5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
10 /* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
12 * a regular expression.
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
20 /* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
24 /* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
29 /* The names of the functions have been changed from regcomp and
30 * regexec to pregcomp and pregexec in order to avoid conflicts
31 * with the POSIX routines of the same names.
34 #ifdef PERL_EXT_RE_BUILD
39 * pregcomp and pregexec -- regsub and regerror are not used in perl
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
59 **** Alterations to Henry's code are...
61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
74 /* Note on debug output:
76 * This is set up so that -Dr turns on debugging like all other flags that are
77 * enabled by -DDEBUGGING. -Drv gives more verbose output. This applies to
78 * all regular expressions encountered in a program, and gives a huge amount of
79 * output for all but the shortest programs.
81 * The ability to output pattern debugging information lexically, and with much
82 * finer grained control was added, with 'use re qw(Debug ....);' available even
83 * in non-DEBUGGING builds. This is accomplished by copying the contents of
84 * regcomp.c to ext/re/re_comp.c, and regexec.c is copied to ext/re/re_exec.c.
85 * Those files are compiled and linked into the perl executable, and they are
86 * compiled essentially as if DEBUGGING were enabled, and controlled by calls
89 * That would normally mean linking errors when two functions of the same name
90 * are attempted to be placed into the same executable. That is solved in one
92 * 1) Static functions aren't known outside the file they are in, so for the
93 * many functions of that type in this file, it just isn't a problem.
94 * 2) Most externally known functions are enclosed in
95 * #ifndef PERL_IN_XSUB_RE
98 * blocks, so there is only one defintion for them in the whole
99 * executable, the one in regcomp.c (or regexec.c). The implication of
100 * that is any debugging info that comes from them is controlled only by
101 * -Dr. Further, any static function they call will also be the version
102 * in regcomp.c (or regexec.c), so its debugging will also be by -Dr.
103 * 3) About a dozen external functions are re-#defined in ext/re/re_top.h, to
104 * have different names, so that what gets loaded in the executable is
105 * 'Perl_foo' from regcomp.c (and regexec.c), and the identical function
106 * from re_comp.c (and re_exec.c), but with the name 'my_foo' Debugging
107 * in the 'Perl_foo' versions is controlled by -Dr, but the 'my_foo'
108 * versions and their callees are under control of re.pm. The catch is
109 * that references to all these go through the regexp_engine structure,
110 * which is initialized in regcomp.h to the Perl_foo versions, and
111 * substituted out in lexical scopes where 'use re' is in effect to the
112 * 'my_foo' ones. That structure is public API, so it would be a hard
113 * sell to add any additional members.
114 * 4) For functions in regcomp.c and re_comp.c that are called only from,
115 * respectively, regexec.c and re_exec.c, they can have two different
116 * names, depending on #ifdef'ing PERL_IN_XSUB_RE, in both regexec.c and
119 * The bottom line is that if you add code to one of the public functions
120 * listed in ext/re/re_top.h, debugging automagically works. But if you write
121 * a new function that needs to do debugging or there is a chain of calls from
122 * it that need to do debugging, all functions in the chain should use options
125 * A function may have to be split so that debugging stuff is static, but it
126 * calls out to some other function that only gets compiled in regcomp.c to
127 * access data that we don't want to duplicate.
131 #define PERL_IN_REGCOMP_C
135 #ifdef PERL_IN_XSUB_RE
136 # include "re_comp.h"
137 EXTERN_C const struct regexp_engine my_reg_engine;
138 EXTERN_C const struct regexp_engine wild_reg_engine;
140 # include "regcomp.h"
143 #include "invlist_inline.h"
144 #include "unicode_constants.h"
147 #define STATIC static
150 /* this is a chain of data about sub patterns we are processing that
151 need to be handled separately/specially in study_chunk. Its so
152 we can simulate recursion without losing state. */
154 typedef struct scan_frame {
155 regnode *last_regnode; /* last node to process in this frame */
156 regnode *next_regnode; /* next node to process when last is reached */
157 U32 prev_recursed_depth;
158 I32 stopparen; /* what stopparen do we use */
159 bool in_gosub; /* this or an outer frame is for GOSUB */
161 struct scan_frame *this_prev_frame; /* this previous frame */
162 struct scan_frame *prev_frame; /* previous frame */
163 struct scan_frame *next_frame; /* next frame */
166 /* Certain characters are output as a sequence with the first being a
168 #define isBACKSLASHED_PUNCT(c) memCHRs("-[]\\^", c)
171 struct RExC_state_t {
172 U32 flags; /* RXf_* are we folding, multilining? */
173 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
174 char *precomp; /* uncompiled string. */
175 char *precomp_end; /* pointer to end of uncompiled string. */
176 REGEXP *rx_sv; /* The SV that is the regexp. */
177 regexp *rx; /* perl core regexp structure */
178 regexp_internal *rxi; /* internal data for regexp object
180 char *start; /* Start of input for compile */
181 char *end; /* End of input for compile */
182 char *parse; /* Input-scan pointer. */
183 char *copy_start; /* start of copy of input within
184 constructed parse string */
185 char *save_copy_start; /* Provides one level of saving
186 and restoring 'copy_start' */
187 char *copy_start_in_input; /* Position in input string
188 corresponding to copy_start */
189 SSize_t whilem_seen; /* number of WHILEM in this expr */
190 regnode *emit_start; /* Start of emitted-code area */
191 regnode_offset emit; /* Code-emit pointer */
192 I32 naughty; /* How bad is this pattern? */
193 I32 sawback; /* Did we see \1, ...? */
194 SSize_t size; /* Number of regnode equivalents in
196 Size_t sets_depth; /* Counts recursion depth of already-
197 compiled regex set patterns */
200 I32 parens_buf_size; /* #slots malloced open/close_parens */
201 regnode_offset *open_parens; /* offsets to open parens */
202 regnode_offset *close_parens; /* offsets to close parens */
203 HV *paren_names; /* Paren names */
205 /* position beyond 'precomp' of the warning message furthest away from
206 * 'precomp'. During the parse, no warnings are raised for any problems
207 * earlier in the parse than this position. This works if warnings are
208 * raised the first time a given spot is parsed, and if only one
209 * independent warning is raised for any given spot */
210 Size_t latest_warn_offset;
212 I32 npar; /* Capture buffer count so far in the
213 parse, (OPEN) plus one. ("par" 0 is
215 I32 total_par; /* During initial parse, is either 0,
216 or -1; the latter indicating a
217 reparse is needed. After that pass,
218 it is what 'npar' became after the
219 pass. Hence, it being > 0 indicates
220 we are in a reparse situation */
221 I32 nestroot; /* root parens we are in - used by
224 regnode *end_op; /* END node in program */
225 I32 utf8; /* whether the pattern is utf8 or not */
226 I32 orig_utf8; /* whether the pattern was originally in utf8 */
227 /* XXX use this for future optimisation of case
228 * where pattern must be upgraded to utf8. */
229 I32 uni_semantics; /* If a d charset modifier should use unicode
230 rules, even if the pattern is not in
233 I32 recurse_count; /* Number of recurse regops we have generated */
234 regnode **recurse; /* Recurse regops */
235 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
237 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
240 I32 override_recoding;
241 I32 recode_x_to_native;
242 I32 in_multi_char_class;
243 int code_index; /* next code_blocks[] slot */
244 struct reg_code_blocks *code_blocks;/* positions of literal (?{})
246 SSize_t maxlen; /* mininum possible number of chars in string to match */
247 scan_frame *frame_head;
248 scan_frame *frame_last;
252 SV *runtime_code_qr; /* qr with the runtime code blocks */
254 const char *lastparse;
256 U32 study_chunk_recursed_count;
257 AV *paren_name_list; /* idx -> name */
261 #define RExC_lastparse (pRExC_state->lastparse)
262 #define RExC_lastnum (pRExC_state->lastnum)
263 #define RExC_paren_name_list (pRExC_state->paren_name_list)
264 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
265 #define RExC_mysv (pRExC_state->mysv1)
266 #define RExC_mysv1 (pRExC_state->mysv1)
267 #define RExC_mysv2 (pRExC_state->mysv2)
275 bool sWARN_EXPERIMENTAL__VLB;
276 bool sWARN_EXPERIMENTAL__REGEX_SETS;
279 #define RExC_flags (pRExC_state->flags)
280 #define RExC_pm_flags (pRExC_state->pm_flags)
281 #define RExC_precomp (pRExC_state->precomp)
282 #define RExC_copy_start_in_input (pRExC_state->copy_start_in_input)
283 #define RExC_copy_start_in_constructed (pRExC_state->copy_start)
284 #define RExC_save_copy_start_in_constructed (pRExC_state->save_copy_start)
285 #define RExC_precomp_end (pRExC_state->precomp_end)
286 #define RExC_rx_sv (pRExC_state->rx_sv)
287 #define RExC_rx (pRExC_state->rx)
288 #define RExC_rxi (pRExC_state->rxi)
289 #define RExC_start (pRExC_state->start)
290 #define RExC_end (pRExC_state->end)
291 #define RExC_parse (pRExC_state->parse)
292 #define RExC_latest_warn_offset (pRExC_state->latest_warn_offset )
293 #define RExC_whilem_seen (pRExC_state->whilem_seen)
294 #define RExC_seen_d_op (pRExC_state->seen_d_op) /* Seen something that differs
295 under /d from /u ? */
297 #ifdef RE_TRACK_PATTERN_OFFSETS
298 # define RExC_offsets (RExC_rxi->u.offsets) /* I am not like the
301 #define RExC_emit (pRExC_state->emit)
302 #define RExC_emit_start (pRExC_state->emit_start)
303 #define RExC_sawback (pRExC_state->sawback)
304 #define RExC_seen (pRExC_state->seen)
305 #define RExC_size (pRExC_state->size)
306 #define RExC_maxlen (pRExC_state->maxlen)
307 #define RExC_npar (pRExC_state->npar)
308 #define RExC_total_parens (pRExC_state->total_par)
309 #define RExC_parens_buf_size (pRExC_state->parens_buf_size)
310 #define RExC_nestroot (pRExC_state->nestroot)
311 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
312 #define RExC_utf8 (pRExC_state->utf8)
313 #define RExC_uni_semantics (pRExC_state->uni_semantics)
314 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
315 #define RExC_open_parens (pRExC_state->open_parens)
316 #define RExC_close_parens (pRExC_state->close_parens)
317 #define RExC_end_op (pRExC_state->end_op)
318 #define RExC_paren_names (pRExC_state->paren_names)
319 #define RExC_recurse (pRExC_state->recurse)
320 #define RExC_recurse_count (pRExC_state->recurse_count)
321 #define RExC_sets_depth (pRExC_state->sets_depth)
322 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
323 #define RExC_study_chunk_recursed_bytes \
324 (pRExC_state->study_chunk_recursed_bytes)
325 #define RExC_in_lookaround (pRExC_state->in_lookaround)
326 #define RExC_contains_locale (pRExC_state->contains_locale)
327 #define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
330 # define SET_recode_x_to_native(x) \
331 STMT_START { RExC_recode_x_to_native = (x); } STMT_END
333 # define SET_recode_x_to_native(x) NOOP
336 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
337 #define RExC_frame_head (pRExC_state->frame_head)
338 #define RExC_frame_last (pRExC_state->frame_last)
339 #define RExC_frame_count (pRExC_state->frame_count)
340 #define RExC_strict (pRExC_state->strict)
341 #define RExC_study_started (pRExC_state->study_started)
342 #define RExC_warn_text (pRExC_state->warn_text)
343 #define RExC_in_script_run (pRExC_state->in_script_run)
344 #define RExC_use_BRANCHJ (pRExC_state->use_BRANCHJ)
345 #define RExC_warned_WARN_EXPERIMENTAL__VLB (pRExC_state->sWARN_EXPERIMENTAL__VLB)
346 #define RExC_warned_WARN_EXPERIMENTAL__REGEX_SETS (pRExC_state->sWARN_EXPERIMENTAL__REGEX_SETS)
347 #define RExC_unlexed_names (pRExC_state->unlexed_names)
349 /* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
350 * a flag to disable back-off on the fixed/floating substrings - if it's
351 * a high complexity pattern we assume the benefit of avoiding a full match
352 * is worth the cost of checking for the substrings even if they rarely help.
354 #define RExC_naughty (pRExC_state->naughty)
355 #define TOO_NAUGHTY (10)
356 #define MARK_NAUGHTY(add) \
357 if (RExC_naughty < TOO_NAUGHTY) \
358 RExC_naughty += (add)
359 #define MARK_NAUGHTY_EXP(exp, add) \
360 if (RExC_naughty < TOO_NAUGHTY) \
361 RExC_naughty += RExC_naughty / (exp) + (add)
363 #define isNON_BRACE_QUANTIFIER(c) ((c) == '*' || (c) == '+' || (c) == '?')
364 #define isQUANTIFIER(s,e) ( isNON_BRACE_QUANTIFIER(*s) \
365 || ((*s) == '{' && regcurly(s, e, NULL)))
368 * Flags to be passed up and down.
370 #define HASWIDTH 0x01 /* Known to not match null strings, could match
372 #define SIMPLE 0x02 /* Exactly one character wide */
373 /* (or LNBREAK as a special case) */
374 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
375 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
376 #define RESTART_PARSE 0x20 /* Need to redo the parse */
377 #define NEED_UTF8 0x40 /* In conjunction with RESTART_PARSE, need to
378 calcuate sizes as UTF-8 */
380 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
382 /* whether trie related optimizations are enabled */
383 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
384 #define TRIE_STUDY_OPT
385 #define FULL_TRIE_STUDY
391 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
392 #define PBITVAL(paren) (1 << ((paren) & 7))
393 #define PAREN_OFFSET(depth) \
394 (RExC_study_chunk_recursed + (depth) * RExC_study_chunk_recursed_bytes)
395 #define PAREN_TEST(depth, paren) \
396 (PBYTE(PAREN_OFFSET(depth), paren) & PBITVAL(paren))
397 #define PAREN_SET(depth, paren) \
398 (PBYTE(PAREN_OFFSET(depth), paren) |= PBITVAL(paren))
399 #define PAREN_UNSET(depth, paren) \
400 (PBYTE(PAREN_OFFSET(depth), paren) &= ~PBITVAL(paren))
402 #define REQUIRE_UTF8(flagp) STMT_START { \
404 *flagp = RESTART_PARSE|NEED_UTF8; \
409 /* /u is to be chosen if we are supposed to use Unicode rules, or if the
410 * pattern is in UTF-8. This latter condition is in case the outermost rules
411 * are locale. See GH #17278 */
412 #define toUSE_UNI_CHARSET_NOT_DEPENDS (RExC_uni_semantics || UTF)
414 /* Change from /d into /u rules, and restart the parse. RExC_uni_semantics is
415 * a flag that indicates we need to override /d with /u as a result of
416 * something in the pattern. It should only be used in regards to calling
417 * set_regex_charset() or get_regex_charset() */
418 #define REQUIRE_UNI_RULES(flagp, restart_retval) \
420 if (DEPENDS_SEMANTICS) { \
421 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
422 RExC_uni_semantics = 1; \
423 if (RExC_seen_d_op && LIKELY(! IN_PARENS_PASS)) { \
424 /* No need to restart the parse if we haven't seen \
425 * anything that differs between /u and /d, and no need \
426 * to restart immediately if we're going to reparse \
427 * anyway to count parens */ \
428 *flagp |= RESTART_PARSE; \
429 return restart_retval; \
434 #define REQUIRE_BRANCHJ(flagp, restart_retval) \
436 RExC_use_BRANCHJ = 1; \
437 *flagp |= RESTART_PARSE; \
438 return restart_retval; \
441 /* Until we have completed the parse, we leave RExC_total_parens at 0 or
442 * less. After that, it must always be positive, because the whole re is
443 * considered to be surrounded by virtual parens. Setting it to negative
444 * indicates there is some construct that needs to know the actual number of
445 * parens to be properly handled. And that means an extra pass will be
446 * required after we've counted them all */
447 #define ALL_PARENS_COUNTED (RExC_total_parens > 0)
448 #define REQUIRE_PARENS_PASS \
449 STMT_START { /* No-op if have completed a pass */ \
450 if (! ALL_PARENS_COUNTED) RExC_total_parens = -1; \
452 #define IN_PARENS_PASS (RExC_total_parens < 0)
455 /* This is used to return failure (zero) early from the calling function if
456 * various flags in 'flags' are set. Two flags always cause a return:
457 * 'RESTART_PARSE' and 'NEED_UTF8'. 'extra' can be used to specify any
458 * additional flags that should cause a return; 0 if none. If the return will
459 * be done, '*flagp' is first set to be all of the flags that caused the
461 #define RETURN_FAIL_ON_RESTART_OR_FLAGS(flags,flagp,extra) \
463 if ((flags) & (RESTART_PARSE|NEED_UTF8|(extra))) { \
464 *(flagp) = (flags) & (RESTART_PARSE|NEED_UTF8|(extra)); \
469 #define MUST_RESTART(flags) ((flags) & (RESTART_PARSE))
471 #define RETURN_FAIL_ON_RESTART(flags,flagp) \
472 RETURN_FAIL_ON_RESTART_OR_FLAGS( flags, flagp, 0)
473 #define RETURN_FAIL_ON_RESTART_FLAGP(flagp) \
474 if (MUST_RESTART(*(flagp))) return 0
476 /* This converts the named class defined in regcomp.h to its equivalent class
477 * number defined in handy.h. */
478 #define namedclass_to_classnum(class) ((int) ((class) / 2))
479 #define classnum_to_namedclass(classnum) ((classnum) * 2)
481 #define _invlist_union_complement_2nd(a, b, output) \
482 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
483 #define _invlist_intersection_complement_2nd(a, b, output) \
484 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
486 /* We add a marker if we are deferring expansion of a property that is both
487 * 1) potentiallly user-defined; and
488 * 2) could also be an official Unicode property.
490 * Without this marker, any deferred expansion can only be for a user-defined
491 * one. This marker shouldn't conflict with any that could be in a legal name,
492 * and is appended to its name to indicate this. There is a string and
494 #define DEFERRED_COULD_BE_OFFICIAL_MARKERs "~"
495 #define DEFERRED_COULD_BE_OFFICIAL_MARKERc '~'
497 /* What is infinity for optimization purposes */
498 #define OPTIMIZE_INFTY SSize_t_MAX
500 /* About scan_data_t.
502 During optimisation we recurse through the regexp program performing
503 various inplace (keyhole style) optimisations. In addition study_chunk
504 and scan_commit populate this data structure with information about
505 what strings MUST appear in the pattern. We look for the longest
506 string that must appear at a fixed location, and we look for the
507 longest string that may appear at a floating location. So for instance
512 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
513 strings (because they follow a .* construct). study_chunk will identify
514 both FOO and BAR as being the longest fixed and floating strings respectively.
516 The strings can be composites, for instance
520 will result in a composite fixed substring 'foo'.
522 For each string some basic information is maintained:
525 This is the position the string must appear at, or not before.
526 It also implicitly (when combined with minlenp) tells us how many
527 characters must match before the string we are searching for.
528 Likewise when combined with minlenp and the length of the string it
529 tells us how many characters must appear after the string we have
533 Only used for floating strings. This is the rightmost point that
534 the string can appear at. If set to OPTIMIZE_INFTY it indicates that the
535 string can occur infinitely far to the right.
536 For fixed strings, it is equal to min_offset.
539 A pointer to the minimum number of characters of the pattern that the
540 string was found inside. This is important as in the case of positive
541 lookahead or positive lookbehind we can have multiple patterns
546 The minimum length of the pattern overall is 3, the minimum length
547 of the lookahead part is 3, but the minimum length of the part that
548 will actually match is 1. So 'FOO's minimum length is 3, but the
549 minimum length for the F is 1. This is important as the minimum length
550 is used to determine offsets in front of and behind the string being
551 looked for. Since strings can be composites this is the length of the
552 pattern at the time it was committed with a scan_commit. Note that
553 the length is calculated by study_chunk, so that the minimum lengths
554 are not known until the full pattern has been compiled, thus the
555 pointer to the value.
559 In the case of lookbehind the string being searched for can be
560 offset past the start point of the final matching string.
561 If this value was just blithely removed from the min_offset it would
562 invalidate some of the calculations for how many chars must match
563 before or after (as they are derived from min_offset and minlen and
564 the length of the string being searched for).
565 When the final pattern is compiled and the data is moved from the
566 scan_data_t structure into the regexp structure the information
567 about lookbehind is factored in, with the information that would
568 have been lost precalculated in the end_shift field for the
571 The fields pos_min and pos_delta are used to store the minimum offset
572 and the delta to the maximum offset at the current point in the pattern.
576 struct scan_data_substrs {
577 SV *str; /* longest substring found in pattern */
578 SSize_t min_offset; /* earliest point in string it can appear */
579 SSize_t max_offset; /* latest point in string it can appear */
580 SSize_t *minlenp; /* pointer to the minlen relevant to the string */
581 SSize_t lookbehind; /* is the pos of the string modified by LB */
582 I32 flags; /* per substring SF_* and SCF_* flags */
585 typedef struct scan_data_t {
586 /*I32 len_min; unused */
587 /*I32 len_delta; unused */
591 SSize_t last_end; /* min value, <0 unless valid. */
592 SSize_t last_start_min;
593 SSize_t last_start_max;
594 U8 cur_is_floating; /* whether the last_* values should be set as
595 * the next fixed (0) or floating (1)
598 /* [0] is longest fixed substring so far, [1] is longest float so far */
599 struct scan_data_substrs substrs[2];
601 I32 flags; /* common SF_* and SCF_* flags */
603 SSize_t *last_closep;
604 regnode_ssc *start_class;
608 * Forward declarations for pregcomp()'s friends.
611 static const scan_data_t zero_scan_data = {
612 0, 0, NULL, 0, 0, 0, 0,
614 { NULL, 0, 0, 0, 0, 0 },
615 { NULL, 0, 0, 0, 0, 0 },
622 #define SF_BEFORE_SEOL 0x0001
623 #define SF_BEFORE_MEOL 0x0002
624 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
626 #define SF_IS_INF 0x0040
627 #define SF_HAS_PAR 0x0080
628 #define SF_IN_PAR 0x0100
629 #define SF_HAS_EVAL 0x0200
632 /* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
633 * longest substring in the pattern. When it is not set the optimiser keeps
634 * track of position, but does not keep track of the actual strings seen,
636 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
639 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
640 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
641 * turned off because of the alternation (BRANCH). */
642 #define SCF_DO_SUBSTR 0x0400
644 #define SCF_DO_STCLASS_AND 0x0800
645 #define SCF_DO_STCLASS_OR 0x1000
646 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
647 #define SCF_WHILEM_VISITED_POS 0x2000
649 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
650 #define SCF_SEEN_ACCEPT 0x8000
651 #define SCF_TRIE_DOING_RESTUDY 0x10000
652 #define SCF_IN_DEFINE 0x20000
657 #define UTF cBOOL(RExC_utf8)
659 /* The enums for all these are ordered so things work out correctly */
660 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
661 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
662 == REGEX_DEPENDS_CHARSET)
663 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
664 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
665 >= REGEX_UNICODE_CHARSET)
666 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
667 == REGEX_ASCII_RESTRICTED_CHARSET)
668 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
669 >= REGEX_ASCII_RESTRICTED_CHARSET)
670 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
671 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
673 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
675 /* For programs that want to be strictly Unicode compatible by dying if any
676 * attempt is made to match a non-Unicode code point against a Unicode
678 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
680 #define OOB_NAMEDCLASS -1
682 /* There is no code point that is out-of-bounds, so this is problematic. But
683 * its only current use is to initialize a variable that is always set before
685 #define OOB_UNICODE 0xDEADBEEF
687 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
690 /* length of regex to show in messages that don't mark a position within */
691 #define RegexLengthToShowInErrorMessages 127
694 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
695 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
696 * op/pragma/warn/regcomp.
698 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
699 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
701 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
702 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
704 /* The code in this file in places uses one level of recursion with parsing
705 * rebased to an alternate string constructed by us in memory. This can take
706 * the form of something that is completely different from the input, or
707 * something that uses the input as part of the alternate. In the first case,
708 * there should be no possibility of an error, as we are in complete control of
709 * the alternate string. But in the second case we don't completely control
710 * the input portion, so there may be errors in that. Here's an example:
712 * is handled specially because \x{df} folds to a sequence of more than one
713 * character: 'ss'. What is done is to create and parse an alternate string,
714 * which looks like this:
715 * /(?:\x{DF}|[abc\x{DF}def])/ui
716 * where it uses the input unchanged in the middle of something it constructs,
717 * which is a branch for the DF outside the character class, and clustering
718 * parens around the whole thing. (It knows enough to skip the DF inside the
719 * class while in this substitute parse.) 'abc' and 'def' may have errors that
720 * need to be reported. The general situation looks like this:
722 * |<------- identical ------>|
724 * Input: ---------------------------------------------------------------
725 * Constructed: ---------------------------------------------------
727 * |<------- identical ------>|
729 * sI..eI is the portion of the input pattern we are concerned with here.
730 * sC..EC is the constructed substitute parse string.
731 * sC..tC is constructed by us
732 * tC..eC is an exact duplicate of the portion of the input pattern tI..eI.
733 * In the diagram, these are vertically aligned.
734 * eC..EC is also constructed by us.
735 * xC is the position in the substitute parse string where we found a
737 * xI is the position in the original pattern corresponding to xC.
739 * We want to display a message showing the real input string. Thus we need to
740 * translate from xC to xI. We know that xC >= tC, since the portion of the
741 * string sC..tC has been constructed by us, and so shouldn't have errors. We
743 * xI = tI + (xC - tC)
745 * When the substitute parse is constructed, the code needs to set:
748 * RExC_copy_start_in_input (tI)
749 * RExC_copy_start_in_constructed (tC)
750 * and restore them when done.
752 * During normal processing of the input pattern, both
753 * 'RExC_copy_start_in_input' and 'RExC_copy_start_in_constructed' are set to
754 * sI, so that xC equals xI.
757 #define sI RExC_precomp
758 #define eI RExC_precomp_end
759 #define sC RExC_start
761 #define tI RExC_copy_start_in_input
762 #define tC RExC_copy_start_in_constructed
763 #define xI(xC) (tI + (xC - tC))
764 #define xI_offset(xC) (xI(xC) - sI)
766 #define REPORT_LOCATION_ARGS(xC) \
768 (xI(xC) > eI) /* Don't run off end */ \
769 ? eI - sI /* Length before the <--HERE */ \
770 : ((xI_offset(xC) >= 0) \
772 : (Perl_croak(aTHX_ "panic: %s: %d: negative offset: %" \
773 IVdf " trying to output message for " \
775 __FILE__, __LINE__, (IV) xI_offset(xC), \
776 ((int) (eC - sC)), sC), 0)), \
777 sI), /* The input pattern printed up to the <--HERE */ \
779 (xI(xC) > eI) ? 0 : eI - xI(xC), /* Length after <--HERE */ \
780 (xI(xC) > eI) ? eI : xI(xC)) /* pattern after <--HERE */
782 /* Used to point after bad bytes for an error message, but avoid skipping
783 * past a nul byte. */
784 #define SKIP_IF_CHAR(s, e) (!*(s) ? 0 : UTF ? UTF8_SAFE_SKIP(s, e) : 1)
786 /* Set up to clean up after our imminent demise */
787 #define PREPARE_TO_DIE \
790 SAVEFREESV(RExC_rx_sv); \
791 if (RExC_open_parens) \
792 SAVEFREEPV(RExC_open_parens); \
793 if (RExC_close_parens) \
794 SAVEFREEPV(RExC_close_parens); \
798 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
799 * arg. Show regex, up to a maximum length. If it's too long, chop and add
802 #define _FAIL(code) STMT_START { \
803 const char *ellipses = ""; \
804 IV len = RExC_precomp_end - RExC_precomp; \
807 if (len > RegexLengthToShowInErrorMessages) { \
808 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
809 len = RegexLengthToShowInErrorMessages - 10; \
815 #define FAIL(msg) _FAIL( \
816 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
817 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
819 #define FAIL2(msg,arg) _FAIL( \
820 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
821 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
823 #define FAIL3(msg,arg1,arg2) _FAIL( \
824 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
825 arg1, arg2, UTF8fARG(UTF, len, RExC_precomp), ellipses))
828 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
830 #define Simple_vFAIL(m) STMT_START { \
831 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
832 m, REPORT_LOCATION_ARGS(RExC_parse)); \
836 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
838 #define vFAIL(m) STMT_START { \
844 * Like Simple_vFAIL(), but accepts two arguments.
846 #define Simple_vFAIL2(m,a1) STMT_START { \
847 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, \
848 REPORT_LOCATION_ARGS(RExC_parse)); \
852 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
854 #define vFAIL2(m,a1) STMT_START { \
856 Simple_vFAIL2(m, a1); \
861 * Like Simple_vFAIL(), but accepts three arguments.
863 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
864 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, a2, \
865 REPORT_LOCATION_ARGS(RExC_parse)); \
869 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
871 #define vFAIL3(m,a1,a2) STMT_START { \
873 Simple_vFAIL3(m, a1, a2); \
877 * Like Simple_vFAIL(), but accepts four arguments.
879 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
880 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, a2, a3, \
881 REPORT_LOCATION_ARGS(RExC_parse)); \
884 #define vFAIL4(m,a1,a2,a3) STMT_START { \
886 Simple_vFAIL4(m, a1, a2, a3); \
889 /* A specialized version of vFAIL2 that works with UTF8f */
890 #define vFAIL2utf8f(m, a1) STMT_START { \
892 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, \
893 REPORT_LOCATION_ARGS(RExC_parse)); \
896 #define vFAIL3utf8f(m, a1, a2) STMT_START { \
898 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, a2, \
899 REPORT_LOCATION_ARGS(RExC_parse)); \
902 /* Setting this to NULL is a signal to not output warnings */
903 #define TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE \
905 RExC_save_copy_start_in_constructed = RExC_copy_start_in_constructed;\
906 RExC_copy_start_in_constructed = NULL; \
908 #define RESTORE_WARNINGS \
909 RExC_copy_start_in_constructed = RExC_save_copy_start_in_constructed
911 /* Since a warning can be generated multiple times as the input is reparsed, we
912 * output it the first time we come to that point in the parse, but suppress it
913 * otherwise. 'RExC_copy_start_in_constructed' being NULL is a flag to not
914 * generate any warnings */
915 #define TO_OUTPUT_WARNINGS(loc) \
916 ( RExC_copy_start_in_constructed \
917 && ((xI(loc)) - RExC_precomp) > (Ptrdiff_t) RExC_latest_warn_offset)
919 /* After we've emitted a warning, we save the position in the input so we don't
921 #define UPDATE_WARNINGS_LOC(loc) \
923 if (TO_OUTPUT_WARNINGS(loc)) { \
924 RExC_latest_warn_offset = MAX(sI, MIN(eI, xI(loc))) \
929 /* 'warns' is the output of the packWARNx macro used in 'code' */
930 #define _WARN_HELPER(loc, warns, code) \
932 if (! RExC_copy_start_in_constructed) { \
933 Perl_croak( aTHX_ "panic! %s: %d: Tried to warn when none" \
934 " expected at '%s'", \
935 __FILE__, __LINE__, loc); \
937 if (TO_OUTPUT_WARNINGS(loc)) { \
941 UPDATE_WARNINGS_LOC(loc); \
945 /* m is not necessarily a "literal string", in this macro */
946 #define warn_non_literal_string(loc, packed_warn, m) \
947 _WARN_HELPER(loc, packed_warn, \
948 Perl_warner(aTHX_ packed_warn, \
949 "%s" REPORT_LOCATION, \
950 m, REPORT_LOCATION_ARGS(loc)))
951 #define reg_warn_non_literal_string(loc, m) \
952 warn_non_literal_string(loc, packWARN(WARN_REGEXP), m)
954 #define ckWARN2_non_literal_string(loc, packwarn, m, a1) \
957 Size_t format_size = strlen(m) + strlen(REPORT_LOCATION)+ 1;\
958 Newx(format, format_size, char); \
959 my_strlcpy(format, m, format_size); \
960 my_strlcat(format, REPORT_LOCATION, format_size); \
961 SAVEFREEPV(format); \
962 _WARN_HELPER(loc, packwarn, \
963 Perl_ck_warner(aTHX_ packwarn, \
965 a1, REPORT_LOCATION_ARGS(loc))); \
968 #define ckWARNreg(loc,m) \
969 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
970 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
972 REPORT_LOCATION_ARGS(loc)))
974 #define vWARN(loc, m) \
975 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
976 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
978 REPORT_LOCATION_ARGS(loc))) \
980 #define vWARN_dep(loc, m) \
981 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
982 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
984 REPORT_LOCATION_ARGS(loc)))
986 #define ckWARNdep(loc,m) \
987 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
988 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
990 REPORT_LOCATION_ARGS(loc)))
992 #define ckWARNregdep(loc,m) \
993 _WARN_HELPER(loc, packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
994 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
997 REPORT_LOCATION_ARGS(loc)))
999 #define ckWARN2reg_d(loc,m, a1) \
1000 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1001 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
1002 m REPORT_LOCATION, \
1003 a1, REPORT_LOCATION_ARGS(loc)))
1005 #define ckWARN2reg(loc, m, a1) \
1006 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1007 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
1008 m REPORT_LOCATION, \
1009 a1, REPORT_LOCATION_ARGS(loc)))
1011 #define vWARN3(loc, m, a1, a2) \
1012 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1013 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
1014 m REPORT_LOCATION, \
1015 a1, a2, REPORT_LOCATION_ARGS(loc)))
1017 #define ckWARN3reg(loc, m, a1, a2) \
1018 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1019 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
1020 m REPORT_LOCATION, \
1022 REPORT_LOCATION_ARGS(loc)))
1024 #define vWARN4(loc, m, a1, a2, a3) \
1025 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1026 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
1027 m REPORT_LOCATION, \
1029 REPORT_LOCATION_ARGS(loc)))
1031 #define ckWARN4reg(loc, m, a1, a2, a3) \
1032 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1033 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
1034 m REPORT_LOCATION, \
1036 REPORT_LOCATION_ARGS(loc)))
1038 #define vWARN5(loc, m, a1, a2, a3, a4) \
1039 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1040 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
1041 m REPORT_LOCATION, \
1043 REPORT_LOCATION_ARGS(loc)))
1045 #define ckWARNexperimental(loc, class, m) \
1047 if (! RExC_warned_ ## class) { /* warn once per compilation */ \
1048 RExC_warned_ ## class = 1; \
1049 _WARN_HELPER(loc, packWARN(class), \
1050 Perl_ck_warner_d(aTHX_ packWARN(class), \
1051 m REPORT_LOCATION, \
1052 REPORT_LOCATION_ARGS(loc)));\
1056 /* Convert between a pointer to a node and its offset from the beginning of the
1058 #define REGNODE_p(offset) (RExC_emit_start + (offset))
1059 #define REGNODE_OFFSET(node) ((node) - RExC_emit_start)
1061 /* Macros for recording node offsets. 20001227 mjd@plover.com
1062 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
1063 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
1064 * Element 0 holds the number n.
1065 * Position is 1 indexed.
1067 #ifndef RE_TRACK_PATTERN_OFFSETS
1068 #define Set_Node_Offset_To_R(offset,byte)
1069 #define Set_Node_Offset(node,byte)
1070 #define Set_Cur_Node_Offset
1071 #define Set_Node_Length_To_R(node,len)
1072 #define Set_Node_Length(node,len)
1073 #define Set_Node_Cur_Length(node,start)
1074 #define Node_Offset(n)
1075 #define Node_Length(n)
1076 #define Set_Node_Offset_Length(node,offset,len)
1077 #define ProgLen(ri) ri->u.proglen
1078 #define SetProgLen(ri,x) ri->u.proglen = x
1079 #define Track_Code(code)
1081 #define ProgLen(ri) ri->u.offsets[0]
1082 #define SetProgLen(ri,x) ri->u.offsets[0] = x
1083 #define Set_Node_Offset_To_R(offset,byte) STMT_START { \
1084 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
1085 __LINE__, (int)(offset), (int)(byte))); \
1086 if((offset) < 0) { \
1087 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
1090 RExC_offsets[2*(offset)-1] = (byte); \
1094 #define Set_Node_Offset(node,byte) \
1095 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (byte)-RExC_start)
1096 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
1098 #define Set_Node_Length_To_R(node,len) STMT_START { \
1099 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
1100 __LINE__, (int)(node), (int)(len))); \
1102 Perl_croak(aTHX_ "value of node is %d in Length macro", \
1105 RExC_offsets[2*(node)] = (len); \
1109 #define Set_Node_Length(node,len) \
1110 Set_Node_Length_To_R(REGNODE_OFFSET(node), len)
1111 #define Set_Node_Cur_Length(node, start) \
1112 Set_Node_Length(node, RExC_parse - start)
1114 /* Get offsets and lengths */
1115 #define Node_Offset(n) (RExC_offsets[2*(REGNODE_OFFSET(n))-1])
1116 #define Node_Length(n) (RExC_offsets[2*(REGNODE_OFFSET(n))])
1118 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
1119 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (offset)); \
1120 Set_Node_Length_To_R(REGNODE_OFFSET(node), (len)); \
1123 #define Track_Code(code) STMT_START { code } STMT_END
1126 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
1127 #define EXPERIMENTAL_INPLACESCAN
1128 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
1132 Perl_re_printf(pTHX_ const char *fmt, ...)
1136 PerlIO *f= Perl_debug_log;
1137 PERL_ARGS_ASSERT_RE_PRINTF;
1139 result = PerlIO_vprintf(f, fmt, ap);
1145 Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
1149 PerlIO *f= Perl_debug_log;
1150 PERL_ARGS_ASSERT_RE_INDENTF;
1151 va_start(ap, depth);
1152 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
1153 result = PerlIO_vprintf(f, fmt, ap);
1157 #endif /* DEBUGGING */
1159 #define DEBUG_RExC_seen() \
1160 DEBUG_OPTIMISE_MORE_r({ \
1161 Perl_re_printf( aTHX_ "RExC_seen: "); \
1163 if (RExC_seen & REG_ZERO_LEN_SEEN) \
1164 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
1166 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
1167 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
1169 if (RExC_seen & REG_GPOS_SEEN) \
1170 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
1172 if (RExC_seen & REG_RECURSE_SEEN) \
1173 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
1175 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
1176 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
1178 if (RExC_seen & REG_VERBARG_SEEN) \
1179 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
1181 if (RExC_seen & REG_CUTGROUP_SEEN) \
1182 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
1184 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
1185 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
1187 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
1188 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
1190 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
1191 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
1193 Perl_re_printf( aTHX_ "\n"); \
1196 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
1197 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
1202 S_debug_show_study_flags(pTHX_ U32 flags, const char *open_str,
1203 const char *close_str)
1208 Perl_re_printf( aTHX_ "%s", open_str);
1209 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_SEOL);
1210 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_MEOL);
1211 DEBUG_SHOW_STUDY_FLAG(flags, SF_IS_INF);
1212 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_PAR);
1213 DEBUG_SHOW_STUDY_FLAG(flags, SF_IN_PAR);
1214 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_EVAL);
1215 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_SUBSTR);
1216 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_AND);
1217 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_OR);
1218 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS);
1219 DEBUG_SHOW_STUDY_FLAG(flags, SCF_WHILEM_VISITED_POS);
1220 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_RESTUDY);
1221 DEBUG_SHOW_STUDY_FLAG(flags, SCF_SEEN_ACCEPT);
1222 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_DOING_RESTUDY);
1223 DEBUG_SHOW_STUDY_FLAG(flags, SCF_IN_DEFINE);
1224 Perl_re_printf( aTHX_ "%s", close_str);
1229 S_debug_studydata(pTHX_ const char *where, scan_data_t *data,
1230 U32 depth, int is_inf)
1232 DECLARE_AND_GET_RE_DEBUG_FLAGS;
1234 DEBUG_OPTIMISE_MORE_r({
1237 Perl_re_indentf(aTHX_ "%s: Pos:%" IVdf "/%" IVdf " Flags: 0x%" UVXf,
1241 (IV)data->pos_delta,
1245 S_debug_show_study_flags(aTHX_ data->flags," [","]");
1247 Perl_re_printf( aTHX_
1248 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s",
1250 (IV)(data->last_closep ? *((data)->last_closep) : -1),
1251 is_inf ? "INF " : ""
1254 if (data->last_found) {
1256 Perl_re_printf(aTHX_
1257 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf,
1258 SvPVX_const(data->last_found),
1260 (IV)data->last_start_min,
1261 (IV)data->last_start_max
1264 for (i = 0; i < 2; i++) {
1265 Perl_re_printf(aTHX_
1266 " %s%s: '%s' @ %" IVdf "/%" IVdf,
1267 data->cur_is_floating == i ? "*" : "",
1268 i ? "Float" : "Fixed",
1269 SvPVX_const(data->substrs[i].str),
1270 (IV)data->substrs[i].min_offset,
1271 (IV)data->substrs[i].max_offset
1273 S_debug_show_study_flags(aTHX_ data->substrs[i].flags," [","]");
1277 Perl_re_printf( aTHX_ "\n");
1283 S_debug_peep(pTHX_ const char *str, const RExC_state_t *pRExC_state,
1284 regnode *scan, U32 depth, U32 flags)
1286 DECLARE_AND_GET_RE_DEBUG_FLAGS;
1293 Next = regnext(scan);
1294 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
1295 Perl_re_indentf( aTHX_ "%s>%3d: %s (%d)",
1298 REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),
1299 Next ? (REG_NODE_NUM(Next)) : 0 );
1300 S_debug_show_study_flags(aTHX_ flags," [ ","]");
1301 Perl_re_printf( aTHX_ "\n");
1306 # define DEBUG_STUDYDATA(where, data, depth, is_inf) \
1307 S_debug_studydata(aTHX_ where, data, depth, is_inf)
1309 # define DEBUG_PEEP(str, scan, depth, flags) \
1310 S_debug_peep(aTHX_ str, pRExC_state, scan, depth, flags)
1313 # define DEBUG_STUDYDATA(where, data, depth, is_inf) NOOP
1314 # define DEBUG_PEEP(str, scan, depth, flags) NOOP
1318 /* =========================================================
1319 * BEGIN edit_distance stuff.
1321 * This calculates how many single character changes of any type are needed to
1322 * transform a string into another one. It is taken from version 3.1 of
1324 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1327 /* Our unsorted dictionary linked list. */
1328 /* Note we use UVs, not chars. */
1333 struct dictionary* next;
1335 typedef struct dictionary item;
1338 PERL_STATIC_INLINE item*
1339 push(UV key, item* curr)
1342 Newx(head, 1, item);
1350 PERL_STATIC_INLINE item*
1351 find(item* head, UV key)
1353 item* iterator = head;
1355 if (iterator->key == key){
1358 iterator = iterator->next;
1364 PERL_STATIC_INLINE item*
1365 uniquePush(item* head, UV key)
1367 item* iterator = head;
1370 if (iterator->key == key) {
1373 iterator = iterator->next;
1376 return push(key, head);
1379 PERL_STATIC_INLINE void
1380 dict_free(item* head)
1382 item* iterator = head;
1385 item* temp = iterator;
1386 iterator = iterator->next;
1393 /* End of Dictionary Stuff */
1395 /* All calculations/work are done here */
1397 S_edit_distance(const UV* src,
1399 const STRLEN x, /* length of src[] */
1400 const STRLEN y, /* length of tgt[] */
1401 const SSize_t maxDistance
1405 UV swapCount, swapScore, targetCharCount, i, j;
1407 UV score_ceil = x + y;
1409 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1411 /* intialize matrix start values */
1412 Newx(scores, ( (x + 2) * (y + 2)), UV);
1413 scores[0] = score_ceil;
1414 scores[1 * (y + 2) + 0] = score_ceil;
1415 scores[0 * (y + 2) + 1] = score_ceil;
1416 scores[1 * (y + 2) + 1] = 0;
1417 head = uniquePush(uniquePush(head, src[0]), tgt[0]);
1422 for (i=1;i<=x;i++) {
1424 head = uniquePush(head, src[i]);
1425 scores[(i+1) * (y + 2) + 1] = i;
1426 scores[(i+1) * (y + 2) + 0] = score_ceil;
1429 for (j=1;j<=y;j++) {
1432 head = uniquePush(head, tgt[j]);
1433 scores[1 * (y + 2) + (j + 1)] = j;
1434 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1437 targetCharCount = find(head, tgt[j-1])->value;
1438 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1440 if (src[i-1] != tgt[j-1]){
1441 scores[(i+1) * (y + 2) + (j + 1)] = MIN(swapScore,(MIN(scores[i * (y + 2) + j], MIN(scores[(i+1) * (y + 2) + j], scores[i * (y + 2) + (j + 1)])) + 1));
1445 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1449 find(head, src[i-1])->value = i;
1453 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1456 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1460 /* END of edit_distance() stuff
1461 * ========================================================= */
1463 /* Mark that we cannot extend a found fixed substring at this point.
1464 Update the longest found anchored substring or the longest found
1465 floating substrings if needed. */
1468 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1469 SSize_t *minlenp, int is_inf)
1471 const STRLEN l = CHR_SVLEN(data->last_found);
1472 SV * const longest_sv = data->substrs[data->cur_is_floating].str;
1473 const STRLEN old_l = CHR_SVLEN(longest_sv);
1474 DECLARE_AND_GET_RE_DEBUG_FLAGS;
1476 PERL_ARGS_ASSERT_SCAN_COMMIT;
1478 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
1479 const U8 i = data->cur_is_floating;
1480 SvSetMagicSV(longest_sv, data->last_found);
1481 data->substrs[i].min_offset = l ? data->last_start_min : data->pos_min;
1484 data->substrs[0].max_offset = data->substrs[0].min_offset;
1486 data->substrs[1].max_offset =
1490 ? data->last_start_max
1491 /* temporary underflow guard for 5.32 */
1492 : data->pos_delta < 0 ? OPTIMIZE_INFTY
1493 : (data->pos_delta > OPTIMIZE_INFTY - data->pos_min
1495 : data->pos_min + data->pos_delta));
1498 data->substrs[i].flags &= ~SF_BEFORE_EOL;
1499 data->substrs[i].flags |= data->flags & SF_BEFORE_EOL;
1500 data->substrs[i].minlenp = minlenp;
1501 data->substrs[i].lookbehind = 0;
1504 SvCUR_set(data->last_found, 0);
1506 SV * const sv = data->last_found;
1507 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1508 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1513 data->last_end = -1;
1514 data->flags &= ~SF_BEFORE_EOL;
1515 DEBUG_STUDYDATA("commit", data, 0, is_inf);
1518 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1519 * list that describes which code points it matches */
1522 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1524 /* Set the SSC 'ssc' to match an empty string or any code point */
1526 PERL_ARGS_ASSERT_SSC_ANYTHING;
1528 assert(is_ANYOF_SYNTHETIC(ssc));
1530 /* mortalize so won't leak */
1531 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
1532 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1536 S_ssc_is_anything(const regnode_ssc *ssc)
1538 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1539 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1540 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1541 * in any way, so there's no point in using it */
1546 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1548 assert(is_ANYOF_SYNTHETIC(ssc));
1550 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1554 /* See if the list consists solely of the range 0 - Infinity */
1555 invlist_iterinit(ssc->invlist);
1556 ret = invlist_iternext(ssc->invlist, &start, &end)
1560 invlist_iterfinish(ssc->invlist);
1566 /* If e.g., both \w and \W are set, matches everything */
1567 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1569 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1570 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1580 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1582 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1583 * string, any code point, or any posix class under locale */
1585 PERL_ARGS_ASSERT_SSC_INIT;
1587 Zero(ssc, 1, regnode_ssc);
1588 set_ANYOF_SYNTHETIC(ssc);
1589 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1592 /* If any portion of the regex is to operate under locale rules that aren't
1593 * fully known at compile time, initialization includes it. The reason
1594 * this isn't done for all regexes is that the optimizer was written under
1595 * the assumption that locale was all-or-nothing. Given the complexity and
1596 * lack of documentation in the optimizer, and that there are inadequate
1597 * test cases for locale, many parts of it may not work properly, it is
1598 * safest to avoid locale unless necessary. */
1599 if (RExC_contains_locale) {
1600 ANYOF_POSIXL_SETALL(ssc);
1603 ANYOF_POSIXL_ZERO(ssc);
1608 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1609 const regnode_ssc *ssc)
1611 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1612 * to the list of code points matched, and locale posix classes; hence does
1613 * not check its flags) */
1618 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1620 assert(is_ANYOF_SYNTHETIC(ssc));
1622 invlist_iterinit(ssc->invlist);
1623 ret = invlist_iternext(ssc->invlist, &start, &end)
1627 invlist_iterfinish(ssc->invlist);
1633 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1640 #define INVLIST_INDEX 0
1641 #define ONLY_LOCALE_MATCHES_INDEX 1
1642 #define DEFERRED_USER_DEFINED_INDEX 2
1645 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1646 const regnode_charclass* const node)
1648 /* Returns a mortal inversion list defining which code points are matched
1649 * by 'node', which is of type ANYOF. Handles complementing the result if
1650 * appropriate. If some code points aren't knowable at this time, the
1651 * returned list must, and will, contain every code point that is a
1655 SV* only_utf8_locale_invlist = NULL;
1657 const U32 n = ARG(node);
1658 bool new_node_has_latin1 = FALSE;
1659 const U8 flags = (inRANGE(OP(node), ANYOFH, ANYOFRb))
1661 : ANYOF_FLAGS(node);
1663 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1665 /* Look at the data structure created by S_set_ANYOF_arg() */
1666 if (n != ANYOF_ONLY_HAS_BITMAP) {
1667 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1668 AV * const av = MUTABLE_AV(SvRV(rv));
1669 SV **const ary = AvARRAY(av);
1670 assert(RExC_rxi->data->what[n] == 's');
1672 if (av_tindex_skip_len_mg(av) >= DEFERRED_USER_DEFINED_INDEX) {
1674 /* Here there are things that won't be known until runtime -- we
1675 * have to assume it could be anything */
1676 invlist = sv_2mortal(_new_invlist(1));
1677 return _add_range_to_invlist(invlist, 0, UV_MAX);
1679 else if (ary[INVLIST_INDEX]) {
1681 /* Use the node's inversion list */
1682 invlist = sv_2mortal(invlist_clone(ary[INVLIST_INDEX], NULL));
1685 /* Get the code points valid only under UTF-8 locales */
1686 if ( (flags & ANYOFL_FOLD)
1687 && av_tindex_skip_len_mg(av) >= ONLY_LOCALE_MATCHES_INDEX)
1689 only_utf8_locale_invlist = ary[ONLY_LOCALE_MATCHES_INDEX];
1694 invlist = sv_2mortal(_new_invlist(0));
1697 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1698 * code points, and an inversion list for the others, but if there are code
1699 * points that should match only conditionally on the target string being
1700 * UTF-8, those are placed in the inversion list, and not the bitmap.
1701 * Since there are circumstances under which they could match, they are
1702 * included in the SSC. But if the ANYOF node is to be inverted, we have
1703 * to exclude them here, so that when we invert below, the end result
1704 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1705 * have to do this here before we add the unconditionally matched code
1707 if (flags & ANYOF_INVERT) {
1708 _invlist_intersection_complement_2nd(invlist,
1713 /* Add in the points from the bit map */
1714 if (! inRANGE(OP(node), ANYOFH, ANYOFRb)) {
1715 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1716 if (ANYOF_BITMAP_TEST(node, i)) {
1717 unsigned int start = i++;
1719 for (; i < NUM_ANYOF_CODE_POINTS
1720 && ANYOF_BITMAP_TEST(node, i); ++i)
1724 invlist = _add_range_to_invlist(invlist, start, i-1);
1725 new_node_has_latin1 = TRUE;
1730 /* If this can match all upper Latin1 code points, have to add them
1731 * as well. But don't add them if inverting, as when that gets done below,
1732 * it would exclude all these characters, including the ones it shouldn't
1733 * that were added just above */
1734 if (! (flags & ANYOF_INVERT) && OP(node) == ANYOFD
1735 && (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1737 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1740 /* Similarly for these */
1741 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1742 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1745 if (flags & ANYOF_INVERT) {
1746 _invlist_invert(invlist);
1748 else if (flags & ANYOFL_FOLD) {
1749 if (new_node_has_latin1) {
1751 /* Under /li, any 0-255 could fold to any other 0-255, depending on
1752 * the locale. We can skip this if there are no 0-255 at all. */
1753 _invlist_union(invlist, PL_Latin1, &invlist);
1755 invlist = add_cp_to_invlist(invlist, LATIN_SMALL_LETTER_DOTLESS_I);
1756 invlist = add_cp_to_invlist(invlist, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
1759 if (_invlist_contains_cp(invlist, LATIN_SMALL_LETTER_DOTLESS_I)) {
1760 invlist = add_cp_to_invlist(invlist, 'I');
1762 if (_invlist_contains_cp(invlist,
1763 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE))
1765 invlist = add_cp_to_invlist(invlist, 'i');
1770 /* Similarly add the UTF-8 locale possible matches. These have to be
1771 * deferred until after the non-UTF-8 locale ones are taken care of just
1772 * above, or it leads to wrong results under ANYOF_INVERT */
1773 if (only_utf8_locale_invlist) {
1774 _invlist_union_maybe_complement_2nd(invlist,
1775 only_utf8_locale_invlist,
1776 flags & ANYOF_INVERT,
1783 /* These two functions currently do the exact same thing */
1784 #define ssc_init_zero ssc_init
1786 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1787 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1789 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1790 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1791 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1794 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1795 const regnode_charclass *and_with)
1797 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1798 * another SSC or a regular ANYOF class. Can create false positives. */
1801 U8 and_with_flags = inRANGE(OP(and_with), ANYOFH, ANYOFRb)
1803 : ANYOF_FLAGS(and_with);
1806 PERL_ARGS_ASSERT_SSC_AND;
1808 assert(is_ANYOF_SYNTHETIC(ssc));
1810 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1811 * the code point inversion list and just the relevant flags */
1812 if (is_ANYOF_SYNTHETIC(and_with)) {
1813 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1814 anded_flags = and_with_flags;
1816 /* XXX This is a kludge around what appears to be deficiencies in the
1817 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1818 * there are paths through the optimizer where it doesn't get weeded
1819 * out when it should. And if we don't make some extra provision for
1820 * it like the code just below, it doesn't get added when it should.
1821 * This solution is to add it only when AND'ing, which is here, and
1822 * only when what is being AND'ed is the pristine, original node
1823 * matching anything. Thus it is like adding it to ssc_anything() but
1824 * only when the result is to be AND'ed. Probably the same solution
1825 * could be adopted for the same problem we have with /l matching,
1826 * which is solved differently in S_ssc_init(), and that would lead to
1827 * fewer false positives than that solution has. But if this solution
1828 * creates bugs, the consequences are only that a warning isn't raised
1829 * that should be; while the consequences for having /l bugs is
1830 * incorrect matches */
1831 if (ssc_is_anything((regnode_ssc *)and_with)) {
1832 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1836 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1837 if (OP(and_with) == ANYOFD) {
1838 anded_flags = and_with_flags & ANYOF_COMMON_FLAGS;
1841 anded_flags = and_with_flags
1842 &( ANYOF_COMMON_FLAGS
1843 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1844 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1845 if (ANYOFL_UTF8_LOCALE_REQD(and_with_flags)) {
1847 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1852 ANYOF_FLAGS(ssc) &= anded_flags;
1854 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1855 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1856 * 'and_with' may be inverted. When not inverted, we have the situation of
1858 * (C1 | P1) & (C2 | P2)
1859 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1860 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1861 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1862 * <= ((C1 & C2) | P1 | P2)
1863 * Alternatively, the last few steps could be:
1864 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1865 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1866 * <= (C1 | C2 | (P1 & P2))
1867 * We favor the second approach if either P1 or P2 is non-empty. This is
1868 * because these components are a barrier to doing optimizations, as what
1869 * they match cannot be known until the moment of matching as they are
1870 * dependent on the current locale, 'AND"ing them likely will reduce or
1872 * But we can do better if we know that C1,P1 are in their initial state (a
1873 * frequent occurrence), each matching everything:
1874 * (<everything>) & (C2 | P2) = C2 | P2
1875 * Similarly, if C2,P2 are in their initial state (again a frequent
1876 * occurrence), the result is a no-op
1877 * (C1 | P1) & (<everything>) = C1 | P1
1880 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1881 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1882 * <= (C1 & ~C2) | (P1 & ~P2)
1885 if ((and_with_flags & ANYOF_INVERT)
1886 && ! is_ANYOF_SYNTHETIC(and_with))
1890 ssc_intersection(ssc,
1892 FALSE /* Has already been inverted */
1895 /* If either P1 or P2 is empty, the intersection will be also; can skip
1897 if (! (and_with_flags & ANYOF_MATCHES_POSIXL)) {
1898 ANYOF_POSIXL_ZERO(ssc);
1900 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1902 /* Note that the Posix class component P from 'and_with' actually
1904 * P = Pa | Pb | ... | Pn
1905 * where each component is one posix class, such as in [\w\s].
1907 * ~P = ~(Pa | Pb | ... | Pn)
1908 * = ~Pa & ~Pb & ... & ~Pn
1909 * <= ~Pa | ~Pb | ... | ~Pn
1910 * The last is something we can easily calculate, but unfortunately
1911 * is likely to have many false positives. We could do better
1912 * in some (but certainly not all) instances if two classes in
1913 * P have known relationships. For example
1914 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1916 * :lower: & :print: = :lower:
1917 * And similarly for classes that must be disjoint. For example,
1918 * since \s and \w can have no elements in common based on rules in
1919 * the POSIX standard,
1920 * \w & ^\S = nothing
1921 * Unfortunately, some vendor locales do not meet the Posix
1922 * standard, in particular almost everything by Microsoft.
1923 * The loop below just changes e.g., \w into \W and vice versa */
1925 regnode_charclass_posixl temp;
1926 int add = 1; /* To calculate the index of the complement */
1928 Zero(&temp, 1, regnode_charclass_posixl);
1929 ANYOF_POSIXL_ZERO(&temp);
1930 for (i = 0; i < ANYOF_MAX; i++) {
1932 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1933 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1935 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1936 ANYOF_POSIXL_SET(&temp, i + add);
1938 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1940 ANYOF_POSIXL_AND(&temp, ssc);
1942 } /* else ssc already has no posixes */
1943 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1944 in its initial state */
1945 else if (! is_ANYOF_SYNTHETIC(and_with)
1946 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1948 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1949 * copy it over 'ssc' */
1950 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1951 if (is_ANYOF_SYNTHETIC(and_with)) {
1952 StructCopy(and_with, ssc, regnode_ssc);
1955 ssc->invlist = anded_cp_list;
1956 ANYOF_POSIXL_ZERO(ssc);
1957 if (and_with_flags & ANYOF_MATCHES_POSIXL) {
1958 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1962 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1963 || (and_with_flags & ANYOF_MATCHES_POSIXL))
1965 /* One or the other of P1, P2 is non-empty. */
1966 if (and_with_flags & ANYOF_MATCHES_POSIXL) {
1967 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1969 ssc_union(ssc, anded_cp_list, FALSE);
1971 else { /* P1 = P2 = empty */
1972 ssc_intersection(ssc, anded_cp_list, FALSE);
1978 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1979 const regnode_charclass *or_with)
1981 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1982 * another SSC or a regular ANYOF class. Can create false positives if
1983 * 'or_with' is to be inverted. */
1987 U8 or_with_flags = inRANGE(OP(or_with), ANYOFH, ANYOFRb)
1989 : ANYOF_FLAGS(or_with);
1991 PERL_ARGS_ASSERT_SSC_OR;
1993 assert(is_ANYOF_SYNTHETIC(ssc));
1995 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1996 * the code point inversion list and just the relevant flags */
1997 if (is_ANYOF_SYNTHETIC(or_with)) {
1998 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1999 ored_flags = or_with_flags;
2002 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
2003 ored_flags = or_with_flags & ANYOF_COMMON_FLAGS;
2004 if (OP(or_with) != ANYOFD) {
2007 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
2008 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
2009 if (ANYOFL_UTF8_LOCALE_REQD(or_with_flags)) {
2011 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
2016 ANYOF_FLAGS(ssc) |= ored_flags;
2018 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
2019 * C2 is the list of code points in 'or-with'; P2, its posix classes.
2020 * 'or_with' may be inverted. When not inverted, we have the simple
2021 * situation of computing:
2022 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
2023 * If P1|P2 yields a situation with both a class and its complement are
2024 * set, like having both \w and \W, this matches all code points, and we
2025 * can delete these from the P component of the ssc going forward. XXX We
2026 * might be able to delete all the P components, but I (khw) am not certain
2027 * about this, and it is better to be safe.
2030 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
2031 * <= (C1 | P1) | ~C2
2032 * <= (C1 | ~C2) | P1
2033 * (which results in actually simpler code than the non-inverted case)
2036 if ((or_with_flags & ANYOF_INVERT)
2037 && ! is_ANYOF_SYNTHETIC(or_with))
2039 /* We ignore P2, leaving P1 going forward */
2040 } /* else Not inverted */
2041 else if (or_with_flags & ANYOF_MATCHES_POSIXL) {
2042 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
2043 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
2045 for (i = 0; i < ANYOF_MAX; i += 2) {
2046 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
2048 ssc_match_all_cp(ssc);
2049 ANYOF_POSIXL_CLEAR(ssc, i);
2050 ANYOF_POSIXL_CLEAR(ssc, i+1);
2058 FALSE /* Already has been inverted */
2063 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
2065 PERL_ARGS_ASSERT_SSC_UNION;
2067 assert(is_ANYOF_SYNTHETIC(ssc));
2069 _invlist_union_maybe_complement_2nd(ssc->invlist,
2076 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
2078 const bool invert2nd)
2080 PERL_ARGS_ASSERT_SSC_INTERSECTION;
2082 assert(is_ANYOF_SYNTHETIC(ssc));
2084 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
2091 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
2093 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
2095 assert(is_ANYOF_SYNTHETIC(ssc));
2097 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
2101 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
2103 /* AND just the single code point 'cp' into the SSC 'ssc' */
2105 SV* cp_list = _new_invlist(2);
2107 PERL_ARGS_ASSERT_SSC_CP_AND;
2109 assert(is_ANYOF_SYNTHETIC(ssc));
2111 cp_list = add_cp_to_invlist(cp_list, cp);
2112 ssc_intersection(ssc, cp_list,
2113 FALSE /* Not inverted */
2115 SvREFCNT_dec_NN(cp_list);
2119 S_ssc_clear_locale(regnode_ssc *ssc)
2121 /* Set the SSC 'ssc' to not match any locale things */
2122 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
2124 assert(is_ANYOF_SYNTHETIC(ssc));
2126 ANYOF_POSIXL_ZERO(ssc);
2127 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
2131 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
2133 /* The synthetic start class is used to hopefully quickly winnow down
2134 * places where a pattern could start a match in the target string. If it
2135 * doesn't really narrow things down that much, there isn't much point to
2136 * having the overhead of using it. This function uses some very crude
2137 * heuristics to decide if to use the ssc or not.
2139 * It returns TRUE if 'ssc' rules out more than half what it considers to
2140 * be the "likely" possible matches, but of course it doesn't know what the
2141 * actual things being matched are going to be; these are only guesses
2143 * For /l matches, it assumes that the only likely matches are going to be
2144 * in the 0-255 range, uniformly distributed, so half of that is 127
2145 * For /a and /d matches, it assumes that the likely matches will be just
2146 * the ASCII range, so half of that is 63
2147 * For /u and there isn't anything matching above the Latin1 range, it
2148 * assumes that that is the only range likely to be matched, and uses
2149 * half that as the cut-off: 127. If anything matches above Latin1,
2150 * it assumes that all of Unicode could match (uniformly), except for
2151 * non-Unicode code points and things in the General Category "Other"
2152 * (unassigned, private use, surrogates, controls and formats). This
2153 * is a much large number. */
2155 U32 count = 0; /* Running total of number of code points matched by
2157 UV start, end; /* Start and end points of current range in inversion
2158 XXX outdated. UTF-8 locales are common, what about invert? list */
2159 const U32 max_code_points = (LOC)
2161 : (( ! UNI_SEMANTICS
2162 || invlist_highest(ssc->invlist) < 256)
2165 const U32 max_match = max_code_points / 2;
2167 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
2169 invlist_iterinit(ssc->invlist);
2170 while (invlist_iternext(ssc->invlist, &start, &end)) {
2171 if (start >= max_code_points) {
2174 end = MIN(end, max_code_points - 1);
2175 count += end - start + 1;
2176 if (count >= max_match) {
2177 invlist_iterfinish(ssc->invlist);
2187 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
2189 /* The inversion list in the SSC is marked mortal; now we need a more
2190 * permanent copy, which is stored the same way that is done in a regular
2191 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
2194 SV* invlist = invlist_clone(ssc->invlist, NULL);
2196 PERL_ARGS_ASSERT_SSC_FINALIZE;
2198 assert(is_ANYOF_SYNTHETIC(ssc));
2200 /* The code in this file assumes that all but these flags aren't relevant
2201 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
2202 * by the time we reach here */
2203 assert(! (ANYOF_FLAGS(ssc)
2204 & ~( ANYOF_COMMON_FLAGS
2205 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
2206 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
2208 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
2210 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist, NULL, NULL);
2211 SvREFCNT_dec(invlist);
2213 /* Make sure is clone-safe */
2214 ssc->invlist = NULL;
2216 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
2217 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
2218 OP(ssc) = ANYOFPOSIXL;
2220 else if (RExC_contains_locale) {
2224 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
2227 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
2228 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
2229 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
2230 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
2231 ? (TRIE_LIST_CUR( idx ) - 1) \
2237 dump_trie(trie,widecharmap,revcharmap)
2238 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
2239 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
2241 These routines dump out a trie in a somewhat readable format.
2242 The _interim_ variants are used for debugging the interim
2243 tables that are used to generate the final compressed
2244 representation which is what dump_trie expects.
2246 Part of the reason for their existence is to provide a form
2247 of documentation as to how the different representations function.
2252 Dumps the final compressed table form of the trie to Perl_debug_log.
2253 Used for debugging make_trie().
2257 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
2258 AV *revcharmap, U32 depth)
2261 SV *sv=sv_newmortal();
2262 int colwidth= widecharmap ? 6 : 4;
2264 DECLARE_AND_GET_RE_DEBUG_FLAGS;
2266 PERL_ARGS_ASSERT_DUMP_TRIE;
2268 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
2269 depth+1, "Match","Base","Ofs" );
2271 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2272 SV ** const tmp = av_fetch( revcharmap, state, 0);
2274 Perl_re_printf( aTHX_ "%*s",
2276 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2277 PL_colors[0], PL_colors[1],
2278 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2279 PERL_PV_ESCAPE_FIRSTCHAR
2284 Perl_re_printf( aTHX_ "\n");
2285 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
2287 for( state = 0 ; state < trie->uniquecharcount ; state++ )
2288 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2289 Perl_re_printf( aTHX_ "\n");
2291 for( state = 1 ; state < trie->statecount ; state++ ) {
2292 const U32 base = trie->states[ state ].trans.base;
2294 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
2296 if ( trie->states[ state ].wordnum ) {
2297 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
2299 Perl_re_printf( aTHX_ "%6s", "" );
2302 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
2307 while( ( base + ofs < trie->uniquecharcount ) ||
2308 ( base + ofs - trie->uniquecharcount < trie->lasttrans
2309 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2313 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
2315 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2316 if ( ( base + ofs >= trie->uniquecharcount )
2317 && ( base + ofs - trie->uniquecharcount
2319 && trie->trans[ base + ofs
2320 - trie->uniquecharcount ].check == state )
2322 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
2323 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2326 Perl_re_printf( aTHX_ "%*s", colwidth," ." );
2330 Perl_re_printf( aTHX_ "]");
2333 Perl_re_printf( aTHX_ "\n" );
2335 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
2337 for (word=1; word <= trie->wordcount; word++) {
2338 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2339 (int)word, (int)(trie->wordinfo[word].prev),
2340 (int)(trie->wordinfo[word].len));
2342 Perl_re_printf( aTHX_ "\n" );
2345 Dumps a fully constructed but uncompressed trie in list form.
2346 List tries normally only are used for construction when the number of
2347 possible chars (trie->uniquecharcount) is very high.
2348 Used for debugging make_trie().
2351 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2352 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2356 SV *sv=sv_newmortal();
2357 int colwidth= widecharmap ? 6 : 4;
2358 DECLARE_AND_GET_RE_DEBUG_FLAGS;
2360 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2362 /* print out the table precompression. */
2363 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
2365 Perl_re_indentf( aTHX_ "%s",
2366 depth+1, "------:-----+-----------------\n" );
2368 for( state=1 ; state < next_alloc ; state ++ ) {
2371 Perl_re_indentf( aTHX_ " %4" UVXf " :",
2372 depth+1, (UV)state );
2373 if ( ! trie->states[ state ].wordnum ) {
2374 Perl_re_printf( aTHX_ "%5s| ","");
2376 Perl_re_printf( aTHX_ "W%4x| ",
2377 trie->states[ state ].wordnum
2380 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2381 SV ** const tmp = av_fetch( revcharmap,
2382 TRIE_LIST_ITEM(state, charid).forid, 0);
2384 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
2386 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2388 PL_colors[0], PL_colors[1],
2389 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2390 | PERL_PV_ESCAPE_FIRSTCHAR
2392 TRIE_LIST_ITEM(state, charid).forid,
2393 (UV)TRIE_LIST_ITEM(state, charid).newstate
2396 Perl_re_printf( aTHX_ "\n%*s| ",
2397 (int)((depth * 2) + 14), "");
2400 Perl_re_printf( aTHX_ "\n");
2405 Dumps a fully constructed but uncompressed trie in table form.
2406 This is the normal DFA style state transition table, with a few
2407 twists to facilitate compression later.
2408 Used for debugging make_trie().
2411 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2412 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2417 SV *sv=sv_newmortal();
2418 int colwidth= widecharmap ? 6 : 4;
2419 DECLARE_AND_GET_RE_DEBUG_FLAGS;
2421 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2424 print out the table precompression so that we can do a visual check
2425 that they are identical.
2428 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
2430 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2431 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2433 Perl_re_printf( aTHX_ "%*s",
2435 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2436 PL_colors[0], PL_colors[1],
2437 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2438 PERL_PV_ESCAPE_FIRSTCHAR
2444 Perl_re_printf( aTHX_ "\n");
2445 Perl_re_indentf( aTHX_ "State+-", depth+1 );
2447 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2448 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
2451 Perl_re_printf( aTHX_ "\n" );
2453 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2455 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
2457 (UV)TRIE_NODENUM( state ) );
2459 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2460 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2462 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
2464 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
2466 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2467 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
2468 (UV)trie->trans[ state ].check );
2470 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
2471 (UV)trie->trans[ state ].check,
2472 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2480 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2481 startbranch: the first branch in the whole branch sequence
2482 first : start branch of sequence of branch-exact nodes.
2483 May be the same as startbranch
2484 last : Thing following the last branch.
2485 May be the same as tail.
2486 tail : item following the branch sequence
2487 count : words in the sequence
2488 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2489 depth : indent depth
2491 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2493 A trie is an N'ary tree where the branches are determined by digital
2494 decomposition of the key. IE, at the root node you look up the 1st character and
2495 follow that branch repeat until you find the end of the branches. Nodes can be
2496 marked as "accepting" meaning they represent a complete word. Eg:
2500 would convert into the following structure. Numbers represent states, letters
2501 following numbers represent valid transitions on the letter from that state, if
2502 the number is in square brackets it represents an accepting state, otherwise it
2503 will be in parenthesis.
2505 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2509 (1) +-i->(6)-+-s->[7]
2511 +-s->(3)-+-h->(4)-+-e->[5]
2513 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2515 This shows that when matching against the string 'hers' we will begin at state 1
2516 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2517 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2518 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2519 single traverse. We store a mapping from accepting to state to which word was
2520 matched, and then when we have multiple possibilities we try to complete the
2521 rest of the regex in the order in which they occurred in the alternation.
2523 The only prior NFA like behaviour that would be changed by the TRIE support is
2524 the silent ignoring of duplicate alternations which are of the form:
2526 / (DUPE|DUPE) X? (?{ ... }) Y /x
2528 Thus EVAL blocks following a trie may be called a different number of times with
2529 and without the optimisation. With the optimisations dupes will be silently
2530 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2531 the following demonstrates:
2533 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2535 which prints out 'word' three times, but
2537 'words'=~/(word|word|word)(?{ print $1 })S/
2539 which doesnt print it out at all. This is due to other optimisations kicking in.
2541 Example of what happens on a structural level:
2543 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2545 1: CURLYM[1] {1,32767}(18)
2556 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2557 and should turn into:
2559 1: CURLYM[1] {1,32767}(18)
2561 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2569 Cases where tail != last would be like /(?foo|bar)baz/:
2579 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2580 and would end up looking like:
2583 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2590 d = uvchr_to_utf8_flags(d, uv, 0);
2592 is the recommended Unicode-aware way of saying
2597 #define TRIE_STORE_REVCHAR(val) \
2600 SV *zlopp = newSV(UTF8_MAXBYTES); \
2601 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2602 unsigned char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2604 SvCUR_set(zlopp, kapow - flrbbbbb); \
2607 av_push(revcharmap, zlopp); \
2609 char ooooff = (char)val; \
2610 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2614 /* This gets the next character from the input, folding it if not already
2616 #define TRIE_READ_CHAR STMT_START { \
2619 /* if it is UTF then it is either already folded, or does not need \
2621 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2623 else if (folder == PL_fold_latin1) { \
2624 /* This folder implies Unicode rules, which in the range expressible \
2625 * by not UTF is the lower case, with the two exceptions, one of \
2626 * which should have been taken care of before calling this */ \
2627 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2628 uvc = toLOWER_L1(*uc); \
2629 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2632 /* raw data, will be folded later if needed */ \
2640 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2641 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2642 U32 ging = TRIE_LIST_LEN( state ) * 2; \
2643 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2644 TRIE_LIST_LEN( state ) = ging; \
2646 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2647 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2648 TRIE_LIST_CUR( state )++; \
2651 #define TRIE_LIST_NEW(state) STMT_START { \
2652 Newx( trie->states[ state ].trans.list, \
2653 4, reg_trie_trans_le ); \
2654 TRIE_LIST_CUR( state ) = 1; \
2655 TRIE_LIST_LEN( state ) = 4; \
2658 #define TRIE_HANDLE_WORD(state) STMT_START { \
2659 U16 dupe= trie->states[ state ].wordnum; \
2660 regnode * const noper_next = regnext( noper ); \
2663 /* store the word for dumping */ \
2665 if (OP(noper) != NOTHING) \
2666 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2668 tmp = newSVpvn_utf8( "", 0, UTF ); \
2669 av_push( trie_words, tmp ); \
2673 trie->wordinfo[curword].prev = 0; \
2674 trie->wordinfo[curword].len = wordlen; \
2675 trie->wordinfo[curword].accept = state; \
2677 if ( noper_next < tail ) { \
2679 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2681 trie->jump[curword] = (U16)(noper_next - convert); \
2683 jumper = noper_next; \
2685 nextbranch= regnext(cur); \
2689 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2690 /* chain, so that when the bits of chain are later */\
2691 /* linked together, the dups appear in the chain */\
2692 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2693 trie->wordinfo[dupe].prev = curword; \
2695 /* we haven't inserted this word yet. */ \
2696 trie->states[ state ].wordnum = curword; \
2701 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2702 ( ( base + charid >= ucharcount \
2703 && base + charid < ubound \
2704 && state == trie->trans[ base - ucharcount + charid ].check \
2705 && trie->trans[ base - ucharcount + charid ].next ) \
2706 ? trie->trans[ base - ucharcount + charid ].next \
2707 : ( state==1 ? special : 0 ) \
2710 #define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2712 TRIE_BITMAP_SET(trie, uvc); \
2713 /* store the folded codepoint */ \
2715 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2718 /* store first byte of utf8 representation of */ \
2719 /* variant codepoints */ \
2720 if (! UVCHR_IS_INVARIANT(uvc)) { \
2721 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2726 #define MADE_JUMP_TRIE 2
2727 #define MADE_EXACT_TRIE 4
2730 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2731 regnode *first, regnode *last, regnode *tail,
2732 U32 word_count, U32 flags, U32 depth)
2734 /* first pass, loop through and scan words */
2735 reg_trie_data *trie;
2736 HV *widecharmap = NULL;
2737 AV *revcharmap = newAV();
2743 regnode *jumper = NULL;
2744 regnode *nextbranch = NULL;
2745 regnode *convert = NULL;
2746 U32 *prev_states; /* temp array mapping each state to previous one */
2747 /* we just use folder as a flag in utf8 */
2748 const U8 * folder = NULL;
2750 /* in the below add_data call we are storing either 'tu' or 'tuaa'
2751 * which stands for one trie structure, one hash, optionally followed
2754 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuaa"));
2755 AV *trie_words = NULL;
2756 /* along with revcharmap, this only used during construction but both are
2757 * useful during debugging so we store them in the struct when debugging.
2760 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2761 STRLEN trie_charcount=0;
2763 SV *re_trie_maxbuff;
2764 DECLARE_AND_GET_RE_DEBUG_FLAGS;
2766 PERL_ARGS_ASSERT_MAKE_TRIE;
2768 PERL_UNUSED_ARG(depth);
2772 case EXACT: case EXACT_REQ8: case EXACTL: break;
2776 case EXACTFLU8: folder = PL_fold_latin1; break;
2777 case EXACTF: folder = PL_fold; break;
2778 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2781 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2783 trie->startstate = 1;
2784 trie->wordcount = word_count;
2785 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2786 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2787 if (flags == EXACT || flags == EXACT_REQ8 || flags == EXACTL)
2788 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2789 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2790 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2793 trie_words = newAV();
2796 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, GV_ADD);
2797 assert(re_trie_maxbuff);
2798 if (!SvIOK(re_trie_maxbuff)) {
2799 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2801 DEBUG_TRIE_COMPILE_r({
2802 Perl_re_indentf( aTHX_
2803 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2805 REG_NODE_NUM(startbranch), REG_NODE_NUM(first),
2806 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2809 /* Find the node we are going to overwrite */
2810 if ( first == startbranch && OP( last ) != BRANCH ) {
2811 /* whole branch chain */
2814 /* branch sub-chain */
2815 convert = NEXTOPER( first );
2818 /* -- First loop and Setup --
2820 We first traverse the branches and scan each word to determine if it
2821 contains widechars, and how many unique chars there are, this is
2822 important as we have to build a table with at least as many columns as we
2825 We use an array of integers to represent the character codes 0..255
2826 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2827 the native representation of the character value as the key and IV's for
2830 *TODO* If we keep track of how many times each character is used we can
2831 remap the columns so that the table compression later on is more
2832 efficient in terms of memory by ensuring the most common value is in the
2833 middle and the least common are on the outside. IMO this would be better
2834 than a most to least common mapping as theres a decent chance the most
2835 common letter will share a node with the least common, meaning the node
2836 will not be compressible. With a middle is most common approach the worst
2837 case is when we have the least common nodes twice.
2841 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2842 regnode *noper = NEXTOPER( cur );
2846 U32 wordlen = 0; /* required init */
2847 STRLEN minchars = 0;
2848 STRLEN maxchars = 0;
2849 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2852 if (OP(noper) == NOTHING) {
2853 /* skip past a NOTHING at the start of an alternation
2854 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2856 * If the next node is not something we are supposed to process
2857 * we will just ignore it due to the condition guarding the
2861 regnode *noper_next= regnext(noper);
2862 if (noper_next < tail)
2867 && ( OP(noper) == flags
2868 || (flags == EXACT && OP(noper) == EXACT_REQ8)
2869 || (flags == EXACTFU && ( OP(noper) == EXACTFU_REQ8
2870 || OP(noper) == EXACTFUP))))
2872 uc= (U8*)STRING(noper);
2873 e= uc + STR_LEN(noper);
2880 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2881 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2882 regardless of encoding */
2883 if (OP( noper ) == EXACTFUP) {
2884 /* false positives are ok, so just set this */
2885 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2889 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2891 TRIE_CHARCOUNT(trie)++;
2894 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2895 * is in effect. Under /i, this character can match itself, or
2896 * anything that folds to it. If not under /i, it can match just
2897 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2898 * all fold to k, and all are single characters. But some folds
2899 * expand to more than one character, so for example LATIN SMALL
2900 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2901 * the string beginning at 'uc' is 'ffi', it could be matched by
2902 * three characters, or just by the one ligature character. (It
2903 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2904 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2905 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2906 * match.) The trie needs to know the minimum and maximum number
2907 * of characters that could match so that it can use size alone to
2908 * quickly reject many match attempts. The max is simple: it is
2909 * the number of folded characters in this branch (since a fold is
2910 * never shorter than what folds to it. */
2914 /* And the min is equal to the max if not under /i (indicated by
2915 * 'folder' being NULL), or there are no multi-character folds. If
2916 * there is a multi-character fold, the min is incremented just
2917 * once, for the character that folds to the sequence. Each
2918 * character in the sequence needs to be added to the list below of
2919 * characters in the trie, but we count only the first towards the
2920 * min number of characters needed. This is done through the
2921 * variable 'foldlen', which is returned by the macros that look
2922 * for these sequences as the number of bytes the sequence
2923 * occupies. Each time through the loop, we decrement 'foldlen' by
2924 * how many bytes the current char occupies. Only when it reaches
2925 * 0 do we increment 'minchars' or look for another multi-character
2927 if (folder == NULL) {
2930 else if (foldlen > 0) {
2931 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2936 /* See if *uc is the beginning of a multi-character fold. If
2937 * so, we decrement the length remaining to look at, to account
2938 * for the current character this iteration. (We can use 'uc'
2939 * instead of the fold returned by TRIE_READ_CHAR because the
2940 * macro is smart enough to account for any unfolded
2943 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2944 foldlen -= UTF8SKIP(uc);
2947 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2952 /* The current character (and any potential folds) should be added
2953 * to the possible matching characters for this position in this
2957 U8 folded= folder[ (U8) uvc ];
2958 if ( !trie->charmap[ folded ] ) {
2959 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2960 TRIE_STORE_REVCHAR( folded );
2963 if ( !trie->charmap[ uvc ] ) {
2964 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2965 TRIE_STORE_REVCHAR( uvc );
2968 /* store the codepoint in the bitmap, and its folded
2970 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
2971 set_bit = 0; /* We've done our bit :-) */
2975 /* XXX We could come up with the list of code points that fold
2976 * to this using PL_utf8_foldclosures, except not for
2977 * multi-char folds, as there may be multiple combinations
2978 * there that could work, which needs to wait until runtime to
2979 * resolve (The comment about LIGATURE FFI above is such an
2984 widecharmap = newHV();
2986 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2989 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
2991 if ( !SvTRUE( *svpp ) ) {
2992 sv_setiv( *svpp, ++trie->uniquecharcount );
2993 TRIE_STORE_REVCHAR(uvc);
2996 } /* end loop through characters in this branch of the trie */
2998 /* We take the min and max for this branch and combine to find the min
2999 * and max for all branches processed so far */
3000 if( cur == first ) {
3001 trie->minlen = minchars;
3002 trie->maxlen = maxchars;
3003 } else if (minchars < trie->minlen) {
3004 trie->minlen = minchars;
3005 } else if (maxchars > trie->maxlen) {
3006 trie->maxlen = maxchars;
3008 } /* end first pass */
3009 DEBUG_TRIE_COMPILE_r(
3010 Perl_re_indentf( aTHX_
3011 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
3013 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
3014 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
3015 (int)trie->minlen, (int)trie->maxlen )
3019 We now know what we are dealing with in terms of unique chars and
3020 string sizes so we can calculate how much memory a naive
3021 representation using a flat table will take. If it's over a reasonable
3022 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
3023 conservative but potentially much slower representation using an array
3026 At the end we convert both representations into the same compressed
3027 form that will be used in regexec.c for matching with. The latter
3028 is a form that cannot be used to construct with but has memory
3029 properties similar to the list form and access properties similar
3030 to the table form making it both suitable for fast searches and
3031 small enough that its feasable to store for the duration of a program.
3033 See the comment in the code where the compressed table is produced
3034 inplace from the flat tabe representation for an explanation of how
3035 the compression works.
3040 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
3043 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
3044 > SvIV(re_trie_maxbuff) )
3047 Second Pass -- Array Of Lists Representation
3049 Each state will be represented by a list of charid:state records
3050 (reg_trie_trans_le) the first such element holds the CUR and LEN
3051 points of the allocated array. (See defines above).
3053 We build the initial structure using the lists, and then convert
3054 it into the compressed table form which allows faster lookups
3055 (but cant be modified once converted).
