by the /i status, so
24585 * that there is a key for /i and a different key for non-/i */
24586 key = newSVpvn(((to_fold) ? "1" : "0"), 1);
24587 fq_name = S_get_fq_name(aTHX_ name, name_len, is_utf8,
24588 non_pkg_begin != 0);
24589 sv_catsv(key, fq_name);
24592 /* We only call the sub once throughout the life of the program
24593 * (with the /i, non-/i exception noted above). That means the
24594 * hash must be global and accessible to all threads. It is
24595 * created at program start-up, before any threads are created, so
24596 * is accessible to all children. But this creates some
24599 * 1) The keys can't be shared, or else problems arise; sharing is
24600 * turned off at hash creation time
24601 * 2) All SVs in it are there for the remainder of the life of the
24602 * program, and must be created in the same interpreter context
24603 * as the hash, or else they will be freed from the wrong pool
24604 * at global destruction time. This is handled by switching to
24605 * the hash's context to create each SV going into it, and then
24606 * immediately switching back
24607 * 3) All accesses to the hash must be controlled by a mutex, to
24608 * prevent two threads from getting an unstable state should
24609 * they simultaneously be accessing it. The code below is
24610 * crafted so that the mutex is locked whenever there is an
24611 * access and unlocked only when the next stable state is
24614 * The hash stores either the definition of the property if it was
24615 * valid, or, if invalid, the error message that was raised. We
24616 * use the type of SV to distinguish.
24618 * There's also the need to guard against the definition expansion
24619 * from infinitely recursing. This is handled by storing the aTHX
24620 * of the expanding thread during the expansion. Again the SV type
24621 * is used to distinguish this from the other two cases. If we
24622 * come to here and the hash entry for this property is our aTHX,
24623 * it means we have recursed, and the code assumes that we would
24624 * infinitely recurse, so instead stops and raises an error.
24625 * (Any recursion has always been treated as infinite recursion in
24628 * If instead, the entry is for a different aTHX, it means that
24629 * that thread has gotten here first, and hasn't finished expanding
24630 * the definition yet. We just have to wait until it is done. We
24631 * sleep and retry a few times, returning an error if the other
24632 * thread doesn't complete. */
24635 USER_PROP_MUTEX_LOCK;
24637 /* If we have an entry for this key, the subroutine has already
24638 * been called once with this /i status. */
24639 saved_user_prop_ptr = hv_fetch(PL_user_def_props,
24640 SvPVX(key), SvCUR(key), 0);
24641 if (saved_user_prop_ptr) {
24643 /* If the saved result is an inversion list, it is the valid
24644 * definition of this property */
24645 if (is_invlist(*saved_user_prop_ptr)) {
24646 prop_definition = *saved_user_prop_ptr;
24648 /* The SV in the hash won't be removed until global
24649 * destruction, so it is stable and we can unlock */
24650 USER_PROP_MUTEX_UNLOCK;
24652 /* The caller shouldn't try to free this SV */
24653 return prop_definition;
24656 /* Otherwise, if it is a string, it is the error message
24657 * that was returned when we first tried to evaluate this
24658 * property. Fail, and append the message */
24659 if (SvPOK(*saved_user_prop_ptr)) {
24660 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24661 sv_catsv(msg, *saved_user_prop_ptr);
24663 /* The SV in the hash won't be removed until global
24664 * destruction, so it is stable and we can unlock */
24665 USER_PROP_MUTEX_UNLOCK;
24670 assert(SvIOK(*saved_user_prop_ptr));
24672 /* Here, we have an unstable entry in the hash. Either another
24673 * thread is in the middle of expanding the property's
24674 * definition, or we are ourselves recursing. We use the aTHX
24675 * in it to distinguish */
24676 if (SvIV(*saved_user_prop_ptr) != PTR2IV(CUR_CONTEXT)) {
24678 /* Here, it's another thread doing the expanding. We've
24679 * looked as much as we are going to at the contents of the
24680 * hash entry. It's safe to unlock. */
24681 USER_PROP_MUTEX_UNLOCK;
24683 /* Retry a few times */
24684 if (retry_countdown-- > 0) {
24689 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24690 sv_catpvs(msg, "Timeout waiting for another thread to "
24692 goto append_name_to_msg;
24695 /* Here, we are recursing; don't dig any deeper */
24696 USER_PROP_MUTEX_UNLOCK;
24698 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24700 "Infinite recursion in user-defined property");
24701 goto append_name_to_msg;
24704 /* Here, this thread has exclusive control, and there is no entry
24705 * for this property in the hash. So we have the go ahead to
24706 * expand the definition ourselves. */
24708 PUSHSTACKi(PERLSI_REGCOMP);
24711 /* Create a temporary placeholder in the hash to detect recursion
24713 SWITCH_TO_GLOBAL_CONTEXT;
24714 placeholder= newSVuv(PTR2IV(ORIGINAL_CONTEXT));
24715 (void) hv_store_ent(PL_user_def_props, key, placeholder, 0);
24718 /* Now that we have a placeholder, we can let other threads
24720 USER_PROP_MUTEX_UNLOCK;
24722 /* Make sure the placeholder always gets destroyed */
24723 SAVEDESTRUCTOR_X(S_delete_recursion_entry, SvPVX(key));
24728 /* Call the user's function, with the /i status as a parameter.