3058 STRLEN transcount = 1;
3060 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
3063 trie->states = (reg_trie_state *)
3064 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3065 sizeof(reg_trie_state) );
3069 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3071 regnode *noper = NEXTOPER( cur );
3072 U32 state = 1; /* required init */
3073 U16 charid = 0; /* sanity init */
3074 U32 wordlen = 0; /* required init */
3076 if (OP(noper) == NOTHING) {
3077 regnode *noper_next= regnext(noper);
3078 if (noper_next < tail)
3080 /* we will undo this assignment if noper does not
3081 * point at a trieable type in the else clause of
3082 * the following statement. */
3086 && ( OP(noper) == flags
3087 || (flags == EXACT && OP(noper) == EXACT_REQ8)
3088 || (flags == EXACTFU && ( OP(noper) == EXACTFU_REQ8
3089 || OP(noper) == EXACTFUP))))
3091 const U8 *uc= (U8*)STRING(noper);
3092 const U8 *e= uc + STR_LEN(noper);
3094 for ( ; uc < e ; uc += len ) {
3099 charid = trie->charmap[ uvc ];
3101 SV** const svpp = hv_fetch( widecharmap,
3108 charid=(U16)SvIV( *svpp );
3111 /* charid is now 0 if we dont know the char read, or
3112 * nonzero if we do */
3119 if ( !trie->states[ state ].trans.list ) {
3120 TRIE_LIST_NEW( state );
3123 check <= TRIE_LIST_USED( state );
3126 if ( TRIE_LIST_ITEM( state, check ).forid
3129 newstate = TRIE_LIST_ITEM( state, check ).newstate;
3134 newstate = next_alloc++;
3135 prev_states[newstate] = state;
3136 TRIE_LIST_PUSH( state, charid, newstate );
3141 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3145 /* If we end up here it is because we skipped past a NOTHING, but did not end up
3146 * on a trieable type. So we need to reset noper back to point at the first regop
3147 * in the branch before we call TRIE_HANDLE_WORD()
3149 noper= NEXTOPER(cur);
3151 TRIE_HANDLE_WORD(state);
3153 } /* end second pass */
3155 /* next alloc is the NEXT state to be allocated */
3156 trie->statecount = next_alloc;
3157 trie->states = (reg_trie_state *)
3158 PerlMemShared_realloc( trie->states,
3160 * sizeof(reg_trie_state) );
3162 /* and now dump it out before we compress it */
3163 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
3164 revcharmap, next_alloc,
3168 trie->trans = (reg_trie_trans *)
3169 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
3176 for( state=1 ; state < next_alloc ; state ++ ) {
3180 DEBUG_TRIE_COMPILE_MORE_r(
3181 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
3185 if (trie->states[state].trans.list) {
3186 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
3190 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3191 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
3192 if ( forid < minid ) {
3194 } else if ( forid > maxid ) {
3198 if ( transcount < tp + maxid - minid + 1) {
3200 trie->trans = (reg_trie_trans *)
3201 PerlMemShared_realloc( trie->trans,
3203 * sizeof(reg_trie_trans) );
3204 Zero( trie->trans + (transcount / 2),
3208 base = trie->uniquecharcount + tp - minid;
3209 if ( maxid == minid ) {
3211 for ( ; zp < tp ; zp++ ) {
3212 if ( ! trie->trans[ zp ].next ) {
3213 base = trie->uniquecharcount + zp - minid;
3214 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
3216 trie->trans[ zp ].check = state;
3222 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
3224 trie->trans[ tp ].check = state;
3229 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3230 const U32 tid = base
3231 - trie->uniquecharcount
3232 + TRIE_LIST_ITEM( state, idx ).forid;
3233 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
3235 trie->trans[ tid ].check = state;
3237 tp += ( maxid - minid + 1 );
3239 Safefree(trie->states[ state ].trans.list);
3242 DEBUG_TRIE_COMPILE_MORE_r(
3243 Perl_re_printf( aTHX_ " base: %d\n",base);
3246 trie->states[ state ].trans.base=base;
3248 trie->lasttrans = tp + 1;
3252 Second Pass -- Flat Table Representation.
3254 we dont use the 0 slot of either trans[] or states[] so we add 1 to
3255 each. We know that we will need Charcount+1 trans at most to store
3256 the data (one row per char at worst case) So we preallocate both
3257 structures assuming worst case.
3259 We then construct the trie using only the .next slots of the entry
3262 We use the .check field of the first entry of the node temporarily
3263 to make compression both faster and easier by keeping track of how
3264 many non zero fields are in the node.
3266 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
3269 There are two terms at use here: state as a TRIE_NODEIDX() which is
3270 a number representing the first entry of the node, and state as a
3271 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
3272 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
3273 if there are 2 entrys per node. eg:
3281 The table is internally in the right hand, idx form. However as we
3282 also have to deal with the states array which is indexed by nodenum
3283 we have to use TRIE_NODENUM() to convert.
3286 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
3289 trie->trans = (reg_trie_trans *)
3290 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
3291 * trie->uniquecharcount + 1,
3292 sizeof(reg_trie_trans) );
3293 trie->states = (reg_trie_state *)
3294 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3295 sizeof(reg_trie_state) );
3296 next_alloc = trie->uniquecharcount + 1;
3299 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3301 regnode *noper = NEXTOPER( cur );
3303 U32 state = 1; /* required init */
3305 U16 charid = 0; /* sanity init */
3306 U32 accept_state = 0; /* sanity init */
3308 U32 wordlen = 0; /* required init */
3310 if (OP(noper) == NOTHING) {
3311 regnode *noper_next= regnext(noper);
3312 if (noper_next < tail)
3314 /* we will undo this assignment if noper does not
3315 * point at a trieable type in the else clause of
3316 * the following statement. */
3320 && ( OP(noper) == flags
3321 || (flags == EXACT && OP(noper) == EXACT_REQ8)
3322 || (flags == EXACTFU && ( OP(noper) == EXACTFU_REQ8
3323 || OP(noper) == EXACTFUP))))
3325 const U8 *uc= (U8*)STRING(noper);
3326 const U8 *e= uc + STR_LEN(noper);
3328 for ( ; uc < e ; uc += len ) {
3333 charid = trie->charmap[ uvc ];
3335 SV* const * const svpp = hv_fetch( widecharmap,
3339 charid = svpp ? (U16)SvIV(*svpp) : 0;
3343 if ( !trie->trans[ state + charid ].next ) {
3344 trie->trans[ state + charid ].next = next_alloc;
3345 trie->trans[ state ].check++;
3346 prev_states[TRIE_NODENUM(next_alloc)]
3347 = TRIE_NODENUM(state);
3348 next_alloc += trie->uniquecharcount;
3350 state = trie->trans[ state + charid ].next;
3352 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3354 /* charid is now 0 if we dont know the char read, or
3355 * nonzero if we do */
3358 /* If we end up here it is because we skipped past a NOTHING, but did not end up
3359 * on a trieable type. So we need to reset noper back to point at the first regop
3360 * in the branch before we call TRIE_HANDLE_WORD().
3362 noper= NEXTOPER(cur);
3364 accept_state = TRIE_NODENUM( state );
3365 TRIE_HANDLE_WORD(accept_state);
3367 } /* end second pass */
3369 /* and now dump it out before we compress it */
3370 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3372 next_alloc, depth+1));
3376 * Inplace compress the table.*
3378 For sparse data sets the table constructed by the trie algorithm will
3379 be mostly 0/FAIL transitions or to put it another way mostly empty.
3380 (Note that leaf nodes will not contain any transitions.)
3382 This algorithm compresses the tables by eliminating most such
3383 transitions, at the cost of a modest bit of extra work during lookup:
3385 - Each states[] entry contains a .base field which indicates the
3386 index in the state[] array wheres its transition data is stored.
3388 - If .base is 0 there are no valid transitions from that node.
3390 - If .base is nonzero then charid is added to it to find an entry in
3393 -If trans[states[state].base+charid].check!=state then the
3394 transition is taken to be a 0/Fail transition. Thus if there are fail
3395 transitions at the front of the node then the .base offset will point
3396 somewhere inside the previous nodes data (or maybe even into a node
3397 even earlier), but the .check field determines if the transition is
3401 The following process inplace converts the table to the compressed
3402 table: We first do not compress the root node 1,and mark all its
3403 .check pointers as 1 and set its .base pointer as 1 as well. This
3404 allows us to do a DFA construction from the compressed table later,
3405 and ensures that any .base pointers we calculate later are greater
3408 - We set 'pos' to indicate the first entry of the second node.
3410 - We then iterate over the columns of the node, finding the first and
3411 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3412 and set the .check pointers accordingly, and advance pos
3413 appropriately and repreat for the next node. Note that when we copy
3414 the next pointers we have to convert them from the original
3415 NODEIDX form to NODENUM form as the former is not valid post
3418 - If a node has no transitions used we mark its base as 0 and do not
3419 advance the pos pointer.
3421 - If a node only has one transition we use a second pointer into the
3422 structure to fill in allocated fail transitions from other states.
3423 This pointer is independent of the main pointer and scans forward
3424 looking for null transitions that are allocated to a state. When it
3425 finds one it writes the single transition into the "hole". If the
3426 pointer doesnt find one the single transition is appended as normal.
3428 - Once compressed we can Renew/realloc the structures to release the
3431 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3432 specifically Fig 3.47 and the associated pseudocode.
3436 const U32 laststate = TRIE_NODENUM( next_alloc );
3439 trie->statecount = laststate;
3441 for ( state = 1 ; state < laststate ; state++ ) {
3443 const U32 stateidx = TRIE_NODEIDX( state );
3444 const U32 o_used = trie->trans[ stateidx ].check;
3445 U32 used = trie->trans[ stateidx ].check;
3446 trie->trans[ stateidx ].check = 0;
3449 used && charid < trie->uniquecharcount;
3452 if ( flag || trie->trans[ stateidx + charid ].next ) {
3453 if ( trie->trans[ stateidx + charid ].next ) {
3455 for ( ; zp < pos ; zp++ ) {
3456 if ( ! trie->trans[ zp ].next ) {
3460 trie->states[ state ].trans.base
3462 + trie->uniquecharcount
3464 trie->trans[ zp ].next
3465 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3467 trie->trans[ zp ].check = state;
3468 if ( ++zp > pos ) pos = zp;
3475 trie->states[ state ].trans.base
3476 = pos + trie->uniquecharcount - charid ;
3478 trie->trans[ pos ].next
3479 = SAFE_TRIE_NODENUM(
3480 trie->trans[ stateidx + charid ].next );
3481 trie->trans[ pos ].check = state;
3486 trie->lasttrans = pos + 1;
3487 trie->states = (reg_trie_state *)
3488 PerlMemShared_realloc( trie->states, laststate
3489 * sizeof(reg_trie_state) );
3490 DEBUG_TRIE_COMPILE_MORE_r(
3491 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
3493 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3497 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3500 } /* end table compress */
3502 DEBUG_TRIE_COMPILE_MORE_r(
3503 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
3505 (UV)trie->statecount,
3506 (UV)trie->lasttrans)
3508 /* resize the trans array to remove unused space */
3509 trie->trans = (reg_trie_trans *)
3510 PerlMemShared_realloc( trie->trans, trie->lasttrans
3511 * sizeof(reg_trie_trans) );
3513 { /* Modify the program and insert the new TRIE node */
3514 U8 nodetype =(U8)(flags & 0xFF);
3518 regnode *optimize = NULL;
3519 #ifdef RE_TRACK_PATTERN_OFFSETS
3522 U32 mjd_nodelen = 0;
3523 #endif /* RE_TRACK_PATTERN_OFFSETS */
3524 #endif /* DEBUGGING */
3526 This means we convert either the first branch or the first Exact,
3527 depending on whether the thing following (in 'last') is a branch
3528 or not and whther first is the startbranch (ie is it a sub part of
3529 the alternation or is it the whole thing.)
3530 Assuming its a sub part we convert the EXACT otherwise we convert
3531 the whole branch sequence, including the first.
3533 /* Find the node we are going to overwrite */
3534 if ( first != startbranch || OP( last ) == BRANCH ) {
3535 /* branch sub-chain */
3536 NEXT_OFF( first ) = (U16)(last - first);
3537 #ifdef RE_TRACK_PATTERN_OFFSETS
3539 mjd_offset= Node_Offset((convert));
3540 mjd_nodelen= Node_Length((convert));
3543 /* whole branch chain */
3545 #ifdef RE_TRACK_PATTERN_OFFSETS
3548 const regnode *nop = NEXTOPER( convert );
3549 mjd_offset= Node_Offset((nop));
3550 mjd_nodelen= Node_Length((nop));
3554 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
3556 (UV)mjd_offset, (UV)mjd_nodelen)
3559 /* But first we check to see if there is a common prefix we can
3560 split out as an EXACT and put in front of the TRIE node. */
3561 trie->startstate= 1;
3562 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3563 /* we want to find the first state that has more than
3564 * one transition, if that state is not the first state
3565 * then we have a common prefix which we can remove.
3568 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3570 I32 first_ofs = -1; /* keeps track of the ofs of the first
3571 transition, -1 means none */
3573 const U32 base = trie->states[ state ].trans.base;
3575 /* does this state terminate an alternation? */
3576 if ( trie->states[state].wordnum )
3579 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3580 if ( ( base + ofs >= trie->uniquecharcount ) &&
3581 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3582 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3584 if ( ++count > 1 ) {
3585 /* we have more than one transition */
3588 /* if this is the first state there is no common prefix
3589 * to extract, so we can exit */
3590 if ( state == 1 ) break;
3591 tmp = av_fetch( revcharmap, ofs, 0);
3592 ch = (U8*)SvPV_nolen_const( *tmp );
3594 /* if we are on count 2 then we need to initialize the
3595 * bitmap, and store the previous char if there was one
3598 /* clear the bitmap */
3599 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3601 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
3604 if (first_ofs >= 0) {
3605 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
3606 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3608 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3610 Perl_re_printf( aTHX_ "%s", (char*)ch)
3614 /* store the current firstchar in the bitmap */
3615 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3616 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
3622 /* This state has only one transition, its transition is part
3623 * of a common prefix - we need to concatenate the char it
3624 * represents to what we have so far. */
3625 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
3627 char *ch = SvPV( *tmp, len );
3629 SV *sv=sv_newmortal();
3630 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
3632 (UV)state, (UV)first_ofs,
3633 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3634 PL_colors[0], PL_colors[1],
3635 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3636 PERL_PV_ESCAPE_FIRSTCHAR
3641 OP( convert ) = nodetype;
3642 str=STRING(convert);
3643 setSTR_LEN(convert, 0);
3645 assert( ( STR_LEN(convert) + len ) < 256 );
3646 setSTR_LEN(convert, (U8)(STR_LEN(convert) + len));
3652 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
3657 trie->prefixlen = (state-1);
3659 regnode *n = convert+NODE_SZ_STR(convert);
3660 assert( NODE_SZ_STR(convert) <= U16_MAX );
3661 NEXT_OFF(convert) = (U16)(NODE_SZ_STR(convert));
3662 trie->startstate = state;
3663 trie->minlen -= (state - 1);
3664 trie->maxlen -= (state - 1);
3666 /* At least the UNICOS C compiler choked on this
3667 * being argument to DEBUG_r(), so let's just have
3670 #ifdef PERL_EXT_RE_BUILD
3676 regnode *fix = convert;
3677 U32 word = trie->wordcount;
3678 #ifdef RE_TRACK_PATTERN_OFFSETS
3681 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3682 while( ++fix < n ) {
3683 Set_Node_Offset_Length(fix, 0, 0);
3686 SV ** const tmp = av_fetch( trie_words, word, 0 );
3688 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3689 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3691 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3699 NEXT_OFF(convert) = (U16)(tail - convert);
3700 DEBUG_r(optimize= n);
3706 if ( trie->maxlen ) {
3707 NEXT_OFF( convert ) = (U16)(tail - convert);
3708 ARG_SET( convert, data_slot );
3709 /* Store the offset to the first unabsorbed branch in
3710 jump[0], which is otherwise unused by the jump logic.
3711 We use this when dumping a trie and during optimisation. */
3713 trie->jump[0] = (U16)(nextbranch - convert);
3715 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3716 * and there is a bitmap
3717 * and the first "jump target" node we found leaves enough room
3718 * then convert the TRIE node into a TRIEC node, with the bitmap
3719 * embedded inline in the opcode - this is hypothetically faster.
3721 if ( !trie->states[trie->startstate].wordnum
3723 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3725 OP( convert ) = TRIEC;
3726 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3727 PerlMemShared_free(trie->bitmap);
3730 OP( convert ) = TRIE;
3732 /* store the type in the flags */
3733 convert->flags = nodetype;
3737 + regarglen[ OP( convert ) ];
3739 /* XXX We really should free up the resource in trie now,
3740 as we won't use them - (which resources?) dmq */
3742 /* needed for dumping*/
3743 DEBUG_r(if (optimize) {
3744 regnode *opt = convert;
3746 while ( ++opt < optimize) {
3747 Set_Node_Offset_Length(opt, 0, 0);
3750 Try to clean up some of the debris left after the
3753 while( optimize < jumper ) {
3754 Track_Code( mjd_nodelen += Node_Length((optimize)); );
3755 OP( optimize ) = OPTIMIZED;
3756 Set_Node_Offset_Length(optimize, 0, 0);
3759 Set_Node_Offset_Length(convert, mjd_offset, mjd_nodelen);
3761 } /* end node insert */
3763 /* Finish populating the prev field of the wordinfo array. Walk back
3764 * from each accept state until we find another accept state, and if
3765 * so, point the first word's .prev field at the second word. If the
3766 * second already has a .prev field set, stop now. This will be the
3767 * case either if we've already processed that word's accept state,
3768 * or that state had multiple words, and the overspill words were
3769 * already linked up earlier.
3776 for (word=1; word <= trie->wordcount; word++) {
3778 if (trie->wordinfo[word].prev)
3780 state = trie->wordinfo[word].accept;
3782 state = prev_states[state];
3785 prev = trie->states[state].wordnum;
3789 trie->wordinfo[word].prev = prev;
3791 Safefree(prev_states);
3795 /* and now dump out the compressed format */
3796 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3798 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3800 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3801 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3803 SvREFCNT_dec_NN(revcharmap);
3807 : trie->startstate>1
3813 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3815 /* The Trie is constructed and compressed now so we can build a fail array if
3818 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3820 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3824 We find the fail state for each state in the trie, this state is the longest
3825 proper suffix of the current state's 'word' that is also a proper prefix of
3826 another word in our trie. State 1 represents the word '' and is thus the
3827 default fail state. This allows the DFA not to have to restart after its
3828 tried and failed a word at a given point, it simply continues as though it
3829 had been matching the other word in the first place.
3831 'abcdgu'=~/abcdefg|cdgu/
3832 When we get to 'd' we are still matching the first word, we would encounter
3833 'g' which would fail, which would bring us to the state representing 'd' in
3834 the second word where we would try 'g' and succeed, proceeding to match
3837 /* add a fail transition */
3838 const U32 trie_offset = ARG(source);
3839 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3841 const U32 ucharcount = trie->uniquecharcount;
3842 const U32 numstates = trie->statecount;
3843 const U32 ubound = trie->lasttrans + ucharcount;
3847 U32 base = trie->states[ 1 ].trans.base;
3850 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3852 DECLARE_AND_GET_RE_DEBUG_FLAGS;
3854 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3855 PERL_UNUSED_CONTEXT;
3857 PERL_UNUSED_ARG(depth);
3860 if ( OP(source) == TRIE ) {
3861 struct regnode_1 *op = (struct regnode_1 *)
3862 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3863 StructCopy(source, op, struct regnode_1);
3864 stclass = (regnode *)op;
3866 struct regnode_charclass *op = (struct regnode_charclass *)
3867 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3868 StructCopy(source, op, struct regnode_charclass);
3869 stclass = (regnode *)op;
3871 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3873 ARG_SET( stclass, data_slot );
3874 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3875 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3876 aho->trie=trie_offset;
3877 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3878 Copy( trie->states, aho->states, numstates, reg_trie_state );
3879 Newx( q, numstates, U32);
3880 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3883 /* initialize fail[0..1] to be 1 so that we always have
3884 a valid final fail state */
3885 fail[ 0 ] = fail[ 1 ] = 1;
3887 for ( charid = 0; charid < ucharcount ; charid++ ) {
3888 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3890 q[ q_write ] = newstate;
3891 /* set to point at the root */
3892 fail[ q[ q_write++ ] ]=1;
3895 while ( q_read < q_write) {
3896 const U32 cur = q[ q_read++ % numstates ];
3897 base = trie->states[ cur ].trans.base;
3899 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3900 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3902 U32 fail_state = cur;
3905 fail_state = fail[ fail_state ];
3906 fail_base = aho->states[ fail_state ].trans.base;
3907 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3909 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3910 fail[ ch_state ] = fail_state;
3911 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3913 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3915 q[ q_write++ % numstates] = ch_state;
3919 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3920 when we fail in state 1, this allows us to use the
3921 charclass scan to find a valid start char. This is based on the principle
3922 that theres a good chance the string being searched contains lots of stuff
3923 that cant be a start char.
3925 fail[ 0 ] = fail[ 1 ] = 0;
3926 DEBUG_TRIE_COMPILE_r({
3927 Perl_re_indentf( aTHX_ "Stclass Failtable (%" UVuf " states): 0",
3928 depth, (UV)numstates
3930 for( q_read=1; q_read<numstates; q_read++ ) {
3931 Perl_re_printf( aTHX_ ", %" UVuf, (UV)fail[q_read]);
3933 Perl_re_printf( aTHX_ "\n");
3936 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3941 /* The below joins as many adjacent EXACTish nodes as possible into a single
3942 * one. The regop may be changed if the node(s) contain certain sequences that
3943 * require special handling. The joining is only done if:
3944 * 1) there is room in the current conglomerated node to entirely contain the
3946 * 2) they are compatible node types
3948 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3949 * these get optimized out
3951 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3952 * as possible, even if that means splitting an existing node so that its first
3953 * part is moved to the preceeding node. This would maximise the efficiency of
3954 * memEQ during matching.
3956 * If a node is to match under /i (folded), the number of characters it matches
3957 * can be different than its character length if it contains a multi-character
3958 * fold. *min_subtract is set to the total delta number of characters of the
3961 * And *unfolded_multi_char is set to indicate whether or not the node contains
3962 * an unfolded multi-char fold. This happens when it won't be known until
3963 * runtime whether the fold is valid or not; namely
3964 * 1) for EXACTF nodes that contain LATIN SMALL LETTER SHARP S, as only if the
3965 * target string being matched against turns out to be UTF-8 is that fold
3967 * 2) for EXACTFL nodes whose folding rules depend on the locale in force at
3969 * (Multi-char folds whose components are all above the Latin1 range are not
3970 * run-time locale dependent, and have already been folded by the time this
3971 * function is called.)
3973 * This is as good a place as any to discuss the design of handling these
3974 * multi-character fold sequences. It's been wrong in Perl for a very long
3975 * time. There are three code points in Unicode whose multi-character folds
3976 * were long ago discovered to mess things up. The previous designs for
3977 * dealing with these involved assigning a special node for them. This
3978 * approach doesn't always work, as evidenced by this example:
3979 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3980 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3981 * would match just the \xDF, it won't be able to handle the case where a
3982 * successful match would have to cross the node's boundary. The new approach
3983 * that hopefully generally solves the problem generates an EXACTFUP node
3984 * that is "sss" in this case.
3986 * It turns out that there are problems with all multi-character folds, and not
3987 * just these three. Now the code is general, for all such cases. The
3988 * approach taken is:
3989 * 1) This routine examines each EXACTFish node that could contain multi-
3990 * character folded sequences. Since a single character can fold into
3991 * such a sequence, the minimum match length for this node is less than
3992 * the number of characters in the node. This routine returns in
3993 * *min_subtract how many characters to subtract from the actual
3994 * length of the string to get a real minimum match length; it is 0 if
3995 * there are no multi-char foldeds. This delta is used by the caller to
3996 * adjust the min length of the match, and the delta between min and max,
3997 * so that the optimizer doesn't reject these possibilities based on size
4000 * 2) For the sequence involving the LATIN SMALL LETTER SHARP S (U+00DF)
4001 * under /u, we fold it to 'ss' in regatom(), and in this routine, after
4002 * joining, we scan for occurrences of the sequence 'ss' in non-UTF-8
4003 * EXACTFU nodes. The node type of such nodes is then changed to
4004 * EXACTFUP, indicating it is problematic, and needs careful handling.
4005 * (The procedures in step 1) above are sufficient to handle this case in
4006 * UTF-8 encoded nodes.) The reason this is problematic is that this is
4007 * the only case where there is a possible fold length change in non-UTF-8
4008 * patterns. By reserving a special node type for problematic cases, the
4009 * far more common regular EXACTFU nodes can be processed faster.
4010 * regexec.c takes advantage of this.
4012 * EXACTFUP has been created as a grab-bag for (hopefully uncommon)
4013 * problematic cases. These all only occur when the pattern is not
4014 * UTF-8. In addition to the 'ss' sequence where there is a possible fold
4015 * length change, it handles the situation where the string cannot be
4016 * entirely folded. The strings in an EXACTFish node are folded as much
4017 * as possible during compilation in regcomp.c. This saves effort in
4018 * regex matching. By using an EXACTFUP node when it is not possible to
4019 * fully fold at compile time, regexec.c can know that everything in an
4020 * EXACTFU node is folded, so folding can be skipped at runtime. The only
4021 * case where folding in EXACTFU nodes can't be done at compile time is
4022 * the presumably uncommon MICRO SIGN, when the pattern isn't UTF-8. This
4023 * is because its fold requires UTF-8 to represent. Thus EXACTFUP nodes
4024 * handle two very different cases. Alternatively, there could have been
4025 * a node type where there are length changes, one for unfolded, and one
4026 * for both. If yet another special case needed to be created, the number
4027 * of required node types would have to go to 7. khw figures that even
4028 * though there are plenty of node types to spare, that the maintenance
4029 * cost wasn't worth the small speedup of doing it that way, especially
4030 * since he thinks the MICRO SIGN is rarely encountered in practice.
4032 * There are other cases where folding isn't done at compile time, but
4033 * none of them are under /u, and hence not for EXACTFU nodes. The folds
4034 * in EXACTFL nodes aren't known until runtime, and vary as the locale
4035 * changes. Some folds in EXACTF depend on if the runtime target string
4036 * is UTF-8 or not. (regatom() will create an EXACTFU node even under /di
4037 * when no fold in it depends on the UTF-8ness of the target string.)
4039 * 3) A problem remains for unfolded multi-char folds. (These occur when the
4040 * validity of the fold won't be known until runtime, and so must remain
4041 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFAA
4042 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
4043 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
4044 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
4045 * The reason this is a problem is that the optimizer part of regexec.c
4046 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
4047 * that a character in the pattern corresponds to at most a single
4048 * character in the target string. (And I do mean character, and not byte
4049 * here, unlike other parts of the documentation that have never been
4050 * updated to account for multibyte Unicode.) Sharp s in EXACTF and
4051 * EXACTFL nodes can match the two character string 'ss'; in EXACTFAA
4052 * nodes it can match "\x{17F}\x{17F}". These, along with other ones in
4053 * EXACTFL nodes, violate the assumption, and they are the only instances
4054 * where it is violated. I'm reluctant to try to change the assumption,
4055 * as the code involved is impenetrable to me (khw), so instead the code
4056 * here punts. This routine examines EXACTFL nodes, and (when the pattern
4057 * isn't UTF-8) EXACTF and EXACTFAA for such unfolded folds, and returns a
4058 * boolean indicating whether or not the node contains such a fold. When
4059 * it is true, the caller sets a flag that later causes the optimizer in
4060 * this file to not set values for the floating and fixed string lengths,
4061 * and thus avoids the optimizer code in regexec.c that makes the invalid
4062 * assumption. Thus, there is no optimization based on string lengths for
4063 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
4064 * EXACTF and EXACTFAA nodes that contain the sharp s. (The reason the
4065 * assumption is wrong only in these cases is that all other non-UTF-8
4066 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
4067 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
4068 * EXACTF nodes because we don't know at compile time if it actually
4069 * matches 'ss' or not. For EXACTF nodes it will match iff the target
4070 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
4071 * always matches; and EXACTFAA where it never does. In an EXACTFAA node
4072 * in a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
4073 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
4074 * string would require the pattern to be forced into UTF-8, the overhead
4075 * of which we want to avoid. Similarly the unfolded multi-char folds in
4076 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
4079 * Similarly, the code that generates tries doesn't currently handle
4080 * not-already-folded multi-char folds, and it looks like a pain to change
4081 * that. Therefore, trie generation of EXACTFAA nodes with the sharp s
4082 * doesn't work. Instead, such an EXACTFAA is turned into a new regnode,
4083 * EXACTFAA_NO_TRIE, which the trie code knows not to handle. Most people
4084 * using /iaa matching will be doing so almost entirely with ASCII
4085 * strings, so this should rarely be encountered in practice */
4088 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
4089 UV *min_subtract, bool *unfolded_multi_char,
4090 U32 flags, regnode *val, U32 depth)
4092 /* Merge several consecutive EXACTish nodes into one. */
4094 regnode *n = regnext(scan);
4096 regnode *next = scan + NODE_SZ_STR(scan);
4100 regnode *stop = scan;
4101 DECLARE_AND_GET_RE_DEBUG_FLAGS;
4103 PERL_UNUSED_ARG(depth);
4106 PERL_ARGS_ASSERT_JOIN_EXACT;
4107 #ifndef EXPERIMENTAL_INPLACESCAN
4108 PERL_UNUSED_ARG(flags);
4109 PERL_UNUSED_ARG(val);
4111 DEBUG_PEEP("join", scan, depth, 0);
4113 assert(PL_regkind[OP(scan)] == EXACT);
4115 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
4116 * EXACT ones that are mergeable to the current one. */
4118 && ( PL_regkind[OP(n)] == NOTHING
4119 || (stringok && PL_regkind[OP(n)] == EXACT))
4121 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
4124 if (OP(n) == TAIL || n > next)
4126 if (PL_regkind[OP(n)] == NOTHING) {
4127 DEBUG_PEEP("skip:", n, depth, 0);
4128 NEXT_OFF(scan) += NEXT_OFF(n);
4129 next = n + NODE_STEP_REGNODE;
4136 else if (stringok) {
4137 const unsigned int oldl = STR_LEN(scan);
4138 regnode * const nnext = regnext(n);
4140 /* XXX I (khw) kind of doubt that this works on platforms (should
4141 * Perl ever run on one) where U8_MAX is above 255 because of lots
4142 * of other assumptions */
4143 /* Don't join if the sum can't fit into a single node */
4144 if (oldl + STR_LEN(n) > U8_MAX)
4147 /* Joining something that requires UTF-8 with something that
4148 * doesn't, means the result requires UTF-8. */
4149 if (OP(scan) == EXACT && (OP(n) == EXACT_REQ8)) {
4150 OP(scan) = EXACT_REQ8;
4152 else if (OP(scan) == EXACT_REQ8 && (OP(n) == EXACT)) {
4153 ; /* join is compatible, no need to change OP */
4155 else if ((OP(scan) == EXACTFU) && (OP(n) == EXACTFU_REQ8)) {
4156 OP(scan) = EXACTFU_REQ8;
4158 else if ((OP(scan) == EXACTFU_REQ8) && (OP(n) == EXACTFU)) {
4159 ; /* join is compatible, no need to change OP */
4161 else if (OP(scan) == EXACTFU && OP(n) == EXACTFU) {
4162 ; /* join is compatible, no need to change OP */
4164 else if (OP(scan) == EXACTFU && OP(n) == EXACTFU_S_EDGE) {
4166 /* Under /di, temporary EXACTFU_S_EDGE nodes are generated,
4167 * which can join with EXACTFU ones. We check for this case
4168 * here. These need to be resolved to either EXACTFU or
4169 * EXACTF at joining time. They have nothing in them that
4170 * would forbid them from being the more desirable EXACTFU
4171 * nodes except that they begin and/or end with a single [Ss].
4172 * The reason this is problematic is because they could be
4173 * joined in this loop with an adjacent node that ends and/or
4174 * begins with [Ss] which would then form the sequence 'ss',
4175 * which matches differently under /di than /ui, in which case
4176 * EXACTFU can't be used. If the 'ss' sequence doesn't get
4177 * formed, the nodes get absorbed into any adjacent EXACTFU
4178 * node. And if the only adjacent node is EXACTF, they get
4179 * absorbed into that, under the theory that a longer node is
4180 * better than two shorter ones, even if one is EXACTFU. Note
4181 * that EXACTFU_REQ8 is generated only for UTF-8 patterns,
4182 * and the EXACTFU_S_EDGE ones only for non-UTF-8. */
4184 if (STRING(n)[STR_LEN(n)-1] == 's') {
4186 /* Here the joined node would end with 's'. If the node
4187 * following the combination is an EXACTF one, it's better to
4188 * join this trailing edge 's' node with that one, leaving the
4189 * current one in 'scan' be the more desirable EXACTFU */
4190 if (OP(nnext) == EXACTF) {
4194 OP(scan) = EXACTFU_S_EDGE;
4196 } /* Otherwise, the beginning 's' of the 2nd node just
4197 becomes an interior 's' in 'scan' */
4199 else if (OP(scan) == EXACTF && OP(n) == EXACTF) {
4200 ; /* join is compatible, no need to change OP */
4202 else if (OP(scan) == EXACTF && OP(n) == EXACTFU_S_EDGE) {
4204 /* EXACTF nodes are compatible for joining with EXACTFU_S_EDGE
4205 * nodes. But the latter nodes can be also joined with EXACTFU
4206 * ones, and that is a better outcome, so if the node following
4207 * 'n' is EXACTFU, quit now so that those two can be joined
4209 if (OP(nnext) == EXACTFU) {
4213 /* The join is compatible, and the combined node will be
4214 * EXACTF. (These don't care if they begin or end with 's' */
4216 else if (OP(scan) == EXACTFU_S_EDGE && OP(n) == EXACTFU_S_EDGE) {
4217 if ( STRING(scan)[STR_LEN(scan)-1] == 's'
4218 && STRING(n)[0] == 's')
4220 /* When combined, we have the sequence 'ss', which means we
4221 * have to remain /di */
4225 else if (OP(scan) == EXACTFU_S_EDGE && OP(n) == EXACTFU) {
4226 if (STRING(n)[0] == 's') {
4227 ; /* Here the join is compatible and the combined node
4228 starts with 's', no need to change OP */
4230 else { /* Now the trailing 's' is in the interior */
4234 else if (OP(scan) == EXACTFU_S_EDGE && OP(n) == EXACTF) {
4236 /* The join is compatible, and the combined node will be
4237 * EXACTF. (These don't care if they begin or end with 's' */
4240 else if (OP(scan) != OP(n)) {
4242 /* The only other compatible joinings are the same node type */
4246 DEBUG_PEEP("merg", n, depth, 0);
4249 NEXT_OFF(scan) += NEXT_OFF(n);
4250 assert( ( STR_LEN(scan) + STR_LEN(n) ) < 256 );
4251 setSTR_LEN(scan, (U8)(STR_LEN(scan) + STR_LEN(n)));
4252 next = n + NODE_SZ_STR(n);
4253 /* Now we can overwrite *n : */
4254 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
4262 #ifdef EXPERIMENTAL_INPLACESCAN
4263 if (flags && !NEXT_OFF(n)) {
4264 DEBUG_PEEP("atch", val, depth, 0);
4265 if (reg_off_by_arg[OP(n)]) {
4266 ARG_SET(n, val - n);
4269 NEXT_OFF(n) = val - n;
4276 /* This temporary node can now be turned into EXACTFU, and must, as
4277 * regexec.c doesn't handle it */
4278 if (OP(scan) == EXACTFU_S_EDGE) {
4283 *unfolded_multi_char = FALSE;
4285 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
4286 * can now analyze for sequences of problematic code points. (Prior to
4287 * this final joining, sequences could have been split over boundaries, and
4288 * hence missed). The sequences only happen in folding, hence for any
4289 * non-EXACT EXACTish node */
4290 if (OP(scan) != EXACT && OP(scan) != EXACT_REQ8 && OP(scan) != EXACTL) {
4291 U8* s0 = (U8*) STRING(scan);
4293 U8* s_end = s0 + STR_LEN(scan);
4295 int total_count_delta = 0; /* Total delta number of characters that
4296 multi-char folds expand to */
4298 /* One pass is made over the node's string looking for all the
4299 * possibilities. To avoid some tests in the loop, there are two main
4300 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
4305 if (OP(scan) == EXACTFL) {
4308 /* An EXACTFL node would already have been changed to another
4309 * node type unless there is at least one character in it that
4310 * is problematic; likely a character whose fold definition
4311 * won't be known until runtime, and so has yet to be folded.
4312 * For all but the UTF-8 locale, folds are 1-1 in length, but
4313 * to handle the UTF-8 case, we need to create a temporary
4314 * folded copy using UTF-8 locale rules in order to analyze it.
4315 * This is because our macros that look to see if a sequence is
4316 * a multi-char fold assume everything is folded (otherwise the
4317 * tests in those macros would be too complicated and slow).
4318 * Note that here, the non-problematic folds will have already
4319 * been done, so we can just copy such characters. We actually
4320 * don't completely fold the EXACTFL string. We skip the
4321 * unfolded multi-char folds, as that would just create work
4322 * below to figure out the size they already are */
4324 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
4327 STRLEN s_len = UTF8SKIP(s);
4328 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
4329 Copy(s, d, s_len, U8);
4332 else if (is_FOLDS_TO_MULTI_utf8(s)) {
4333 *unfolded_multi_char = TRUE;
4334 Copy(s, d, s_len, U8);
4337 else if (isASCII(*s)) {
4338 *(d++) = toFOLD(*s);
4342 _toFOLD_utf8_flags(s, s_end, d, &len, FOLD_FLAGS_FULL);
4348 /* Point the remainder of the routine to look at our temporary
4352 } /* End of creating folded copy of EXACTFL string */
4354 /* Examine the string for a multi-character fold sequence. UTF-8
4355 * patterns have all characters pre-folded by the time this code is
4357 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
4358 length sequence we are looking for is 2 */
4360 int count = 0; /* How many characters in a multi-char fold */
4361 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
4362 if (! len) { /* Not a multi-char fold: get next char */
4367 { /* Here is a generic multi-char fold. */
4368 U8* multi_end = s + len;
4370 /* Count how many characters are in it. In the case of
4371 * /aa, no folds which contain ASCII code points are
4372 * allowed, so check for those, and skip if found. */
4373 if (OP(scan) != EXACTFAA && OP(scan) != EXACTFAA_NO_TRIE) {
4374 count = utf8_length(s, multi_end);
4378 while (s < multi_end) {
4381 goto next_iteration;
4391 /* The delta is how long the sequence is minus 1 (1 is how long
4392 * the character that folds to the sequence is) */
4393 total_count_delta += count - 1;
4397 /* We created a temporary folded copy of the string in EXACTFL
4398 * nodes. Therefore we need to be sure it doesn't go below zero,
4399 * as the real string could be shorter */
4400 if (OP(scan) == EXACTFL) {
4401 int total_chars = utf8_length((U8*) STRING(scan),
4402 (U8*) STRING(scan) + STR_LEN(scan));
4403 if (total_count_delta > total_chars) {
4404 total_count_delta = total_chars;
4408 *min_subtract += total_count_delta;
4411 else if (OP(scan) == EXACTFAA) {
4413 /* Non-UTF-8 pattern, EXACTFAA node. There can't be a multi-char
4414 * fold to the ASCII range (and there are no existing ones in the
4415 * upper latin1 range). But, as outlined in the comments preceding
4416 * this function, we need to flag any occurrences of the sharp s.
4417 * This character forbids trie formation (because of added
4419 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
4420 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
4421 || UNICODE_DOT_DOT_VERSION > 0)
4423 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
4424 OP(scan) = EXACTFAA_NO_TRIE;
4425 *unfolded_multi_char = TRUE;
4431 else if (OP(scan) != EXACTFAA_NO_TRIE) {
4433 /* Non-UTF-8 pattern, not EXACTFAA node. Look for the multi-char
4434 * folds that are all Latin1. As explained in the comments
4435 * preceding this function, we look also for the sharp s in EXACTF
4436 * and EXACTFL nodes; it can be in the final position. Otherwise
4437 * we can stop looking 1 byte earlier because have to find at least
4438 * two characters for a multi-fold */
4439 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
4444 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
4445 if (! len) { /* Not a multi-char fold. */
4446 if (*s == LATIN_SMALL_LETTER_SHARP_S
4447 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
4449 *unfolded_multi_char = TRUE;
4456 && isALPHA_FOLD_EQ(*s, 's')
4457 && isALPHA_FOLD_EQ(*(s+1), 's'))
4460 /* EXACTF nodes need to know that the minimum length
4461 * changed so that a sharp s in the string can match this
4462 * ss in the pattern, but they remain EXACTF nodes, as they
4463 * won't match this unless the target string is in UTF-8,
4464 * which we don't know until runtime. EXACTFL nodes can't
4465 * transform into EXACTFU nodes */
4466 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
4467 OP(scan) = EXACTFUP;
4471 *min_subtract += len - 1;
4479 /* Allow dumping but overwriting the collection of skipped
4480 * ops and/or strings with fake optimized ops */
4481 n = scan + NODE_SZ_STR(scan);
4489 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl", scan, depth, 0);});
4493 /* REx optimizer. Converts nodes into quicker variants "in place".
4494 Finds fixed substrings. */
4496 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
4497 to the position after last scanned or to NULL. */
4499 #define INIT_AND_WITHP \
4500 assert(!and_withp); \
4501 Newx(and_withp, 1, regnode_ssc); \
4502 SAVEFREEPV(and_withp)
4506 S_unwind_scan_frames(pTHX_ const void *p)
4508 scan_frame *f= (scan_frame *)p;
4510 scan_frame *n= f->next_frame;
4516 /* Follow the next-chain of the current node and optimize away
4517 all the NOTHINGs from it.
4520 S_rck_elide_nothing(pTHX_ regnode *node)
4522 PERL_ARGS_ASSERT_RCK_ELIDE_NOTHING;
4524 if (OP(node) != CURLYX) {
4525 const int max = (reg_off_by_arg[OP(node)]
4527 /* I32 may be smaller than U16 on CRAYs! */
4528 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
4529 int off = (reg_off_by_arg[OP(node)] ? ARG(node) : NEXT_OFF(node));
4533 /* Skip NOTHING and LONGJMP. */
4537 (PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
4538 || ((OP(n) == LONGJMP) && (noff = ARG(n)))
4544 if (reg_off_by_arg[OP(node)])
4547 NEXT_OFF(node) = off;
4552 /* the return from this sub is the minimum length that could possibly match */
4554 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
4555 SSize_t *minlenp, SSize_t *deltap,
4560 regnode_ssc *and_withp,
4561 U32 flags, U32 depth, bool was_mutate_ok)
4562 /* scanp: Start here (read-write). */
4563 /* deltap: Write maxlen-minlen here. */
4564 /* last: Stop before this one. */
4565 /* data: string data about the pattern */
4566 /* stopparen: treat close N as END */
4567 /* recursed: which subroutines have we recursed into */
4568 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
4570 SSize_t final_minlen;
4571 /* There must be at least this number of characters to match */
4574 regnode *scan = *scanp, *next;
4576 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
4577 int is_inf_internal = 0; /* The studied chunk is infinite */
4578 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
4579 scan_data_t data_fake;
4580 SV *re_trie_maxbuff = NULL;
4581 regnode *first_non_open = scan;
4582 SSize_t stopmin = OPTIMIZE_INFTY;
4583 scan_frame *frame = NULL;
4584 DECLARE_AND_GET_RE_DEBUG_FLAGS;
4586 PERL_ARGS_ASSERT_STUDY_CHUNK;
4587 RExC_study_started= 1;
4589 Zero(&data_fake, 1, scan_data_t);
4592 while (first_non_open && OP(first_non_open) == OPEN)
4593 first_non_open=regnext(first_non_open);
4599 RExC_study_chunk_recursed_count++;
4601 DEBUG_OPTIMISE_MORE_r(
4603 Perl_re_indentf( aTHX_ "study_chunk stopparen=%ld recursed_count=%lu depth=%lu recursed_depth=%lu scan=%p last=%p",
4604 depth, (long)stopparen,
4605 (unsigned long)RExC_study_chunk_recursed_count,
4606 (unsigned long)depth, (unsigned long)recursed_depth,
4609 if (recursed_depth) {
4612 for ( j = 0 ; j < recursed_depth ; j++ ) {
4613 for ( i = 0 ; i < (U32)RExC_total_parens ; i++ ) {
4614 if (PAREN_TEST(j, i) && (!j || !PAREN_TEST(j - 1, i))) {
4615 Perl_re_printf( aTHX_ " %d",(int)i);
4619 if ( j + 1 < recursed_depth ) {
4620 Perl_re_printf( aTHX_ ",");
4624 Perl_re_printf( aTHX_ "\n");
4627 while ( scan && OP(scan) != END && scan < last ){
4628 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
4629 node length to get a real minimum (because
4630 the folded version may be shorter) */
4631 bool unfolded_multi_char = FALSE;
4632 /* avoid mutating ops if we are anywhere within the recursed or
4633 * enframed handling for a GOSUB: the outermost level will handle it.
4635 bool mutate_ok = was_mutate_ok && !(frame && frame->in_gosub);
4636 /* Peephole optimizer: */
4637 DEBUG_STUDYDATA("Peep", data, depth, is_inf);
4638 DEBUG_PEEP("Peep", scan, depth, flags);
4641 /* The reason we do this here is that we need to deal with things like
4642 * /(?:f)(?:o)(?:o)/ which cant be dealt with by the normal EXACT
4643 * parsing code, as each (?:..) is handled by a different invocation of
4646 if (PL_regkind[OP(scan)] == EXACT
4647 && OP(scan) != LEXACT
4648 && OP(scan) != LEXACT_REQ8
4651 join_exact(pRExC_state, scan, &min_subtract, &unfolded_multi_char,
4652 0, NULL, depth + 1);
4655 /* Follow the next-chain of the current node and optimize
4656 away all the NOTHINGs from it.
4658 rck_elide_nothing(scan);
4660 /* The principal pseudo-switch. Cannot be a switch, since we look into
4661 * several different things. */
4662 if ( OP(scan) == DEFINEP ) {
4664 SSize_t deltanext = 0;
4665 SSize_t fake_last_close = 0;
4666 I32 f = SCF_IN_DEFINE;
4668 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4669 scan = regnext(scan);
4670 assert( OP(scan) == IFTHEN );
4671 DEBUG_PEEP("expect IFTHEN", scan, depth, flags);
4673 data_fake.last_closep= &fake_last_close;
4675 next = regnext(scan);
4676 scan = NEXTOPER(NEXTOPER(scan));
4677 DEBUG_PEEP("scan", scan, depth, flags);
4678 DEBUG_PEEP("next", next, depth, flags);
4680 /* we suppose the run is continuous, last=next...
4681 * NOTE we dont use the return here! */
4682 /* DEFINEP study_chunk() recursion */
4683 (void)study_chunk(pRExC_state, &scan, &minlen,
4684 &deltanext, next, &data_fake, stopparen,
4685 recursed_depth, NULL, f, depth+1, mutate_ok);
4690 OP(scan) == BRANCH ||
4691 OP(scan) == BRANCHJ ||
4694 next = regnext(scan);
4697 /* The op(next)==code check below is to see if we
4698 * have "BRANCH-BRANCH", "BRANCHJ-BRANCHJ", "IFTHEN-IFTHEN"
4699 * IFTHEN is special as it might not appear in pairs.
4700 * Not sure whether BRANCH-BRANCHJ is possible, regardless
4701 * we dont handle it cleanly. */
4702 if (OP(next) == code || code == IFTHEN) {
4703 /* NOTE - There is similar code to this block below for
4704 * handling TRIE nodes on a re-study. If you change stuff here
4705 * check there too. */
4706 SSize_t max1 = 0, min1 = OPTIMIZE_INFTY, num = 0;
4708 regnode * const startbranch=scan;
4710 if (flags & SCF_DO_SUBSTR) {
4711 /* Cannot merge strings after this. */
4712 scan_commit(pRExC_state, data, minlenp, is_inf);
4715 if (flags & SCF_DO_STCLASS)
4716 ssc_init_zero(pRExC_state, &accum);
4718 while (OP(scan) == code) {
4719 SSize_t deltanext, minnext, fake;
4721 regnode_ssc this_class;
4723 DEBUG_PEEP("Branch", scan, depth, flags);
4726 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4728 data_fake.whilem_c = data->whilem_c;
4729 data_fake.last_closep = data->last_closep;
4732 data_fake.last_closep = &fake;
4734 data_fake.pos_delta = delta;
4735 next = regnext(scan);
4737 scan = NEXTOPER(scan); /* everything */
4738 if (code != BRANCH) /* everything but BRANCH */
4739 scan = NEXTOPER(scan);
4741 if (flags & SCF_DO_STCLASS) {
4742 ssc_init(pRExC_state, &this_class);
4743 data_fake.start_class = &this_class;
4744 f = SCF_DO_STCLASS_AND;
4746 if (flags & SCF_WHILEM_VISITED_POS)
4747 f |= SCF_WHILEM_VISITED_POS;
4749 /* we suppose the run is continuous, last=next...*/
4750 /* recurse study_chunk() for each BRANCH in an alternation */
4751 minnext = study_chunk(pRExC_state, &scan, minlenp,
4752 &deltanext, next, &data_fake, stopparen,
4753 recursed_depth, NULL, f, depth+1,
4758 if (deltanext == OPTIMIZE_INFTY) {
4759 is_inf = is_inf_internal = 1;
4760 max1 = OPTIMIZE_INFTY;
4761 } else if (max1 < minnext + deltanext)
4762 max1 = minnext + deltanext;
4764 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4766 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4767 if ( stopmin > minnext)
4768 stopmin = min + min1;
4769 flags &= ~SCF_DO_SUBSTR;
4771 data->flags |= SCF_SEEN_ACCEPT;
4774 if (data_fake.flags & SF_HAS_EVAL)
4775 data->flags |= SF_HAS_EVAL;
4776 data->whilem_c = data_fake.whilem_c;
4778 if (flags & SCF_DO_STCLASS)
4779 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
4781 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
4783 if (flags & SCF_DO_SUBSTR) {
4784 data->pos_min += min1;
4785 if (data->pos_delta >= OPTIMIZE_INFTY - (max1 - min1))
4786 data->pos_delta = OPTIMIZE_INFTY;
4788 data->pos_delta += max1 - min1;
4789 if (max1 != min1 || is_inf)
4790 data->cur_is_floating = 1;
4793 if (delta == OPTIMIZE_INFTY
4794 || OPTIMIZE_INFTY - delta - (max1 - min1) < 0)
4795 delta = OPTIMIZE_INFTY;
4797 delta += max1 - min1;
4798 if (flags & SCF_DO_STCLASS_OR) {
4799 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
4801 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4802 flags &= ~SCF_DO_STCLASS;
4805 else if (flags & SCF_DO_STCLASS_AND) {
4807 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
4808 flags &= ~SCF_DO_STCLASS;
4811 /* Switch to OR mode: cache the old value of
4812 * data->start_class */
4814 StructCopy(data->start_class, and_withp, regnode_ssc);
4815 flags &= ~SCF_DO_STCLASS_AND;
4816 StructCopy(&accum, data->start_class, regnode_ssc);
4817 flags |= SCF_DO_STCLASS_OR;
4821 if (PERL_ENABLE_TRIE_OPTIMISATION
4822 && OP(startbranch) == BRANCH
4827 Assuming this was/is a branch we are dealing with: 'scan'
4828 now points at the item that follows the branch sequence,
4829 whatever it is. We now start at the beginning of the
4830 sequence and look for subsequences of
4836 which would be constructed from a pattern like
4839 If we can find such a subsequence we need to turn the first
4840 element into a trie and then add the subsequent branch exact
4841 strings to the trie.
4845 1. patterns where the whole set of branches can be
4848 2. patterns where only a subset can be converted.
4850 In case 1 we can replace the whole set with a single regop
4851 for the trie. In case 2 we need to keep the start and end
4854 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
4855 becomes BRANCH TRIE; BRANCH X;
4857 There is an additional case, that being where there is a
4858 common prefix, which gets split out into an EXACT like node
4859 preceding the TRIE node.
4861 If x(1..n)==tail then we can do a simple trie, if not we make
4862 a "jump" trie, such that when we match the appropriate word
4863 we "jump" to the appropriate tail node. Essentially we turn
4864 a nested if into a case structure of sorts.
4869 if (!re_trie_maxbuff) {
4870 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
4871 if (!SvIOK(re_trie_maxbuff))
4872 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
4874 if ( SvIV(re_trie_maxbuff)>=0 ) {
4876 regnode *first = (regnode *)NULL;
4877 regnode *prev = (regnode *)NULL;
4878 regnode *tail = scan;
4882 /* var tail is used because there may be a TAIL
4883 regop in the way. Ie, the exacts will point to the
4884 thing following the TAIL, but the last branch will
4885 point at the TAIL. So we advance tail. If we
4886 have nested (?:) we may have to move through several
4890 while ( OP( tail ) == TAIL ) {
4891 /* this is the TAIL generated by (?:) */
4892 tail = regnext( tail );
4896 DEBUG_TRIE_COMPILE_r({
4897 regprop(RExC_rx, RExC_mysv, tail, NULL, pRExC_state);
4898 Perl_re_indentf( aTHX_ "%s %" UVuf ":%s\n",
4900 "Looking for TRIE'able sequences. Tail node is ",
4901 (UV) REGNODE_OFFSET(tail),
4902 SvPV_nolen_const( RExC_mysv )
4908 Step through the branches
4909 cur represents each branch,
4910 noper is the first thing to be matched as part
4912 noper_next is the regnext() of that node.
4914 We normally handle a case like this
4915 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
4916 support building with NOJUMPTRIE, which restricts
4917 the trie logic to structures like /FOO|BAR/.
4919 If noper is a trieable nodetype then the branch is
4920 a possible optimization target. If we are building
4921 under NOJUMPTRIE then we require that noper_next is
4922 the same as scan (our current position in the regex
4925 Once we have two or more consecutive such branches
4926 we can create a trie of the EXACT's contents and
4927 stitch it in place into the program.
4929 If the sequence represents all of the branches in
4930 the alternation we replace the entire thing with a
4933 Otherwise when it is a subsequence we need to
4934 stitch it in place and replace only the relevant
4935 branches. This means the first branch has to remain
4936 as it is used by the alternation logic, and its
4937 next pointer, and needs to be repointed at the item
4938 on the branch chain following the last branch we
4939 have optimized away.
4941 This could be either a BRANCH, in which case the
4942 subsequence is internal, or it could be the item
4943 following the branch sequence in which case the
4944 subsequence is at the end (which does not
4945 necessarily mean the first node is the start of the
4948 TRIE_TYPE(X) is a define which maps the optype to a
4952 ----------------+-----------
4957 EXACTFU_REQ8 | EXACTFU
4961 EXACTFLU8 | EXACTFLU8
4965 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) \
4967 : ( EXACT == (X) || EXACT_REQ8 == (X) ) \
4969 : ( EXACTFU == (X) \
4970 || EXACTFU_REQ8 == (X) \
4971 || EXACTFUP == (X) ) \
4973 : ( EXACTFAA == (X) ) \
4975 : ( EXACTL == (X) ) \
4977 : ( EXACTFLU8 == (X) ) \
4981 /* dont use tail as the end marker for this traverse */
4982 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
4983 regnode * const noper = NEXTOPER( cur );
4984 U8 noper_type = OP( noper );
4985 U8 noper_trietype = TRIE_TYPE( noper_type );
4986 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
4987 regnode * const noper_next = regnext( noper );
4988 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4989 U8 noper_next_trietype = (noper_next && noper_next < tail) ? TRIE_TYPE( noper_next_type ) :0;
4992 DEBUG_TRIE_COMPILE_r({
4993 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4994 Perl_re_indentf( aTHX_ "- %d:%s (%d)",
4996 REG_NODE_NUM(cur), SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur) );
4998 regprop(RExC_rx, RExC_mysv, noper, NULL, pRExC_state);
4999 Perl_re_printf( aTHX_ " -> %d:%s",
5000 REG_NODE_NUM(noper), SvPV_nolen_const(RExC_mysv));
5003 regprop(RExC_rx, RExC_mysv, noper_next, NULL, pRExC_state);
5004 Perl_re_printf( aTHX_ "\t=> %d:%s\t",
5005 REG_NODE_NUM(noper_next), SvPV_nolen_const(RExC_mysv));
5007 Perl_re_printf( aTHX_ "(First==%d,Last==%d,Cur==%d,tt==%s,ntt==%s,nntt==%s)\n",
5008 REG_NODE_NUM(first), REG_NODE_NUM(prev), REG_NODE_NUM(cur),
5009 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
5013 /* Is noper a trieable nodetype that can be merged
5014 * with the current trie (if there is one)? */
5018 ( noper_trietype == NOTHING )
5019 || ( trietype == NOTHING )
5020 || ( trietype == noper_trietype )
5023 && noper_next >= tail
5027 /* Handle mergable triable node Either we are
5028 * the first node in a new trieable sequence,
5029 * in which case we do some bookkeeping,
5030 * otherwise we update the end pointer. */
5033 if ( noper_trietype == NOTHING ) {
5034 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
5035 regnode * const noper_next = regnext( noper );
5036 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
5037 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
5040 if ( noper_next_trietype ) {
5041 trietype = noper_next_trietype;
5042 } else if (noper_next_type) {
5043 /* a NOTHING regop is 1 regop wide.
5044 * We need at least two for a trie
5045 * so we can't merge this in */
5049 trietype = noper_trietype;
5052 if ( trietype == NOTHING )
5053 trietype = noper_trietype;
5058 } /* end handle mergable triable node */
5060 /* handle unmergable node -
5061 * noper may either be a triable node which can
5062 * not be tried together with the current trie,
5063 * or a non triable node */
5065 /* If last is set and trietype is not
5066 * NOTHING then we have found at least two
5067 * triable branch sequences in a row of a
5068 * similar trietype so we can turn them
5069 * into a trie. If/when we allow NOTHING to
5070 * start a trie sequence this condition
5071 * will be required, and it isn't expensive
5072 * so we leave it in for now. */
5073 if ( trietype && trietype != NOTHING )
5074 make_trie( pRExC_state,
5075 startbranch, first, cur, tail,
5076 count, trietype, depth+1 );
5077 prev = NULL; /* note: we clear/update
5078 first, trietype etc below,
5079 so we dont do it here */
5083 && noper_next >= tail
5086 /* noper is triable, so we can start a new
5090 trietype = noper_trietype;
5092 /* if we already saw a first but the
5093 * current node is not triable then we have
5094 * to reset the first information. */
5099 } /* end handle unmergable node */
5100 } /* loop over branches */
5101 DEBUG_TRIE_COMPILE_r({
5102 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
5103 Perl_re_indentf( aTHX_ "- %s (%d) <SCAN FINISHED> ",
5104 depth+1, SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur));
5105 Perl_re_printf( aTHX_ "(First==%d, Last==%d, Cur==%d, tt==%s)\n",
5106 REG_NODE_NUM(first), REG_NODE_NUM(prev), REG_NODE_NUM(cur),
5107 PL_reg_name[trietype]
5111 if ( prev && trietype ) {
5112 if ( trietype != NOTHING ) {
5113 /* the last branch of the sequence was part of
5114 * a trie, so we have to construct it here
5115 * outside of the loop */
5116 made= make_trie( pRExC_state, startbranch,
5117 first, scan, tail, count,
5118 trietype, depth+1 );
5119 #ifdef TRIE_STUDY_OPT
5120 if ( ((made == MADE_EXACT_TRIE &&
5121 startbranch == first)
5122 || ( first_non_open == first )) &&
5124 flags |= SCF_TRIE_RESTUDY;
5125 if ( startbranch == first
5128 RExC_seen &=~REG_TOP_LEVEL_BRANCHES_SEEN;
5133 /* at this point we know whatever we have is a
5134 * NOTHING sequence/branch AND if 'startbranch'
5135 * is 'first' then we can turn the whole thing
5138 if ( startbranch == first ) {
5140 /* the entire thing is a NOTHING sequence,
5141 * something like this: (?:|) So we can
5142 * turn it into a plain NOTHING op. */
5143 DEBUG_TRIE_COMPILE_r({
5144 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
5145 Perl_re_indentf( aTHX_ "- %s (%d) <NOTHING BRANCH SEQUENCE>\n",
5147 SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur));
5150 OP(startbranch)= NOTHING;
5151 NEXT_OFF(startbranch)= tail - startbranch;
5152 for ( opt= startbranch + 1; opt < tail ; opt++ )
5156 } /* end if ( prev) */
5157 } /* TRIE_MAXBUF is non zero */
5161 else if ( code == BRANCHJ ) { /* single branch is optimized. */
5162 scan = NEXTOPER(NEXTOPER(scan));
5163 } else /* single branch is optimized. */
5164 scan = NEXTOPER(scan);
5166 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB) {
5168 regnode *start = NULL;
5169 regnode *end = NULL;
5170 U32 my_recursed_depth= recursed_depth;
5172 if (OP(scan) != SUSPEND) { /* GOSUB */
5173 /* Do setup, note this code has side effects beyond
5174 * the rest of this block. Specifically setting
5175 * RExC_recurse[] must happen at least once during
5178 RExC_recurse[ARG2L(scan)] = scan;
5179 start = REGNODE_p(RExC_open_parens[paren]);
5180 end = REGNODE_p(RExC_close_parens[paren]);
5182 /* NOTE we MUST always execute the above code, even
5183 * if we do nothing with a GOSUB */
5185 ( flags & SCF_IN_DEFINE )
5188 (is_inf_internal || is_inf || (data && data->flags & SF_IS_INF))
5190 ( (flags & (SCF_DO_STCLASS | SCF_DO_SUBSTR)) == 0 )
5193 /* no need to do anything here if we are in a define. */
5194 /* or we are after some kind of infinite construct
5195 * so we can skip recursing into this item.
5196 * Since it is infinite we will not change the maxlen
5197 * or delta, and if we miss something that might raise
5198 * the minlen it will merely pessimise a little.
5200 * Iow /(?(DEFINE)(?<foo>foo|food))a+(?&foo)/
5201 * might result in a minlen of 1 and not of 4,
5202 * but this doesn't make us mismatch, just try a bit
5203 * harder than we should.
5205 * However we must assume this GOSUB is infinite, to
5206 * avoid wrongly applying other optimizations in the
5207 * enclosing scope - see GH 18096, for example.
5209 is_inf = is_inf_internal = 1;
5210 scan= regnext(scan);
5216 || !PAREN_TEST(recursed_depth - 1, paren)
5218 /* it is quite possible that there are more efficient ways
5219 * to do this. We maintain a bitmap per level of recursion
5220 * of which patterns we have entered so we can detect if a
5221 * pattern creates a possible infinite loop. When we
5222 * recurse down a level we copy the previous levels bitmap
5223 * down. When we are at recursion level 0 we zero the top
5224 * level bitmap. It would be nice to implement a different
5225 * more efficient way of doing this. In particular the top
5226 * level bitmap may be unnecessary.
5228 if (!recursed_depth) {
5229 Zero(RExC_study_chunk_recursed, RExC_study_chunk_recursed_bytes, U8);
5231 Copy(PAREN_OFFSET(recursed_depth - 1),
5232 PAREN_OFFSET(recursed_depth),
5233 RExC_study_chunk_recursed_bytes, U8);
5235 /* we havent recursed into this paren yet, so recurse into it */
5236 DEBUG_STUDYDATA("gosub-set", data, depth, is_inf);
5237 PAREN_SET(recursed_depth, paren);
5238 my_recursed_depth= recursed_depth + 1;
5240 DEBUG_STUDYDATA("gosub-inf", data, depth, is_inf);
5241 /* some form of infinite recursion, assume infinite length
5243 if (flags & SCF_DO_SUBSTR) {
5244 scan_commit(pRExC_state, data, minlenp, is_inf);
5245 data->cur_is_floating = 1;
5247 is_inf = is_inf_internal = 1;
5248 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5249 ssc_anything(data->start_class);
5250 flags &= ~SCF_DO_STCLASS;
5252 start= NULL; /* reset start so we dont recurse later on. */
5257 end = regnext(scan);
5260 scan_frame *newframe;
5262 if (!RExC_frame_last) {
5263 Newxz(newframe, 1, scan_frame);
5264 SAVEDESTRUCTOR_X(S_unwind_scan_frames, newframe);
5265 RExC_frame_head= newframe;
5267 } else if (!RExC_frame_last->next_frame) {
5268 Newxz(newframe, 1, scan_frame);
5269 RExC_frame_last->next_frame= newframe;
5270 newframe->prev_frame= RExC_frame_last;
5273 newframe= RExC_frame_last->next_frame;
5275 RExC_frame_last= newframe;
5277 newframe->next_regnode = regnext(scan);
5278 newframe->last_regnode = last;
5279 newframe->stopparen = stopparen;
5280 newframe->prev_recursed_depth = recursed_depth;
5281 newframe->this_prev_frame= frame;
5282 newframe->in_gosub = (
5283 (frame && frame->in_gosub) || OP(scan) == GOSUB
5286 DEBUG_STUDYDATA("frame-new", data, depth, is_inf);
5287 DEBUG_PEEP("fnew", scan, depth, flags);
5294 recursed_depth= my_recursed_depth;
5299 else if (PL_regkind[OP(scan)] == EXACT && ! isEXACTFish(OP(scan))) {
5300 SSize_t bytelen = STR_LEN(scan), charlen;
5304 const U8 * const s = (U8*)STRING(scan);
5305 uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
5306 charlen = utf8_length(s, s + bytelen);
5308 uc = *((U8*)STRING(scan));
5312 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
5313 /* The code below prefers earlier match for fixed
5314 offset, later match for variable offset. */
5315 if (data->last_end == -1) { /* Update the start info. */
5316 data->last_start_min = data->pos_min;
5317 data->last_start_max =
5318 is_inf ? OPTIMIZE_INFTY
5319 : (data->pos_delta > OPTIMIZE_INFTY - data->pos_min)
5320 ? OPTIMIZE_INFTY : data->pos_min + data->pos_delta;
5322 sv_catpvn(data->last_found, STRING(scan), bytelen);
5324 SvUTF8_on(data->last_found);
5326 SV * const sv = data->last_found;
5327 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
5328 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5329 if (mg && mg->mg_len >= 0)
5330 mg->mg_len += charlen;
5332 data->last_end = data->pos_min + charlen;
5333 data->pos_min += charlen; /* As in the first entry. */
5334 data->flags &= ~SF_BEFORE_EOL;
5337 /* ANDing the code point leaves at most it, and not in locale, and
5338 * can't match null string */
5339 if (flags & SCF_DO_STCLASS_AND) {
5340 ssc_cp_and(data->start_class, uc);
5341 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5342 ssc_clear_locale(data->start_class);
5344 else if (flags & SCF_DO_STCLASS_OR) {
5345 ssc_add_cp(data->start_class, uc);
5346 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5348 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5349 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5351 flags &= ~SCF_DO_STCLASS;
5353 else if (PL_regkind[OP(scan)] == EXACT) {
5354 /* But OP != EXACT!, so is EXACTFish */
5355 SSize_t bytelen = STR_LEN(scan), charlen;
5356 const U8 * s = (U8*)STRING(scan);
5358 /* Replace a length 1 ASCII fold pair node with an ANYOFM node,
5359 * with the mask set to the complement of the bit that differs
5360 * between upper and lower case, and the lowest code point of the
5361 * pair (which the '&' forces) */
5364 && ( OP(scan) == EXACTFAA
5365 || ( OP(scan) == EXACTFU
5366 && ! HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(*s)))
5369 U8 mask = ~ ('A' ^ 'a'); /* These differ in just one bit */
5372 ARG_SET(scan, *s & mask);
5374 /* we're not EXACTFish any more, so restudy */
5378 /* Search for fixed substrings supports EXACT only. */
5379 if (flags & SCF_DO_SUBSTR) {
5381 scan_commit(pRExC_state, data, minlenp, is_inf);
5383 charlen = UTF ? (SSize_t) utf8_length(s, s + bytelen) : bytelen;
5384 if (unfolded_multi_char) {
5385 RExC_seen |= REG_UNFOLDED_MULTI_SEEN;
5387 min += charlen - min_subtract;
5389 if ((SSize_t)min_subtract < OPTIMIZE_INFTY
5390 && delta < OPTIMIZE_INFTY - (SSize_t)min_subtract
5392 delta += min_subtract;
5394 delta = OPTIMIZE_INFTY;
5396 if (flags & SCF_DO_SUBSTR) {
5397 data->pos_min += charlen - min_subtract;
5398 if (data->pos_min < 0) {
5401 if ((SSize_t)min_subtract < OPTIMIZE_INFTY
5402 && data->pos_delta < OPTIMIZE_INFTY - (SSize_t)min_subtract
5404 data->pos_delta += min_subtract;
5406 data->pos_delta = OPTIMIZE_INFTY;
5409 data->cur_is_floating = 1; /* float */
5413 if (flags & SCF_DO_STCLASS) {
5414 SV* EXACTF_invlist = make_exactf_invlist(pRExC_state, scan);
5416 assert(EXACTF_invlist);
5417 if (flags & SCF_DO_STCLASS_AND) {
5418 if (OP(scan) != EXACTFL)
5419 ssc_clear_locale(data->start_class);
5420 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5421 ANYOF_POSIXL_ZERO(data->start_class);
5422 ssc_intersection(data->start_class, EXACTF_invlist, FALSE);
5424 else { /* SCF_DO_STCLASS_OR */
5425 ssc_union(data->start_class, EXACTF_invlist, FALSE);
5426 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5428 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5429 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5431 flags &= ~SCF_DO_STCLASS;
5432 SvREFCNT_dec(EXACTF_invlist);
5435 else if (REGNODE_VARIES(OP(scan))) {
5436 SSize_t mincount, maxcount, minnext, deltanext, pos_before = 0;
5437 I32 fl = 0, f = flags;
5438 regnode * const oscan = scan;
5439 regnode_ssc this_class;
5440 regnode_ssc *oclass = NULL;
5441 I32 next_is_eval = 0;
5443 switch (PL_regkind[OP(scan)]) {
5444 case WHILEM: /* End of (?:...)* . */
5445 scan = NEXTOPER(scan);
5448 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
5449 next = NEXTOPER(scan);
5450 if ( ( PL_regkind[OP(next)] == EXACT
5451 && ! isEXACTFish(OP(next)))
5452 || (flags & SCF_DO_STCLASS))
5455 maxcount = REG_INFTY;
5456 next = regnext(scan);
5457 scan = NEXTOPER(scan);
5461 if (flags & SCF_DO_SUBSTR)
5463 /* This will bypass the formal 'min += minnext * mincount'
5464 * calculation in the do_curly path, so assumes min width
5465 * of the PLUS payload is exactly one. */
5469 next = NEXTOPER(scan);
5471 /* This temporary node can now be turned into EXACTFU, and
5472 * must, as regexec.c doesn't handle it */
5473 if (OP(next) == EXACTFU_S_EDGE && mutate_ok) {
5477 if ( STR_LEN(next) == 1
5478 && isALPHA_A(* STRING(next))
5479 && ( OP(next) == EXACTFAA
5480 || ( OP(next) == EXACTFU
5481 && ! HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(* STRING(next))))
5484 /* These differ in just one bit */
5485 U8 mask = ~ ('A' ^ 'a');
5487 assert(isALPHA_A(* STRING(next)));
5489 /* Then replace it by an ANYOFM node, with
5490 * the mask set to the complement of the
5491 * bit that differs between upper and lower
5492 * case, and the lowest code point of the
5493 * pair (which the '&' forces) */
5495 ARG_SET(next, *STRING(next) & mask);
5499 if (flags & SCF_DO_STCLASS) {
5501 maxcount = REG_INFTY;
5502 next = regnext(scan);
5503 scan = NEXTOPER(scan);
5506 if (flags & SCF_DO_SUBSTR) {
5507 scan_commit(pRExC_state, data, minlenp, is_inf);
5508 /* Cannot extend fixed substrings */
5509 data->cur_is_floating = 1; /* float */
5511 is_inf = is_inf_internal = 1;
5512 scan = regnext(scan);
5513 goto optimize_curly_tail;
5515 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
5516 && (scan->flags == stopparen))
5521 mincount = ARG1(scan);
5522 maxcount = ARG2(scan);
5524 next = regnext(scan);
5525 if (OP(scan) == CURLYX) {
5526 I32 lp = (data ? *(data->last_closep) : 0);
5527 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
5529 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
5530 next_is_eval = (OP(scan) == EVAL);
5532 if (flags & SCF_DO_SUBSTR) {
5534 scan_commit(pRExC_state, data, minlenp, is_inf);
5535 /* Cannot extend fixed substrings */
5536 pos_before = data->pos_min;
5540 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
5542 data->flags |= SF_IS_INF;
5544 if (flags & SCF_DO_STCLASS) {
5545 ssc_init(pRExC_state, &this_class);
5546 oclass = data->start_class;
5547 data->start_class = &this_class;
5548 f |= SCF_DO_STCLASS_AND;
5549 f &= ~SCF_DO_STCLASS_OR;
5551 /* Exclude from super-linear cache processing any {n,m}
5552 regops for which the combination of input pos and regex
5553 pos is not enough information to determine if a match
5556 For example, in the regex /foo(bar\s*){4,8}baz/ with the
5557 regex pos at the \s*, the prospects for a match depend not
5558 only on the input position but also on how many (bar\s*)
5559 repeats into the {4,8} we are. */
5560 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
5561 f &= ~SCF_WHILEM_VISITED_POS;
5563 /* This will finish on WHILEM, setting scan, or on NULL: */
5564 /* recurse study_chunk() on loop bodies */
5565 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
5566 last, data, stopparen, recursed_depth, NULL,
5568 ? (f & ~SCF_DO_SUBSTR)
5570 , depth+1, mutate_ok);
5572 if (flags & SCF_DO_STCLASS)
5573 data->start_class = oclass;
5574 if (mincount == 0 || minnext == 0) {
5575 if (flags & SCF_DO_STCLASS_OR) {
5576 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5578 else if (flags & SCF_DO_STCLASS_AND) {
5579 /* Switch to OR mode: cache the old value of
5580 * data->start_class */
5582 StructCopy(data->start_class, and_withp, regnode_ssc);
5583 flags &= ~SCF_DO_STCLASS_AND;
5584 StructCopy(&this_class, data->start_class, regnode_ssc);
5585 flags |= SCF_DO_STCLASS_OR;
5586 ANYOF_FLAGS(data->start_class)
5587 |= SSC_MATCHES_EMPTY_STRING;
5589 } else { /* Non-zero len */
5590 if (flags & SCF_DO_STCLASS_OR) {
5591 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5592 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5594 else if (flags & SCF_DO_STCLASS_AND)
5595 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5596 flags &= ~SCF_DO_STCLASS;
5598 if (!scan) /* It was not CURLYX, but CURLY. */
5600 if (((flags & (SCF_TRIE_DOING_RESTUDY|SCF_DO_SUBSTR))==SCF_DO_SUBSTR)
5601 /* ? quantifier ok, except for (?{ ... }) */
5602 && (next_is_eval || !(mincount == 0 && maxcount == 1))
5603 && (minnext == 0) && (deltanext == 0)
5604 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
5605 && maxcount <= REG_INFTY/3) /* Complement check for big
5608 _WARN_HELPER(RExC_precomp_end, packWARN(WARN_REGEXP),
5609 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP),
5610 "Quantifier unexpected on zero-length expression "
5611 "in regex m/%" UTF8f "/",
5612 UTF8fARG(UTF, RExC_precomp_end - RExC_precomp,
5616 if ( ( minnext > 0 && mincount >= SSize_t_MAX / minnext )
5617 || min >= SSize_t_MAX - minnext * mincount )
5619 FAIL("Regexp out of space");
5622 min += minnext * mincount;
5623 is_inf_internal |= deltanext == OPTIMIZE_INFTY
5624 || (maxcount == REG_INFTY && minnext + deltanext > 0);
5625 is_inf |= is_inf_internal;
5627 delta = OPTIMIZE_INFTY;
5629 delta += (minnext + deltanext) * maxcount
5630 - minnext * mincount;
5632 /* Try powerful optimization CURLYX => CURLYN. */
5633 if ( OP(oscan) == CURLYX && data
5634 && data->flags & SF_IN_PAR
5635 && !(data->flags & SF_HAS_EVAL)
5636 && !deltanext && minnext == 1
5639 /* Try to optimize to CURLYN. */
5640 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
5641 regnode * const nxt1 = nxt;
5648 if (!REGNODE_SIMPLE(OP(nxt))
5649 && !(PL_regkind[OP(nxt)] == EXACT
5650 && STR_LEN(nxt) == 1))
5656 if (OP(nxt) != CLOSE)
5658 if (RExC_open_parens) {
5661 RExC_open_parens[ARG(nxt1)] = REGNODE_OFFSET(oscan);
5664 RExC_close_parens[ARG(nxt1)] = REGNODE_OFFSET(nxt) + 2;
5666 /* Now we know that nxt2 is the only contents: */
5667 oscan->flags = (U8)ARG(nxt);
5669 OP(nxt1) = NOTHING; /* was OPEN. */
5672 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5673 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
5674 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
5675 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5676 OP(nxt + 1) = OPTIMIZED; /* was count. */
5677 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
5682 /* Try optimization CURLYX => CURLYM. */
5683 if ( OP(oscan) == CURLYX && data
5684 && !(data->flags & SF_HAS_PAR)
5685 && !(data->flags & SF_HAS_EVAL)
5686 && !deltanext /* atom is fixed width */
5687 && minnext != 0 /* CURLYM can't handle zero width */
5688 /* Nor characters whose fold at run-time may be
5689 * multi-character */
5690 && ! (RExC_seen & REG_UNFOLDED_MULTI_SEEN)
5693 /* XXXX How to optimize if data == 0? */
5694 /* Optimize to a simpler form. */
5695 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
5699 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
5700 && (OP(nxt2) != WHILEM))
5702 OP(nxt2) = SUCCEED; /* Whas WHILEM */
5703 /* Need to optimize away parenths. */
5704 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
5705 /* Set the parenth number. */
5706 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
5708 oscan->flags = (U8)ARG(nxt);
5709 if (RExC_open_parens) {
5711 RExC_open_parens[ARG(nxt1)] = REGNODE_OFFSET(oscan);
5714 RExC_close_parens[ARG(nxt1)] = REGNODE_OFFSET(nxt2)
5717 OP(nxt1) = OPTIMIZED; /* was OPEN. */
5718 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5721 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5722 OP(nxt + 1) = OPTIMIZED; /* was count. */
5723 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
5724 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
5727 while ( nxt1 && (OP(nxt1) != WHILEM)) {
5728 regnode *nnxt = regnext(nxt1);
5730 if (reg_off_by_arg[OP(nxt1)])
5731 ARG_SET(nxt1, nxt2 - nxt1);
5732 else if (nxt2 - nxt1 < U16_MAX)
5733 NEXT_OFF(nxt1) = nxt2 - nxt1;
5735 OP(nxt) = NOTHING; /* Cannot beautify */
5740 /* Optimize again: */
5741 /* recurse study_chunk() on optimised CURLYX => CURLYM */
5742 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
5743 NULL, stopparen, recursed_depth, NULL, 0,
5744 depth+1, mutate_ok);
5749 else if ((OP(oscan) == CURLYX)
5750 && (flags & SCF_WHILEM_VISITED_POS)
5751 /* See the comment on a similar expression above.
5752 However, this time it's not a subexpression
5753 we care about, but the expression itself. */
5754 && (maxcount == REG_INFTY)
5756 /* This stays as CURLYX, we can put the count/of pair. */
5757 /* Find WHILEM (as in regexec.c) */
5758 regnode *nxt = oscan + NEXT_OFF(oscan);
5760 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
5762 nxt = PREVOPER(nxt);
5763 if (nxt->flags & 0xf) {
5764 /* we've already set whilem count on this node */
5765 } else if (++data->whilem_c < 16) {
5766 assert(data->whilem_c <= RExC_whilem_seen);
5767 nxt->flags = (U8)(data->whilem_c
5768 | (RExC_whilem_seen << 4)); /* On WHILEM */
5771 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
5773 if (flags & SCF_DO_SUBSTR) {
5774 SV *last_str = NULL;
5775 STRLEN last_chrs = 0;
5776 int counted = mincount != 0;
5778 if (data->last_end > 0 && mincount != 0) { /* Ends with a
5780 SSize_t b = pos_before >= data->last_start_min
5781 ? pos_before : data->last_start_min;
5783 const char * const s = SvPV_const(data->last_found, l);
5784 SSize_t old = b - data->last_start_min;
5788 old = utf8_hop_forward((U8*)s, old,
5789 (U8 *) SvEND(data->last_found))
5792 /* Get the added string: */
5793 last_str = newSVpvn_utf8(s + old, l, UTF);
5794 last_chrs = UTF ? utf8_length((U8*)(s + old),
5795 (U8*)(s + old + l)) : l;
5796 if (deltanext == 0 && pos_before == b) {
5797 /* What was added is a constant string */
5800 SvGROW(last_str, (mincount * l) + 1);
5801 repeatcpy(SvPVX(last_str) + l,
5802 SvPVX_const(last_str), l,
5804 SvCUR_set(last_str, SvCUR(last_str) * mincount);
5805 /* Add additional parts. */
5806 SvCUR_set(data->last_found,
5807 SvCUR(data->last_found) - l);
5808 sv_catsv(data->last_found, last_str);
5810 SV * sv = data->last_found;
5812 SvUTF8(sv) && SvMAGICAL(sv) ?
5813 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5814 if (mg && mg->mg_len >= 0)
5815 mg->mg_len += last_chrs * (mincount-1);
5817 last_chrs *= mincount;
5818 data->last_end += l * (mincount - 1);
5821 /* start offset must point into the last copy */
5822 data->last_start_min += minnext * (mincount - 1);
5823 data->last_start_max =
5826 : data->last_start_max +
5827 (maxcount - 1) * (minnext + data->pos_delta);
5830 /* It is counted once already... */
5831 data->pos_min += minnext * (mincount - counted);
5833 Perl_re_printf( aTHX_ "counted=%" UVuf " deltanext=%" UVuf
5834 " OPTIMIZE_INFTY=%" UVuf " minnext=%" UVuf
5835 " maxcount=%" UVuf " mincount=%" UVuf
5836 " data->pos_delta=%" UVuf "\n",
5837 (UV)counted, (UV)deltanext, (UV)OPTIMIZE_INFTY, (UV)minnext, (UV)maxcount,
5838 (UV)mincount, (UV)data->pos_delta);
5839 if (deltanext != OPTIMIZE_INFTY)
5840 Perl_re_printf( aTHX_ "LHS=%" UVuf " RHS=%" UVuf "\n",
5841 (UV)(-counted * deltanext + (minnext + deltanext) * maxcount
5842 - minnext * mincount), (UV)(OPTIMIZE_INFTY - data->pos_delta));
5844 if (deltanext == OPTIMIZE_INFTY
5845 || data->pos_delta == OPTIMIZE_INFTY
5846 || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= OPTIMIZE_INFTY - data->pos_delta)
5847 data->pos_delta = OPTIMIZE_INFTY;
5849 data->pos_delta += - counted * deltanext +
5850 (minnext + deltanext) * maxcount - minnext * mincount;
5851 if (mincount != maxcount) {
5852 /* Cannot extend fixed substrings found inside
5854 scan_commit(pRExC_state, data, minlenp, is_inf);
5855 if (mincount && last_str) {
5856 SV * const sv = data->last_found;
5857 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
5858 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5862 sv_setsv(sv, last_str);
5863 data->last_end = data->pos_min;
5864 data->last_start_min = data->pos_min - last_chrs;
5865 data->last_start_max = is_inf
5867 : data->pos_min + data->pos_delta - last_chrs;
5869 data->cur_is_floating = 1; /* float */
5871 SvREFCNT_dec(last_str);
5873 if (data && (fl & SF_HAS_EVAL))
5874 data->flags |= SF_HAS_EVAL;
5875 optimize_curly_tail:
5876 rck_elide_nothing(oscan);
5880 Perl_croak(aTHX_ "panic: unexpected varying REx opcode %d",
5884 if (flags & SCF_DO_SUBSTR) {
5885 /* Cannot expect anything... */
5886 scan_commit(pRExC_state, data, minlenp, is_inf);
5887 data->cur_is_floating = 1; /* float */
5889 is_inf = is_inf_internal = 1;
5890 if (flags & SCF_DO_STCLASS_OR) {
5891 if (OP(scan) == CLUMP) {
5892 /* Actually is any start char, but very few code points
5893 * aren't start characters */
5894 ssc_match_all_cp(data->start_class);
5897 ssc_anything(data->start_class);
5900 flags &= ~SCF_DO_STCLASS;
5904 else if (OP(scan) == LNBREAK) {
5905 if (flags & SCF_DO_STCLASS) {
5906 if (flags & SCF_DO_STCLASS_AND) {
5907 ssc_intersection(data->start_class,
5908 PL_XPosix_ptrs[_CC_VERTSPACE], FALSE);
5909 ssc_clear_locale(data->start_class);
5910 ANYOF_FLAGS(data->start_class)
5911 &= ~SSC_MATCHES_EMPTY_STRING;
5913 else if (flags & SCF_DO_STCLASS_OR) {
5914 ssc_union(data->start_class,
5915 PL_XPosix_ptrs[_CC_VERTSPACE],
5917 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5919 /* See commit msg for
5920 * 749e076fceedeb708a624933726e7989f2302f6a */
5921 ANYOF_FLAGS(data->start_class)
5922 &= ~SSC_MATCHES_EMPTY_STRING;
5924 flags &= ~SCF_DO_STCLASS;
5927 if (delta != OPTIMIZE_INFTY)
5928 delta++; /* Because of the 2 char string cr-lf */
5929 if (flags & SCF_DO_SUBSTR) {
5930 /* Cannot expect anything... */
5931 scan_commit(pRExC_state, data, minlenp, is_inf);
5933 if (data->pos_delta != OPTIMIZE_INFTY) {
5934 data->pos_delta += 1;
5936 data->cur_is_floating = 1; /* float */
5939 else if (REGNODE_SIMPLE(OP(scan))) {
5941 if (flags & SCF_DO_SUBSTR) {
5942 scan_commit(pRExC_state, data, minlenp, is_inf);
5946 if (flags & SCF_DO_STCLASS) {
5948 SV* my_invlist = NULL;
5951 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5952 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5954 /* Some of the logic below assumes that switching
5955 locale on will only add false positives. */
5960 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d",
5964 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5965 ssc_match_all_cp(data->start_class);
5970 SV* REG_ANY_invlist = _new_invlist(2);
5971 REG_ANY_invlist = add_cp_to_invlist(REG_ANY_invlist,
5973 if (flags & SCF_DO_STCLASS_OR) {
5974 ssc_union(data->start_class,
5976 TRUE /* TRUE => invert, hence all but \n
5980 else if (flags & SCF_DO_STCLASS_AND) {
5981 ssc_intersection(data->start_class,
5983 TRUE /* TRUE => invert */
5985 ssc_clear_locale(data->start_class);
5987 SvREFCNT_dec_NN(REG_ANY_invlist);
5999 if (flags & SCF_DO_STCLASS_AND)
6000 ssc_and(pRExC_state, data->start_class,
6001 (regnode_charclass *) scan);
6003 ssc_or(pRExC_state, data->start_class,
6004 (regnode_charclass *) scan);
6007 case NANYOFM: /* NANYOFM already contains the inversion of the
6008 input ANYOF data, so, unlike things like
6009 NPOSIXA, don't change 'invert' to TRUE */
6013 SV* cp_list = get_ANYOFM_contents(scan);
6015 if (flags & SCF_DO_STCLASS_OR) {
6016 ssc_union(data->start_class, cp_list, invert);
6018 else if (flags & SCF_DO_STCLASS_AND) {
6019 ssc_intersection(data->start_class, cp_list, invert);
6022 SvREFCNT_dec_NN(cp_list);
6031 cp_list = _add_range_to_invlist(cp_list,
6033 ANYOFRbase(scan) + ANYOFRdelta(scan));
6035 if (flags & SCF_DO_STCLASS_OR) {
6036 ssc_union(data->start_class, cp_list, invert);
6038 else if (flags & SCF_DO_STCLASS_AND) {
6039 ssc_intersection(data->start_class, cp_list, invert);
6042 SvREFCNT_dec_NN(cp_list);
6051 namedclass = classnum_to_namedclass(FLAGS(scan)) + invert;
6052 if (flags & SCF_DO_STCLASS_AND) {
6053 bool was_there = cBOOL(
6054 ANYOF_POSIXL_TEST(data->start_class,
6056 ANYOF_POSIXL_ZERO(data->start_class);
6057 if (was_there) { /* Do an AND */
6058 ANYOF_POSIXL_SET(data->start_class, namedclass);
6060 /* No individual code points can now match */
6061 data->start_class->invlist
6062 = sv_2mortal(_new_invlist(0));
6065 int complement = namedclass + ((invert) ? -1 : 1);
6067 assert(flags & SCF_DO_STCLASS_OR);
6069 /* If the complement of this class was already there,
6070 * the result is that they match all code points,
6071 * (\d + \D == everything). Remove the classes from
6072 * future consideration. Locale is not relevant in
6074 if (ANYOF_POSIXL_TEST(data->start_class, complement)) {
6075 ssc_match_all_cp(data->start_class);
6076 ANYOF_POSIXL_CLEAR(data->start_class, namedclass);
6077 ANYOF_POSIXL_CLEAR(data->start_class, complement);
6079 else { /* The usual case; just add this class to the
6081 ANYOF_POSIXL_SET(data->start_class, namedclass);
6086 case NPOSIXA: /* For these, we always know the exact set of
6091 my_invlist = invlist_clone(PL_Posix_ptrs[FLAGS(scan)], NULL);
6092 goto join_posix_and_ascii;
6100 my_invlist = invlist_clone(PL_XPosix_ptrs[FLAGS(scan)], NULL);
6102 /* NPOSIXD matches all upper Latin1 code points unless the
6103 * target string being matched is UTF-8, which is
6104 * unknowable until match time. Since we are going to
6105 * invert, we want to get rid of all of them so that the
6106 * inversion will match all */
6107 if (OP(scan) == NPOSIXD) {
6108 _invlist_subtract(my_invlist, PL_UpperLatin1,
6112 join_posix_and_ascii:
6114 if (flags & SCF_DO_STCLASS_AND) {
6115 ssc_intersection(data->start_class, my_invlist, invert);
6116 ssc_clear_locale(data->start_class);
6119 assert(flags & SCF_DO_STCLASS_OR);
6120 ssc_union(data->start_class, my_invlist, invert);
6122 SvREFCNT_dec(my_invlist);
6124 if (flags & SCF_DO_STCLASS_OR)
6125 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6126 flags &= ~SCF_DO_STCLASS;
6129 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
6130 data->flags |= (OP(scan) == MEOL
6133 scan_commit(pRExC_state, data, minlenp, is_inf);
6136 else if ( PL_regkind[OP(scan)] == BRANCHJ
6137 /* Lookbehind, or need to calculate parens/evals/stclass: */
6138 && (scan->flags || data || (flags & SCF_DO_STCLASS))
6139 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM))
6141 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
6142 || OP(scan) == UNLESSM )
6144 /* Negative Lookahead/lookbehind
6145 In this case we can't do fixed string optimisation.
6148 SSize_t deltanext, minnext, fake = 0;
6153 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
6155 data_fake.whilem_c = data->whilem_c;
6156 data_fake.last_closep = data->last_closep;
6159 data_fake.last_closep = &fake;
6160 data_fake.pos_delta = delta;
6161 if ( flags & SCF_DO_STCLASS && !scan->flags
6162 && OP(scan) == IFMATCH ) { /* Lookahead */
6163 ssc_init(pRExC_state, &intrnl);
6164 data_fake.start_class = &intrnl;
6165 f |= SCF_DO_STCLASS_AND;
6167 if (flags & SCF_WHILEM_VISITED_POS)
6168 f |= SCF_WHILEM_VISITED_POS;
6169 next = regnext(scan);
6170 nscan = NEXTOPER(NEXTOPER(scan));
6172 /* recurse study_chunk() for lookahead body */
6173 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
6174 last, &data_fake, stopparen,
6175 recursed_depth, NULL, f, depth+1,
6179 || deltanext > (I32) U8_MAX
6180 || minnext > (I32)U8_MAX
6181 || minnext + deltanext > (I32)U8_MAX)
6183 FAIL2("Lookbehind longer than %" UVuf " not implemented",
6187 /* The 'next_off' field has been repurposed to count the
6188 * additional starting positions to try beyond the initial
6189 * one. (This leaves it at 0 for non-variable length
6190 * matches to avoid breakage for those not using this
6193 scan->next_off = deltanext;
6194 ckWARNexperimental(RExC_parse,
6195 WARN_EXPERIMENTAL__VLB,
6196 "Variable length lookbehind is experimental");
6198 scan->flags = (U8)minnext + deltanext;
6201 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
6203 if (data_fake.flags & SF_HAS_EVAL)
6204 data->flags |= SF_HAS_EVAL;
6205 data->whilem_c = data_fake.whilem_c;
6207 if (f & SCF_DO_STCLASS_AND) {
6208 if (flags & SCF_DO_STCLASS_OR) {
6209 /* OR before, AND after: ideally we would recurse with
6210 * data_fake to get the AND applied by study of the
6211 * remainder of the pattern, and then derecurse;
6212 * *** HACK *** for now just treat as "no information".
6213 * See [perl #56690].
6215 ssc_init(pRExC_state, data->start_class);
6217 /* AND before and after: combine and continue. These
6218 * assertions are zero-length, so can match an EMPTY
6220 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
6221 ANYOF_FLAGS(data->start_class)
6222 |= SSC_MATCHES_EMPTY_STRING;
6226 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
6228 /* Positive Lookahead/lookbehind
6229 In this case we can do fixed string optimisation,
6230 but we must be careful about it. Note in the case of
6231 lookbehind the positions will be offset by the minimum
6232 length of the pattern, something we won't know about
6233 until after the recurse.
6235 SSize_t deltanext, fake = 0;
6239 /* We use SAVEFREEPV so that when the full compile
6240 is finished perl will clean up the allocated
6241 minlens when it's all done. This way we don't
6242 have to worry about freeing them when we know
6243 they wont be used, which would be a pain.
6246 Newx( minnextp, 1, SSize_t );
6247 SAVEFREEPV(minnextp);
6250 StructCopy(data, &data_fake, scan_data_t);
6251 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
6254 scan_commit(pRExC_state, &data_fake, minlenp, is_inf);
6255 data_fake.last_found=newSVsv(data->last_found);
6259 data_fake.last_closep = &fake;
6260 data_fake.flags = 0;
6261 data_fake.substrs[0].flags = 0;
6262 data_fake.substrs[1].flags = 0;
6263 data_fake.pos_delta = delta;
6265 data_fake.flags |= SF_IS_INF;
6266 if ( flags & SCF_DO_STCLASS && !scan->flags
6267 && OP(scan) == IFMATCH ) { /* Lookahead */
6268 ssc_init(pRExC_state, &intrnl);
6269 data_fake.start_class = &intrnl;
6270 f |= SCF_DO_STCLASS_AND;
6272 if (flags & SCF_WHILEM_VISITED_POS)
6273 f |= SCF_WHILEM_VISITED_POS;
6274 next = regnext(scan);
6275 nscan = NEXTOPER(NEXTOPER(scan));
6277 /* positive lookahead study_chunk() recursion */
6278 *minnextp = study_chunk(pRExC_state, &nscan, minnextp,
6279 &deltanext, last, &data_fake,
6280 stopparen, recursed_depth, NULL,
6281 f, depth+1, mutate_ok);
6283 assert(0); /* This code has never been tested since this
6284 is normally not compiled */
6286 || deltanext > (I32) U8_MAX
6287 || *minnextp > (I32)U8_MAX
6288 || *minnextp + deltanext > (I32)U8_MAX)
6290 FAIL2("Lookbehind longer than %" UVuf " not implemented",
6295 scan->next_off = deltanext;
6297 scan->flags = (U8)*minnextp + deltanext;
6302 if (f & SCF_DO_STCLASS_AND) {
6303 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
6304 ANYOF_FLAGS(data->start_class) |= SSC_MATCHES_EMPTY_STRING;
6307 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
6309 if (data_fake.flags & SF_HAS_EVAL)
6310 data->flags |= SF_HAS_EVAL;
6311 data->whilem_c = data_fake.whilem_c;
6312 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
6314 if (RExC_rx->minlen<*minnextp)
6315 RExC_rx->minlen=*minnextp;
6316 scan_commit(pRExC_state, &data_fake, minnextp, is_inf);
6317 SvREFCNT_dec_NN(data_fake.last_found);
6319 for (i = 0; i < 2; i++) {
6320 if (data_fake.substrs[i].minlenp != minlenp) {
6321 data->substrs[i].min_offset =
6322 data_fake.substrs[i].min_offset;
6323 data->substrs[i].max_offset =
6324 data_fake.substrs[i].max_offset;
6325 data->substrs[i].minlenp =
6326 data_fake.substrs[i].minlenp;
6327 data->substrs[i].lookbehind += scan->flags;
6335 else if (OP(scan) == OPEN) {
6336 if (stopparen != (I32)ARG(scan))
6339 else if (OP(scan) == CLOSE) {
6340 if (stopparen == (I32)ARG(scan)) {
6343 if ((I32)ARG(scan) == is_par) {
6344 next = regnext(scan);
6346 if ( next && (OP(next) != WHILEM) && next < last)
6347 is_par = 0; /* Disable optimization */
6350 *(data->last_closep) = ARG(scan);
6352 else if (OP(scan) == EVAL) {
6354 data->flags |= SF_HAS_EVAL;
6356 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
6357 if (flags & SCF_DO_SUBSTR) {
6358 scan_commit(pRExC_state, data, minlenp, is_inf);
6359 flags &= ~SCF_DO_SUBSTR;
6361 if (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;
6551 else if (OP(scan) == REGEX_SET) {
6552 Perl_croak(aTHX_ "panic: %s regnode should be resolved"
6553 " before optimization", reg_name[REGEX_SET]);
6556 #endif /* old or new */
6557 #endif /* TRIE_STUDY_OPT */
6559 /* Else: zero-length, ignore. */
6560 scan = regnext(scan);
6565 /* we need to unwind recursion. */
6568 DEBUG_STUDYDATA("frame-end", data, depth, is_inf);
6569 DEBUG_PEEP("fend", scan, depth, flags);
6571 /* restore previous context */
6572 last = frame->last_regnode;
6573 scan = frame->next_regnode;
6574 stopparen = frame->stopparen;
6575 recursed_depth = frame->prev_recursed_depth;
6577 RExC_frame_last = frame->prev_frame;
6578 frame = frame->this_prev_frame;
6579 goto fake_study_recurse;
6583 DEBUG_STUDYDATA("pre-fin", data, depth, is_inf);
6586 *deltap = is_inf_internal ? OPTIMIZE_INFTY : delta;
6588 if (flags & SCF_DO_SUBSTR && is_inf)
6589 data->pos_delta = OPTIMIZE_INFTY - data->pos_min;
6590 if (is_par > (I32)U8_MAX)
6592 if (is_par && pars==1 && data) {
6593 data->flags |= SF_IN_PAR;
6594 data->flags &= ~SF_HAS_PAR;
6596 else if (pars && data) {
6597 data->flags |= SF_HAS_PAR;
6598 data->flags &= ~SF_IN_PAR;
6600 if (flags & SCF_DO_STCLASS_OR)
6601 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6602 if (flags & SCF_TRIE_RESTUDY)
6603 data->flags |= SCF_TRIE_RESTUDY;
6605 DEBUG_STUDYDATA("post-fin", data, depth, is_inf);
6607 final_minlen = min < stopmin
6610 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN)) {
6611 if (final_minlen > OPTIMIZE_INFTY - delta)
6612 RExC_maxlen = OPTIMIZE_INFTY;
6613 else if (RExC_maxlen < final_minlen + delta)
6614 RExC_maxlen = final_minlen + delta;
6616 return final_minlen;
6620 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
6622 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
6624 PERL_ARGS_ASSERT_ADD_DATA;
6626 Renewc(RExC_rxi->data,
6627 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
6628 char, struct reg_data);
6630 Renew(RExC_rxi->data->what, count + n, U8);
6632 Newx(RExC_rxi->data->what, n, U8);
6633 RExC_rxi->data->count = count + n;
6634 Copy(s, RExC_rxi->data->what + count, n, U8);
6638 /*XXX: todo make this not included in a non debugging perl, but appears to be
6639 * used anyway there, in 'use re' */
6640 #ifndef PERL_IN_XSUB_RE
6642 Perl_reginitcolors(pTHX)
6644 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
6646 char *t = savepv(s);
6650 t = strchr(t, '\t');
6656 PL_colors[i] = t = (char *)"";
6661 PL_colors[i++] = (char *)"";
6668 #ifdef TRIE_STUDY_OPT
6669 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
6672 (data.flags & SCF_TRIE_RESTUDY) \
6680 #define CHECK_RESTUDY_GOTO_butfirst
6684 * pregcomp - compile a regular expression into internal code
6686 * Decides which engine's compiler to call based on the hint currently in
6690 #ifndef PERL_IN_XSUB_RE
6692 /* return the currently in-scope regex engine (or the default if none) */
6694 regexp_engine const *
6695 Perl_current_re_engine(pTHX)
6697 if (IN_PERL_COMPILETIME) {
6698 HV * const table = GvHV(PL_hintgv);
6701 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
6702 return &PL_core_reg_engine;
6703 ptr = hv_fetchs(table, "regcomp", FALSE);
6704 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
6705 return &PL_core_reg_engine;
6706 return INT2PTR(regexp_engine*, SvIV(*ptr));
6710 if (!PL_curcop->cop_hints_hash)
6711 return &PL_core_reg_engine;
6712 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
6713 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
6714 return &PL_core_reg_engine;
6715 return INT2PTR(regexp_engine*, SvIV(ptr));
6721 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
6723 regexp_engine const *eng = current_re_engine();
6724 DECLARE_AND_GET_RE_DEBUG_FLAGS;
6726 PERL_ARGS_ASSERT_PREGCOMP;
6728 /* Dispatch a request to compile a regexp to correct regexp engine. */
6730 Perl_re_printf( aTHX_ "Using engine %" UVxf "\n",
6733 return CALLREGCOMP_ENG(eng, pattern, flags);
6737 /* public(ish) entry point for the perl core's own regex compiling code.
6738 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
6739 * pattern rather than a list of OPs, and uses the internal engine rather
6740 * than the current one */
6743 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
6745 SV *pat = pattern; /* defeat constness! */
6747 PERL_ARGS_ASSERT_RE_COMPILE;
6749 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
6750 #ifdef PERL_IN_XSUB_RE
6753 &PL_core_reg_engine,
6755 NULL, NULL, rx_flags, 0);
6759 S_free_codeblocks(pTHX_ struct reg_code_blocks *cbs)
6763 if (--cbs->refcnt > 0)
6765 for (n = 0; n < cbs->count; n++) {
6766 REGEXP *rx = cbs->cb[n].src_regex;
6768 cbs->cb[n].src_regex = NULL;
6769 SvREFCNT_dec_NN(rx);
6777 static struct reg_code_blocks *
6778 S_alloc_code_blocks(pTHX_ int ncode)
6780 struct reg_code_blocks *cbs;
6781 Newx(cbs, 1, struct reg_code_blocks);
6784 SAVEDESTRUCTOR_X(S_free_codeblocks, cbs);
6786 Newx(cbs->cb, ncode, struct reg_code_block);
6793 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
6794 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
6795 * point to the realloced string and length.
6797 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
6801 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
6802 char **pat_p, STRLEN *plen_p, int num_code_blocks)
6804 U8 *const src = (U8*)*pat_p;
6809 DECLARE_AND_GET_RE_DEBUG_FLAGS;
6811 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6812 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
6814 /* 1 for each byte + 1 for each byte that expands to two, + trailing NUL */
6815 Newx(dst, *plen_p + variant_under_utf8_count(src, src + *plen_p) + 1, U8);
6818 while (s < *plen_p) {
6819 append_utf8_from_native_byte(src[s], &d);
6821 if (n < num_code_blocks) {
6822 assert(pRExC_state->code_blocks);
6823 if (!do_end && pRExC_state->code_blocks->cb[n].start == s) {
6824 pRExC_state->code_blocks->cb[n].start = d - dst - 1;
6825 assert(*(d - 1) == '(');
6828 else if (do_end && pRExC_state->code_blocks->cb[n].end == s) {
6829 pRExC_state->code_blocks->cb[n].end = d - dst - 1;
6830 assert(*(d - 1) == ')');
6839 *pat_p = (char*) dst;
6841 RExC_orig_utf8 = RExC_utf8 = 1;
6846 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
6847 * while recording any code block indices, and handling overloading,
6848 * nested qr// objects etc. If pat is null, it will allocate a new
6849 * string, or just return the first arg, if there's only one.
6851 * Returns the malloced/updated pat.
6852 * patternp and pat_count is the array of SVs to be concatted;
6853 * oplist is the optional list of ops that generated the SVs;
6854 * recompile_p is a pointer to a boolean that will be set if
6855 * the regex will need to be recompiled.
6856 * delim, if non-null is an SV that will be inserted between each element
6860 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
6861 SV *pat, SV ** const patternp, int pat_count,
6862 OP *oplist, bool *recompile_p, SV *delim)
6866 bool use_delim = FALSE;
6867 bool alloced = FALSE;
6869 /* if we know we have at least two args, create an empty string,
6870 * then concatenate args to that. For no args, return an empty string */
6871 if (!pat && pat_count != 1) {
6877 for (svp = patternp; svp < patternp + pat_count; svp++) {
6880 STRLEN orig_patlen = 0;
6882 SV *msv = use_delim ? delim : *svp;
6883 if (!msv) msv = &PL_sv_undef;
6885 /* if we've got a delimiter, we go round the loop twice for each
6886 * svp slot (except the last), using the delimiter the second
6895 if (SvTYPE(msv) == SVt_PVAV) {
6896 /* we've encountered an interpolated array within
6897 * the pattern, e.g. /...@a..../. Expand the list of elements,
6898 * then recursively append elements.
6899 * The code in this block is based on S_pushav() */
6901 AV *const av = (AV*)msv;
6902 const SSize_t maxarg = AvFILL(av) + 1;
6906 assert(oplist->op_type == OP_PADAV
6907 || oplist->op_type == OP_RV2AV);
6908 oplist = OpSIBLING(oplist);
6911 if (SvRMAGICAL(av)) {
6914 Newx(array, maxarg, SV*);
6916 for (i=0; i < maxarg; i++) {
6917 SV ** const svp = av_fetch(av, i, FALSE);
6918 array[i] = svp ? *svp : &PL_sv_undef;
6922 array = AvARRAY(av);
6924 pat = S_concat_pat(aTHX_ pRExC_state, pat,
6925 array, maxarg, NULL, recompile_p,
6927 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
6933 /* we make the assumption here that each op in the list of
6934 * op_siblings maps to one SV pushed onto the stack,
6935 * except for code blocks, with have both an OP_NULL and
6937 * This allows us to match up the list of SVs against the
6938 * list of OPs to find the next code block.
6940 * Note that PUSHMARK PADSV PADSV ..
6942 * PADRANGE PADSV PADSV ..
6943 * so the alignment still works. */
6946 if (oplist->op_type == OP_NULL
6947 && (oplist->op_flags & OPf_SPECIAL))
6949 assert(n < pRExC_state->code_blocks->count);
6950 pRExC_state->code_blocks->cb[n].start = pat ? SvCUR(pat) : 0;
6951 pRExC_state->code_blocks->cb[n].block = oplist;
6952 pRExC_state->code_blocks->cb[n].src_regex = NULL;
6955 oplist = OpSIBLING(oplist); /* skip CONST */
6958 oplist = OpSIBLING(oplist);;
6961 /* apply magic and QR overloading to arg */
6964 if (SvROK(msv) && SvAMAGIC(msv)) {
6965 SV *sv = AMG_CALLunary(msv, regexp_amg);
6969 if (SvTYPE(sv) != SVt_REGEXP)
6970 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
6975 /* try concatenation overload ... */
6976 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
6977 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
6980 /* overloading involved: all bets are off over literal
6981 * code. Pretend we haven't seen it */
6983 pRExC_state->code_blocks->count -= n;
6987 /* ... or failing that, try "" overload */
6988 while (SvAMAGIC(msv)
6989 && (sv = AMG_CALLunary(msv, string_amg))
6993 && SvRV(msv) == SvRV(sv))
6998 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
7002 /* this is a partially unrolled
7003 * sv_catsv_nomg(pat, msv);
7004 * that allows us to adjust code block indices if
7007 char *dst = SvPV_force_nomg(pat, dlen);
7009 if (SvUTF8(msv) && !SvUTF8(pat)) {
7010 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
7011 sv_setpvn(pat, dst, dlen);
7014 sv_catsv_nomg(pat, msv);
7018 /* We have only one SV to process, but we need to verify
7019 * it is properly null terminated or we will fail asserts
7020 * later. In theory we probably shouldn't get such SV's,
7021 * but if we do we should handle it gracefully. */
7022 if ( SvTYPE(msv) != SVt_PV || (SvLEN(msv) > SvCUR(msv) && *(SvEND(msv)) == 0) || SvIsCOW_shared_hash(msv) ) {
7023 /* not a string, or a string with a trailing null */
7026 /* a string with no trailing null, we need to copy it
7027 * so it has a trailing null */
7028 pat = sv_2mortal(newSVsv(msv));
7033 pRExC_state->code_blocks->cb[n-1].end = SvCUR(pat)-1;
7036 /* extract any code blocks within any embedded qr//'s */
7037 if (rx && SvTYPE(rx) == SVt_REGEXP
7038 && RX_ENGINE((REGEXP*)rx)->op_comp)
7041 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
7042 if (ri->code_blocks && ri->code_blocks->count) {
7044 /* the presence of an embedded qr// with code means
7045 * we should always recompile: the text of the
7046 * qr// may not have changed, but it may be a
7047 * different closure than last time */
7049 if (pRExC_state->code_blocks) {
7050 int new_count = pRExC_state->code_blocks->count
7051 + ri->code_blocks->count;
7052 Renew(pRExC_state->code_blocks->cb,
7053 new_count, struct reg_code_block);
7054 pRExC_state->code_blocks->count = new_count;
7057 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_
7058 ri->code_blocks->count);
7060 for (i=0; i < ri->code_blocks->count; i++) {
7061 struct reg_code_block *src, *dst;
7062 STRLEN offset = orig_patlen
7063 + ReANY((REGEXP *)rx)->pre_prefix;
7064 assert(n < pRExC_state->code_blocks->count);
7065 src = &ri->code_blocks->cb[i];
7066 dst = &pRExC_state->code_blocks->cb[n];
7067 dst->start = src->start + offset;
7068 dst->end = src->end + offset;
7069 dst->block = src->block;
7070 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
7079 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
7088 /* see if there are any run-time code blocks in the pattern.
7089 * False positives are allowed */
7092 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
7093 char *pat, STRLEN plen)
7098 PERL_UNUSED_CONTEXT;
7100 for (s = 0; s < plen; s++) {
7101 if ( pRExC_state->code_blocks
7102 && n < pRExC_state->code_blocks->count
7103 && s == pRExC_state->code_blocks->cb[n].start)
7105 s = pRExC_state->code_blocks->cb[n].end;
7109 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
7111 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
7113 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
7120 /* Handle run-time code blocks. We will already have compiled any direct
7121 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
7122 * copy of it, but with any literal code blocks blanked out and
7123 * appropriate chars escaped; then feed it into
7125 * eval "qr'modified_pattern'"
7129 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
7133 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
7135 * After eval_sv()-ing that, grab any new code blocks from the returned qr
7136 * and merge them with any code blocks of the original regexp.
7138 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
7139 * instead, just save the qr and return FALSE; this tells our caller that
7140 * the original pattern needs upgrading to utf8.
7144 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
7145 char *pat, STRLEN plen)
7149 DECLARE_AND_GET_RE_DEBUG_FLAGS;
7151 if (pRExC_state->runtime_code_qr) {
7152 /* this is the second time we've been called; this should
7153 * only happen if the main pattern got upgraded to utf8
7154 * during compilation; re-use the qr we compiled first time
7155 * round (which should be utf8 too)
7157 qr = pRExC_state->runtime_code_qr;
7158 pRExC_state->runtime_code_qr = NULL;
7159 assert(RExC_utf8 && SvUTF8(qr));
7165 int newlen = plen + 7; /* allow for "qr''xx\0" extra chars */
7169 /* determine how many extra chars we need for ' and \ escaping */
7170 for (s = 0; s < plen; s++) {
7171 if (pat[s] == '\'' || pat[s] == '\\')
7175 Newx(newpat, newlen, char);
7177 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
7179 for (s = 0; s < plen; s++) {
7180 if ( pRExC_state->code_blocks
7181 && n < pRExC_state->code_blocks->count
7182 && s == pRExC_state->code_blocks->cb[n].start)
7184 /* blank out literal code block so that they aren't
7185 * recompiled: eg change from/to:
7195 assert(pat[s] == '(');
7196 assert(pat[s+1] == '?');
7200 while (s < pRExC_state->code_blocks->cb[n].end) {
7208 if (pat[s] == '\'' || pat[s] == '\\')
7213 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED) {
7215 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED_MORE) {
7221 Perl_re_printf( aTHX_
7222 "%sre-parsing pattern for runtime code:%s %s\n",
7223 PL_colors[4], PL_colors[5], newpat);
7226 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
7232 PUSHSTACKi(PERLSI_REQUIRE);
7233 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
7234 * parsing qr''; normally only q'' does this. It also alters
7236 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
7237 SvREFCNT_dec_NN(sv);
7242 SV * const errsv = ERRSV;
7243 if (SvTRUE_NN(errsv))
7244 /* use croak_sv ? */
7245 Perl_croak_nocontext("%" SVf, SVfARG(errsv));
7247 assert(SvROK(qr_ref));
7249 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
7250 /* the leaving below frees the tmp qr_ref.
7251 * Give qr a life of its own */
7259 if (!RExC_utf8 && SvUTF8(qr)) {
7260 /* first time through; the pattern got upgraded; save the
7261 * qr for the next time through */
7262 assert(!pRExC_state->runtime_code_qr);
7263 pRExC_state->runtime_code_qr = qr;
7268 /* extract any code blocks within the returned qr// */
7271 /* merge the main (r1) and run-time (r2) code blocks into one */
7273 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
7274 struct reg_code_block *new_block, *dst;
7275 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
7279 if (!r2->code_blocks || !r2->code_blocks->count) /* we guessed wrong */
7281 SvREFCNT_dec_NN(qr);
7285 if (!r1->code_blocks)
7286 r1->code_blocks = S_alloc_code_blocks(aTHX_ 0);
7288 r1c = r1->code_blocks->count;
7289 r2c = r2->code_blocks->count;
7291 Newx(new_block, r1c + r2c, struct reg_code_block);
7295 while (i1 < r1c || i2 < r2c) {
7296 struct reg_code_block *src;
7300 src = &r2->code_blocks->cb[i2++];
7304 src = &r1->code_blocks->cb[i1++];
7305 else if ( r1->code_blocks->cb[i1].start
7306 < r2->code_blocks->cb[i2].start)
7308 src = &r1->code_blocks->cb[i1++];
7309 assert(src->end < r2->code_blocks->cb[i2].start);
7312 assert( r1->code_blocks->cb[i1].start
7313 > r2->code_blocks->cb[i2].start);
7314 src = &r2->code_blocks->cb[i2++];
7316 assert(src->end < r1->code_blocks->cb[i1].start);
7319 assert(pat[src->start] == '(');
7320 assert(pat[src->end] == ')');
7321 dst->start = src->start;
7322 dst->end = src->end;
7323 dst->block = src->block;
7324 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
7328 r1->code_blocks->count += r2c;
7329 Safefree(r1->code_blocks->cb);
7330 r1->code_blocks->cb = new_block;
7333 SvREFCNT_dec_NN(qr);
7339 S_setup_longest(pTHX_ RExC_state_t *pRExC_state,
7340 struct reg_substr_datum *rsd,
7341 struct scan_data_substrs *sub,
7342 STRLEN longest_length)
7344 /* This is the common code for setting up the floating and fixed length
7345 * string data extracted from Perl_re_op_compile() below. Returns a boolean
7346 * as to whether succeeded or not */
7350 bool eol = cBOOL(sub->flags & SF_BEFORE_EOL);
7351 bool meol = cBOOL(sub->flags & SF_BEFORE_MEOL);
7353 if (! (longest_length
7354 || (eol /* Can't have SEOL and MULTI */
7355 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
7357 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
7358 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
7363 /* copy the information about the longest from the reg_scan_data
7364 over to the program. */
7365 if (SvUTF8(sub->str)) {
7367 rsd->utf8_substr = sub->str;
7369 rsd->substr = sub->str;
7370 rsd->utf8_substr = NULL;
7372 /* end_shift is how many chars that must be matched that
7373 follow this item. We calculate it ahead of time as once the
7374 lookbehind offset is added in we lose the ability to correctly
7376 ml = sub->minlenp ? *(sub->minlenp) : (SSize_t)longest_length;
7377 rsd->end_shift = ml - sub->min_offset
7379 /* XXX SvTAIL is always false here - did you mean FBMcf_TAIL
7381 + (SvTAIL(sub->str) != 0)
7385 t = (eol/* Can't have SEOL and MULTI */
7386 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
7387 fbm_compile(sub->str, t ? FBMcf_TAIL : 0);
7393 S_set_regex_pv(pTHX_ RExC_state_t *pRExC_state, REGEXP *Rx)
7395 /* Calculates and sets in the compiled pattern 'Rx' the string to compile,
7396 * properly wrapped with the right modifiers */
7398 bool has_p = ((RExC_rx->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
7399 bool has_charset = RExC_utf8 || (get_regex_charset(RExC_rx->extflags)
7400 != REGEX_DEPENDS_CHARSET);
7402 /* The caret is output if there are any defaults: if not all the STD
7403 * flags are set, or if no character set specifier is needed */
7405 (((RExC_rx->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
7407 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
7408 == REG_RUN_ON_COMMENT_SEEN);
7409 U8 reganch = (U8)((RExC_rx->extflags & RXf_PMf_STD_PMMOD)
7410 >> RXf_PMf_STD_PMMOD_SHIFT);
7411 const char *fptr = STD_PAT_MODS; /*"msixxn"*/
7413 STRLEN pat_len = RExC_precomp_end - RExC_precomp;
7415 /* We output all the necessary flags; we never output a minus, as all
7416 * those are defaults, so are
7417 * covered by the caret */
7418 const STRLEN wraplen = pat_len + has_p + has_runon
7419 + has_default /* If needs a caret */
7420 + PL_bitcount[reganch] /* 1 char for each set standard flag */
7422 /* If needs a character set specifier */
7423 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
7424 + (sizeof("(?:)") - 1);
7426 PERL_ARGS_ASSERT_SET_REGEX_PV;
7428 /* make sure PL_bitcount bounds not exceeded */
7429 STATIC_ASSERT_STMT(sizeof(STD_PAT_MODS) <= 8);
7431 p = sv_grow(MUTABLE_SV(Rx), wraplen + 1); /* +1 for the ending NUL */
7434 SvFLAGS(Rx) |= SVf_UTF8;
7437 /* If a default, cover it using the caret */
7439 *p++= DEFAULT_PAT_MOD;
7445 name = get_regex_charset_name(RExC_rx->extflags, &len);
7446 if (strEQ(name, DEPENDS_PAT_MODS)) { /* /d under UTF-8 => /u */
7448 name = UNICODE_PAT_MODS;
7449 len = sizeof(UNICODE_PAT_MODS) - 1;
7451 Copy(name, p, len, char);
7455 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
7458 while((ch = *fptr++)) {
7466 Copy(RExC_precomp, p, pat_len, char);
7467 assert ((RX_WRAPPED(Rx) - p) < 16);
7468 RExC_rx->pre_prefix = p - RX_WRAPPED(Rx);
7471 /* Adding a trailing \n causes this to compile properly:
7472 my $R = qr / A B C # D E/x; /($R)/
7473 Otherwise the parens are considered part of the comment */
7478 SvCUR_set(Rx, p - RX_WRAPPED(Rx));
7482 * Perl_re_op_compile - the perl internal RE engine's function to compile a
7483 * regular expression into internal code.
7484 * The pattern may be passed either as:
7485 * a list of SVs (patternp plus pat_count)
7486 * a list of OPs (expr)
7487 * If both are passed, the SV list is used, but the OP list indicates
7488 * which SVs are actually pre-compiled code blocks
7490 * The SVs in the list have magic and qr overloading applied to them (and
7491 * the list may be modified in-place with replacement SVs in the latter
7494 * If the pattern hasn't changed from old_re, then old_re will be
7497 * eng is the current engine. If that engine has an op_comp method, then
7498 * handle directly (i.e. we assume that op_comp was us); otherwise, just
7499 * do the initial concatenation of arguments and pass on to the external
7502 * If is_bare_re is not null, set it to a boolean indicating whether the
7503 * arg list reduced (after overloading) to a single bare regex which has
7504 * been returned (i.e. /$qr/).
7506 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
7508 * pm_flags contains the PMf_* flags, typically based on those from the
7509 * pm_flags field of the related PMOP. Currently we're only interested in
7510 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL, PMf_WILDCARD.
7512 * For many years this code had an initial sizing pass that calculated
7513 * (sometimes incorrectly, leading to security holes) the size needed for the
7514 * compiled pattern. That was changed by commit
7515 * 7c932d07cab18751bfc7515b4320436273a459e2 in 5.29, which reallocs the size, a
7516 * node at a time, as parsing goes along. Patches welcome to fix any obsolete
7517 * references to this sizing pass.
7519 * Now, an initial crude guess as to the size needed is made, based on the
7520 * length of the pattern. Patches welcome to improve that guess. That amount
7521 * of space is malloc'd and then immediately freed, and then clawed back node
7522 * by node. This design is to minimze, to the extent possible, memory churn
7523 * when doing the reallocs.
7525 * A separate parentheses counting pass may be needed in some cases.
7526 * (Previously the sizing pass did this.) Patches welcome to reduce the number
7529 * The existence of a sizing pass necessitated design decisions that are no
7530 * longer needed. There are potential areas of simplification.
7532 * Beware that the optimization-preparation code in here knows about some
7533 * of the structure of the compiled regexp. [I'll say.]
7537 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
7538 OP *expr, const regexp_engine* eng, REGEXP *old_re,
7539 bool *is_bare_re, const U32 orig_rx_flags, const U32 pm_flags)
7541 REGEXP *Rx; /* Capital 'R' means points to a REGEXP */
7549 SV** new_patternp = patternp;
7551 /* these are all flags - maybe they should be turned
7552 * into a single int with different bit masks */
7553 I32 sawlookahead = 0;
7558 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
7560 bool runtime_code = 0;
7562 RExC_state_t RExC_state;
7563 RExC_state_t * const pRExC_state = &RExC_state;
7564 #ifdef TRIE_STUDY_OPT
7566 RExC_state_t copyRExC_state;
7568 DECLARE_AND_GET_RE_DEBUG_FLAGS;
7570 PERL_ARGS_ASSERT_RE_OP_COMPILE;
7572 DEBUG_r(if (!PL_colorset) reginitcolors());
7575 pRExC_state->warn_text = NULL;
7576 pRExC_state->unlexed_names = NULL;
7577 pRExC_state->code_blocks = NULL;
7580 *is_bare_re = FALSE;
7582 if (expr && (expr->op_type == OP_LIST ||
7583 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
7584 /* allocate code_blocks if needed */
7588 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o))
7589 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
7590 ncode++; /* count of DO blocks */
7593 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_ ncode);
7597 /* compile-time pattern with just OP_CONSTs and DO blocks */
7602 /* find how many CONSTs there are */
7605 if (expr->op_type == OP_CONST)
7608 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
7609 if (o->op_type == OP_CONST)
7613 /* fake up an SV array */
7615 assert(!new_patternp);
7616 Newx(new_patternp, n, SV*);
7617 SAVEFREEPV(new_patternp);
7621 if (expr->op_type == OP_CONST)
7622 new_patternp[n] = cSVOPx_sv(expr);
7624 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
7625 if (o->op_type == OP_CONST)
7626 new_patternp[n++] = cSVOPo_sv;
7631 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7632 "Assembling pattern from %d elements%s\n", pat_count,
7633 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
7635 /* set expr to the first arg op */
7637 if (pRExC_state->code_blocks && pRExC_state->code_blocks->count
7638 && expr->op_type != OP_CONST)
7640 expr = cLISTOPx(expr)->op_first;
7641 assert( expr->op_type == OP_PUSHMARK
7642 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
7643 || expr->op_type == OP_PADRANGE);
7644 expr = OpSIBLING(expr);
7647 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
7648 expr, &recompile, NULL);
7650 /* handle bare (possibly after overloading) regex: foo =~ $re */
7655 if (SvTYPE(re) == SVt_REGEXP) {
7659 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7660 "Precompiled pattern%s\n",
7661 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
7667 exp = SvPV_nomg(pat, plen);
7669 if (!eng->op_comp) {
7670 if ((SvUTF8(pat) && IN_BYTES)
7671 || SvGMAGICAL(pat) || SvAMAGIC(pat))
7673 /* make a temporary copy; either to convert to bytes,
7674 * or to avoid repeating get-magic / overloaded stringify */
7675 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
7676 (IN_BYTES ? 0 : SvUTF8(pat)));
7678 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
7681 /* ignore the utf8ness if the pattern is 0 length */
7682 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
7683 RExC_uni_semantics = 0;
7684 RExC_contains_locale = 0;
7685 RExC_strict = cBOOL(pm_flags & RXf_PMf_STRICT);
7686 RExC_in_script_run = 0;
7687 RExC_study_started = 0;
7688 pRExC_state->runtime_code_qr = NULL;
7689 RExC_frame_head= NULL;
7690 RExC_frame_last= NULL;
7691 RExC_frame_count= 0;
7692 RExC_latest_warn_offset = 0;
7693 RExC_use_BRANCHJ = 0;
7694 RExC_warned_WARN_EXPERIMENTAL__VLB = 0;
7695 RExC_warned_WARN_EXPERIMENTAL__REGEX_SETS = 0;
7696 RExC_total_parens = 0;
7697 RExC_open_parens = NULL;
7698 RExC_close_parens = NULL;
7699 RExC_paren_names = NULL;
7701 RExC_seen_d_op = FALSE;
7703 RExC_paren_name_list = NULL;
7707 RExC_mysv1= sv_newmortal();
7708 RExC_mysv2= sv_newmortal();
7712 SV *dsv= sv_newmortal();
7713 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, PL_dump_re_max_len);
7714 Perl_re_printf( aTHX_ "%sCompiling REx%s %s\n",
7715 PL_colors[4], PL_colors[5], s);
7718 /* we jump here if we have to recompile, e.g., from upgrading the pattern
7721 if ((pm_flags & PMf_USE_RE_EVAL)
7722 /* this second condition covers the non-regex literal case,
7723 * i.e. $foo =~ '(?{})'. */
7724 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
7726 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
7729 /* return old regex if pattern hasn't changed */
7730 /* XXX: note in the below we have to check the flags as well as the
7733 * Things get a touch tricky as we have to compare the utf8 flag
7734 * independently from the compile flags. */
7738 && !!RX_UTF8(old_re) == !!RExC_utf8
7739 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
7740 && RX_PRECOMP(old_re)
7741 && RX_PRELEN(old_re) == plen
7742 && memEQ(RX_PRECOMP(old_re), exp, plen)
7743 && !runtime_code /* with runtime code, always recompile */ )
7746 SV *dsv= sv_newmortal();
7747 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, PL_dump_re_max_len);
7748 Perl_re_printf( aTHX_ "%sSkipping recompilation of unchanged REx%s %s\n",
7749 PL_colors[4], PL_colors[5], s);
7754 /* Allocate the pattern's SV */
7755 RExC_rx_sv = Rx = (REGEXP*) newSV_type(SVt_REGEXP);
7756 RExC_rx = ReANY(Rx);
7757 if ( RExC_rx == NULL )
7758 FAIL("Regexp out of space");
7760 rx_flags = orig_rx_flags;
7762 if ( toUSE_UNI_CHARSET_NOT_DEPENDS
7763 && initial_charset == REGEX_DEPENDS_CHARSET)
7766 /* Set to use unicode semantics if the pattern is in utf8 and has the
7767 * 'depends' charset specified, as it means unicode when utf8 */
7768 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
7769 RExC_uni_semantics = 1;
7772 RExC_pm_flags = pm_flags;
7775 assert(TAINTING_get || !TAINT_get);
7777 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
7779 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
7780 /* whoops, we have a non-utf8 pattern, whilst run-time code
7781 * got compiled as utf8. Try again with a utf8 pattern */
7782 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7783 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7787 assert(!pRExC_state->runtime_code_qr);
7793 RExC_in_lookaround = 0;
7794 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
7795 RExC_recode_x_to_native = 0;
7796 RExC_in_multi_char_class = 0;
7798 RExC_start = RExC_copy_start_in_constructed = RExC_copy_start_in_input = RExC_precomp = exp;
7799 RExC_precomp_end = RExC_end = exp + plen;
7801 RExC_whilem_seen = 0;
7803 RExC_recurse = NULL;
7804 RExC_study_chunk_recursed = NULL;
7805 RExC_study_chunk_recursed_bytes= 0;
7806 RExC_recurse_count = 0;
7807 RExC_sets_depth = 0;
7808 pRExC_state->code_index = 0;
7810 /* Initialize the string in the compiled pattern. This is so that there is
7811 * something to output if necessary */
7812 set_regex_pv(pRExC_state, Rx);
7815 Perl_re_printf( aTHX_
7816 "Starting parse and generation\n");
7818 RExC_lastparse=NULL;
7821 /* Allocate space and zero-initialize. Note, the two step process
7822 of zeroing when in debug mode, thus anything assigned has to
7823 happen after that */
7826 /* On the first pass of the parse, we guess how big this will be. Then
7827 * we grow in one operation to that amount and then give it back. As
7828 * we go along, we re-allocate what we need.
7830 * XXX Currently the guess is essentially that the pattern will be an
7831 * EXACT node with one byte input, one byte output. This is crude, and
7832 * better heuristics are welcome.
7834 * On any subsequent passes, we guess what we actually computed in the
7835 * latest earlier pass. Such a pass probably didn't complete so is
7836 * missing stuff. We could improve those guesses by knowing where the
7837 * parse stopped, and use the length so far plus apply the above
7838 * assumption to what's left. */
7839 RExC_size = STR_SZ(RExC_end - RExC_start);
7842 Newxc(RExC_rxi, sizeof(regexp_internal) + RExC_size, char, regexp_internal);
7843 if ( RExC_rxi == NULL )
7844 FAIL("Regexp out of space");
7846 Zero(RExC_rxi, sizeof(regexp_internal) + RExC_size, char);
7847 RXi_SET( RExC_rx, RExC_rxi );
7849 /* We start from 0 (over from 0 in the case this is a reparse. The first
7850 * node parsed will give back any excess memory we have allocated so far).
7854 /* non-zero initialization begins here */
7855 RExC_rx->engine= eng;
7856 RExC_rx->extflags = rx_flags;
7857 RXp_COMPFLAGS(RExC_rx) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
7859 if (pm_flags & PMf_IS_QR) {
7860 RExC_rxi->code_blocks = pRExC_state->code_blocks;
7861 if (RExC_rxi->code_blocks) {
7862 RExC_rxi->code_blocks->refcnt++;
7866 RExC_rx->intflags = 0;
7868 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
7871 /* This NUL is guaranteed because the pattern comes from an SV*, and the sv
7872 * code makes sure the final byte is an uncounted NUL. But should this
7873 * ever not be the case, lots of things could read beyond the end of the
7874 * buffer: loops like
7875 * while(isFOO(*RExC_parse)) RExC_parse++;
7876 * strchr(RExC_parse, "foo");
7877 * etc. So it is worth noting. */
7878 assert(*RExC_end == '\0');
7882 RExC_parens_buf_size = 0;
7883 RExC_emit_start = RExC_rxi->program;
7884 pRExC_state->code_index = 0;
7886 *((char*) RExC_emit_start) = (char) REG_MAGIC;
7890 if (reg(pRExC_state, 0, &flags, 1)) {
7892 /* Success!, But we may need to redo the parse knowing how many parens
7893 * there actually are */
7894 if (IN_PARENS_PASS) {
7895 flags |= RESTART_PARSE;
7898 /* We have that number in RExC_npar */
7899 RExC_total_parens = RExC_npar;
7901 else if (! MUST_RESTART(flags)) {
7903 Perl_croak(aTHX_ "panic: reg returned failure to re_op_compile, flags=%#" UVxf, (UV) flags);
7906 /* Here, we either have success, or we have to redo the parse for some reason */
7907 if (MUST_RESTART(flags)) {
7909 /* It's possible to write a regexp in ascii that represents Unicode
7910 codepoints outside of the byte range, such as via \x{100}. If we
7911 detect such a sequence we have to convert the entire pattern to utf8
7912 and then recompile, as our sizing calculation will have been based
7913 on 1 byte == 1 character, but we will need to use utf8 to encode
7914 at least some part of the pattern, and therefore must convert the whole
7917 if (flags & NEED_UTF8) {
7919 /* We have stored the offset of the final warning output so far.
7920 * That must be adjusted. Any variant characters between the start
7921 * of the pattern and this warning count for 2 bytes in the final,
7922 * so just add them again */
7923 if (UNLIKELY(RExC_latest_warn_offset > 0)) {
7924 RExC_latest_warn_offset +=
7925 variant_under_utf8_count((U8 *) exp, (U8 *) exp
7926 + RExC_latest_warn_offset);
7928 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7929 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7930 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "Need to redo parse after upgrade\n"));
7933 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "Need to redo parse\n"));
7936 if (ALL_PARENS_COUNTED) {
7937 /* Make enough room for all the known parens, and zero it */
7938 Renew(RExC_open_parens, RExC_total_parens, regnode_offset);
7939 Zero(RExC_open_parens, RExC_total_parens, regnode_offset);
7940 RExC_open_parens[0] = 1; /* +1 for REG_MAGIC */
7942 Renew(RExC_close_parens, RExC_total_parens, regnode_offset);
7943 Zero(RExC_close_parens, RExC_total_parens, regnode_offset);
7945 else { /* Parse did not complete. Reinitialize the parentheses
7947 RExC_total_parens = 0;
7948 if (RExC_open_parens) {
7949 Safefree(RExC_open_parens);
7950 RExC_open_parens = NULL;
7952 if (RExC_close_parens) {
7953 Safefree(RExC_close_parens);
7954 RExC_close_parens = NULL;
7958 /* Clean up what we did in this parse */
7959 SvREFCNT_dec_NN(RExC_rx_sv);
7964 /* Here, we have successfully parsed and generated the pattern's program
7965 * for the regex engine. We are ready to finish things up and look for
7968 /* Update the string to compile, with correct modifiers, etc */
7969 set_regex_pv(pRExC_state, Rx);
7971 RExC_rx->nparens = RExC_total_parens - 1;
7973 /* Uses the upper 4 bits of the FLAGS field, so keep within that size */
7974 if (RExC_whilem_seen > 15)
7975 RExC_whilem_seen = 15;
7978 Perl_re_printf( aTHX_
7979 "Required size %" IVdf " nodes\n", (IV)RExC_size);
7981 RExC_lastparse=NULL;
7984 #ifdef RE_TRACK_PATTERN_OFFSETS
7985 DEBUG_OFFSETS_r(Perl_re_printf( aTHX_
7986 "%s %" UVuf " bytes for offset annotations.\n",
7987 RExC_offsets ? "Got" : "Couldn't get",
7988 (UV)((RExC_offsets[0] * 2 + 1))));
7989 DEBUG_OFFSETS_r(if (RExC_offsets) {
7990 const STRLEN len = RExC_offsets[0];
7992 DECLARE_AND_GET_RE_DEBUG_FLAGS;
7993 Perl_re_printf( aTHX_
7994 "Offsets: [%" UVuf "]\n\t", (UV)RExC_offsets[0]);
7995 for (i = 1; i <= len; i++) {
7996 if (RExC_offsets[i*2-1] || RExC_offsets[i*2])
7997 Perl_re_printf( aTHX_ "%" UVuf ":%" UVuf "[%" UVuf "] ",
7998 (UV)i, (UV)RExC_offsets[i*2-1], (UV)RExC_offsets[i*2]);
8000 Perl_re_printf( aTHX_ "\n");
8004 SetProgLen(RExC_rxi,RExC_size);
8007 DEBUG_DUMP_PRE_OPTIMIZE_r({
8008 SV * const sv = sv_newmortal();
8009 RXi_GET_DECL(RExC_rx, ri);
8011 Perl_re_printf( aTHX_ "Program before optimization:\n");
8013 (void)dumpuntil(RExC_rx, ri->program, ri->program + 1, NULL, NULL,
8018 Perl_re_printf( aTHX_ "Starting post parse optimization\n");
8021 /* XXXX To minimize changes to RE engine we always allocate
8022 3-units-long substrs field. */
8023 Newx(RExC_rx->substrs, 1, struct reg_substr_data);
8024 if (RExC_recurse_count) {
8025 Newx(RExC_recurse, RExC_recurse_count, regnode *);
8026 SAVEFREEPV(RExC_recurse);
8029 if (RExC_seen & REG_RECURSE_SEEN) {
8030 /* Note, RExC_total_parens is 1 + the number of parens in a pattern.
8031 * So its 1 if there are no parens. */
8032 RExC_study_chunk_recursed_bytes= (RExC_total_parens >> 3) +
8033 ((RExC_total_parens & 0x07) != 0);
8034 Newx(RExC_study_chunk_recursed,
8035 RExC_study_chunk_recursed_bytes * RExC_total_parens, U8);
8036 SAVEFREEPV(RExC_study_chunk_recursed);
8040 RExC_rx->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
8042 RExC_study_chunk_recursed_count= 0;
8044 Zero(RExC_rx->substrs, 1, struct reg_substr_data);
8045 if (RExC_study_chunk_recursed) {
8046 Zero(RExC_study_chunk_recursed,
8047 RExC_study_chunk_recursed_bytes * RExC_total_parens, U8);
8051 #ifdef TRIE_STUDY_OPT
8053 StructCopy(&zero_scan_data, &data, scan_data_t);
8054 copyRExC_state = RExC_state;
8057 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "Restudying\n"));
8059 RExC_state = copyRExC_state;
8060 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
8061 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
8063 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
8064 StructCopy(&zero_scan_data, &data, scan_data_t);
8067 StructCopy(&zero_scan_data, &data, scan_data_t);
8070 /* Dig out information for optimizations. */
8071 RExC_rx->extflags = RExC_flags; /* was pm_op */
8072 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
8075 SvUTF8_on(Rx); /* Unicode in it? */
8076 RExC_rxi->regstclass = NULL;
8077 if (RExC_naughty >= TOO_NAUGHTY) /* Probably an expensive pattern. */
8078 RExC_rx->intflags |= PREGf_NAUGHTY;
8079 scan = RExC_rxi->program + 1; /* First BRANCH. */
8081 /* testing for BRANCH here tells us whether there is "must appear"
8082 data in the pattern. If there is then we can use it for optimisations */
8083 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
8086 STRLEN longest_length[2];
8087 regnode_ssc ch_class; /* pointed to by data */
8089 SSize_t last_close = 0; /* pointed to by data */
8090 regnode *first= scan;
8091 regnode *first_next= regnext(first);
8095 * Skip introductions and multiplicators >= 1
8096 * so that we can extract the 'meat' of the pattern that must
8097 * match in the large if() sequence following.
8098 * NOTE that EXACT is NOT covered here, as it is normally
8099 * picked up by the optimiser separately.
8101 * This is unfortunate as the optimiser isnt handling lookahead
8102 * properly currently.
8105 while ((OP(first) == OPEN && (sawopen = 1)) ||
8106 /* An OR of *one* alternative - should not happen now. */
8107 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
8108 /* for now we can't handle lookbehind IFMATCH*/
8109 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
8110 (OP(first) == PLUS) ||
8111 (OP(first) == MINMOD) ||
8112 /* An {n,m} with n>0 */
8113 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
8114 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
8117 * the only op that could be a regnode is PLUS, all the rest
8118 * will be regnode_1 or regnode_2.
8120 * (yves doesn't think this is true)
8122 if (OP(first) == PLUS)
8125 if (OP(first) == MINMOD)
8127 first += regarglen[OP(first)];
8129 first = NEXTOPER(first);
8130 first_next= regnext(first);
8133 /* Starting-point info. */
8135 DEBUG_PEEP("first:", first, 0, 0);
8136 /* Ignore EXACT as we deal with it later. */
8137 if (PL_regkind[OP(first)] == EXACT) {
8138 if (! isEXACTFish(OP(first))) {
8139 NOOP; /* Empty, get anchored substr later. */
8142 RExC_rxi->regstclass = first;
8145 else if (PL_regkind[OP(first)] == TRIE &&
8146 ((reg_trie_data *)RExC_rxi->data->data[ ARG(first) ])->minlen>0)
8148 /* this can happen only on restudy */
8149 RExC_rxi->regstclass = construct_ahocorasick_from_trie(pRExC_state, (regnode *)first, 0);
8152 else if (REGNODE_SIMPLE(OP(first)))
8153 RExC_rxi->regstclass = first;
8154 else if (PL_regkind[OP(first)] == BOUND ||
8155 PL_regkind[OP(first)] == NBOUND)
8156 RExC_rxi->regstclass = first;
8157 else if (PL_regkind[OP(first)] == BOL) {
8158 RExC_rx->intflags |= (OP(first) == MBOL
8161 first = NEXTOPER(first);
8164 else if (OP(first) == GPOS) {
8165 RExC_rx->intflags |= PREGf_ANCH_GPOS;
8166 first = NEXTOPER(first);
8169 else if ((!sawopen || !RExC_sawback) &&
8171 (OP(first) == STAR &&
8172 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
8173 !(RExC_rx->intflags & PREGf_ANCH) && !pRExC_state->code_blocks)
8175 /* turn .* into ^.* with an implied $*=1 */
8177 (OP(NEXTOPER(first)) == REG_ANY)
8180 RExC_rx->intflags |= (type | PREGf_IMPLICIT);
8181 first = NEXTOPER(first);
8184 if (sawplus && !sawminmod && !sawlookahead
8185 && (!sawopen || !RExC_sawback)
8186 && !pRExC_state->code_blocks) /* May examine pos and $& */
8187 /* x+ must match at the 1st pos of run of x's */
8188 RExC_rx->intflags |= PREGf_SKIP;
8190 /* Scan is after the zeroth branch, first is atomic matcher. */
8191 #ifdef TRIE_STUDY_OPT
8194 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
8195 (IV)(first - scan + 1))
8199 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
8200 (IV)(first - scan + 1))
8206 * If there's something expensive in the r.e., find the
8207 * longest literal string that must appear and make it the
8208 * regmust. Resolve ties in favor of later strings, since
8209 * the regstart check works with the beginning of the r.e.
8210 * and avoiding duplication strengthens checking. Not a
8211 * strong reason, but sufficient in the absence of others.
8212 * [Now we resolve ties in favor of the earlier string if
8213 * it happens that c_offset_min has been invalidated, since the
8214 * earlier string may buy us something the later one won't.]
8217 data.substrs[0].str = newSVpvs("");
8218 data.substrs[1].str = newSVpvs("");
8219 data.last_found = newSVpvs("");
8220 data.cur_is_floating = 0; /* initially any found substring is fixed */
8221 ENTER_with_name("study_chunk");
8222 SAVEFREESV(data.substrs[0].str);
8223 SAVEFREESV(data.substrs[1].str);
8224 SAVEFREESV(data.last_found);
8226 if (!RExC_rxi->regstclass) {
8227 ssc_init(pRExC_state, &ch_class);
8228 data.start_class = &ch_class;
8229 stclass_flag = SCF_DO_STCLASS_AND;
8230 } else /* XXXX Check for BOUND? */
8232 data.last_closep = &last_close;
8236 * MAIN ENTRY FOR study_chunk() FOR m/PATTERN/
8237 * (NO top level branches)
8239 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
8240 scan + RExC_size, /* Up to end */
8242 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
8243 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
8247 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
8250 if ( RExC_total_parens == 1 && !data.cur_is_floating
8251 && data.last_start_min == 0 && data.last_end > 0
8252 && !RExC_seen_zerolen
8253 && !(RExC_seen & REG_VERBARG_SEEN)
8254 && !(RExC_seen & REG_GPOS_SEEN)
8256 RExC_rx->extflags |= RXf_CHECK_ALL;
8258 scan_commit(pRExC_state, &data,&minlen, 0);
8261 /* XXX this is done in reverse order because that's the way the
8262 * code was before it was parameterised. Don't know whether it
8263 * actually needs doing in reverse order. DAPM */
8264 for (i = 1; i >= 0; i--) {
8265 longest_length[i] = CHR_SVLEN(data.substrs[i].str);
8268 && SvCUR(data.substrs[0].str) /* ok to leave SvCUR */
8269 && data.substrs[0].min_offset
8270 == data.substrs[1].min_offset
8271 && SvCUR(data.substrs[0].str)
8272 == SvCUR(data.substrs[1].str)
8274 && S_setup_longest (aTHX_ pRExC_state,
8275 &(RExC_rx->substrs->data[i]),
8279 RExC_rx->substrs->data[i].min_offset =
8280 data.substrs[i].min_offset - data.substrs[i].lookbehind;
8282 RExC_rx->substrs->data[i].max_offset = data.substrs[i].max_offset;
8283 /* Don't offset infinity */
8284 if (data.substrs[i].max_offset < OPTIMIZE_INFTY)
8285 RExC_rx->substrs->data[i].max_offset -= data.substrs[i].lookbehind;
8286 SvREFCNT_inc_simple_void_NN(data.substrs[i].str);
8289 RExC_rx->substrs->data[i].substr = NULL;
8290 RExC_rx->substrs->data[i].utf8_substr = NULL;
8291 longest_length[i] = 0;
8295 LEAVE_with_name("study_chunk");
8297 if (RExC_rxi->regstclass
8298 && (OP(RExC_rxi->regstclass) == REG_ANY || OP(RExC_rxi->regstclass) == SANY))
8299 RExC_rxi->regstclass = NULL;
8301 if ((!(RExC_rx->substrs->data[0].substr || RExC_rx->substrs->data[0].utf8_substr)
8302 || RExC_rx->substrs->data[0].min_offset)
8304 && ! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
8305 && is_ssc_worth_it(pRExC_state, data.start_class))
8307 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
8309 ssc_finalize(pRExC_state, data.start_class);
8311 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
8312 StructCopy(data.start_class,
8313 (regnode_ssc*)RExC_rxi->data->data[n],
8315 RExC_rxi->regstclass = (regnode*)RExC_rxi->data->data[n];
8316 RExC_rx->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
8317 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
8318 regprop(RExC_rx, sv, (regnode*)data.start_class, NULL, pRExC_state);
8319 Perl_re_printf( aTHX_
8320 "synthetic stclass \"%s\".\n",
8321 SvPVX_const(sv));});
8322 data.start_class = NULL;
8325 /* A temporary algorithm prefers floated substr to fixed one of
8326 * same length to dig more info. */
8327 i = (longest_length[0] <= longest_length[1]);
8328 RExC_rx->substrs->check_ix = i;
8329 RExC_rx->check_end_shift = RExC_rx->substrs->data[i].end_shift;
8330 RExC_rx->check_substr = RExC_rx->substrs->data[i].substr;
8331 RExC_rx->check_utf8 = RExC_rx->substrs->data[i].utf8_substr;
8332 RExC_rx->check_offset_min = RExC_rx->substrs->data[i].min_offset;
8333 RExC_rx->check_offset_max = RExC_rx->substrs->data[i].max_offset;
8334 if (!i && (RExC_rx->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS)))
8335 RExC_rx->intflags |= PREGf_NOSCAN;
8337 if ((RExC_rx->check_substr || RExC_rx->check_utf8) ) {
8338 RExC_rx->extflags |= RXf_USE_INTUIT;
8339 if (SvTAIL(RExC_rx->check_substr ? RExC_rx->check_substr : RExC_rx->check_utf8))
8340 RExC_rx->extflags |= RXf_INTUIT_TAIL;
8343 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
8344 if ( (STRLEN)minlen < longest_length[1] )
8345 minlen= longest_length[1];
8346 if ( (STRLEN)minlen < longest_length[0] )
8347 minlen= longest_length[0];
8351 /* Several toplevels. Best we can is to set minlen. */
8353 regnode_ssc ch_class;
8354 SSize_t last_close = 0;
8356 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "\nMulti Top Level\n"));
8358 scan = RExC_rxi->program + 1;
8359 ssc_init(pRExC_state, &ch_class);
8360 data.start_class = &ch_class;
8361 data.last_closep = &last_close;
8365 * MAIN ENTRY FOR study_chunk() FOR m/P1|P2|.../
8366 * (patterns WITH top level branches)
8368 minlen = study_chunk(pRExC_state,
8369 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
8370 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
8371 ? SCF_TRIE_DOING_RESTUDY
8375 CHECK_RESTUDY_GOTO_butfirst(NOOP);
8377 RExC_rx->check_substr = NULL;
8378 RExC_rx->check_utf8 = NULL;
8379 RExC_rx->substrs->data[0].substr = NULL;
8380 RExC_rx->substrs->data[0].utf8_substr = NULL;
8381 RExC_rx->substrs->data[1].substr = NULL;
8382 RExC_rx->substrs->data[1].utf8_substr = NULL;
8384 if (! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
8385 && is_ssc_worth_it(pRExC_state, data.start_class))
8387 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
8389 ssc_finalize(pRExC_state, data.start_class);
8391 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
8392 StructCopy(data.start_class,
8393 (regnode_ssc*)RExC_rxi->data->data[n],
8395 RExC_rxi->regstclass = (regnode*)RExC_rxi->data->data[n];
8396 RExC_rx->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
8397 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
8398 regprop(RExC_rx, sv, (regnode*)data.start_class, NULL, pRExC_state);
8399 Perl_re_printf( aTHX_
8400 "synthetic stclass \"%s\".\n",
8401 SvPVX_const(sv));});
8402 data.start_class = NULL;
8406 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
8407 RExC_rx->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
8408 RExC_rx->maxlen = REG_INFTY;
8411 RExC_rx->maxlen = RExC_maxlen;
8414 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
8415 the "real" pattern. */
8417 Perl_re_printf( aTHX_ "minlen: %" IVdf " RExC_rx->minlen:%" IVdf " maxlen:%" IVdf "\n",
8418 (IV)minlen, (IV)RExC_rx->minlen, (IV)RExC_maxlen);
8420 RExC_rx->minlenret = minlen;
8421 if (RExC_rx->minlen < minlen)
8422 RExC_rx->minlen = minlen;
8424 if (RExC_seen & REG_RECURSE_SEEN ) {
8425 RExC_rx->intflags |= PREGf_RECURSE_SEEN;
8426 Newx(RExC_rx->recurse_locinput, RExC_rx->nparens + 1, char *);
8428 if (RExC_seen & REG_GPOS_SEEN)
8429 RExC_rx->intflags |= PREGf_GPOS_SEEN;
8430 if (RExC_seen & REG_LOOKBEHIND_SEEN)
8431 RExC_rx->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
8433 if (pRExC_state->code_blocks)
8434 RExC_rx->extflags |= RXf_EVAL_SEEN;
8435 if (RExC_seen & REG_VERBARG_SEEN)
8437 RExC_rx->intflags |= PREGf_VERBARG_SEEN;
8438 RExC_rx->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
8440 if (RExC_seen & REG_CUTGROUP_SEEN)
8441 RExC_rx->intflags |= PREGf_CUTGROUP_SEEN;
8442 if (pm_flags & PMf_USE_RE_EVAL)
8443 RExC_rx->intflags |= PREGf_USE_RE_EVAL;
8444 if (RExC_paren_names)
8445 RXp_PAREN_NAMES(RExC_rx) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
8447 RXp_PAREN_NAMES(RExC_rx) = NULL;
8449 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
8450 * so it can be used in pp.c */
8451 if (RExC_rx->intflags & PREGf_ANCH)
8452 RExC_rx->extflags |= RXf_IS_ANCHORED;
8456 /* this is used to identify "special" patterns that might result
8457 * in Perl NOT calling the regex engine and instead doing the match "itself",
8458 * particularly special cases in split//. By having the regex compiler
8459 * do this pattern matching at a regop level (instead of by inspecting the pattern)
8460 * we avoid weird issues with equivalent patterns resulting in different behavior,
8461 * AND we allow non Perl engines to get the same optimizations by the setting the
8462 * flags appropriately - Yves */
8463 regnode *first = RExC_rxi->program + 1;
8465 regnode *next = NEXTOPER(first);
8466 /* It's safe to read through *next only if OP(first) is a regop of
8467 * the right type (not EXACT, for example).
8469 U8 nop = (fop == NOTHING || fop == MBOL || fop == SBOL || fop == PLUS)
8472 if (PL_regkind[fop] == NOTHING && nop == END)
8473 RExC_rx->extflags |= RXf_NULL;
8474 else if ((fop == MBOL || (fop == SBOL && !first->flags)) && nop == END)
8475 /* when fop is SBOL first->flags will be true only when it was
8476 * produced by parsing /\A/, and not when parsing /^/. This is
8477 * very important for the split code as there we want to
8478 * treat /^/ as /^/m, but we do not want to treat /\A/ as /^/m.
8479 * See rt #122761 for more details. -- Yves */
8480 RExC_rx->extflags |= RXf_START_ONLY;
8481 else if (fop == PLUS
8482 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
8483 && OP(regnext(first)) == END)
8484 RExC_rx->extflags |= RXf_WHITE;
8485 else if ( RExC_rx->extflags & RXf_SPLIT
8486 && (PL_regkind[fop] == EXACT && ! isEXACTFish(fop))
8487 && STR_LEN(first) == 1
8488 && *(STRING(first)) == ' '
8489 && OP(regnext(first)) == END )
8490 RExC_rx->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
8494 if (RExC_contains_locale) {
8495 RXp_EXTFLAGS(RExC_rx) |= RXf_TAINTED;
8499 if (RExC_paren_names) {
8500 RExC_rxi->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
8501 RExC_rxi->data->data[RExC_rxi->name_list_idx]
8502 = (void*)SvREFCNT_inc(RExC_paren_name_list);
8505 RExC_rxi->name_list_idx = 0;
8507 while ( RExC_recurse_count > 0 ) {
8508 const regnode *scan = RExC_recurse[ --RExC_recurse_count ];
8510 * This data structure is set up in study_chunk() and is used
8511 * to calculate the distance between a GOSUB regopcode and
8512 * the OPEN/CURLYM (CURLYM's are special and can act like OPEN's)
8515 * If for some reason someone writes code that optimises
8516 * away a GOSUB opcode then the assert should be changed to
8517 * an if(scan) to guard the ARG2L_SET() - Yves
8520 assert(scan && OP(scan) == GOSUB);
8521 ARG2L_SET( scan, RExC_open_parens[ARG(scan)] - REGNODE_OFFSET(scan));
8524 Newxz(RExC_rx->offs, RExC_total_parens, regexp_paren_pair);
8525 /* assume we don't need to swap parens around before we match */
8527 Perl_re_printf( aTHX_ "study_chunk_recursed_count: %lu\n",
8528 (unsigned long)RExC_study_chunk_recursed_count);
8532 Perl_re_printf( aTHX_ "Final program:\n");
8536 if (RExC_open_parens) {
8537 Safefree(RExC_open_parens);
8538 RExC_open_parens = NULL;
8540 if (RExC_close_parens) {
8541 Safefree(RExC_close_parens);
8542 RExC_close_parens = NULL;
8546 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
8547 * by setting the regexp SV to readonly-only instead. If the
8548 * pattern's been recompiled, the USEDness should remain. */
8549 if (old_re && SvREADONLY(old_re))
8557 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
8560 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
8562 PERL_UNUSED_ARG(value);
8564 if (flags & RXapif_FETCH) {
8565 return reg_named_buff_fetch(rx, key, flags);
8566 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
8567 Perl_croak_no_modify();
8569 } else if (flags & RXapif_EXISTS) {
8570 return reg_named_buff_exists(rx, key, flags)
8573 } else if (flags & RXapif_REGNAMES) {
8574 return reg_named_buff_all(rx, flags);
8575 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
8576 return reg_named_buff_scalar(rx, flags);
8578 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
8584 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
8587 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
8588 PERL_UNUSED_ARG(lastkey);
8590 if (flags & RXapif_FIRSTKEY)
8591 return reg_named_buff_firstkey(rx, flags);
8592 else if (flags & RXapif_NEXTKEY)
8593 return reg_named_buff_nextkey(rx, flags);
8595 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
8602 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
8606 struct regexp *const rx = ReANY(r);
8608 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
8610 if (rx && RXp_PAREN_NAMES(rx)) {
8611 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
8614 SV* sv_dat=HeVAL(he_str);
8615 I32 *nums=(I32*)SvPVX(sv_dat);
8616 AV * const retarray = (flags & RXapif_ALL) ? newAV() : NULL;
8617 for ( i=0; i<SvIVX(sv_dat); i++ ) {
8618 if ((I32)(rx->nparens) >= nums[i]
8619 && rx->offs[nums[i]].start != -1
8620 && rx->offs[nums[i]].end != -1)
8623 CALLREG_NUMBUF_FETCH(r, nums[i], ret);
8628 ret = newSVsv(&PL_sv_undef);
8631 av_push(retarray, ret);
8634 return newRV_noinc(MUTABLE_SV(retarray));
8641 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
8644 struct regexp *const rx = ReANY(r);
8646 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
8648 if (rx && RXp_PAREN_NAMES(rx)) {
8649 if (flags & RXapif_ALL) {
8650 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
8652 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
8654 SvREFCNT_dec_NN(sv);
8666 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
8668 struct regexp *const rx = ReANY(r);
8670 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
8672 if ( rx && RXp_PAREN_NAMES(rx) ) {
8673 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
8675 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
8682 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
8684 struct regexp *const rx = ReANY(r);
8685 DECLARE_AND_GET_RE_DEBUG_FLAGS;
8687 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
8689 if (rx && RXp_PAREN_NAMES(rx)) {
8690 HV *hv = RXp_PAREN_NAMES(rx);
8692 while ( (temphe = hv_iternext_flags(hv, 0)) ) {
8695 SV* sv_dat = HeVAL(temphe);
8696 I32 *nums = (I32*)SvPVX(sv_dat);
8697 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8698 if ((I32)(rx->lastparen) >= nums[i] &&
8699 rx->offs[nums[i]].start != -1 &&
8700 rx->offs[nums[i]].end != -1)
8706 if (parno || flags & RXapif_ALL) {
8707 return newSVhek(HeKEY_hek(temphe));
8715 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
8720 struct regexp *const rx = ReANY(r);
8722 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
8724 if (rx && RXp_PAREN_NAMES(rx)) {
8725 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
8726 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
8727 } else if (flags & RXapif_ONE) {
8728 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
8729 av = MUTABLE_AV(SvRV(ret));
8730 length = av_count(av);
8731 SvREFCNT_dec_NN(ret);
8732 return newSViv(length);
8734 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
8739 return &PL_sv_undef;
8743 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
8745 struct regexp *const rx = ReANY(r);
8748 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
8750 if (rx && RXp_PAREN_NAMES(rx)) {
8751 HV *hv= RXp_PAREN_NAMES(rx);
8753 (void)hv_iterinit(hv);
8754 while ( (temphe = hv_iternext_flags(hv, 0)) ) {
8757 SV* sv_dat = HeVAL(temphe);
8758 I32 *nums = (I32*)SvPVX(sv_dat);
8759 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8760 if ((I32)(rx->lastparen) >= nums[i] &&
8761 rx->offs[nums[i]].start != -1 &&
8762 rx->offs[nums[i]].end != -1)
8768 if (parno || flags & RXapif_ALL) {
8769 av_push(av, newSVhek(HeKEY_hek(temphe)));
8774 return newRV_noinc(MUTABLE_SV(av));
8778 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
8781 struct regexp *const rx = ReANY(r);
8787 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
8789 if ( n == RX_BUFF_IDX_CARET_PREMATCH
8790 || n == RX_BUFF_IDX_CARET_FULLMATCH
8791 || n == RX_BUFF_IDX_CARET_POSTMATCH
8794 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8796 /* on something like
8799 * the KEEPCOPY is set on the PMOP rather than the regex */
8800 if (PL_curpm && r == PM_GETRE(PL_curpm))
8801 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8810 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
8811 /* no need to distinguish between them any more */
8812 n = RX_BUFF_IDX_FULLMATCH;
8814 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
8815 && rx->offs[0].start != -1)
8817 /* $`, ${^PREMATCH} */
8818 i = rx->offs[0].start;
8822 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
8823 && rx->offs[0].end != -1)
8825 /* $', ${^POSTMATCH} */
8826 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
8827 i = rx->sublen + rx->suboffset - rx->offs[0].end;
8830 if (inRANGE(n, 0, (I32)rx->nparens) &&
8831 (s1 = rx->offs[n].start) != -1 &&
8832 (t1 = rx->offs[n].end) != -1)
8834 /* $&, ${^MATCH}, $1 ... */
8836 s = rx->subbeg + s1 - rx->suboffset;
8841 assert(s >= rx->subbeg);
8842 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
8844 #ifdef NO_TAINT_SUPPORT
8845 sv_setpvn(sv, s, i);
8847 const int oldtainted = TAINT_get;
8849 sv_setpvn(sv, s, i);
8850 TAINT_set(oldtainted);
8852 if (RXp_MATCH_UTF8(rx))
8857 if (RXp_MATCH_TAINTED(rx)) {
8858 if (SvTYPE(sv) >= SVt_PVMG) {
8859 MAGIC* const mg = SvMAGIC(sv);
8862 SvMAGIC_set(sv, mg->mg_moremagic);
8864 if ((mgt = SvMAGIC(sv))) {
8865 mg->mg_moremagic = mgt;
8866 SvMAGIC_set(sv, mg);
8883 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
8884 SV const * const value)
8886 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
8888 PERL_UNUSED_ARG(rx);
8889 PERL_UNUSED_ARG(paren);
8890 PERL_UNUSED_ARG(value);
8893 Perl_croak_no_modify();
8897 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
8900 struct regexp *const rx = ReANY(r);
8904 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
8906 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
8907 || paren == RX_BUFF_IDX_CARET_FULLMATCH
8908 || paren == RX_BUFF_IDX_CARET_POSTMATCH
8911 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8913 /* on something like
8916 * the KEEPCOPY is set on the PMOP rather than the regex */
8917 if (PL_curpm && r == PM_GETRE(PL_curpm))
8918 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8924 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
8926 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
8927 case RX_BUFF_IDX_PREMATCH: /* $` */
8928 if (rx->offs[0].start != -1) {
8929 i = rx->offs[0].start;
8938 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
8939 case RX_BUFF_IDX_POSTMATCH: /* $' */
8940 if (rx->offs[0].end != -1) {
8941 i = rx->sublen - rx->offs[0].end;
8943 s1 = rx->offs[0].end;
8950 default: /* $& / ${^MATCH}, $1, $2, ... */
8951 if (paren <= (I32)rx->nparens &&
8952 (s1 = rx->offs[paren].start) != -1 &&
8953 (t1 = rx->offs[paren].end) != -1)
8959 if (ckWARN(WARN_UNINITIALIZED))
8960 report_uninit((const SV *)sv);
8965 if (i > 0 && RXp_MATCH_UTF8(rx)) {
8966 const char * const s = rx->subbeg - rx->suboffset + s1;
8971 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
8978 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
8980 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
8981 PERL_UNUSED_ARG(rx);
8985 return newSVpvs("Regexp");
8988 /* Scans the name of a named buffer from the pattern.
8989 * If flags is REG_RSN_RETURN_NULL returns null.
8990 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
8991 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
8992 * to the parsed name as looked up in the RExC_paren_names hash.
8993 * If there is an error throws a vFAIL().. type exception.
8996 #define REG_RSN_RETURN_NULL 0
8997 #define REG_RSN_RETURN_NAME 1
8998 #define REG_RSN_RETURN_DATA 2
9001 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
9003 char *name_start = RExC_parse;
9006 PERL_ARGS_ASSERT_REG_SCAN_NAME;
9008 assert (RExC_parse <= RExC_end);
9009 if (RExC_parse == RExC_end) NOOP;
9010 else if (isIDFIRST_lazy_if_safe(RExC_parse, RExC_end, UTF)) {
9011 /* Note that the code here assumes well-formed UTF-8. Skip IDFIRST by
9012 * using do...while */
9015 RExC_parse += UTF8SKIP(RExC_parse);
9016 } while ( RExC_parse < RExC_end
9017 && isWORDCHAR_utf8_safe((U8*)RExC_parse, (U8*) RExC_end));
9021 } while (RExC_parse < RExC_end && isWORDCHAR(*RExC_parse));
9023 RExC_parse++; /* so the <- from the vFAIL is after the offending
9025 vFAIL("Group name must start with a non-digit word character");
9027 sv_name = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
9028 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
9029 if ( flags == REG_RSN_RETURN_NAME)
9031 else if (flags==REG_RSN_RETURN_DATA) {
9034 if ( ! sv_name ) /* should not happen*/
9035 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
9036 if (RExC_paren_names)
9037 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
9039 sv_dat = HeVAL(he_str);
9040 if ( ! sv_dat ) { /* Didn't find group */
9042 /* It might be a forward reference; we can't fail until we
9043 * know, by completing the parse to get all the groups, and
9045 if (ALL_PARENS_COUNTED) {
9046 vFAIL("Reference to nonexistent named group");
9049 REQUIRE_PARENS_PASS;
9055 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
9056 (unsigned long) flags);
9059 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
9060 if (RExC_lastparse!=RExC_parse) { \
9061 Perl_re_printf( aTHX_ "%s", \
9062 Perl_pv_pretty(aTHX_ RExC_mysv1, RExC_parse, \
9063 RExC_end - RExC_parse, 16, \
9065 PERL_PV_ESCAPE_UNI_DETECT | \
9066 PERL_PV_PRETTY_ELLIPSES | \
9067 PERL_PV_PRETTY_LTGT | \
9068 PERL_PV_ESCAPE_RE | \
9069 PERL_PV_PRETTY_EXACTSIZE \
9073 Perl_re_printf( aTHX_ "%16s",""); \
9075 if (RExC_lastnum!=RExC_emit) \
9076 Perl_re_printf( aTHX_ "|%4zu", RExC_emit); \
9078 Perl_re_printf( aTHX_ "|%4s",""); \
9079 Perl_re_printf( aTHX_ "|%*s%-4s", \
9080 (int)((depth*2)), "", \
9083 RExC_lastnum=RExC_emit; \
9084 RExC_lastparse=RExC_parse; \
9089 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
9090 DEBUG_PARSE_MSG((funcname)); \
9091 Perl_re_printf( aTHX_ "%4s","\n"); \
9093 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({\
9094 DEBUG_PARSE_MSG((funcname)); \
9095 Perl_re_printf( aTHX_ fmt "\n",args); \
9098 /* This section of code defines the inversion list object and its methods. The
9099 * interfaces are highly subject to change, so as much as possible is static to
9100 * this file. An inversion list is here implemented as a malloc'd C UV array
9101 * as an SVt_INVLIST scalar.
9103 * An inversion list for Unicode is an array of code points, sorted by ordinal
9104 * number. Each element gives the code point that begins a range that extends
9105 * up-to but not including the code point given by the next element. The final
9106 * element gives the first code point of a range that extends to the platform's
9107 * infinity. The even-numbered elements (invlist[0], invlist[2], invlist[4],
9108 * ...) give ranges whose code points are all in the inversion list. We say
9109 * that those ranges are in the set. The odd-numbered elements give ranges
9110 * whose code points are not in the inversion list, and hence not in the set.
9111 * Thus, element [0] is the first code point in the list. Element [1]
9112 * is the first code point beyond that not in the list; and element [2] is the
9113 * first code point beyond that that is in the list. In other words, the first
9114 * range is invlist[0]..(invlist[1]-1), and all code points in that range are
9115 * in the inversion list. The second range is invlist[1]..(invlist[2]-1), and
9116 * all code points in that range are not in the inversion list. The third
9117 * range invlist[2]..(invlist[3]-1) gives code points that are in the inversion
9118 * list, and so forth. Thus every element whose index is divisible by two
9119 * gives the beginning of a range that is in the list, and every element whose
9120 * index is not divisible by two gives the beginning of a range not in the
9121 * list. If the final element's index is divisible by two, the inversion list
9122 * extends to the platform's infinity; otherwise the highest code point in the
9123 * inversion list is the contents of that element minus 1.
9125 * A range that contains just a single code point N will look like
9127 * invlist[i+1] == N+1
9129 * If N is UV_MAX (the highest representable code point on the machine), N+1 is
9130 * impossible to represent, so element [i+1] is omitted. The single element
9132 * invlist[0] == UV_MAX
9133 * contains just UV_MAX, but is interpreted as matching to infinity.
9135 * Taking the complement (inverting) an inversion list is quite simple, if the
9136 * first element is 0, remove it; otherwise add a 0 element at the beginning.
9137 * This implementation reserves an element at the beginning of each inversion
9138 * list to always contain 0; there is an additional flag in the header which
9139 * indicates if the list begins at the 0, or is offset to begin at the next
9140 * element. This means that the inversion list can be inverted without any
9141 * copying; just flip the flag.
9143 * More about inversion lists can be found in "Unicode Demystified"
9144 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
9146 * The inversion list data structure is currently implemented as an SV pointing
9147 * to an array of UVs that the SV thinks are bytes. This allows us to have an
9148 * array of UV whose memory management is automatically handled by the existing
9149 * facilities for SV's.
9151 * Some of the methods should always be private to the implementation, and some
9152 * should eventually be made public */
9154 /* The header definitions are in F<invlist_inline.h> */
9156 #ifndef PERL_IN_XSUB_RE
9158 PERL_STATIC_INLINE UV*
9159 S__invlist_array_init(SV* const invlist, const bool will_have_0)
9161 /* Returns a pointer to the first element in the inversion list's array.