24729 * Note that we have gone to a lot of trouble to keep this call
24730 * from being within the locked mutex region. */
24731 XPUSHs(boolSV(to_fold));
24734 /* The following block was taken from swash_init(). Presumably
24735 * they apply to here as well, though we no longer use a swash --
24739 /* We might get here via a subroutine signature which uses a utf8
24740 * parameter name, at which point PL_subname will have been set
24741 * but not yet used. */
24742 save_item(PL_subname);
24744 /* G_SCALAR guarantees a single return value */
24745 (void) call_sv(user_sub_sv, G_EVAL|G_SCALAR);
24750 if (TAINT_get || SvTRUE(error)) {
24751 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24752 if (SvTRUE(error)) {
24753 sv_catpvs(msg, "Error \"");
24754 sv_catsv(msg, error);
24755 sv_catpvs(msg, "\"");
24758 if (SvTRUE(error)) sv_catpvs(msg, "; ");
24759 sv_catpvn(msg, insecure, sizeof(insecure) - 1);
24762 if (name_len > 0) {
24763 sv_catpvs(msg, " in expansion of ");
24764 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8,
24770 prop_definition = NULL;
24773 SV * contents = POPs;
24775 /* The contents is supposed to be the expansion of the property
24776 * definition. If the definition is deferrable, and we got an
24777 * empty string back, set a flag to later defer it (after clean
24780 && (! SvPOK(contents) || SvCUR(contents) == 0))
24782 empty_return = TRUE;
24784 else { /* Otherwise, call a function to check for valid syntax,
24787 prop_definition = handle_user_defined_property(
24789 is_utf8, to_fold, runtime,
24791 contents, user_defined_ptr,
24797 /* Here, we have the results of the expansion. Delete the
24798 * placeholder, and if the definition is now known, replace it with
24799 * that definition. We need exclusive access to the hash, and we
24800 * can't let anyone else in, between when we delete the placeholder
24801 * and add the permanent entry */
24802 USER_PROP_MUTEX_LOCK;
24804 S_delete_recursion_entry(aTHX_ SvPVX(key));
24806 if ( ! empty_return
24807 && (! prop_definition || is_invlist(prop_definition)))
24809 /* If we got success we use the inversion list defining the
24810 * property; otherwise use the error message */
24811 SWITCH_TO_GLOBAL_CONTEXT;
24812 (void) hv_store_ent(PL_user_def_props,
24815 ? newSVsv(prop_definition)
24821 /* All done, and the hash now has a permanent entry for this
24822 * property. Give up exclusive control */
24823 USER_PROP_MUTEX_UNLOCK;
24829 if (empty_return) {
24830 goto definition_deferred;
24833 if (prop_definition) {
24835 /* If the definition is for something not known at this time,
24836 * we toss it, and go return the main property name, as that's
24837 * the one the user will be aware of */
24838 if (! is_invlist(prop_definition)) {
24839 SvREFCNT_dec_NN(prop_definition);
24840 goto definition_deferred;
24843 sv_2mortal(prop_definition);
24847 return prop_definition;
24849 } /* End of calling the subroutine for the user-defined property */
24850 } /* End of it could be a user-defined property */
24852 /* Here it wasn't a user-defined property that is known at this time. See
24853 * if it is a Unicode property */
24855 lookup_len = j; /* This is a more mnemonic name than 'j' */
24857 /* Get the index into our pointer table of the inversion list corresponding
24858 * to the property */
24859 table_index = do_uniprop_match(lookup_name, lookup_len);