9162 * This is called upon initialization of an inversion list. Where the
9163 * array begins depends on whether the list has the code point U+0000 in it
9164 * or not. The other parameter tells it whether the code that follows this
9165 * call is about to put a 0 in the inversion list or not. The first
9166 * element is either the element reserved for 0, if TRUE, or the element
9167 * after it, if FALSE */
9169 bool* offset = get_invlist_offset_addr(invlist);
9170 UV* zero_addr = (UV *) SvPVX(invlist);
9172 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
9175 assert(! _invlist_len(invlist));
9179 /* 1^1 = 0; 1^0 = 1 */
9180 *offset = 1 ^ will_have_0;
9181 return zero_addr + *offset;
9185 S_invlist_replace_list_destroys_src(pTHX_ SV * dest, SV * src)
9187 /* Replaces the inversion list in 'dest' with the one from 'src'. It
9188 * steals the list from 'src', so 'src' is made to have a NULL list. This
9189 * is similar to what SvSetMagicSV() would do, if it were implemented on
9190 * inversion lists, though this routine avoids a copy */
9192 const UV src_len = _invlist_len(src);
9193 const bool src_offset = *get_invlist_offset_addr(src);
9194 const STRLEN src_byte_len = SvLEN(src);
9195 char * array = SvPVX(src);
9197 const int oldtainted = TAINT_get;
9199 PERL_ARGS_ASSERT_INVLIST_REPLACE_LIST_DESTROYS_SRC;
9201 assert(is_invlist(src));
9202 assert(is_invlist(dest));
9203 assert(! invlist_is_iterating(src));
9204 assert(SvCUR(src) == 0 || SvCUR(src) < SvLEN(src));
9206 /* Make sure it ends in the right place with a NUL, as our inversion list
9207 * manipulations aren't careful to keep this true, but sv_usepvn_flags()
9209 array[src_byte_len - 1] = '\0';
9211 TAINT_NOT; /* Otherwise it breaks */
9212 sv_usepvn_flags(dest,
9216 /* This flag is documented to cause a copy to be avoided */
9217 SV_HAS_TRAILING_NUL);
9218 TAINT_set(oldtainted);
9223 /* Finish up copying over the other fields in an inversion list */
9224 *get_invlist_offset_addr(dest) = src_offset;
9225 invlist_set_len(dest, src_len, src_offset);
9226 *get_invlist_previous_index_addr(dest) = 0;
9227 invlist_iterfinish(dest);
9230 PERL_STATIC_INLINE IV*
9231 S_get_invlist_previous_index_addr(SV* invlist)
9233 /* Return the address of the IV that is reserved to hold the cached index
9235 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
9237 assert(is_invlist(invlist));
9239 return &(((XINVLIST*) SvANY(invlist))->prev_index);
9242 PERL_STATIC_INLINE IV
9243 S_invlist_previous_index(SV* const invlist)
9245 /* Returns cached index of previous search */
9247 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
9249 return *get_invlist_previous_index_addr(invlist);
9252 PERL_STATIC_INLINE void
9253 S_invlist_set_previous_index(SV* const invlist, const IV index)
9255 /* Caches <index> for later retrieval */
9257 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
9259 assert(index == 0 || index < (int) _invlist_len(invlist));
9261 *get_invlist_previous_index_addr(invlist) = index;
9264 PERL_STATIC_INLINE void
9265 S_invlist_trim(SV* invlist)
9267 /* Free the not currently-being-used space in an inversion list */
9269 /* But don't free up the space needed for the 0 UV that is always at the
9270 * beginning of the list, nor the trailing NUL */
9271 const UV min_size = TO_INTERNAL_SIZE(1) + 1;
9273 PERL_ARGS_ASSERT_INVLIST_TRIM;
9275 assert(is_invlist(invlist));
9277 SvPV_renew(invlist, MAX(min_size, SvCUR(invlist) + 1));
9280 PERL_STATIC_INLINE void
9281 S_invlist_clear(pTHX_ SV* invlist) /* Empty the inversion list */
9283 PERL_ARGS_ASSERT_INVLIST_CLEAR;
9285 assert(is_invlist(invlist));
9287 invlist_set_len(invlist, 0, 0);
9288 invlist_trim(invlist);
9291 #endif /* ifndef PERL_IN_XSUB_RE */
9293 PERL_STATIC_INLINE bool
9294 S_invlist_is_iterating(SV* const invlist)
9296 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
9298 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
9301 #ifndef PERL_IN_XSUB_RE
9303 PERL_STATIC_INLINE UV
9304 S_invlist_max(SV* const invlist)
9306 /* Returns the maximum number of elements storable in the inversion list's
9307 * array, without having to realloc() */
9309 PERL_ARGS_ASSERT_INVLIST_MAX;
9311 assert(is_invlist(invlist));
9313 /* Assumes worst case, in which the 0 element is not counted in the
9314 * inversion list, so subtracts 1 for that */
9315 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
9316 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
9317 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
9321 S_initialize_invlist_guts(pTHX_ SV* invlist, const Size_t initial_size)
9323 PERL_ARGS_ASSERT_INITIALIZE_INVLIST_GUTS;
9325 /* First 1 is in case the zero element isn't in the list; second 1 is for
9327 SvGROW(invlist, TO_INTERNAL_SIZE(initial_size + 1) + 1);
9328 invlist_set_len(invlist, 0, 0);
9330 /* Force iterinit() to be used to get iteration to work */
9331 invlist_iterfinish(invlist);
9333 *get_invlist_previous_index_addr(invlist) = 0;
9334 SvPOK_on(invlist); /* This allows B to extract the PV */
9338 Perl__new_invlist(pTHX_ IV initial_size)
9341 /* Return a pointer to a newly constructed inversion list, with enough
9342 * space to store 'initial_size' elements. If that number is negative, a
9343 * system default is used instead */
9347 if (initial_size < 0) {
9351 new_list = newSV_type(SVt_INVLIST);
9352 initialize_invlist_guts(new_list, initial_size);
9358 Perl__new_invlist_C_array(pTHX_ const UV* const list)
9360 /* Return a pointer to a newly constructed inversion list, initialized to
9361 * point to <list>, which has to be in the exact correct inversion list
9362 * form, including internal fields. Thus this is a dangerous routine that
9363 * should not be used in the wrong hands. The passed in 'list' contains
9364 * several header fields at the beginning that are not part of the
9365 * inversion list body proper */
9367 const STRLEN length = (STRLEN) list[0];
9368 const UV version_id = list[1];
9369 const bool offset = cBOOL(list[2]);
9370 #define HEADER_LENGTH 3
9371 /* If any of the above changes in any way, you must change HEADER_LENGTH
9372 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
9373 * perl -E 'say int(rand 2**31-1)'
9375 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
9376 data structure type, so that one being
9377 passed in can be validated to be an
9378 inversion list of the correct vintage.
9381 SV* invlist = newSV_type(SVt_INVLIST);
9383 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
9385 if (version_id != INVLIST_VERSION_ID) {
9386 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
9389 /* The generated array passed in includes header elements that aren't part
9390 * of the list proper, so start it just after them */
9391 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
9393 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
9394 shouldn't touch it */
9396 *(get_invlist_offset_addr(invlist)) = offset;
9398 /* The 'length' passed to us is the physical number of elements in the
9399 * inversion list. But if there is an offset the logical number is one
9401 invlist_set_len(invlist, length - offset, offset);
9403 invlist_set_previous_index(invlist, 0);
9405 /* Initialize the iteration pointer. */
9406 invlist_iterfinish(invlist);
9408 SvREADONLY_on(invlist);
9415 S__append_range_to_invlist(pTHX_ SV* const invlist,
9416 const UV start, const UV end)
9418 /* Subject to change or removal. Append the range from 'start' to 'end' at
9419 * the end of the inversion list. The range must be above any existing
9423 UV max = invlist_max(invlist);
9424 UV len = _invlist_len(invlist);
9427 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
9429 if (len == 0) { /* Empty lists must be initialized */
9430 offset = start != 0;
9431 array = _invlist_array_init(invlist, ! offset);
9434 /* Here, the existing list is non-empty. The current max entry in the
9435 * list is generally the first value not in the set, except when the
9436 * set extends to the end of permissible values, in which case it is
9437 * the first entry in that final set, and so this call is an attempt to
9438 * append out-of-order */
9440 UV final_element = len - 1;
9441 array = invlist_array(invlist);
9442 if ( array[final_element] > start
9443 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
9445 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",
9446 array[final_element], start,
9447 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
9450 /* Here, it is a legal append. If the new range begins 1 above the end
9451 * of the range below it, it is extending the range below it, so the
9452 * new first value not in the set is one greater than the newly
9453 * extended range. */
9454 offset = *get_invlist_offset_addr(invlist);
9455 if (array[final_element] == start) {
9456 if (end != UV_MAX) {
9457 array[final_element] = end + 1;
9460 /* But if the end is the maximum representable on the machine,
9461 * assume that infinity was actually what was meant. Just let
9462 * the range that this would extend to have no end */
9463 invlist_set_len(invlist, len - 1, offset);
9469 /* Here the new range doesn't extend any existing set. Add it */
9471 len += 2; /* Includes an element each for the start and end of range */
9473 /* If wll overflow the existing space, extend, which may cause the array to
9476 invlist_extend(invlist, len);
9478 /* Have to set len here to avoid assert failure in invlist_array() */
9479 invlist_set_len(invlist, len, offset);
9481 array = invlist_array(invlist);
9484 invlist_set_len(invlist, len, offset);
9487 /* The next item on the list starts the range, the one after that is
9488 * one past the new range. */
9489 array[len - 2] = start;
9490 if (end != UV_MAX) {
9491 array[len - 1] = end + 1;
9494 /* But if the end is the maximum representable on the machine, just let
9495 * the range have no end */
9496 invlist_set_len(invlist, len - 1, offset);
9501 Perl__invlist_search(SV* const invlist, const UV cp)
9503 /* Searches the inversion list for the entry that contains the input code
9504 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
9505 * return value is the index into the list's array of the range that
9506 * contains <cp>, that is, 'i' such that
9507 * array[i] <= cp < array[i+1]
9512 IV high = _invlist_len(invlist);
9513 const IV highest_element = high - 1;
9516 PERL_ARGS_ASSERT__INVLIST_SEARCH;
9518 /* If list is empty, return failure. */
9523 /* (We can't get the array unless we know the list is non-empty) */
9524 array = invlist_array(invlist);
9526 mid = invlist_previous_index(invlist);
9528 if (mid > highest_element) {
9529 mid = highest_element;
9532 /* <mid> contains the cache of the result of the previous call to this
9533 * function (0 the first time). See if this call is for the same result,
9534 * or if it is for mid-1. This is under the theory that calls to this
9535 * function will often be for related code points that are near each other.
9536 * And benchmarks show that caching gives better results. We also test
9537 * here if the code point is within the bounds of the list. These tests
9538 * replace others that would have had to be made anyway to make sure that
9539 * the array bounds were not exceeded, and these give us extra information
9540 * at the same time */
9541 if (cp >= array[mid]) {
9542 if (cp >= array[highest_element]) {
9543 return highest_element;
9546 /* Here, array[mid] <= cp < array[highest_element]. This means that
9547 * the final element is not the answer, so can exclude it; it also
9548 * means that <mid> is not the final element, so can refer to 'mid + 1'
9550 if (cp < array[mid + 1]) {
9556 else { /* cp < aray[mid] */
9557 if (cp < array[0]) { /* Fail if outside the array */
9561 if (cp >= array[mid - 1]) {
9566 /* Binary search. What we are looking for is <i> such that
9567 * array[i] <= cp < array[i+1]
9568 * The loop below converges on the i+1. Note that there may not be an
9569 * (i+1)th element in the array, and things work nonetheless */
9570 while (low < high) {
9571 mid = (low + high) / 2;
9572 assert(mid <= highest_element);
9573 if (array[mid] <= cp) { /* cp >= array[mid] */
9576 /* We could do this extra test to exit the loop early.
9577 if (cp < array[low]) {
9582 else { /* cp < array[mid] */
9589 invlist_set_previous_index(invlist, high);
9594 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9595 const bool complement_b, SV** output)
9597 /* Take the union of two inversion lists and point '*output' to it. On
9598 * input, '*output' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9599 * even 'a' or 'b'). If to an inversion list, the contents of the original
9600 * list will be replaced by the union. The first list, 'a', may be
9601 * NULL, in which case a copy of the second list is placed in '*output'.
9602 * If 'complement_b' is TRUE, the union is taken of the complement
9603 * (inversion) of 'b' instead of b itself.
9605 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9606 * Richard Gillam, published by Addison-Wesley, and explained at some
9607 * length there. The preface says to incorporate its examples into your
9608 * code at your own risk.
9610 * The algorithm is like a merge sort. */
9612 const UV* array_a; /* a's array */
9614 UV len_a; /* length of a's array */
9617 SV* u; /* the resulting union */
9621 UV i_a = 0; /* current index into a's array */
9625 /* running count, as explained in the algorithm source book; items are
9626 * stopped accumulating and are output when the count changes to/from 0.
9627 * The count is incremented when we start a range that's in an input's set,
9628 * and decremented when we start a range that's not in a set. So this
9629 * variable can be 0, 1, or 2. When it is 0 neither input is in their set,
9630 * and hence nothing goes into the union; 1, just one of the inputs is in
9631 * its set (and its current range gets added to the union); and 2 when both
9632 * inputs are in their sets. */
9635 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
9637 assert(*output == NULL || is_invlist(*output));
9639 len_b = _invlist_len(b);
9642 /* Here, 'b' is empty, hence it's complement is all possible code
9643 * points. So if the union includes the complement of 'b', it includes
9644 * everything, and we need not even look at 'a'. It's easiest to
9645 * create a new inversion list that matches everything. */
9647 SV* everything = _add_range_to_invlist(NULL, 0, UV_MAX);
9649 if (*output == NULL) { /* If the output didn't exist, just point it
9651 *output = everything;
9653 else { /* Otherwise, replace its contents with the new list */
9654 invlist_replace_list_destroys_src(*output, everything);
9655 SvREFCNT_dec_NN(everything);
9661 /* Here, we don't want the complement of 'b', and since 'b' is empty,
9662 * the union will come entirely from 'a'. If 'a' is NULL or empty, the
9663 * output will be empty */
9665 if (a == NULL || _invlist_len(a) == 0) {
9666 if (*output == NULL) {
9667 *output = _new_invlist(0);
9670 invlist_clear(*output);
9675 /* Here, 'a' is not empty, but 'b' is, so 'a' entirely determines the
9676 * union. We can just return a copy of 'a' if '*output' doesn't point
9677 * to an existing list */
9678 if (*output == NULL) {
9679 *output = invlist_clone(a, NULL);
9683 /* If the output is to overwrite 'a', we have a no-op, as it's
9689 /* Here, '*output' is to be overwritten by 'a' */
9690 u = invlist_clone(a, NULL);
9691 invlist_replace_list_destroys_src(*output, u);
9697 /* Here 'b' is not empty. See about 'a' */
9699 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
9701 /* Here, 'a' is empty (and b is not). That means the union will come
9702 * entirely from 'b'. If '*output' is NULL, we can directly return a
9703 * clone of 'b'. Otherwise, we replace the contents of '*output' with
9706 SV ** dest = (*output == NULL) ? output : &u;
9707 *dest = invlist_clone(b, NULL);
9709 _invlist_invert(*dest);
9713 invlist_replace_list_destroys_src(*output, u);
9720 /* Here both lists exist and are non-empty */
9721 array_a = invlist_array(a);
9722 array_b = invlist_array(b);
9724 /* If are to take the union of 'a' with the complement of b, set it
9725 * up so are looking at b's complement. */
9728 /* To complement, we invert: if the first element is 0, remove it. To
9729 * do this, we just pretend the array starts one later */
9730 if (array_b[0] == 0) {
9736 /* But if the first element is not zero, we pretend the list starts
9737 * at the 0 that is always stored immediately before the array. */
9743 /* Size the union for the worst case: that the sets are completely
9745 u = _new_invlist(len_a + len_b);
9747 /* Will contain U+0000 if either component does */
9748 array_u = _invlist_array_init(u, ( len_a > 0 && array_a[0] == 0)
9749 || (len_b > 0 && array_b[0] == 0));
9751 /* Go through each input list item by item, stopping when have exhausted
9753 while (i_a < len_a && i_b < len_b) {
9754 UV cp; /* The element to potentially add to the union's array */
9755 bool cp_in_set; /* is it in the input list's set or not */
9757 /* We need to take one or the other of the two inputs for the union.
9758 * Since we are merging two sorted lists, we take the smaller of the
9759 * next items. In case of a tie, we take first the one that is in its
9760 * set. If we first took the one not in its set, it would decrement
9761 * the count, possibly to 0 which would cause it to be output as ending
9762 * the range, and the next time through we would take the same number,
9763 * and output it again as beginning the next range. By doing it the
9764 * opposite way, there is no possibility that the count will be
9765 * momentarily decremented to 0, and thus the two adjoining ranges will
9766 * be seamlessly merged. (In a tie and both are in the set or both not
9767 * in the set, it doesn't matter which we take first.) */
9768 if ( array_a[i_a] < array_b[i_b]
9769 || ( array_a[i_a] == array_b[i_b]
9770 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9772 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9773 cp = array_a[i_a++];
9776 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9777 cp = array_b[i_b++];
9780 /* Here, have chosen which of the two inputs to look at. Only output
9781 * if the running count changes to/from 0, which marks the
9782 * beginning/end of a range that's in the set */
9785 array_u[i_u++] = cp;
9792 array_u[i_u++] = cp;
9798 /* The loop above increments the index into exactly one of the input lists
9799 * each iteration, and ends when either index gets to its list end. That
9800 * means the other index is lower than its end, and so something is
9801 * remaining in that one. We decrement 'count', as explained below, if
9802 * that list is in its set. (i_a and i_b each currently index the element
9803 * beyond the one we care about.) */
9804 if ( (i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9805 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9810 /* Above we decremented 'count' if the list that had unexamined elements in
9811 * it was in its set. This has made it so that 'count' being non-zero
9812 * means there isn't anything left to output; and 'count' equal to 0 means
9813 * that what is left to output is precisely that which is left in the
9814 * non-exhausted input list.
9816 * To see why, note first that the exhausted input obviously has nothing
9817 * left to add to the union. If it was in its set at its end, that means
9818 * the set extends from here to the platform's infinity, and hence so does
9819 * the union and the non-exhausted set is irrelevant. The exhausted set
9820 * also contributed 1 to 'count'. If 'count' was 2, it got decremented to
9821 * 1, but if it was 1, the non-exhausted set wasn't in its set, and so
9822 * 'count' remains at 1. This is consistent with the decremented 'count'
9823 * != 0 meaning there's nothing left to add to the union.
9825 * But if the exhausted input wasn't in its set, it contributed 0 to
9826 * 'count', and the rest of the union will be whatever the other input is.
9827 * If 'count' was 0, neither list was in its set, and 'count' remains 0;
9828 * otherwise it gets decremented to 0. This is consistent with 'count'
9829 * == 0 meaning the remainder of the union is whatever is left in the
9830 * non-exhausted list. */
9835 IV copy_count = len_a - i_a;
9836 if (copy_count > 0) { /* The non-exhausted input is 'a' */
9837 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
9839 else { /* The non-exhausted input is b */
9840 copy_count = len_b - i_b;
9841 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
9843 len_u = i_u + copy_count;
9846 /* Set the result to the final length, which can change the pointer to
9847 * array_u, so re-find it. (Note that it is unlikely that this will
9848 * change, as we are shrinking the space, not enlarging it) */
9849 if (len_u != _invlist_len(u)) {
9850 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
9852 array_u = invlist_array(u);
9855 if (*output == NULL) { /* Simply return the new inversion list */
9859 /* Otherwise, overwrite the inversion list that was in '*output'. We
9860 * could instead free '*output', and then set it to 'u', but experience
9861 * has shown [perl #127392] that if the input is a mortal, we can get a
9862 * huge build-up of these during regex compilation before they get
9864 invlist_replace_list_destroys_src(*output, u);
9872 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9873 const bool complement_b, SV** i)
9875 /* Take the intersection of two inversion lists and point '*i' to it. On
9876 * input, '*i' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9877 * even 'a' or 'b'). If to an inversion list, the contents of the original
9878 * list will be replaced by the intersection. The first list, 'a', may be
9879 * NULL, in which case '*i' will be an empty list. If 'complement_b' is
9880 * TRUE, the result will be the intersection of 'a' and the complement (or
9881 * inversion) of 'b' instead of 'b' directly.
9883 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9884 * Richard Gillam, published by Addison-Wesley, and explained at some
9885 * length there. The preface says to incorporate its examples into your
9886 * code at your own risk. In fact, it had bugs
9888 * The algorithm is like a merge sort, and is essentially the same as the
9892 const UV* array_a; /* a's array */
9894 UV len_a; /* length of a's array */
9897 SV* r; /* the resulting intersection */
9901 UV i_a = 0; /* current index into a's array */
9905 /* running count of how many of the two inputs are postitioned at ranges
9906 * that are in their sets. As explained in the algorithm source book,
9907 * items are stopped accumulating and are output when the count changes
9908 * to/from 2. The count is incremented when we start a range that's in an
9909 * input's set, and decremented when we start a range that's not in a set.
9910 * Only when it is 2 are we in the intersection. */
9913 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
9915 assert(*i == NULL || is_invlist(*i));
9917 /* Special case if either one is empty */
9918 len_a = (a == NULL) ? 0 : _invlist_len(a);
9919 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
9920 if (len_a != 0 && complement_b) {
9922 /* Here, 'a' is not empty, therefore from the enclosing 'if', 'b'
9923 * must be empty. Here, also we are using 'b's complement, which
9924 * hence must be every possible code point. Thus the intersection
9927 if (*i == a) { /* No-op */
9932 *i = invlist_clone(a, NULL);
9936 r = invlist_clone(a, NULL);
9937 invlist_replace_list_destroys_src(*i, r);
9942 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
9943 * intersection must be empty */
9945 *i = _new_invlist(0);
9953 /* Here both lists exist and are non-empty */
9954 array_a = invlist_array(a);
9955 array_b = invlist_array(b);
9957 /* If are to take the intersection of 'a' with the complement of b, set it
9958 * up so are looking at b's complement. */
9961 /* To complement, we invert: if the first element is 0, remove it. To
9962 * do this, we just pretend the array starts one later */
9963 if (array_b[0] == 0) {
9969 /* But if the first element is not zero, we pretend the list starts
9970 * at the 0 that is always stored immediately before the array. */
9976 /* Size the intersection for the worst case: that the intersection ends up
9977 * fragmenting everything to be completely disjoint */
9978 r= _new_invlist(len_a + len_b);
9980 /* Will contain U+0000 iff both components do */
9981 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
9982 && len_b > 0 && array_b[0] == 0);
9984 /* Go through each list item by item, stopping when have exhausted one of
9986 while (i_a < len_a && i_b < len_b) {
9987 UV cp; /* The element to potentially add to the intersection's
9989 bool cp_in_set; /* Is it in the input list's set or not */
9991 /* We need to take one or the other of the two inputs for the
9992 * intersection. Since we are merging two sorted lists, we take the
9993 * smaller of the next items. In case of a tie, we take first the one
9994 * that is not in its set (a difference from the union algorithm). If
9995 * we first took the one in its set, it would increment the count,
9996 * possibly to 2 which would cause it to be output as starting a range
9997 * in the intersection, and the next time through we would take that
9998 * same number, and output it again as ending the set. By doing the
9999 * opposite of this, there is no possibility that the count will be
10000 * momentarily incremented to 2. (In a tie and both are in the set or
10001 * both not in the set, it doesn't matter which we take first.) */
10002 if ( array_a[i_a] < array_b[i_b]
10003 || ( array_a[i_a] == array_b[i_b]
10004 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
10006 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
10007 cp = array_a[i_a++];
10010 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
10011 cp= array_b[i_b++];
10014 /* Here, have chosen which of the two inputs to look at. Only output
10015 * if the running count changes to/from 2, which marks the
10016 * beginning/end of a range that's in the intersection */
10020 array_r[i_r++] = cp;
10025 array_r[i_r++] = cp;
10032 /* The loop above increments the index into exactly one of the input lists
10033 * each iteration, and ends when either index gets to its list end. That
10034 * means the other index is lower than its end, and so something is
10035 * remaining in that one. We increment 'count', as explained below, if the
10036 * exhausted list was in its set. (i_a and i_b each currently index the
10037 * element beyond the one we care about.) */
10038 if ( (i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
10039 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
10044 /* Above we incremented 'count' if the exhausted list was in its set. This
10045 * has made it so that 'count' being below 2 means there is nothing left to
10046 * output; otheriwse what's left to add to the intersection is precisely
10047 * that which is left in the non-exhausted input list.
10049 * To see why, note first that the exhausted input obviously has nothing
10050 * left to affect the intersection. If it was in its set at its end, that
10051 * means the set extends from here to the platform's infinity, and hence
10052 * anything in the non-exhausted's list will be in the intersection, and
10053 * anything not in it won't be. Hence, the rest of the intersection is
10054 * precisely what's in the non-exhausted list The exhausted set also
10055 * contributed 1 to 'count', meaning 'count' was at least 1. Incrementing
10056 * it means 'count' is now at least 2. This is consistent with the
10057 * incremented 'count' being >= 2 means to add the non-exhausted list to
10058 * the intersection.
10060 * But if the exhausted input wasn't in its set, it contributed 0 to
10061 * 'count', and the intersection can't include anything further; the
10062 * non-exhausted set is irrelevant. 'count' was at most 1, and doesn't get
10063 * incremented. This is consistent with 'count' being < 2 meaning nothing
10064 * further to add to the intersection. */
10065 if (count < 2) { /* Nothing left to put in the intersection. */
10068 else { /* copy the non-exhausted list, unchanged. */
10069 IV copy_count = len_a - i_a;
10070 if (copy_count > 0) { /* a is the one with stuff left */
10071 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
10073 else { /* b is the one with stuff left */
10074 copy_count = len_b - i_b;
10075 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
10077 len_r = i_r + copy_count;
10080 /* Set the result to the final length, which can change the pointer to
10081 * array_r, so re-find it. (Note that it is unlikely that this will
10082 * change, as we are shrinking the space, not enlarging it) */
10083 if (len_r != _invlist_len(r)) {
10084 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
10086 array_r = invlist_array(r);
10089 if (*i == NULL) { /* Simply return the calculated intersection */
10092 else { /* Otherwise, replace the existing inversion list in '*i'. We could
10093 instead free '*i', and then set it to 'r', but experience has
10094 shown [perl #127392] that if the input is a mortal, we can get a
10095 huge build-up of these during regex compilation before they get
10098 invlist_replace_list_destroys_src(*i, r);
10103 SvREFCNT_dec_NN(r);
10110 Perl__add_range_to_invlist(pTHX_ SV* invlist, UV start, UV end)
10112 /* Add the range from 'start' to 'end' inclusive to the inversion list's
10113 * set. A pointer to the inversion list is returned. This may actually be
10114 * a new list, in which case the passed in one has been destroyed. The
10115 * passed-in inversion list can be NULL, in which case a new one is created
10116 * with just the one range in it. The new list is not necessarily
10117 * NUL-terminated. Space is not freed if the inversion list shrinks as a
10118 * result of this function. The gain would not be large, and in many
10119 * cases, this is called multiple times on a single inversion list, so
10120 * anything freed may almost immediately be needed again.
10122 * This used to mostly call the 'union' routine, but that is much more
10123 * heavyweight than really needed for a single range addition */
10125 UV* array; /* The array implementing the inversion list */
10126 UV len; /* How many elements in 'array' */
10127 SSize_t i_s; /* index into the invlist array where 'start'
10129 SSize_t i_e = 0; /* And the index where 'end' should go */
10130 UV cur_highest; /* The highest code point in the inversion list
10131 upon entry to this function */
10133 /* This range becomes the whole inversion list if none already existed */
10134 if (invlist == NULL) {
10135 invlist = _new_invlist(2);
10136 _append_range_to_invlist(invlist, start, end);
10140 /* Likewise, if the inversion list is currently empty */
10141 len = _invlist_len(invlist);
10143 _append_range_to_invlist(invlist, start, end);
10147 /* Starting here, we have to know the internals of the list */
10148 array = invlist_array(invlist);
10150 /* If the new range ends higher than the current highest ... */
10151 cur_highest = invlist_highest(invlist);
10152 if (end > cur_highest) {
10154 /* If the whole range is higher, we can just append it */
10155 if (start > cur_highest) {
10156 _append_range_to_invlist(invlist, start, end);
10160 /* Otherwise, add the portion that is higher ... */
10161 _append_range_to_invlist(invlist, cur_highest + 1, end);
10163 /* ... and continue on below to handle the rest. As a result of the
10164 * above append, we know that the index of the end of the range is the
10165 * final even numbered one of the array. Recall that the final element
10166 * always starts a range that extends to infinity. If that range is in
10167 * the set (meaning the set goes from here to infinity), it will be an
10168 * even index, but if it isn't in the set, it's odd, and the final
10169 * range in the set is one less, which is even. */
10170 if (end == UV_MAX) {
10178 /* We have dealt with appending, now see about prepending. If the new
10179 * range starts lower than the current lowest ... */
10180 if (start < array[0]) {
10182 /* Adding something which has 0 in it is somewhat tricky, and uncommon.
10183 * Let the union code handle it, rather than having to know the
10184 * trickiness in two code places. */
10185 if (UNLIKELY(start == 0)) {
10188 range_invlist = _new_invlist(2);
10189 _append_range_to_invlist(range_invlist, start, end);
10191 _invlist_union(invlist, range_invlist, &invlist);
10193 SvREFCNT_dec_NN(range_invlist);
10198 /* If the whole new range comes before the first entry, and doesn't
10199 * extend it, we have to insert it as an additional range */
10200 if (end < array[0] - 1) {
10202 goto splice_in_new_range;
10205 /* Here the new range adjoins the existing first range, extending it
10209 /* And continue on below to handle the rest. We know that the index of
10210 * the beginning of the range is the first one of the array */
10213 else { /* Not prepending any part of the new range to the existing list.
10214 * Find where in the list it should go. This finds i_s, such that:
10215 * invlist[i_s] <= start < array[i_s+1]
10217 i_s = _invlist_search(invlist, start);
10220 /* At this point, any extending before the beginning of the inversion list
10221 * and/or after the end has been done. This has made it so that, in the
10222 * code below, each endpoint of the new range is either in a range that is
10223 * in the set, or is in a gap between two ranges that are. This means we
10224 * don't have to worry about exceeding the array bounds.
10226 * Find where in the list the new range ends (but we can skip this if we
10227 * have already determined what it is, or if it will be the same as i_s,
10228 * which we already have computed) */
10230 i_e = (start == end)
10232 : _invlist_search(invlist, end);
10235 /* Here generally invlist[i_e] <= end < array[i_e+1]. But if invlist[i_e]
10236 * is a range that goes to infinity there is no element at invlist[i_e+1],
10237 * so only the first relation holds. */
10239 if ( ! ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
10241 /* Here, the ranges on either side of the beginning of the new range
10242 * are in the set, and this range starts in the gap between them.
10244 * The new range extends the range above it downwards if the new range
10245 * ends at or above that range's start */
10246 const bool extends_the_range_above = ( end == UV_MAX
10247 || end + 1 >= array[i_s+1]);
10249 /* The new range extends the range below it upwards if it begins just
10250 * after where that range ends */
10251 if (start == array[i_s]) {
10253 /* If the new range fills the entire gap between the other ranges,
10254 * they will get merged together. Other ranges may also get
10255 * merged, depending on how many of them the new range spans. In
10256 * the general case, we do the merge later, just once, after we
10257 * figure out how many to merge. But in the case where the new
10258 * range exactly spans just this one gap (possibly extending into
10259 * the one above), we do the merge here, and an early exit. This
10260 * is done here to avoid having to special case later. */
10261 if (i_e - i_s <= 1) {
10263 /* If i_e - i_s == 1, it means that the new range terminates
10264 * within the range above, and hence 'extends_the_range_above'
10265 * must be true. (If the range above it extends to infinity,
10266 * 'i_s+2' will be above the array's limit, but 'len-i_s-2'
10267 * will be 0, so no harm done.) */
10268 if (extends_the_range_above) {
10269 Move(array + i_s + 2, array + i_s, len - i_s - 2, UV);
10270 invlist_set_len(invlist,
10272 *(get_invlist_offset_addr(invlist)));
10276 /* Here, i_e must == i_s. We keep them in sync, as they apply
10277 * to the same range, and below we are about to decrement i_s
10282 /* Here, the new range is adjacent to the one below. (It may also
10283 * span beyond the range above, but that will get resolved later.)
10284 * Extend the range below to include this one. */
10285 array[i_s] = (end == UV_MAX) ? UV_MAX : end + 1;
10287 start = array[i_s];
10289 else if (extends_the_range_above) {
10291 /* Here the new range only extends the range above it, but not the
10292 * one below. It merges with the one above. Again, we keep i_e
10293 * and i_s in sync if they point to the same range */
10298 array[i_s] = start;
10302 /* Here, we've dealt with the new range start extending any adjoining
10305 * If the new range extends to infinity, it is now the final one,
10306 * regardless of what was there before */
10307 if (UNLIKELY(end == UV_MAX)) {
10308 invlist_set_len(invlist, i_s + 1, *(get_invlist_offset_addr(invlist)));
10312 /* If i_e started as == i_s, it has also been dealt with,
10313 * and been updated to the new i_s, which will fail the following if */
10314 if (! ELEMENT_RANGE_MATCHES_INVLIST(i_e)) {
10316 /* Here, the ranges on either side of the end of the new range are in
10317 * the set, and this range ends in the gap between them.
10319 * If this range is adjacent to (hence extends) the range above it, it
10320 * becomes part of that range; likewise if it extends the range below,
10321 * it becomes part of that range */
10322 if (end + 1 == array[i_e+1]) {
10324 array[i_e] = start;
10326 else if (start <= array[i_e]) {
10327 array[i_e] = end + 1;
10334 /* If the range fits entirely in an existing range (as possibly already
10335 * extended above), it doesn't add anything new */
10336 if (ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
10340 /* Here, no part of the range is in the list. Must add it. It will
10341 * occupy 2 more slots */
10342 splice_in_new_range:
10344 invlist_extend(invlist, len + 2);
10345 array = invlist_array(invlist);
10346 /* Move the rest of the array down two slots. Don't include any
10348 Move(array + i_e + 1, array + i_e + 3, len - i_e - 1, UV);
10350 /* Do the actual splice */
10351 array[i_e+1] = start;
10352 array[i_e+2] = end + 1;
10353 invlist_set_len(invlist, len + 2, *(get_invlist_offset_addr(invlist)));
10357 /* Here the new range crossed the boundaries of a pre-existing range. The
10358 * code above has adjusted things so that both ends are in ranges that are
10359 * in the set. This means everything in between must also be in the set.
10360 * Just squash things together */
10361 Move(array + i_e + 1, array + i_s + 1, len - i_e - 1, UV);
10362 invlist_set_len(invlist,
10364 *(get_invlist_offset_addr(invlist)));
10370 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
10371 UV** other_elements_ptr)
10373 /* Create and return an inversion list whose contents are to be populated
10374 * by the caller. The caller gives the number of elements (in 'size') and
10375 * the very first element ('element0'). This function will set
10376 * '*other_elements_ptr' to an array of UVs, where the remaining elements
10377 * are to be placed.
10379 * Obviously there is some trust involved that the caller will properly
10380 * fill in the other elements of the array.
10382 * (The first element needs to be passed in, as the underlying code does
10383 * things differently depending on whether it is zero or non-zero) */
10385 SV* invlist = _new_invlist(size);
10388 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
10390 invlist = add_cp_to_invlist(invlist, element0);
10391 offset = *get_invlist_offset_addr(invlist);
10393 invlist_set_len(invlist, size, offset);
10394 *other_elements_ptr = invlist_array(invlist) + 1;
10400 #ifndef PERL_IN_XSUB_RE
10402 Perl__invlist_invert(pTHX_ SV* const invlist)
10404 /* Complement the input inversion list. This adds a 0 if the list didn't
10405 * have a zero; removes it otherwise. As described above, the data
10406 * structure is set up so that this is very efficient */
10408 PERL_ARGS_ASSERT__INVLIST_INVERT;
10410 assert(! invlist_is_iterating(invlist));
10412 /* The inverse of matching nothing is matching everything */
10413 if (_invlist_len(invlist) == 0) {
10414 _append_range_to_invlist(invlist, 0, UV_MAX);
10418 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
10422 Perl_invlist_clone(pTHX_ SV* const invlist, SV* new_invlist)
10424 /* Return a new inversion list that is a copy of the input one, which is
10425 * unchanged. The new list will not be mortal even if the old one was. */
10427 const STRLEN nominal_length = _invlist_len(invlist);
10428 const STRLEN physical_length = SvCUR(invlist);
10429 const bool offset = *(get_invlist_offset_addr(invlist));
10431 PERL_ARGS_ASSERT_INVLIST_CLONE;
10433 if (new_invlist == NULL) {
10434 new_invlist = _new_invlist(nominal_length);
10437 sv_upgrade(new_invlist, SVt_INVLIST);
10438 initialize_invlist_guts(new_invlist, nominal_length);
10441 *(get_invlist_offset_addr(new_invlist)) = offset;
10442 invlist_set_len(new_invlist, nominal_length, offset);
10443 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
10445 return new_invlist;
10450 PERL_STATIC_INLINE UV
10451 S_invlist_lowest(SV* const invlist)
10453 /* Returns the lowest code point that matches an inversion list. This API
10454 * has an ambiguity, as it returns 0 under either the lowest is actually
10455 * 0, or if the list is empty. If this distinction matters to you, check
10456 * for emptiness before calling this function */
10458 UV len = _invlist_len(invlist);
10461 PERL_ARGS_ASSERT_INVLIST_LOWEST;
10467 array = invlist_array(invlist);
10473 S_invlist_contents(pTHX_ SV* const invlist, const bool traditional_style)
10475 /* Get the contents of an inversion list into a string SV so that they can
10476 * be printed out. If 'traditional_style' is TRUE, it uses the format
10477 * traditionally done for debug tracing; otherwise it uses a format
10478 * suitable for just copying to the output, with blanks between ranges and
10479 * a dash between range components */
10483 const char intra_range_delimiter = (traditional_style ? '\t' : '-');
10484 const char inter_range_delimiter = (traditional_style ? '\n' : ' ');
10486 if (traditional_style) {
10487 output = newSVpvs("\n");
10490 output = newSVpvs("");
10493 PERL_ARGS_ASSERT_INVLIST_CONTENTS;
10495 assert(! invlist_is_iterating(invlist));
10497 invlist_iterinit(invlist);
10498 while (invlist_iternext(invlist, &start, &end)) {
10499 if (end == UV_MAX) {
10500 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%cINFTY%c",
10501 start, intra_range_delimiter,
10502 inter_range_delimiter);
10504 else if (end != start) {
10505 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c%04" UVXf "%c",
10507 intra_range_delimiter,
10508 end, inter_range_delimiter);
10511 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c",
10512 start, inter_range_delimiter);
10516 if (SvCUR(output) && ! traditional_style) {/* Get rid of trailing blank */
10517 SvCUR_set(output, SvCUR(output) - 1);
10523 #ifndef PERL_IN_XSUB_RE
10525 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
10526 const char * const indent, SV* const invlist)
10528 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
10529 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
10530 * the string 'indent'. The output looks like this:
10531 [0] 0x000A .. 0x000D
10533 [4] 0x2028 .. 0x2029
10534 [6] 0x3104 .. INFTY
10535 * This means that the first range of code points matched by the list are
10536 * 0xA through 0xD; the second range contains only the single code point
10537 * 0x85, etc. An inversion list is an array of UVs. Two array elements
10538 * are used to define each range (except if the final range extends to
10539 * infinity, only a single element is needed). The array index of the
10540 * first element for the corresponding range is given in brackets. */
10545 PERL_ARGS_ASSERT__INVLIST_DUMP;
10547 if (invlist_is_iterating(invlist)) {
10548 Perl_dump_indent(aTHX_ level, file,
10549 "%sCan't dump inversion list because is in middle of iterating\n",
10554 invlist_iterinit(invlist);
10555 while (invlist_iternext(invlist, &start, &end)) {
10556 if (end == UV_MAX) {
10557 Perl_dump_indent(aTHX_ level, file,
10558 "%s[%" UVuf "] 0x%04" UVXf " .. INFTY\n",
10559 indent, (UV)count, start);
10561 else if (end != start) {
10562 Perl_dump_indent(aTHX_ level, file,
10563 "%s[%" UVuf "] 0x%04" UVXf " .. 0x%04" UVXf "\n",
10564 indent, (UV)count, start, end);
10567 Perl_dump_indent(aTHX_ level, file, "%s[%" UVuf "] 0x%04" UVXf "\n",
10568 indent, (UV)count, start);
10576 #if defined(PERL_ARGS_ASSERT__INVLISTEQ) && !defined(PERL_IN_XSUB_RE)
10578 Perl__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
10580 /* Return a boolean as to if the two passed in inversion lists are
10581 * identical. The final argument, if TRUE, says to take the complement of
10582 * the second inversion list before doing the comparison */
10584 const UV len_a = _invlist_len(a);
10585 UV len_b = _invlist_len(b);
10587 const UV* array_a = NULL;
10588 const UV* array_b = NULL;
10590 PERL_ARGS_ASSERT__INVLISTEQ;
10592 /* This code avoids accessing the arrays unless it knows the length is
10597 return ! complement_b;
10601 array_a = invlist_array(a);
10605 array_b = invlist_array(b);
10608 /* If are to compare 'a' with the complement of b, set it
10609 * up so are looking at b's complement. */
10610 if (complement_b) {
10612 /* The complement of nothing is everything, so <a> would have to have
10613 * just one element, starting at zero (ending at infinity) */
10615 return (len_a == 1 && array_a[0] == 0);
10617 if (array_b[0] == 0) {
10619 /* Otherwise, to complement, we invert. Here, the first element is
10620 * 0, just remove it. To do this, we just pretend the array starts
10628 /* But if the first element is not zero, we pretend the list starts
10629 * at the 0 that is always stored immediately before the array. */
10635 return len_a == len_b
10636 && memEQ(array_a, array_b, len_a * sizeof(array_a[0]));
10642 * As best we can, determine the characters that can match the start of
10643 * the given EXACTF-ish node. This is for use in creating ssc nodes, so there
10644 * can be false positive matches
10646 * Returns the invlist as a new SV*; it is the caller's responsibility to
10647 * call SvREFCNT_dec() when done with it.
10650 S_make_exactf_invlist(pTHX_ RExC_state_t *pRExC_state, regnode *node)
10652 const U8 * s = (U8*)STRING(node);
10653 SSize_t bytelen = STR_LEN(node);
10655 /* Start out big enough for 2 separate code points */
10656 SV* invlist = _new_invlist(4);
10658 PERL_ARGS_ASSERT_MAKE_EXACTF_INVLIST;
10663 /* We punt and assume can match anything if the node begins
10664 * with a multi-character fold. Things are complicated. For
10665 * example, /ffi/i could match any of:
10666 * "\N{LATIN SMALL LIGATURE FFI}"
10667 * "\N{LATIN SMALL LIGATURE FF}I"
10668 * "F\N{LATIN SMALL LIGATURE FI}"
10669 * plus several other things; and making sure we have all the
10670 * possibilities is hard. */
10671 if (is_MULTI_CHAR_FOLD_latin1_safe(s, s + bytelen)) {
10672 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10675 /* Any Latin1 range character can potentially match any
10676 * other depending on the locale, and in Turkic locales, U+130 and
10678 if (OP(node) == EXACTFL) {
10679 _invlist_union(invlist, PL_Latin1, &invlist);
10680 invlist = add_cp_to_invlist(invlist,
10681 LATIN_SMALL_LETTER_DOTLESS_I);
10682 invlist = add_cp_to_invlist(invlist,
10683 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
10686 /* But otherwise, it matches at least itself. We can
10687 * quickly tell if it has a distinct fold, and if so,
10688 * it matches that as well */
10689 invlist = add_cp_to_invlist(invlist, uc);
10690 if (IS_IN_SOME_FOLD_L1(uc))
10691 invlist = add_cp_to_invlist(invlist, PL_fold_latin1[uc]);
10694 /* Some characters match above-Latin1 ones under /i. This
10695 * is true of EXACTFL ones when the locale is UTF-8 */
10696 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(uc)
10697 && (! isASCII(uc) || ! inRANGE(OP(node), EXACTFAA,
10698 EXACTFAA_NO_TRIE)))
10700 add_above_Latin1_folds(pRExC_state, (U8) uc, &invlist);
10704 else { /* Pattern is UTF-8 */
10705 U8 folded[UTF8_MAX_FOLD_CHAR_EXPAND * UTF8_MAXBYTES_CASE + 1] = { '\0' };
10706 const U8* e = s + bytelen;
10709 fc = uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
10711 /* The only code points that aren't folded in a UTF EXACTFish
10712 * node are the problematic ones in EXACTFL nodes */
10713 if (OP(node) == EXACTFL && is_PROBLEMATIC_LOCALE_FOLDEDS_START_cp(uc)) {
10714 /* We need to check for the possibility that this EXACTFL
10715 * node begins with a multi-char fold. Therefore we fold
10716 * the first few characters of it so that we can make that
10722 for (i = 0; i < UTF8_MAX_FOLD_CHAR_EXPAND && s < e; i++) {
10724 *(d++) = (U8) toFOLD(*s);
10725 if (fc < 0) { /* Save the first fold */
10732 UV fold = toFOLD_utf8_safe(s, e, d, &len);
10733 if (fc < 0) { /* Save the first fold */
10741 /* And set up so the code below that looks in this folded
10742 * buffer instead of the node's string */
10747 /* When we reach here 's' points to the fold of the first
10748 * character(s) of the node; and 'e' points to far enough along
10749 * the folded string to be just past any possible multi-char
10752 * Like the non-UTF case above, we punt if the node begins with a
10753 * multi-char fold */
10755 if (is_MULTI_CHAR_FOLD_utf8_safe(s, e)) {
10756 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10758 else { /* Single char fold */
10761 const U32 * remaining_folds;
10762 Size_t folds_count;
10764 /* It matches itself */
10765 invlist = add_cp_to_invlist(invlist, fc);
10767 /* ... plus all the things that fold to it, which are found in
10768 * PL_utf8_foldclosures */
10769 folds_count = _inverse_folds(fc, &first_fold,
10771 for (k = 0; k < folds_count; k++) {
10772 UV c = (k == 0) ? first_fold : remaining_folds[k-1];
10774 /* /aa doesn't allow folds between ASCII and non- */
10775 if ( inRANGE(OP(node), EXACTFAA, EXACTFAA_NO_TRIE)
10776 && isASCII(c) != isASCII(fc))
10781 invlist = add_cp_to_invlist(invlist, c);
10784 if (OP(node) == EXACTFL) {
10786 /* If either [iI] are present in an EXACTFL node the above code
10787 * should have added its normal case pair, but under a Turkish
10788 * locale they could match instead the case pairs from it. Add
10789 * those as potential matches as well */
10790 if (isALPHA_FOLD_EQ(fc, 'I')) {
10791 invlist = add_cp_to_invlist(invlist,
10792 LATIN_SMALL_LETTER_DOTLESS_I);
10793 invlist = add_cp_to_invlist(invlist,
10794 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
10796 else if (fc == LATIN_SMALL_LETTER_DOTLESS_I) {
10797 invlist = add_cp_to_invlist(invlist, 'I');
10799 else if (fc == LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE) {
10800 invlist = add_cp_to_invlist(invlist, 'i');
10809 #undef HEADER_LENGTH
10810 #undef TO_INTERNAL_SIZE
10811 #undef FROM_INTERNAL_SIZE
10812 #undef INVLIST_VERSION_ID
10814 /* End of inversion list object */
10817 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
10819 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
10820 * constructs, and updates RExC_flags with them. On input, RExC_parse
10821 * should point to the first flag; it is updated on output to point to the
10822 * final ')' or ':'. There needs to be at least one flag, or this will
10825 /* for (?g), (?gc), and (?o) warnings; warning
10826 about (?c) will warn about (?g) -- japhy */
10828 #define WASTED_O 0x01
10829 #define WASTED_G 0x02
10830 #define WASTED_C 0x04
10831 #define WASTED_GC (WASTED_G|WASTED_C)
10832 I32 wastedflags = 0x00;
10833 U32 posflags = 0, negflags = 0;
10834 U32 *flagsp = &posflags;
10835 char has_charset_modifier = '\0';
10837 bool has_use_defaults = FALSE;
10838 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
10839 int x_mod_count = 0;
10841 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
10843 /* '^' as an initial flag sets certain defaults */
10844 if (UCHARAT(RExC_parse) == '^') {
10846 has_use_defaults = TRUE;
10847 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
10848 cs = (toUSE_UNI_CHARSET_NOT_DEPENDS)
10849 ? REGEX_UNICODE_CHARSET
10850 : REGEX_DEPENDS_CHARSET;
10851 set_regex_charset(&RExC_flags, cs);
10854 cs = get_regex_charset(RExC_flags);
10855 if ( cs == REGEX_DEPENDS_CHARSET
10856 && (toUSE_UNI_CHARSET_NOT_DEPENDS))
10858 cs = REGEX_UNICODE_CHARSET;
10862 while (RExC_parse < RExC_end) {
10863 /* && memCHRs("iogcmsx", *RExC_parse) */
10864 /* (?g), (?gc) and (?o) are useless here
10865 and must be globally applied -- japhy */
10866 if ((RExC_pm_flags & PMf_WILDCARD)) {
10867 if (flagsp == & negflags) {
10868 if (*RExC_parse == 'm') {
10870 /* diag_listed_as: Use of %s is not allowed in Unicode
10871 property wildcard subpatterns in regex; marked by <--
10873 vFAIL("Use of modifier '-m' is not allowed in Unicode"
10874 " property wildcard subpatterns");
10878 if (*RExC_parse == 's') {
10879 goto modifier_illegal_in_wildcard;
10884 switch (*RExC_parse) {
10886 /* Code for the imsxn flags */
10887 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp, x_mod_count);
10889 case LOCALE_PAT_MOD:
10890 if (has_charset_modifier) {
10891 goto excess_modifier;
10893 else if (flagsp == &negflags) {
10896 cs = REGEX_LOCALE_CHARSET;
10897 has_charset_modifier = LOCALE_PAT_MOD;
10899 case UNICODE_PAT_MOD:
10900 if (has_charset_modifier) {
10901 goto excess_modifier;
10903 else if (flagsp == &negflags) {
10906 cs = REGEX_UNICODE_CHARSET;
10907 has_charset_modifier = UNICODE_PAT_MOD;
10909 case ASCII_RESTRICT_PAT_MOD:
10910 if (flagsp == &negflags) {
10913 if (has_charset_modifier) {
10914 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
10915 goto excess_modifier;
10917 /* Doubled modifier implies more restricted */
10918 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
10921 cs = REGEX_ASCII_RESTRICTED_CHARSET;
10923 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
10925 case DEPENDS_PAT_MOD:
10926 if (has_use_defaults) {
10927 goto fail_modifiers;
10929 else if (flagsp == &negflags) {
10932 else if (has_charset_modifier) {
10933 goto excess_modifier;
10936 /* The dual charset means unicode semantics if the
10937 * pattern (or target, not known until runtime) are
10938 * utf8, or something in the pattern indicates unicode
10940 cs = (toUSE_UNI_CHARSET_NOT_DEPENDS)
10941 ? REGEX_UNICODE_CHARSET
10942 : REGEX_DEPENDS_CHARSET;
10943 has_charset_modifier = DEPENDS_PAT_MOD;
10947 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
10948 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
10950 else if (has_charset_modifier == *(RExC_parse - 1)) {
10951 vFAIL2("Regexp modifier \"%c\" may not appear twice",
10952 *(RExC_parse - 1));
10955 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
10957 NOT_REACHED; /*NOTREACHED*/
10960 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
10961 *(RExC_parse - 1));
10962 NOT_REACHED; /*NOTREACHED*/
10963 case GLOBAL_PAT_MOD: /* 'g' */
10964 if (RExC_pm_flags & PMf_WILDCARD) {
10965 goto modifier_illegal_in_wildcard;
10968 case ONCE_PAT_MOD: /* 'o' */
10969 if (ckWARN(WARN_REGEXP)) {
10970 const I32 wflagbit = *RExC_parse == 'o'
10973 if (! (wastedflags & wflagbit) ) {
10974 wastedflags |= wflagbit;
10975 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10978 "Useless (%s%c) - %suse /%c modifier",
10979 flagsp == &negflags ? "?-" : "?",
10981 flagsp == &negflags ? "don't " : "",
10988 case CONTINUE_PAT_MOD: /* 'c' */
10989 if (RExC_pm_flags & PMf_WILDCARD) {
10990 goto modifier_illegal_in_wildcard;
10992 if (ckWARN(WARN_REGEXP)) {
10993 if (! (wastedflags & WASTED_C) ) {
10994 wastedflags |= WASTED_GC;
10995 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10998 "Useless (%sc) - %suse /gc modifier",
10999 flagsp == &negflags ? "?-" : "?",
11000 flagsp == &negflags ? "don't " : ""
11005 case KEEPCOPY_PAT_MOD: /* 'p' */
11006 if (RExC_pm_flags & PMf_WILDCARD) {
11007 goto modifier_illegal_in_wildcard;
11009 if (flagsp == &negflags) {
11010 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
11012 *flagsp |= RXf_PMf_KEEPCOPY;
11016 /* A flag is a default iff it is following a minus, so
11017 * if there is a minus, it means will be trying to
11018 * re-specify a default which is an error */
11019 if (has_use_defaults || flagsp == &negflags) {
11020 goto fail_modifiers;
11022 flagsp = &negflags;
11023 wastedflags = 0; /* reset so (?g-c) warns twice */
11029 if ( (RExC_pm_flags & PMf_WILDCARD)
11030 && cs != REGEX_ASCII_MORE_RESTRICTED_CHARSET)
11033 /* diag_listed_as: Use of %s is not allowed in Unicode
11034 property wildcard subpatterns in regex; marked by <--
11036 vFAIL2("Use of modifier '%c' is not allowed in Unicode"
11037 " property wildcard subpatterns",
11038 has_charset_modifier);
11041 if ((posflags & (RXf_PMf_EXTENDED|RXf_PMf_EXTENDED_MORE)) == RXf_PMf_EXTENDED) {
11042 negflags |= RXf_PMf_EXTENDED_MORE;
11044 RExC_flags |= posflags;
11046 if (negflags & RXf_PMf_EXTENDED) {
11047 negflags |= RXf_PMf_EXTENDED_MORE;
11049 RExC_flags &= ~negflags;
11050 set_regex_charset(&RExC_flags, cs);
11055 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
11056 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11057 vFAIL2utf8f("Sequence (%" UTF8f "...) not recognized",
11058 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
11059 NOT_REACHED; /*NOTREACHED*/
11062 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11065 vFAIL("Sequence (?... not terminated");
11067 modifier_illegal_in_wildcard:
11069 /* diag_listed_as: Use of %s is not allowed in Unicode property wildcard
11070 subpatterns in regex; marked by <-- HERE in m/%s/ */
11071 vFAIL2("Use of modifier '%c' is not allowed in Unicode property wildcard"
11072 " subpatterns", *(RExC_parse - 1));
11076 - reg - regular expression, i.e. main body or parenthesized thing
11078 * Caller must absorb opening parenthesis.
11080 * Combining parenthesis handling with the base level of regular expression
11081 * is a trifle forced, but the need to tie the tails of the branches to what
11082 * follows makes it hard to avoid.
11084 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
11086 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
11088 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
11091 STATIC regnode_offset
11092 S_handle_named_backref(pTHX_ RExC_state_t *pRExC_state,
11094 char * parse_start,
11098 regnode_offset ret;
11099 char* name_start = RExC_parse;
11101 SV *sv_dat = reg_scan_name(pRExC_state, REG_RSN_RETURN_DATA);
11102 DECLARE_AND_GET_RE_DEBUG_FLAGS;
11104 PERL_ARGS_ASSERT_HANDLE_NAMED_BACKREF;
11106 if (RExC_parse != name_start && ch == '}') {
11107 while (isBLANK(*RExC_parse)) {
11111 if (RExC_parse == name_start || *RExC_parse != ch) {
11112 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
11113 vFAIL2("Sequence %.3s... not terminated", parse_start);
11117 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11118 RExC_rxi->data->data[num]=(void*)sv_dat;
11119 SvREFCNT_inc_simple_void_NN(sv_dat);
11122 ret = reganode(pRExC_state,
11125 : (ASCII_FOLD_RESTRICTED)
11127 : (AT_LEAST_UNI_SEMANTICS)
11133 *flagp |= HASWIDTH;
11135 Set_Node_Offset(REGNODE_p(ret), parse_start+1);
11136 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
11138 nextchar(pRExC_state);
11142 /* On success, returns the offset at which any next node should be placed into
11143 * the regex engine program being compiled.
11145 * Returns 0 otherwise, with *flagp set to indicate why:
11146 * TRYAGAIN at the end of (?) that only sets flags.
11147 * RESTART_PARSE if the parse needs to be restarted, or'd with
11148 * NEED_UTF8 if the pattern needs to be upgraded to UTF-8.
11149 * Otherwise would only return 0 if regbranch() returns 0, which cannot
11151 STATIC regnode_offset
11152 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp, U32 depth)
11153 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
11154 * 2 is like 1, but indicates that nextchar() has been called to advance
11155 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
11156 * this flag alerts us to the need to check for that */
11158 regnode_offset ret = 0; /* Will be the head of the group. */
11160 regnode_offset lastbr;
11161 regnode_offset ender = 0;
11164 U32 oregflags = RExC_flags;
11165 bool have_branch = 0;
11167 I32 freeze_paren = 0;
11168 I32 after_freeze = 0;
11169 I32 num; /* numeric backreferences */
11170 SV * max_open; /* Max number of unclosed parens */
11171 I32 was_in_lookaround = RExC_in_lookaround;
11173 char * parse_start = RExC_parse; /* MJD */
11174 char * const oregcomp_parse = RExC_parse;
11176 DECLARE_AND_GET_RE_DEBUG_FLAGS;
11178 PERL_ARGS_ASSERT_REG;
11179 DEBUG_PARSE("reg ");
11181 max_open = get_sv(RE_COMPILE_RECURSION_LIMIT, GV_ADD);
11183 if (!SvIOK(max_open)) {
11184 sv_setiv(max_open, RE_COMPILE_RECURSION_INIT);
11186 if (depth > 4 * (UV) SvIV(max_open)) { /* We increase depth by 4 for each
11188 vFAIL("Too many nested open parens");
11191 *flagp = 0; /* Initialize. */
11193 /* Having this true makes it feasible to have a lot fewer tests for the
11194 * parse pointer being in scope. For example, we can write
11195 * while(isFOO(*RExC_parse)) RExC_parse++;
11197 * while(RExC_parse < RExC_end && isFOO(*RExC_parse)) RExC_parse++;
11199 assert(*RExC_end == '\0');
11201 /* Make an OPEN node, if parenthesized. */
11204 /* Under /x, space and comments can be gobbled up between the '(' and
11205 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
11206 * intervening space, as the sequence is a token, and a token should be
11208 bool has_intervening_patws = (paren == 2)
11209 && *(RExC_parse - 1) != '(';
11211 if (RExC_parse >= RExC_end) {
11212 vFAIL("Unmatched (");
11215 if (paren == 'r') { /* Atomic script run */
11219 else if ( *RExC_parse == '*') { /* (*VERB:ARG), (*construct:...) */
11220 char *start_verb = RExC_parse + 1;
11222 char *start_arg = NULL;
11223 unsigned char op = 0;
11224 int arg_required = 0;
11225 int internal_argval = -1; /* if >-1 we are not allowed an argument*/
11226 bool has_upper = FALSE;
11228 if (has_intervening_patws) {
11229 RExC_parse++; /* past the '*' */
11231 /* For strict backwards compatibility, don't change the message
11232 * now that we also have lowercase operands */
11233 if (isUPPER(*RExC_parse)) {
11234 vFAIL("In '(*VERB...)', the '(' and '*' must be adjacent");
11237 vFAIL("In '(*...)', the '(' and '*' must be adjacent");
11240 while (RExC_parse < RExC_end && *RExC_parse != ')' ) {
11241 if ( *RExC_parse == ':' ) {
11242 start_arg = RExC_parse + 1;
11246 if (isUPPER(*RExC_parse)) {
11252 RExC_parse += UTF8SKIP(RExC_parse);
11255 verb_len = RExC_parse - start_verb;
11257 if (RExC_parse >= RExC_end) {
11258 goto unterminated_verb_pattern;
11261 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11262 while ( RExC_parse < RExC_end && *RExC_parse != ')' ) {
11263 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11265 if ( RExC_parse >= RExC_end || *RExC_parse != ')' ) {
11266 unterminated_verb_pattern:
11268 vFAIL("Unterminated verb pattern argument");
11271 vFAIL("Unterminated '(*...' argument");
11275 if ( RExC_parse >= RExC_end || *RExC_parse != ')' ) {
11277 vFAIL("Unterminated verb pattern");
11280 vFAIL("Unterminated '(*...' construct");
11285 /* Here, we know that RExC_parse < RExC_end */
11287 switch ( *start_verb ) {
11288 case 'A': /* (*ACCEPT) */
11289 if ( memEQs(start_verb, verb_len,"ACCEPT") ) {
11291 internal_argval = RExC_nestroot;
11294 case 'C': /* (*COMMIT) */
11295 if ( memEQs(start_verb, verb_len,"COMMIT") )
11298 case 'F': /* (*FAIL) */
11299 if ( verb_len==1 || memEQs(start_verb, verb_len,"FAIL") ) {
11303 case ':': /* (*:NAME) */
11304 case 'M': /* (*MARK:NAME) */
11305 if ( verb_len==0 || memEQs(start_verb, verb_len,"MARK") ) {
11310 case 'P': /* (*PRUNE) */
11311 if ( memEQs(start_verb, verb_len,"PRUNE") )
11314 case 'S': /* (*SKIP) */
11315 if ( memEQs(start_verb, verb_len,"SKIP") )
11318 case 'T': /* (*THEN) */
11319 /* [19:06] <TimToady> :: is then */
11320 if ( memEQs(start_verb, verb_len,"THEN") ) {
11322 RExC_seen |= REG_CUTGROUP_SEEN;
11326 if ( memEQs(start_verb, verb_len, "asr")
11327 || memEQs(start_verb, verb_len, "atomic_script_run"))
11329 paren = 'r'; /* Mnemonic: recursed run */
11332 else if (memEQs(start_verb, verb_len, "atomic")) {
11333 paren = 't'; /* AtOMIC */
11334 goto alpha_assertions;
11338 if ( memEQs(start_verb, verb_len, "plb")
11339 || memEQs(start_verb, verb_len, "positive_lookbehind"))
11342 goto lookbehind_alpha_assertions;
11344 else if ( memEQs(start_verb, verb_len, "pla")
11345 || memEQs(start_verb, verb_len, "positive_lookahead"))
11348 goto alpha_assertions;
11352 if ( memEQs(start_verb, verb_len, "nlb")
11353 || memEQs(start_verb, verb_len, "negative_lookbehind"))
11356 goto lookbehind_alpha_assertions;
11358 else if ( memEQs(start_verb, verb_len, "nla")
11359 || memEQs(start_verb, verb_len, "negative_lookahead"))
11362 goto alpha_assertions;
11366 if ( memEQs(start_verb, verb_len, "sr")
11367 || memEQs(start_verb, verb_len, "script_run"))
11369 regnode_offset atomic;
11375 /* This indicates Unicode rules. */
11376 REQUIRE_UNI_RULES(flagp, 0);
11382 RExC_parse = start_arg;
11384 if (RExC_in_script_run) {
11386 /* Nested script runs are treated as no-ops, because
11387 * if the nested one fails, the outer one must as
11388 * well. It could fail sooner, and avoid (??{} with
11389 * side effects, but that is explicitly documented as
11390 * undefined behavior. */
11394 if (paren == 's') {
11399 /* But, the atomic part of a nested atomic script run
11400 * isn't a no-op, but can be treated just like a '(?>'
11406 if (paren == 's') {
11407 /* Here, we're starting a new regular script run */
11408 ret = reg_node(pRExC_state, SROPEN);
11409 RExC_in_script_run = 1;
11414 /* Here, we are starting an atomic script run. This is
11415 * handled by recursing to deal with the atomic portion
11416 * separately, enclosed in SROPEN ... SRCLOSE nodes */
11418 ret = reg_node(pRExC_state, SROPEN);
11420 RExC_in_script_run = 1;
11422 atomic = reg(pRExC_state, 'r', &flags, depth);
11423 if (flags & (RESTART_PARSE|NEED_UTF8)) {
11424 *flagp = flags & (RESTART_PARSE|NEED_UTF8);
11428 if (! REGTAIL(pRExC_state, ret, atomic)) {
11429 REQUIRE_BRANCHJ(flagp, 0);
11432 if (! REGTAIL(pRExC_state, atomic, reg_node(pRExC_state,
11435 REQUIRE_BRANCHJ(flagp, 0);
11438 RExC_in_script_run = 0;
11444 lookbehind_alpha_assertions:
11445 RExC_seen |= REG_LOOKBEHIND_SEEN;
11450 RExC_in_lookaround++;
11451 RExC_seen_zerolen++;
11457 /* An empty negative lookahead assertion simply is failure */
11458 if (paren == 'A' && RExC_parse == start_arg) {
11459 ret=reganode(pRExC_state, OPFAIL, 0);
11460 nextchar(pRExC_state);
11464 RExC_parse = start_arg;
11469 "'(*%" UTF8f "' requires a terminating ':'",
11470 UTF8fARG(UTF, verb_len, start_verb));
11471 NOT_REACHED; /*NOTREACHED*/
11473 } /* End of switch */
11476 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
11478 if (has_upper || verb_len == 0) {
11480 "Unknown verb pattern '%" UTF8f "'",
11481 UTF8fARG(UTF, verb_len, start_verb));
11485 "Unknown '(*...)' construct '%" UTF8f "'",
11486 UTF8fARG(UTF, verb_len, start_verb));
11489 if ( RExC_parse == start_arg ) {
11492 if ( arg_required && !start_arg ) {
11493 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
11494 (int) verb_len, start_verb);
11496 if (internal_argval == -1) {
11497 ret = reganode(pRExC_state, op, 0);
11499 ret = reg2Lanode(pRExC_state, op, 0, internal_argval);
11501 RExC_seen |= REG_VERBARG_SEEN;
11503 SV *sv = newSVpvn( start_arg,
11504 RExC_parse - start_arg);
11505 ARG(REGNODE_p(ret)) = add_data( pRExC_state,
11506 STR_WITH_LEN("S"));
11507 RExC_rxi->data->data[ARG(REGNODE_p(ret))]=(void*)sv;
11508 FLAGS(REGNODE_p(ret)) = 1;
11510 FLAGS(REGNODE_p(ret)) = 0;
11512 if ( internal_argval != -1 )
11513 ARG2L_SET(REGNODE_p(ret), internal_argval);
11514 nextchar(pRExC_state);
11517 else if (*RExC_parse == '?') { /* (?...) */
11518 bool is_logical = 0;
11519 const char * const seqstart = RExC_parse;
11520 const char * endptr;
11521 const char non_existent_group_msg[]
11522 = "Reference to nonexistent group";
11523 const char impossible_group[] = "Invalid reference to group";
11525 if (has_intervening_patws) {
11527 vFAIL("In '(?...)', the '(' and '?' must be adjacent");
11530 RExC_parse++; /* past the '?' */
11531 paren = *RExC_parse; /* might be a trailing NUL, if not
11533 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11534 if (RExC_parse > RExC_end) {
11537 ret = 0; /* For look-ahead/behind. */
11540 case 'P': /* (?P...) variants for those used to PCRE/Python */
11541 paren = *RExC_parse;
11542 if ( paren == '<') { /* (?P<...>) named capture */
11544 if (RExC_parse >= RExC_end) {
11545 vFAIL("Sequence (?P<... not terminated");
11547 goto named_capture;
11549 else if (paren == '>') { /* (?P>name) named recursion */
11551 if (RExC_parse >= RExC_end) {
11552 vFAIL("Sequence (?P>... not terminated");
11554 goto named_recursion;
11556 else if (paren == '=') { /* (?P=...) named backref */
11558 return handle_named_backref(pRExC_state, flagp,
11561 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
11562 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11563 vFAIL3("Sequence (%.*s...) not recognized",
11564 (int) (RExC_parse - seqstart), seqstart);
11565 NOT_REACHED; /*NOTREACHED*/
11566 case '<': /* (?<...) */
11567 /* If you want to support (?<*...), first reconcile with GH #17363 */
11568 if (*RExC_parse == '!')
11570 else if (*RExC_parse != '=')
11577 case '\'': /* (?'...') */
11578 name_start = RExC_parse;
11579 svname = reg_scan_name(pRExC_state, REG_RSN_RETURN_NAME);
11580 if ( RExC_parse == name_start
11581 || RExC_parse >= RExC_end
11582 || *RExC_parse != paren)
11584 vFAIL2("Sequence (?%c... not terminated",
11585 paren=='>' ? '<' : (char) paren);
11590 if (!svname) /* shouldn't happen */
11592 "panic: reg_scan_name returned NULL");
11593 if (!RExC_paren_names) {
11594 RExC_paren_names= newHV();
11595 sv_2mortal(MUTABLE_SV(RExC_paren_names));
11597 RExC_paren_name_list= newAV();
11598 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
11601 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
11603 sv_dat = HeVAL(he_str);
11605 /* croak baby croak */
11607 "panic: paren_name hash element allocation failed");
11608 } else if ( SvPOK(sv_dat) ) {
11609 /* (?|...) can mean we have dupes so scan to check
11610 its already been stored. Maybe a flag indicating
11611 we are inside such a construct would be useful,
11612 but the arrays are likely to be quite small, so
11613 for now we punt -- dmq */
11614 IV count = SvIV(sv_dat);
11615 I32 *pv = (I32*)SvPVX(sv_dat);
11617 for ( i = 0 ; i < count ; i++ ) {
11618 if ( pv[i] == RExC_npar ) {
11624 pv = (I32*)SvGROW(sv_dat,
11625 SvCUR(sv_dat) + sizeof(I32)+1);
11626 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
11627 pv[count] = RExC_npar;
11628 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
11631 (void)SvUPGRADE(sv_dat, SVt_PVNV);
11632 sv_setpvn(sv_dat, (char *)&(RExC_npar),
11635 SvIV_set(sv_dat, 1);
11638 /* Yes this does cause a memory leak in debugging Perls
11640 if (!av_store(RExC_paren_name_list,
11641 RExC_npar, SvREFCNT_inc_NN(svname)))
11642 SvREFCNT_dec_NN(svname);
11645 /*sv_dump(sv_dat);*/
11647 nextchar(pRExC_state);
11649 goto capturing_parens;
11652 RExC_seen |= REG_LOOKBEHIND_SEEN;
11653 RExC_in_lookaround++;
11655 if (RExC_parse >= RExC_end) {
11656 vFAIL("Sequence (?... not terminated");
11658 RExC_seen_zerolen++;
11660 case '=': /* (?=...) */
11661 RExC_seen_zerolen++;
11662 RExC_in_lookaround++;
11664 case '!': /* (?!...) */
11665 RExC_seen_zerolen++;
11666 /* check if we're really just a "FAIL" assertion */
11667 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
11668 FALSE /* Don't force to /x */ );
11669 if (*RExC_parse == ')') {
11670 ret=reganode(pRExC_state, OPFAIL, 0);
11671 nextchar(pRExC_state);
11674 RExC_in_lookaround++;
11676 case '|': /* (?|...) */
11677 /* branch reset, behave like a (?:...) except that
11678 buffers in alternations share the same numbers */
11680 after_freeze = freeze_paren = RExC_npar;
11682 /* XXX This construct currently requires an extra pass.
11683 * Investigation would be required to see if that could be
11685 REQUIRE_PARENS_PASS;
11687 case ':': /* (?:...) */
11688 case '>': /* (?>...) */
11690 case '$': /* (?$...) */
11691 case '@': /* (?@...) */
11692 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
11694 case '0' : /* (?0) */
11695 case 'R' : /* (?R) */
11696 if (RExC_parse == RExC_end || *RExC_parse != ')')
11697 FAIL("Sequence (?R) not terminated");
11699 RExC_seen |= REG_RECURSE_SEEN;
11701 /* XXX These constructs currently require an extra pass.
11702 * It probably could be changed */
11703 REQUIRE_PARENS_PASS;
11705 *flagp |= POSTPONED;
11706 goto gen_recurse_regop;
11708 /* named and numeric backreferences */
11709 case '&': /* (?&NAME) */
11710 parse_start = RExC_parse - 1;
11713 SV *sv_dat = reg_scan_name(pRExC_state,
11714 REG_RSN_RETURN_DATA);
11715 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
11717 if (RExC_parse >= RExC_end || *RExC_parse != ')')
11718 vFAIL("Sequence (?&... not terminated");
11719 goto gen_recurse_regop;
11722 if (! inRANGE(RExC_parse[0], '1', '9')) {
11724 vFAIL("Illegal pattern");
11726 goto parse_recursion;
11728 case '-': /* (?-1) */
11729 if (! inRANGE(RExC_parse[0], '1', '9')) {
11730 RExC_parse--; /* rewind to let it be handled later */
11734 case '1': case '2': case '3': case '4': /* (?1) */
11735 case '5': case '6': case '7': case '8': case '9':
11736 RExC_parse = (char *) seqstart + 1; /* Point to the digit */
11739 bool is_neg = FALSE;
11741 parse_start = RExC_parse - 1; /* MJD */
11742 if (*RExC_parse == '-') {
11747 if (grok_atoUV(RExC_parse, &unum, &endptr)
11751 RExC_parse = (char*)endptr;
11753 else { /* Overflow, or something like that. Position
11754 beyond all digits for the message */
11755 while (RExC_parse < RExC_end && isDIGIT(*RExC_parse)) {
11758 vFAIL(impossible_group);
11761 /* -num is always representable on 1 and 2's complement
11766 if (*RExC_parse!=')')
11767 vFAIL("Expecting close bracket");
11770 if (paren == '-' || paren == '+') {
11772 /* Don't overflow */
11773 if (UNLIKELY(I32_MAX - RExC_npar < num)) {
11775 vFAIL(impossible_group);
11779 Diagram of capture buffer numbering.
11780 Top line is the normal capture buffer numbers
11781 Bottom line is the negative indexing as from
11785 /(a(x)y)(a(b(c(?+2)d)e)f)(g(h))/
11786 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
11787 - 5 4 3 2 1 X Y x x
11789 Resolve to absolute group. Recall that RExC_npar is +1 of
11790 the actual parenthesis group number. For lookahead, we
11791 have to compensate for that. Using the above example, when
11792 we get to Y in the parse, num is 2 and RExC_npar is 6. We
11793 want 7 for +2, and 4 for -2.
11795 if ( paren == '+' ) {
11801 if (paren == '-' && num < 1) {
11803 vFAIL(non_existent_group_msg);
11807 if (num >= RExC_npar) {
11809 /* It might be a forward reference; we can't fail until we
11810 * know, by completing the parse to get all the groups, and
11811 * then reparsing */
11812 if (ALL_PARENS_COUNTED) {
11813 if (num >= RExC_total_parens) {
11815 vFAIL(non_existent_group_msg);
11819 REQUIRE_PARENS_PASS;
11823 /* We keep track how many GOSUB items we have produced.
11824 To start off the ARG2L() of the GOSUB holds its "id",
11825 which is used later in conjunction with RExC_recurse
11826 to calculate the offset we need to jump for the GOSUB,
11827 which it will store in the final representation.
11828 We have to defer the actual calculation until much later
11829 as the regop may move.
11831 ret = reg2Lanode(pRExC_state, GOSUB, num, RExC_recurse_count);
11832 RExC_recurse_count++;
11833 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11834 "%*s%*s Recurse #%" UVuf " to %" IVdf "\n",
11835 22, "| |", (int)(depth * 2 + 1), "",
11836 (UV)ARG(REGNODE_p(ret)),
11837 (IV)ARG2L(REGNODE_p(ret))));
11838 RExC_seen |= REG_RECURSE_SEEN;
11840 Set_Node_Length(REGNODE_p(ret),
11841 1 + regarglen[OP(REGNODE_p(ret))]); /* MJD */
11842 Set_Node_Offset(REGNODE_p(ret), parse_start); /* MJD */
11844 *flagp |= POSTPONED;
11845 assert(*RExC_parse == ')');
11846 nextchar(pRExC_state);
11851 case '?': /* (??...) */
11853 if (*RExC_parse != '{') {
11854 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
11855 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11857 "Sequence (%" UTF8f "...) not recognized",
11858 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
11859 NOT_REACHED; /*NOTREACHED*/
11861 *flagp |= POSTPONED;
11865 case '{': /* (?{...}) */
11868 struct reg_code_block *cb;
11871 RExC_seen_zerolen++;
11873 if ( !pRExC_state->code_blocks
11874 || pRExC_state->code_index
11875 >= pRExC_state->code_blocks->count
11876 || pRExC_state->code_blocks->cb[pRExC_state->code_index].start
11877 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
11880 if (RExC_pm_flags & PMf_USE_RE_EVAL)
11881 FAIL("panic: Sequence (?{...}): no code block found\n");
11882 FAIL("Eval-group not allowed at runtime, use re 'eval'");
11884 /* this is a pre-compiled code block (?{...}) */
11885 cb = &pRExC_state->code_blocks->cb[pRExC_state->code_index];
11886 RExC_parse = RExC_start + cb->end;
11888 if (cb->src_regex) {
11889 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
11890 RExC_rxi->data->data[n] =
11891 (void*)SvREFCNT_inc((SV*)cb->src_regex);
11892 RExC_rxi->data->data[n+1] = (void*)o;
11895 n = add_data(pRExC_state,
11896 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
11897 RExC_rxi->data->data[n] = (void*)o;
11899 pRExC_state->code_index++;
11900 nextchar(pRExC_state);
11903 regnode_offset eval;
11904 ret = reg_node(pRExC_state, LOGICAL);
11906 eval = reg2Lanode(pRExC_state, EVAL,
11909 /* for later propagation into (??{})
11911 RExC_flags & RXf_PMf_COMPILETIME
11913 FLAGS(REGNODE_p(ret)) = 2;
11914 if (! REGTAIL(pRExC_state, ret, eval)) {
11915 REQUIRE_BRANCHJ(flagp, 0);
11917 /* deal with the length of this later - MJD */
11920 ret = reg2Lanode(pRExC_state, EVAL, n, 0);
11921 Set_Node_Length(REGNODE_p(ret), RExC_parse - parse_start + 1);
11922 Set_Node_Offset(REGNODE_p(ret), parse_start);
11925 case '(': /* (?(?{...})...) and (?(?=...)...) */
11928 const int DEFINE_len = sizeof("DEFINE") - 1;
11929 if ( RExC_parse < RExC_end - 1
11930 && ( ( RExC_parse[0] == '?' /* (?(?...)) */
11931 && ( RExC_parse[1] == '='
11932 || RExC_parse[1] == '!'