24861 /* If it didn't find the property ... */
24862 if (table_index == 0) {
24864 /* Try again stripping off any initial 'Is'. This is because we
24865 * promise that an initial Is is optional. The same isn't true of
24866 * names that start with 'In'. Those can match only blocks, and the
24867 * lookup table already has those accounted for. The lookup table also
24868 * has already accounted for Perl extensions (without and = sign)
24869 * starting with 'i's'. */
24870 if (starts_with_Is && equals_pos >= 0) {
24876 table_index = do_uniprop_match(lookup_name, lookup_len);
24879 if (table_index == 0) {
24882 /* Here, we didn't find it. If not a numeric type property, and
24883 * can't be a user-defined one, it isn't a legal property */
24884 if (! is_nv_type) {
24885 if (! could_be_user_defined) {
24889 /* Here, the property name is legal as a user-defined one. At
24890 * compile time, it might just be that the subroutine for that
24891 * property hasn't been encountered yet, but at runtime, it's
24892 * an error to try to use an undefined one */
24893 if (! deferrable) {
24894 goto unknown_user_defined;;
24897 goto definition_deferred;
24898 } /* End of isn't a numeric type property */
24900 /* The numeric type properties need more work to decide. What we
24901 * do is make sure we have the number in canonical form and look
24904 if (slash_pos < 0) { /* No slash */
24906 /* When it isn't a rational, take the input, convert it to a
24907 * NV, then create a canonical string representation of that
24911 SSize_t value_len = lookup_len - equals_pos;
24913 /* Get the value */
24914 if ( value_len <= 0
24915 || my_atof3(lookup_name + equals_pos, &value,
24917 != lookup_name + lookup_len)
24922 /* If the value is an integer, the canonical value is integral
24924 if (Perl_ceil(value) == value) {
24925 canonical = Perl_form(aTHX_ "%.*s%.0" NVff,
24926 equals_pos, lookup_name, value);
24928 else { /* Otherwise, it is %e with a known precision */
24931 canonical = Perl_form(aTHX_ "%.*s%.*" NVef,
24932 equals_pos, lookup_name,
24933 PL_E_FORMAT_PRECISION, value);
24935 /* The exponent generated is expecting two digits, whereas
24936 * %e on some systems will generate three. Remove leading
24937 * zeros in excess of 2 from the exponent. We start
24938 * looking for them after the '=' */
24939 exp_ptr = strchr(canonical + equals_pos, 'e');
24941 char * cur_ptr = exp_ptr + 2; /* past the 'e[+-]' */
24942 SSize_t excess_exponent_len = strlen(cur_ptr) - 2;
24944 assert(*(cur_ptr - 1) == '-' || *(cur_ptr - 1) == '+');
24946 if (excess_exponent_len > 0) {
24947 SSize_t leading_zeros = strspn(cur_ptr, "0");
24948 SSize_t excess_leading_zeros
24949 = MIN(leading_zeros, excess_exponent_len);
24950 if (excess_leading_zeros > 0) {
24951 Move(cur_ptr + excess_leading_zeros,
24953 strlen(cur_ptr) - excess_leading_zeros
24954 + 1, /* Copy the NUL as well */
24961 else { /* Has a slash. Create a rational in canonical form */
24962 UV numerator, denominator, gcd, trial;
24963 const char * end_ptr;
24964 const char * sign = "";
24966 /* We can't just find the numerator, denominator, and do the
24967 * division, then use the method above, because that is
24968 * inexact. And the input could be a rational that is within
24969 * epsilon (given our precision) of a valid rational, and would
24970 * then incorrectly compare valid.
24972 * We're only interested in the part after the '=' */
24973 const char * this_lookup_name = lookup_name + equals_pos;
24974 lookup_len -= equals_pos;
24975 slash_pos -= equals_pos;
24977 /* Handle any leading minus */
24978 if (this_lookup_name[0] == '-') {
24980 this_lookup_name++;
24985 /* Convert the numerator to numeric */
24986 end_ptr = this_lookup_name + slash_pos;
24987 if (! grok_atoUV(this_lookup_name, &numerator, &end_ptr)) {
24991 /* It better have included all characters before the slash */
24992 if (*end_ptr != '/') {
24996 /* Set to look at just the denominator */
24997 this_lookup_name += slash_pos;
24998 lookup_len -= slash_pos;
24999 end_ptr = this_lookup_name + lookup_len;
25001 /* Convert the denominator to numeric */
25002 if (! grok_atoUV(this_lookup_name, &denominator, &end_ptr)) {
25006 /* It better be the rest of the characters, and don't divide by
25008 if ( end_ptr != this_lookup_name + lookup_len
25009 || denominator == 0)
25014 /* Get the greatest common denominator using
25015 http://en.wikipedia.org/wiki/Euclidean_algorithm */
25017 trial = denominator;
25018 while (trial != 0) {
25020 trial = gcd % trial;
25024 /* If already in lowest possible terms, we have already tried
25025 * looking this up */
25030 /* Reduce the rational, which should put it in canonical form
25033 denominator /= gcd;
25035 canonical = Perl_form(aTHX_ "%.*s%s%" UVuf "/%" UVuf,
25036 equals_pos, lookup_name, sign, numerator, denominator);
25039 /* Here, we have the number in canonical form. Try that */
25040 table_index = do_uniprop_match(canonical, strlen(canonical));
25041 if (table_index == 0) {
25044 } /* End of still didn't find the property in our table */
25045 } /* End of didn't find the property in our table */
25047 /* Here, we have a non-zero return, which is an index into a table of ptrs.
25048 * A negative return signifies that the real index is the absolute value,
25049 * but the result needs to be inverted */
25050 if (table_index < 0) {
25051 invert_return = TRUE;
25052 table_index = -table_index;
25055 /* Out-of band indices indicate a deprecated property. The proper index is
25056 * modulo it with the table size. And dividing by the table size yields
25057 * an offset into a table constructed by regen/mk_invlists.pl to contain
25058 * the corresponding warning message */
25059 if (table_index > MAX_UNI_KEYWORD_INDEX) {
25060 Size_t warning_offset = table_index / MAX_UNI_KEYWORD_INDEX;
25061 table_index %= MAX_UNI_KEYWORD_INDEX;
25062 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED),
25063 "Use of '%.*s' in \\p{} or \\P{} is deprecated because: %s",
25064 (int) name_len, name,
25065 get_deprecated_property_msg(warning_offset));
25068 /* In a few properties, a different property is used under /i. These are
25069 * unlikely to change, so are hard-coded here. */
25071 if ( table_index == UNI_XPOSIXUPPER
25072 || table_index == UNI_XPOSIXLOWER
25073 || table_index == UNI_TITLE)
25075 table_index = UNI_CASED;
25077 else if ( table_index == UNI_UPPERCASELETTER
25078 || table_index == UNI_LOWERCASELETTER
25079 # ifdef UNI_TITLECASELETTER /* Missing from early Unicodes */
25080 || table_index == UNI_TITLECASELETTER
25083 table_index = UNI_CASEDLETTER;
25085 else if ( table_index == UNI_POSIXUPPER
25086 || table_index == UNI_POSIXLOWER)
25088 table_index = UNI_POSIXALPHA;
25092 /* Create and return the inversion list */
25093 prop_definition = get_prop_definition(table_index);
25094 sv_2mortal(prop_definition);
25096 /* See if there is a private use override to add to this definition */
25098 COPHH * hinthash = (IN_PERL_COMPILETIME)
25099 ? CopHINTHASH_get(&PL_compiling)
25100 : CopHINTHASH_get(PL_curcop);
25101 SV * pu_overrides = cophh_fetch_pv(hinthash, "private_use", 0, 0);
25103 if (UNLIKELY(pu_overrides && SvPOK(pu_overrides))) {
25105 /* See if there is an element in the hints hash for this table */
25106 SV * pu_lookup = Perl_newSVpvf(aTHX_ "%d=", table_index);
25107 const char * pos = strstr(SvPVX(pu_overrides), SvPVX(pu_lookup));
25111 SV * pu_definition;
25113 SV * expanded_prop_definition =