11933 || RExC_parse[1] == '<'
11934 || RExC_parse[1] == '{'))
11935 || ( RExC_parse[0] == '*' /* (?(*...)) */
11936 && ( memBEGINs(RExC_parse + 1,
11937 (Size_t) (RExC_end - (RExC_parse + 1)),
11939 || memBEGINs(RExC_parse + 1,
11940 (Size_t) (RExC_end - (RExC_parse + 1)),
11942 || memBEGINs(RExC_parse + 1,
11943 (Size_t) (RExC_end - (RExC_parse + 1)),
11945 || memBEGINs(RExC_parse + 1,
11946 (Size_t) (RExC_end - (RExC_parse + 1)),
11948 || memBEGINs(RExC_parse + 1,
11949 (Size_t) (RExC_end - (RExC_parse + 1)),
11950 "positive_lookahead:")
11951 || memBEGINs(RExC_parse + 1,
11952 (Size_t) (RExC_end - (RExC_parse + 1)),
11953 "positive_lookbehind:")
11954 || memBEGINs(RExC_parse + 1,
11955 (Size_t) (RExC_end - (RExC_parse + 1)),
11956 "negative_lookahead:")
11957 || memBEGINs(RExC_parse + 1,
11958 (Size_t) (RExC_end - (RExC_parse + 1)),
11959 "negative_lookbehind:"))))
11960 ) { /* Lookahead or eval. */
11962 regnode_offset tail;
11964 ret = reg_node(pRExC_state, LOGICAL);
11965 FLAGS(REGNODE_p(ret)) = 1;
11967 tail = reg(pRExC_state, 1, &flag, depth+1);
11968 RETURN_FAIL_ON_RESTART(flag, flagp);
11969 if (! REGTAIL(pRExC_state, ret, tail)) {
11970 REQUIRE_BRANCHJ(flagp, 0);
11974 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
11975 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
11977 char ch = RExC_parse[0] == '<' ? '>' : '\'';
11978 char *name_start= RExC_parse++;
11980 SV *sv_dat=reg_scan_name(pRExC_state, REG_RSN_RETURN_DATA);
11981 if ( RExC_parse == name_start
11982 || RExC_parse >= RExC_end
11983 || *RExC_parse != ch)
11985 vFAIL2("Sequence (?(%c... not terminated",
11986 (ch == '>' ? '<' : ch));
11990 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11991 RExC_rxi->data->data[num]=(void*)sv_dat;
11992 SvREFCNT_inc_simple_void_NN(sv_dat);
11994 ret = reganode(pRExC_state, GROUPPN, num);
11995 goto insert_if_check_paren;
11997 else if (memBEGINs(RExC_parse,
11998 (STRLEN) (RExC_end - RExC_parse),
12001 ret = reganode(pRExC_state, DEFINEP, 0);
12002 RExC_parse += DEFINE_len;
12004 goto insert_if_check_paren;
12006 else if (RExC_parse[0] == 'R') {
12008 /* parno == 0 => /(?(R)YES|NO)/ "in any form of recursion OR eval"
12009 * parno == 1 => /(?(R0)YES|NO)/ "in GOSUB (?0) / (?R)"
12010 * parno == 2 => /(?(R1)YES|NO)/ "in GOSUB (?1) (parno-1)"
12013 if (RExC_parse[0] == '0') {
12017 else if (inRANGE(RExC_parse[0], '1', '9')) {
12020 if (grok_atoUV(RExC_parse, &uv, &endptr)
12023 parno = (I32)uv + 1;
12024 RExC_parse = (char*)endptr;
12026 /* else "Switch condition not recognized" below */
12027 } else if (RExC_parse[0] == '&') {
12030 sv_dat = reg_scan_name(pRExC_state,
12031 REG_RSN_RETURN_DATA);
12033 parno = 1 + *((I32 *)SvPVX(sv_dat));
12035 ret = reganode(pRExC_state, INSUBP, parno);
12036 goto insert_if_check_paren;
12038 else if (inRANGE(RExC_parse[0], '1', '9')) {
12043 if (grok_atoUV(RExC_parse, &uv, &endptr)
12047 RExC_parse = (char*)endptr;
12050 vFAIL("panic: grok_atoUV returned FALSE");
12052 ret = reganode(pRExC_state, GROUPP, parno);
12054 insert_if_check_paren:
12055 if (UCHARAT(RExC_parse) != ')') {
12057 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
12059 vFAIL("Switch condition not recognized");
12061 nextchar(pRExC_state);
12063 if (! REGTAIL(pRExC_state, ret, reganode(pRExC_state,
12066 REQUIRE_BRANCHJ(flagp, 0);
12068 br = regbranch(pRExC_state, &flags, 1, depth+1);
12070 RETURN_FAIL_ON_RESTART(flags,flagp);
12071 FAIL2("panic: regbranch returned failure, flags=%#" UVxf,
12074 if (! REGTAIL(pRExC_state, br, reganode(pRExC_state,
12077 REQUIRE_BRANCHJ(flagp, 0);
12079 c = UCHARAT(RExC_parse);
12080 nextchar(pRExC_state);
12081 if (flags&HASWIDTH)
12082 *flagp |= HASWIDTH;
12085 vFAIL("(?(DEFINE)....) does not allow branches");
12087 /* Fake one for optimizer. */
12088 lastbr = reganode(pRExC_state, IFTHEN, 0);
12090 if (!regbranch(pRExC_state, &flags, 1, depth+1)) {
12091 RETURN_FAIL_ON_RESTART(flags, flagp);
12092 FAIL2("panic: regbranch returned failure, flags=%#" UVxf,
12095 if (! REGTAIL(pRExC_state, ret, lastbr)) {
12096 REQUIRE_BRANCHJ(flagp, 0);
12098 if (flags&HASWIDTH)
12099 *flagp |= HASWIDTH;
12100 c = UCHARAT(RExC_parse);
12101 nextchar(pRExC_state);
12106 if (RExC_parse >= RExC_end)
12107 vFAIL("Switch (?(condition)... not terminated");
12109 vFAIL("Switch (?(condition)... contains too many branches");
12111 ender = reg_node(pRExC_state, TAIL);
12112 if (! REGTAIL(pRExC_state, br, ender)) {
12113 REQUIRE_BRANCHJ(flagp, 0);
12116 if (! REGTAIL(pRExC_state, lastbr, ender)) {
12117 REQUIRE_BRANCHJ(flagp, 0);
12119 if (! REGTAIL(pRExC_state,
12122 NEXTOPER(REGNODE_p(lastbr)))),
12125 REQUIRE_BRANCHJ(flagp, 0);
12129 if (! REGTAIL(pRExC_state, ret, ender)) {
12130 REQUIRE_BRANCHJ(flagp, 0);
12132 #if 0 /* Removing this doesn't cause failures in the test suite -- khw */
12133 RExC_size++; /* XXX WHY do we need this?!!
12134 For large programs it seems to be required
12135 but I can't figure out why. -- dmq*/
12140 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
12142 vFAIL("Unknown switch condition (?(...))");
12144 case '[': /* (?[ ... ]) */
12145 return handle_regex_sets(pRExC_state, NULL, flagp, depth+1,
12147 case 0: /* A NUL */
12148 RExC_parse--; /* for vFAIL to print correctly */
12149 vFAIL("Sequence (? incomplete");
12153 if (RExC_strict) { /* [perl #132851] */
12154 ckWARNreg(RExC_parse, "Empty (?) without any modifiers");
12157 case '*': /* If you want to support (?*...), first reconcile with GH #17363 */
12159 default: /* e.g., (?i) */
12160 RExC_parse = (char *) seqstart + 1;
12162 parse_lparen_question_flags(pRExC_state);
12163 if (UCHARAT(RExC_parse) != ':') {
12164 if (RExC_parse < RExC_end)
12165 nextchar(pRExC_state);
12170 nextchar(pRExC_state);
12175 else if (!(RExC_flags & RXf_PMf_NOCAPTURE)) { /* (...) */
12179 if (! ALL_PARENS_COUNTED) {
12180 /* If we are in our first pass through (and maybe only pass),
12181 * we need to allocate memory for the capturing parentheses
12185 if (!RExC_parens_buf_size) {
12186 /* first guess at number of parens we might encounter */
12187 RExC_parens_buf_size = 10;
12189 /* setup RExC_open_parens, which holds the address of each
12190 * OPEN tag, and to make things simpler for the 0 index the
12191 * start of the program - this is used later for offsets */
12192 Newxz(RExC_open_parens, RExC_parens_buf_size,
12194 RExC_open_parens[0] = 1; /* +1 for REG_MAGIC */
12196 /* setup RExC_close_parens, which holds the address of each
12197 * CLOSE tag, and to make things simpler for the 0 index
12198 * the end of the program - this is used later for offsets
12200 Newxz(RExC_close_parens, RExC_parens_buf_size,
12202 /* we dont know where end op starts yet, so we dont need to
12203 * set RExC_close_parens[0] like we do RExC_open_parens[0]
12206 else if (RExC_npar > RExC_parens_buf_size) {
12207 I32 old_size = RExC_parens_buf_size;
12209 RExC_parens_buf_size *= 2;
12211 Renew(RExC_open_parens, RExC_parens_buf_size,
12213 Zero(RExC_open_parens + old_size,
12214 RExC_parens_buf_size - old_size, regnode_offset);
12216 Renew(RExC_close_parens, RExC_parens_buf_size,
12218 Zero(RExC_close_parens + old_size,
12219 RExC_parens_buf_size - old_size, regnode_offset);
12223 ret = reganode(pRExC_state, OPEN, parno);
12224 if (!RExC_nestroot)
12225 RExC_nestroot = parno;
12226 if (RExC_open_parens && !RExC_open_parens[parno])
12228 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12229 "%*s%*s Setting open paren #%" IVdf " to %zu\n",
12230 22, "| |", (int)(depth * 2 + 1), "",
12232 RExC_open_parens[parno]= ret;
12235 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
12236 Set_Node_Offset(REGNODE_p(ret), RExC_parse); /* MJD */
12239 /* with RXf_PMf_NOCAPTURE treat (...) as (?:...) */
12248 /* Pick up the branches, linking them together. */
12249 parse_start = RExC_parse; /* MJD */
12250 br = regbranch(pRExC_state, &flags, 1, depth+1);
12252 /* branch_len = (paren != 0); */
12255 RETURN_FAIL_ON_RESTART(flags, flagp);
12256 FAIL2("panic: regbranch returned failure, flags=%#" UVxf, (UV) flags);
12258 if (*RExC_parse == '|') {
12259 if (RExC_use_BRANCHJ) {
12260 reginsert(pRExC_state, BRANCHJ, br, depth+1);
12263 reginsert(pRExC_state, BRANCH, br, depth+1);
12264 Set_Node_Length(REGNODE_p(br), paren != 0);
12265 Set_Node_Offset_To_R(br, parse_start-RExC_start);
12269 else if (paren == ':') {
12270 *flagp |= flags&SIMPLE;
12272 if (is_open) { /* Starts with OPEN. */
12273 if (! REGTAIL(pRExC_state, ret, br)) { /* OPEN -> first. */
12274 REQUIRE_BRANCHJ(flagp, 0);
12277 else if (paren != '?') /* Not Conditional */
12279 *flagp |= flags & (HASWIDTH | POSTPONED);
12281 while (*RExC_parse == '|') {
12282 if (RExC_use_BRANCHJ) {
12285 ender = reganode(pRExC_state, LONGJMP, 0);
12287 /* Append to the previous. */
12288 shut_gcc_up = REGTAIL(pRExC_state,
12289 REGNODE_OFFSET(NEXTOPER(NEXTOPER(REGNODE_p(lastbr)))),
12291 PERL_UNUSED_VAR(shut_gcc_up);
12293 nextchar(pRExC_state);
12294 if (freeze_paren) {
12295 if (RExC_npar > after_freeze)
12296 after_freeze = RExC_npar;
12297 RExC_npar = freeze_paren;
12299 br = regbranch(pRExC_state, &flags, 0, depth+1);
12302 RETURN_FAIL_ON_RESTART(flags, flagp);
12303 FAIL2("panic: regbranch returned failure, flags=%#" UVxf, (UV) flags);
12305 if (! REGTAIL(pRExC_state, lastbr, br)) { /* BRANCH -> BRANCH. */
12306 REQUIRE_BRANCHJ(flagp, 0);
12309 *flagp |= flags & (HASWIDTH | POSTPONED);
12312 if (have_branch || paren != ':') {
12315 /* Make a closing node, and hook it on the end. */
12318 ender = reg_node(pRExC_state, TAIL);
12321 ender = reganode(pRExC_state, CLOSE, parno);
12322 if ( RExC_close_parens ) {
12323 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12324 "%*s%*s Setting close paren #%" IVdf " to %zu\n",
12325 22, "| |", (int)(depth * 2 + 1), "",
12326 (IV)parno, ender));
12327 RExC_close_parens[parno]= ender;
12328 if (RExC_nestroot == parno)
12331 Set_Node_Offset(REGNODE_p(ender), RExC_parse+1); /* MJD */
12332 Set_Node_Length(REGNODE_p(ender), 1); /* MJD */
12335 ender = reg_node(pRExC_state, SRCLOSE);
12336 RExC_in_script_run = 0;
12346 *flagp &= ~HASWIDTH;
12348 case 't': /* aTomic */
12350 ender = reg_node(pRExC_state, SUCCEED);
12353 ender = reg_node(pRExC_state, END);
12354 assert(!RExC_end_op); /* there can only be one! */
12355 RExC_end_op = REGNODE_p(ender);
12356 if (RExC_close_parens) {
12357 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12358 "%*s%*s Setting close paren #0 (END) to %zu\n",
12359 22, "| |", (int)(depth * 2 + 1), "",
12362 RExC_close_parens[0]= ender;
12367 DEBUG_PARSE_MSG("lsbr");
12368 regprop(RExC_rx, RExC_mysv1, REGNODE_p(lastbr), NULL, pRExC_state);
12369 regprop(RExC_rx, RExC_mysv2, REGNODE_p(ender), NULL, pRExC_state);
12370 Perl_re_printf( aTHX_ "~ tying lastbr %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
12371 SvPV_nolen_const(RExC_mysv1),
12373 SvPV_nolen_const(RExC_mysv2),
12375 (IV)(ender - lastbr)
12378 if (! REGTAIL(pRExC_state, lastbr, ender)) {
12379 REQUIRE_BRANCHJ(flagp, 0);
12383 char is_nothing= 1;
12385 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
12387 /* Hook the tails of the branches to the closing node. */
12388 for (br = REGNODE_p(ret); br; br = regnext(br)) {
12389 const U8 op = PL_regkind[OP(br)];
12390 if (op == BRANCH) {
12391 if (! REGTAIL_STUDY(pRExC_state,
12392 REGNODE_OFFSET(NEXTOPER(br)),
12395 REQUIRE_BRANCHJ(flagp, 0);
12397 if ( OP(NEXTOPER(br)) != NOTHING
12398 || regnext(NEXTOPER(br)) != REGNODE_p(ender))
12401 else if (op == BRANCHJ) {
12402 bool shut_gcc_up = REGTAIL_STUDY(pRExC_state,
12403 REGNODE_OFFSET(NEXTOPER(NEXTOPER(br))),
12405 PERL_UNUSED_VAR(shut_gcc_up);
12406 /* for now we always disable this optimisation * /
12407 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
12408 || regnext(NEXTOPER(NEXTOPER(br))) != REGNODE_p(ender))
12414 regnode * ret_as_regnode = REGNODE_p(ret);
12415 br= PL_regkind[OP(ret_as_regnode)] != BRANCH
12416 ? regnext(ret_as_regnode)
12419 DEBUG_PARSE_MSG("NADA");
12420 regprop(RExC_rx, RExC_mysv1, ret_as_regnode,
12421 NULL, pRExC_state);
12422 regprop(RExC_rx, RExC_mysv2, REGNODE_p(ender),
12423 NULL, pRExC_state);
12424 Perl_re_printf( aTHX_ "~ converting ret %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
12425 SvPV_nolen_const(RExC_mysv1),
12426 (IV)REG_NODE_NUM(ret_as_regnode),
12427 SvPV_nolen_const(RExC_mysv2),
12433 if (OP(REGNODE_p(ender)) == TAIL) {
12435 RExC_emit= REGNODE_OFFSET(br) + 1;
12438 for ( opt= br + 1; opt < REGNODE_p(ender) ; opt++ )
12439 OP(opt)= OPTIMIZED;
12440 NEXT_OFF(br)= REGNODE_p(ender) - br;
12448 /* Even/odd or x=don't care: 010101x10x */
12449 static const char parens[] = "=!aA<,>Bbt";
12450 /* flag below is set to 0 up through 'A'; 1 for larger */
12452 if (paren && (p = strchr(parens, paren))) {
12453 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
12454 int flag = (p - parens) > 3;
12456 if (paren == '>' || paren == 't') {
12457 node = SUSPEND, flag = 0;
12460 reginsert(pRExC_state, node, ret, depth+1);
12461 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
12462 Set_Node_Offset(REGNODE_p(ret), parse_start + 1);
12463 FLAGS(REGNODE_p(ret)) = flag;
12464 if (! REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL)))
12466 REQUIRE_BRANCHJ(flagp, 0);
12471 /* Check for proper termination. */
12473 /* restore original flags, but keep (?p) and, if we've encountered
12474 * something in the parse that changes /d rules into /u, keep the /u */
12475 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
12476 if (DEPENDS_SEMANTICS && toUSE_UNI_CHARSET_NOT_DEPENDS) {
12477 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
12479 if (RExC_parse >= RExC_end || UCHARAT(RExC_parse) != ')') {
12480 RExC_parse = oregcomp_parse;
12481 vFAIL("Unmatched (");
12483 nextchar(pRExC_state);
12485 else if (!paren && RExC_parse < RExC_end) {
12486 if (*RExC_parse == ')') {
12488 vFAIL("Unmatched )");
12491 FAIL("Junk on end of regexp"); /* "Can't happen". */
12492 NOT_REACHED; /* NOTREACHED */
12495 if (after_freeze > RExC_npar)
12496 RExC_npar = after_freeze;
12498 RExC_in_lookaround = was_in_lookaround;
12504 - regbranch - one alternative of an | operator
12506 * Implements the concatenation operator.
12508 * On success, returns the offset at which any next node should be placed into
12509 * the regex engine program being compiled.
12511 * Returns 0 otherwise, setting flagp to RESTART_PARSE if the parse needs
12512 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to
12515 STATIC regnode_offset
12516 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
12518 regnode_offset ret;
12519 regnode_offset chain = 0;
12520 regnode_offset latest;
12521 I32 flags = 0, c = 0;
12522 DECLARE_AND_GET_RE_DEBUG_FLAGS;
12524 PERL_ARGS_ASSERT_REGBRANCH;
12526 DEBUG_PARSE("brnc");
12531 if (RExC_use_BRANCHJ)
12532 ret = reganode(pRExC_state, BRANCHJ, 0);
12534 ret = reg_node(pRExC_state, BRANCH);
12535 Set_Node_Length(REGNODE_p(ret), 1);
12539 *flagp = 0; /* Initialize. */
12541 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
12542 FALSE /* Don't force to /x */ );
12543 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
12544 flags &= ~TRYAGAIN;
12545 latest = regpiece(pRExC_state, &flags, depth+1);
12547 if (flags & TRYAGAIN)
12549 RETURN_FAIL_ON_RESTART(flags, flagp);
12550 FAIL2("panic: regpiece returned failure, flags=%#" UVxf, (UV) flags);
12554 *flagp |= flags&(HASWIDTH|POSTPONED);
12556 /* FIXME adding one for every branch after the first is probably
12557 * excessive now we have TRIE support. (hv) */
12559 if (! REGTAIL(pRExC_state, chain, latest)) {
12560 /* XXX We could just redo this branch, but figuring out what
12561 * bookkeeping needs to be reset is a pain, and it's likely
12562 * that other branches that goto END will also be too large */
12563 REQUIRE_BRANCHJ(flagp, 0);
12569 if (chain == 0) { /* Loop ran zero times. */
12570 chain = reg_node(pRExC_state, NOTHING);
12575 *flagp |= flags&SIMPLE;
12587 #ifndef PERL_IN_XSUB_RE
12589 Perl_regcurly(const char *s, const char *e, const char * result[5])
12591 /* This function matches a {m,n} quantifier. When called with a NULL final
12592 * argument, it simply parses the input from 's' up through 'e-1', and
12593 * returns a boolean as to whether or not this input is syntactically a
12594 * {m,n} quantifier.
12596 * When called with a non-NULL final parameter, and when the function
12597 * returns TRUE, it additionally stores information into the array
12598 * specified by that parameter about what it found in the parse. The
12599 * parameter must be a pointer into a 5 element array of 'const char *'
12600 * elements. The returned information is as follows:
12601 * result[RBRACE] points to the closing brace
12602 * result[MIN_S] points to the first byte of the lower bound
12603 * result[MIN_E] points to one beyond the final byte of the lower bound
12604 * result[MAX_S] points to the first byte of the upper bound
12605 * result[MAX_E] points to one beyond the final byte of the upper bound
12607 * If the quantifier is of the form {m,} (meaning an infinite upper
12608 * bound), result[MAX_E] is set to result[MAX_S]; what they actually point
12609 * to is irrelevant, just that it's the same place
12611 * If instead the quantifier is of the form {m} there is actually only
12612 * one bound, and both the upper and lower result[] elements are set to
12615 * This function checks only for syntactic validity; it leaves checking for
12616 * semantic validity and raising any diagnostics to the caller. This
12617 * function is called in multiple places to check for syntax, but only from
12618 * one for semantics. It makes it as simple as possible for the
12619 * syntax-only callers, while furnishing just enough information for the
12623 const char * min_start = NULL;
12624 const char * max_start = NULL;
12625 const char * min_end = NULL;
12626 const char * max_end = NULL;
12628 bool has_comma = FALSE;
12630 PERL_ARGS_ASSERT_REGCURLY;
12632 if (s >= e || *s++ != '{')
12635 while (s < e && isBLANK(*s)) {
12643 } while (s < e && isDIGIT(*s));
12647 while (s < e && isBLANK(*s)) {
12655 while (s < e && isBLANK(*s)) {
12663 } while (s < e && isDIGIT(*s));
12668 while (s < e && isBLANK(*s)) {
12671 /* Need at least one number */
12672 if (s >= e || *s != '}' || (! min_start && ! max_end)) {
12678 result[RBRACE] = s;
12680 result[MIN_S] = min_start;
12681 result[MIN_E] = min_end;
12684 result[MAX_S] = max_start;
12685 result[MAX_E] = max_end;
12688 /* Having no value after the comma is signalled by setting
12689 * start and end to the same value. What that value is isn't
12690 * relevant; NULL is chosen simply because it will fail if the
12691 * caller mistakenly uses it */
12692 result[MAX_S] = result[MAX_E] = NULL;
12695 else { /* No comma means lower and upper bounds are the same */
12696 result[MAX_S] = min_start;
12697 result[MAX_E] = min_end;
12706 S_get_quantifier_value(pTHX_ RExC_state_t *pRExC_state,
12707 const char * start, const char * end)
12709 /* This is a helper function for regpiece() to compute, given the
12710 * quantifier {m,n}, the value of either m or n, based on the starting
12711 * position 'start' in the string, through the byte 'end-1', returning it
12712 * if valid, and failing appropriately if not. It knows the restrictions
12713 * imposed on quantifier values */
12716 STATIC_ASSERT_DECL(REG_INFTY <= U32_MAX);
12718 PERL_ARGS_ASSERT_GET_QUANTIFIER_VALUE;
12720 if (grok_atoUV(start, &uv, &end)) {
12721 if (uv < REG_INFTY) { /* A valid, small-enough number */
12725 else if (*start == '0') { /* grok_atoUV() fails for only two reasons:
12726 leading zeros or overflow */
12727 RExC_parse = (char * ) end;
12729 /* Perhaps too generic a msg for what is only failure from having
12730 * leading zeros, but this is how it's always behaved. */
12731 vFAIL("Invalid quantifier in {,}");
12732 NOT_REACHED; /*NOTREACHED*/
12735 /* Here, found a quantifier, but was too large; either it overflowed or was
12736 * too big a legal number */
12737 RExC_parse = (char * ) end;
12738 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
12740 NOT_REACHED; /*NOTREACHED*/
12741 return U32_MAX; /* Perhaps some compilers will be expecting a return */
12745 - regpiece - something followed by possible quantifier * + ? {n,m}
12747 * Note that the branching code sequences used for ? and the general cases
12748 * of * and + are somewhat optimized: they use the same NOTHING node as
12749 * both the endmarker for their branch list and the body of the last branch.
12750 * It might seem that this node could be dispensed with entirely, but the
12751 * endmarker role is not redundant.
12753 * On success, returns the offset at which any next node should be placed into
12754 * the regex engine program being compiled.
12756 * Returns 0 otherwise, with *flagp set to indicate why:
12757 * TRYAGAIN if regatom() returns 0 with TRYAGAIN.
12758 * RESTART_PARSE if the parse needs to be restarted, or'd with
12759 * NEED_UTF8 if the pattern needs to be upgraded to UTF-8.
12761 STATIC regnode_offset
12762 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
12764 regnode_offset ret;
12767 const char * const origparse = RExC_parse;
12769 I32 max = REG_INFTY;
12770 #ifdef RE_TRACK_PATTERN_OFFSETS
12774 /* Save the original in case we change the emitted regop to a FAIL. */
12775 const regnode_offset orig_emit = RExC_emit;
12777 DECLARE_AND_GET_RE_DEBUG_FLAGS;
12779 PERL_ARGS_ASSERT_REGPIECE;
12781 DEBUG_PARSE("piec");
12783 ret = regatom(pRExC_state, &flags, depth+1);
12785 RETURN_FAIL_ON_RESTART_OR_FLAGS(flags, flagp, TRYAGAIN);
12786 FAIL2("panic: regatom returned failure, flags=%#" UVxf, (UV) flags);
12789 #ifdef RE_TRACK_PATTERN_OFFSETS
12790 parse_start = RExC_parse;
12795 const char * regcurly_return[5];
12798 nextchar(pRExC_state);
12803 nextchar(pRExC_state);
12808 nextchar(pRExC_state);
12812 case '{': /* A '{' may or may not indicate a quantifier; call regcurly()
12813 to determine which */
12814 if (regcurly(RExC_parse, RExC_end, regcurly_return)) {
12815 const char * min_start = regcurly_return[MIN_S];
12816 const char * min_end = regcurly_return[MIN_E];
12817 const char * max_start = regcurly_return[MAX_S];
12818 const char * max_end = regcurly_return[MAX_E];
12821 min = get_quantifier_value(pRExC_state, min_start, min_end);
12827 if (max_start == max_end) { /* Was of the form {m,} */
12830 else if (max_start == min_start) { /* Was of the form {m} */
12833 else { /* Was of the form {m,n} */
12834 assert(max_end >= max_start);
12836 max = get_quantifier_value(pRExC_state, max_start, max_end);
12839 RExC_parse = (char *) regcurly_return[RBRACE];
12840 nextchar(pRExC_state);
12842 if (max < min) { /* If can't match, warn and optimize to fail
12844 reginsert(pRExC_state, OPFAIL, orig_emit, depth+1);
12845 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
12846 NEXT_OFF(REGNODE_p(orig_emit)) =
12847 regarglen[OPFAIL] + NODE_STEP_REGNODE;
12850 else if (min == max && *RExC_parse == '?') {
12851 ckWARN2reg(RExC_parse + 1,
12852 "Useless use of greediness modifier '%c'",
12857 } /* End of is {m,n} */
12859 /* Here was a '{', but what followed it didn't form a quantifier. */
12865 NOT_REACHED; /*NOTREACHED*/
12868 /* Here we have a quantifier, and have calculated 'min' and 'max'.
12870 * Check and possibly adjust a zero width operand */
12871 if (! (flags & (HASWIDTH|POSTPONED))) {
12872 if (max > REG_INFTY/3) {
12873 if (origparse[0] == '\\' && origparse[1] == 'K') {
12875 "%" UTF8f " is forbidden - matches null string"
12877 UTF8fARG(UTF, (RExC_parse >= origparse
12878 ? RExC_parse - origparse
12882 ckWARN2reg(RExC_parse,
12883 "%" UTF8f " matches null string many times",
12884 UTF8fARG(UTF, (RExC_parse >= origparse
12885 ? RExC_parse - origparse
12891 /* There's no point in trying to match something 0 length more than
12892 * once except for extra side effects, which we don't have here since
12902 /* If this is a code block pass it up */
12903 *flagp |= (flags & POSTPONED);
12906 *flagp |= (flags & HASWIDTH);
12907 if (max == REG_INFTY)
12908 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
12911 /* 'SIMPLE' operands don't require full generality */
12912 if ((flags&SIMPLE)) {
12913 if (max == REG_INFTY) {
12915 if (UNLIKELY(RExC_pm_flags & PMf_WILDCARD)) {
12916 goto min0_maxINF_wildcard_forbidden;
12919 reginsert(pRExC_state, STAR, ret, depth+1);
12923 else if (min == 1) {
12924 reginsert(pRExC_state, PLUS, ret, depth+1);
12930 /* Here, SIMPLE, but not the '*' and '+' special cases */
12932 MARK_NAUGHTY_EXP(2, 2);
12933 reginsert(pRExC_state, CURLY, ret, depth+1);
12934 Set_Node_Offset(REGNODE_p(ret), parse_start+1); /* MJD */
12935 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
12937 else { /* not SIMPLE */
12938 const regnode_offset w = reg_node(pRExC_state, WHILEM);
12940 FLAGS(REGNODE_p(w)) = 0;
12941 if (! REGTAIL(pRExC_state, ret, w)) {
12942 REQUIRE_BRANCHJ(flagp, 0);
12944 if (RExC_use_BRANCHJ) {
12945 reginsert(pRExC_state, LONGJMP, ret, depth+1);
12946 reginsert(pRExC_state, NOTHING, ret, depth+1);
12947 NEXT_OFF(REGNODE_p(ret)) = 3; /* Go over LONGJMP. */
12949 reginsert(pRExC_state, CURLYX, ret, depth+1);
12951 Set_Node_Offset(REGNODE_p(ret), parse_start+1);
12952 Set_Node_Length(REGNODE_p(ret),
12953 op == '{' ? (RExC_parse - parse_start) : 1);
12955 if (RExC_use_BRANCHJ)
12956 NEXT_OFF(REGNODE_p(ret)) = 3; /* Go over NOTHING to
12958 if (! REGTAIL(pRExC_state, ret, reg_node(pRExC_state,
12961 REQUIRE_BRANCHJ(flagp, 0);
12963 RExC_whilem_seen++;
12964 MARK_NAUGHTY_EXP(1, 4); /* compound interest */
12967 /* Finish up the CURLY/CURLYX case */
12968 FLAGS(REGNODE_p(ret)) = 0;
12970 ARG1_SET(REGNODE_p(ret), (U16)min);
12971 ARG2_SET(REGNODE_p(ret), (U16)max);
12975 /* Process any greediness modifiers */
12976 if (*RExC_parse == '?') {
12977 nextchar(pRExC_state);
12978 reginsert(pRExC_state, MINMOD, ret, depth+1);
12979 if (! REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE)) {
12980 REQUIRE_BRANCHJ(flagp, 0);
12983 else if (*RExC_parse == '+') {
12984 regnode_offset ender;
12985 nextchar(pRExC_state);
12986 ender = reg_node(pRExC_state, SUCCEED);
12987 if (! REGTAIL(pRExC_state, ret, ender)) {
12988 REQUIRE_BRANCHJ(flagp, 0);
12990 reginsert(pRExC_state, SUSPEND, ret, depth+1);
12991 ender = reg_node(pRExC_state, TAIL);
12992 if (! REGTAIL(pRExC_state, ret, ender)) {
12993 REQUIRE_BRANCHJ(flagp, 0);
12997 /* Forbid extra quantifiers */
12998 if (isQUANTIFIER(RExC_parse, RExC_end)) {
13000 vFAIL("Nested quantifiers");
13005 min0_maxINF_wildcard_forbidden:
13007 /* Here we are in a wildcard match, and the minimum match length is 0, and
13008 * the max could be infinity. This is currently forbidden. The only
13009 * reason is to make it harder to write patterns that take a long long time
13010 * to halt, and because the use of this construct isn't necessary in
13011 * matching Unicode property values */
13013 /* diag_listed_as: Use of %s is not allowed in Unicode property wildcard
13014 subpatterns in regex; marked by <-- HERE in m/%s/
13016 vFAIL("Use of quantifier '*' is not allowed in Unicode property wildcard"
13019 /* Note, don't need to worry about the input being '{0,}', as a '}' isn't
13020 * legal at all in wildcards, so can't get this far */
13022 NOT_REACHED; /*NOTREACHED*/
13026 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state,
13027 regnode_offset * node_p,
13035 /* This routine teases apart the various meanings of \N and returns
13036 * accordingly. The input parameters constrain which meaning(s) is/are valid
13037 * in the current context.
13039 * Exactly one of <node_p> and <code_point_p> must be non-NULL.
13041 * If <code_point_p> is not NULL, the context is expecting the result to be a
13042 * single code point. If this \N instance turns out to a single code point,
13043 * the function returns TRUE and sets *code_point_p to that code point.
13045 * If <node_p> is not NULL, the context is expecting the result to be one of
13046 * the things representable by a regnode. If this \N instance turns out to be
13047 * one such, the function generates the regnode, returns TRUE and sets *node_p
13048 * to point to the offset of that regnode into the regex engine program being
13051 * If this instance of \N isn't legal in any context, this function will
13052 * generate a fatal error and not return.
13054 * On input, RExC_parse should point to the first char following the \N at the
13055 * time of the call. On successful return, RExC_parse will have been updated
13056 * to point to just after the sequence identified by this routine. Also
13057 * *flagp has been updated as needed.
13059 * When there is some problem with the current context and this \N instance,
13060 * the function returns FALSE, without advancing RExC_parse, nor setting
13061 * *node_p, nor *code_point_p, nor *flagp.
13063 * If <cp_count> is not NULL, the caller wants to know the length (in code
13064 * points) that this \N sequence matches. This is set, and the input is
13065 * parsed for errors, even if the function returns FALSE, as detailed below.
13067 * There are 6 possibilities here, as detailed in the next 6 paragraphs.
13069 * Probably the most common case is for the \N to specify a single code point.
13070 * *cp_count will be set to 1, and *code_point_p will be set to that code
13073 * Another possibility is for the input to be an empty \N{}. This is no
13074 * longer accepted, and will generate a fatal error.
13076 * Another possibility is for a custom charnames handler to be in effect which
13077 * translates the input name to an empty string. *cp_count will be set to 0.
13078 * *node_p will be set to a generated NOTHING node.
13080 * Still another possibility is for the \N to mean [^\n]. *cp_count will be
13081 * set to 0. *node_p will be set to a generated REG_ANY node.
13083 * The fifth possibility is that \N resolves to a sequence of more than one
13084 * code points. *cp_count will be set to the number of code points in the
13085 * sequence. *node_p will be set to a generated node returned by this
13086 * function calling S_reg().
13088 * The sixth and final possibility is that it is premature to be calling this
13089 * function; the parse needs to be restarted. This can happen when this
13090 * changes from /d to /u rules, or when the pattern needs to be upgraded to
13091 * UTF-8. The latter occurs only when the fifth possibility would otherwise
13092 * be in effect, and is because one of those code points requires the pattern
13093 * to be recompiled as UTF-8. The function returns FALSE, and sets the
13094 * RESTART_PARSE and NEED_UTF8 flags in *flagp, as appropriate. When this
13095 * happens, the caller needs to desist from continuing parsing, and return
13096 * this information to its caller. This is not set for when there is only one
13097 * code point, as this can be called as part of an ANYOF node, and they can
13098 * store above-Latin1 code points without the pattern having to be in UTF-8.
13100 * For non-single-quoted regexes, the tokenizer has resolved character and
13101 * sequence names inside \N{...} into their Unicode values, normalizing the
13102 * result into what we should see here: '\N{U+c1.c2...}', where c1... are the
13103 * hex-represented code points in the sequence. This is done there because
13104 * the names can vary based on what charnames pragma is in scope at the time,
13105 * so we need a way to take a snapshot of what they resolve to at the time of
13106 * the original parse. [perl #56444].
13108 * That parsing is skipped for single-quoted regexes, so here we may get
13109 * '\N{NAME}', which is parsed now. If the single-quoted regex is something
13110 * like '\N{U+41}', that code point is Unicode, and has to be translated into
13111 * the native character set for non-ASCII platforms. The other possibilities
13112 * are already native, so no translation is done. */
13114 char * endbrace; /* points to '}' following the name */
13115 char * e; /* points to final non-blank before endbrace */
13116 char* p = RExC_parse; /* Temporary */
13118 SV * substitute_parse = NULL;
13123 DECLARE_AND_GET_RE_DEBUG_FLAGS;
13125 PERL_ARGS_ASSERT_GROK_BSLASH_N;
13127 assert(cBOOL(node_p) ^ cBOOL(code_point_p)); /* Exactly one should be set */
13128 assert(! (node_p && cp_count)); /* At most 1 should be set */
13130 if (cp_count) { /* Initialize return for the most common case */
13134 /* The [^\n] meaning of \N ignores spaces and comments under the /x
13135 * modifier. The other meanings do not (except blanks adjacent to and
13136 * within the braces), so use a temporary until we find out which we are
13137 * being called with */
13138 skip_to_be_ignored_text(pRExC_state, &p,
13139 FALSE /* Don't force to /x */ );
13141 /* Disambiguate between \N meaning a named character versus \N meaning
13142 * [^\n]. The latter is assumed when the {...} following the \N is a legal
13143 * quantifier, or if there is no '{' at all */
13144 if (*p != '{' || regcurly(p, RExC_end, NULL)) {
13154 *node_p = reg_node(pRExC_state, REG_ANY);
13155 *flagp |= HASWIDTH|SIMPLE;
13157 Set_Node_Length(REGNODE_p(*(node_p)), 1); /* MJD */
13161 /* The test above made sure that the next real character is a '{', but
13162 * under the /x modifier, it could be separated by space (or a comment and
13163 * \n) and this is not allowed (for consistency with \x{...} and the
13164 * tokenizer handling of \N{NAME}). */
13165 if (*RExC_parse != '{') {
13166 vFAIL("Missing braces on \\N{}");
13169 RExC_parse++; /* Skip past the '{' */
13171 endbrace = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
13172 if (! endbrace) { /* no trailing brace */
13173 vFAIL2("Missing right brace on \\%c{}", 'N');
13176 /* Here, we have decided it should be a named character or sequence. These
13177 * imply Unicode semantics */
13178 REQUIRE_UNI_RULES(flagp, FALSE);
13180 /* \N{_} is what toke.c returns to us to indicate a name that evaluates to
13181 * nothing at all (not allowed under strict) */
13182 if (endbrace - RExC_parse == 1 && *RExC_parse == '_') {
13183 RExC_parse = endbrace;
13185 RExC_parse++; /* Position after the "}" */
13186 vFAIL("Zero length \\N{}");
13192 nextchar(pRExC_state);
13197 *node_p = reg_node(pRExC_state, NOTHING);
13201 while (isBLANK(*RExC_parse)) {
13206 while (RExC_parse < e && isBLANK(*(e-1))) {
13210 if (e - RExC_parse < 2 || ! strBEGINs(RExC_parse, "U+")) {
13212 /* Here, the name isn't of the form U+.... This can happen if the
13213 * pattern is single-quoted, so didn't get evaluated in toke.c. Now
13214 * is the time to find out what the name means */
13216 const STRLEN name_len = e - RExC_parse;
13217 SV * value_sv; /* What does this name evaluate to */
13219 const U8 * value; /* string of name's value */
13220 STRLEN value_len; /* and its length */
13222 /* RExC_unlexed_names is a hash of names that weren't evaluated by
13223 * toke.c, and their values. Make sure is initialized */
13224 if (! RExC_unlexed_names) {
13225 RExC_unlexed_names = newHV();
13228 /* If we have already seen this name in this pattern, use that. This
13229 * allows us to only call the charnames handler once per name per
13230 * pattern. A broken or malicious handler could return something
13231 * different each time, which could cause the results to vary depending
13232 * on if something gets added or subtracted from the pattern that
13233 * causes the number of passes to change, for example */
13234 if ((value_svp = hv_fetch(RExC_unlexed_names, RExC_parse,
13237 value_sv = *value_svp;
13239 else { /* Otherwise we have to go out and get the name */
13240 const char * error_msg = NULL;
13241 value_sv = get_and_check_backslash_N_name(RExC_parse, e,
13245 RExC_parse = endbrace;
13249 /* If no error message, should have gotten a valid return */
13252 /* Save the name's meaning for later use */
13253 if (! hv_store(RExC_unlexed_names, RExC_parse, name_len,
13256 Perl_croak(aTHX_ "panic: hv_store() unexpectedly failed");
13260 /* Here, we have the value the name evaluates to in 'value_sv' */
13261 value = (U8 *) SvPV(value_sv, value_len);
13263 /* See if the result is one code point vs 0 or multiple */
13264 if (inRANGE(value_len, 1, ((UV) SvUTF8(value_sv)
13268 /* Here, exactly one code point. If that isn't what is wanted,
13270 if (! code_point_p) {
13275 /* Convert from string to numeric code point */
13276 *code_point_p = (SvUTF8(value_sv))
13277 ? valid_utf8_to_uvchr(value, NULL)
13280 /* Have parsed this entire single code point \N{...}. *cp_count
13281 * has already been set to 1, so don't do it again. */
13282 RExC_parse = endbrace;
13283 nextchar(pRExC_state);
13285 } /* End of is a single code point */
13287 /* Count the code points, if caller desires. The API says to do this
13288 * even if we will later return FALSE */
13292 *cp_count = (SvUTF8(value_sv))
13293 ? utf8_length(value, value + value_len)
13297 /* Fail if caller doesn't want to handle a multi-code-point sequence.
13298 * But don't back the pointer up if the caller wants to know how many
13299 * code points there are (they need to handle it themselves in this
13308 /* Convert this to a sub-pattern of the form "(?: ... )", and then call
13309 * reg recursively to parse it. That way, it retains its atomicness,
13310 * while not having to worry about any special handling that some code
13311 * points may have. */
13313 substitute_parse = newSVpvs("?:");
13314 sv_catsv(substitute_parse, value_sv);
13315 sv_catpv(substitute_parse, ")");
13317 /* The value should already be native, so no need to convert on EBCDIC
13319 assert(! RExC_recode_x_to_native);
13322 else { /* \N{U+...} */
13323 Size_t count = 0; /* code point count kept internally */
13325 /* We can get to here when the input is \N{U+...} or when toke.c has
13326 * converted a name to the \N{U+...} form. This include changing a
13327 * name that evaluates to multiple code points to \N{U+c1.c2.c3 ...} */
13329 RExC_parse += 2; /* Skip past the 'U+' */
13331 /* Code points are separated by dots. The '}' terminates the whole
13334 do { /* Loop until the ending brace */
13335 I32 flags = PERL_SCAN_SILENT_OVERFLOW
13336 | PERL_SCAN_SILENT_ILLDIGIT
13337 | PERL_SCAN_NOTIFY_ILLDIGIT
13338 | PERL_SCAN_ALLOW_MEDIAL_UNDERSCORES
13339 | PERL_SCAN_DISALLOW_PREFIX;
13340 STRLEN len = e - RExC_parse;
13342 char * start_digit = RExC_parse;
13343 UV cp = grok_hex(RExC_parse, &len, &flags, &overflow_value);
13348 vFAIL("Invalid hexadecimal number in \\N{U+...}");
13353 if (cp > MAX_LEGAL_CP) {
13354 vFAIL(form_cp_too_large_msg(16, start_digit, len, 0));
13357 if (RExC_parse >= e) { /* Got to the closing '}' */
13362 /* Here, is a single code point; fail if doesn't want that */
13363 if (! code_point_p) {
13368 /* A single code point is easy to handle; just return it */
13369 *code_point_p = UNI_TO_NATIVE(cp);
13370 RExC_parse = endbrace;
13371 nextchar(pRExC_state);
13375 /* Here, the parse stopped bfore the ending brace. This is legal
13376 * only if that character is a dot separating code points, like a
13377 * multiple character sequence (of the form "\N{U+c1.c2. ... }".
13378 * So the next character must be a dot (and the one after that
13379 * can't be the ending brace, or we'd have something like
13382 if (*RExC_parse != '.' || RExC_parse + 1 >= e) {
13383 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
13384 ? UTF8SKIP(RExC_parse)
13386 RExC_parse = MIN(e, RExC_parse);/* Guard against malformed utf8
13391 /* Here, looks like its really a multiple character sequence. Fail
13392 * if that's not what the caller wants. But continue with counting
13393 * and error checking if they still want a count */
13394 if (! node_p && ! cp_count) {
13398 /* What is done here is to convert this to a sub-pattern of the
13399 * form \x{char1}\x{char2}... and then call reg recursively to
13400 * parse it (enclosing in "(?: ... )" ). That way, it retains its
13401 * atomicness, while not having to worry about special handling
13402 * that some code points may have. We don't create a subpattern,
13403 * but go through the motions of code point counting and error
13404 * checking, if the caller doesn't want a node returned. */
13406 if (node_p && ! substitute_parse) {
13407 substitute_parse = newSVpvs("?:");
13413 /* Convert to notation the rest of the code understands */
13414 sv_catpvs(substitute_parse, "\\x{");
13415 sv_catpvn(substitute_parse, start_digit,
13416 RExC_parse - start_digit);
13417 sv_catpvs(substitute_parse, "}");
13420 /* Move to after the dot (or ending brace the final time through.)
13425 } while (RExC_parse < e);
13427 if (! node_p) { /* Doesn't want the node */
13434 sv_catpvs(substitute_parse, ")");
13436 /* The values are Unicode, and therefore have to be converted to native
13437 * on a non-Unicode (meaning non-ASCII) platform. */
13438 SET_recode_x_to_native(1);
13441 /* Here, we have the string the name evaluates to, ready to be parsed,
13442 * stored in 'substitute_parse' as a series of valid "\x{...}\x{...}"
13443 * constructs. This can be called from within a substitute parse already.
13444 * The error reporting mechanism doesn't work for 2 levels of this, but the
13445 * code above has validated this new construct, so there should be no
13446 * errors generated by the below. And this isn't an exact copy, so the
13447 * mechanism to seamlessly deal with this won't work, so turn off warnings
13449 save_start = RExC_start;
13450 orig_end = RExC_end;
13452 RExC_parse = RExC_start = SvPVX(substitute_parse);
13453 RExC_end = RExC_parse + SvCUR(substitute_parse);
13454 TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE;
13456 *node_p = reg(pRExC_state, 1, &flags, depth+1);
13458 /* Restore the saved values */
13460 RExC_start = save_start;
13461 RExC_parse = endbrace;
13462 RExC_end = orig_end;
13463 SET_recode_x_to_native(0);
13465 SvREFCNT_dec_NN(substitute_parse);
13468 RETURN_FAIL_ON_RESTART(flags, flagp);
13469 FAIL2("panic: reg returned failure to grok_bslash_N, flags=%#" UVxf,
13472 *flagp |= flags&(HASWIDTH|SIMPLE|POSTPONED);
13474 nextchar(pRExC_state);
13481 S_compute_EXACTish(RExC_state_t *pRExC_state)
13485 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
13493 op = get_regex_charset(RExC_flags);
13494 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
13495 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
13496 been, so there is no hole */
13499 return op + EXACTF;
13502 /* Parse backref decimal value, unless it's too big to sensibly be a backref,
13503 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
13506 S_backref_value(char *p, char *e)
13508 const char* endptr = e;
13510 if (grok_atoUV(p, &val, &endptr) && val <= I32_MAX)
13517 - regatom - the lowest level
13519 Try to identify anything special at the start of the current parse position.
13520 If there is, then handle it as required. This may involve generating a
13521 single regop, such as for an assertion; or it may involve recursing, such as
13522 to handle a () structure.
13524 If the string doesn't start with something special then we gobble up
13525 as much literal text as we can. If we encounter a quantifier, we have to
13526 back off the final literal character, as that quantifier applies to just it
13527 and not to the whole string of literals.
13529 Once we have been able to handle whatever type of thing started the
13530 sequence, we return the offset into the regex engine program being compiled
13531 at which any next regnode should be placed.
13533 Returns 0, setting *flagp to TRYAGAIN if reg() returns 0 with TRYAGAIN.
13534 Returns 0, setting *flagp to RESTART_PARSE if the parse needs to be
13535 restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
13536 Otherwise does not return 0.
13538 Note: we have to be careful with escapes, as they can be both literal
13539 and special, and in the case of \10 and friends, context determines which.
13541 A summary of the code structure is:
13543 switch (first_byte) {
13544 cases for each special:
13545 handle this special;
13548 switch (2nd byte) {
13549 cases for each unambiguous special:
13550 handle this special;
13552 cases for each ambigous special/literal:
13554 if (special) handle here
13556 default: // unambiguously literal:
13559 default: // is a literal char
13562 create EXACTish node for literal;
13563 while (more input and node isn't full) {
13564 switch (input_byte) {
13565 cases for each special;
13566 make sure parse pointer is set so that the next call to
13567 regatom will see this special first
13568 goto loopdone; // EXACTish node terminated by prev. char
13570 append char to EXACTISH node;
13572 get next input byte;
13576 return the generated node;
13578 Specifically there are two separate switches for handling
13579 escape sequences, with the one for handling literal escapes requiring
13580 a dummy entry for all of the special escapes that are actually handled
13585 STATIC regnode_offset
13586 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
13588 regnode_offset ret = 0;
13594 DECLARE_AND_GET_RE_DEBUG_FLAGS;
13596 *flagp = 0; /* Initialize. */
13598 DEBUG_PARSE("atom");
13600 PERL_ARGS_ASSERT_REGATOM;
13603 parse_start = RExC_parse;
13604 assert(RExC_parse < RExC_end);
13605 switch ((U8)*RExC_parse) {
13607 RExC_seen_zerolen++;
13608 nextchar(pRExC_state);
13609 if (RExC_flags & RXf_PMf_MULTILINE)
13610 ret = reg_node(pRExC_state, MBOL);
13612 ret = reg_node(pRExC_state, SBOL);
13613 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13616 nextchar(pRExC_state);
13618 RExC_seen_zerolen++;
13619 if (RExC_flags & RXf_PMf_MULTILINE)
13620 ret = reg_node(pRExC_state, MEOL);
13622 ret = reg_node(pRExC_state, SEOL);
13623 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13626 nextchar(pRExC_state);
13627 if (RExC_flags & RXf_PMf_SINGLELINE)
13628 ret = reg_node(pRExC_state, SANY);
13630 ret = reg_node(pRExC_state, REG_ANY);
13631 *flagp |= HASWIDTH|SIMPLE;
13633 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13637 char * const oregcomp_parse = ++RExC_parse;
13638 ret = regclass(pRExC_state, flagp, depth+1,
13639 FALSE, /* means parse the whole char class */
13640 TRUE, /* allow multi-char folds */
13641 FALSE, /* don't silence non-portable warnings. */
13642 (bool) RExC_strict,
13643 TRUE, /* Allow an optimized regnode result */
13646 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13647 FAIL2("panic: regclass returned failure to regatom, flags=%#" UVxf,
13650 if (*RExC_parse != ']') {
13651 RExC_parse = oregcomp_parse;
13652 vFAIL("Unmatched [");
13654 nextchar(pRExC_state);
13655 Set_Node_Length(REGNODE_p(ret), RExC_parse - oregcomp_parse + 1); /* MJD */
13659 nextchar(pRExC_state);
13660 ret = reg(pRExC_state, 2, &flags, depth+1);
13662 if (flags & TRYAGAIN) {
13663 if (RExC_parse >= RExC_end) {
13664 /* Make parent create an empty node if needed. */
13665 *flagp |= TRYAGAIN;
13670 RETURN_FAIL_ON_RESTART(flags, flagp);
13671 FAIL2("panic: reg returned failure to regatom, flags=%#" UVxf,
13674 *flagp |= flags&(HASWIDTH|SIMPLE|POSTPONED);
13678 if (flags & TRYAGAIN) {
13679 *flagp |= TRYAGAIN;
13682 vFAIL("Internal urp");
13683 /* Supposed to be caught earlier. */
13689 vFAIL("Quantifier follows nothing");
13694 This switch handles escape sequences that resolve to some kind
13695 of special regop and not to literal text. Escape sequences that
13696 resolve to literal text are handled below in the switch marked
13699 Every entry in this switch *must* have a corresponding entry
13700 in the literal escape switch. However, the opposite is not
13701 required, as the default for this switch is to jump to the
13702 literal text handling code.
13705 switch ((U8)*RExC_parse) {
13706 /* Special Escapes */
13708 RExC_seen_zerolen++;
13709 /* Under wildcards, this is changed to match \n; should be
13710 * invisible to the user, as they have to compile under /m */
13711 if (RExC_pm_flags & PMf_WILDCARD) {
13712 ret = reg_node(pRExC_state, MBOL);
13715 ret = reg_node(pRExC_state, SBOL);
13716 /* SBOL is shared with /^/ so we set the flags so we can tell
13717 * /\A/ from /^/ in split. */
13718 FLAGS(REGNODE_p(ret)) = 1;
13720 goto finish_meta_pat;
13722 if (RExC_pm_flags & PMf_WILDCARD) {
13724 /* diag_listed_as: Use of %s is not allowed in Unicode property
13725 wildcard subpatterns in regex; marked by <-- HERE in m/%s/
13727 vFAIL("Use of '\\G' is not allowed in Unicode property"
13728 " wildcard subpatterns");
13730 ret = reg_node(pRExC_state, GPOS);
13731 RExC_seen |= REG_GPOS_SEEN;
13732 goto finish_meta_pat;
13734 if (!RExC_in_lookaround) {
13735 RExC_seen_zerolen++;
13736 ret = reg_node(pRExC_state, KEEPS);
13737 /* XXX:dmq : disabling in-place substitution seems to
13738 * be necessary here to avoid cases of memory corruption, as
13739 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
13741 RExC_seen |= REG_LOOKBEHIND_SEEN;
13742 goto finish_meta_pat;
13745 ++RExC_parse; /* advance past the 'K' */
13746 vFAIL("\\K not permitted in lookahead/lookbehind");
13749 if (RExC_pm_flags & PMf_WILDCARD) {
13750 /* See comment under \A above */
13751 ret = reg_node(pRExC_state, MEOL);
13754 ret = reg_node(pRExC_state, SEOL);
13756 RExC_seen_zerolen++; /* Do not optimize RE away */
13757 goto finish_meta_pat;
13759 if (RExC_pm_flags & PMf_WILDCARD) {
13760 /* See comment under \A above */
13761 ret = reg_node(pRExC_state, MEOL);
13764 ret = reg_node(pRExC_state, EOS);
13766 RExC_seen_zerolen++; /* Do not optimize RE away */
13767 goto finish_meta_pat;
13769 vFAIL("\\C no longer supported");
13771 ret = reg_node(pRExC_state, CLUMP);
13772 *flagp |= HASWIDTH;
13773 goto finish_meta_pat;
13781 regex_charset charset = get_regex_charset(RExC_flags);
13783 RExC_seen_zerolen++;
13784 RExC_seen |= REG_LOOKBEHIND_SEEN;
13785 op = BOUND + charset;
13787 if (RExC_parse >= RExC_end || *(RExC_parse + 1) != '{') {
13788 flags = TRADITIONAL_BOUND;
13789 if (op > BOUNDA) { /* /aa is same as /a */
13795 char name = *RExC_parse;
13796 char * endbrace = (char *) memchr(RExC_parse, '}',
13797 RExC_end - RExC_parse);
13798 char * e = endbrace;
13803 vFAIL2("Missing right brace on \\%c{}", name);
13806 while (isBLANK(*RExC_parse)) {
13810 while (RExC_parse < e && isBLANK(*(e - 1))) {
13814 if (e == RExC_parse) {
13815 RExC_parse = endbrace + 1; /* After the '}' */
13816 vFAIL2("Empty \\%c{}", name);
13819 length = e - RExC_parse;
13821 switch (*RExC_parse) {
13824 && (memNEs(RExC_parse + 1, length - 1, "cb")))
13826 goto bad_bound_type;
13831 if (length != 2 || *(RExC_parse + 1) != 'b') {
13832 goto bad_bound_type;
13837 if (length != 2 || *(RExC_parse + 1) != 'b') {
13838 goto bad_bound_type;
13843 if (length != 2 || *(RExC_parse + 1) != 'b') {
13844 goto bad_bound_type;
13852 "'%" UTF8f "' is an unknown bound type",
13853 UTF8fARG(UTF, length, e - length));
13854 NOT_REACHED; /*NOTREACHED*/
13856 RExC_parse = endbrace;
13857 REQUIRE_UNI_RULES(flagp, 0);
13862 else if (op >= BOUNDA) { /* /aa is same as /a */
13866 /* Don't have to worry about UTF-8, in this message because
13867 * to get here the contents of the \b must be ASCII */
13868 ckWARN4reg(RExC_parse + 1, /* Include the '}' in msg */
13869 "Using /u for '%.*s' instead of /%s",
13871 endbrace - length + 1,
13872 (charset == REGEX_ASCII_RESTRICTED_CHARSET)
13873 ? ASCII_RESTRICT_PAT_MODS
13874 : ASCII_MORE_RESTRICT_PAT_MODS);
13879 RExC_seen_d_op = TRUE;
13881 else if (op == BOUNDL) {
13882 RExC_contains_locale = 1;
13886 op += NBOUND - BOUND;
13889 ret = reg_node(pRExC_state, op);
13890 FLAGS(REGNODE_p(ret)) = flags;
13892 goto finish_meta_pat;
13896 ret = reg_node(pRExC_state, LNBREAK);
13897 *flagp |= HASWIDTH|SIMPLE;
13898 goto finish_meta_pat;
13912 /* These all have the same meaning inside [brackets], and it knows
13913 * how to do the best optimizations for them. So, pretend we found
13914 * these within brackets, and let it do the work */
13917 ret = regclass(pRExC_state, flagp, depth+1,
13918 TRUE, /* means just parse this element */
13919 FALSE, /* don't allow multi-char folds */
13920 FALSE, /* don't silence non-portable warnings. It
13921 would be a bug if these returned
13923 (bool) RExC_strict,
13924 TRUE, /* Allow an optimized regnode result */
13926 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13927 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
13928 * multi-char folds are allowed. */
13930 FAIL2("panic: regclass returned failure to regatom, flags=%#" UVxf,
13933 RExC_parse--; /* regclass() leaves this one too far ahead */
13936 /* The escapes above that don't take a parameter can't be
13937 * followed by a '{'. But 'pX', 'p{foo}' and
13938 * correspondingly 'P' can be */
13939 if ( RExC_parse - parse_start == 1
13940 && UCHARAT(RExC_parse + 1) == '{'
13941 && UNLIKELY(! regcurly(RExC_parse + 1, RExC_end, NULL)))
13944 vFAIL("Unescaped left brace in regex is illegal here");
13946 Set_Node_Offset(REGNODE_p(ret), parse_start);
13947 Set_Node_Length(REGNODE_p(ret), RExC_parse - parse_start + 1); /* MJD */
13948 nextchar(pRExC_state);
13951 /* Handle \N, \N{} and \N{NAMED SEQUENCE} (the latter meaning the
13952 * \N{...} evaluates to a sequence of more than one code points).
13953 * The function call below returns a regnode, which is our result.
13954 * The parameters cause it to fail if the \N{} evaluates to a
13955 * single code point; we handle those like any other literal. The
13956 * reason that the multicharacter case is handled here and not as
13957 * part of the EXACtish code is because of quantifiers. In
13958 * /\N{BLAH}+/, the '+' applies to the whole thing, and doing it
13959 * this way makes that Just Happen. dmq.
13960 * join_exact() will join this up with adjacent EXACTish nodes
13961 * later on, if appropriate. */
13963 if (grok_bslash_N(pRExC_state,
13964 &ret, /* Want a regnode returned */
13965 NULL, /* Fail if evaluates to a single code
13967 NULL, /* Don't need a count of how many code
13976 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13978 /* Here, evaluates to a single code point. Go get that */
13979 RExC_parse = parse_start;
13982 case 'k': /* Handle \k<NAME> and \k'NAME' and \k{NAME} */
13983 parse_named_seq: /* Also handle non-numeric \g{...} */
13986 if ( RExC_parse >= RExC_end - 1
13987 || (( ch = RExC_parse[1]) != '<'
13992 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
13993 vFAIL2("Sequence %.2s... not terminated", parse_start);
13997 while (isBLANK(*RExC_parse)) {
14001 ret = handle_named_backref(pRExC_state,
14013 case '1': case '2': case '3': case '4':
14014 case '5': case '6': case '7': case '8': case '9':
14017 char * endbrace = NULL;
14018 char * s = RExC_parse;
14019 char * e = RExC_end;
14026 endbrace = (char *) memchr(s, '}', RExC_end - s);
14029 /* Missing '}'. Position after the number to give
14030 * a better indication to the user of where the
14037 /* If it looks to be a name and not a number, go
14038 * handle it there */
14039 if (! isDIGIT(*s)) {
14040 goto parse_named_seq;
14045 } while isDIGIT(*s);
14048 vFAIL("Unterminated \\g{...} pattern");
14051 s++; /* Past the '{' */
14053 while (isBLANK(*s)) {
14057 /* Ignore trailing blanks */
14059 while (s < e && isBLANK(*(e - 1))) {
14064 /* Here, have isolated the meat of the construct from any
14065 * surrounding braces */
14072 if (endbrace && !isDIGIT(*s)) {
14073 goto parse_named_seq;
14077 num = S_backref_value(RExC_parse, RExC_end);
14079 vFAIL("Reference to invalid group 0");
14080 else if (num == I32_MAX) {
14081 if (isDIGIT(*RExC_parse))
14082 vFAIL("Reference to nonexistent group");
14084 vFAIL("Unterminated \\g... pattern");
14088 num = RExC_npar - num;
14090 vFAIL("Reference to nonexistent or unclosed group");
14094 num = S_backref_value(RExC_parse, RExC_end);
14095 /* bare \NNN might be backref or octal - if it is larger
14096 * than or equal RExC_npar then it is assumed to be an
14097 * octal escape. Note RExC_npar is +1 from the actual
14098 * number of parens. */
14099 /* Note we do NOT check if num == I32_MAX here, as that is
14100 * handled by the RExC_npar check */
14102 if ( /* any numeric escape < 10 is always a backref */
14104 /* any numeric escape < RExC_npar is a backref */
14105 && num >= RExC_npar
14106 /* cannot be an octal escape if it starts with [89]
14108 && ! inRANGE(*RExC_parse, '8', '9')
14110 /* Probably not meant to be a backref, instead likely
14111 * to be an octal character escape, e.g. \35 or \777.
14112 * The above logic should make it obvious why using
14113 * octal escapes in patterns is problematic. - Yves */
14114 RExC_parse = parse_start;
14119 /* At this point RExC_parse points at a numeric escape like
14120 * \12 or \88 or the digits in \g{34} or \g34 or something
14121 * similar, which we should NOT treat as an octal escape. It
14122 * may or may not be a valid backref escape. For instance
14123 * \88888888 is unlikely to be a valid backref.
14125 * We've already figured out what value the digits represent.
14126 * Now, move the parse to beyond them. */
14128 RExC_parse = endbrace + 1;
14130 else while (isDIGIT(*RExC_parse)) {
14134 if (num >= (I32)RExC_npar) {
14136 /* It might be a forward reference; we can't fail until we
14137 * know, by completing the parse to get all the groups, and
14138 * then reparsing */
14139 if (ALL_PARENS_COUNTED) {
14140 if (num >= RExC_total_parens) {
14141 vFAIL("Reference to nonexistent group");
14145 REQUIRE_PARENS_PASS;
14149 ret = reganode(pRExC_state,
14152 : (ASCII_FOLD_RESTRICTED)
14154 : (AT_LEAST_UNI_SEMANTICS)
14160 if (OP(REGNODE_p(ret)) == REFF) {
14161 RExC_seen_d_op = TRUE;
14163 *flagp |= HASWIDTH;
14165 /* override incorrect value set in reganode MJD */
14166 Set_Node_Offset(REGNODE_p(ret), parse_start);
14167 Set_Node_Cur_Length(REGNODE_p(ret), parse_start-1);
14168 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14169 FALSE /* Don't force to /x */ );
14173 if (RExC_parse >= RExC_end)
14174 FAIL("Trailing \\");
14177 /* Do not generate "unrecognized" warnings here, we fall
14178 back into the quick-grab loop below */
14179 RExC_parse = parse_start;
14181 } /* end of switch on a \foo sequence */
14186 /* '#' comments should have been spaced over before this function was
14188 assert((RExC_flags & RXf_PMf_EXTENDED) == 0);
14190 if (RExC_flags & RXf_PMf_EXTENDED) {
14191 RExC_parse = reg_skipcomment( pRExC_state, RExC_parse );
14192 if (RExC_parse < RExC_end)
14202 /* Here, we have determined that the next thing is probably a
14203 * literal character. RExC_parse points to the first byte of its
14204 * definition. (It still may be an escape sequence that evaluates
14205 * to a single character) */
14210 char *s, *old_s = NULL, *old_old_s = NULL;
14212 U32 max_string_len = 255;
14214 /* We may have to reparse the node, artificially stopping filling
14215 * it early, based on info gleaned in the first parse. This
14216 * variable gives where we stop. Make it above the normal stopping
14217 * place first time through; otherwise it would stop too early */
14218 U32 upper_fill = max_string_len + 1;
14220 /* We start out as an EXACT node, even if under /i, until we find a
14221 * character which is in a fold. The algorithm now segregates into
14222 * separate nodes, characters that fold from those that don't under
14223 * /i. (This hopefully will create nodes that are fixed strings
14224 * even under /i, giving the optimizer something to grab on to.)
14225 * So, if a node has something in it and the next character is in
14226 * the opposite category, that node is closed up, and the function
14227 * returns. Then regatom is called again, and a new node is
14228 * created for the new category. */
14229 U8 node_type = EXACT;
14231 /* Assume the node will be fully used; the excess is given back at
14232 * the end. Under /i, we may need to temporarily add the fold of
14233 * an extra character or two at the end to check for splitting
14234 * multi-char folds, so allocate extra space for that. We can't
14235 * make any other length assumptions, as a byte input sequence
14236 * could shrink down. */
14237 Ptrdiff_t current_string_nodes = STR_SZ(max_string_len
14241 ? UTF8_MAXBYTES_CASE
14242 /* Max non-UTF-8 expansion is 2 */ : 2)));
14244 bool next_is_quantifier;
14245 char * oldp = NULL;
14247 /* We can convert EXACTF nodes to EXACTFU if they contain only
14248 * characters that match identically regardless of the target
14249 * string's UTF8ness. The reason to do this is that EXACTF is not
14250 * trie-able, EXACTFU is, and EXACTFU requires fewer operations at
14253 * Similarly, we can convert EXACTFL nodes to EXACTFLU8 if they
14254 * contain only above-Latin1 characters (hence must be in UTF8),
14255 * which don't participate in folds with Latin1-range characters,
14256 * as the latter's folds aren't known until runtime. */
14257 bool maybe_exactfu = FOLD && (DEPENDS_SEMANTICS || LOC);
14259 /* Single-character EXACTish nodes are almost always SIMPLE. This
14260 * allows us to override this as encountered */
14261 U8 maybe_SIMPLE = SIMPLE;
14263 /* Does this node contain something that can't match unless the
14264 * target string is (also) in UTF-8 */
14265 bool requires_utf8_target = FALSE;
14267 /* The sequence 'ss' is problematic in non-UTF-8 patterns. */
14268 bool has_ss = FALSE;
14270 /* So is the MICRO SIGN */
14271 bool has_micro_sign = FALSE;
14273 /* Set when we fill up the current node and there is still more
14274 * text to process */
14277 /* Allocate an EXACT node. The node_type may change below to
14278 * another EXACTish node, but since the size of the node doesn't
14279 * change, it works */
14280 ret = regnode_guts(pRExC_state, node_type, current_string_nodes,
14282 FILL_NODE(ret, node_type);
14285 s = STRING(REGNODE_p(ret));
14296 maybe_exactfu = FOLD && (DEPENDS_SEMANTICS || LOC);
14297 maybe_SIMPLE = SIMPLE;
14298 requires_utf8_target = FALSE;
14300 has_micro_sign = FALSE;
14304 /* This breaks under rare circumstances. If folding, we do not
14305 * want to split a node at a character that is a non-final in a
14306 * multi-char fold, as an input string could just happen to want to
14307 * match across the node boundary. The code at the end of the loop
14308 * looks for this, and backs off until it finds not such a
14309 * character, but it is possible (though extremely, extremely
14310 * unlikely) for all characters in the node to be non-final fold
14311 * ones, in which case we just leave the node fully filled, and
14312 * hope that it doesn't match the string in just the wrong place */
14314 assert( ! UTF /* Is at the beginning of a character */
14315 || UTF8_IS_INVARIANT(UCHARAT(RExC_parse))
14316 || UTF8_IS_START(UCHARAT(RExC_parse)));
14318 overflowed = FALSE;
14320 /* Here, we have a literal character. Find the maximal string of
14321 * them in the input that we can fit into a single EXACTish node.
14322 * We quit at the first non-literal or when the node gets full, or
14323 * under /i the categorization of folding/non-folding character
14325 while (p < RExC_end && len < upper_fill) {
14327 /* In most cases each iteration adds one byte to the output.
14328 * The exceptions override this */
14329 Size_t added_len = 1;
14335 /* White space has already been ignored */
14336 assert( (RExC_flags & RXf_PMf_EXTENDED) == 0
14337 || ! is_PATWS_safe((p), RExC_end, UTF));
14340 const char* message;
14353 /* Literal Escapes Switch
14355 This switch is meant to handle escape sequences that
14356 resolve to a literal character.
14358 Every escape sequence that represents something
14359 else, like an assertion or a char class, is handled
14360 in the switch marked 'Special Escapes' above in this
14361 routine, but also has an entry here as anything that
14362 isn't explicitly mentioned here will be treated as
14363 an unescaped equivalent literal.
14366 switch ((U8)*++p) {
14368 /* These are all the special escapes. */
14369 case 'A': /* Start assertion */
14370 case 'b': case 'B': /* Word-boundary assertion*/
14371 case 'C': /* Single char !DANGEROUS! */
14372 case 'd': case 'D': /* digit class */
14373 case 'g': case 'G': /* generic-backref, pos assertion */
14374 case 'h': case 'H': /* HORIZWS */
14375 case 'k': case 'K': /* named backref, keep marker */
14376 case 'p': case 'P': /* Unicode property */
14377 case 'R': /* LNBREAK */
14378 case 's': case 'S': /* space class */
14379 case 'v': case 'V': /* VERTWS */
14380 case 'w': case 'W': /* word class */
14381 case 'X': /* eXtended Unicode "combining
14382 character sequence" */
14383 case 'z': case 'Z': /* End of line/string assertion */
14387 /* Anything after here is an escape that resolves to a
14388 literal. (Except digits, which may or may not)
14394 case 'N': /* Handle a single-code point named character. */
14395 RExC_parse = p + 1;
14396 if (! grok_bslash_N(pRExC_state,
14397 NULL, /* Fail if evaluates to
14398 anything other than a
14399 single code point */
14400 &ender, /* The returned single code
14402 NULL, /* Don't need a count of
14403 how many code points */
14408 if (*flagp & NEED_UTF8)
14409 FAIL("panic: grok_bslash_N set NEED_UTF8");
14410 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
14412 /* Here, it wasn't a single code point. Go close
14413 * up this EXACTish node. The switch() prior to
14414 * this switch handles the other cases */
14415 RExC_parse = p = oldp;
14419 RExC_parse = parse_start;
14421 /* The \N{} means the pattern, if previously /d,
14422 * becomes /u. That means it can't be an EXACTF node,
14423 * but an EXACTFU */
14424 if (node_type == EXACTF) {
14425 node_type = EXACTFU;
14427 /* If the node already contains something that
14428 * differs between EXACTF and EXACTFU, reparse it
14430 if (! maybe_exactfu) {
14451 ender = ESC_NATIVE;
14459 if (! grok_bslash_o(&p,
14464 (bool) RExC_strict,
14465 FALSE, /* No illegal cp's */
14468 RExC_parse = p; /* going to die anyway; point to
14469 exact spot of failure */
14473 if (message && TO_OUTPUT_WARNINGS(p)) {
14474 warn_non_literal_string(p, packed_warn, message);
14478 if (! grok_bslash_x(&p,
14483 (bool) RExC_strict,
14484 FALSE, /* No illegal cp's */
14487 RExC_parse = p; /* going to die anyway; point
14488 to exact spot of failure */
14492 if (message && TO_OUTPUT_WARNINGS(p)) {
14493 warn_non_literal_string(p, packed_warn, message);
14497 if (ender < 0x100) {
14498 if (RExC_recode_x_to_native) {
14499 ender = LATIN1_TO_NATIVE(ender);
14506 if (! grok_bslash_c(*p, &grok_c_char,
14507 &message, &packed_warn))
14509 /* going to die anyway; point to exact spot of
14511 RExC_parse = p + ((UTF)
14512 ? UTF8_SAFE_SKIP(p, RExC_end)
14517 ender = grok_c_char;
14519 if (message && TO_OUTPUT_WARNINGS(p)) {
14520 warn_non_literal_string(p, packed_warn, message);
14524 case '8': case '9': /* must be a backreference */
14526 /* we have an escape like \8 which cannot be an octal escape
14527 * so we exit the loop, and let the outer loop handle this
14528 * escape which may or may not be a legitimate backref. */
14530 case '1': case '2': case '3':case '4':
14531 case '5': case '6': case '7':
14533 /* When we parse backslash escapes there is ambiguity
14534 * between backreferences and octal escapes. Any escape
14535 * from \1 - \9 is a backreference, any multi-digit
14536 * escape which does not start with 0 and which when
14537 * evaluated as decimal could refer to an already
14538 * parsed capture buffer is a back reference. Anything
14541 * Note this implies that \118 could be interpreted as
14542 * 118 OR as "\11" . "8" depending on whether there
14543 * were 118 capture buffers defined already in the
14546 /* NOTE, RExC_npar is 1 more than the actual number of
14547 * parens we have seen so far, hence the "<" as opposed
14549 if ( !isDIGIT(p[1]) || S_backref_value(p, RExC_end) < RExC_npar)
14550 { /* Not to be treated as an octal constant, go
14558 I32 flags = PERL_SCAN_SILENT_ILLDIGIT
14559 | PERL_SCAN_NOTIFY_ILLDIGIT;
14561 ender = grok_oct(p, &numlen, &flags, NULL);
14563 if ( (flags & PERL_SCAN_NOTIFY_ILLDIGIT)
14564 && isDIGIT(*p) /* like \08, \178 */
14565 && ckWARN(WARN_REGEXP))
14567 reg_warn_non_literal_string(
14569 form_alien_digit_msg(8, numlen, p,
14570 RExC_end, UTF, FALSE));
14576 FAIL("Trailing \\");
14579 if (isALPHANUMERIC(*p)) {
14580 /* An alpha followed by '{' is going to fail next
14581 * iteration, so don't output this warning in that
14583 if (! isALPHA(*p) || *(p + 1) != '{') {
14584 ckWARN2reg(p + 1, "Unrecognized escape \\%.1s"
14585 " passed through", p);
14588 goto normal_default;
14589 } /* End of switch on '\' */
14592 /* Trying to gain new uses for '{' without breaking too
14593 * much existing code is hard. The solution currently
14595 * 1) If there is no ambiguity that a '{' should always
14596 * be taken literally, at the start of a construct, we
14598 * 2) If the literal '{' conflicts with our desired use
14599 * of it as a metacharacter, we die. The deprecation
14600 * cycles for this have come and gone.
14601 * 3) If there is ambiguity, we raise a simple warning.
14602 * This could happen, for example, if the user
14603 * intended it to introduce a quantifier, but slightly
14604 * misspelled the quantifier. Without this warning,
14605 * the quantifier would silently be taken as a literal
14606 * string of characters instead of a meta construct */
14607 if (len || (p > RExC_start && isALPHA_A(*(p - 1)))) {
14609 || ( p > parse_start + 1
14610 && isALPHA_A(*(p - 1))
14611 && *(p - 2) == '\\'))
14613 RExC_parse = p + 1;
14614 vFAIL("Unescaped left brace in regex is "
14617 ckWARNreg(p + 1, "Unescaped left brace in regex is"
14618 " passed through");
14620 goto normal_default;
14623 if (p > RExC_parse && RExC_strict) {
14624 ckWARN2reg(p + 1, "Unescaped literal '%c'", *p);
14627 default: /* A literal character */
14629 if (! UTF8_IS_INVARIANT(*p) && UTF) {
14631 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
14632 &numlen, UTF8_ALLOW_DEFAULT);
14638 } /* End of switch on the literal */
14640 /* Here, have looked at the literal character, and <ender>
14641 * contains its ordinal; <p> points to the character after it.
14645 REQUIRE_UTF8(flagp);
14646 if ( UNICODE_IS_PERL_EXTENDED(ender)
14647 && TO_OUTPUT_WARNINGS(p))
14649 ckWARN2_non_literal_string(p,
14650 packWARN(WARN_PORTABLE),
14651 PL_extended_cp_format,
14656 /* We need to check if the next non-ignored thing is a
14657 * quantifier. Move <p> to after anything that should be
14658 * ignored, which, as a side effect, positions <p> for the next
14659 * loop iteration */
14660 skip_to_be_ignored_text(pRExC_state, &p,
14661 FALSE /* Don't force to /x */ );
14663 /* If the next thing is a quantifier, it applies to this
14664 * character only, which means that this character has to be in
14665 * its own node and can't just be appended to the string in an
14666 * existing node, so if there are already other characters in
14667 * the node, close the node with just them, and set up to do
14668 * this character again next time through, when it will be the
14669 * only thing in its new node */
14671 next_is_quantifier = LIKELY(p < RExC_end)
14672 && UNLIKELY(isQUANTIFIER(p, RExC_end));
14674 if (next_is_quantifier && LIKELY(len)) {
14679 /* Ready to add 'ender' to the node */
14681 if (! FOLD) { /* The simple case, just append the literal */
14684 /* Don't output if it would overflow */
14685 if (UNLIKELY(len > max_string_len - ((UTF)
14686 ? UVCHR_SKIP(ender)
14693 if (UVCHR_IS_INVARIANT(ender) || ! UTF) {
14694 *(s++) = (char) ender;
14697 U8 * new_s = uvchr_to_utf8((U8*)s, ender);
14698 added_len = (char *) new_s - s;
14699 s = (char *) new_s;
14702 requires_utf8_target = TRUE;
14706 else if (LOC && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)) {
14708 /* Here are folding under /l, and the code point is
14709 * problematic. If this is the first character in the
14710 * node, change the node type to folding. Otherwise, if
14711 * this is the first problematic character, close up the
14712 * existing node, so can start a new node with this one */
14714 node_type = EXACTFL;
14715 RExC_contains_locale = 1;
14717 else if (node_type == EXACT) {
14722 /* This problematic code point means we can't simplify
14724 maybe_exactfu = FALSE;
14726 /* Although these two characters have folds that are
14727 * locale-problematic, they also have folds to above Latin1
14728 * that aren't a problem. Doing these now helps at
14730 if (UNLIKELY( ender == GREEK_CAPITAL_LETTER_MU
14731 || ender == LATIN_CAPITAL_LETTER_SHARP_S))
14736 /* Here, we are adding a problematic fold character.
14737 * "Problematic" in this context means that its fold isn't
14738 * known until runtime. (The non-problematic code points
14739 * are the above-Latin1 ones that fold to also all
14740 * above-Latin1. Their folds don't vary no matter what the
14741 * locale is.) But here we have characters whose fold
14742 * depends on the locale. We just add in the unfolded
14743 * character, and wait until runtime to fold it */
14744 goto not_fold_common;
14746 else /* regular fold; see if actually is in a fold */
14747 if ( (ender < 256 && ! IS_IN_SOME_FOLD_L1(ender))
14749 && ! _invlist_contains_cp(PL_in_some_fold, ender)))
14751 /* Here, folding, but the character isn't in a fold.
14753 * Start a new node if previous characters in the node were
14755 if (len && node_type != EXACT) {
14760 /* Here, continuing a node with non-folded characters. Add
14762 goto not_fold_common;
14764 else { /* Here, does participate in some fold */
14766 /* If this is the first character in the node, change its
14767 * type to folding. Otherwise, if this is the first
14768 * folding character in the node, close up the existing
14769 * node, so can start a new node with this one. */
14771 node_type = compute_EXACTish(pRExC_state);
14773 else if (node_type == EXACT) {
14778 if (UTF) { /* Alway use the folded value for UTF-8
14780 if (UVCHR_IS_INVARIANT(ender)) {
14781 if (UNLIKELY(len + 1 > max_string_len)) {
14786 *(s)++ = (U8) toFOLD(ender);
14792 folded = _to_uni_fold_flags(
14794 (U8 *) s, /* We have allocated extra space
14795 in 's' so can't run off the
14799 | (( ASCII_FOLD_RESTRICTED
14800 || node_type == EXACTFL)
14801 ? FOLD_FLAGS_NOMIX_ASCII
14803 if (UNLIKELY(len + added_len > max_string_len)) {
14811 && LIKELY(folded != GREEK_SMALL_LETTER_MU))
14813 /* U+B5 folds to the MU, so its possible for a
14814 * non-UTF-8 target to match it */
14815 requires_utf8_target = TRUE;
14819 else { /* Here is non-UTF8. */
14821 /* The fold will be one or (rarely) two characters.
14822 * Check that there's room for at least a single one
14823 * before setting any flags, etc. Because otherwise an
14824 * overflowing character could cause a flag to be set
14825 * even though it doesn't end up in this node. (For
14826 * the two character fold, we check again, before
14827 * setting any flags) */
14828 if (UNLIKELY(len + 1 > max_string_len)) {
14833 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
14834 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
14835 || UNICODE_DOT_DOT_VERSION > 0)
14837 /* On non-ancient Unicodes, check for the only possible
14838 * multi-char fold */
14839 if (UNLIKELY(ender == LATIN_SMALL_LETTER_SHARP_S)) {
14841 /* This potential multi-char fold means the node
14842 * can't be simple (because it could match more
14843 * than a single char). And in some cases it will
14844 * match 'ss', so set that flag */
14848 /* It can't change to be an EXACTFU (unless already
14849 * is one). We fold it iff under /u rules. */
14850 if (node_type != EXACTFU) {
14851 maybe_exactfu = FALSE;
14854 if (UNLIKELY(len + 2 > max_string_len)) {
14863 goto done_with_this_char;
14866 else if ( UNLIKELY(isALPHA_FOLD_EQ(ender, 's'))
14868 && UNLIKELY(isALPHA_FOLD_EQ(*(s-1), 's')))
14870 /* Also, the sequence 'ss' is special when not
14871 * under /u. If the target string is UTF-8, it
14872 * should match SHARP S; otherwise it won't. So,
14873 * here we have to exclude the possibility of this
14874 * node moving to /u.*/
14876 maybe_exactfu = FALSE;
14879 /* Here, the fold will be a single character */
14881 if (UNLIKELY(ender == MICRO_SIGN)) {
14882 has_micro_sign = TRUE;
14884 else if (PL_fold[ender] != PL_fold_latin1[ender]) {
14886 /* If the character's fold differs between /d and
14887 * /u, this can't change to be an EXACTFU node */
14888 maybe_exactfu = FALSE;
14891 *(s++) = (DEPENDS_SEMANTICS)
14892 ? (char) toFOLD(ender)
14894 /* Under /u, the fold of any character in
14895 * the 0-255 range happens to be its
14896 * lowercase equivalent, except for LATIN
14897 * SMALL LETTER SHARP S, which was handled
14898 * above, and the MICRO SIGN, whose fold
14899 * requires UTF-8 to represent. */
14900 : (char) toLOWER_L1(ender);
14902 } /* End of adding current character to the node */
14904 done_with_this_char:
14908 if (next_is_quantifier) {
14910 /* Here, the next input is a quantifier, and to get here,
14911 * the current character is the only one in the node. */
14915 } /* End of loop through literal characters */
14917 /* Here we have either exhausted the input or run out of room in
14918 * the node. If the former, we are done. (If we encountered a
14919 * character that can't be in the node, transfer is made directly
14920 * to <loopdone>, and so we wouldn't have fallen off the end of the
14922 if (LIKELY(! overflowed)) {
14926 /* Here we have run out of room. We can grow plain EXACT and
14927 * LEXACT nodes. If the pattern is gigantic enough, though,
14928 * eventually we'll have to artificially chunk the pattern into
14929 * multiple nodes. */
14930 if (! LOC && (node_type == EXACT || node_type == LEXACT)) {
14931 Size_t overhead = 1 + regarglen[OP(REGNODE_p(ret))];
14932 Size_t overhead_expansion = 0;
14934 Size_t max_nodes_for_string;
14938 /* Here we couldn't fit the final character in the current
14939 * node, so it will have to be reparsed, no matter what else we
14943 /* If would have overflowed a regular EXACT node, switch
14944 * instead to an LEXACT. The code below is structured so that
14945 * the actual growing code is common to changing from an EXACT
14946 * or just increasing the LEXACT size. This means that we have
14947 * to save the string in the EXACT case before growing, and
14948 * then copy it afterwards to its new location */
14949 if (node_type == EXACT) {
14950 overhead_expansion = regarglen[LEXACT] - regarglen[EXACT];
14951 RExC_emit += overhead_expansion;
14952 Copy(s0, temp, len, char);
14955 /* Ready to grow. If it was a plain EXACT, the string was
14956 * saved, and the first few bytes of it overwritten by adding
14957 * an argument field. We assume, as we do elsewhere in this
14958 * file, that one byte of remaining input will translate into
14959 * one byte of output, and if that's too small, we grow again,
14960 * if too large the excess memory is freed at the end */
14962 max_nodes_for_string = U16_MAX - overhead - overhead_expansion;
14963 achievable = MIN(max_nodes_for_string,
14964 current_string_nodes + STR_SZ(RExC_end - p));
14965 delta = achievable - current_string_nodes;
14967 /* If there is just no more room, go finish up this chunk of
14973 change_engine_size(pRExC_state, delta + overhead_expansion);
14974 current_string_nodes += delta;
14976 = sizeof(struct regnode) * current_string_nodes;
14977 upper_fill = max_string_len + 1;
14979 /* If the length was small, we know this was originally an
14980 * EXACT node now converted to LEXACT, and the string has to be
14981 * restored. Otherwise the string was untouched. 260 is just
14982 * a number safely above 255 so don't have to worry about
14983 * getting it precise */
14985 node_type = LEXACT;
14986 FILL_NODE(ret, node_type);
14987 s0 = STRING(REGNODE_p(ret));
14988 Copy(temp, s0, len, char);
14992 goto continue_parse;
14995 bool splittable = FALSE;
14996 bool backed_up = FALSE;
14997 char * e; /* should this be U8? */
14998 char * s_start; /* should this be U8? */
15000 /* Here is /i. Running out of room creates a problem if we are
15001 * folding, and the split happens in the middle of a
15002 * multi-character fold, as a match that should have occurred,
15003 * won't, due to the way nodes are matched, and our artificial
15004 * boundary. So back off until we aren't splitting such a
15005 * fold. If there is no such place to back off to, we end up
15006 * taking the entire node as-is. This can happen if the node
15007 * consists entirely of 'f' or entirely of 's' characters (or
15008 * things that fold to them) as 'ff' and 'ss' are
15009 * multi-character folds.
15011 * The Unicode standard says that multi character folds consist
15012 * of either two or three characters. That means we would be
15013 * splitting one if the final character in the node is at the
15014 * beginning of either type, or is the second of a three
15018 * ender is the code point of the character that won't fit
15020 * s points to just beyond the final byte in the node.
15021 * It's where we would place ender if there were
15022 * room, and where in fact we do place ender's fold
15023 * in the code below, as we've over-allocated space
15024 * for s0 (hence s) to allow for this
15025 * e starts at 's' and advances as we append things.
15026 * old_s is the same as 's'. (If ender had fit, 's' would
15027 * have been advanced to beyond it).
15028 * old_old_s points to the beginning byte of the final
15029 * character in the node
15030 * p points to the beginning byte in the input of the
15031 * character beyond 'ender'.
15032 * oldp points to the beginning byte in the input of
15035 * In the case of /il, we haven't folded anything that could be
15036 * affected by the locale. That means only above-Latin1
15037 * characters that fold to other above-latin1 characters get
15038 * folded at compile time. To check where a good place to
15039 * split nodes is, everything in it will have to be folded.
15040 * The boolean 'maybe_exactfu' keeps track in /il if there are
15041 * any unfolded characters in the node. */
15042 bool need_to_fold_loc = LOC && ! maybe_exactfu;
15044 /* If we do need to fold the node, we need a place to store the
15045 * folded copy, and a way to map back to the unfolded original
15047 char * locfold_buf = NULL;
15048 Size_t * loc_correspondence = NULL;
15050 if (! need_to_fold_loc) { /* The normal case. Just
15051 initialize to the actual node */
15054 s = old_old_s; /* Point to the beginning of the final char
15055 that fits in the node */
15059 /* Here, we have filled a /il node, and there are unfolded
15060 * characters in it. If the runtime locale turns out to be
15061 * UTF-8, there are possible multi-character folds, just
15062 * like when not under /l. The node hence can't terminate
15063 * in the middle of such a fold. To determine this, we
15064 * have to create a folded copy of this node. That means
15065 * reparsing the node, folding everything assuming a UTF-8
15066 * locale. (If at runtime it isn't such a locale, the
15067 * actions here wouldn't have been necessary, but we have
15068 * to assume the worst case.) If we find we need to back
15069 * off the folded string, we do so, and then map that
15070 * position back to the original unfolded node, which then
15071 * gets output, truncated at that spot */
15073 char * redo_p = RExC_parse;
15077 /* Allow enough space assuming a single byte input folds to
15078 * a single byte output, plus assume that the two unparsed
15079 * characters (that we may need) fold to the largest number
15080 * of bytes possible, plus extra for one more worst case
15081 * scenario. In the loop below, if we start eating into
15082 * that final spare space, we enlarge this initial space */
15083 Size_t size = max_string_len + (3 * UTF8_MAXBYTES_CASE) + 1;
15085 Newxz(locfold_buf, size, char);
15086 Newxz(loc_correspondence, size, Size_t);
15088 /* Redo this node's parse, folding into 'locfold_buf' */
15089 redo_p = RExC_parse;
15090 old_redo_e = redo_e = locfold_buf;
15091 while (redo_p <= oldp) {
15093 old_redo_e = redo_e;
15094 loc_correspondence[redo_e - locfold_buf]
15095 = redo_p - RExC_parse;
15100 (void) _to_utf8_fold_flags((U8 *) redo_p,
15105 redo_e += added_len;
15106 redo_p += UTF8SKIP(redo_p);
15110 /* Note that if this code is run on some ancient
15111 * Unicode versions, SHARP S doesn't fold to 'ss',
15112 * but rather than clutter the code with #ifdef's,
15113 * as is done above, we ignore that possibility.
15114 * This is ok because this code doesn't affect what
15115 * gets matched, but merely where the node gets
15117 if (UCHARAT(redo_p) != LATIN_SMALL_LETTER_SHARP_S) {
15118 *redo_e++ = toLOWER_L1(UCHARAT(redo_p));
15128 /* If we're getting so close to the end that a
15129 * worst-case fold in the next character would cause us
15130 * to overflow, increase, assuming one byte output byte
15131 * per one byte input one, plus room for another worst
15133 if ( redo_p <= oldp
15134 && redo_e > locfold_buf + size
15135 - (UTF8_MAXBYTES_CASE + 1))
15137 Size_t new_size = size
15139 + UTF8_MAXBYTES_CASE + 1;
15140 Ptrdiff_t e_offset = redo_e - locfold_buf;
15142 Renew(locfold_buf, new_size, char);
15143 Renew(loc_correspondence, new_size, Size_t);
15146 redo_e = locfold_buf + e_offset;
15150 /* Set so that things are in terms of the folded, temporary
15153 s_start = locfold_buf;
15158 /* Here, we have 's', 's_start' and 'e' set up to point to the
15159 * input that goes into the node, folded.