25114 sv_2mortal(invlist_clone(prop_definition, NULL));
25116 /* If so, it's definition is the string from here to the next
25117 * \a character. And its format is the same as a user-defined
25119 pos += SvCUR(pu_lookup);
25120 pu_definition = newSVpvn(pos, strchr(pos, '\a') - pos);
25121 pu_invlist = handle_user_defined_property(lookup_name,
25124 0, /* Not folded */
25132 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
25133 sv_catpvs(msg, "Insecure private-use override");
25134 goto append_name_to_msg;
25137 /* For now, as a safety measure, make sure that it doesn't
25138 * override non-private use code points */
25139 _invlist_intersection(pu_invlist, PL_Private_Use, &pu_invlist);
25141 /* Add it to the list to be returned */
25142 _invlist_union(prop_definition, pu_invlist,
25143 &expanded_prop_definition);
25144 prop_definition = expanded_prop_definition;
25145 Perl_ck_warner_d(aTHX_ packWARN(WARN_EXPERIMENTAL__PRIVATE_USE), "The private_use feature is experimental");
25150 if (invert_return) {
25151 _invlist_invert(prop_definition);
25153 return prop_definition;
25155 unknown_user_defined:
25156 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
25157 sv_catpvs(msg, "Unknown user-defined property name");
25158 goto append_name_to_msg;
25161 if (non_pkg_begin != 0) {
25162 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
25163 sv_catpvs(msg, "Illegal user-defined property name");
25166 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
25167 sv_catpvs(msg, "Can't find Unicode property definition");
25171 append_name_to_msg:
25173 const char * prefix = (runtime && level == 0) ? " \\p{" : " \"";
25174 const char * suffix = (runtime && level == 0) ? "}" : "\"";
25176 sv_catpv(msg, prefix);
25177 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8, name_len, name));
25178 sv_catpv(msg, suffix);
25183 definition_deferred:
25186 bool is_qualified = non_pkg_begin != 0; /* If has "::" */
25188 /* Here it could yet to be defined, so defer evaluation of this until
25189 * its needed at runtime. We need the fully qualified property name to
25190 * avoid ambiguity */
25192 fq_name = S_get_fq_name(aTHX_ name, name_len, is_utf8,
25196 /* If it didn't come with a package, or the package is utf8::, this
25197 * actually could be an official Unicode property whose inclusion we
25198 * are deferring until runtime to make sure that it isn't overridden by
25199 * a user-defined property of the same name (which we haven't
25200 * encountered yet). Add a marker to indicate this possibility, for
25201 * use at such time when we first need the definition during pattern
25202 * matching execution */
25203 if (! is_qualified || memBEGINPs(name, non_pkg_begin, "utf8::")) {
25204 sv_catpvs(fq_name, DEFERRED_COULD_BE_OFFICIAL_MARKERs);
25207 /* We also need a trailing newline */
25208 sv_catpvs(fq_name, "\n");
25210 *user_defined_ptr = TRUE;
25216 S_handle_names_wildcard(pTHX_ const char * wname, /* wildcard name to match */
25217 const STRLEN wname_len, /* Its length */
25218 SV ** prop_definition,
25221 /* Deal with Name property wildcard subpatterns; returns TRUE if there were
25222 * any matches, adding them to prop_definition */
25226 CV * get_names_info; /* entry to charnames.pm to get info we need */
25227 SV * names_string; /* Contains all character names, except algo */
25228 SV * algorithmic_names; /* Contains info about algorithmically
25229 generated character names */
25230 REGEXP * subpattern_re; /* The user's pattern to match with */
25231 struct regexp * prog; /* The compiled pattern */
25232 char * all_names_start; /* lib/unicore/Name.pl string of every
25233 (non-algorithmic) character name */
25234 char * cur_pos; /* We match, effectively using /gc; this is
25235 where we are now */
25236 bool found_matches = FALSE; /* Did any name match so far? */
25237 SV * empty; /* For matching zero length names */
25238 SV * must_sv; /* Contains the substring, if any, that must be
25239 in a name for the subpattern to match */
25240 const char * must; /* The PV of 'must' */
25241 STRLEN must_len; /* And its length */