15161 * If the final character of the node and the fold of ender
15162 * form the first two characters of a three character fold, we
15163 * need to peek ahead at the next (unparsed) character in the
15164 * input to determine if the three actually do form such a
15165 * fold. Just looking at that character is not generally
15166 * sufficient, as it could be, for example, an escape sequence
15167 * that evaluates to something else, and it needs to be folded.
15169 * khw originally thought to just go through the parse loop one
15170 * extra time, but that doesn't work easily as that iteration
15171 * could cause things to think that the parse is over and to
15172 * goto loopdone. The character could be a '$' for example, or
15173 * the character beyond could be a quantifier, and other
15174 * glitches as well.
15176 * The solution used here for peeking ahead is to look at that
15177 * next character. If it isn't ASCII punctuation, then it will
15178 * be something that would continue on in an EXACTish node if
15179 * there were space. We append the fold of it to s, having
15180 * reserved enough room in s0 for the purpose. If we can't
15181 * reasonably peek ahead, we instead assume the worst case:
15182 * that it is something that would form the completion of a
15185 * If we can't split between s and ender, we work backwards
15186 * character-by-character down to s0. At each current point
15187 * see if we are at the beginning of a multi-char fold. If so,
15188 * that means we would be splitting the fold across nodes, and
15189 * so we back up one and try again.
15191 * If we're not at the beginning, we still could be at the
15192 * final two characters of a (rare) three character fold. We
15193 * check if the sequence starting at the character before the
15194 * current position (and including the current and next
15195 * characters) is a three character fold. If not, the node can
15196 * be split here. If it is, we have to backup two characters
15199 * Otherwise, the node can be split at the current position.
15201 * The same logic is used for UTF-8 patterns and not */
15205 /* Append the fold of ender */
15206 (void) _to_uni_fold_flags(
15210 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
15211 ? FOLD_FLAGS_NOMIX_ASCII
15215 /* 's' and the character folded to by ender may be the
15216 * first two of a three-character fold, in which case the
15217 * node should not be split here. That may mean examining
15218 * the so-far unparsed character starting at 'p'. But if
15219 * ender folded to more than one character, we already have
15220 * three characters to look at. Also, we first check if
15221 * the sequence consisting of s and the next character form
15222 * the first two of some three character fold. If not,
15223 * there's no need to peek ahead. */
15224 if ( added_len <= UTF8SKIP(e - added_len)
15225 && UNLIKELY(is_THREE_CHAR_FOLD_HEAD_utf8_safe(s, e)))
15227 /* Here, the two do form the beginning of a potential
15228 * three character fold. The unexamined character may
15229 * or may not complete it. Peek at it. It might be
15230 * something that ends the node or an escape sequence,
15231 * in which case we don't know without a lot of work
15232 * what it evaluates to, so we have to assume the worst
15233 * case: that it does complete the fold, and so we
15234 * can't split here. All such instances will have
15235 * that character be an ASCII punctuation character,
15236 * like a backslash. So, for that case, backup one and
15237 * drop down to try at that position */
15239 s = (char *) utf8_hop_back((U8 *) s, -1,
15244 /* Here, since it's not punctuation, it must be a
15245 * real character, and we can append its fold to
15246 * 'e' (having deliberately reserved enough space
15247 * for this eventuality) and drop down to check if
15248 * the three actually do form a folded sequence */
15249 (void) _to_utf8_fold_flags(
15250 (U8 *) p, (U8 *) RExC_end,
15253 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
15254 ? FOLD_FLAGS_NOMIX_ASCII
15260 /* Here, we either have three characters available in
15261 * sequence starting at 's', or we have two characters and
15262 * know that the following one can't possibly be part of a
15263 * three character fold. We go through the node backwards
15264 * until we find a place where we can split it without
15265 * breaking apart a multi-character fold. At any given
15266 * point we have to worry about if such a fold begins at
15267 * the current 's', and also if a three-character fold
15268 * begins at s-1, (containing s and s+1). Splitting in
15269 * either case would break apart a fold */
15271 char *prev_s = (char *) utf8_hop_back((U8 *) s, -1,
15274 /* If is a multi-char fold, can't split here. Backup
15275 * one char and try again */
15276 if (UNLIKELY(is_MULTI_CHAR_FOLD_utf8_safe(s, e))) {
15282 /* If the two characters beginning at 's' are part of a
15283 * three character fold starting at the character
15284 * before s, we can't split either before or after s.
15285 * Backup two chars and try again */
15286 if ( LIKELY(s > s_start)
15287 && UNLIKELY(is_THREE_CHAR_FOLD_utf8_safe(prev_s, e)))
15290 s = (char *) utf8_hop_back((U8 *) s, -1, (U8 *) s_start);
15295 /* Here there's no multi-char fold between s and the
15296 * next character following it. We can split */
15300 } while (s > s_start); /* End of loops backing up through the node */
15302 /* Here we either couldn't find a place to split the node,
15303 * or else we broke out of the loop setting 'splittable' to
15304 * true. In the latter case, the place to split is between
15305 * the first and second characters in the sequence starting
15311 else { /* Pattern not UTF-8 */
15312 if ( ender != LATIN_SMALL_LETTER_SHARP_S
15313 || ASCII_FOLD_RESTRICTED)
15315 assert( toLOWER_L1(ender) < 256 );
15316 *e++ = (char)(toLOWER_L1(ender)); /* should e and the cast be U8? */
15324 && UNLIKELY(is_THREE_CHAR_FOLD_HEAD_latin1_safe(s, e)))
15331 if ( UCHARAT(p) != LATIN_SMALL_LETTER_SHARP_S
15332 || ASCII_FOLD_RESTRICTED)
15334 assert( toLOWER_L1(ender) < 256 );
15335 *e++ = (char)(toLOWER_L1(ender)); /* should e and the cast be U8? */
15345 if (UNLIKELY(is_MULTI_CHAR_FOLD_latin1_safe(s, e))) {
15351 if ( LIKELY(s > s_start)
15352 && UNLIKELY(is_THREE_CHAR_FOLD_latin1_safe(s - 1, e)))
15362 } while (s > s_start);
15369 /* Here, we are done backing up. If we didn't backup at all
15370 * (the likely case), just proceed */
15373 /* If we did find a place to split, reparse the entire node
15374 * stopping where we have calculated. */
15377 /* If we created a temporary folded string under /l, we
15378 * have to map that back to the original */
15379 if (need_to_fold_loc) {
15380 upper_fill = loc_correspondence[s - s_start];
15381 if (upper_fill == 0) {
15382 FAIL2("panic: loc_correspondence[%d] is 0",
15383 (int) (s - s_start));
15385 Safefree(locfold_buf);
15386 Safefree(loc_correspondence);
15389 upper_fill = s - s0;
15394 /* Here the node consists entirely of non-final multi-char
15395 * folds. (Likely it is all 'f's or all 's's.) There's no
15396 * decent place to split it, so give up and just take the
15401 if (need_to_fold_loc) {
15402 Safefree(locfold_buf);
15403 Safefree(loc_correspondence);
15405 } /* End of verifying node ends with an appropriate char */
15407 /* We need to start the next node at the character that didn't fit
15411 loopdone: /* Jumped to when encounters something that shouldn't be
15414 /* Free up any over-allocated space; cast is to silence bogus
15415 * warning in MS VC */
15416 change_engine_size(pRExC_state,
15417 - (Ptrdiff_t) (current_string_nodes - STR_SZ(len)));
15419 /* I (khw) don't know if you can get here with zero length, but the
15420 * old code handled this situation by creating a zero-length EXACT
15421 * node. Might as well be NOTHING instead */
15423 OP(REGNODE_p(ret)) = NOTHING;
15427 /* If the node type is EXACT here, check to see if it
15428 * should be EXACTL, or EXACT_REQ8. */
15429 if (node_type == EXACT) {
15431 node_type = EXACTL;
15433 else if (requires_utf8_target) {
15434 node_type = EXACT_REQ8;
15437 else if (node_type == LEXACT) {
15438 if (requires_utf8_target) {
15439 node_type = LEXACT_REQ8;
15443 if ( UNLIKELY(has_micro_sign || has_ss)
15444 && (node_type == EXACTFU || ( node_type == EXACTF
15445 && maybe_exactfu)))
15446 { /* These two conditions are problematic in non-UTF-8
15449 node_type = EXACTFUP;
15451 else if (node_type == EXACTFL) {
15453 /* 'maybe_exactfu' is deliberately set above to
15454 * indicate this node type, where all code points in it
15456 if (maybe_exactfu) {
15457 node_type = EXACTFLU8;
15460 _invlist_contains_cp(PL_HasMultiCharFold, ender)))
15462 /* A character that folds to more than one will
15463 * match multiple characters, so can't be SIMPLE.
15464 * We don't have to worry about this with EXACTFLU8
15465 * nodes just above, as they have already been
15466 * folded (since the fold doesn't vary at run
15467 * time). Here, if the final character in the node
15468 * folds to multiple, it can't be simple. (This
15469 * only has an effect if the node has only a single
15470 * character, hence the final one, as elsewhere we
15471 * turn off simple for nodes whose length > 1 */
15475 else if (node_type == EXACTF) { /* Means is /di */
15477 /* This intermediate variable is needed solely because
15478 * the asserts in the macro where used exceed Win32's
15479 * literal string capacity */
15480 char first_char = * STRING(REGNODE_p(ret));
15482 /* If 'maybe_exactfu' is clear, then we need to stay
15483 * /di. If it is set, it means there are no code
15484 * points that match differently depending on UTF8ness
15485 * of the target string, so it can become an EXACTFU
15487 if (! maybe_exactfu) {
15488 RExC_seen_d_op = TRUE;
15490 else if ( isALPHA_FOLD_EQ(first_char, 's')
15491 || isALPHA_FOLD_EQ(ender, 's'))
15493 /* But, if the node begins or ends in an 's' we
15494 * have to defer changing it into an EXACTFU, as
15495 * the node could later get joined with another one
15496 * that ends or begins with 's' creating an 'ss'
15497 * sequence which would then wrongly match the
15498 * sharp s without the target being UTF-8. We
15499 * create a special node that we resolve later when
15500 * we join nodes together */
15502 node_type = EXACTFU_S_EDGE;
15505 node_type = EXACTFU;
15509 if (requires_utf8_target && node_type == EXACTFU) {
15510 node_type = EXACTFU_REQ8;
15514 OP(REGNODE_p(ret)) = node_type;
15515 setSTR_LEN(REGNODE_p(ret), len);
15516 RExC_emit += STR_SZ(len);
15518 /* If the node isn't a single character, it can't be SIMPLE */
15519 if (len > (Size_t) ((UTF) ? UTF8SKIP(STRING(REGNODE_p(ret))) : 1)) {
15523 *flagp |= HASWIDTH | maybe_SIMPLE;
15526 Set_Node_Length(REGNODE_p(ret), p - parse_start - 1);
15530 /* len is STRLEN which is unsigned, need to copy to signed */
15533 vFAIL("Internal disaster");
15536 } /* End of label 'defchar:' */
15538 } /* End of giant switch on input character */
15540 /* Position parse to next real character */
15541 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
15542 FALSE /* Don't force to /x */ );
15543 if ( *RExC_parse == '{'
15544 && OP(REGNODE_p(ret)) != SBOL && ! regcurly(RExC_parse, RExC_end, NULL))
15548 vFAIL("Unescaped left brace in regex is illegal here");
15550 ckWARNreg(RExC_parse + 1, "Unescaped left brace in regex is"
15551 " passed through");
15559 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
15561 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
15562 * sets up the bitmap and any flags, removing those code points from the
15563 * inversion list, setting it to NULL should it become completely empty */
15566 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
15567 assert(PL_regkind[OP(node)] == ANYOF);
15569 /* There is no bitmap for this node type */
15570 if (inRANGE(OP(node), ANYOFH, ANYOFRb)) {
15574 ANYOF_BITMAP_ZERO(node);
15575 if (*invlist_ptr) {
15577 /* This gets set if we actually need to modify things */
15578 bool change_invlist = FALSE;
15582 /* Start looking through *invlist_ptr */
15583 invlist_iterinit(*invlist_ptr);
15584 while (invlist_iternext(*invlist_ptr, &start, &end)) {
15588 if (end == UV_MAX && start <= NUM_ANYOF_CODE_POINTS) {
15589 ANYOF_FLAGS(node) |= ANYOF_MATCHES_ALL_ABOVE_BITMAP;
15592 /* Quit if are above what we should change */
15593 if (start >= NUM_ANYOF_CODE_POINTS) {
15597 change_invlist = TRUE;
15599 /* Set all the bits in the range, up to the max that we are doing */
15600 high = (end < NUM_ANYOF_CODE_POINTS - 1)
15602 : NUM_ANYOF_CODE_POINTS - 1;
15603 for (i = start; i <= (int) high; i++) {
15604 ANYOF_BITMAP_SET(node, i);
15607 invlist_iterfinish(*invlist_ptr);
15609 /* Done with loop; remove any code points that are in the bitmap from
15610 * *invlist_ptr; similarly for code points above the bitmap if we have
15611 * a flag to match all of them anyways */
15612 if (change_invlist) {
15613 _invlist_subtract(*invlist_ptr, PL_InBitmap, invlist_ptr);
15615 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
15616 _invlist_intersection(*invlist_ptr, PL_InBitmap, invlist_ptr);
15619 /* If have completely emptied it, remove it completely */
15620 if (_invlist_len(*invlist_ptr) == 0) {
15621 SvREFCNT_dec_NN(*invlist_ptr);
15622 *invlist_ptr = NULL;
15627 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
15628 Character classes ([:foo:]) can also be negated ([:^foo:]).
15629 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
15630 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
15631 but trigger failures because they are currently unimplemented. */
15633 #define POSIXCC_DONE(c) ((c) == ':')
15634 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
15635 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
15636 #define MAYBE_POSIXCC(c) (POSIXCC(c) || (c) == '^' || (c) == ';')
15638 #define WARNING_PREFIX "Assuming NOT a POSIX class since "
15639 #define NO_BLANKS_POSIX_WARNING "no blanks are allowed in one"
15640 #define SEMI_COLON_POSIX_WARNING "a semi-colon was found instead of a colon"
15642 #define NOT_MEANT_TO_BE_A_POSIX_CLASS (OOB_NAMEDCLASS - 1)
15644 /* 'posix_warnings' and 'warn_text' are names of variables in the following
15646 #define ADD_POSIX_WARNING(p, text) STMT_START { \
15647 if (posix_warnings) { \
15648 if (! RExC_warn_text ) RExC_warn_text = \
15649 (AV *) sv_2mortal((SV *) newAV()); \
15650 av_push(RExC_warn_text, Perl_newSVpvf(aTHX_ \
15654 REPORT_LOCATION_ARGS(p))); \
15657 #define CLEAR_POSIX_WARNINGS() \
15659 if (posix_warnings && RExC_warn_text) \
15660 av_clear(RExC_warn_text); \
15663 #define CLEAR_POSIX_WARNINGS_AND_RETURN(ret) \
15665 CLEAR_POSIX_WARNINGS(); \
15670 S_handle_possible_posix(pTHX_ RExC_state_t *pRExC_state,
15672 const char * const s, /* Where the putative posix class begins.
15673 Normally, this is one past the '['. This
15674 parameter exists so it can be somewhere
15675 besides RExC_parse. */
15676 char ** updated_parse_ptr, /* Where to set the updated parse pointer, or
15678 AV ** posix_warnings, /* Where to place any generated warnings, or
15680 const bool check_only /* Don't die if error */
15683 /* This parses what the caller thinks may be one of the three POSIX
15685 * 1) a character class, like [:blank:]
15686 * 2) a collating symbol, like [. .]
15687 * 3) an equivalence class, like [= =]
15688 * In the latter two cases, it croaks if it finds a syntactically legal
15689 * one, as these are not handled by Perl.
15691 * The main purpose is to look for a POSIX character class. It returns:
15692 * a) the class number
15693 * if it is a completely syntactically and semantically legal class.
15694 * 'updated_parse_ptr', if not NULL, is set to point to just after the
15695 * closing ']' of the class
15696 * b) OOB_NAMEDCLASS
15697 * if it appears that one of the three POSIX constructs was meant, but
15698 * its specification was somehow defective. 'updated_parse_ptr', if
15699 * not NULL, is set to point to the character just after the end
15700 * character of the class. See below for handling of warnings.
15701 * c) NOT_MEANT_TO_BE_A_POSIX_CLASS
15702 * if it doesn't appear that a POSIX construct was intended.
15703 * 'updated_parse_ptr' is not changed. No warnings nor errors are
15706 * In b) there may be errors or warnings generated. If 'check_only' is
15707 * TRUE, then any errors are discarded. Warnings are returned to the
15708 * caller via an AV* created into '*posix_warnings' if it is not NULL. If
15709 * instead it is NULL, warnings are suppressed.
15711 * The reason for this function, and its complexity is that a bracketed
15712 * character class can contain just about anything. But it's easy to
15713 * mistype the very specific posix class syntax but yielding a valid
15714 * regular bracketed class, so it silently gets compiled into something
15715 * quite unintended.
15717 * The solution adopted here maintains backward compatibility except that
15718 * it adds a warning if it looks like a posix class was intended but
15719 * improperly specified. The warning is not raised unless what is input
15720 * very closely resembles one of the 14 legal posix classes. To do this,
15721 * it uses fuzzy parsing. It calculates how many single-character edits it
15722 * would take to transform what was input into a legal posix class. Only
15723 * if that number is quite small does it think that the intention was a
15724 * posix class. Obviously these are heuristics, and there will be cases
15725 * where it errs on one side or another, and they can be tweaked as
15726 * experience informs.
15728 * The syntax for a legal posix class is:
15730 * qr/(?xa: \[ : \^? [[:lower:]]{4,6} : \] )/
15732 * What this routine considers syntactically to be an intended posix class
15733 * is this (the comments indicate some restrictions that the pattern
15736 * qr/(?x: \[? # The left bracket, possibly
15738 * \h* # possibly followed by blanks
15739 * (?: \^ \h* )? # possibly a misplaced caret
15740 * [:;]? # The opening class character,
15741 * # possibly omitted. A typo
15742 * # semi-colon can also be used.
15744 * \^? # possibly a correctly placed
15745 * # caret, but not if there was also
15746 * # a misplaced one
15748 * .{3,15} # The class name. If there are
15749 * # deviations from the legal syntax,
15750 * # its edit distance must be close
15751 * # to a real class name in order
15752 * # for it to be considered to be
15753 * # an intended posix class.
15755 * [[:punct:]]? # The closing class character,
15756 * # possibly omitted. If not a colon
15757 * # nor semi colon, the class name
15758 * # must be even closer to a valid
15761 * \]? # The right bracket, possibly
15765 * In the above, \h must be ASCII-only.
15767 * These are heuristics, and can be tweaked as field experience dictates.
15768 * There will be cases when someone didn't intend to specify a posix class
15769 * that this warns as being so. The goal is to minimize these, while
15770 * maximizing the catching of things intended to be a posix class that
15771 * aren't parsed as such.
15775 const char * const e = RExC_end;
15776 unsigned complement = 0; /* If to complement the class */
15777 bool found_problem = FALSE; /* Assume OK until proven otherwise */
15778 bool has_opening_bracket = FALSE;
15779 bool has_opening_colon = FALSE;
15780 int class_number = OOB_NAMEDCLASS; /* Out-of-bounds until find
15782 const char * possible_end = NULL; /* used for a 2nd parse pass */
15783 const char* name_start; /* ptr to class name first char */
15785 /* If the number of single-character typos the input name is away from a
15786 * legal name is no more than this number, it is considered to have meant
15787 * the legal name */
15788 int max_distance = 2;
15790 /* to store the name. The size determines the maximum length before we
15791 * decide that no posix class was intended. Should be at least
15792 * sizeof("alphanumeric") */
15794 STATIC_ASSERT_DECL(C_ARRAY_LENGTH(input_text) >= sizeof "alphanumeric");
15796 PERL_ARGS_ASSERT_HANDLE_POSSIBLE_POSIX;
15798 CLEAR_POSIX_WARNINGS();
15801 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
15804 if (*(p - 1) != '[') {
15805 ADD_POSIX_WARNING(p, "it doesn't start with a '['");
15806 found_problem = TRUE;
15809 has_opening_bracket = TRUE;
15812 /* They could be confused and think you can put spaces between the
15815 found_problem = TRUE;
15819 } while (p < e && isBLANK(*p));
15821 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15824 /* For [. .] and [= =]. These are quite different internally from [: :],
15825 * so they are handled separately. */
15826 if (POSIXCC_NOTYET(*p) && p < e - 3) /* 1 for the close, and 1 for the ']'
15827 and 1 for at least one char in it
15830 const char open_char = *p;
15831 const char * temp_ptr = p + 1;
15833 /* These two constructs are not handled by perl, and if we find a
15834 * syntactically valid one, we croak. khw, who wrote this code, finds
15835 * this explanation of them very unclear:
15836 * http://pubs.opengroup.org/onlinepubs/009696899/basedefs/xbd_chap09.html
15837 * And searching the rest of the internet wasn't very helpful either.
15838 * It looks like just about any byte can be in these constructs,
15839 * depending on the locale. But unless the pattern is being compiled
15840 * under /l, which is very rare, Perl runs under the C or POSIX locale.
15841 * In that case, it looks like [= =] isn't allowed at all, and that
15842 * [. .] could be any single code point, but for longer strings the
15843 * constituent characters would have to be the ASCII alphabetics plus
15844 * the minus-hyphen. Any sensible locale definition would limit itself
15845 * to these. And any portable one definitely should. Trying to parse
15846 * the general case is a nightmare (see [perl #127604]). So, this code
15847 * looks only for interiors of these constructs that match:
15849 * Using \w relaxes the apparent rules a little, without adding much
15850 * danger of mistaking something else for one of these constructs.
15852 * [. .] in some implementations described on the internet is usable to
15853 * escape a character that otherwise is special in bracketed character
15854 * classes. For example [.].] means a literal right bracket instead of
15855 * the ending of the class
15857 * [= =] can legitimately contain a [. .] construct, but we don't
15858 * handle this case, as that [. .] construct will later get parsed
15859 * itself and croak then. And [= =] is checked for even when not under
15860 * /l, as Perl has long done so.
15862 * The code below relies on there being a trailing NUL, so it doesn't
15863 * have to keep checking if the parse ptr < e.
15865 if (temp_ptr[1] == open_char) {
15868 else while ( temp_ptr < e
15869 && (isWORDCHAR(*temp_ptr) || *temp_ptr == '-'))
15874 if (*temp_ptr == open_char) {
15876 if (*temp_ptr == ']') {
15878 if (! found_problem && ! check_only) {
15879 RExC_parse = (char *) temp_ptr;
15880 vFAIL3("POSIX syntax [%c %c] is reserved for future "
15881 "extensions", open_char, open_char);
15884 /* Here, the syntax wasn't completely valid, or else the call
15885 * is to check-only */
15886 if (updated_parse_ptr) {
15887 *updated_parse_ptr = (char *) temp_ptr;
15890 CLEAR_POSIX_WARNINGS_AND_RETURN(OOB_NAMEDCLASS);
15894 /* If we find something that started out to look like one of these
15895 * constructs, but isn't, we continue below so that it can be checked
15896 * for being a class name with a typo of '.' or '=' instead of a colon.
15900 /* Here, we think there is a possibility that a [: :] class was meant, and
15901 * we have the first real character. It could be they think the '^' comes
15904 found_problem = TRUE;
15905 ADD_POSIX_WARNING(p + 1, "the '^' must come after the colon");
15910 found_problem = TRUE;
15914 } while (p < e && isBLANK(*p));
15916 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15920 /* But the first character should be a colon, which they could have easily
15921 * mistyped on a qwerty keyboard as a semi-colon (and which may be hard to
15922 * distinguish from a colon, so treat that as a colon). */
15925 has_opening_colon = TRUE;
15927 else if (*p == ';') {
15928 found_problem = TRUE;
15930 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
15931 has_opening_colon = TRUE;
15934 found_problem = TRUE;
15935 ADD_POSIX_WARNING(p, "there must be a starting ':'");
15937 /* Consider an initial punctuation (not one of the recognized ones) to
15938 * be a left terminator */
15939 if (*p != '^' && *p != ']' && isPUNCT(*p)) {
15944 /* They may think that you can put spaces between the components */
15946 found_problem = TRUE;
15950 } while (p < e && isBLANK(*p));
15952 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15957 /* We consider something like [^:^alnum:]] to not have been intended to
15958 * be a posix class, but XXX maybe we should */
15960 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15967 /* Again, they may think that you can put spaces between the components */
15969 found_problem = TRUE;
15973 } while (p < e && isBLANK(*p));
15975 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15980 /* XXX This ']' may be a typo, and something else was meant. But
15981 * treating it as such creates enough complications, that that
15982 * possibility isn't currently considered here. So we assume that the
15983 * ']' is what is intended, and if we've already found an initial '[',
15984 * this leaves this construct looking like [:] or [:^], which almost
15985 * certainly weren't intended to be posix classes */
15986 if (has_opening_bracket) {
15987 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15990 /* But this function can be called when we parse the colon for
15991 * something like qr/[alpha:]]/, so we back up to look for the
15996 found_problem = TRUE;
15997 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
15999 else if (*p != ':') {
16001 /* XXX We are currently very restrictive here, so this code doesn't
16002 * consider the possibility that, say, /[alpha.]]/ was intended to
16003 * be a posix class. */
16004 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
16007 /* Here we have something like 'foo:]'. There was no initial colon,
16008 * and we back up over 'foo. XXX Unlike the going forward case, we
16009 * don't handle typos of non-word chars in the middle */
16010 has_opening_colon = FALSE;
16013 while (p > RExC_start && isWORDCHAR(*p)) {
16018 /* Here, we have positioned ourselves to where we think the first
16019 * character in the potential class is */
16022 /* Now the interior really starts. There are certain key characters that
16023 * can end the interior, or these could just be typos. To catch both
16024 * cases, we may have to do two passes. In the first pass, we keep on
16025 * going unless we come to a sequence that matches
16026 * qr/ [[:punct:]] [[:blank:]]* \] /xa
16027 * This means it takes a sequence to end the pass, so two typos in a row if
16028 * that wasn't what was intended. If the class is perfectly formed, just
16029 * this one pass is needed. We also stop if there are too many characters
16030 * being accumulated, but this number is deliberately set higher than any
16031 * real class. It is set high enough so that someone who thinks that
16032 * 'alphanumeric' is a correct name would get warned that it wasn't.
16033 * While doing the pass, we keep track of where the key characters were in
16034 * it. If we don't find an end to the class, and one of the key characters
16035 * was found, we redo the pass, but stop when we get to that character.
16036 * Thus the key character was considered a typo in the first pass, but a
16037 * terminator in the second. If two key characters are found, we stop at
16038 * the second one in the first pass. Again this can miss two typos, but
16039 * catches a single one
16041 * In the first pass, 'possible_end' starts as NULL, and then gets set to
16042 * point to the first key character. For the second pass, it starts as -1.
16048 bool has_blank = FALSE;
16049 bool has_upper = FALSE;
16050 bool has_terminating_colon = FALSE;
16051 bool has_terminating_bracket = FALSE;
16052 bool has_semi_colon = FALSE;
16053 unsigned int name_len = 0;
16054 int punct_count = 0;
16058 /* Squeeze out blanks when looking up the class name below */
16059 if (isBLANK(*p) ) {
16061 found_problem = TRUE;
16066 /* The name will end with a punctuation */
16068 const char * peek = p + 1;
16070 /* Treat any non-']' punctuation followed by a ']' (possibly
16071 * with intervening blanks) as trying to terminate the class.
16072 * ']]' is very likely to mean a class was intended (but
16073 * missing the colon), but the warning message that gets
16074 * generated shows the error position better if we exit the
16075 * loop at the bottom (eventually), so skip it here. */
16077 if (peek < e && isBLANK(*peek)) {
16079 found_problem = TRUE;
16082 } while (peek < e && isBLANK(*peek));
16085 if (peek < e && *peek == ']') {
16086 has_terminating_bracket = TRUE;
16088 has_terminating_colon = TRUE;
16090 else if (*p == ';') {
16091 has_semi_colon = TRUE;
16092 has_terminating_colon = TRUE;
16095 found_problem = TRUE;
16102 /* Here we have punctuation we thought didn't end the class.
16103 * Keep track of the position of the key characters that are
16104 * more likely to have been class-enders */
16105 if (*p == ']' || *p == '[' || *p == ':' || *p == ';') {
16107 /* Allow just one such possible class-ender not actually
16108 * ending the class. */
16109 if (possible_end) {
16115 /* If we have too many punctuation characters, no use in
16117 if (++punct_count > max_distance) {
16121 /* Treat the punctuation as a typo. */
16122 input_text[name_len++] = *p;
16125 else if (isUPPER(*p)) { /* Use lowercase for lookup */
16126 input_text[name_len++] = toLOWER(*p);
16128 found_problem = TRUE;
16130 } else if (! UTF || UTF8_IS_INVARIANT(*p)) {
16131 input_text[name_len++] = *p;
16135 input_text[name_len++] = utf8_to_uvchr_buf((U8 *) p, e, NULL);
16139 /* The declaration of 'input_text' is how long we allow a potential
16140 * class name to be, before saying they didn't mean a class name at
16142 if (name_len >= C_ARRAY_LENGTH(input_text)) {
16147 /* We get to here when the possible class name hasn't been properly
16148 * terminated before:
16149 * 1) we ran off the end of the pattern; or
16150 * 2) found two characters, each of which might have been intended to
16151 * be the name's terminator
16152 * 3) found so many punctuation characters in the purported name,
16153 * that the edit distance to a valid one is exceeded
16154 * 4) we decided it was more characters than anyone could have
16155 * intended to be one. */
16157 found_problem = TRUE;
16159 /* In the final two cases, we know that looking up what we've
16160 * accumulated won't lead to a match, even a fuzzy one. */
16161 if ( name_len >= C_ARRAY_LENGTH(input_text)
16162 || punct_count > max_distance)
16164 /* If there was an intermediate key character that could have been
16165 * an intended end, redo the parse, but stop there */
16166 if (possible_end && possible_end != (char *) -1) {
16167 possible_end = (char *) -1; /* Special signal value to say
16168 we've done a first pass */
16173 /* Otherwise, it can't have meant to have been a class */
16174 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
16177 /* If we ran off the end, and the final character was a punctuation
16178 * one, back up one, to look at that final one just below. Later, we
16179 * will restore the parse pointer if appropriate */
16180 if (name_len && p == e && isPUNCT(*(p-1))) {
16185 if (p < e && isPUNCT(*p)) {
16187 has_terminating_bracket = TRUE;
16189 /* If this is a 2nd ']', and the first one is just below this
16190 * one, consider that to be the real terminator. This gives a
16191 * uniform and better positioning for the warning message */
16193 && possible_end != (char *) -1
16194 && *possible_end == ']'
16195 && name_len && input_text[name_len - 1] == ']')
16200 /* And this is actually equivalent to having done the 2nd
16201 * pass now, so set it to not try again */
16202 possible_end = (char *) -1;
16207 has_terminating_colon = TRUE;
16209 else if (*p == ';') {
16210 has_semi_colon = TRUE;
16211 has_terminating_colon = TRUE;
16219 /* Here, we have a class name to look up. We can short circuit the
16220 * stuff below for short names that can't possibly be meant to be a
16221 * class name. (We can do this on the first pass, as any second pass
16222 * will yield an even shorter name) */
16223 if (name_len < 3) {
16224 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
16227 /* Find which class it is. Initially switch on the length of the name.
16229 switch (name_len) {
16231 if (memEQs(name_start, 4, "word")) {
16232 /* this is not POSIX, this is the Perl \w */
16233 class_number = ANYOF_WORDCHAR;
16237 /* Names all of length 5: alnum alpha ascii blank cntrl digit
16238 * graph lower print punct space upper
16239 * Offset 4 gives the best switch position. */
16240 switch (name_start[4]) {
16242 if (memBEGINs(name_start, 5, "alph")) /* alpha */
16243 class_number = ANYOF_ALPHA;
16246 if (memBEGINs(name_start, 5, "spac")) /* space */
16247 class_number = ANYOF_SPACE;
16250 if (memBEGINs(name_start, 5, "grap")) /* graph */
16251 class_number = ANYOF_GRAPH;
16254 if (memBEGINs(name_start, 5, "asci")) /* ascii */
16255 class_number = ANYOF_ASCII;
16258 if (memBEGINs(name_start, 5, "blan")) /* blank */
16259 class_number = ANYOF_BLANK;
16262 if (memBEGINs(name_start, 5, "cntr")) /* cntrl */
16263 class_number = ANYOF_CNTRL;
16266 if (memBEGINs(name_start, 5, "alnu")) /* alnum */
16267 class_number = ANYOF_ALPHANUMERIC;
16270 if (memBEGINs(name_start, 5, "lowe")) /* lower */
16271 class_number = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
16272 else if (memBEGINs(name_start, 5, "uppe")) /* upper */
16273 class_number = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
16276 if (memBEGINs(name_start, 5, "digi")) /* digit */
16277 class_number = ANYOF_DIGIT;
16278 else if (memBEGINs(name_start, 5, "prin")) /* print */
16279 class_number = ANYOF_PRINT;
16280 else if (memBEGINs(name_start, 5, "punc")) /* punct */
16281 class_number = ANYOF_PUNCT;
16286 if (memEQs(name_start, 6, "xdigit"))
16287 class_number = ANYOF_XDIGIT;
16291 /* If the name exactly matches a posix class name the class number will
16292 * here be set to it, and the input almost certainly was meant to be a
16293 * posix class, so we can skip further checking. If instead the syntax
16294 * is exactly correct, but the name isn't one of the legal ones, we
16295 * will return that as an error below. But if neither of these apply,
16296 * it could be that no posix class was intended at all, or that one
16297 * was, but there was a typo. We tease these apart by doing fuzzy
16298 * matching on the name */
16299 if (class_number == OOB_NAMEDCLASS && found_problem) {
16300 const UV posix_names[][6] = {
16301 { 'a', 'l', 'n', 'u', 'm' },
16302 { 'a', 'l', 'p', 'h', 'a' },
16303 { 'a', 's', 'c', 'i', 'i' },
16304 { 'b', 'l', 'a', 'n', 'k' },
16305 { 'c', 'n', 't', 'r', 'l' },
16306 { 'd', 'i', 'g', 'i', 't' },
16307 { 'g', 'r', 'a', 'p', 'h' },
16308 { 'l', 'o', 'w', 'e', 'r' },
16309 { 'p', 'r', 'i', 'n', 't' },
16310 { 'p', 'u', 'n', 'c', 't' },
16311 { 's', 'p', 'a', 'c', 'e' },
16312 { 'u', 'p', 'p', 'e', 'r' },
16313 { 'w', 'o', 'r', 'd' },
16314 { 'x', 'd', 'i', 'g', 'i', 't' }
16316 /* The names of the above all have added NULs to make them the same
16317 * size, so we need to also have the real lengths */
16318 const UV posix_name_lengths[] = {
16319 sizeof("alnum") - 1,
16320 sizeof("alpha") - 1,
16321 sizeof("ascii") - 1,
16322 sizeof("blank") - 1,
16323 sizeof("cntrl") - 1,
16324 sizeof("digit") - 1,
16325 sizeof("graph") - 1,
16326 sizeof("lower") - 1,
16327 sizeof("print") - 1,
16328 sizeof("punct") - 1,
16329 sizeof("space") - 1,
16330 sizeof("upper") - 1,
16331 sizeof("word") - 1,
16332 sizeof("xdigit")- 1
16335 int temp_max = max_distance; /* Use a temporary, so if we
16336 reparse, we haven't changed the
16339 /* Use a smaller max edit distance if we are missing one of the
16341 if ( has_opening_bracket + has_opening_colon < 2
16342 || has_terminating_bracket + has_terminating_colon < 2)
16347 /* See if the input name is close to a legal one */
16348 for (i = 0; i < C_ARRAY_LENGTH(posix_names); i++) {
16350 /* Short circuit call if the lengths are too far apart to be
16352 if (abs( (int) (name_len - posix_name_lengths[i]))
16358 if (edit_distance(input_text,
16361 posix_name_lengths[i],
16365 { /* If it is close, it probably was intended to be a class */
16366 goto probably_meant_to_be;
16370 /* Here the input name is not close enough to a valid class name
16371 * for us to consider it to be intended to be a posix class. If
16372 * we haven't already done so, and the parse found a character that
16373 * could have been terminators for the name, but which we absorbed
16374 * as typos during the first pass, repeat the parse, signalling it
16375 * to stop at that character */
16376 if (possible_end && possible_end != (char *) -1) {
16377 possible_end = (char *) -1;
16382 /* Here neither pass found a close-enough class name */
16383 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
16386 probably_meant_to_be:
16388 /* Here we think that a posix specification was intended. Update any
16390 if (updated_parse_ptr) {
16391 *updated_parse_ptr = (char *) p;
16394 /* If a posix class name was intended but incorrectly specified, we
16395 * output or return the warnings */
16396 if (found_problem) {
16398 /* We set flags for these issues in the parse loop above instead of
16399 * adding them to the list of warnings, because we can parse it
16400 * twice, and we only want one warning instance */
16402 ADD_POSIX_WARNING(p, "the name must be all lowercase letters");
16405 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
16407 if (has_semi_colon) {
16408 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
16410 else if (! has_terminating_colon) {
16411 ADD_POSIX_WARNING(p, "there is no terminating ':'");
16413 if (! has_terminating_bracket) {
16414 ADD_POSIX_WARNING(p, "there is no terminating ']'");
16417 if ( posix_warnings
16419 && av_count(RExC_warn_text) > 0)
16421 *posix_warnings = RExC_warn_text;
16424 else if (class_number != OOB_NAMEDCLASS) {
16425 /* If it is a known class, return the class. The class number
16426 * #defines are structured so each complement is +1 to the normal
16428 CLEAR_POSIX_WARNINGS_AND_RETURN(class_number + complement);
16430 else if (! check_only) {
16432 /* Here, it is an unrecognized class. This is an error (unless the
16433 * call is to check only, which we've already handled above) */
16434 const char * const complement_string = (complement)
16437 RExC_parse = (char *) p;
16438 vFAIL3utf8f("POSIX class [:%s%" UTF8f ":] unknown",
16440 UTF8fARG(UTF, RExC_parse - name_start - 2, name_start));
16444 return OOB_NAMEDCLASS;
16446 #undef ADD_POSIX_WARNING
16448 STATIC unsigned int
16449 S_regex_set_precedence(const U8 my_operator) {
16451 /* Returns the precedence in the (?[...]) construct of the input operator,
16452 * specified by its character representation. The precedence follows
16453 * general Perl rules, but it extends this so that ')' and ']' have (low)
16454 * precedence even though they aren't really operators */
16456 switch (my_operator) {
16472 NOT_REACHED; /* NOTREACHED */
16473 return 0; /* Silence compiler warning */
16476 STATIC regnode_offset
16477 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
16478 I32 *flagp, U32 depth,
16479 char * const oregcomp_parse)
16481 /* Handle the (?[...]) construct to do set operations */
16483 U8 curchar; /* Current character being parsed */
16484 UV start, end; /* End points of code point ranges */
16485 SV* final = NULL; /* The end result inversion list */
16486 SV* result_string; /* 'final' stringified */
16487 AV* stack; /* stack of operators and operands not yet
16489 AV* fence_stack = NULL; /* A stack containing the positions in
16490 'stack' of where the undealt-with left
16491 parens would be if they were actually
16493 /* The 'volatile' is a workaround for an optimiser bug
16494 * in Solaris Studio 12.3. See RT #127455 */
16495 volatile IV fence = 0; /* Position of where most recent undealt-
16496 with left paren in stack is; -1 if none.
16498 STRLEN len; /* Temporary */
16499 regnode_offset node; /* Temporary, and final regnode returned by
16501 const bool save_fold = FOLD; /* Temporary */
16502 char *save_end, *save_parse; /* Temporaries */
16503 const bool in_locale = LOC; /* we turn off /l during processing */
16505 DECLARE_AND_GET_RE_DEBUG_FLAGS;
16507 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
16508 PERL_UNUSED_ARG(oregcomp_parse); /* Only for Set_Node_Length */
16510 DEBUG_PARSE("xcls");
16513 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
16516 /* The use of this operator implies /u. This is required so that the
16517 * compile time values are valid in all runtime cases */
16518 REQUIRE_UNI_RULES(flagp, 0);
16520 ckWARNexperimental(RExC_parse,
16521 WARN_EXPERIMENTAL__REGEX_SETS,
16522 "The regex_sets feature is experimental");
16524 /* Everything in this construct is a metacharacter. Operands begin with
16525 * either a '\' (for an escape sequence), or a '[' for a bracketed
16526 * character class. Any other character should be an operator, or
16527 * parenthesis for grouping. Both types of operands are handled by calling
16528 * regclass() to parse them. It is called with a parameter to indicate to
16529 * return the computed inversion list. The parsing here is implemented via
16530 * a stack. Each entry on the stack is a single character representing one
16531 * of the operators; or else a pointer to an operand inversion list. */
16533 #define IS_OPERATOR(a) SvIOK(a)
16534 #define IS_OPERAND(a) (! IS_OPERATOR(a))
16536 /* The stack is kept in Łukasiewicz order. (That's pronounced similar
16537 * to luke-a-shave-itch (or -itz), but people who didn't want to bother
16538 * with pronouncing it called it Reverse Polish instead, but now that YOU
16539 * know how to pronounce it you can use the correct term, thus giving due
16540 * credit to the person who invented it, and impressing your geek friends.
16541 * Wikipedia says that the pronounciation of "Ł" has been changing so that
16542 * it is now more like an English initial W (as in wonk) than an L.)
16544 * This means that, for example, 'a | b & c' is stored on the stack as
16552 * where the numbers in brackets give the stack [array] element number.
16553 * In this implementation, parentheses are not stored on the stack.
16554 * Instead a '(' creates a "fence" so that the part of the stack below the
16555 * fence is invisible except to the corresponding ')' (this allows us to
16556 * replace testing for parens, by using instead subtraction of the fence
16557 * position). As new operands are processed they are pushed onto the stack
16558 * (except as noted in the next paragraph). New operators of higher
16559 * precedence than the current final one are inserted on the stack before
16560 * the lhs operand (so that when the rhs is pushed next, everything will be
16561 * in the correct positions shown above. When an operator of equal or
16562 * lower precedence is encountered in parsing, all the stacked operations
16563 * of equal or higher precedence are evaluated, leaving the result as the
16564 * top entry on the stack. This makes higher precedence operations
16565 * evaluate before lower precedence ones, and causes operations of equal
16566 * precedence to left associate.
16568 * The only unary operator '!' is immediately pushed onto the stack when
16569 * encountered. When an operand is encountered, if the top of the stack is
16570 * a '!", the complement is immediately performed, and the '!' popped. The
16571 * resulting value is treated as a new operand, and the logic in the
16572 * previous paragraph is executed. Thus in the expression
16574 * the stack looks like
16580 * as 'b' gets parsed, the latter gets evaluated to '!b', and the stack
16587 * A ')' is treated as an operator with lower precedence than all the
16588 * aforementioned ones, which causes all operations on the stack above the
16589 * corresponding '(' to be evaluated down to a single resultant operand.
16590 * Then the fence for the '(' is removed, and the operand goes through the
16591 * algorithm above, without the fence.
16593 * A separate stack is kept of the fence positions, so that the position of
16594 * the latest so-far unbalanced '(' is at the top of it.
16596 * The ']' ending the construct is treated as the lowest operator of all,
16597 * so that everything gets evaluated down to a single operand, which is the
16600 sv_2mortal((SV *)(stack = newAV()));
16601 sv_2mortal((SV *)(fence_stack = newAV()));
16603 while (RExC_parse < RExC_end) {
16604 I32 top_index; /* Index of top-most element in 'stack' */
16605 SV** top_ptr; /* Pointer to top 'stack' element */
16606 SV* current = NULL; /* To contain the current inversion list
16608 SV* only_to_avoid_leaks;
16610 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
16611 TRUE /* Force /x */ );
16612 if (RExC_parse >= RExC_end) { /* Fail */
16616 curchar = UCHARAT(RExC_parse);
16620 #ifdef ENABLE_REGEX_SETS_DEBUGGING
16621 /* Enable with -Accflags=-DENABLE_REGEX_SETS_DEBUGGING */
16622 DEBUG_U(dump_regex_sets_structures(pRExC_state,
16623 stack, fence, fence_stack));
16626 top_index = av_tindex_skip_len_mg(stack);
16629 SV** stacked_ptr; /* Ptr to something already on 'stack' */
16630 char stacked_operator; /* The topmost operator on the 'stack'. */
16631 SV* lhs; /* Operand to the left of the operator */
16632 SV* rhs; /* Operand to the right of the operator */
16633 SV* fence_ptr; /* Pointer to top element of the fence
16637 if ( RExC_parse < RExC_end - 2
16638 && UCHARAT(RExC_parse + 1) == '?'
16639 && UCHARAT(RExC_parse + 2) == '^')
16641 const regnode_offset orig_emit = RExC_emit;
16642 SV * resultant_invlist;
16644 /* If is a '(?^', could be an embedded '(?^flags:(?[...])'.
16645 * This happens when we have some thing like
16647 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
16649 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
16651 * Here we would be handling the interpolated
16652 * '$thai_or_lao'. We handle this by a recursive call to
16653 * reg which returns the inversion list the
16654 * interpolated expression evaluates to. Actually, the
16655 * return is a special regnode containing a pointer to that
16656 * inversion list. If the return isn't that regnode alone,
16657 * we know that this wasn't such an interpolation, which is
16658 * an error: we need to get a single inversion list back
16659 * from the recursion */
16664 node = reg(pRExC_state, 2, flagp, depth+1);
16665 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16667 if ( OP(REGNODE_p(node)) != REGEX_SET
16668 /* If more than a single node returned, the nested
16669 * parens evaluated to more than just a (?[...]),
16670 * which isn't legal */
16671 || RExC_emit != orig_emit
16672 + NODE_STEP_REGNODE
16673 + regarglen[REGEX_SET])
16675 vFAIL("Expecting interpolated extended charclass");
16677 resultant_invlist = (SV *) ARGp(REGNODE_p(node));
16678 current = invlist_clone(resultant_invlist, NULL);
16679 SvREFCNT_dec(resultant_invlist);
16682 RExC_emit = orig_emit;
16683 goto handle_operand;
16686 /* A regular '('. Look behind for illegal syntax */
16687 if (top_index - fence >= 0) {
16688 /* If the top entry on the stack is an operator, it had
16689 * better be a '!', otherwise the entry below the top
16690 * operand should be an operator */
16691 if ( ! (top_ptr = av_fetch(stack, top_index, FALSE))
16692 || (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) != '!')
16693 || ( IS_OPERAND(*top_ptr)
16694 && ( top_index - fence < 1
16695 || ! (stacked_ptr = av_fetch(stack,
16698 || ! IS_OPERATOR(*stacked_ptr))))
16701 vFAIL("Unexpected '(' with no preceding operator");
16705 /* Stack the position of this undealt-with left paren */
16706 av_push(fence_stack, newSViv(fence));
16707 fence = top_index + 1;
16711 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
16712 * multi-char folds are allowed. */
16713 if (!regclass(pRExC_state, flagp, depth+1,
16714 TRUE, /* means parse just the next thing */
16715 FALSE, /* don't allow multi-char folds */
16716 FALSE, /* don't silence non-portable warnings. */
16718 FALSE, /* Require return to be an ANYOF */
16721 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16722 goto regclass_failed;
16727 /* regclass() will return with parsing just the \ sequence,
16728 * leaving the parse pointer at the next thing to parse */
16730 goto handle_operand;
16732 case '[': /* Is a bracketed character class */
16734 /* See if this is a [:posix:] class. */
16735 bool is_posix_class = (OOB_NAMEDCLASS
16736 < handle_possible_posix(pRExC_state,
16740 TRUE /* checking only */));
16741 /* If it is a posix class, leave the parse pointer at the '['
16742 * to fool regclass() into thinking it is part of a
16743 * '[[:posix:]]'. */
16744 if (! is_posix_class) {
16748 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
16749 * multi-char folds are allowed. */
16750 if (!regclass(pRExC_state, flagp, depth+1,
16751 is_posix_class, /* parse the whole char
16752 class only if not a
16754 FALSE, /* don't allow multi-char folds */
16755 TRUE, /* silence non-portable warnings. */
16757 FALSE, /* Require return to be an ANYOF */
16760 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16761 goto regclass_failed;
16766 /* function call leaves parse pointing to the ']', except if we
16768 if (is_posix_class) {
16772 goto handle_operand;
16776 if (top_index >= 1) {
16777 goto join_operators;
16780 /* Only a single operand on the stack: are done */
16784 if (av_tindex_skip_len_mg(fence_stack) < 0) {
16785 if (UCHARAT(RExC_parse - 1) == ']') {
16789 vFAIL("Unexpected ')'");
16792 /* If nothing after the fence, is missing an operand */
16793 if (top_index - fence < 0) {
16797 /* If at least two things on the stack, treat this as an
16799 if (top_index - fence >= 1) {
16800 goto join_operators;
16803 /* Here only a single thing on the fenced stack, and there is a
16804 * fence. Get rid of it */
16805 fence_ptr = av_pop(fence_stack);
16807 fence = SvIV(fence_ptr);
16808 SvREFCNT_dec_NN(fence_ptr);
16815 /* Having gotten rid of the fence, we pop the operand at the
16816 * stack top and process it as a newly encountered operand */
16817 current = av_pop(stack);
16818 if (IS_OPERAND(current)) {
16819 goto handle_operand;
16831 /* These binary operators should have a left operand already
16833 if ( top_index - fence < 0
16834 || top_index - fence == 1
16835 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
16836 || ! IS_OPERAND(*top_ptr))
16838 goto unexpected_binary;
16841 /* If only the one operand is on the part of the stack visible
16842 * to us, we just place this operator in the proper position */
16843 if (top_index - fence < 2) {
16845 /* Place the operator before the operand */
16847 SV* lhs = av_pop(stack);
16848 av_push(stack, newSVuv(curchar));
16849 av_push(stack, lhs);
16853 /* But if there is something else on the stack, we need to
16854 * process it before this new operator if and only if the
16855 * stacked operation has equal or higher precedence than the
16860 /* The operator on the stack is supposed to be below both its
16862 if ( ! (stacked_ptr = av_fetch(stack, top_index - 2, FALSE))
16863 || IS_OPERAND(*stacked_ptr))
16865 /* But if not, it's legal and indicates we are completely
16866 * done if and only if we're currently processing a ']',
16867 * which should be the final thing in the expression */
16868 if (curchar == ']') {
16874 vFAIL2("Unexpected binary operator '%c' with no "
16875 "preceding operand", curchar);
16877 stacked_operator = (char) SvUV(*stacked_ptr);
16879 if (regex_set_precedence(curchar)
16880 > regex_set_precedence(stacked_operator))
16882 /* Here, the new operator has higher precedence than the
16883 * stacked one. This means we need to add the new one to
16884 * the stack to await its rhs operand (and maybe more
16885 * stuff). We put it before the lhs operand, leaving
16886 * untouched the stacked operator and everything below it
16888 lhs = av_pop(stack);
16889 assert(IS_OPERAND(lhs));
16891 av_push(stack, newSVuv(curchar));
16892 av_push(stack, lhs);
16896 /* Here, the new operator has equal or lower precedence than
16897 * what's already there. This means the operation already
16898 * there should be performed now, before the new one. */
16900 rhs = av_pop(stack);
16901 if (! IS_OPERAND(rhs)) {
16903 /* This can happen when a ! is not followed by an operand,
16904 * like in /(?[\t &!])/ */
16908 lhs = av_pop(stack);
16910 if (! IS_OPERAND(lhs)) {
16912 /* This can happen when there is an empty (), like in
16913 * /(?[[0]+()+])/ */
16917 switch (stacked_operator) {
16919 _invlist_intersection(lhs, rhs, &rhs);
16924 _invlist_union(lhs, rhs, &rhs);
16928 _invlist_subtract(lhs, rhs, &rhs);
16931 case '^': /* The union minus the intersection */
16936 _invlist_union(lhs, rhs, &u);
16937 _invlist_intersection(lhs, rhs, &i);
16938 _invlist_subtract(u, i, &rhs);
16939 SvREFCNT_dec_NN(i);
16940 SvREFCNT_dec_NN(u);
16946 /* Here, the higher precedence operation has been done, and the
16947 * result is in 'rhs'. We overwrite the stacked operator with
16948 * the result. Then we redo this code to either push the new
16949 * operator onto the stack or perform any higher precedence
16950 * stacked operation */
16951 only_to_avoid_leaks = av_pop(stack);
16952 SvREFCNT_dec(only_to_avoid_leaks);
16953 av_push(stack, rhs);
16956 case '!': /* Highest priority, right associative */
16958 /* If what's already at the top of the stack is another '!",
16959 * they just cancel each other out */
16960 if ( (top_ptr = av_fetch(stack, top_index, FALSE))
16961 && (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) == '!'))
16963 only_to_avoid_leaks = av_pop(stack);
16964 SvREFCNT_dec(only_to_avoid_leaks);
16966 else { /* Otherwise, since it's right associative, just push
16968 av_push(stack, newSVuv(curchar));
16973 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16974 if (RExC_parse >= RExC_end) {
16977 vFAIL("Unexpected character");
16981 /* Here 'current' is the operand. If something is already on the
16982 * stack, we have to check if it is a !. But first, the code above
16983 * may have altered the stack in the time since we earlier set
16986 top_index = av_tindex_skip_len_mg(stack);
16987 if (top_index - fence >= 0) {
16988 /* If the top entry on the stack is an operator, it had better
16989 * be a '!', otherwise the entry below the top operand should
16990 * be an operator */
16991 top_ptr = av_fetch(stack, top_index, FALSE);
16993 if (IS_OPERATOR(*top_ptr)) {
16995 /* The only permissible operator at the top of the stack is
16996 * '!', which is applied immediately to this operand. */
16997 curchar = (char) SvUV(*top_ptr);
16998 if (curchar != '!') {
16999 SvREFCNT_dec(current);
17000 vFAIL2("Unexpected binary operator '%c' with no "
17001 "preceding operand", curchar);
17004 _invlist_invert(current);
17006 only_to_avoid_leaks = av_pop(stack);
17007 SvREFCNT_dec(only_to_avoid_leaks);
17009 /* And we redo with the inverted operand. This allows
17010 * handling multiple ! in a row */
17011 goto handle_operand;
17013 /* Single operand is ok only for the non-binary ')'
17015 else if ((top_index - fence == 0 && curchar != ')')
17016 || (top_index - fence > 0
17017 && (! (stacked_ptr = av_fetch(stack,
17020 || IS_OPERAND(*stacked_ptr))))
17022 SvREFCNT_dec(current);
17023 vFAIL("Operand with no preceding operator");
17027 /* Here there was nothing on the stack or the top element was
17028 * another operand. Just add this new one */
17029 av_push(stack, current);
17031 } /* End of switch on next parse token */
17033 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
17034 } /* End of loop parsing through the construct */
17036 vFAIL("Syntax error in (?[...])");
17040 if (RExC_parse >= RExC_end || RExC_parse[1] != ')') {
17041 if (RExC_parse < RExC_end) {
17045 vFAIL("Unexpected ']' with no following ')' in (?[...");
17048 if (av_tindex_skip_len_mg(fence_stack) >= 0) {
17049 vFAIL("Unmatched (");
17052 if (av_tindex_skip_len_mg(stack) < 0 /* Was empty */
17053 || ((final = av_pop(stack)) == NULL)
17054 || ! IS_OPERAND(final)
17055 || ! is_invlist(final)
17056 || av_tindex_skip_len_mg(stack) >= 0) /* More left on stack */
17059 SvREFCNT_dec(final);
17060 vFAIL("Incomplete expression within '(?[ ])'");
17063 /* Here, 'final' is the resultant inversion list from evaluating the
17064 * expression. Return it if so requested */
17065 if (return_invlist) {
17066 *return_invlist = final;
17070 if (RExC_sets_depth) { /* If within a recursive call, return in a special
17073 node = regpnode(pRExC_state, REGEX_SET, final);
17077 /* Otherwise generate a resultant node, based on 'final'. regclass()
17078 * is expecting a string of ranges and individual code points */
17079 invlist_iterinit(final);
17080 result_string = newSVpvs("");
17081 while (invlist_iternext(final, &start, &end)) {
17082 if (start == end) {
17083 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}", start);
17086 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}-\\x{%"
17087 UVXf "}", start, end);
17091 /* About to generate an ANYOF (or similar) node from the inversion list
17092 * we have calculated */
17093 save_parse = RExC_parse;
17094 RExC_parse = SvPV(result_string, len);
17095 save_end = RExC_end;
17096 RExC_end = RExC_parse + len;
17097 TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE;
17099 /* We turn off folding around the call, as the class we have
17100 * constructed already has all folding taken into consideration, and we
17101 * don't want regclass() to add to that */
17102 RExC_flags &= ~RXf_PMf_FOLD;
17103 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if multi-char
17104 * folds are allowed. */
17105 node = regclass(pRExC_state, flagp, depth+1,
17106 FALSE, /* means parse the whole char class */
17107 FALSE, /* don't allow multi-char folds */
17108 TRUE, /* silence non-portable warnings. The above may
17109 very well have generated non-portable code
17110 points, but they're valid on this machine */
17111 FALSE, /* similarly, no need for strict */
17113 /* We can optimize into something besides an ANYOF,
17114 * except under /l, which needs to be ANYOF because of
17115 * runtime checks for locale sanity, etc */
17121 RExC_parse = save_parse + 1;
17122 RExC_end = save_end;
17123 SvREFCNT_dec_NN(final);
17124 SvREFCNT_dec_NN(result_string);
17127 RExC_flags |= RXf_PMf_FOLD;
17131 RETURN_FAIL_ON_RESTART(*flagp, flagp);
17132 goto regclass_failed;
17135 /* Fix up the node type if we are in locale. (We have pretended we are
17136 * under /u for the purposes of regclass(), as this construct will only
17137 * work under UTF-8 locales. But now we change the opcode to be ANYOFL
17138 * (so as to cause any warnings about bad locales to be output in
17139 * regexec.c), and add the flag that indicates to check if not in a
17140 * UTF-8 locale. The reason we above forbid optimization into
17141 * something other than an ANYOF node is simply to minimize the number
17142 * of code changes in regexec.c. Otherwise we would have to create new
17143 * EXACTish node types and deal with them. This decision could be
17144 * revisited should this construct become popular.
17146 * (One might think we could look at the resulting ANYOF node and
17147 * suppress the flag if everything is above 255, as those would be
17148 * UTF-8 only, but this isn't true, as the components that led to that
17149 * result could have been locale-affected, and just happen to cancel
17150 * each other out under UTF-8 locales.) */
17152 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
17154 assert(OP(REGNODE_p(node)) == ANYOF);
17156 OP(REGNODE_p(node)) = ANYOFL;
17157 ANYOF_FLAGS(REGNODE_p(node))
17158 |= ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
17162 nextchar(pRExC_state);
17163 Set_Node_Length(REGNODE_p(node), RExC_parse - oregcomp_parse + 1); /* MJD */
17167 FAIL2("panic: regclass returned failure to handle_sets, " "flags=%#" UVxf,
17171 #ifdef ENABLE_REGEX_SETS_DEBUGGING
17174 S_dump_regex_sets_structures(pTHX_ RExC_state_t *pRExC_state,
17175 AV * stack, const IV fence, AV * fence_stack)
17176 { /* Dumps the stacks in handle_regex_sets() */
17178 const SSize_t stack_top = av_tindex_skip_len_mg(stack);
17179 const SSize_t fence_stack_top = av_tindex_skip_len_mg(fence_stack);
17182 PERL_ARGS_ASSERT_DUMP_REGEX_SETS_STRUCTURES;
17184 PerlIO_printf(Perl_debug_log, "\nParse position is:%s\n", RExC_parse);
17186 if (stack_top < 0) {
17187 PerlIO_printf(Perl_debug_log, "Nothing on stack\n");
17190 PerlIO_printf(Perl_debug_log, "Stack: (fence=%d)\n", (int) fence);
17191 for (i = stack_top; i >= 0; i--) {
17192 SV ** element_ptr = av_fetch(stack, i, FALSE);
17193 if (! element_ptr) {
17196 if (IS_OPERATOR(*element_ptr)) {
17197 PerlIO_printf(Perl_debug_log, "[%d]: %c\n",
17198 (int) i, (int) SvIV(*element_ptr));
17201 PerlIO_printf(Perl_debug_log, "[%d] ", (int) i);
17202 sv_dump(*element_ptr);
17207 if (fence_stack_top < 0) {
17208 PerlIO_printf(Perl_debug_log, "Nothing on fence_stack\n");
17211 PerlIO_printf(Perl_debug_log, "Fence_stack: \n");
17212 for (i = fence_stack_top; i >= 0; i--) {
17213 SV ** element_ptr = av_fetch(fence_stack, i, FALSE);
17214 if (! element_ptr) {
17217 PerlIO_printf(Perl_debug_log, "[%d]: %d\n",
17218 (int) i, (int) SvIV(*element_ptr));
17229 S_add_above_Latin1_folds(pTHX_ RExC_state_t *pRExC_state, const U8 cp, SV** invlist)
17231 /* This adds the Latin1/above-Latin1 folding rules.
17233 * This should be called only for a Latin1-range code points, cp, which is
17234 * known to be involved in a simple fold with other code points above
17235 * Latin1. It would give false results if /aa has been specified.
17236 * Multi-char folds are outside the scope of this, and must be handled
17239 PERL_ARGS_ASSERT_ADD_ABOVE_LATIN1_FOLDS;
17241 assert(HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(cp));
17243 /* The rules that are valid for all Unicode versions are hard-coded in */
17248 add_cp_to_invlist(*invlist, KELVIN_SIGN);
17252 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_LONG_S);
17255 *invlist = add_cp_to_invlist(*invlist, GREEK_CAPITAL_LETTER_MU);
17256 *invlist = add_cp_to_invlist(*invlist, GREEK_SMALL_LETTER_MU);
17258 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
17259 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
17260 *invlist = add_cp_to_invlist(*invlist, ANGSTROM_SIGN);
17262 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
17263 *invlist = add_cp_to_invlist(*invlist,
17264 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
17267 default: /* Other code points are checked against the data for the
17268 current Unicode version */
17270 Size_t folds_count;
17272 const U32 * remaining_folds;
17276 folded_cp = toFOLD(cp);
17279 U8 dummy_fold[UTF8_MAXBYTES_CASE+1];
17281 folded_cp = _to_fold_latin1(cp, dummy_fold, &dummy_len, 0);
17284 if (folded_cp > 255) {
17285 *invlist = add_cp_to_invlist(*invlist, folded_cp);
17288 folds_count = _inverse_folds(folded_cp, &first_fold,
17290 if (folds_count == 0) {
17292 /* Use deprecated warning to increase the chances of this being
17294 ckWARN2reg_d(RExC_parse,
17295 "Perl folding rules are not up-to-date for 0x%02X;"
17296 " please use the perlbug utility to report;", cp);
17301 if (first_fold > 255) {
17302 *invlist = add_cp_to_invlist(*invlist, first_fold);
17304 for (i = 0; i < folds_count - 1; i++) {
17305 if (remaining_folds[i] > 255) {
17306 *invlist = add_cp_to_invlist(*invlist,
17307 remaining_folds[i]);
17317 S_output_posix_warnings(pTHX_ RExC_state_t *pRExC_state, AV* posix_warnings)
17319 /* Output the elements of the array given by '*posix_warnings' as REGEXP
17323 const bool first_is_fatal = ckDEAD(packWARN(WARN_REGEXP));
17325 PERL_ARGS_ASSERT_OUTPUT_POSIX_WARNINGS;
17327 if (! TO_OUTPUT_WARNINGS(RExC_parse)) {
17328 CLEAR_POSIX_WARNINGS();
17332 while ((msg = av_shift(posix_warnings)) != &PL_sv_undef) {
17333 if (first_is_fatal) { /* Avoid leaking this */
17334 av_undef(posix_warnings); /* This isn't necessary if the
17335 array is mortal, but is a
17337 (void) sv_2mortal(msg);
17340 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s", SvPVX(msg));
17341 SvREFCNT_dec_NN(msg);
17344 UPDATE_WARNINGS_LOC(RExC_parse);
17347 PERL_STATIC_INLINE Size_t
17348 S_find_first_differing_byte_pos(const U8 * s1, const U8 * s2, const Size_t max)
17350 const U8 * const start = s1;
17351 const U8 * const send = start + max;
17353 PERL_ARGS_ASSERT_FIND_FIRST_DIFFERING_BYTE_POS;
17355 while (s1 < send && *s1 == *s2) {
17364 S_add_multi_match(pTHX_ AV* multi_char_matches, SV* multi_string, const STRLEN cp_count)
17366 /* This adds the string scalar <multi_string> to the array
17367 * <multi_char_matches>. <multi_string> is known to have exactly
17368 * <cp_count> code points in it. This is used when constructing a
17369 * bracketed character class and we find something that needs to match more
17370 * than a single character.
17372 * <multi_char_matches> is actually an array of arrays. Each top-level
17373 * element is an array that contains all the strings known so far that are
17374 * the same length. And that length (in number of code points) is the same
17375 * as the index of the top-level array. Hence, the [2] element is an
17376 * array, each element thereof is a string containing TWO code points;
17377 * while element [3] is for strings of THREE characters, and so on. Since
17378 * this is for multi-char strings there can never be a [0] nor [1] element.
17380 * When we rewrite the character class below, we will do so such that the
17381 * longest strings are written first, so that it prefers the longest
17382 * matching strings first. This is done even if it turns out that any
17383 * quantifier is non-greedy, out of this programmer's (khw) laziness. Tom
17384 * Christiansen has agreed that this is ok. This makes the test for the
17385 * ligature 'ffi' come before the test for 'ff', for example */
17388 AV** this_array_ptr;
17390 PERL_ARGS_ASSERT_ADD_MULTI_MATCH;
17392 if (! multi_char_matches) {
17393 multi_char_matches = newAV();
17396 if (av_exists(multi_char_matches, cp_count)) {
17397 this_array_ptr = (AV**) av_fetch(multi_char_matches, cp_count, FALSE);
17398 this_array = *this_array_ptr;
17401 this_array = newAV();
17402 av_store(multi_char_matches, cp_count,
17405 av_push(this_array, multi_string);
17407 return multi_char_matches;
17410 /* The names of properties whose definitions are not known at compile time are
17411 * stored in this SV, after a constant heading. So if the length has been
17412 * changed since initialization, then there is a run-time definition. */
17413 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
17414 (SvCUR(listsv) != initial_listsv_len)
17416 /* There is a restricted set of white space characters that are legal when
17417 * ignoring white space in a bracketed character class. This generates the
17418 * code to skip them.
17420 * There is a line below that uses the same white space criteria but is outside
17421 * this macro. Both here and there must use the same definition */
17422 #define SKIP_BRACKETED_WHITE_SPACE(do_skip, p, stop_p) \
17425 while (p < stop_p && isBLANK_A(UCHARAT(p))) \
17432 STATIC regnode_offset
17433 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
17434 const bool stop_at_1, /* Just parse the next thing, don't
17435 look for a full character class */
17436 bool allow_mutiple_chars,
17437 const bool silence_non_portable, /* Don't output warnings
17441 bool optimizable, /* ? Allow a non-ANYOF return
17443 SV** ret_invlist /* Return an inversion list, not a node */
17446 /* parse a bracketed class specification. Most of these will produce an
17447 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
17448 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
17449 * under /i with multi-character folds: it will be rewritten following the
17450 * paradigm of this example, where the <multi-fold>s are characters which
17451 * fold to multiple character sequences:
17452 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
17453 * gets effectively rewritten as:
17454 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
17455 * reg() gets called (recursively) on the rewritten version, and this
17456 * function will return what it constructs. (Actually the <multi-fold>s
17457 * aren't physically removed from the [abcdefghi], it's just that they are
17458 * ignored in the recursion by means of a flag:
17459 * <RExC_in_multi_char_class>.)
17461 * ANYOF nodes contain a bit map for the first NUM_ANYOF_CODE_POINTS
17462 * characters, with the corresponding bit set if that character is in the
17463 * list. For characters above this, an inversion list is used. There
17464 * are extra bits for \w, etc. in locale ANYOFs, as what these match is not
17465 * determinable at compile time
17467 * On success, returns the offset at which any next node should be placed
17468 * into the regex engine program being compiled.
17470 * Returns 0 otherwise, setting flagp to RESTART_PARSE if the parse needs
17471 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to
17475 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
17477 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
17478 regnode_offset ret = -1; /* Initialized to an illegal value */
17480 int namedclass = OOB_NAMEDCLASS;
17481 char *rangebegin = NULL;
17482 SV *listsv = NULL; /* List of \p{user-defined} whose definitions
17483 aren't available at the time this was called */
17484 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
17485 than just initialized. */
17486 SV* properties = NULL; /* Code points that match \p{} \P{} */
17487 SV* posixes = NULL; /* Code points that match classes like [:word:],
17488 extended beyond the Latin1 range. These have to
17489 be kept separate from other code points for much
17490 of this function because their handling is
17491 different under /i, and for most classes under
17493 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
17494 separate for a while from the non-complemented
17495 versions because of complications with /d
17497 SV* simple_posixes = NULL; /* But under some conditions, the classes can be
17498 treated more simply than the general case,
17499 leading to less compilation and execution
17501 UV element_count = 0; /* Number of distinct elements in the class.
17502 Optimizations may be possible if this is tiny */
17503 AV * multi_char_matches = NULL; /* Code points that fold to more than one
17504 character; used under /i */
17506 char * stop_ptr = RExC_end; /* where to stop parsing */
17508 /* ignore unescaped whitespace? */
17509 const bool skip_white = cBOOL( ret_invlist
17510 || (RExC_flags & RXf_PMf_EXTENDED_MORE));
17512 /* inversion list of code points this node matches only when the target
17513 * string is in UTF-8. These are all non-ASCII, < 256. (Because is under
17515 SV* upper_latin1_only_utf8_matches = NULL;
17517 /* Inversion list of code points this node matches regardless of things
17518 * like locale, folding, utf8ness of the target string */
17519 SV* cp_list = NULL;
17521 /* Like cp_list, but code points on this list need to be checked for things
17522 * that fold to/from them under /i */
17523 SV* cp_foldable_list = NULL;
17525 /* Like cp_list, but code points on this list are valid only when the
17526 * runtime locale is UTF-8 */
17527 SV* only_utf8_locale_list = NULL;
17529 /* In a range, if one of the endpoints is non-character-set portable,
17530 * meaning that it hard-codes a code point that may mean a different
17531 * charactger in ASCII vs. EBCDIC, as opposed to, say, a literal 'A' or a
17532 * mnemonic '\t' which each mean the same character no matter which
17533 * character set the platform is on. */
17534 unsigned int non_portable_endpoint = 0;
17536 /* Is the range unicode? which means on a platform that isn't 1-1 native
17537 * to Unicode (i.e. non-ASCII), each code point in it should be considered
17538 * to be a Unicode value. */
17539 bool unicode_range = FALSE;
17540 bool invert = FALSE; /* Is this class to be complemented */
17542 bool warn_super = ALWAYS_WARN_SUPER;
17544 const char * orig_parse = RExC_parse;
17546 /* This variable is used to mark where the end in the input is of something
17547 * that looks like a POSIX construct but isn't. During the parse, when
17548 * something looks like it could be such a construct is encountered, it is
17549 * checked for being one, but not if we've already checked this area of the
17550 * input. Only after this position is reached do we check again */
17551 char *not_posix_region_end = RExC_parse - 1;
17553 AV* posix_warnings = NULL;
17554 const bool do_posix_warnings = ckWARN(WARN_REGEXP);
17555 U8 op = ANYOF; /* The returned node-type, initialized the expected type.
17557 U8 anyof_flags = 0; /* flag bits if the node is an ANYOF-type */
17558 U32 posixl = 0; /* bit field of posix classes matched under /l */
17561 /* Flags as to what things aren't knowable until runtime. (Note that these are
17562 * mutually exclusive.) */
17563 #define HAS_USER_DEFINED_PROPERTY 0x01 /* /u any user-defined properties that
17564 haven't been defined as of yet */
17565 #define HAS_D_RUNTIME_DEPENDENCY 0x02 /* /d if the target being matched is
17567 #define HAS_L_RUNTIME_DEPENDENCY 0x04 /* /l what the posix classes match and
17568 what gets folded */
17569 U32 has_runtime_dependency = 0; /* OR of the above flags */
17571 DECLARE_AND_GET_RE_DEBUG_FLAGS;
17573 PERL_ARGS_ASSERT_REGCLASS;
17575 PERL_UNUSED_ARG(depth);
17578 assert(! (ret_invlist && allow_mutiple_chars));
17580 /* If wants an inversion list returned, we can't optimize to something
17583 optimizable = FALSE;
17586 DEBUG_PARSE("clas");
17588 #if UNICODE_MAJOR_VERSION < 3 /* no multifolds in early Unicode */ \
17589 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0 \
17590 && UNICODE_DOT_DOT_VERSION == 0)
17591 allow_mutiple_chars = FALSE;
17594 /* We include the /i status at the beginning of this so that we can
17595 * know it at runtime */
17596 listsv = sv_2mortal(Perl_newSVpvf(aTHX_ "#%d\n", cBOOL(FOLD)));
17597 initial_listsv_len = SvCUR(listsv);
17598 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
17600 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse, RExC_end);
17602 assert(RExC_parse <= RExC_end);
17604 if (UCHARAT(RExC_parse) == '^') { /* Complement the class */
17607 allow_mutiple_chars = FALSE;
17609 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse, RExC_end);
17612 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
17613 if (! ret_invlist && MAYBE_POSIXCC(UCHARAT(RExC_parse))) {
17614 int maybe_class = handle_possible_posix(pRExC_state,
17616 ¬_posix_region_end,
17618 TRUE /* checking only */);
17619 if (maybe_class >= OOB_NAMEDCLASS && do_posix_warnings) {
17620 ckWARN4reg(not_posix_region_end,
17621 "POSIX syntax [%c %c] belongs inside character classes%s",
17622 *RExC_parse, *RExC_parse,
17623 (maybe_class == OOB_NAMEDCLASS)
17624 ? ((POSIXCC_NOTYET(*RExC_parse))
17625 ? " (but this one isn't implemented)"
17626 : " (but this one isn't fully valid)")
17632 /* If the caller wants us to just parse a single element, accomplish this
17633 * by faking the loop ending condition */
17634 if (stop_at_1 && RExC_end > RExC_parse) {
17635 stop_ptr = RExC_parse + 1;
17638 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
17639 if (UCHARAT(RExC_parse) == ']')
17640 goto charclassloop;
17644 if ( posix_warnings
17645 && av_tindex_skip_len_mg(posix_warnings) >= 0
17646 && RExC_parse > not_posix_region_end)
17648 /* Warnings about posix class issues are considered tentative until
17649 * we are far enough along in the parse that we can no longer
17650 * change our mind, at which point we output them. This is done
17651 * each time through the loop so that a later class won't zap them
17652 * before they have been dealt with. */
17653 output_posix_warnings(pRExC_state, posix_warnings);
17656 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse, RExC_end);
17658 if (RExC_parse >= stop_ptr) {
17662 if (UCHARAT(RExC_parse) == ']') {
17668 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
17669 save_value = value;
17670 save_prevvalue = prevvalue;
17673 rangebegin = RExC_parse;
17675 non_portable_endpoint = 0;
17677 if (UTF && ! UTF8_IS_INVARIANT(* RExC_parse)) {
17678 value = utf8n_to_uvchr((U8*)RExC_parse,
17679 RExC_end - RExC_parse,
17680 &numlen, UTF8_ALLOW_DEFAULT);
17681 RExC_parse += numlen;
17684 value = UCHARAT(RExC_parse++);
17686 if (value == '[') {
17687 char * posix_class_end;
17688 namedclass = handle_possible_posix(pRExC_state,
17691 do_posix_warnings ? &posix_warnings : NULL,
17692 FALSE /* die if error */);
17693 if (namedclass > OOB_NAMEDCLASS) {
17695 /* If there was an earlier attempt to parse this particular
17696 * posix class, and it failed, it was a false alarm, as this
17697 * successful one proves */
17698 if ( posix_warnings
17699 && av_tindex_skip_len_mg(posix_warnings) >= 0
17700 && not_posix_region_end >= RExC_parse
17701 && not_posix_region_end <= posix_class_end)
17703 av_undef(posix_warnings);
17706 RExC_parse = posix_class_end;
17708 else if (namedclass == OOB_NAMEDCLASS) {
17709 not_posix_region_end = posix_class_end;
17712 namedclass = OOB_NAMEDCLASS;
17715 else if ( RExC_parse - 1 > not_posix_region_end
17716 && MAYBE_POSIXCC(value))
17718 (void) handle_possible_posix(
17720 RExC_parse - 1, /* -1 because parse has already been
17722 ¬_posix_region_end,
17723 do_posix_warnings ? &posix_warnings : NULL,
17724 TRUE /* checking only */);
17726 else if ( strict && ! skip_white
17727 && ( _generic_isCC(value, _CC_VERTSPACE)
17728 || is_VERTWS_cp_high(value)))
17730 vFAIL("Literal vertical space in [] is illegal except under /x");
17732 else if (value == '\\') {
17733 /* Is a backslash; get the code point of the char after it */
17735 if (RExC_parse >= RExC_end) {
17736 vFAIL("Unmatched [");
17739 if (UTF && ! UTF8_IS_INVARIANT(UCHARAT(RExC_parse))) {
17740 value = utf8n_to_uvchr((U8*)RExC_parse,
17741 RExC_end - RExC_parse,
17742 &numlen, UTF8_ALLOW_DEFAULT);
17743 RExC_parse += numlen;
17746 value = UCHARAT(RExC_parse++);
17748 /* Some compilers cannot handle switching on 64-bit integer
17749 * values, therefore value cannot be an UV. Yes, this will
17750 * be a problem later if we want switch on Unicode.
17751 * A similar issue a little bit later when switching on
17752 * namedclass. --jhi */
17754 /* If the \ is escaping white space when white space is being
17755 * skipped, it means that that white space is wanted literally, and
17756 * is already in 'value'. Otherwise, need to translate the escape
17757 * into what it signifies. */
17758 if (! skip_white || ! isBLANK_A(value)) switch ((I32)value) {
17759 const char * message;
17763 case 'w': namedclass = ANYOF_WORDCHAR; break;
17764 case 'W': namedclass = ANYOF_NWORDCHAR; break;
17765 case 's': namedclass = ANYOF_SPACE; break;
17766 case 'S': namedclass = ANYOF_NSPACE; break;
17767 case 'd': namedclass = ANYOF_DIGIT; break;
17768 case 'D': namedclass = ANYOF_NDIGIT; break;
17769 case 'v': namedclass = ANYOF_VERTWS; break;
17770 case 'V': namedclass = ANYOF_NVERTWS; break;
17771 case 'h': namedclass = ANYOF_HORIZWS; break;
17772 case 'H': namedclass = ANYOF_NHORIZWS; break;
17773 case 'N': /* Handle \N{NAME} in class */
17775 const char * const backslash_N_beg = RExC_parse - 2;
17778 if (! grok_bslash_N(pRExC_state,
17779 NULL, /* No regnode */
17780 &value, /* Yes single value */
17781 &cp_count, /* Multiple code pt count */
17787 if (*flagp & NEED_UTF8)
17788 FAIL("panic: grok_bslash_N set NEED_UTF8");
17790 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
17792 if (cp_count < 0) {
17793 vFAIL("\\N in a character class must be a named character: \\N{...}");
17795 else if (cp_count == 0) {
17796 ckWARNreg(RExC_parse,
17797 "Ignoring zero length \\N{} in character class");
17799 else { /* cp_count > 1 */
17800 assert(cp_count > 1);
17801 if (! RExC_in_multi_char_class) {
17802 if ( ! allow_mutiple_chars
17805 || *RExC_parse == '-')
17809 vFAIL("\\N{} here is restricted to one character");
17811 ckWARNreg(RExC_parse, "Using just the first character returned by \\N{} in character class");
17812 break; /* <value> contains the first code
17813 point. Drop out of the switch to
17817 SV * multi_char_N = newSVpvn(backslash_N_beg,
17818 RExC_parse - backslash_N_beg);
17820 = add_multi_match(multi_char_matches,
17825 } /* End of cp_count != 1 */
17827 /* This element should not be processed further in this
17830 value = save_value;
17831 prevvalue = save_prevvalue;
17832 continue; /* Back to top of loop to get next char */
17835 /* Here, is a single code point, and <value> contains it */
17836 unicode_range = TRUE; /* \N{} are Unicode */
17844 if (RExC_pm_flags & PMf_WILDCARD) {
17846 /* diag_listed_as: Use of %s is not allowed in Unicode
17847 property wildcard subpatterns in regex; marked by <--
17849 vFAIL3("Use of '\\%c%c' is not allowed in Unicode property"
17850 " wildcard subpatterns", (char) value, *(RExC_parse - 1));
17853 /* \p means they want Unicode semantics */
17854 REQUIRE_UNI_RULES(flagp, 0);
17856 if (RExC_parse >= RExC_end)
17857 vFAIL2("Empty \\%c", (U8)value);
17858 if (*RExC_parse == '{') {
17859 const U8 c = (U8)value;
17860 e = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
17863 vFAIL2("Missing right brace on \\%c{}", c);
17868 /* White space is allowed adjacent to the braces and after
17869 * any '^', even when not under /x */
17870 while (isSPACE(*RExC_parse)) {
17874 if (UCHARAT(RExC_parse) == '^') {
17876 /* toggle. (The rhs xor gets the single bit that
17877 * differs between P and p; the other xor inverts just
17879 value ^= 'P' ^ 'p';
17882 while (isSPACE(*RExC_parse)) {
17887 if (e == RExC_parse)
17888 vFAIL2("Empty \\%c{}", c);
17890 n = e - RExC_parse;
17891 while (isSPACE(*(RExC_parse + n - 1)))
17894 } /* The \p isn't immediately followed by a '{' */
17895 else if (! isALPHA(*RExC_parse)) {
17896 RExC_parse += (UTF)
17897 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
17899 vFAIL2("Character following \\%c must be '{' or a "
17900 "single-character Unicode property name",
17908 char* name = RExC_parse;
17910 /* Any message returned about expanding the definition */
17911 SV* msg = newSVpvs_flags("", SVs_TEMP);
17913 /* If set TRUE, the property is user-defined as opposed to
17914 * official Unicode */
17915 bool user_defined = FALSE;
17916 AV * strings = NULL;
17918 SV * prop_definition = parse_uniprop_string(
17919 name, n, UTF, FOLD,
17920 FALSE, /* This is compile-time */
17922 /* We can't defer this defn when
17923 * the full result is required in
17925 ! cBOOL(ret_invlist),
17932 if (SvCUR(msg)) { /* Assumes any error causes a msg */
17933 assert(prop_definition == NULL);
17934 RExC_parse = e + 1;
17935 if (SvUTF8(msg)) { /* msg being UTF-8 makes the whole
17936 thing so, or else the display is
17940 /* diag_listed_as: Can't find Unicode property definition "%s" in regex; marked by <-- HERE in m/%s/ */
17941 vFAIL2utf8f("%" UTF8f, UTF8fARG(SvUTF8(msg),
17942 SvCUR(msg), SvPVX(msg)));
17945 assert(prop_definition || strings);
17949 if (! prop_definition) {
17950 RExC_parse = e + 1;
17951 vFAIL("Unicode string properties are not implemented in (?[...])");
17955 "Using just the single character results"
17956 " returned by \\p{} in (?[...])");
17959 else if (! RExC_in_multi_char_class) {
17960 if (invert ^ (value == 'P')) {
17961 RExC_parse = e + 1;
17962 vFAIL("Inverting a character class which contains"
17963 " a multi-character sequence is illegal");
17966 /* For each multi-character string ... */
17967 while (av_count(strings) > 0) {
17968 /* ... Each entry is itself an array of code
17970 AV * this_string = (AV *) av_shift( strings);
17971 STRLEN cp_count = av_count(this_string);
17972 SV * final = newSV(cp_count * 4);
17975 /* Create another string of sequences of \x{...} */
17976 while (av_count(this_string) > 0) {
17977 SV * character = av_shift(this_string);
17978 UV cp = SvUV(character);
17981 REQUIRE_UTF8(flagp);
17983 Perl_sv_catpvf(aTHX_ final, "\\x{%" UVXf "}",
17985 SvREFCNT_dec_NN(character);
17987 SvREFCNT_dec_NN(this_string);
17989 /* And add that to the list of such things */
17991 = add_multi_match(multi_char_matches,
17996 SvREFCNT_dec_NN(strings);
17999 if (! prop_definition) { /* If we got only a string,
18000 this iteration didn't really
18001 find a character */
18004 else if (! is_invlist(prop_definition)) {
18006 /* Here, the definition isn't known, so we have gotten
18007 * returned a string that will be evaluated if and when
18008 * encountered at runtime. We add it to the list of
18009 * such properties, along with whether it should be
18010 * complemented or not */
18011 if (value == 'P') {
18012 sv_catpvs(listsv, "!");
18015 sv_catpvs(listsv, "+");
18017 sv_catsv(listsv, prop_definition);
18019 has_runtime_dependency |= HAS_USER_DEFINED_PROPERTY;
18021 /* We don't know yet what this matches, so have to flag
18023 anyof_flags |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
18026 assert (prop_definition && is_invlist(prop_definition));
18028 /* Here we do have the complete property definition
18030 * Temporary workaround for [perl #133136]. For this
18031 * precise input that is in the .t that is failing,
18032 * load utf8.pm, which is what the test wants, so that
18033 * that .t passes */
18034 if ( memEQs(RExC_start, e + 1 - RExC_start,
18036 && ! hv_common(GvHVn(PL_incgv),
18038 "utf8.pm", sizeof("utf8.pm") - 1,
18039 0, HV_FETCH_ISEXISTS, NULL, 0))
18041 require_pv("utf8.pm");
18044 if (! user_defined &&
18045 /* We warn on matching an above-Unicode code point
18046 * if the match would return true, except don't
18047 * warn for \p{All}, which has exactly one element
18049 (_invlist_contains_cp(prop_definition, 0x110000)
18050 && (! (_invlist_len(prop_definition) == 1
18051 && *invlist_array(prop_definition) == 0))))
18056 /* Invert if asking for the complement */
18057 if (value == 'P') {
18058 _invlist_union_complement_2nd(properties,
18063 _invlist_union(properties, prop_definition, &properties);
18068 RExC_parse = e + 1;
18069 namedclass = ANYOF_UNIPROP; /* no official name, but it's
18073 case 'n': value = '\n'; break;
18074 case 'r': value = '\r'; break;
18075 case 't': value = '\t'; break;
18076 case 'f': value = '\f'; break;
18077 case 'b': value = '\b'; break;
18078 case 'e': value = ESC_NATIVE; break;
18079 case 'a': value = '\a'; break;
18081 RExC_parse--; /* function expects to be pointed at the 'o' */
18082 if (! grok_bslash_o(&RExC_parse,
18088 cBOOL(range), /* MAX_UV allowed for range
18094 else if (message && TO_OUTPUT_WARNINGS(RExC_parse)) {
18095 warn_non_literal_string(RExC_parse, packed_warn, message);
18099 non_portable_endpoint++;
18103 RExC_parse--; /* function expects to be pointed at the 'x' */
18104 if (! grok_bslash_x(&RExC_parse,
18110 cBOOL(range), /* MAX_UV allowed for range
18116 else if (message && TO_OUTPUT_WARNINGS(RExC_parse)) {
18117 warn_non_literal_string(RExC_parse, packed_warn, message);
18121 non_portable_endpoint++;
18125 if (! grok_bslash_c(*RExC_parse, &grok_c_char, &message,
18128 /* going to die anyway; point to exact spot of
18130 RExC_parse += (UTF)
18131 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
18136 value = grok_c_char;
18138 if (message && TO_OUTPUT_WARNINGS(RExC_parse)) {
18139 warn_non_literal_string(RExC_parse, packed_warn, message);
18142 non_portable_endpoint++;
18144 case '0': case '1': case '2': case '3': case '4':
18145 case '5': case '6': case '7':
18147 /* Take 1-3 octal digits */
18148 I32 flags = PERL_SCAN_SILENT_ILLDIGIT
18149 | PERL_SCAN_NOTIFY_ILLDIGIT;
18150 numlen = (strict) ? 4 : 3;
18151 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
18152 RExC_parse += numlen;
18155 RExC_parse += (UTF)
18156 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
18158 vFAIL("Need exactly 3 octal digits");
18160 else if ( (flags & PERL_SCAN_NOTIFY_ILLDIGIT)
18161 && RExC_parse < RExC_end
18162 && isDIGIT(*RExC_parse)
18163 && ckWARN(WARN_REGEXP))
18165 reg_warn_non_literal_string(
18167 form_alien_digit_msg(8, numlen, RExC_parse,
18168 RExC_end, UTF, FALSE));
18172 non_portable_endpoint++;
18177 /* Allow \_ to not give an error */
18178 if (isWORDCHAR(value) && value != '_') {
18180 vFAIL2("Unrecognized escape \\%c in character class",
18184 ckWARN2reg(RExC_parse,
18185 "Unrecognized escape \\%c in character class passed through",
18190 } /* End of switch on char following backslash */
18191 } /* end of handling backslash escape sequences */
18193 /* Here, we have the current token in 'value' */
18195 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
18198 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
18199 * literal, as is the character that began the false range, i.e.
18200 * the 'a' in the examples */
18202 const int w = (RExC_parse >= rangebegin)
18203 ? RExC_parse - rangebegin
18207 "False [] range \"%" UTF8f "\"",
18208 UTF8fARG(UTF, w, rangebegin));
18211 ckWARN2reg(RExC_parse,
18212 "False [] range \"%" UTF8f "\"",
18213 UTF8fARG(UTF, w, rangebegin));
18214 cp_list = add_cp_to_invlist(cp_list, '-');
18215 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
18219 range = 0; /* this was not a true range */
18220 element_count += 2; /* So counts for three values */
18223 classnum = namedclass_to_classnum(namedclass);
18225 if (LOC && namedclass < ANYOF_POSIXL_MAX
18226 #ifndef HAS_ISASCII
18227 && classnum != _CC_ASCII
18230 SV* scratch_list = NULL;
18232 /* What the Posix classes (like \w, [:space:]) match isn't
18233 * generally knowable under locale until actual match time. A
18234 * special node is used for these which has extra space for a
18235 * bitmap, with a bit reserved for each named class that is to
18236 * be matched against. (This isn't needed for \p{} and
18237 * pseudo-classes, as they are not affected by locale, and
18238 * hence are dealt with separately.) However, if a named class
18239 * and its complement are both present, then it matches
18240 * everything, and there is no runtime dependency. Odd numbers
18241 * are the complements of the next lower number, so xor works.
18242 * (Note that something like [\w\D] should match everything,
18243 * because \d should be a proper subset of \w. But rather than
18244 * trust that the locale is well behaved, we leave this to
18245 * runtime to sort out) */
18246 if (POSIXL_TEST(posixl, namedclass ^ 1)) {
18247 cp_list = _add_range_to_invlist(cp_list, 0, UV_MAX);
18248 POSIXL_ZERO(posixl);
18249 has_runtime_dependency &= ~HAS_L_RUNTIME_DEPENDENCY;
18250 anyof_flags &= ~ANYOF_MATCHES_POSIXL;
18251 continue; /* We could ignore the rest of the class, but
18252 best to parse it for any errors */
18254 else { /* Here, isn't the complement of any already parsed
18256 POSIXL_SET(posixl, namedclass);
18257 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
18258 anyof_flags |= ANYOF_MATCHES_POSIXL;
18260 /* The above-Latin1 characters are not subject to locale
18261 * rules. Just add them to the unconditionally-matched
18264 /* Get the list of the above-Latin1 code points this
18266 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
18267 PL_XPosix_ptrs[classnum],
18269 /* Odd numbers are complements,
18270 * like NDIGIT, NASCII, ... */
18271 namedclass % 2 != 0,
18273 /* Checking if 'cp_list' is NULL first saves an extra
18274 * clone. Its reference count will be decremented at the
18275 * next union, etc, or if this is the only instance, at the
18276 * end of the routine */
18278 cp_list = scratch_list;
18281 _invlist_union(cp_list, scratch_list, &cp_list);
18282 SvREFCNT_dec_NN(scratch_list);
18284 continue; /* Go get next character */
18289 /* Here, is not /l, or is a POSIX class for which /l doesn't
18290 * matter (or is a Unicode property, which is skipped here). */
18291 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
18292 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
18294 /* Here, should be \h, \H, \v, or \V. None of /d, /i
18295 * nor /l make a difference in what these match,
18296 * therefore we just add what they match to cp_list. */
18297 if (classnum != _CC_VERTSPACE) {
18298 assert( namedclass == ANYOF_HORIZWS
18299 || namedclass == ANYOF_NHORIZWS);
18301 /* It turns out that \h is just a synonym for
18303 classnum = _CC_BLANK;
18306 _invlist_union_maybe_complement_2nd(
18308 PL_XPosix_ptrs[classnum],
18309 namedclass % 2 != 0, /* Complement if odd
18310 (NHORIZWS, NVERTWS)
18315 else if ( AT_LEAST_UNI_SEMANTICS
18316 || classnum == _CC_ASCII
18317 || (DEPENDS_SEMANTICS && ( classnum == _CC_DIGIT
18318 || classnum == _CC_XDIGIT)))
18320 /* We usually have to worry about /d affecting what POSIX
18321 * classes match, with special code needed because we won't
18322 * know until runtime what all matches. But there is no
18323 * extra work needed under /u and /a; and [:ascii:] is
18324 * unaffected by /d; and :digit: and :xdigit: don't have
18325 * runtime differences under /d. So we can special case
18326 * these, and avoid some extra work below, and at runtime.
18328 _invlist_union_maybe_complement_2nd(
18330 ((AT_LEAST_ASCII_RESTRICTED)
18331 ? PL_Posix_ptrs[classnum]
18332 : PL_XPosix_ptrs[classnum]),
18333 namedclass % 2 != 0,
18336 else { /* Garden variety class. If is NUPPER, NALPHA, ...
18337 complement and use nposixes */
18338 SV** posixes_ptr = namedclass % 2 == 0
18341 _invlist_union_maybe_complement_2nd(
18343 PL_XPosix_ptrs[classnum],
18344 namedclass % 2 != 0,
18348 } /* end of namedclass \blah */
18350 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse, RExC_end);
18352 /* If 'range' is set, 'value' is the ending of a range--check its
18353 * validity. (If value isn't a single code point in the case of a
18354 * range, we should have figured that out above in the code that
18355 * catches false ranges). Later, we will handle each individual code
18356 * point in the range. If 'range' isn't set, this could be the
18357 * beginning of a range, so check for that by looking ahead to see if
18358 * the next real character to be processed is the range indicator--the
18363 /* For unicode ranges, we have to test that the Unicode as opposed
18364 * to the native values are not decreasing. (Above 255, there is
18365 * no difference between native and Unicode) */
18366 if (unicode_range && prevvalue < 255 && value < 255) {
18367 if (NATIVE_TO_LATIN1(prevvalue) > NATIVE_TO_LATIN1(value)) {
18368 goto backwards_range;
18373 if (prevvalue > value) /* b-a */ {
18378 w = RExC_parse - rangebegin;
18380 "Invalid [] range \"%" UTF8f "\"",
18381 UTF8fARG(UTF, w, rangebegin));
18382 NOT_REACHED; /* NOTREACHED */
18386 prevvalue = value; /* save the beginning of the potential range */
18387 if (! stop_at_1 /* Can't be a range if parsing just one thing */
18388 && *RExC_parse == '-')
18390 char* next_char_ptr = RExC_parse + 1;
18392 /* Get the next real char after the '-' */
18393 SKIP_BRACKETED_WHITE_SPACE(skip_white, next_char_ptr, RExC_end);
18395 /* If the '-' is at the end of the class (just before the ']',
18396 * it is a literal minus; otherwise it is a range */
18397 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
18398 RExC_parse = next_char_ptr;
18400 /* a bad range like \w-, [:word:]- ? */
18401 if (namedclass > OOB_NAMEDCLASS) {
18402 if (strict || ckWARN(WARN_REGEXP)) {
18403 const int w = RExC_parse >= rangebegin
18404 ? RExC_parse - rangebegin
18407 vFAIL4("False [] range \"%*.*s\"",
18412 "False [] range \"%*.*s\"",
18416 cp_list = add_cp_to_invlist(cp_list, '-');
18419 range = 1; /* yeah, it's a range! */
18420 continue; /* but do it the next time */
18425 if (namedclass > OOB_NAMEDCLASS) {
18429 /* Here, we have a single value this time through the loop, and
18430 * <prevvalue> is the beginning of the range, if any; or <value> if
18433 /* non-Latin1 code point implies unicode semantics. */
18435 if (value > MAX_LEGAL_CP && ( value != UV_MAX
18436 || prevvalue > MAX_LEGAL_CP))
18438 vFAIL(form_cp_too_large_msg(16, NULL, 0, value));
18440 REQUIRE_UNI_RULES(flagp, 0);
18441 if ( ! silence_non_portable
18442 && UNICODE_IS_PERL_EXTENDED(value)
18443 && TO_OUTPUT_WARNINGS(RExC_parse))
18445 ckWARN2_non_literal_string(RExC_parse,
18446 packWARN(WARN_PORTABLE),
18447 PL_extended_cp_format,
18452 /* Ready to process either the single value, or the completed range.
18453 * For single-valued non-inverted ranges, we consider the possibility
18454 * of multi-char folds. (We made a conscious decision to not do this
18455 * for the other cases because it can often lead to non-intuitive
18456 * results. For example, you have the peculiar case that:
18457 * "s s" =~ /^[^\xDF]+$/i => Y
18458 * "ss" =~ /^[^\xDF]+$/i => N
18460 * See [perl #89750] */
18461 if (FOLD && allow_mutiple_chars && value == prevvalue) {
18462 if ( value == LATIN_SMALL_LETTER_SHARP_S
18463 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
18466 /* Here <value> is indeed a multi-char fold. Get what it is */
18468 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
18471 UV folded = _to_uni_fold_flags(
18475 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
18476 ? FOLD_FLAGS_NOMIX_ASCII
18480 /* Here, <folded> should be the first character of the
18481 * multi-char fold of <value>, with <foldbuf> containing the
18482 * whole thing. But, if this fold is not allowed (because of
18483 * the flags), <fold> will be the same as <value>, and should
18484 * be processed like any other character, so skip the special
18486 if (folded != value) {
18488 /* Skip if we are recursed, currently parsing the class
18489 * again. Otherwise add this character to the list of
18490 * multi-char folds. */
18491 if (! RExC_in_multi_char_class) {
18492 STRLEN cp_count = utf8_length(foldbuf,
18493 foldbuf + foldlen);
18494 SV* multi_fold = sv_2mortal(newSVpvs(""));
18496 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%" UVXf "}", value);
18499 = add_multi_match(multi_char_matches,
18505 /* This element should not be processed further in this
18508 value = save_value;
18509 prevvalue = save_prevvalue;
18515 if (strict && ckWARN(WARN_REGEXP)) {
18518 /* If the range starts above 255, everything is portable and
18519 * likely to be so for any forseeable character set, so don't
18521 if (unicode_range && non_portable_endpoint && prevvalue < 256) {
18522 vWARN(RExC_parse, "Both or neither range ends should be Unicode");
18524 else if (prevvalue != value) {
18526 /* Under strict, ranges that stop and/or end in an ASCII
18527 * printable should have each end point be a portable value
18528 * for it (preferably like 'A', but we don't warn if it is
18529 * a (portable) Unicode name or code point), and the range
18530 * must be all digits or all letters of the same case.
18531 * Otherwise, the range is non-portable and unclear as to
18532 * what it contains */
18533 if ( (isPRINT_A(prevvalue) || isPRINT_A(value))
18534 && ( non_portable_endpoint
18535 || ! ( (isDIGIT_A(prevvalue) && isDIGIT_A(value))
18536 || (isLOWER_A(prevvalue) && isLOWER_A(value))
18537 || (isUPPER_A(prevvalue) && isUPPER_A(value))
18539 vWARN(RExC_parse, "Ranges of ASCII printables should"
18540 " be some subset of \"0-9\","
18541 " \"A-Z\", or \"a-z\"");
18543 else if (prevvalue >= FIRST_NON_ASCII_DECIMAL_DIGIT) {
18544 SSize_t index_start;
18545 SSize_t index_final;
18547 /* But the nature of Unicode and languages mean we
18548 * can't do the same checks for above-ASCII ranges,
18549 * except in the case of digit ones. These should
18550 * contain only digits from the same group of 10. The
18551 * ASCII case is handled just above. Hence here, the
18552 * range could be a range of digits. First some
18553 * unlikely special cases. Grandfather in that a range
18554 * ending in 19DA (NEW TAI LUE THAM DIGIT ONE) is bad
18555 * if its starting value is one of the 10 digits prior
18556 * to it. This is because it is an alternate way of
18557 * writing 19D1, and some people may expect it to be in
18558 * that group. But it is bad, because it won't give
18559 * the expected results. In Unicode 5.2 it was
18560 * considered to be in that group (of 11, hence), but
18561 * this was fixed in the next version */
18563 if (UNLIKELY(value == 0x19DA && prevvalue >= 0x19D0)) {
18564 goto warn_bad_digit_range;
18566 else if (UNLIKELY( prevvalue >= 0x1D7CE
18567 && value <= 0x1D7FF))
18569 /* This is the only other case currently in Unicode
18570 * where the algorithm below fails. The code
18571 * points just above are the end points of a single
18572 * range containing only decimal digits. It is 5
18573 * different series of 0-9. All other ranges of
18574 * digits currently in Unicode are just a single
18575 * series. (And mktables will notify us if a later
18576 * Unicode version breaks this.)
18578 * If the range being checked is at most 9 long,
18579 * and the digit values represented are in
18580 * numerical order, they are from the same series.
18582 if ( value - prevvalue > 9
18583 || ((( value - 0x1D7CE) % 10)
18584 <= (prevvalue - 0x1D7CE) % 10))
18586 goto warn_bad_digit_range;
18591 /* For all other ranges of digits in Unicode, the
18592 * algorithm is just to check if both end points
18593 * are in the same series, which is the same range.
18595 index_start = _invlist_search(
18596 PL_XPosix_ptrs[_CC_DIGIT],
18599 /* Warn if the range starts and ends with a digit,
18600 * and they are not in the same group of 10. */
18601 if ( index_start >= 0
18602 && ELEMENT_RANGE_MATCHES_INVLIST(index_start)
18604 _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
18605 value)) != index_start
18606 && index_final >= 0
18607 && ELEMENT_RANGE_MATCHES_INVLIST(index_final))
18609 warn_bad_digit_range:
18610 vWARN(RExC_parse, "Ranges of digits should be"
18611 " from the same group of"
18618 if ((! range || prevvalue == value) && non_portable_endpoint) {
18619 if (isPRINT_A(value)) {
18622 if (isBACKSLASHED_PUNCT(value)) {
18623 literal[d++] = '\\';
18625 literal[d++] = (char) value;
18626 literal[d++] = '\0';
18629 "\"%.*s\" is more clearly written simply as \"%s\"",
18630 (int) (RExC_parse - rangebegin),
18635 else if (isMNEMONIC_CNTRL(value)) {
18637 "\"%.*s\" is more clearly written simply as \"%s\"",
18638 (int) (RExC_parse - rangebegin),
18640 cntrl_to_mnemonic((U8) value)
18646 /* Deal with this element of the class */
18649 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
18652 /* On non-ASCII platforms, for ranges that span all of 0..255, and ones
18653 * that don't require special handling, we can just add the range like
18654 * we do for ASCII platforms */
18655 if ((UNLIKELY(prevvalue == 0) && value >= 255)
18656 || ! (prevvalue < 256
18658 || (! non_portable_endpoint
18659 && ((isLOWER_A(prevvalue) && isLOWER_A(value))
18660 || (isUPPER_A(prevvalue)
18661 && isUPPER_A(value)))))))
18663 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
18667 /* Here, requires special handling. This can be because it is a
18668 * range whose code points are considered to be Unicode, and so
18669 * must be individually translated into native, or because its a
18670 * subrange of 'A-Z' or 'a-z' which each aren't contiguous in
18671 * EBCDIC, but we have defined them to include only the "expected"
18672 * upper or lower case ASCII alphabetics. Subranges above 255 are
18673 * the same in native and Unicode, so can be added as a range */
18674 U8 start = NATIVE_TO_LATIN1(prevvalue);
18676 U8 end = (value < 256) ? NATIVE_TO_LATIN1(value) : 255;
18677 for (j = start; j <= end; j++) {
18678 cp_foldable_list = add_cp_to_invlist(cp_foldable_list, LATIN1_TO_NATIVE(j));
18681 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
18687 range = 0; /* this range (if it was one) is done now */
18688 } /* End of loop through all the text within the brackets */
18690 if ( posix_warnings && av_tindex_skip_len_mg(posix_warnings) >= 0) {
18691 output_posix_warnings(pRExC_state, posix_warnings);
18694 /* If anything in the class expands to more than one character, we have to
18695 * deal with them by building up a substitute parse string, and recursively
18696 * calling reg() on it, instead of proceeding */
18697 if (multi_char_matches) {
18698 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
18701 char *save_end = RExC_end;
18702 char *save_parse = RExC_parse;
18703 char *save_start = RExC_start;
18704 Size_t constructed_prefix_len = 0; /* This gives the length of the
18705 constructed portion of the
18706 substitute parse. */
18707 bool first_time = TRUE; /* First multi-char occurrence doesn't get
18712 /* Only one level of recursion allowed */
18713 assert(RExC_copy_start_in_constructed == RExC_precomp);
18715 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
18716 because too confusing */
18718 sv_catpvs(substitute_parse, "(?:");
18722 /* Look at the longest strings first */
18723 for (cp_count = av_tindex_skip_len_mg(multi_char_matches);
18728 if (av_exists(multi_char_matches, cp_count)) {
18729 AV** this_array_ptr;
18732 this_array_ptr = (AV**) av_fetch(multi_char_matches,
18734 while ((this_sequence = av_pop(*this_array_ptr)) !=
18737 if (! first_time) {
18738 sv_catpvs(substitute_parse, "|");
18740 first_time = FALSE;
18742 sv_catpv(substitute_parse, SvPVX(this_sequence));
18747 /* If the character class contains anything else besides these
18748 * multi-character strings, have to include it in recursive parsing */
18749 if (element_count) {
18750 bool has_l_bracket = orig_parse > RExC_start && *(orig_parse - 1) == '[';
18752 sv_catpvs(substitute_parse, "|");
18753 if (has_l_bracket) { /* Add an [ if the original had one */
18754 sv_catpvs(substitute_parse, "[");
18756 constructed_prefix_len = SvCUR(substitute_parse);
18757 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
18759 /* Put in a closing ']' to match any opening one, but not if going
18760 * off the end, as otherwise we are adding something that really
18762 if (has_l_bracket && RExC_parse < RExC_end) {
18763 sv_catpvs(substitute_parse, "]");
18767 sv_catpvs(substitute_parse, ")");
18770 /* This is a way to get the parse to skip forward a whole named
18771 * sequence instead of matching the 2nd character when it fails the
18773 sv_catpvs(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
18777 /* Set up the data structure so that any errors will be properly
18778 * reported. See the comments at the definition of
18779 * REPORT_LOCATION_ARGS for details */
18780 RExC_copy_start_in_input = (char *) orig_parse;
18781 RExC_start = RExC_parse = SvPV(substitute_parse, len);
18782 RExC_copy_start_in_constructed = RExC_start + constructed_prefix_len;
18783 RExC_end = RExC_parse + len;
18784 RExC_in_multi_char_class = 1;
18786 ret = reg(pRExC_state, 1, ®_flags, depth+1);
18788 *flagp |= reg_flags & (HASWIDTH|SIMPLE|POSTPONED|RESTART_PARSE|NEED_UTF8);
18790 /* And restore so can parse the rest of the pattern */
18791 RExC_parse = save_parse;
18792 RExC_start = RExC_copy_start_in_constructed = RExC_copy_start_in_input = save_start;
18793 RExC_end = save_end;
18794 RExC_in_multi_char_class = 0;
18795 SvREFCNT_dec_NN(multi_char_matches);
18796 SvREFCNT_dec(properties);
18797 SvREFCNT_dec(cp_list);
18798 SvREFCNT_dec(simple_posixes);
18799 SvREFCNT_dec(posixes);
18800 SvREFCNT_dec(nposixes);
18801 SvREFCNT_dec(cp_foldable_list);
18805 /* If folding, we calculate all characters that could fold to or from the
18806 * ones already on the list */
18807 if (cp_foldable_list) {
18809 UV start, end; /* End points of code point ranges */
18811 SV* fold_intersection = NULL;
18814 /* Our calculated list will be for Unicode rules. For locale
18815 * matching, we have to keep a separate list that is consulted at
18816 * runtime only when the locale indicates Unicode rules (and we
18817 * don't include potential matches in the ASCII/Latin1 range, as
18818 * any code point could fold to any other, based on the run-time
18819 * locale). For non-locale, we just use the general list */
18821 use_list = &only_utf8_locale_list;
18824 use_list = &cp_list;
18827 /* Only the characters in this class that participate in folds need
18828 * be checked. Get the intersection of this class and all the
18829 * possible characters that are foldable. This can quickly narrow
18830 * down a large class */
18831 _invlist_intersection(PL_in_some_fold, cp_foldable_list,
18832 &fold_intersection);
18834 /* Now look at the foldable characters in this class individually */
18835 invlist_iterinit(fold_intersection);
18836 while (invlist_iternext(fold_intersection, &start, &end)) {
18840 /* Look at every character in the range */
18841 for (j = start; j <= end; j++) {
18842 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
18845 Size_t folds_count;
18847 const U32 * remaining_folds;
18851 /* Under /l, we don't know what code points below 256
18852 * fold to, except we do know the MICRO SIGN folds to
18853 * an above-255 character if the locale is UTF-8, so we
18854 * add it to the special list (in *use_list) Otherwise
18855 * we know now what things can match, though some folds
18856 * are valid under /d only if the target is UTF-8.
18857 * Those go in a separate list */
18858 if ( IS_IN_SOME_FOLD_L1(j)
18859 && ! (LOC && j != MICRO_SIGN))
18862 /* ASCII is always matched; non-ASCII is matched
18863 * only under Unicode rules (which could happen
18864 * under /l if the locale is a UTF-8 one */
18865 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
18866 *use_list = add_cp_to_invlist(*use_list,
18867 PL_fold_latin1[j]);
18869 else if (j != PL_fold_latin1[j]) {
18870 upper_latin1_only_utf8_matches
18871 = add_cp_to_invlist(
18872 upper_latin1_only_utf8_matches,
18873 PL_fold_latin1[j]);
18877 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(j)
18878 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
18880 add_above_Latin1_folds(pRExC_state,
18887 /* Here is an above Latin1 character. We don't have the
18888 * rules hard-coded for it. First, get its fold. This is
18889 * the simple fold, as the multi-character folds have been
18890 * handled earlier and separated out */
18891 folded = _to_uni_fold_flags(j, foldbuf, &foldlen,
18892 (ASCII_FOLD_RESTRICTED)
18893 ? FOLD_FLAGS_NOMIX_ASCII
18896 /* Single character fold of above Latin1. Add everything
18897 * in its fold closure to the list that this node should
18899 folds_count = _inverse_folds(folded, &first_fold,
18901 for (k = 0; k <= folds_count; k++) {
18902 UV c = (k == 0) /* First time through use itself */
18904 : (k == 1) /* 2nd time use, the first fold */
18907 /* Then the remaining ones */
18908 : remaining_folds[k-2];
18910 /* /aa doesn't allow folds between ASCII and non- */
18911 if (( ASCII_FOLD_RESTRICTED
18912 && (isASCII(c) != isASCII(j))))
18917 /* Folds under /l which cross the 255/256 boundary are
18918 * added to a separate list. (These are valid only
18919 * when the locale is UTF-8.) */
18920 if (c < 256 && LOC) {
18921 *use_list = add_cp_to_invlist(*use_list, c);
18925 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
18927 cp_list = add_cp_to_invlist(cp_list, c);
18930 /* Similarly folds involving non-ascii Latin1
18931 * characters under /d are added to their list */
18932 upper_latin1_only_utf8_matches
18933 = add_cp_to_invlist(
18934 upper_latin1_only_utf8_matches,
18940 SvREFCNT_dec_NN(fold_intersection);
18943 /* Now that we have finished adding all the folds, there is no reason
18944 * to keep the foldable list separate */
18945 _invlist_union(cp_list, cp_foldable_list, &cp_list);
18946 SvREFCNT_dec_NN(cp_foldable_list);
18949 /* And combine the result (if any) with any inversion lists from posix
18950 * classes. The lists are kept separate up to now because we don't want to
18951 * fold the classes */
18952 if (simple_posixes) { /* These are the classes known to be unaffected by
18955 _invlist_union(cp_list, simple_posixes, &cp_list);
18956 SvREFCNT_dec_NN(simple_posixes);
18959 cp_list = simple_posixes;
18962 if (posixes || nposixes) {
18963 if (! DEPENDS_SEMANTICS) {
18965 /* For everything but /d, we can just add the current 'posixes' and
18966 * 'nposixes' to the main list */
18969 _invlist_union(cp_list, posixes, &cp_list);
18970 SvREFCNT_dec_NN(posixes);
18978 _invlist_union(cp_list, nposixes, &cp_list);
18979 SvREFCNT_dec_NN(nposixes);
18982 cp_list = nposixes;
18987 /* Under /d, things like \w match upper Latin1 characters only if
18988 * the target string is in UTF-8. But things like \W match all the
18989 * upper Latin1 characters if the target string is not in UTF-8.
18991 * Handle the case with something like \W separately */
18993 SV* only_non_utf8_list = invlist_clone(PL_UpperLatin1, NULL);
18995 /* A complemented posix class matches all upper Latin1
18996 * characters if not in UTF-8. And it matches just certain
18997 * ones when in UTF-8. That means those certain ones are
18998 * matched regardless, so can just be added to the
18999 * unconditional list */
19001 _invlist_union(cp_list, nposixes, &cp_list);
19002 SvREFCNT_dec_NN(nposixes);
19006 cp_list = nposixes;
19009 /* Likewise for 'posixes' */
19010 _invlist_union(posixes, cp_list, &cp_list);
19011 SvREFCNT_dec(posixes);
19013 /* Likewise for anything else in the range that matched only
19015 if (upper_latin1_only_utf8_matches) {
19016 _invlist_union(cp_list,
19017 upper_latin1_only_utf8_matches,
19019 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
19020 upper_latin1_only_utf8_matches = NULL;
19023 /* If we don't match all the upper Latin1 characters regardless
19024 * of UTF-8ness, we have to set a flag to match the rest when
19026 _invlist_subtract(only_non_utf8_list, cp_list,
19027 &only_non_utf8_list);
19028 if (_invlist_len(only_non_utf8_list) != 0) {
19029 anyof_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
19031 SvREFCNT_dec_NN(only_non_utf8_list);
19034 /* Here there were no complemented posix classes. That means
19035 * the upper Latin1 characters in 'posixes' match only when the
19036 * target string is in UTF-8. So we have to add them to the
19037 * list of those types of code points, while adding the
19038 * remainder to the unconditional list.
19040 * First calculate what they are */
19041 SV* nonascii_but_latin1_properties = NULL;
19042 _invlist_intersection(posixes, PL_UpperLatin1,
19043 &nonascii_but_latin1_properties);
19045 /* And add them to the final list of such characters. */
19046 _invlist_union(upper_latin1_only_utf8_matches,
19047 nonascii_but_latin1_properties,
19048 &upper_latin1_only_utf8_matches);
19050 /* Remove them from what now becomes the unconditional list */
19051 _invlist_subtract(posixes, nonascii_but_latin1_properties,
19054 /* And add those unconditional ones to the final list */
19056 _invlist_union(cp_list, posixes, &cp_list);
19057 SvREFCNT_dec_NN(posixes);
19064 SvREFCNT_dec(nonascii_but_latin1_properties);
19066 /* Get rid of any characters from the conditional list that we
19067 * now know are matched unconditionally, which may make that
19069 _invlist_subtract(upper_latin1_only_utf8_matches,
19071 &upper_latin1_only_utf8_matches);
19072 if (_invlist_len(upper_latin1_only_utf8_matches) == 0) {
19073 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
19074 upper_latin1_only_utf8_matches = NULL;
19080 /* And combine the result (if any) with any inversion list from properties.
19081 * The lists are kept separate up to now so that we can distinguish the two
19082 * in regards to matching above-Unicode. A run-time warning is generated
19083 * if a Unicode property is matched against a non-Unicode code point. But,
19084 * we allow user-defined properties to match anything, without any warning,
19085 * and we also suppress the warning if there is a portion of the character
19086 * class that isn't a Unicode property, and which matches above Unicode, \W
19087 * or [\x{110000}] for example.
19088 * (Note that in this case, unlike the Posix one above, there is no
19089 * <upper_latin1_only_utf8_matches>, because having a Unicode property
19090 * forces Unicode semantics */
19094 /* If it matters to the final outcome, see if a non-property
19095 * component of the class matches above Unicode. If so, the
19096 * warning gets suppressed. This is true even if just a single
19097 * such code point is specified, as, though not strictly correct if
19098 * another such code point is matched against, the fact that they
19099 * are using above-Unicode code points indicates they should know
19100 * the issues involved */
19102 warn_super = ! (invert
19103 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
19106 _invlist_union(properties, cp_list, &cp_list);
19107 SvREFCNT_dec_NN(properties);
19110 cp_list = properties;
19115 |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
19117 /* Because an ANYOF node is the only one that warns, this node
19118 * can't be optimized into something else */
19119 optimizable = FALSE;
19123 /* Here, we have calculated what code points should be in the character
19126 * Now we can see about various optimizations. Fold calculation (which we
19127 * did above) needs to take place before inversion. Otherwise /[^k]/i
19128 * would invert to include K, which under /i would match k, which it
19129 * shouldn't. Therefore we can't invert folded locale now, as it won't be
19130 * folded until runtime */
19132 /* If we didn't do folding, it's because some information isn't available
19133 * until runtime; set the run-time fold flag for these We know to set the
19134 * flag if we have a non-NULL list for UTF-8 locales, or the class matches
19135 * at least one 0-255 range code point */
19138 /* Some things on the list might be unconditionally included because of
19139 * other components. Remove them, and clean up the list if it goes to
19141 if (only_utf8_locale_list && cp_list) {
19142 _invlist_subtract(only_utf8_locale_list, cp_list,
19143 &only_utf8_locale_list);
19145 if (_invlist_len(only_utf8_locale_list) == 0) {
19146 SvREFCNT_dec_NN(only_utf8_locale_list);
19147 only_utf8_locale_list = NULL;
19150 if ( only_utf8_locale_list
19151 || (cp_list && ( _invlist_contains_cp(cp_list, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE)
19152 || _invlist_contains_cp(cp_list, LATIN_SMALL_LETTER_DOTLESS_I))))
19154 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
19157 | ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
19159 else if (cp_list && invlist_lowest(cp_list) < 256) {
19160 /* If nothing is below 256, has no locale dependency; otherwise it
19162 anyof_flags |= ANYOFL_FOLD;
19163 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
19166 else if ( DEPENDS_SEMANTICS
19167 && ( upper_latin1_only_utf8_matches
19168 || (anyof_flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)))
19170 RExC_seen_d_op = TRUE;
19171 has_runtime_dependency |= HAS_D_RUNTIME_DEPENDENCY;
19174 /* Optimize inverted patterns (e.g. [^a-z]) when everything is known at
19178 && ! has_runtime_dependency)
19180 _invlist_invert(cp_list);
19182 /* Clear the invert flag since have just done it here */
19186 /* All possible optimizations below still have these characteristics.
19187 * (Multi-char folds aren't SIMPLE, but they don't get this far in this
19189 *flagp |= HASWIDTH|SIMPLE;
19192 *ret_invlist = cp_list;
19194 return (cp_list) ? RExC_emit : 0;
19197 if (anyof_flags & ANYOF_LOCALE_FLAGS) {
19198 RExC_contains_locale = 1;
19203 /* Some character classes are equivalent to other nodes. Such nodes
19204 * take up less room, and some nodes require fewer operations to
19205 * execute, than ANYOF nodes. EXACTish nodes may be joinable with
19206 * adjacent nodes to improve efficiency. */
19207 op = optimize_regclass(pRExC_state, cp_list,
19208 only_utf8_locale_list,
19209 upper_latin1_only_utf8_matches,
19210 has_runtime_dependency,
19212 &anyof_flags, &invert, &ret, flagp);
19213 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
19215 /* If optimized to something else, finish up and return */
19217 Set_Node_Offset_Length(REGNODE_p(ret), orig_parse - RExC_start,
19218 RExC_parse - orig_parse);;
19219 SvREFCNT_dec(cp_list);;
19220 SvREFCNT_dec(only_utf8_locale_list);
19221 SvREFCNT_dec(upper_latin1_only_utf8_matches);
19226 /* Here didn't optimize, or optimized to a specialized ANYOF node. If the
19227 * former, 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;
19402 /* Compute which bit is set, which is the same thing as, e.g.,
19403 * ANYOF_CNTRL. From
19404 * https://graphics.stanford.edu/~seander/bithacks.html#IntegerLogDeBruijn
19406 static const int MultiplyDeBruijnBitPosition2[32] = {
19407 0, 1, 28, 2, 29, 14, 24, 3, 30, 22, 20, 15, 25, 17, 4, 8,
19408 31, 27, 13, 23, 21, 19, 16, 7, 26, 12, 18, 6, 11, 5, 10, 9
19411 namedclass = MultiplyDeBruijnBitPosition2[(posixl
19412 * 0x077CB531U) >> 27];
19413 classnum = namedclass_to_classnum(namedclass);
19415 /* The named classes are such that the inverted number is one
19416 * larger than the non-inverted one */
19417 already_inverted = namedclass - classnum_to_namedclass(classnum);
19419 /* Create an inversion list of the official property, inverted if
19420 * the constructed node list is inverted, and restricted to only
19421 * the above latin1 code points, which are the only ones known at
19423 _invlist_intersection_maybe_complement_2nd(
19425 PL_XPosix_ptrs[classnum],
19427 &class_above_latin1);
19428 are_equivalent = _invlistEQ(class_above_latin1, cp_list, FALSE);
19429 SvREFCNT_dec_NN(class_above_latin1);
19431 if (are_equivalent) {
19433 /* Resolve the run-time inversion flag with this possibly
19434 * inverted class */
19435 *invert = *invert ^ already_inverted;
19437 op = POSIXL + *invert * (NPOSIXL - POSIXL);
19438 *ret = reg_node(pRExC_state, op);
19439 FLAGS(REGNODE_p(*ret)) = classnum;
19445 /* khw can't think of any other possible transformation involving these. */
19446 if (has_runtime_dependency & HAS_USER_DEFINED_PROPERTY) {
19450 if (! has_runtime_dependency) {
19452 /* If the list is empty, nothing matches. This happens, for example,
19453 * when a Unicode property that doesn't match anything is the only
19454 * element in the character class (perluniprops.pod notes such
19456 if (partial_cp_count == 0) {
19459 *ret = reg_node(pRExC_state, op);
19463 *ret = reganode(pRExC_state, op, 0);
19469 /* If matches everything but \n */
19470 if ( start[0] == 0 && end[0] == '\n' - 1
19471 && start[1] == '\n' + 1 && end[1] == UV_MAX)
19473 assert (! *invert);
19475 *ret = reg_node(pRExC_state, op);
19481 /* Next see if can optimize classes that contain just a few code points
19482 * into an EXACTish node. The reason to do this is to let the optimizer
19483 * join this node with adjacent EXACTish ones, and ANYOF nodes require
19484 * runtime conversion to code point from UTF-8.
19486 * An EXACTFish node can be generated even if not under /i, and vice versa.
19487 * But care must be taken. An EXACTFish node has to be such that it only
19488 * matches precisely the code points in the class, but we want to generate
19489 * the least restrictive one that does that, to increase the odds of being
19490 * able to join with an adjacent node. For example, if the class contains
19491 * [kK], we have to make it an EXACTFAA node to prevent the KELVIN SIGN
19492 * from matching. Whether we are under /i or not is irrelevant in this
19493 * case. Less obvious is the pattern qr/[\x{02BC}]n/i. U+02BC is MODIFIER
19494 * LETTER APOSTROPHE. That is supposed to match the single character U+0149
19495 * LATIN SMALL LETTER N PRECEDED BY APOSTROPHE. And so even though there
19496 * is no simple fold that includes \X{02BC}, there is a multi-char fold
19497 * that does, and so the node generated for it must be an EXACTFish one.
19498 * On the other hand qr/:/i should generate a plain EXACT node since the
19499 * colon participates in no fold whatsoever, and having it EXACT tells the
19500 * optimizer the target string cannot match unless it has a colon in it.
19505 /* Only try if there are no more code points in the class than in
19506 * the max possible fold */
19507 && inRANGE(partial_cp_count, 1, MAX_FOLD_FROMS + 1))
19509 /* We can always make a single code point class into an EXACTish node.
19511 if (partial_cp_count == 1 && ! upper_latin1_only_utf8_matches) {
19514 /* Here is /l: Use EXACTL, except if there is a fold not known
19515 * until runtime so shows as only a single code point here.
19516 * For code points above 255, we know which can cause problems
19517 * by having a potential fold to the Latin1 range. */
19519 || ( start[0] > 255
19520 && ! is_PROBLEMATIC_LOCALE_FOLD_cp(start[0])))
19528 else if (! FOLD) { /* Not /l and not /i */
19529 op = (start[0] < 256) ? EXACT : EXACT_REQ8;
19531 else if (start[0] < 256) { /* /i, not /l, and the code point is
19534 /* Under /i, it gets a little tricky. A code point that
19535 * doesn't participate in a fold should be an EXACT node. We
19536 * know this one isn't the result of a simple fold, or there'd
19537 * be more than one code point in the list, but it could be
19538 * part of a multi-character fold. In that case we better not
19539 * create an EXACT node, as we would wrongly be telling the
19540 * optimizer that this code point must be in the target string,
19541 * and that is wrong. This is because if the sequence around
19542 * this code point forms a multi-char fold, what needs to be in
19543 * the string could be the code point that folds to the
19546 * This handles the case of below-255 code points, as we have
19547 * an easy look up for those. The next clause handles the
19549 op = IS_IN_SOME_FOLD_L1(start[0])
19553 else { /* /i, larger code point. Since we are under /i, and have
19554 just this code point, we know that it can't fold to
19555 something else, so PL_InMultiCharFold applies to it */
19556 op = (_invlist_contains_cp(PL_InMultiCharFold, start[0]))
19563 else if ( ! (has_runtime_dependency & ~HAS_D_RUNTIME_DEPENDENCY)
19564 && _invlist_contains_cp(PL_in_some_fold, start[0]))
19566 /* Here, the only runtime dependency, if any, is from /d, and the
19567 * class matches more than one code point, and the lowest code
19568 * point participates in some fold. It might be that the other
19569 * code points are /i equivalent to this one, and hence they would
19570 * representable by an EXACTFish node. Above, we eliminated
19571 * classes that contain too many code points to be EXACTFish, with
19572 * the test for MAX_FOLD_FROMS
19574 * First, special case the ASCII fold pairs, like 'B' and 'b'. We
19575 * do this because we have EXACTFAA at our disposal for the ASCII
19577 if (partial_cp_count == 2 && isASCII(start[0])) {
19579 /* The only ASCII characters that participate in folds are
19581 assert(isALPHA(start[0]));
19582 if ( end[0] == start[0] /* First range is a single
19583 character, so 2nd exists */
19584 && isALPHA_FOLD_EQ(start[0], start[1]))
19587 /* Here, is part of an ASCII fold pair */
19589 if ( ASCII_FOLD_RESTRICTED
19590 || HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(start[0]))
19592 /* If the second clause just above was true, it means
19593 * we can't be under /i, or else the list would have
19594 * included more than this fold pair. Therefore we
19595 * have to exclude the possibility of whatever else it
19596 * is that folds to these, by using EXACTFAA */
19599 else if (HAS_NONLATIN1_FOLD_CLOSURE(start[0])) {
19601 /* Here, there's no simple fold that start[0] is part
19602 * of, but there is a multi-character one. If we are
19603 * not under /i, we want to exclude that possibility;
19604 * if under /i, we want to include it */
19605 op = (FOLD) ? EXACTFU : EXACTFAA;
19609 /* Here, the only possible fold start[0] particpates in
19610 * is with start[1]. /i or not isn't relevant */
19614 value = toFOLD(start[0]);
19617 else if ( ! upper_latin1_only_utf8_matches
19618 || ( _invlist_len(upper_latin1_only_utf8_matches) == 2
19620 invlist_highest(upper_latin1_only_utf8_matches)]
19623 /* Here, the smallest character is non-ascii or there are more
19624 * than 2 code points matched by this node. Also, we either
19625 * don't have /d UTF-8 dependent matches, or if we do, they
19626 * look like they could be a single character that is the fold
19627 * of the lowest one is in the always-match list. This test
19628 * quickly excludes most of the false positives when there are
19629 * /d UTF-8 depdendent matches. These are like LATIN CAPITAL
19630 * LETTER A WITH GRAVE matching LATIN SMALL LETTER A WITH GRAVE
19631 * iff the target string is UTF-8. (We don't have to worry
19632 * above about exceeding the array bounds of PL_fold_latin1[]
19633 * because any code point in 'upper_latin1_only_utf8_matches'
19636 * EXACTFAA would apply only to pairs (hence exactly 2 code
19637 * points) in the ASCII range, so we can't use it here to
19638 * artificially restrict the fold domain, so we check if the
19639 * class does or does not match some EXACTFish node. Further,
19640 * if we aren't under /i, and and the folded-to character is
19641 * part of a multi-character fold, we can't do this
19642 * optimization, as the sequence around it could be that
19643 * multi-character fold, and we don't here know the context, so
19644 * we have to assume it is that multi-char fold, to prevent
19647 * To do the general case, we first find the fold of the lowest
19648 * code point (which may be higher than the lowest one), then
19649 * find everything that folds to it. (The data structure we
19650 * have only maps from the folded code points, so we have to do
19651 * the earlier step.) */
19654 U8 foldbuf[UTF8_MAXBYTES_CASE];
19655 UV folded = _to_uni_fold_flags(start[0], foldbuf, &foldlen, 0);
19657 const U32 * remaining_folds;
19658 Size_t folds_to_this_cp_count = _inverse_folds(
19662 Size_t folds_count = folds_to_this_cp_count + 1;
19663 SV * fold_list = _new_invlist(folds_count);
19666 /* If there are UTF-8 dependent matches, create a temporary
19667 * list of what this node matches, including them. */
19668 SV * all_cp_list = NULL;
19669 SV ** use_this_list = &cp_list;
19671 if (upper_latin1_only_utf8_matches) {
19672 all_cp_list = _new_invlist(0);
19673 use_this_list = &all_cp_list;
19674 _invlist_union(cp_list,
19675 upper_latin1_only_utf8_matches,
19679 /* Having gotten everything that participates in the fold
19680 * containing the lowest code point, we turn that into an
19681 * inversion list, making sure everything is included. */
19682 fold_list = add_cp_to_invlist(fold_list, start[0]);
19683 fold_list = add_cp_to_invlist(fold_list, folded);
19684 if (folds_to_this_cp_count > 0) {
19685 fold_list = add_cp_to_invlist(fold_list, first_fold);
19686 for (i = 0; i + 1 < folds_to_this_cp_count; i++) {
19687 fold_list = add_cp_to_invlist(fold_list,
19688 remaining_folds[i]);
19692 /* If the fold list is identical to what's in this ANYOF node,
19693 * the node can be represented by an EXACTFish one instead */
19694 if (_invlistEQ(*use_this_list, fold_list,
19695 0 /* Don't complement */ )
19698 /* But, we have to be careful, as mentioned above. Just
19699 * the right sequence of characters could match this if it
19700 * is part of a multi-character fold. That IS what we want
19701 * if we are under /i. But it ISN'T what we want if not
19702 * under /i, as it could match when it shouldn't. So, when
19703 * we aren't under /i and this character participates in a
19704 * multi-char fold, we don't optimize into an EXACTFish
19705 * node. So, for each case below we have to check if we
19706 * are folding and if not, if it is not part of a
19707 * multi-char fold. */
19708 if (start[0] > 255) { /* Highish code point */
19709 if (FOLD || ! _invlist_contains_cp(
19710 PL_InMultiCharFold, folded))
19714 : (ASCII_FOLD_RESTRICTED)
19719 } /* Below, the lowest code point < 256 */
19722 && DEPENDS_SEMANTICS)
19723 { /* An EXACTF node containing a single character 's',
19724 can be an EXACTFU if it doesn't get joined with an
19726 op = EXACTFU_S_EDGE;
19730 || ! HAS_NONLATIN1_FOLD_CLOSURE(start[0]))
19732 if (upper_latin1_only_utf8_matches) {
19735 /* We can't use the fold, as that only matches
19739 else if ( UNLIKELY(start[0] == MICRO_SIGN)
19741 { /* EXACTFUP is a special node for this character */
19742 op = (ASCII_FOLD_RESTRICTED)
19745 value = MICRO_SIGN;
19747 else if ( ASCII_FOLD_RESTRICTED
19748 && ! isASCII(start[0]))
19749 { /* For ASCII under /iaa, we can use EXACTFU below
19761 SvREFCNT_dec_NN(fold_list);
19762 SvREFCNT_dec(all_cp_list);
19769 /* Here, we have calculated what EXACTish node to use. Have to
19770 * convert to UTF-8 if not already there */
19773 SvREFCNT_dec(cp_list);;
19774 REQUIRE_UTF8(flagp);
19777 /* This is a kludge to the special casing issues with this
19778 * ligature under /aa. FB05 should fold to FB06, but the call
19779 * above to _to_uni_fold_flags() didn't find this, as it didn't
19780 * use the /aa restriction in order to not miss other folds
19781 * that would be affected. This is the only instance likely to
19782 * ever be a problem in all of Unicode. So special case it. */
19783 if ( value == LATIN_SMALL_LIGATURE_LONG_S_T
19784 && ASCII_FOLD_RESTRICTED)
19786 value = LATIN_SMALL_LIGATURE_ST;
19790 len = (UTF) ? UVCHR_SKIP(value) : 1;
19792 *ret = regnode_guts(pRExC_state, op, len, "exact");
19793 FILL_NODE(*ret, op);
19794 RExC_emit += 1 + STR_SZ(len);
19795 setSTR_LEN(REGNODE_p(*ret), len);
19797 *STRINGs(REGNODE_p(*ret)) = (U8) value;
19800 uvchr_to_utf8((U8 *) STRINGs(REGNODE_p(*ret)), value);
19806 if (! has_runtime_dependency) {
19808 /* See if this can be turned into an ANYOFM node. Think about the bit
19809 * patterns in two different bytes. In some positions, the bits in
19810 * each will be 1; and in other positions both will be 0; and in some
19811 * positions the bit will be 1 in one byte, and 0 in the other. Let
19812 * 'n' be the number of positions where the bits differ. We create a
19813 * mask which has exactly 'n' 0 bits, each in a position where the two
19814 * bytes differ. Now take the set of all bytes that when ANDed with
19815 * the mask yield the same result. That set has 2**n elements, and is
19816 * representable by just two 8 bit numbers: the result and the mask.
19817 * Importantly, matching the set can be vectorized by creating a word
19818 * full of the result bytes, and a word full of the mask bytes,
19819 * yielding a significant speed up. Here, see if this node matches
19820 * such a set. As a concrete example consider [01], and the byte
19821 * representing '0' which is 0x30 on ASCII machines. It has the bits
19822 * 0011 0000. Take the mask 1111 1110. If we AND 0x31 and 0x30 with
19823 * that mask we get 0x30. Any other bytes ANDed yield something else.
19824 * So [01], which is a common usage, is optimizable into ANYOFM, and
19825 * can benefit from the speed up. We can only do this on UTF-8
19826 * invariant bytes, because they have the same bit patterns under UTF-8
19828 PERL_UINT_FAST8_T inverted = 0;
19830 const PERL_UINT_FAST8_T max_permissible = 0xFF;
19832 const PERL_UINT_FAST8_T max_permissible = 0x7F;
19834 /* If doesn't fit the criteria for ANYOFM, invert and try again. If
19835 * that works we will instead later generate an NANYOFM, and invert
19836 * back when through */
19837 if (invlist_highest(cp_list) > max_permissible) {
19838 _invlist_invert(cp_list);
19842 if (invlist_highest(cp_list) <= max_permissible) {
19843 UV this_start, this_end;
19844 UV lowest_cp = UV_MAX; /* init'ed to suppress compiler warn */
19845 U8 bits_differing = 0;
19846 Size_t full_cp_count = 0;
19847 bool first_time = TRUE;
19849 /* Go through the bytes and find the bit positions that differ */
19850 invlist_iterinit(cp_list);
19851 while (invlist_iternext(cp_list, &this_start, &this_end)) {
19852 unsigned int i = this_start;
19855 if (! UVCHR_IS_INVARIANT(i)) {
19859 first_time = FALSE;
19860 lowest_cp = this_start;
19862 /* We have set up the code point to compare with. Don't
19863 * compare it with itself */
19867 /* Find the bit positions that differ from the lowest code
19868 * point in the node. Keep track of all such positions by
19870 for (; i <= this_end; i++) {
19871 if (! UVCHR_IS_INVARIANT(i)) {
19875 bits_differing |= i ^ lowest_cp;
19878 full_cp_count += this_end - this_start + 1;
19881 /* At the end of the loop, we count how many bits differ from the
19882 * bits in lowest code point, call the count 'd'. If the set we
19883 * found contains 2**d elements, it is the closure of all code
19884 * points that differ only in those bit positions. To convince
19885 * yourself of that, first note that the number in the closure must
19886 * be a power of 2, which we test for. The only way we could have
19887 * that count and it be some differing set, is if we got some code
19888 * points that don't differ from the lowest code point in any
19889 * position, but do differ from each other in some other position.
19890 * That means one code point has a 1 in that position, and another
19891 * has a 0. But that would mean that one of them differs from the
19892 * lowest code point in that position, which possibility we've
19893 * already excluded. */
19894 if ( (inverted || full_cp_count > 1)
19895 && full_cp_count == 1U << PL_bitcount[bits_differing])
19899 op = ANYOFM + inverted;;
19901 /* We need to make the bits that differ be 0's */
19902 ANYOFM_mask = ~ bits_differing; /* This goes into FLAGS */
19904 /* The argument is the lowest code point */
19905 *ret = reganode(pRExC_state, op, lowest_cp);
19906 FLAGS(REGNODE_p(*ret)) = ANYOFM_mask;
19910 invlist_iterfinish(cp_list);
19914 _invlist_invert(cp_list);
19921 /* XXX We could create an ANYOFR_LOW node here if we saved above if all
19922 * were invariants, it wasn't inverted, and there is a single range.
19923 * This would be faster than some of the posix nodes we create below
19924 * like /\d/a, but would be twice the size. Without having actually
19925 * measured the gain, khw doesn't think the tradeoff is really worth it
19929 if (! (*anyof_flags & ANYOF_LOCALE_FLAGS)) {
19930 PERL_UINT_FAST8_T type;
19931 SV * intersection = NULL;
19932 SV* d_invlist = NULL;
19934 /* See if this matches any of the POSIX classes. The POSIXA and POSIXD
19935 * ones are about the same speed as ANYOF ops, but take less room; the
19936 * ones that have above-Latin1 code point matches are somewhat faster
19939 for (type = POSIXA; type >= POSIXD; type--) {
19942 if (type == POSIXL) { /* But not /l posix classes */
19946 for (posix_class = 0;
19947 posix_class <= _HIGHEST_REGCOMP_DOT_H_SYNC;
19950 SV** our_code_points = &cp_list;
19951 SV** official_code_points;
19954 if (type == POSIXA) {
19955 official_code_points = &PL_Posix_ptrs[posix_class];
19958 official_code_points = &PL_XPosix_ptrs[posix_class];
19961 /* Skip non-existent classes of this type. e.g. \v only has an
19962 * entry in PL_XPosix_ptrs */
19963 if (! *official_code_points) {
19967 /* Try both the regular class, and its inversion */
19968 for (try_inverted = 0; try_inverted < 2; try_inverted++) {
19969 bool this_inverted = *invert ^ try_inverted;
19971 if (type != POSIXD) {
19973 /* This class that isn't /d can't match if we have /d
19975 if (has_runtime_dependency
19976 & HAS_D_RUNTIME_DEPENDENCY)
19981 else /* is /d */ if (! this_inverted) {
19983 /* /d classes don't match anything non-ASCII below 256
19984 * unconditionally (which cp_list contains) */
19985 _invlist_intersection(cp_list, PL_UpperLatin1,
19987 if (_invlist_len(intersection) != 0) {
19991 SvREFCNT_dec(d_invlist);
19992 d_invlist = invlist_clone(cp_list, NULL);
19994 /* But under UTF-8 it turns into using /u rules. Add
19995 * the things it matches under these conditions so that
19996 * we check below that these are identical to what the
19997 * tested class should match */
19998 if (upper_latin1_only_utf8_matches) {
20001 upper_latin1_only_utf8_matches,
20004 our_code_points = &d_invlist;
20006 else { /* POSIXD, inverted. If this doesn't have this
20007 flag set, it isn't /d. */
20008 if (! (*anyof_flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
20012 our_code_points = &cp_list;
20015 /* Here, have weeded out some things. We want to see if
20016 * the list of characters this node contains
20017 * ('*our_code_points') precisely matches those of the
20018 * class we are currently checking against
20019 * ('*official_code_points'). */
20020 if (_invlistEQ(*our_code_points,
20021 *official_code_points,
20024 /* Here, they precisely match. Optimize this ANYOF
20025 * node into its equivalent POSIX one of the correct
20026 * type, possibly inverted */
20027 op = (try_inverted)
20028 ? type + NPOSIXA - POSIXA
20030 *ret = reg_node(pRExC_state, op);
20031 FLAGS(REGNODE_p(*ret)) = posix_class;
20032 SvREFCNT_dec(d_invlist);
20033 SvREFCNT_dec(intersection);
20039 SvREFCNT_dec(d_invlist);
20040 SvREFCNT_dec(intersection);
20043 /* If it is a single contiguous range, ANYOFR is an efficient regnode, both
20044 * in size and speed. Currently, a 20 bit range base (smallest code point
20045 * in the range), and a 12 bit maximum delta are packed into a 32 bit word.
20046 * This allows for using it on all of the Unicode code points except for
20047 * the highest plane, which is only for private use code points. khw
20048 * doubts that a bigger delta is likely in real world applications */
20050 && ! has_runtime_dependency
20051 && *anyof_flags == 0
20052 && start[0] < (1 << ANYOFR_BASE_BITS)
20053 && end[0] - start[0]
20054 < ((1U << (sizeof(((struct regnode_1 *)NULL)->arg1)
20055 * CHARBITS - ANYOFR_BASE_BITS))))
20058 U8 low_utf8[UTF8_MAXBYTES+1];
20059 U8 high_utf8[UTF8_MAXBYTES+1];
20062 *ret = reganode(pRExC_state, op,
20063 (start[0] | (end[0] - start[0]) << ANYOFR_BASE_BITS));
20065 /* Place the lowest UTF-8 start byte in the flags field, so as to allow
20066 * efficient ruling out at run time of many possible inputs. */
20067 (void) uvchr_to_utf8(low_utf8, start[0]);
20068 (void) uvchr_to_utf8(high_utf8, end[0]);
20070 /* If all code points share the same first byte, this can be an
20071 * ANYOFRb. Otherwise store the lowest UTF-8 start byte which can
20072 * quickly rule out many inputs at run-time without having to compute
20073 * the code point from UTF-8. For EBCDIC, we use I8, as not doing that
20074 * transformation would not rule out nearly so many things */
20075 if (low_utf8[0] == high_utf8[0]) {
20077 OP(REGNODE_p(*ret)) = op;
20078 ANYOF_FLAGS(REGNODE_p(*ret)) = low_utf8[0];
20081 ANYOF_FLAGS(REGNODE_p(*ret)) = NATIVE_UTF8_TO_I8(low_utf8[0]);
20087 /* If didn't find an optimization and there is no need for a bitmap,
20088 * optimize to indicate that */
20089 if ( start[0] >= NUM_ANYOF_CODE_POINTS
20091 && ! upper_latin1_only_utf8_matches
20092 && *anyof_flags == 0)
20094 U8 low_utf8[UTF8_MAXBYTES+1];
20095 UV highest_cp = invlist_highest(cp_list);
20097 /* Currently the maximum allowed code point by the system is IV_MAX.
20098 * Higher ones are reserved for future internal use. This particular
20099 * regnode can be used for higher ones, but we can't calculate the code
20100 * point of those. IV_MAX suffices though, as it will be a large first
20102 Size_t low_len = uvchr_to_utf8(low_utf8, MIN(start[0], IV_MAX))
20105 /* We store the lowest possible first byte of the UTF-8 representation,
20106 * using the flags field. This allows for quick ruling out of some
20107 * inputs without having to convert from UTF-8 to code point. For
20108 * EBCDIC, we use I8, as not doing that transformation would not rule
20109 * out nearly so many things */
20110 *anyof_flags = NATIVE_UTF8_TO_I8(low_utf8[0]);
20114 /* If the first UTF-8 start byte for the highest code point in the
20115 * range is suitably small, we may be able to get an upper bound as
20117 if (highest_cp <= IV_MAX) {
20118 U8 high_utf8[UTF8_MAXBYTES+1];
20119 Size_t high_len = uvchr_to_utf8(high_utf8, highest_cp) - high_utf8;
20121 /* If the lowest and highest are the same, we can get an exact
20122 * first byte instead of a just minimum or even a sequence of exact
20123 * leading bytes. We signal these with different regnodes */
20124 if (low_utf8[0] == high_utf8[0]) {
20125 Size_t len = find_first_differing_byte_pos(low_utf8,
20127 MIN(low_len, high_len));
20131 /* No need to convert to I8 for EBCDIC as this is an exact
20133 *anyof_flags = low_utf8[0];
20138 *ret = regnode_guts(pRExC_state, op,
20139 regarglen[op] + STR_SZ(len),
20141 FILL_NODE(*ret, op);
20142 ((struct regnode_anyofhs *) REGNODE_p(*ret))->str_len
20144 Copy(low_utf8, /* Add the common bytes */
20145 ((struct regnode_anyofhs *) REGNODE_p(*ret))->string,
20147 RExC_emit += NODE_SZ_STR(REGNODE_p(*ret));
20148 set_ANYOF_arg(pRExC_state, REGNODE_p(*ret), cp_list,
20149 NULL, only_utf8_locale_list);
20153 else if (NATIVE_UTF8_TO_I8(high_utf8[0]) <= MAX_ANYOF_HRx_BYTE) {
20155 /* Here, the high byte is not the same as the low, but is small
20156 * enough that its reasonable to have a loose upper bound,
20157 * which is packed in with the strict lower bound. See
20158 * comments at the definition of MAX_ANYOF_HRx_BYTE. On EBCDIC
20159 * platforms, I8 is used. On ASCII platforms I8 is the same
20160 * thing as UTF-8 */
20163 U8 max_range_diff = MAX_ANYOF_HRx_BYTE - *anyof_flags;
20164 U8 range_diff = NATIVE_UTF8_TO_I8(high_utf8[0])
20167 if (range_diff <= max_range_diff / 8) {
20170 else if (range_diff <= max_range_diff / 4) {
20173 else if (range_diff <= max_range_diff / 2) {
20176 *anyof_flags = (*anyof_flags - 0xC0) << 2 | bits;
20185 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
20188 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
20189 regnode* const node,
20191 SV* const runtime_defns,
20192 SV* const only_utf8_locale_list)
20194 /* Sets the arg field of an ANYOF-type node 'node', using information about
20195 * the node passed-in. If there is nothing outside the node's bitmap, the
20196 * arg is set to ANYOF_ONLY_HAS_BITMAP. Otherwise, it sets the argument to
20197 * the count returned by add_data(), having allocated and stored an array,
20200 * av[0] stores the inversion list defining this class as far as known at
20201 * this time, or PL_sv_undef if nothing definite is now known.
20202 * av[1] stores the inversion list of code points that match only if the
20203 * current locale is UTF-8, or if none, PL_sv_undef if there is an
20204 * av[2], or no entry otherwise.
20205 * av[2] stores the list of user-defined properties whose subroutine
20206 * definitions aren't known at this time, or no entry if none. */
20210 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
20212 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
20213 assert(! (ANYOF_FLAGS(node)
20214 & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP));
20215 ARG_SET(node, ANYOF_ONLY_HAS_BITMAP);
20218 AV * const av = newAV();
20222 av_store(av, INVLIST_INDEX, SvREFCNT_inc_NN(cp_list));
20225 /* (Note that if any of this changes, the size calculations in
20226 * S_optimize_regclass() might need to be updated.) */
20228 if (only_utf8_locale_list) {
20229 av_store(av, ONLY_LOCALE_MATCHES_INDEX,
20230 SvREFCNT_inc_NN(only_utf8_locale_list));
20233 if (runtime_defns) {
20234 av_store(av, DEFERRED_USER_DEFINED_INDEX,
20235 SvREFCNT_inc_NN(runtime_defns));
20238 rv = newRV_noinc(MUTABLE_SV(av));
20239 n = add_data(pRExC_state, STR_WITH_LEN("s"));
20240 RExC_rxi->data->data[n] = (void*)rv;
20247 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
20248 Perl_get_regclass_nonbitmap_data(pTHX_ const regexp *prog, const regnode* node, bool doinit, SV** listsvp, SV** only_utf8_locale_ptr, SV** output_invlist)
20250 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)
20254 /* For internal core use only.
20255 * Returns the inversion list for the input 'node' in the regex 'prog'.
20256 * If <doinit> is 'true', will attempt to create the inversion list if not
20258 * If <listsvp> is non-null, will return the printable contents of the
20259 * property definition. This can be used to get debugging information
20260 * even before the inversion list exists, by calling this function with
20261 * 'doinit' set to false, in which case the components that will be used
20262 * to eventually create the inversion list are returned (in a printable
20264 * If <only_utf8_locale_ptr> is not NULL, it is where this routine is to
20265 * store an inversion list of code points that should match only if the
20266 * execution-time locale is a UTF-8 one.
20267 * If <output_invlist> is not NULL, it is where this routine is to store an
20268 * inversion list of the code points that would be instead returned in
20269 * <listsvp> if this were NULL. Thus, what gets output in <listsvp>
20270 * when this parameter is used, is just the non-code point data that
20271 * will go into creating the inversion list. This currently should be just
20272 * user-defined properties whose definitions were not known at compile
20273 * time. Using this parameter allows for easier manipulation of the
20274 * inversion list's data by the caller. It is illegal to call this
20275 * function with this parameter set, but not <listsvp>
20277 * Tied intimately to how S_set_ANYOF_arg sets up the data structure. Note
20278 * that, in spite of this function's name, the inversion list it returns
20279 * may include the bitmap data as well */
20281 SV *si = NULL; /* Input initialization string */
20282 SV* invlist = NULL;
20284 RXi_GET_DECL(prog, progi);
20285 const struct reg_data * const data = prog ? progi->data : NULL;
20287 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
20288 PERL_ARGS_ASSERT_GET_REGCLASS_NONBITMAP_DATA;
20290 PERL_ARGS_ASSERT_GET_RE_GCLASS_NONBITMAP_DATA;
20292 assert(! output_invlist || listsvp);
20294 if (data && data->count) {
20295 const U32 n = ARG(node);
20297 if (data->what[n] == 's') {
20298 SV * const rv = MUTABLE_SV(data->data[n]);
20299 AV * const av = MUTABLE_AV(SvRV(rv));
20300 SV **const ary = AvARRAY(av);
20302 invlist = ary[INVLIST_INDEX];
20304 if (av_tindex_skip_len_mg(av) >= ONLY_LOCALE_MATCHES_INDEX) {
20305 *only_utf8_locale_ptr = ary[ONLY_LOCALE_MATCHES_INDEX];
20308 if (av_tindex_skip_len_mg(av) >= DEFERRED_USER_DEFINED_INDEX) {
20309 si = ary[DEFERRED_USER_DEFINED_INDEX];
20312 if (doinit && (si || invlist)) {
20315 SV * msg = newSVpvs_flags("", SVs_TEMP);
20317 SV * prop_definition = handle_user_defined_property(
20318 "", 0, FALSE, /* There is no \p{}, \P{} */
20319 SvPVX_const(si)[1] - '0', /* /i or not has been
20320 stored here for just
20322 TRUE, /* run time */
20323 FALSE, /* This call must find the defn */
20324 si, /* The property definition */
20327 0 /* base level call */
20331 assert(prop_definition == NULL);
20333 Perl_croak(aTHX_ "%" UTF8f,
20334 UTF8fARG(SvUTF8(msg), SvCUR(msg), SvPVX(msg)));
20338 _invlist_union(invlist, prop_definition, &invlist);
20339 SvREFCNT_dec_NN(prop_definition);
20342 invlist = prop_definition;
20345 STATIC_ASSERT_STMT(ONLY_LOCALE_MATCHES_INDEX == 1 + INVLIST_INDEX);
20346 STATIC_ASSERT_STMT(DEFERRED_USER_DEFINED_INDEX == 1 + ONLY_LOCALE_MATCHES_INDEX);
20348 ary[INVLIST_INDEX] = invlist;
20349 av_fill(av, (ary[ONLY_LOCALE_MATCHES_INDEX])
20350 ? ONLY_LOCALE_MATCHES_INDEX
20358 /* If requested, return a printable version of what this ANYOF node matches
20361 SV* matches_string = NULL;
20363 /* This function can be called at compile-time, before everything gets
20364 * resolved, in which case we return the currently best available
20365 * information, which is the string that will eventually be used to do
20366 * that resolving, 'si' */
20368 /* Here, we only have 'si' (and possibly some passed-in data in
20369 * 'invlist', which is handled below) If the caller only wants
20370 * 'si', use that. */
20371 if (! output_invlist) {
20372 matches_string = newSVsv(si);
20375 /* But if the caller wants an inversion list of the node, we
20376 * need to parse 'si' and place as much as possible in the
20377 * desired output inversion list, making 'matches_string' only
20378 * contain the currently unresolvable things */
20379 const char *si_string = SvPVX(si);
20380 STRLEN remaining = SvCUR(si);
20384 /* Ignore everything before and including the first new-line */
20385 si_string = (const char *) memchr(si_string, '\n', SvCUR(si));
20386 assert (si_string != NULL);
20388 remaining = SvPVX(si) + SvCUR(si) - si_string;
20390 while (remaining > 0) {
20392 /* The data consists of just strings defining user-defined
20393 * property names, but in prior incarnations, and perhaps
20394 * somehow from pluggable regex engines, it could still
20395 * hold hex code point definitions, all of which should be
20396 * legal (or it wouldn't have gotten this far). Each
20397 * component of a range would be separated by a tab, and
20398 * each range by a new-line. If these are found, instead
20399 * add them to the inversion list */
20400 I32 grok_flags = PERL_SCAN_SILENT_ILLDIGIT
20401 |PERL_SCAN_SILENT_NON_PORTABLE;
20402 STRLEN len = remaining;
20403 UV cp = grok_hex(si_string, &len, &grok_flags, NULL);
20405 /* If the hex decode routine found something, it should go
20406 * up to the next \n */
20407 if ( *(si_string + len) == '\n') {
20408 if (count) { /* 2nd code point on line */
20409 *output_invlist = _add_range_to_invlist(*output_invlist, prev_cp, cp);
20412 *output_invlist = add_cp_to_invlist(*output_invlist, cp);
20415 goto prepare_for_next_iteration;
20418 /* If the hex decode was instead for the lower range limit,
20419 * save it, and go parse the upper range limit */
20420 if (*(si_string + len) == '\t') {
20421 assert(count == 0);
20425 prepare_for_next_iteration:
20426 si_string += len + 1;
20427 remaining -= len + 1;
20431 /* Here, didn't find a legal hex number. Just add the text
20432 * from here up to the next \n, omitting any trailing
20436 len = strcspn(si_string,
20437 DEFERRED_COULD_BE_OFFICIAL_MARKERs "\n");
20439 if (matches_string) {
20440 sv_catpvn(matches_string, si_string, len);
20443 matches_string = newSVpvn(si_string, len);
20445 sv_catpvs(matches_string, " ");
20449 && UCHARAT(si_string)
20450 == DEFERRED_COULD_BE_OFFICIAL_MARKERc)
20455 if (remaining && UCHARAT(si_string) == '\n') {
20459 } /* end of loop through the text */
20461 assert(matches_string);
20462 if (SvCUR(matches_string)) { /* Get rid of trailing blank */
20463 SvCUR_set(matches_string, SvCUR(matches_string) - 1);
20465 } /* end of has an 'si' */
20468 /* Add the stuff that's already known */
20471 /* Again, if the caller doesn't want the output inversion list, put
20472 * everything in 'matches-string' */
20473 if (! output_invlist) {
20474 if ( ! matches_string) {
20475 matches_string = newSVpvs("\n");
20477 sv_catsv(matches_string, invlist_contents(invlist,
20478 TRUE /* traditional style */
20481 else if (! *output_invlist) {
20482 *output_invlist = invlist_clone(invlist, NULL);
20485 _invlist_union(*output_invlist, invlist, output_invlist);
20489 *listsvp = matches_string;
20495 /* reg_skipcomment()
20497 Absorbs an /x style # comment from the input stream,
20498 returning a pointer to the first character beyond the comment, or if the
20499 comment terminates the pattern without anything following it, this returns
20500 one past the final character of the pattern (in other words, RExC_end) and
20501 sets the REG_RUN_ON_COMMENT_SEEN flag.
20503 Note it's the callers responsibility to ensure that we are
20504 actually in /x mode
20508 PERL_STATIC_INLINE char*
20509 S_reg_skipcomment(RExC_state_t *pRExC_state, char* p)
20511 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
20515 while (p < RExC_end) {
20516 if (*(++p) == '\n') {
20521 /* we ran off the end of the pattern without ending the comment, so we have
20522 * to add an \n when wrapping */
20523 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
20528 S_skip_to_be_ignored_text(pTHX_ RExC_state_t *pRExC_state,
20530 const bool force_to_xmod
20533 /* If the text at the current parse position '*p' is a '(?#...)' comment,
20534 * or if we are under /x or 'force_to_xmod' is TRUE, and the text at '*p'
20535 * is /x whitespace, advance '*p' so that on exit it points to the first
20536 * byte past all such white space and comments */
20538 const bool use_xmod = force_to_xmod || (RExC_flags & RXf_PMf_EXTENDED);
20540 PERL_ARGS_ASSERT_SKIP_TO_BE_IGNORED_TEXT;
20542 assert( ! UTF || UTF8_IS_INVARIANT(**p) || UTF8_IS_START(**p));
20545 if (RExC_end - (*p) >= 3
20547 && *(*p + 1) == '?'
20548 && *(*p + 2) == '#')
20550 while (*(*p) != ')') {
20551 if ((*p) == RExC_end)
20552 FAIL("Sequence (?#... not terminated");
20560 const char * save_p = *p;
20561 while ((*p) < RExC_end) {
20563 if ((len = is_PATWS_safe((*p), RExC_end, UTF))) {
20566 else if (*(*p) == '#') {
20567 (*p) = reg_skipcomment(pRExC_state, (*p));
20573 if (*p != save_p) {
20586 Advances the parse position by one byte, unless that byte is the beginning
20587 of a '(?#...)' style comment, or is /x whitespace and /x is in effect. In
20588 those two cases, the parse position is advanced beyond all such comments and
20591 This is the UTF, (?#...), and /x friendly way of saying RExC_parse++.
20595 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
20597 PERL_ARGS_ASSERT_NEXTCHAR;
20599 if (RExC_parse < RExC_end) {
20601 || UTF8_IS_INVARIANT(*RExC_parse)
20602 || UTF8_IS_START(*RExC_parse));
20604 RExC_parse += (UTF)
20605 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
20608 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
20609 FALSE /* Don't force /x */ );
20614 S_change_engine_size(pTHX_ RExC_state_t *pRExC_state, const Ptrdiff_t size)
20616 /* 'size' is the delta number of smallest regnode equivalents to add or
20617 * subtract from the current memory allocated to the regex engine being
20620 PERL_ARGS_ASSERT_CHANGE_ENGINE_SIZE;
20625 sizeof(regexp_internal) + (RExC_size + 1) * sizeof(regnode),
20626 /* +1 for REG_MAGIC */
20629 if ( RExC_rxi == NULL )
20630 FAIL("Regexp out of space");
20631 RXi_SET(RExC_rx, RExC_rxi);
20633 RExC_emit_start = RExC_rxi->program;
20635 Zero(REGNODE_p(RExC_emit), size, regnode);
20638 #ifdef RE_TRACK_PATTERN_OFFSETS
20639 Renew(RExC_offsets, 2*RExC_size+1, U32);
20641 Zero(RExC_offsets + 2*(RExC_size - size) + 1, 2 * size, U32);
20643 RExC_offsets[0] = RExC_size;
20647 STATIC regnode_offset
20648 S_regnode_guts(pTHX_ RExC_state_t *pRExC_state, const U8 op, const STRLEN extra_size, const char* const name)
20650 /* Allocate a regnode for 'op', with 'extra_size' extra (smallest) regnode
20651 * equivalents space. It aligns and increments RExC_size
20653 * It returns the regnode's offset into the regex engine program */
20655 const regnode_offset ret = RExC_emit;
20657 DECLARE_AND_GET_RE_DEBUG_FLAGS;
20659 PERL_ARGS_ASSERT_REGNODE_GUTS;
20661 SIZE_ALIGN(RExC_size);
20662 change_engine_size(pRExC_state, (Ptrdiff_t) 1 + extra_size);
20663 NODE_ALIGN_FILL(REGNODE_p(ret));
20664 #ifndef RE_TRACK_PATTERN_OFFSETS
20665 PERL_UNUSED_ARG(name);
20666 PERL_UNUSED_ARG(op);
20668 assert(extra_size >= regarglen[op] || PL_regkind[op] == ANYOF);
20670 if (RExC_offsets) { /* MJD */
20672 ("%s:%d: (op %s) %s %" UVuf " (len %" UVuf ") (max %" UVuf ").\n",
20675 (UV)(RExC_emit) > RExC_offsets[0]
20676 ? "Overwriting end of array!\n" : "OK",
20678 (UV)(RExC_parse - RExC_start),
20679 (UV)RExC_offsets[0]));
20680 Set_Node_Offset(REGNODE_p(RExC_emit), RExC_parse + (op == END));
20687 - reg_node - emit a node
20689 STATIC regnode_offset /* Location. */
20690 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
20692 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reg_node");
20693 regnode_offset ptr = ret;
20695 PERL_ARGS_ASSERT_REG_NODE;
20697 assert(regarglen[op] == 0);
20699 FILL_ADVANCE_NODE(ptr, op);
20705 - reganode - emit a node with an argument
20707 STATIC regnode_offset /* Location. */
20708 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
20710 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reganode");
20711 regnode_offset ptr = ret;
20713 PERL_ARGS_ASSERT_REGANODE;
20715 /* ANYOF are special cased to allow non-length 1 args */
20716 assert(regarglen[op] == 1);
20718 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
20724 - regpnode - emit a temporary node with a SV* argument
20726 STATIC regnode_offset /* Location. */
20727 S_regpnode(pTHX_ RExC_state_t *pRExC_state, U8 op, SV * arg)
20729 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "regpnode");
20730 regnode_offset ptr = ret;
20732 PERL_ARGS_ASSERT_REGPNODE;
20734 FILL_ADVANCE_NODE_ARGp(ptr, op, arg);
20739 STATIC regnode_offset
20740 S_reg2Lanode(pTHX_ RExC_state_t *pRExC_state, const U8 op, const U32 arg1, const I32 arg2)
20742 /* emit a node with U32 and I32 arguments */
20744 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reg2Lanode");
20745 regnode_offset ptr = ret;
20747 PERL_ARGS_ASSERT_REG2LANODE;
20749 assert(regarglen[op] == 2);
20751 FILL_ADVANCE_NODE_2L_ARG(ptr, op, arg1, arg2);
20757 - reginsert - insert an operator in front of already-emitted operand
20759 * That means that on exit 'operand' is the offset of the newly inserted
20760 * operator, and the original operand has been relocated.
20762 * IMPORTANT NOTE - it is the *callers* responsibility to correctly
20763 * set up NEXT_OFF() of the inserted node if needed. Something like this:
20765 * reginsert(pRExC, OPFAIL, orig_emit, depth+1);
20766 * NEXT_OFF(orig_emit) = regarglen[OPFAIL] + NODE_STEP_REGNODE;
20768 * ALSO NOTE - FLAGS(newly-inserted-operator) will be set to 0 as well.
20771 S_reginsert(pTHX_ RExC_state_t *pRExC_state, const U8 op,
20772 const regnode_offset operand, const U32 depth)
20777 const int offset = regarglen[(U8)op];
20778 const int size = NODE_STEP_REGNODE + offset;
20779 DECLARE_AND_GET_RE_DEBUG_FLAGS;
20781 PERL_ARGS_ASSERT_REGINSERT;
20782 PERL_UNUSED_CONTEXT;
20783 PERL_UNUSED_ARG(depth);
20784 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
20785 DEBUG_PARSE_FMT("inst"," - %s", PL_reg_name[op]);
20786 assert(!RExC_study_started); /* I believe we should never use reginsert once we have started
20787 studying. If this is wrong then we need to adjust RExC_recurse
20788 below like we do with RExC_open_parens/RExC_close_parens. */
20789 change_engine_size(pRExC_state, (Ptrdiff_t) size);
20790 src = REGNODE_p(RExC_emit);
20792 dst = REGNODE_p(RExC_emit);
20794 /* If we are in a "count the parentheses" pass, the numbers are unreliable,
20795 * and [perl #133871] shows this can lead to problems, so skip this
20796 * realignment of parens until a later pass when they are reliable */
20797 if (! IN_PARENS_PASS && RExC_open_parens) {
20799 /*DEBUG_PARSE_FMT("inst"," - %" IVdf, (IV)RExC_npar);*/
20800 /* remember that RExC_npar is rex->nparens + 1,
20801 * iow it is 1 more than the number of parens seen in
20802 * the pattern so far. */
20803 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
20804 /* note, RExC_open_parens[0] is the start of the
20805 * regex, it can't move. RExC_close_parens[0] is the end
20806 * of the regex, it *can* move. */
20807 if ( paren && RExC_open_parens[paren] >= operand ) {
20808 /*DEBUG_PARSE_FMT("open"," - %d", size);*/
20809 RExC_open_parens[paren] += size;
20811 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
20813 if ( RExC_close_parens[paren] >= operand ) {
20814 /*DEBUG_PARSE_FMT("close"," - %d", size);*/
20815 RExC_close_parens[paren] += size;
20817 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
20822 RExC_end_op += size;
20824 while (src > REGNODE_p(operand)) {
20825 StructCopy(--src, --dst, regnode);
20826 #ifdef RE_TRACK_PATTERN_OFFSETS
20827 if (RExC_offsets) { /* MJD 20010112 */
20829 ("%s(%d): (op %s) %s copy %" UVuf " -> %" UVuf " (max %" UVuf ").\n",
20833 (UV)(REGNODE_OFFSET(dst)) > RExC_offsets[0]
20834 ? "Overwriting end of array!\n" : "OK",
20835 (UV)REGNODE_OFFSET(src),
20836 (UV)REGNODE_OFFSET(dst),
20837 (UV)RExC_offsets[0]));
20838 Set_Node_Offset_To_R(REGNODE_OFFSET(dst), Node_Offset(src));
20839 Set_Node_Length_To_R(REGNODE_OFFSET(dst), Node_Length(src));
20844 place = REGNODE_p(operand); /* Op node, where operand used to be. */
20845 #ifdef RE_TRACK_PATTERN_OFFSETS
20846 if (RExC_offsets) { /* MJD */
20848 ("%s(%d): (op %s) %s %" UVuf " <- %" UVuf " (max %" UVuf ").\n",
20852 (UV)REGNODE_OFFSET(place) > RExC_offsets[0]
20853 ? "Overwriting end of array!\n" : "OK",
20854 (UV)REGNODE_OFFSET(place),
20855 (UV)(RExC_parse - RExC_start),
20856 (UV)RExC_offsets[0]));
20857 Set_Node_Offset(place, RExC_parse);
20858 Set_Node_Length(place, 1);
20861 src = NEXTOPER(place);
20863 FILL_NODE(operand, op);
20865 /* Zero out any arguments in the new node */
20866 Zero(src, offset, regnode);
20870 - regtail - set the next-pointer at the end of a node chain of p to val. If
20871 that value won't fit in the space available, instead returns FALSE.
20872 (Except asserts if we can't fit in the largest space the regex
20873 engine is designed for.)
20874 - SEE ALSO: regtail_study
20877 S_regtail(pTHX_ RExC_state_t * pRExC_state,
20878 const regnode_offset p,
20879 const regnode_offset val,
20882 regnode_offset scan;
20883 DECLARE_AND_GET_RE_DEBUG_FLAGS;
20885 PERL_ARGS_ASSERT_REGTAIL;
20887 PERL_UNUSED_ARG(depth);
20890 /* The final node in the chain is the first one with a nonzero next pointer
20892 scan = (regnode_offset) p;
20894 regnode * const temp = regnext(REGNODE_p(scan));
20896 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
20897 regprop(RExC_rx, RExC_mysv, REGNODE_p(scan), NULL, pRExC_state);
20898 Perl_re_printf( aTHX_ "~ %s (%zu) %s %s\n",
20899 SvPV_nolen_const(RExC_mysv), scan,
20900 (temp == NULL ? "->" : ""),
20901 (temp == NULL ? PL_reg_name[OP(REGNODE_p(val))] : "")
20906 scan = REGNODE_OFFSET(temp);
20909 /* Populate this node's next pointer */
20910 assert(val >= scan);
20911 if (reg_off_by_arg[OP(REGNODE_p(scan))]) {
20912 assert((UV) (val - scan) <= U32_MAX);
20913 ARG_SET(REGNODE_p(scan), val - scan);
20916 if (val - scan > U16_MAX) {
20917 /* Populate this with something that won't loop and will likely
20918 * lead to a crash if the caller ignores the failure return, and
20919 * execution continues */
20920 NEXT_OFF(REGNODE_p(scan)) = U16_MAX;
20923 NEXT_OFF(REGNODE_p(scan)) = val - scan;
20931 - regtail_study - set the next-pointer at the end of a node chain of p to val.
20932 - Look for optimizable sequences at the same time.
20933 - currently only looks for EXACT chains.
20935 This is experimental code. The idea is to use this routine to perform
20936 in place optimizations on branches and groups as they are constructed,
20937 with the long term intention of removing optimization from study_chunk so
20938 that it is purely analytical.
20940 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
20941 to control which is which.
20943 This used to return a value that was ignored. It was a problem that it is
20944 #ifdef'd to be another function that didn't return a value. khw has changed it
20945 so both currently return a pass/fail return.
20948 /* TODO: All four parms should be const */
20951 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode_offset p,
20952 const regnode_offset val, U32 depth)
20954 regnode_offset scan;
20956 #ifdef EXPERIMENTAL_INPLACESCAN
20959 DECLARE_AND_GET_RE_DEBUG_FLAGS;
20961 PERL_ARGS_ASSERT_REGTAIL_STUDY;
20964 /* Find last node. */
20968 regnode * const temp = regnext(REGNODE_p(scan));
20969 #ifdef EXPERIMENTAL_INPLACESCAN
20970 if (PL_regkind[OP(REGNODE_p(scan))] == EXACT) {
20971 bool unfolded_multi_char; /* Unexamined in this routine */
20972 if (join_exact(pRExC_state, scan, &min,
20973 &unfolded_multi_char, 1, REGNODE_p(val), depth+1))
20974 return TRUE; /* Was return EXACT */
20978 if (PL_regkind[OP(REGNODE_p(scan))] == EXACT) {
20979 if (exact == PSEUDO )
20980 exact= OP(REGNODE_p(scan));
20981 else if (exact != OP(REGNODE_p(scan)) )
20984 else if (OP(REGNODE_p(scan)) != NOTHING) {
20989 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
20990 regprop(RExC_rx, RExC_mysv, REGNODE_p(scan), NULL, pRExC_state);
20991 Perl_re_printf( aTHX_ "~ %s (%zu) -> %s\n",
20992 SvPV_nolen_const(RExC_mysv),
20994 PL_reg_name[exact]);
20998 scan = REGNODE_OFFSET(temp);
21001 DEBUG_PARSE_MSG("");
21002 regprop(RExC_rx, RExC_mysv, REGNODE_p(val), NULL, pRExC_state);
21003 Perl_re_printf( aTHX_
21004 "~ attach to %s (%" IVdf ") offset to %" IVdf "\n",
21005 SvPV_nolen_const(RExC_mysv),
21010 if (reg_off_by_arg[OP(REGNODE_p(scan))]) {
21011 assert((UV) (val - scan) <= U32_MAX);
21012 ARG_SET(REGNODE_p(scan), val - scan);
21015 if (val - scan > U16_MAX) {
21016 /* Populate this with something that won't loop and will likely
21017 * lead to a crash if the caller ignores the failure return, and
21018 * execution continues */
21019 NEXT_OFF(REGNODE_p(scan)) = U16_MAX;
21022 NEXT_OFF(REGNODE_p(scan)) = val - scan;
21025 return TRUE; /* Was 'return exact' */
21030 S_get_ANYOFM_contents(pTHX_ const regnode * n) {
21032 /* Returns an inversion list of all the code points matched by the
21033 * ANYOFM/NANYOFM node 'n' */
21035 SV * cp_list = _new_invlist(-1);
21036 const U8 lowest = (U8) ARG(n);
21039 U8 needed = 1U << PL_bitcount[ (U8) ~ FLAGS(n)];
21041 PERL_ARGS_ASSERT_GET_ANYOFM_CONTENTS;
21043 /* Starting with the lowest code point, any code point that ANDed with the
21044 * mask yields the lowest code point is in the set */
21045 for (i = lowest; i <= 0xFF; i++) {
21046 if ((i & FLAGS(n)) == ARG(n)) {
21047 cp_list = add_cp_to_invlist(cp_list, i);
21050 /* We know how many code points (a power of two) that are in the
21051 * set. No use looking once we've got that number */
21052 if (count >= needed) break;
21056 if (OP(n) == NANYOFM) {
21057 _invlist_invert(cp_list);
21063 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
21068 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
21073 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
21075 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
21076 if (flags & (1<<bit)) {
21077 if (!set++ && lead)
21078 Perl_re_printf( aTHX_ "%s", lead);
21079 Perl_re_printf( aTHX_ "%s ", PL_reg_intflags_name[bit]);
21084 Perl_re_printf( aTHX_ "\n");
21086 Perl_re_printf( aTHX_ "%s[none-set]\n", lead);
21091 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
21097 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
21099 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
21100 if (flags & (1<<bit)) {
21101 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
21104 if (!set++ && lead)
21105 Perl_re_printf( aTHX_ "%s", lead);
21106 Perl_re_printf( aTHX_ "%s ", PL_reg_extflags_name[bit]);
21109 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
21110 if (!set++ && lead) {
21111 Perl_re_printf( aTHX_ "%s", lead);
21114 case REGEX_UNICODE_CHARSET:
21115 Perl_re_printf( aTHX_ "UNICODE");
21117 case REGEX_LOCALE_CHARSET:
21118 Perl_re_printf( aTHX_ "LOCALE");
21120 case REGEX_ASCII_RESTRICTED_CHARSET:
21121 Perl_re_printf( aTHX_ "ASCII-RESTRICTED");
21123 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
21124 Perl_re_printf( aTHX_ "ASCII-MORE_RESTRICTED");
21127 Perl_re_printf( aTHX_ "UNKNOWN CHARACTER SET");
21133 Perl_re_printf( aTHX_ "\n");
21135 Perl_re_printf( aTHX_ "%s[none-set]\n", lead);
21141 Perl_regdump(pTHX_ const regexp *r)
21145 SV * const sv = sv_newmortal();
21146 SV *dsv= sv_newmortal();
21147 RXi_GET_DECL(r, ri);
21148 DECLARE_AND_GET_RE_DEBUG_FLAGS;
21150 PERL_ARGS_ASSERT_REGDUMP;
21152 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
21154 /* Header fields of interest. */
21155 for (i = 0; i < 2; i++) {
21156 if (r->substrs->data[i].substr) {
21157 RE_PV_QUOTED_DECL(s, 0, dsv,
21158 SvPVX_const(r->substrs->data[i].substr),
21159 RE_SV_DUMPLEN(r->substrs->data[i].substr),
21160 PL_dump_re_max_len);
21161 Perl_re_printf( aTHX_
21162 "%s %s%s at %" IVdf "..%" UVuf " ",
21163 i ? "floating" : "anchored",
21165 RE_SV_TAIL(r->substrs->data[i].substr),
21166 (IV)r->substrs->data[i].min_offset,
21167 (UV)r->substrs->data[i].max_offset);
21169 else if (r->substrs->data[i].utf8_substr) {
21170 RE_PV_QUOTED_DECL(s, 1, dsv,
21171 SvPVX_const(r->substrs->data[i].utf8_substr),
21172 RE_SV_DUMPLEN(r->substrs->data[i].utf8_substr),
21174 Perl_re_printf( aTHX_
21175 "%s utf8 %s%s at %" IVdf "..%" UVuf " ",
21176 i ? "floating" : "anchored",
21178 RE_SV_TAIL(r->substrs->data[i].utf8_substr),
21179 (IV)r->substrs->data[i].min_offset,
21180 (UV)r->substrs->data[i].max_offset);
21184 if (r->check_substr || r->check_utf8)
21185 Perl_re_printf( aTHX_
21187 ( r->check_substr == r->substrs->data[1].substr
21188 && r->check_utf8 == r->substrs->data[1].utf8_substr
21189 ? "(checking floating" : "(checking anchored"));
21190 if (r->intflags & PREGf_NOSCAN)
21191 Perl_re_printf( aTHX_ " noscan");
21192 if (r->extflags & RXf_CHECK_ALL)
21193 Perl_re_printf( aTHX_ " isall");
21194 if (r->check_substr || r->check_utf8)
21195 Perl_re_printf( aTHX_ ") ");
21197 if (ri->regstclass) {
21198 regprop(r, sv, ri->regstclass, NULL, NULL);
21199 Perl_re_printf( aTHX_ "stclass %s ", SvPVX_const(sv));
21201 if (r->intflags & PREGf_ANCH) {
21202 Perl_re_printf( aTHX_ "anchored");
21203 if (r->intflags & PREGf_ANCH_MBOL)
21204 Perl_re_printf( aTHX_ "(MBOL)");
21205 if (r->intflags & PREGf_ANCH_SBOL)
21206 Perl_re_printf( aTHX_ "(SBOL)");
21207 if (r->intflags & PREGf_ANCH_GPOS)
21208 Perl_re_printf( aTHX_ "(GPOS)");
21209 Perl_re_printf( aTHX_ " ");
21211 if (r->intflags & PREGf_GPOS_SEEN)
21212 Perl_re_printf( aTHX_ "GPOS:%" UVuf " ", (UV)r->gofs);
21213 if (r->intflags & PREGf_SKIP)
21214 Perl_re_printf( aTHX_ "plus ");
21215 if (r->intflags & PREGf_IMPLICIT)
21216 Perl_re_printf( aTHX_ "implicit ");
21217 Perl_re_printf( aTHX_ "minlen %" IVdf " ", (IV)r->minlen);
21218 if (r->extflags & RXf_EVAL_SEEN)
21219 Perl_re_printf( aTHX_ "with eval ");
21220 Perl_re_printf( aTHX_ "\n");
21222 regdump_extflags("r->extflags: ", r->extflags);
21223 regdump_intflags("r->intflags: ", r->intflags);
21226 PERL_ARGS_ASSERT_REGDUMP;
21227 PERL_UNUSED_CONTEXT;
21228 PERL_UNUSED_ARG(r);
21229 #endif /* DEBUGGING */
21232 /* Should be synchronized with ANYOF_ #defines in regcomp.h */
21235 # if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 \
21236 || _CC_LOWER != 3 || _CC_UPPER != 4 || _CC_PUNCT != 5 \
21237 || _CC_PRINT != 6 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 \
21238 || _CC_CASED != 9 || _CC_SPACE != 10 || _CC_BLANK != 11 \
21239 || _CC_XDIGIT != 12 || _CC_CNTRL != 13 || _CC_ASCII != 14 \
21240 || _CC_VERTSPACE != 15
21241 # error Need to adjust order of anyofs[]
21243 static const char * const anyofs[] = {
21280 - regprop - printable representation of opcode, with run time support
21284 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo, const RExC_state_t *pRExC_state)
21288 RXi_GET_DECL(prog, progi);
21289 DECLARE_AND_GET_RE_DEBUG_FLAGS;
21291 PERL_ARGS_ASSERT_REGPROP;
21295 if (OP(o) > REGNODE_MAX) { /* regnode.type is unsigned */
21296 if (pRExC_state) { /* This gives more info, if we have it */
21297 FAIL3("panic: corrupted regexp opcode %d > %d",
21298 (int)OP(o), (int)REGNODE_MAX);
21301 Perl_croak(aTHX_ "panic: corrupted regexp opcode %d > %d",
21302 (int)OP(o), (int)REGNODE_MAX);
21305 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
21307 k = PL_regkind[OP(o)];
21310 sv_catpvs(sv, " ");
21311 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
21312 * is a crude hack but it may be the best for now since
21313 * we have no flag "this EXACTish node was UTF-8"
21315 pv_pretty(sv, STRING(o), STR_LEN(o), PL_dump_re_max_len,
21316 PL_colors[0], PL_colors[1],
21317 PERL_PV_ESCAPE_UNI_DETECT |
21318 PERL_PV_ESCAPE_NONASCII |
21319 PERL_PV_PRETTY_ELLIPSES |
21320 PERL_PV_PRETTY_LTGT |
21321 PERL_PV_PRETTY_NOCLEAR
21323 } else if (k == TRIE) {
21324 /* print the details of the trie in dumpuntil instead, as
21325 * progi->data isn't available here */
21326 const char op = OP(o);
21327 const U32 n = ARG(o);
21328 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
21329 (reg_ac_data *)progi->data->data[n] :
21331 const reg_trie_data * const trie
21332 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
21334 Perl_sv_catpvf(aTHX_ sv, "-%s", PL_reg_name[o->flags]);
21335 DEBUG_TRIE_COMPILE_r({
21337 sv_catpvs(sv, "(JUMP)");
21338 Perl_sv_catpvf(aTHX_ sv,
21339 "<S:%" UVuf "/%" IVdf " W:%" UVuf " L:%" UVuf "/%" UVuf " C:%" UVuf "/%" UVuf ">",
21340 (UV)trie->startstate,
21341 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
21342 (UV)trie->wordcount,
21345 (UV)TRIE_CHARCOUNT(trie),
21346 (UV)trie->uniquecharcount
21349 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
21350 sv_catpvs(sv, "[");
21351 (void) put_charclass_bitmap_innards(sv,
21352 ((IS_ANYOF_TRIE(op))
21354 : TRIE_BITMAP(trie)),
21361 sv_catpvs(sv, "]");
21363 } else if (k == CURLY) {
21364 U32 lo = ARG1(o), hi = ARG2(o);
21365 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
21366 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
21367 Perl_sv_catpvf(aTHX_ sv, "{%u,", (unsigned) lo);
21368 if (hi == REG_INFTY)
21369 sv_catpvs(sv, "INFTY");
21371 Perl_sv_catpvf(aTHX_ sv, "%u", (unsigned) hi);
21372 sv_catpvs(sv, "}");
21374 else if (k == WHILEM && o->flags) /* Ordinal/of */
21375 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
21376 else if (k == REF || k == OPEN || k == CLOSE
21377 || k == GROUPP || OP(o)==ACCEPT)
21379 AV *name_list= NULL;
21380 U32 parno= OP(o) == ACCEPT ? (U32)ARG2L(o) : ARG(o);
21381 Perl_sv_catpvf(aTHX_ sv, "%" UVuf, (UV)parno); /* Parenth number */
21382 if ( RXp_PAREN_NAMES(prog) ) {
21383 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
21384 } else if ( pRExC_state ) {
21385 name_list= RExC_paren_name_list;
21388 if ( k != REF || (OP(o) < REFN)) {
21389 SV **name= av_fetch(name_list, parno, 0 );
21391 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
21394 SV *sv_dat= MUTABLE_SV(progi->data->data[ parno ]);
21395 I32 *nums=(I32*)SvPVX(sv_dat);
21396 SV **name= av_fetch(name_list, nums[0], 0 );
21399 for ( n=0; n<SvIVX(sv_dat); n++ ) {
21400 Perl_sv_catpvf(aTHX_ sv, "%s%" IVdf,
21401 (n ? "," : ""), (IV)nums[n]);
21403 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
21407 if ( k == REF && reginfo) {
21408 U32 n = ARG(o); /* which paren pair */
21409 I32 ln = prog->offs[n].start;
21410 if (prog->lastparen < n || ln == -1 || prog->offs[n].end == -1)
21411 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
21412 else if (ln == prog->offs[n].end)
21413 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
21415 const char *s = reginfo->strbeg + ln;
21416 Perl_sv_catpvf(aTHX_ sv, ": ");
21417 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
21418 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
21421 } else if (k == GOSUB) {
21422 AV *name_list= NULL;
21423 if ( RXp_PAREN_NAMES(prog) ) {
21424 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
21425 } else if ( pRExC_state ) {
21426 name_list= RExC_paren_name_list;
21429 /* Paren and offset */
21430 Perl_sv_catpvf(aTHX_ sv, "%d[%+d:%d]", (int)ARG(o),(int)ARG2L(o),
21431 (int)((o + (int)ARG2L(o)) - progi->program) );
21433 SV **name= av_fetch(name_list, ARG(o), 0 );
21435 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
21438 else if (k == LOGICAL)
21439 /* 2: embedded, otherwise 1 */
21440 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
21441 else if (k == ANYOF || k == ANYOFR) {
21445 bool do_sep = FALSE; /* Do we need to separate various components of
21447 /* Set if there is still an unresolved user-defined property */
21448 SV *unresolved = NULL;
21450 /* Things that are ignored except when the runtime locale is UTF-8 */
21451 SV *only_utf8_locale_invlist = NULL;
21453 /* Code points that don't fit in the bitmap */
21454 SV *nonbitmap_invlist = NULL;
21456 /* And things that aren't in the bitmap, but are small enough to be */
21457 SV* bitmap_range_not_in_bitmap = NULL;
21461 if (inRANGE(OP(o), ANYOFH, ANYOFRb)) {
21467 flags = ANYOF_FLAGS(o);
21468 bitmap = ANYOF_BITMAP(o);
21472 if (OP(o) == ANYOFL || OP(o) == ANYOFPOSIXL) {
21473 if (ANYOFL_UTF8_LOCALE_REQD(flags)) {
21474 sv_catpvs(sv, "{utf8-locale-reqd}");
21476 if (flags & ANYOFL_FOLD) {
21477 sv_catpvs(sv, "{i}");
21481 inverted = flags & ANYOF_INVERT;
21483 /* If there is stuff outside the bitmap, get it */
21484 if (arg != ANYOF_ONLY_HAS_BITMAP) {
21485 if (inRANGE(OP(o), ANYOFR, ANYOFRb)) {
21486 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
21488 ANYOFRbase(o) + ANYOFRdelta(o));
21491 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
21492 (void) get_regclass_nonbitmap_data(prog, o, FALSE,
21494 &only_utf8_locale_invlist,
21495 &nonbitmap_invlist);
21497 (void) get_re_gclass_nonbitmap_data(prog, o, FALSE,
21499 &only_utf8_locale_invlist,
21500 &nonbitmap_invlist);
21504 /* The non-bitmap data may contain stuff that could fit in the
21505 * bitmap. This could come from a user-defined property being
21506 * finally resolved when this call was done; or much more likely
21507 * because there are matches that require UTF-8 to be valid, and so
21508 * aren't in the bitmap (or ANYOFR). This is teased apart later */
21509 _invlist_intersection(nonbitmap_invlist,
21511 &bitmap_range_not_in_bitmap);
21512 /* Leave just the things that don't fit into the bitmap */
21513 _invlist_subtract(nonbitmap_invlist,
21515 &nonbitmap_invlist);
21518 /* Obey this flag to add all above-the-bitmap code points */
21519 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
21520 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
21521 NUM_ANYOF_CODE_POINTS,
21525 /* Ready to start outputting. First, the initial left bracket */
21526 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
21528 /* ANYOFH by definition doesn't have anything that will fit inside the
21529 * bitmap; ANYOFR may or may not. */
21530 if ( ! inRANGE(OP(o), ANYOFH, ANYOFHr)
21531 && ( ! inRANGE(OP(o), ANYOFR, ANYOFRb)
21532 || ANYOFRbase(o) < NUM_ANYOF_CODE_POINTS))
21534 /* Then all the things that could fit in the bitmap */
21535 do_sep = put_charclass_bitmap_innards(sv,
21537 bitmap_range_not_in_bitmap,
21538 only_utf8_locale_invlist,
21542 /* Can't try inverting for a
21543 * better display if there
21544 * are things that haven't
21547 || inRANGE(OP(o), ANYOFR, ANYOFRb));
21548 SvREFCNT_dec(bitmap_range_not_in_bitmap);
21550 /* If there are user-defined properties which haven't been defined
21551 * yet, output them. If the result is not to be inverted, it is
21552 * clearest to output them in a separate [] from the bitmap range
21553 * stuff. If the result is to be complemented, we have to show
21554 * everything in one [], as the inversion applies to the whole
21555 * thing. Use {braces} to separate them from anything in the
21556 * bitmap and anything above the bitmap. */
21559 if (! do_sep) { /* If didn't output anything in the bitmap
21561 sv_catpvs(sv, "^");
21563 sv_catpvs(sv, "{");
21566 Perl_sv_catpvf(aTHX_ sv,"%s][%s", PL_colors[1],
21569 sv_catsv(sv, unresolved);
21571 sv_catpvs(sv, "}");
21573 do_sep = ! inverted;
21577 /* And, finally, add the above-the-bitmap stuff */
21578 if (nonbitmap_invlist && _invlist_len(nonbitmap_invlist)) {
21581 /* See if truncation size is overridden */
21582 const STRLEN dump_len = (PL_dump_re_max_len > 256)
21583 ? PL_dump_re_max_len
21586 /* This is output in a separate [] */
21588 Perl_sv_catpvf(aTHX_ sv,"%s][%s", PL_colors[1], PL_colors[0]);
21591 /* And, for easy of understanding, it is shown in the
21592 * uncomplemented form if possible. The one exception being if
21593 * there are unresolved items, where the inversion has to be
21594 * delayed until runtime */
21595 if (inverted && ! unresolved) {
21596 _invlist_invert(nonbitmap_invlist);
21597 _invlist_subtract(nonbitmap_invlist, PL_InBitmap, &nonbitmap_invlist);
21600 contents = invlist_contents(nonbitmap_invlist,
21601 FALSE /* output suitable for catsv */
21604 /* If the output is shorter than the permissible maximum, just do it. */
21605 if (SvCUR(contents) <= dump_len) {
21606 sv_catsv(sv, contents);
21609 const char * contents_string = SvPVX(contents);
21610 STRLEN i = dump_len;
21612 /* Otherwise, start at the permissible max and work back to the
21613 * first break possibility */
21614 while (i > 0 && contents_string[i] != ' ') {
21617 if (i == 0) { /* Fail-safe. Use the max if we couldn't
21618 find a legal break */
21622 sv_catpvn(sv, contents_string, i);
21623 sv_catpvs(sv, "...");
21626 SvREFCNT_dec_NN(contents);
21627 SvREFCNT_dec_NN(nonbitmap_invlist);
21630 /* And finally the matching, closing ']' */
21631 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
21633 if (OP(o) == ANYOFHs) {
21634 Perl_sv_catpvf(aTHX_ sv, " (Leading UTF-8 bytes=%s", _byte_dump_string((U8 *) ((struct regnode_anyofhs *) o)->string, FLAGS(o), 1));
21636 else if (inRANGE(OP(o), ANYOFH, ANYOFRb)) {
21637 U8 lowest = (OP(o) != ANYOFHr)
21639 : LOWEST_ANYOF_HRx_BYTE(FLAGS(o));
21640 U8 highest = (OP(o) == ANYOFHr)
21641 ? HIGHEST_ANYOF_HRx_BYTE(FLAGS(o))
21642 : (OP(o) == ANYOFH || OP(o) == ANYOFR)
21646 if (OP(o) != ANYOFR || ! isASCII(ANYOFRbase(o) + ANYOFRdelta(o)))
21649 Perl_sv_catpvf(aTHX_ sv, " (First UTF-8 byte=%02X", lowest);
21650 if (lowest != highest) {
21651 Perl_sv_catpvf(aTHX_ sv, "-%02X", highest);
21653 Perl_sv_catpvf(aTHX_ sv, ")");
21657 SvREFCNT_dec(unresolved);
21659 else if (k == ANYOFM) {
21660 SV * cp_list = get_ANYOFM_contents(o);
21662 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
21663 if (OP(o) == NANYOFM) {
21664 _invlist_invert(cp_list);
21667 put_charclass_bitmap_innards(sv, NULL, cp_list, NULL, NULL, 0, TRUE);
21668 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
21670 SvREFCNT_dec(cp_list);
21672 else if (k == POSIXD || k == NPOSIXD) {
21673 U8 index = FLAGS(o) * 2;
21674 if (index < C_ARRAY_LENGTH(anyofs)) {
21675 if (*anyofs[index] != '[') {
21676 sv_catpvs(sv, "[");
21678 sv_catpv(sv, anyofs[index]);
21679 if (*anyofs[index] != '[') {
21680 sv_catpvs(sv, "]");
21684 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
21687 else if (k == BOUND || k == NBOUND) {
21688 /* Must be synced with order of 'bound_type' in regcomp.h */
21689 const char * const bounds[] = {
21690 "", /* Traditional */
21696 assert(FLAGS(o) < C_ARRAY_LENGTH(bounds));
21697 sv_catpv(sv, bounds[FLAGS(o)]);
21699 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH)) {
21700 Perl_sv_catpvf(aTHX_ sv, "[%d", -(o->flags));
21702 Perl_sv_catpvf(aTHX_ sv, "..-%d", o->flags - o->next_off);
21704 Perl_sv_catpvf(aTHX_ sv, "]");
21706 else if (OP(o) == SBOL)
21707 Perl_sv_catpvf(aTHX_ sv, " /%s/", o->flags ? "\\A" : "^");
21709 /* add on the verb argument if there is one */
21710 if ( ( k == VERB || OP(o) == ACCEPT || OP(o) == OPFAIL ) && o->flags) {
21712 Perl_sv_catpvf(aTHX_ sv, ":%" SVf,
21713 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
21715 sv_catpvs(sv, ":NULL");
21718 PERL_UNUSED_CONTEXT;
21719 PERL_UNUSED_ARG(sv);
21720 PERL_UNUSED_ARG(o);
21721 PERL_UNUSED_ARG(prog);
21722 PERL_UNUSED_ARG(reginfo);
21723 PERL_UNUSED_ARG(pRExC_state);
21724 #endif /* DEBUGGING */
21730 Perl_re_intuit_string(pTHX_ REGEXP * const r)
21731 { /* Assume that RE_INTUIT is set */
21732 /* Returns an SV containing a string that must appear in the target for it
21733 * to match, or NULL if nothing is known that must match.
21735 * CAUTION: the SV can be freed during execution of the regex engine */
21737 struct regexp *const prog = ReANY(r);
21738 DECLARE_AND_GET_RE_DEBUG_FLAGS;
21740 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
21741 PERL_UNUSED_CONTEXT;
21745 if (prog->maxlen > 0) {
21746 const char * const s = SvPV_nolen_const(RX_UTF8(r)
21747 ? prog->check_utf8 : prog->check_substr);
21749 if (!PL_colorset) reginitcolors();
21750 Perl_re_printf( aTHX_
21751 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
21753 RX_UTF8(r) ? "utf8 " : "",
21754 PL_colors[5], PL_colors[0],
21757 (strlen(s) > PL_dump_re_max_len ? "..." : ""));
21761 /* use UTF8 check substring if regexp pattern itself is in UTF8 */
21762 return RX_UTF8(r) ? prog->check_utf8 : prog->check_substr;
21768 handles refcounting and freeing the perl core regexp structure. When
21769 it is necessary to actually free the structure the first thing it
21770 does is call the 'free' method of the regexp_engine associated to
21771 the regexp, allowing the handling of the void *pprivate; member
21772 first. (This routine is not overridable by extensions, which is why
21773 the extensions free is called first.)
21775 See regdupe and regdupe_internal if you change anything here.
21777 #ifndef PERL_IN_XSUB_RE
21779 Perl_pregfree(pTHX_ REGEXP *r)
21785 Perl_pregfree2(pTHX_ REGEXP *rx)
21787 struct regexp *const r = ReANY(rx);
21788 DECLARE_AND_GET_RE_DEBUG_FLAGS;
21790 PERL_ARGS_ASSERT_PREGFREE2;
21795 if (r->mother_re) {
21796 ReREFCNT_dec(r->mother_re);
21798 CALLREGFREE_PVT(rx); /* free the private data */
21799 SvREFCNT_dec(RXp_PAREN_NAMES(r));
21803 for (i = 0; i < 2; i++) {
21804 SvREFCNT_dec(r->substrs->data[i].substr);
21805 SvREFCNT_dec(r->substrs->data[i].utf8_substr);
21807 Safefree(r->substrs);
21809 RX_MATCH_COPY_FREE(rx);
21810 #ifdef PERL_ANY_COW
21811 SvREFCNT_dec(r->saved_copy);
21814 SvREFCNT_dec(r->qr_anoncv);
21815 if (r->recurse_locinput)
21816 Safefree(r->recurse_locinput);
21822 Copy ssv to dsv, both of which should of type SVt_REGEXP or SVt_PVLV,
21823 except that dsv will be created if NULL.
21825 This function is used in two main ways. First to implement
21826 $r = qr/....; $s = $$r;
21828 Secondly, it is used as a hacky workaround to the structural issue of
21830 being stored in the regexp structure which is in turn stored in
21831 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
21832 could be PL_curpm in multiple contexts, and could require multiple
21833 result sets being associated with the pattern simultaneously, such
21834 as when doing a recursive match with (??{$qr})
21836 The solution is to make a lightweight copy of the regexp structure
21837 when a qr// is returned from the code executed by (??{$qr}) this
21838 lightweight copy doesn't actually own any of its data except for
21839 the starp/end and the actual regexp structure itself.
21845 Perl_reg_temp_copy(pTHX_ REGEXP *dsv, REGEXP *ssv)
21847 struct regexp *drx;
21848 struct regexp *const srx = ReANY(ssv);
21849 const bool islv = dsv && SvTYPE(dsv) == SVt_PVLV;
21851 PERL_ARGS_ASSERT_REG_TEMP_COPY;
21854 dsv = (REGEXP*) newSV_type(SVt_REGEXP);
21856 assert(SvTYPE(dsv) == SVt_REGEXP || (SvTYPE(dsv) == SVt_PVLV));
21858 /* our only valid caller, sv_setsv_flags(), should have done
21859 * a SV_CHECK_THINKFIRST_COW_DROP() by now */
21860 assert(!SvOOK(dsv));
21861 assert(!SvIsCOW(dsv));
21862 assert(!SvROK(dsv));
21864 if (SvPVX_const(dsv)) {
21866 Safefree(SvPVX(dsv));
21871 SvOK_off((SV *)dsv);
21874 /* For PVLVs, the head (sv_any) points to an XPVLV, while
21875 * the LV's xpvlenu_rx will point to a regexp body, which
21876 * we allocate here */
21877 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
21878 assert(!SvPVX(dsv));
21879 ((XPV*)SvANY(dsv))->xpv_len_u.xpvlenu_rx = temp->sv_any;
21880 temp->sv_any = NULL;
21881 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
21882 SvREFCNT_dec_NN(temp);
21883 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
21884 ing below will not set it. */
21885 SvCUR_set(dsv, SvCUR(ssv));
21888 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
21889 sv_force_normal(sv) is called. */
21893 SvFLAGS(dsv) |= SvFLAGS(ssv) & (SVf_POK|SVp_POK|SVf_UTF8);
21894 SvPV_set(dsv, RX_WRAPPED(ssv));
21895 /* We share the same string buffer as the original regexp, on which we
21896 hold a reference count, incremented when mother_re is set below.
21897 The string pointer is copied here, being part of the regexp struct.
21899 memcpy(&(drx->xpv_cur), &(srx->xpv_cur),
21900 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
21904 const I32 npar = srx->nparens+1;
21905 Newx(drx->offs, npar, regexp_paren_pair);
21906 Copy(srx->offs, drx->offs, npar, regexp_paren_pair);
21908 if (srx->substrs) {
21910 Newx(drx->substrs, 1, struct reg_substr_data);
21911 StructCopy(srx->substrs, drx->substrs, struct reg_substr_data);
21913 for (i = 0; i < 2; i++) {
21914 SvREFCNT_inc_void(drx->substrs->data[i].substr);
21915 SvREFCNT_inc_void(drx->substrs->data[i].utf8_substr);
21918 /* check_substr and check_utf8, if non-NULL, point to either their
21919 anchored or float namesakes, and don't hold a second reference. */
21921 RX_MATCH_COPIED_off(dsv);
21922 #ifdef PERL_ANY_COW
21923 drx->saved_copy = NULL;
21925 drx->mother_re = ReREFCNT_inc(srx->mother_re ? srx->mother_re : ssv);
21926 SvREFCNT_inc_void(drx->qr_anoncv);
21927 if (srx->recurse_locinput)
21928 Newx(drx->recurse_locinput, srx->nparens + 1, char *);
21935 /* regfree_internal()
21937 Free the private data in a regexp. This is overloadable by
21938 extensions. Perl takes care of the regexp structure in pregfree(),
21939 this covers the *pprivate pointer which technically perl doesn't
21940 know about, however of course we have to handle the
21941 regexp_internal structure when no extension is in use.
21943 Note this is called before freeing anything in the regexp
21948 Perl_regfree_internal(pTHX_ REGEXP * const rx)
21950 struct regexp *const r = ReANY(rx);
21951 RXi_GET_DECL(r, ri);
21952 DECLARE_AND_GET_RE_DEBUG_FLAGS;
21954 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
21964 SV *dsv= sv_newmortal();
21965 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
21966 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), PL_dump_re_max_len);
21967 Perl_re_printf( aTHX_ "%sFreeing REx:%s %s\n",
21968 PL_colors[4], PL_colors[5], s);
21972 #ifdef RE_TRACK_PATTERN_OFFSETS
21974 Safefree(ri->u.offsets); /* 20010421 MJD */
21976 if (ri->code_blocks)
21977 S_free_codeblocks(aTHX_ ri->code_blocks);
21980 int n = ri->data->count;
21983 /* If you add a ->what type here, update the comment in regcomp.h */
21984 switch (ri->data->what[n]) {
21990 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
21993 Safefree(ri->data->data[n]);
21999 { /* Aho Corasick add-on structure for a trie node.
22000 Used in stclass optimization only */
22002 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
22004 refcount = --aho->refcount;
22007 PerlMemShared_free(aho->states);
22008 PerlMemShared_free(aho->fail);
22009 /* do this last!!!! */
22010 PerlMemShared_free(ri->data->data[n]);
22011 /* we should only ever get called once, so
22012 * assert as much, and also guard the free
22013 * which /might/ happen twice. At the least
22014 * it will make code anlyzers happy and it
22015 * doesn't cost much. - Yves */
22016 assert(ri->regstclass);
22017 if (ri->regstclass) {
22018 PerlMemShared_free(ri->regstclass);
22019 ri->regstclass = 0;
22026 /* trie structure. */
22028 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
22030 refcount = --trie->refcount;
22033 PerlMemShared_free(trie->charmap);
22034 PerlMemShared_free(trie->states);
22035 PerlMemShared_free(trie->trans);
22037 PerlMemShared_free(trie->bitmap);
22039 PerlMemShared_free(trie->jump);
22040 PerlMemShared_free(trie->wordinfo);
22041 /* do this last!!!! */
22042 PerlMemShared_free(ri->data->data[n]);
22047 Perl_croak(aTHX_ "panic: regfree data code '%c'",
22048 ri->data->what[n]);
22051 Safefree(ri->data->what);
22052 Safefree(ri->data);
22058 #define av_dup_inc(s, t) MUTABLE_AV(sv_dup_inc((const SV *)s, t))
22059 #define hv_dup_inc(s, t) MUTABLE_HV(sv_dup_inc((const SV *)s, t))
22060 #define SAVEPVN(p, n) ((p) ? savepvn(p, n) : NULL)
22063 =for apidoc re_dup_guts
22064 Duplicate a regexp.
22066 This routine is expected to clone a given regexp structure. It is only
22067 compiled under USE_ITHREADS.
22069 After all of the core data stored in struct regexp is duplicated
22070 the C<regexp_engine.dupe> method is used to copy any private data
22071 stored in the *pprivate pointer. This allows extensions to handle
22072 any duplication they need to do.
22076 See pregfree() and regfree_internal() if you change anything here.
22078 #if defined(USE_ITHREADS)
22079 #ifndef PERL_IN_XSUB_RE
22081 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
22084 const struct regexp *r = ReANY(sstr);
22085 struct regexp *ret = ReANY(dstr);
22087 PERL_ARGS_ASSERT_RE_DUP_GUTS;
22089 npar = r->nparens+1;
22090 Newx(ret->offs, npar, regexp_paren_pair);
22091 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
22093 if (ret->substrs) {
22094 /* Do it this way to avoid reading from *r after the StructCopy().
22095 That way, if any of the sv_dup_inc()s dislodge *r from the L1
22096 cache, it doesn't matter. */
22098 const bool anchored = r->check_substr
22099 ? r->check_substr == r->substrs->data[0].substr
22100 : r->check_utf8 == r->substrs->data[0].utf8_substr;
22101 Newx(ret->substrs, 1, struct reg_substr_data);
22102 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
22104 for (i = 0; i < 2; i++) {
22105 ret->substrs->data[i].substr =
22106 sv_dup_inc(ret->substrs->data[i].substr, param);
22107 ret->substrs->data[i].utf8_substr =
22108 sv_dup_inc(ret->substrs->data[i].utf8_substr, param);
22111 /* check_substr and check_utf8, if non-NULL, point to either their
22112 anchored or float namesakes, and don't hold a second reference. */
22114 if (ret->check_substr) {
22116 assert(r->check_utf8 == r->substrs->data[0].utf8_substr);
22118 ret->check_substr = ret->substrs->data[0].substr;
22119 ret->check_utf8 = ret->substrs->data[0].utf8_substr;
22121 assert(r->check_substr == r->substrs->data[1].substr);
22122 assert(r->check_utf8 == r->substrs->data[1].utf8_substr);
22124 ret->check_substr = ret->substrs->data[1].substr;
22125 ret->check_utf8 = ret->substrs->data[1].utf8_substr;
22127 } else if (ret->check_utf8) {
22129 ret->check_utf8 = ret->substrs->data[0].utf8_substr;
22131 ret->check_utf8 = ret->substrs->data[1].utf8_substr;
22136 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
22137 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
22138 if (r->recurse_locinput)
22139 Newx(ret->recurse_locinput, r->nparens + 1, char *);
22142 RXi_SET(ret, CALLREGDUPE_PVT(dstr, param));
22144 if (RX_MATCH_COPIED(dstr))
22145 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
22147 ret->subbeg = NULL;
22148 #ifdef PERL_ANY_COW
22149 ret->saved_copy = NULL;
22152 /* Whether mother_re be set or no, we need to copy the string. We
22153 cannot refrain from copying it when the storage points directly to
22154 our mother regexp, because that's
22155 1: a buffer in a different thread
22156 2: something we no longer hold a reference on
22157 so we need to copy it locally. */
22158 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED_const(sstr), SvCUR(sstr)+1);
22159 /* set malloced length to a non-zero value so it will be freed
22160 * (otherwise in combination with SVf_FAKE it looks like an alien
22161 * buffer). It doesn't have to be the actual malloced size, since it
22162 * should never be grown */
22163 SvLEN_set(dstr, SvCUR(sstr)+1);
22164 ret->mother_re = NULL;
22166 #endif /* PERL_IN_XSUB_RE */
22171 This is the internal complement to regdupe() which is used to copy
22172 the structure pointed to by the *pprivate pointer in the regexp.
22173 This is the core version of the extension overridable cloning hook.
22174 The regexp structure being duplicated will be copied by perl prior
22175 to this and will be provided as the regexp *r argument, however
22176 with the /old/ structures pprivate pointer value. Thus this routine
22177 may override any copying normally done by perl.
22179 It returns a pointer to the new regexp_internal structure.
22183 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
22185 struct regexp *const r = ReANY(rx);
22186 regexp_internal *reti;
22188 RXi_GET_DECL(r, ri);
22190 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
22194 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
22195 char, regexp_internal);
22196 Copy(ri->program, reti->program, len+1, regnode);
22199 if (ri->code_blocks) {
22201 Newx(reti->code_blocks, 1, struct reg_code_blocks);
22202 Newx(reti->code_blocks->cb, ri->code_blocks->count,
22203 struct reg_code_block);
22204 Copy(ri->code_blocks->cb, reti->code_blocks->cb,
22205 ri->code_blocks->count, struct reg_code_block);
22206 for (n = 0; n < ri->code_blocks->count; n++)
22207 reti->code_blocks->cb[n].src_regex = (REGEXP*)
22208 sv_dup_inc((SV*)(ri->code_blocks->cb[n].src_regex), param);
22209 reti->code_blocks->count = ri->code_blocks->count;
22210 reti->code_blocks->refcnt = 1;
22213 reti->code_blocks = NULL;
22215 reti->regstclass = NULL;
22218 struct reg_data *d;
22219 const int count = ri->data->count;
22222 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
22223 char, struct reg_data);
22224 Newx(d->what, count, U8);
22227 for (i = 0; i < count; i++) {
22228 d->what[i] = ri->data->what[i];
22229 switch (d->what[i]) {
22230 /* see also regcomp.h and regfree_internal() */
22231 case 'a': /* actually an AV, but the dup function is identical.
22232 values seem to be "plain sv's" generally. */
22233 case 'r': /* a compiled regex (but still just another SV) */
22234 case 's': /* an RV (currently only used for an RV to an AV by the ANYOF code)
22235 this use case should go away, the code could have used
22236 'a' instead - see S_set_ANYOF_arg() for array contents. */
22237 case 'S': /* actually an SV, but the dup function is identical. */
22238 case 'u': /* actually an HV, but the dup function is identical.
22239 values are "plain sv's" */
22240 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
22243 /* Synthetic Start Class - "Fake" charclass we generate to optimize
22244 * patterns which could start with several different things. Pre-TRIE
22245 * this was more important than it is now, however this still helps
22246 * in some places, for instance /x?a+/ might produce a SSC equivalent
22247 * to [xa]. This is used by Perl_re_intuit_start() and S_find_byclass()
22250 /* This is cheating. */
22251 Newx(d->data[i], 1, regnode_ssc);
22252 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
22253 reti->regstclass = (regnode*)d->data[i];
22256 /* AHO-CORASICK fail table */
22257 /* Trie stclasses are readonly and can thus be shared
22258 * without duplication. We free the stclass in pregfree
22259 * when the corresponding reg_ac_data struct is freed.
22261 reti->regstclass= ri->regstclass;
22264 /* TRIE transition table */
22266 ((reg_trie_data*)ri->data->data[i])->refcount++;
22269 case 'l': /* (?{...}) or (??{ ... }) code (cb->block) */
22270 case 'L': /* same when RExC_pm_flags & PMf_HAS_CV and code
22271 is not from another regexp */
22272 d->data[i] = ri->data->data[i];
22275 Perl_croak(aTHX_ "panic: re_dup_guts unknown data code '%c'",
22276 ri->data->what[i]);
22285 reti->name_list_idx = ri->name_list_idx;
22287 #ifdef RE_TRACK_PATTERN_OFFSETS
22288 if (ri->u.offsets) {
22289 Newx(reti->u.offsets, 2*len+1, U32);
22290 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
22293 SetProgLen(reti, len);
22296 return (void*)reti;
22299 #endif /* USE_ITHREADS */
22301 #ifndef PERL_IN_XSUB_RE
22304 - regnext - dig the "next" pointer out of a node
22307 Perl_regnext(pTHX_ regnode *p)
22314 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
22315 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
22316 (int)OP(p), (int)REGNODE_MAX);
22319 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
22329 S_re_croak(pTHX_ bool utf8, const char* pat,...)
22332 STRLEN len = strlen(pat);
22335 const char *message;
22337 PERL_ARGS_ASSERT_RE_CROAK;
22341 Copy(pat, buf, len , char);
22343 buf[len + 1] = '\0';
22344 va_start(args, pat);
22345 msv = vmess(buf, &args);
22347 message = SvPV_const(msv, len);
22350 Copy(message, buf, len , char);
22351 /* len-1 to avoid \n */
22352 Perl_croak(aTHX_ "%" UTF8f, UTF8fARG(utf8, len-1, buf));
22355 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
22357 #ifndef PERL_IN_XSUB_RE
22359 Perl_save_re_context(pTHX)
22364 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
22367 const REGEXP * const rx = PM_GETRE(PL_curpm);
22369 nparens = RX_NPARENS(rx);
22372 /* RT #124109. This is a complete hack; in the SWASHNEW case we know
22373 * that PL_curpm will be null, but that utf8.pm and the modules it
22374 * loads will only use $1..$3.
22375 * The t/porting/re_context.t test file checks this assumption.
22380 for (i = 1; i <= nparens; i++) {
22381 char digits[TYPE_CHARS(long)];
22382 const STRLEN len = my_snprintf(digits, sizeof(digits),
22384 GV *const *const gvp
22385 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
22388 GV * const gv = *gvp;
22389 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
22399 S_put_code_point(pTHX_ SV *sv, UV c)
22401 PERL_ARGS_ASSERT_PUT_CODE_POINT;
22404 Perl_sv_catpvf(aTHX_ sv, "\\x{%04" UVXf "}", c);
22406 else if (isPRINT(c)) {
22407 const char string = (char) c;
22409 /* We use {phrase} as metanotation in the class, so also escape literal
22411 if (isBACKSLASHED_PUNCT(c) || c == '{' || c == '}')
22412 sv_catpvs(sv, "\\");
22413 sv_catpvn(sv, &string, 1);
22415 else if (isMNEMONIC_CNTRL(c)) {
22416 Perl_sv_catpvf(aTHX_ sv, "%s", cntrl_to_mnemonic((U8) c));
22419 Perl_sv_catpvf(aTHX_ sv, "\\x%02X", (U8) c);
22424 S_put_range(pTHX_ SV *sv, UV start, const UV end, const bool allow_literals)
22426 /* Appends to 'sv' a displayable version of the range of code points from
22427 * 'start' to 'end'. Mnemonics (like '\r') are used for the few controls
22428 * that have them, when they occur at the beginning or end of the range.
22429 * It uses hex to output the remaining code points, unless 'allow_literals'
22430 * is true, in which case the printable ASCII ones are output as-is (though
22431 * some of these will be escaped by put_code_point()).
22433 * NOTE: This is designed only for printing ranges of code points that fit
22434 * inside an ANYOF bitmap. Higher code points are simply suppressed
22437 const unsigned int min_range_count = 3;
22439 assert(start <= end);
22441 PERL_ARGS_ASSERT_PUT_RANGE;
22443 while (start <= end) {
22445 const char * format;
22447 if ( end - start < min_range_count
22448 && (end - start <= 2 || (isPRINT_A(start) && isPRINT_A(end))))
22450 /* Output a range of 1 or 2 chars individually, or longer ranges
22451 * when printable */
22452 for (; start <= end; start++) {
22453 put_code_point(sv, start);
22458 /* If permitted by the input options, and there is a possibility that
22459 * this range contains a printable literal, look to see if there is
22461 if (allow_literals && start <= MAX_PRINT_A) {
22463 /* If the character at the beginning of the range isn't an ASCII
22464 * printable, effectively split the range into two parts:
22465 * 1) the portion before the first such printable,
22467 * and output them separately. */
22468 if (! isPRINT_A(start)) {
22469 UV temp_end = start + 1;
22471 /* There is no point looking beyond the final possible
22472 * printable, in MAX_PRINT_A */
22473 UV max = MIN(end, MAX_PRINT_A);
22475 while (temp_end <= max && ! isPRINT_A(temp_end)) {
22479 /* Here, temp_end points to one beyond the first printable if
22480 * found, or to one beyond 'max' if not. If none found, make
22481 * sure that we use the entire range */
22482 if (temp_end > MAX_PRINT_A) {
22483 temp_end = end + 1;
22486 /* Output the first part of the split range: the part that
22487 * doesn't have printables, with the parameter set to not look
22488 * for literals (otherwise we would infinitely recurse) */
22489 put_range(sv, start, temp_end - 1, FALSE);
22491 /* The 2nd part of the range (if any) starts here. */
22494 /* We do a continue, instead of dropping down, because even if
22495 * the 2nd part is non-empty, it could be so short that we want
22496 * to output it as individual characters, as tested for at the
22497 * top of this loop. */
22501 /* Here, 'start' is a printable ASCII. If it is an alphanumeric,
22502 * output a sub-range of just the digits or letters, then process
22503 * the remaining portion as usual. */
22504 if (isALPHANUMERIC_A(start)) {
22505 UV mask = (isDIGIT_A(start))
22510 UV temp_end = start + 1;
22512 /* Find the end of the sub-range that includes just the
22513 * characters in the same class as the first character in it */
22514 while (temp_end <= end && _generic_isCC_A(temp_end, mask)) {
22519 /* For short ranges, don't duplicate the code above to output
22520 * them; just call recursively */
22521 if (temp_end - start < min_range_count) {
22522 put_range(sv, start, temp_end, FALSE);
22524 else { /* Output as a range */
22525 put_code_point(sv, start);
22526 sv_catpvs(sv, "-");
22527 put_code_point(sv, temp_end);
22529 start = temp_end + 1;
22533 /* We output any other printables as individual characters */
22534 if (isPUNCT_A(start) || isSPACE_A(start)) {
22535 while (start <= end && (isPUNCT_A(start)
22536 || isSPACE_A(start)))
22538 put_code_point(sv, start);
22543 } /* End of looking for literals */
22545 /* Here is not to output as a literal. Some control characters have
22546 * mnemonic names. Split off any of those at the beginning and end of
22547 * the range to print mnemonically. It isn't possible for many of
22548 * these to be in a row, so this won't overwhelm with output */
22550 && (isMNEMONIC_CNTRL(start) || isMNEMONIC_CNTRL(end)))
22552 while (isMNEMONIC_CNTRL(start) && start <= end) {
22553 put_code_point(sv, start);
22557 /* If this didn't take care of the whole range ... */
22558 if (start <= end) {
22560 /* Look backwards from the end to find the final non-mnemonic
22563 while (isMNEMONIC_CNTRL(temp_end)) {
22567 /* And separately output the interior range that doesn't start
22568 * or end with mnemonics */
22569 put_range(sv, start, temp_end, FALSE);
22571 /* Then output the mnemonic trailing controls */
22572 start = temp_end + 1;
22573 while (start <= end) {
22574 put_code_point(sv, start);
22581 /* As a final resort, output the range or subrange as hex. */
22583 if (start >= NUM_ANYOF_CODE_POINTS) {
22586 else { /* Have to split range at the bitmap boundary */
22587 this_end = (end < NUM_ANYOF_CODE_POINTS)
22589 : NUM_ANYOF_CODE_POINTS - 1;
22591 #if NUM_ANYOF_CODE_POINTS > 256
22592 format = (this_end < 256)
22593 ? "\\x%02" UVXf "-\\x%02" UVXf
22594 : "\\x{%04" UVXf "}-\\x{%04" UVXf "}";
22596 format = "\\x%02" UVXf "-\\x%02" UVXf;
22598 GCC_DIAG_IGNORE_STMT(-Wformat-nonliteral);
22599 Perl_sv_catpvf(aTHX_ sv, format, start, this_end);
22600 GCC_DIAG_RESTORE_STMT;
22606 S_put_charclass_bitmap_innards_invlist(pTHX_ SV *sv, SV* invlist)
22608 /* Concatenate onto the PV in 'sv' a displayable form of the inversion list
22612 bool allow_literals = TRUE;
22614 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_INVLIST;
22616 /* Generally, it is more readable if printable characters are output as
22617 * literals, but if a range (nearly) spans all of them, it's best to output
22618 * it as a single range. This code will use a single range if all but 2
22619 * ASCII printables are in it */
22620 invlist_iterinit(invlist);
22621 while (invlist_iternext(invlist, &start, &end)) {
22623 /* If the range starts beyond the final printable, it doesn't have any
22625 if (start > MAX_PRINT_A) {
22629 /* In both ASCII and EBCDIC, a SPACE is the lowest printable. To span
22630 * all but two, the range must start and end no later than 2 from
22632 if (start < ' ' + 2 && end > MAX_PRINT_A - 2) {
22633 if (end > MAX_PRINT_A) {
22639 if (end - start >= MAX_PRINT_A - ' ' - 2) {
22640 allow_literals = FALSE;
22645 invlist_iterfinish(invlist);
22647 /* Here we have figured things out. Output each range */
22648 invlist_iterinit(invlist);
22649 while (invlist_iternext(invlist, &start, &end)) {
22650 if (start >= NUM_ANYOF_CODE_POINTS) {
22653 put_range(sv, start, end, allow_literals);
22655 invlist_iterfinish(invlist);
22661 S_put_charclass_bitmap_innards_common(pTHX_
22662 SV* invlist, /* The bitmap */
22663 SV* posixes, /* Under /l, things like [:word:], \S */
22664 SV* only_utf8, /* Under /d, matches iff the target is UTF-8 */
22665 SV* not_utf8, /* /d, matches iff the target isn't UTF-8 */
22666 SV* only_utf8_locale, /* Under /l, matches if the locale is UTF-8 */
22667 const bool invert /* Is the result to be inverted? */
22670 /* Create and return an SV containing a displayable version of the bitmap
22671 * and associated information determined by the input parameters. If the
22672 * output would have been only the inversion indicator '^', NULL is instead
22677 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_COMMON;
22680 output = newSVpvs("^");
22683 output = newSVpvs("");
22686 /* First, the code points in the bitmap that are unconditionally there */
22687 put_charclass_bitmap_innards_invlist(output, invlist);
22689 /* Traditionally, these have been placed after the main code points */
22691 sv_catsv(output, posixes);
22694 if (only_utf8 && _invlist_len(only_utf8)) {
22695 Perl_sv_catpvf(aTHX_ output, "%s{utf8}%s", PL_colors[1], PL_colors[0]);
22696 put_charclass_bitmap_innards_invlist(output, only_utf8);
22699 if (not_utf8 && _invlist_len(not_utf8)) {
22700 Perl_sv_catpvf(aTHX_ output, "%s{not utf8}%s", PL_colors[1], PL_colors[0]);
22701 put_charclass_bitmap_innards_invlist(output, not_utf8);
22704 if (only_utf8_locale && _invlist_len(only_utf8_locale)) {
22705 Perl_sv_catpvf(aTHX_ output, "%s{utf8 locale}%s", PL_colors[1], PL_colors[0]);
22706 put_charclass_bitmap_innards_invlist(output, only_utf8_locale);
22708 /* This is the only list in this routine that can legally contain code
22709 * points outside the bitmap range. The call just above to
22710 * 'put_charclass_bitmap_innards_invlist' will simply suppress them, so
22711 * output them here. There's about a half-dozen possible, and none in
22712 * contiguous ranges longer than 2 */
22713 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
22715 SV* above_bitmap = NULL;
22717 _invlist_subtract(only_utf8_locale, PL_InBitmap, &above_bitmap);
22719 invlist_iterinit(above_bitmap);
22720 while (invlist_iternext(above_bitmap, &start, &end)) {
22723 for (i = start; i <= end; i++) {
22724 put_code_point(output, i);
22727 invlist_iterfinish(above_bitmap);
22728 SvREFCNT_dec_NN(above_bitmap);
22732 if (invert && SvCUR(output) == 1) {
22740 S_put_charclass_bitmap_innards(pTHX_ SV *sv,
22742 SV *nonbitmap_invlist,
22743 SV *only_utf8_locale_invlist,
22744 const regnode * const node,
22746 const bool force_as_is_display)
22748 /* Appends to 'sv' a displayable version of the innards of the bracketed
22749 * character class defined by the other arguments:
22750 * 'bitmap' points to the bitmap, or NULL if to ignore that.
22751 * 'nonbitmap_invlist' is an inversion list of the code points that are in
22752 * the bitmap range, but for some reason aren't in the bitmap; NULL if
22753 * none. The reasons for this could be that they require some
22754 * condition such as the target string being or not being in UTF-8
22755 * (under /d), or because they came from a user-defined property that
22756 * was not resolved at the time of the regex compilation (under /u)
22757 * 'only_utf8_locale_invlist' is an inversion list of the code points that
22758 * are valid only if the runtime locale is a UTF-8 one; NULL if none
22759 * 'node' is the regex pattern ANYOF node. It is needed only when the
22760 * above two parameters are not null, and is passed so that this
22761 * routine can tease apart the various reasons for them.
22762 * 'flags' is the flags field of 'node'
22763 * 'force_as_is_display' is TRUE if this routine should definitely NOT try
22764 * to invert things to see if that leads to a cleaner display. If
22765 * FALSE, this routine is free to use its judgment about doing this.
22767 * It returns TRUE if there was actually something output. (It may be that
22768 * the bitmap, etc is empty.)
22770 * When called for outputting the bitmap of a non-ANYOF node, just pass the
22771 * bitmap, with the succeeding parameters set to NULL, and the final one to
22775 /* In general, it tries to display the 'cleanest' representation of the
22776 * innards, choosing whether to display them inverted or not, regardless of
22777 * whether the class itself is to be inverted. However, there are some
22778 * cases where it can't try inverting, as what actually matches isn't known
22779 * until runtime, and hence the inversion isn't either. */
22781 bool inverting_allowed = ! force_as_is_display;
22784 STRLEN orig_sv_cur = SvCUR(sv);
22786 SV* invlist; /* Inversion list we accumulate of code points that
22787 are unconditionally matched */
22788 SV* only_utf8 = NULL; /* Under /d, list of matches iff the target is
22790 SV* not_utf8 = NULL; /* /d, list of matches iff the target isn't UTF-8
22792 SV* posixes = NULL; /* Under /l, string of things like [:word:], \D */
22793 SV* only_utf8_locale = NULL; /* Under /l, list of matches if the locale
22796 SV* as_is_display; /* The output string when we take the inputs
22798 SV* inverted_display; /* The output string when we invert the inputs */
22800 bool invert = cBOOL(flags & ANYOF_INVERT); /* Is the input to be inverted
22802 /* We are biased in favor of displaying things without them being inverted,
22803 * as that is generally easier to understand */
22804 const int bias = 5;
22806 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS;
22808 /* Start off with whatever code points are passed in. (We clone, so we
22809 * don't change the caller's list) */
22810 if (nonbitmap_invlist) {
22811 assert(invlist_highest(nonbitmap_invlist) < NUM_ANYOF_CODE_POINTS);
22812 invlist = invlist_clone(nonbitmap_invlist, NULL);
22814 else { /* Worst case size is every other code point is matched */
22815 invlist = _new_invlist(NUM_ANYOF_CODE_POINTS / 2);
22819 if (OP(node) == ANYOFD) {
22821 /* This flag indicates that the code points below 0x100 in the
22822 * nonbitmap list are precisely the ones that match only when the
22823 * target is UTF-8 (they should all be non-ASCII). */
22824 if (flags & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)
22826 _invlist_intersection(invlist, PL_UpperLatin1, &only_utf8);
22827 _invlist_subtract(invlist, only_utf8, &invlist);
22830 /* And this flag for matching all non-ASCII 0xFF and below */
22831 if (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)
22833 not_utf8 = invlist_clone(PL_UpperLatin1, NULL);
22836 else if (OP(node) == ANYOFL || OP(node) == ANYOFPOSIXL) {
22838 /* If either of these flags are set, what matches isn't
22839 * determinable except during execution, so don't know enough here
22841 if (flags & (ANYOFL_FOLD|ANYOF_MATCHES_POSIXL)) {
22842 inverting_allowed = FALSE;
22845 /* What the posix classes match also varies at runtime, so these
22846 * will be output symbolically. */
22847 if (ANYOF_POSIXL_TEST_ANY_SET(node)) {
22850 posixes = newSVpvs("");
22851 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
22852 if (ANYOF_POSIXL_TEST(node, i)) {
22853 sv_catpv(posixes, anyofs[i]);
22860 /* Accumulate the bit map into the unconditional match list */
22862 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
22863 if (BITMAP_TEST(bitmap, i)) {
22866 i < NUM_ANYOF_CODE_POINTS && BITMAP_TEST(bitmap, i);
22869 invlist = _add_range_to_invlist(invlist, start, i-1);
22874 /* Make sure that the conditional match lists don't have anything in them
22875 * that match unconditionally; otherwise the output is quite confusing.
22876 * This could happen if the code that populates these misses some
22879 _invlist_subtract(only_utf8, invlist, &only_utf8);
22882 _invlist_subtract(not_utf8, invlist, ¬_utf8);
22885 if (only_utf8_locale_invlist) {
22887 /* Since this list is passed in, we have to make a copy before
22889 only_utf8_locale = invlist_clone(only_utf8_locale_invlist, NULL);
22891 _invlist_subtract(only_utf8_locale, invlist, &only_utf8_locale);
22893 /* And, it can get really weird for us to try outputting an inverted
22894 * form of this list when it has things above the bitmap, so don't even
22896 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
22897 inverting_allowed = FALSE;
22901 /* Calculate what the output would be if we take the input as-is */
22902 as_is_display = put_charclass_bitmap_innards_common(invlist,
22909 /* If have to take the output as-is, just do that */
22910 if (! inverting_allowed) {
22911 if (as_is_display) {
22912 sv_catsv(sv, as_is_display);
22913 SvREFCNT_dec_NN(as_is_display);
22916 else { /* But otherwise, create the output again on the inverted input, and
22917 use whichever version is shorter */
22919 int inverted_bias, as_is_bias;
22921 /* We will apply our bias to whichever of the results doesn't have
22931 inverted_bias = bias;
22934 /* Now invert each of the lists that contribute to the output,
22935 * excluding from the result things outside the possible range */
22937 /* For the unconditional inversion list, we have to add in all the
22938 * conditional code points, so that when inverted, they will be gone
22940 _invlist_union(only_utf8, invlist, &invlist);
22941 _invlist_union(not_utf8, invlist, &invlist);
22942 _invlist_union(only_utf8_locale, invlist, &invlist);
22943 _invlist_invert(invlist);
22944 _invlist_intersection(invlist, PL_InBitmap, &invlist);
22947 _invlist_invert(only_utf8);
22948 _invlist_intersection(only_utf8, PL_UpperLatin1, &only_utf8);
22950 else if (not_utf8) {
22952 /* If a code point matches iff the target string is not in UTF-8,
22953 * then complementing the result has it not match iff not in UTF-8,
22954 * which is the same thing as matching iff it is UTF-8. */
22955 only_utf8 = not_utf8;
22959 if (only_utf8_locale) {
22960 _invlist_invert(only_utf8_locale);
22961 _invlist_intersection(only_utf8_locale,
22963 &only_utf8_locale);
22966 inverted_display = put_charclass_bitmap_innards_common(
22971 only_utf8_locale, invert);
22973 /* Use the shortest representation, taking into account our bias
22974 * against showing it inverted */
22975 if ( inverted_display
22976 && ( ! as_is_display
22977 || ( SvCUR(inverted_display) + inverted_bias
22978 < SvCUR(as_is_display) + as_is_bias)))
22980 sv_catsv(sv, inverted_display);
22982 else if (as_is_display) {
22983 sv_catsv(sv, as_is_display);
22986 SvREFCNT_dec(as_is_display);
22987 SvREFCNT_dec(inverted_display);
22990 SvREFCNT_dec_NN(invlist);
22991 SvREFCNT_dec(only_utf8);
22992 SvREFCNT_dec(not_utf8);
22993 SvREFCNT_dec(posixes);
22994 SvREFCNT_dec(only_utf8_locale);
22996 return SvCUR(sv) > orig_sv_cur;
22999 #define CLEAR_OPTSTART \
23000 if (optstart) STMT_START { \
23001 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ \
23002 " (%" IVdf " nodes)\n", (IV)(node - optstart))); \
23006 #define DUMPUNTIL(b,e) \
23008 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
23010 STATIC const regnode *
23011 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
23012 const regnode *last, const regnode *plast,
23013 SV* sv, I32 indent, U32 depth)
23015 U8 op = PSEUDO; /* Arbitrary non-END op. */
23016 const regnode *next;
23017 const regnode *optstart= NULL;
23019 RXi_GET_DECL(r, ri);
23020 DECLARE_AND_GET_RE_DEBUG_FLAGS;
23022 PERL_ARGS_ASSERT_DUMPUNTIL;
23024 #ifdef DEBUG_DUMPUNTIL
23025 Perl_re_printf( aTHX_ "--- %d : %d - %d - %d\n", indent, node-start,
23026 last ? last-start : 0, plast ? plast-start : 0);
23029 if (plast && plast < last)
23032 while (PL_regkind[op] != END && (!last || node < last)) {
23034 /* While that wasn't END last time... */
23037 if (op == CLOSE || op == SRCLOSE || op == WHILEM)
23039 next = regnext((regnode *)node);
23042 if (OP(node) == OPTIMIZED) {
23043 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
23050 regprop(r, sv, node, NULL, NULL);
23051 Perl_re_printf( aTHX_ "%4" IVdf ":%*s%s", (IV)(node - start),
23052 (int)(2*indent + 1), "", SvPVX_const(sv));
23054 if (OP(node) != OPTIMIZED) {
23055 if (next == NULL) /* Next ptr. */
23056 Perl_re_printf( aTHX_ " (0)");
23057 else if (PL_regkind[(U8)op] == BRANCH
23058 && PL_regkind[OP(next)] != BRANCH )
23059 Perl_re_printf( aTHX_ " (FAIL)");
23061 Perl_re_printf( aTHX_ " (%" IVdf ")", (IV)(next - start));
23062 Perl_re_printf( aTHX_ "\n");
23066 if (PL_regkind[(U8)op] == BRANCHJ) {
23069 const regnode *nnode = (OP(next) == LONGJMP
23070 ? regnext((regnode *)next)
23072 if (last && nnode > last)
23074 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
23077 else if (PL_regkind[(U8)op] == BRANCH) {
23079 DUMPUNTIL(NEXTOPER(node), next);
23081 else if ( PL_regkind[(U8)op] == TRIE ) {
23082 const regnode *this_trie = node;
23083 const char op = OP(node);
23084 const U32 n = ARG(node);
23085 const reg_ac_data * const ac = op>=AHOCORASICK ?
23086 (reg_ac_data *)ri->data->data[n] :
23088 const reg_trie_data * const trie =
23089 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
23091 AV *const trie_words
23092 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
23094 const regnode *nextbranch= NULL;
23097 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
23098 SV ** const elem_ptr = av_fetch(trie_words, word_idx, 0);
23100 Perl_re_indentf( aTHX_ "%s ",
23103 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
23104 SvCUR(*elem_ptr), PL_dump_re_max_len,
23105 PL_colors[0], PL_colors[1],
23107 ? PERL_PV_ESCAPE_UNI
23109 | PERL_PV_PRETTY_ELLIPSES
23110 | PERL_PV_PRETTY_LTGT
23115 U16 dist= trie->jump[word_idx+1];
23116 Perl_re_printf( aTHX_ "(%" UVuf ")\n",
23117 (UV)((dist ? this_trie + dist : next) - start));
23120 nextbranch= this_trie + trie->jump[0];
23121 DUMPUNTIL(this_trie + dist, nextbranch);
23123 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
23124 nextbranch= regnext((regnode *)nextbranch);
23126 Perl_re_printf( aTHX_ "\n");
23129 if (last && next > last)
23134 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
23135 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
23136 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
23138 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
23140 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
23142 else if ( op == PLUS || op == STAR) {
23143 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
23145 else if (PL_regkind[(U8)op] == EXACT || op == ANYOFHs) {
23146 /* Literal string, where present. */
23147 node += NODE_SZ_STR(node) - 1;
23148 node = NEXTOPER(node);
23151 node = NEXTOPER(node);
23152 node += regarglen[(U8)op];
23154 if (op == CURLYX || op == OPEN || op == SROPEN)
23158 #ifdef DEBUG_DUMPUNTIL
23159 Perl_re_printf( aTHX_ "--- %d\n", (int)indent);
23164 #endif /* DEBUGGING */
23166 #ifndef PERL_IN_XSUB_RE
23168 # include "uni_keywords.h"
23171 Perl_init_uniprops(pTHX)
23175 char * dump_len_string;
23177 dump_len_string = PerlEnv_getenv("PERL_DUMP_RE_MAX_LEN");
23178 if ( ! dump_len_string
23179 || ! grok_atoUV(dump_len_string, (UV *)&PL_dump_re_max_len, NULL))
23181 PL_dump_re_max_len = 60; /* A reasonable default */
23185 PL_user_def_props = newHV();
23187 # ifdef USE_ITHREADS
23189 HvSHAREKEYS_off(PL_user_def_props);
23190 PL_user_def_props_aTHX = aTHX;
23194 /* Set up the inversion list interpreter-level variables */
23196 PL_XPosix_ptrs[_CC_ASCII] = _new_invlist_C_array(uni_prop_ptrs[UNI_ASCII]);
23197 PL_XPosix_ptrs[_CC_ALPHANUMERIC] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXALNUM]);
23198 PL_XPosix_ptrs[_CC_ALPHA] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXALPHA]);
23199 PL_XPosix_ptrs[_CC_BLANK] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXBLANK]);
23200 PL_XPosix_ptrs[_CC_CASED] = _new_invlist_C_array(uni_prop_ptrs[UNI_CASED]);
23201 PL_XPosix_ptrs[_CC_CNTRL] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXCNTRL]);
23202 PL_XPosix_ptrs[_CC_DIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXDIGIT]);
23203 PL_XPosix_ptrs[_CC_GRAPH] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXGRAPH]);
23204 PL_XPosix_ptrs[_CC_LOWER] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXLOWER]);
23205 PL_XPosix_ptrs[_CC_PRINT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXPRINT]);
23206 PL_XPosix_ptrs[_CC_PUNCT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXPUNCT]);
23207 PL_XPosix_ptrs[_CC_SPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXSPACE]);
23208 PL_XPosix_ptrs[_CC_UPPER] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXUPPER]);
23209 PL_XPosix_ptrs[_CC_VERTSPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_VERTSPACE]);
23210 PL_XPosix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXWORD]);
23211 PL_XPosix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXXDIGIT]);
23213 PL_Posix_ptrs[_CC_ASCII] = _new_invlist_C_array(uni_prop_ptrs[UNI_ASCII]);
23214 PL_Posix_ptrs[_CC_ALPHANUMERIC] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXALNUM]);
23215 PL_Posix_ptrs[_CC_ALPHA] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXALPHA]);
23216 PL_Posix_ptrs[_CC_BLANK] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXBLANK]);
23217 PL_Posix_ptrs[_CC_CASED] = PL_Posix_ptrs[_CC_ALPHA];
23218 PL_Posix_ptrs[_CC_CNTRL] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXCNTRL]);
23219 PL_Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXDIGIT]);
23220 PL_Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXGRAPH]);
23221 PL_Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXLOWER]);
23222 PL_Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXPRINT]);
23223 PL_Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXPUNCT]);
23224 PL_Posix_ptrs[_CC_SPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXSPACE]);
23225 PL_Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXUPPER]);
23226 PL_Posix_ptrs[_CC_VERTSPACE] = NULL;
23227 PL_Posix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXWORD]);
23228 PL_Posix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXXDIGIT]);
23230 PL_GCB_invlist = _new_invlist_C_array(_Perl_GCB_invlist);
23231 PL_SB_invlist = _new_invlist_C_array(_Perl_SB_invlist);
23232 PL_WB_invlist = _new_invlist_C_array(_Perl_WB_invlist);
23233 PL_LB_invlist = _new_invlist_C_array(_Perl_LB_invlist);
23234 PL_SCX_invlist = _new_invlist_C_array(_Perl_SCX_invlist);
23236 PL_InBitmap = _new_invlist_C_array(InBitmap_invlist);
23237 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
23238 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
23239 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
23241 PL_Assigned_invlist = _new_invlist_C_array(uni_prop_ptrs[UNI_ASSIGNED]);
23243 PL_utf8_perl_idstart = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_IDSTART]);
23244 PL_utf8_perl_idcont = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_IDCONT]);
23246 PL_utf8_charname_begin = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_CHARNAME_BEGIN]);
23247 PL_utf8_charname_continue = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_CHARNAME_CONTINUE]);
23249 PL_in_some_fold = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_ANY_FOLDS]);
23250 PL_HasMultiCharFold = _new_invlist_C_array(uni_prop_ptrs[
23251 UNI__PERL_FOLDS_TO_MULTI_CHAR]);
23252 PL_InMultiCharFold = _new_invlist_C_array(uni_prop_ptrs[
23253 UNI__PERL_IS_IN_MULTI_CHAR_FOLD]);
23254 PL_utf8_toupper = _new_invlist_C_array(Uppercase_Mapping_invlist);
23255 PL_utf8_tolower = _new_invlist_C_array(Lowercase_Mapping_invlist);
23256 PL_utf8_totitle = _new_invlist_C_array(Titlecase_Mapping_invlist);
23257 PL_utf8_tofold = _new_invlist_C_array(Case_Folding_invlist);
23258 PL_utf8_tosimplefold = _new_invlist_C_array(Simple_Case_Folding_invlist);
23259 PL_utf8_foldclosures = _new_invlist_C_array(_Perl_IVCF_invlist);
23260 PL_utf8_mark = _new_invlist_C_array(uni_prop_ptrs[UNI_M]);
23261 PL_CCC_non0_non230 = _new_invlist_C_array(_Perl_CCC_non0_non230_invlist);
23262 PL_Private_Use = _new_invlist_C_array(uni_prop_ptrs[UNI_CO]);
23265 /* The below are used only by deprecated functions. They could be removed */
23266 PL_utf8_xidcont = _new_invlist_C_array(uni_prop_ptrs[UNI_XIDC]);
23267 PL_utf8_idcont = _new_invlist_C_array(uni_prop_ptrs[UNI_IDC]);
23268 PL_utf8_xidstart = _new_invlist_C_array(uni_prop_ptrs[UNI_XIDS]);
23272 /* These four functions are compiled only in regcomp.c, where they have access
23273 * to the data they return. They are a way for re_comp.c to get access to that
23274 * data without having to compile the whole data structures. */
23277 Perl_do_uniprop_match(const char * const key, const U16 key_len)
23279 PERL_ARGS_ASSERT_DO_UNIPROP_MATCH;
23281 return match_uniprop((U8 *) key, key_len);
23285 Perl_get_prop_definition(pTHX_ const int table_index)
23287 PERL_ARGS_ASSERT_GET_PROP_DEFINITION;
23289 /* Create and return the inversion list */
23290 return _new_invlist_C_array(uni_prop_ptrs[table_index]);
23293 const char * const *
23294 Perl_get_prop_values(const int table_index)
23296 PERL_ARGS_ASSERT_GET_PROP_VALUES;
23298 return UNI_prop_value_ptrs[table_index];
23302 Perl_get_deprecated_property_msg(const Size_t warning_offset)
23304 PERL_ARGS_ASSERT_GET_DEPRECATED_PROPERTY_MSG;
23306 return deprecated_property_msgs[warning_offset];
23311 This code was mainly added for backcompat to give a warning for non-portable
23312 code points in user-defined properties. But experiments showed that the
23313 warning in earlier perls were only omitted on overflow, which should be an
23314 error, so there really isnt a backcompat issue, and actually adding the
23315 warning when none was present before might cause breakage, for little gain. So
23316 khw left this code in, but not enabled. Tests were never added.
23319 Ei |const char *|get_extended_utf8_msg|const UV cp
23321 PERL_STATIC_INLINE const char *
23322 S_get_extended_utf8_msg(pTHX_ const UV cp)
23324 U8 dummy[UTF8_MAXBYTES + 1];
23328 uvchr_to_utf8_flags_msgs(dummy, cp, UNICODE_WARN_PERL_EXTENDED,
23331 msg = hv_fetchs(msgs, "text", 0);
23334 (void) sv_2mortal((SV *) msgs);
23336 return SvPVX(*msg);
23340 #endif /* end of ! PERL_IN_XSUB_RE */
23343 S_compile_wildcard(pTHX_ const char * subpattern, const STRLEN len,
23344 const bool ignore_case)
23346 /* Pretends that the input subpattern is qr/subpattern/aam, compiling it
23347 * possibly with /i if the 'ignore_case' parameter is true. Use /aa
23348 * because nothing outside of ASCII will match. Use /m because the input
23349 * string may be a bunch of lines strung together.
23351 * Also sets up the debugging info */
23353 U32 flags = PMf_MULTILINE|PMf_WILDCARD;
23355 SV * subpattern_sv = sv_2mortal(newSVpvn(subpattern, len));
23356 REGEXP * subpattern_re;
23357 DECLARE_AND_GET_RE_DEBUG_FLAGS;
23359 PERL_ARGS_ASSERT_COMPILE_WILDCARD;
23364 set_regex_charset(&flags, REGEX_ASCII_MORE_RESTRICTED_CHARSET);
23366 /* Like in op.c, we copy the compile time pm flags to the rx ones */
23367 rx_flags = flags & RXf_PMf_COMPILETIME;
23369 #ifndef PERL_IN_XSUB_RE
23370 /* Use the core engine if this file is regcomp.c. That means no
23371 * 'use re "Debug ..." is in effect, so the core engine is sufficient */
23372 subpattern_re = Perl_re_op_compile(aTHX_ &subpattern_sv, 1, NULL,
23373 &PL_core_reg_engine,
23377 if (isDEBUG_WILDCARD) {
23378 /* Use the special debugging engine if this file is re_comp.c and wants
23379 * to output the wildcard matching. This uses whatever
23380 * 'use re "Debug ..." is in effect */
23381 subpattern_re = Perl_re_op_compile(aTHX_ &subpattern_sv, 1, NULL,
23387 /* Use the special wildcard engine if this file is re_comp.c and
23388 * doesn't want to output the wildcard matching. This uses whatever
23389 * 'use re "Debug ..." is in effect for compilation, but this engine
23390 * structure has been set up so that it uses the core engine for
23391 * execution, so no execution debugging as a result of re.pm will be
23393 subpattern_re = Perl_re_op_compile(aTHX_ &subpattern_sv, 1, NULL,
23397 /* XXX The above has the effect that any user-supplied regex engine
23398 * won't be called for matching wildcards. That might be good, or bad.
23399 * It could be changed in several ways. The reason it is done the
23400 * current way is to avoid having to save and restore
23401 * ^{^RE_DEBUG_FLAGS} around the execution. save_scalar() perhaps
23402 * could be used. Another suggestion is to keep the authoritative
23403 * value of the debug flags in a thread-local variable and add set/get
23404 * magic to ${^RE_DEBUG_FLAGS} to keep the C level variable up to date.
23405 * Still another is to pass a flag, say in the engine's intflags that
23406 * would be checked each time before doing the debug output */
23410 assert(subpattern_re); /* Should have died if didn't compile successfully */
23411 return subpattern_re;
23415 S_execute_wildcard(pTHX_ REGEXP * const prog, char* stringarg, char *strend,
23416 char *strbeg, SSize_t minend, SV *screamer, U32 nosave)
23419 DECLARE_AND_GET_RE_DEBUG_FLAGS;
23421 PERL_ARGS_ASSERT_EXECUTE_WILDCARD;
23425 /* The compilation has set things up so that if the program doesn't want to
23426 * see the wildcard matching procedure, it will get the core execution
23427 * engine, which is subject only to -Dr. So we have to turn that off
23428 * around this procedure */
23429 if (! isDEBUG_WILDCARD) {
23430 /* Note! Casts away 'volatile' */
23432 PL_debug &= ~ DEBUG_r_FLAG;
23435 result = CALLREGEXEC(prog, stringarg, strend, strbeg, minend, screamer,
23443 S_handle_user_defined_property(pTHX_
23445 /* Parses the contents of a user-defined property definition; returning the
23446 * expanded definition if possible. If so, the return is an inversion
23449 * If there are subroutines that are part of the expansion and which aren't
23450 * known at the time of the call to this function, this returns what
23451 * parse_uniprop_string() returned for the first one encountered.
23453 * If an error was found, NULL is returned, and 'msg' gets a suitable
23454 * message appended to it. (Appending allows the back trace of how we got
23455 * to the faulty definition to be displayed through nested calls of
23456 * user-defined subs.)
23458 * The caller IS responsible for freeing any returned SV.
23460 * The syntax of the contents is pretty much described in perlunicode.pod,
23461 * but we also allow comments on each line */
23463 const char * name, /* Name of property */
23464 const STRLEN name_len, /* The name's length in bytes */
23465 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
23466 const bool to_fold, /* ? Is this under /i */
23467 const bool runtime, /* ? Are we in compile- or run-time */
23468 const bool deferrable, /* Is it ok for this property's full definition
23469 to be deferred until later? */
23470 SV* contents, /* The property's definition */
23471 bool *user_defined_ptr, /* This will be set TRUE as we wouldn't be
23472 getting called unless this is thought to be
23473 a user-defined property */
23474 SV * msg, /* Any error or warning msg(s) are appended to
23476 const STRLEN level) /* Recursion level of this call */
23479 const char * string = SvPV_const(contents, len);
23480 const char * const e = string + len;
23481 const bool is_contents_utf8 = cBOOL(SvUTF8(contents));
23482 const STRLEN msgs_length_on_entry = SvCUR(msg);
23484 const char * s0 = string; /* Points to first byte in the current line
23485 being parsed in 'string' */
23486 const char overflow_msg[] = "Code point too large in \"";
23487 SV* running_definition = NULL;
23489 PERL_ARGS_ASSERT_HANDLE_USER_DEFINED_PROPERTY;
23491 *user_defined_ptr = TRUE;
23493 /* Look at each line */
23495 const char * s; /* Current byte */
23496 char op = '+'; /* Default operation is 'union' */
23497 IV min = 0; /* range begin code point */
23498 IV max = -1; /* and range end */
23499 SV* this_definition;
23501 /* Skip comment lines */
23503 s0 = strchr(s0, '\n');
23511 /* For backcompat, allow an empty first line */
23517 /* First character in the line may optionally be the operation */
23526 /* If the line is one or two hex digits separated by blank space, its
23527 * a range; otherwise it is either another user-defined property or an
23532 if (! isXDIGIT(*s)) {
23533 goto check_if_property;
23536 do { /* Each new hex digit will add 4 bits. */
23537 if (min > ( (IV) MAX_LEGAL_CP >> 4)) {
23538 s = strchr(s, '\n');
23542 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23543 sv_catpv(msg, overflow_msg);
23544 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
23545 UTF8fARG(is_contents_utf8, s - s0, s0));
23546 sv_catpvs(msg, "\"");
23547 goto return_failure;
23550 /* Accumulate this digit into the value */
23551 min = (min << 4) + READ_XDIGIT(s);
23552 } while (isXDIGIT(*s));
23554 while (isBLANK(*s)) { s++; }
23556 /* We allow comments at the end of the line */
23558 s = strchr(s, '\n');
23564 else if (s < e && *s != '\n') {
23565 if (! isXDIGIT(*s)) {
23566 goto check_if_property;
23569 /* Look for the high point of the range */
23572 if (max > ( (IV) MAX_LEGAL_CP >> 4)) {
23573 s = strchr(s, '\n');
23577 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23578 sv_catpv(msg, overflow_msg);
23579 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
23580 UTF8fARG(is_contents_utf8, s - s0, s0));
23581 sv_catpvs(msg, "\"");
23582 goto return_failure;
23585 max = (max << 4) + READ_XDIGIT(s);
23586 } while (isXDIGIT(*s));
23588 while (isBLANK(*s)) { s++; }
23591 s = strchr(s, '\n');
23596 else if (s < e && *s != '\n') {
23597 goto check_if_property;
23601 if (max == -1) { /* The line only had one entry */
23604 else if (max < min) {
23605 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23606 sv_catpvs(msg, "Illegal range in \"");
23607 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
23608 UTF8fARG(is_contents_utf8, s - s0, s0));
23609 sv_catpvs(msg, "\"");
23610 goto return_failure;
23613 # if 0 /* See explanation at definition above of get_extended_utf8_msg() */
23615 if ( UNICODE_IS_PERL_EXTENDED(min)
23616 || UNICODE_IS_PERL_EXTENDED(max))
23618 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23620 /* If both code points are non-portable, warn only on the lower
23622 sv_catpv(msg, get_extended_utf8_msg(
23623 (UNICODE_IS_PERL_EXTENDED(min))
23625 sv_catpvs(msg, " in \"");
23626 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
23627 UTF8fARG(is_contents_utf8, s - s0, s0));
23628 sv_catpvs(msg, "\"");
23633 /* Here, this line contains a legal range */
23634 this_definition = sv_2mortal(_new_invlist(2));
23635 this_definition = _add_range_to_invlist(this_definition, min, max);
23640 /* Here it isn't a legal range line. See if it is a legal property
23641 * line. First find the end of the meat of the line */
23642 s = strpbrk(s, "#\n");
23647 /* Ignore trailing blanks in keeping with the requirements of
23648 * parse_uniprop_string() */
23650 while (s > s0 && isBLANK_A(*s)) {
23655 this_definition = parse_uniprop_string(s0, s - s0,
23656 is_utf8, to_fold, runtime,
23659 user_defined_ptr, msg,
23661 ? level /* Don't increase level
23662 if input is empty */
23665 if (this_definition == NULL) {
23666 goto return_failure; /* 'msg' should have had the reason
23667 appended to it by the above call */
23670 if (! is_invlist(this_definition)) { /* Unknown at this time */
23671 return newSVsv(this_definition);
23675 s = strchr(s, '\n');
23685 _invlist_union(running_definition, this_definition,
23686 &running_definition);
23689 _invlist_subtract(running_definition, this_definition,
23690 &running_definition);
23693 _invlist_intersection(running_definition, this_definition,
23694 &running_definition);
23697 _invlist_union_complement_2nd(running_definition,
23698 this_definition, &running_definition);
23701 Perl_croak(aTHX_ "panic: %s: %d: Unexpected operation %d",
23702 __FILE__, __LINE__, op);
23706 /* Position past the '\n' */
23708 } /* End of loop through the lines of 'contents' */
23710 /* Here, we processed all the lines in 'contents' without error. If we
23711 * didn't add any warnings, simply return success */
23712 if (msgs_length_on_entry == SvCUR(msg)) {
23714 /* If the expansion was empty, the answer isn't nothing: its an empty
23715 * inversion list */
23716 if (running_definition == NULL) {
23717 running_definition = _new_invlist(1);
23720 return running_definition;
23723 /* Otherwise, add some explanatory text, but we will return success */
23727 running_definition = NULL;
23731 if (name_len > 0) {
23732 sv_catpvs(msg, " in expansion of ");
23733 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8, name_len, name));
23736 return running_definition;
23739 /* As explained below, certain operations need to take place in the first
23740 * thread created. These macros switch contexts */
23741