25242 SV * syllable_name = NULL; /* For Hangul syllables */
25243 const char hangul_prefix[] = "HANGUL SYLLABLE ";
25244 const STRLEN hangul_prefix_len = sizeof(hangul_prefix) - 1;
25246 /* By inspection, there are a maximum of 7 bytes in the suffix of a hangul
25247 * syllable name, and these are immutable and guaranteed by the Unicode
25248 * standard to never be extended */
25249 const STRLEN syl_max_len = hangul_prefix_len + 7;
25253 PERL_ARGS_ASSERT_HANDLE_NAMES_WILDCARD;
25255 /* Make sure _charnames is loaded. (The parameters give context
25256 * for any errors generated */
25257 get_names_info = get_cv("_charnames::_get_names_info", 0);
25258 if (! get_names_info) {
25259 Perl_croak(aTHX_ "panic: Can't find '_charnames::_get_names_info");
25262 /* Get the charnames data */
25263 PUSHSTACKi(PERLSI_REGCOMP);
25271 /* Special _charnames entry point that returns the info this routine
25273 call_sv(MUTABLE_SV(get_names_info), G_LIST);
25277 /* Data structure for names which end in their very own code points */
25278 algorithmic_names = POPs;
25279 SvREFCNT_inc_simple_void_NN(algorithmic_names);
25281 /* The lib/unicore/Name.pl string */
25282 names_string = POPs;
25283 SvREFCNT_inc_simple_void_NN(names_string);
25290 if ( ! SvROK(names_string)
25291 || ! SvROK(algorithmic_names))
25292 { /* Perhaps should panic instead XXX */
25293 SvREFCNT_dec(names_string);
25294 SvREFCNT_dec(algorithmic_names);
25298 names_string = sv_2mortal(SvRV(names_string));
25299 all_names_start = SvPVX(names_string);
25300 cur_pos = all_names_start;
25302 algorithmic_names= sv_2mortal(SvRV(algorithmic_names));
25304 /* Compile the subpattern consisting of the name being looked for */
25305 subpattern_re = compile_wildcard(wname, wname_len, FALSE /* /-i */ );
25307 must_sv = re_intuit_string(subpattern_re);
25309 /* regexec.c can free the re_intuit_string() return. GH #17734 */
25310 must_sv = sv_2mortal(newSVsv(must_sv));
25311 must = SvPV(must_sv, must_len);
25318 /* (Note: 'must' could contain a NUL. And yet we use strspn() below on it.
25319 * This works because the NUL causes the function to return early, thus
25320 * showing that there are characters in it other than the acceptable ones,
25321 * which is our desired result.) */
25323 prog = ReANY(subpattern_re);
25325 /* If only nothing is matched, skip to where empty names are looked for */
25326 if (prog->maxlen == 0) {
25330 /* And match against the string of all names /gc. Don't even try if it
25331 * must match a character not found in any name. */
25332 if (strspn(must, "\n -0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ()") == must_len)
25334 while (execute_wildcard(subpattern_re,
25336 SvEND(names_string),
25337 all_names_start, 0,
25340 { /* Here, matched. */
25342 /* Note the string entries look like
25343 * 00001\nSTART OF HEADING\n\n
25344 * so we could match anywhere in that string. We have to rule out
25345 * matching a code point line */
25346 char * this_name_start = all_names_start
25347 + RX_OFFS(subpattern_re)->start;
25348 char * this_name_end = all_names_start
25349 + RX_OFFS(subpattern_re)->end;
25352 UV cp = 0; /* Silences some compilers */
25353 AV * this_string = NULL;
25354 bool is_multi = FALSE;
25356 /* If matched nothing, advance to next possible match */
25357 if (this_name_start == this_name_end) {
25358 cur_pos = (char *) memchr(this_name_end + 1, '\n',
25359 SvEND(names_string) - this_name_end);
25360 if (cur_pos == NULL) {
25365 /* Position the next match to start beyond the current returned
25367 cur_pos = (char *) memchr(this_name_end, '\n',
25368 SvEND(names_string) - this_name_end);
25371 /* Back up to the \n just before the beginning of the character. */
25372 cp_end = (char *) my_memrchr(all_names_start,
25374 this_name_start - all_names_start);
25376 /* If we didn't find a \n, it means it matched somewhere in the
25377 * initial '00000' in the string, so isn't a real match */
25378 if (cp_end == NULL) {
25382 this_name_start = cp_end + 1; /* The name starts just after */
25383 cp_end--; /* the \n, and the code point */
25384 /* ends just before it */
25386 /* All code points are 5 digits long */
25387 cp_start = cp_end - 4;
25389 /* This shouldn't happen, as we found a \n, and the first \n is
25390 * further along than what we subtracted */
25391 assert(cp_start >= all_names_start);
25393 if (cp_start == all_names_start) {
25394 *prop_definition = add_cp_to_invlist(*prop_definition, 0);
25398 /* If the character is a blank, we either have a named sequence, or
25399 * something is wrong */
25400 if (*(cp_start - 1) == ' ') {
25401 cp_start = (char *) my_memrchr(all_names_start,
25403 cp_start - all_names_start);
25407 assert(cp_start != NULL && cp_start >= all_names_start + 2);
25409 /* Except for the first line in the string, the sequence before the
25410 * code point is \n\n. If that isn't the case here, we didn't
25411 * match the name of a character. (We could have matched a named
25412 * sequence, not currently handled */
25413 if (*(cp_start - 1) != '\n' || *(cp_start - 2) != '\n') {
25417 /* We matched! Add this to the list */
25418 found_matches = TRUE;
25420 /* Loop through all the code points in the sequence */
25421 while (cp_start < cp_end) {
25423 /* Calculate this code point from its 5 digits */
25424 cp = (XDIGIT_VALUE(cp_start[0]) << 16)
25425 + (XDIGIT_VALUE(cp_start[1]) << 12)
25426 + (XDIGIT_VALUE(cp_start[2]) << 8)
25427 + (XDIGIT_VALUE(cp_start[3]) << 4)
25428 + XDIGIT_VALUE(cp_start[4]);
25430 cp_start += 6; /* Go past any blank */
25432 if (cp_start < cp_end || is_multi) {
25433 if (this_string == NULL) {
25434 this_string = newAV();
25438 av_push(this_string, newSVuv(cp));
25442 if (is_multi) { /* Was more than one code point */
25443 if (*strings == NULL) {
25444 *strings = newAV();
25447 av_push(*strings, (SV *) this_string);
25449 else { /* Only a single code point */
25450 *prop_definition = add_cp_to_invlist(*prop_definition, cp);
25452 } /* End of loop through the non-algorithmic names string */
25455 /* There are also character names not in 'names_string'. These are
25456 * algorithmically generatable. Try this pattern on each possible one.
25457 * (khw originally planned to leave this out given the large number of
25458 * matches attempted; but the speed turned out to be quite acceptable
25460 * There are plenty of opportunities to optimize to skip many of the tests.
25461 * beyond the rudimentary ones already here */
25463 /* First see if the subpattern matches any of the algorithmic generatable
25464 * Hangul syllable names.
25466 * We know none of these syllable names will match if the input pattern
25467 * requires more bytes than any syllable has, or if the input pattern only
25468 * matches an empty name, or if the pattern has something it must match and
25469 * one of the characters in that isn't in any Hangul syllable. */
25470 if ( prog->minlen <= (SSize_t) syl_max_len
25471 && prog->maxlen > 0
25472 && (strspn(must, "\n ABCDEGHIJKLMNOPRSTUWY") == must_len))
25474 /* These constants, names, values, and algorithm are adapted from the
25475 * Unicode standard, version 5.1, section 3.12, and should never
25477 const char * JamoL[] = {
25478 "G", "GG", "N", "D", "DD", "R", "M", "B", "BB",
25479 "S", "SS", "", "J", "JJ", "C", "K", "T", "P", "H"
25481 const int LCount = C_ARRAY_LENGTH(JamoL);
25483 const char * JamoV[] = {
25484 "A", "AE", "YA", "YAE", "EO", "E", "YEO", "YE", "O", "WA",
25485 "WAE", "OE", "YO", "U", "WEO", "WE", "WI", "YU", "EU", "YI",
25488 const int VCount = C_ARRAY_LENGTH(JamoV);
25490 const char * JamoT[] = {
25491 "", "G", "GG", "GS", "N", "NJ", "NH", "D", "L",
25492 "LG", "LM", "LB", "LS", "LT", "LP", "LH", "M", "B",
25493 "BS", "S", "SS", "NG", "J", "C", "K", "T", "P", "H"
25495 const int TCount = C_ARRAY_LENGTH(JamoT);
25499 /* This is the initial Hangul syllable code point; each time through the
25500 * inner loop, it maps to the next higher code point. For more info,
25501 * see the Hangul syllable section of the Unicode standard. */
25504 syllable_name = sv_2mortal(newSV(syl_max_len));
25505 sv_setpvn(syllable_name, hangul_prefix, hangul_prefix_len);
25507 for (L = 0; L < LCount; L++) {
25508 for (V = 0; V < VCount; V++) {
25509 for (T = 0; T < TCount; T++) {
25511 /* Truncate back to the prefix, which is unvarying */
25512 SvCUR_set(syllable_name, hangul_prefix_len);
25514 sv_catpv(syllable_name, JamoL[L]);
25515 sv_catpv(syllable_name, JamoV[V]);
25516 sv_catpv(syllable_name, JamoT[T]);
25518 if (execute_wildcard(subpattern_re,
25519 SvPVX(syllable_name),
25520 SvEND(syllable_name),
25521 SvPVX(syllable_name), 0,
25525 *prop_definition = add_cp_to_invlist(*prop_definition,
25527 found_matches = TRUE;
25536 /* The rest of the algorithmically generatable names are of the form
25537 * "PREFIX-code_point". The prefixes and the code point limits of each
25538 * were returned to us in the array 'algorithmic_names' from data in
25539 * lib/unicore/Name.pm. 'code_point' in the name is expressed in hex. */
25540 for (i = 0; i <= av_top_index((AV *) algorithmic_names); i++) {
25543 /* Each element of the array is a hash, giving the details for the
25544 * series of names it covers. There is the base name of the characters
25545 * in the series, and the low and high code points in the series. And,
25546 * for optimization purposes a string containing all the legal
25547 * characters that could possibly be in a name in this series. */
25548 HV * this_series = (HV *) SvRV(* av_fetch((AV *) algorithmic_names, i, 0));
25549 SV * prefix = * hv_fetchs(this_series, "name", 0);
25550 IV low = SvIV(* hv_fetchs(this_series, "low", 0));
25551 IV high = SvIV(* hv_fetchs(this_series, "high", 0));
25552 char * legal = SvPVX(* hv_fetchs(this_series, "legal", 0));
25554 /* Pre-allocate an SV with enough space */
25555 SV * algo_name = sv_2mortal(Perl_newSVpvf(aTHX_ "%s-0000",
25557 if (high >= 0x10000) {
25558 sv_catpvs(algo_name, "0");
25561 /* This series can be skipped entirely if the pattern requires
25562 * something longer than any name in the series, or can only match an
25563 * empty name, or contains a character not found in any name in the
25565 if ( prog->minlen <= (SSize_t) SvCUR(algo_name)
25566 && prog->maxlen > 0
25567 && (strspn(must, legal) == must_len))
25569 for (j = low; j <= high; j++) { /* For each code point in the series */
25571 /* Get its name, and see if it matches the subpattern */
25572 Perl_sv_setpvf(aTHX_ algo_name, "%s-%X", SvPVX(prefix),
25575 if (execute_wildcard(subpattern_re,
25578 SvPVX(algo_name), 0,
25582 *prop_definition = add_cp_to_invlist(*prop_definition, j);
25583 found_matches = TRUE;
25590 /* Finally, see if the subpattern matches an empty string */
25591 empty = newSVpvs("");
25592 if (execute_wildcard(subpattern_re,
25599 /* Many code points have empty names. Currently these are the \p{GC=C}
25600 * ones, minus CC and CF */
25602 SV * empty_names_ref = get_prop_definition(UNI_C);
25603 SV * empty_names = invlist_clone(empty_names_ref, NULL);
25605 SV * subtract = get_prop_definition(UNI_CC);
25607 _invlist_subtract(empty_names, subtract, &empty_names);
25608 SvREFCNT_dec_NN(empty_names_ref);
25609 SvREFCNT_dec_NN(subtract);
25611 subtract = get_prop_definition(UNI_CF);
25612 _invlist_subtract(empty_names, subtract, &empty_names);
25613 SvREFCNT_dec_NN(subtract);
25615 _invlist_union(*prop_definition, empty_names, prop_definition);
25616 found_matches = TRUE;
25617 SvREFCNT_dec_NN(empty_names);
25619 SvREFCNT_dec_NN(empty);
25622 /* If we ever were to accept aliases for, say private use names, we would
25623 * need to do something fancier to find empty names. The code below works
25624 * (at the time it was written), and is slower than the above */
25625 const char empties_pat[] = "^.";
25626 if (strNE(name, empties_pat)) {
25627 SV * empty = newSVpvs("");
25628 if (execute_wildcard(subpattern_re,
25635 SV * empties = NULL;
25637 (void) handle_names_wildcard(empties_pat, strlen(empties_pat), &empties);
25639 _invlist_union_complement_2nd(*prop_definition, empties, prop_definition);
25640 SvREFCNT_dec_NN(empties);
25642 found_matches = TRUE;
25644 SvREFCNT_dec_NN(empty);
25648 SvREFCNT_dec_NN(subpattern_re);
25649 return found_matches;
25653 * ex: set ts=8 sts=4 sw=4 et: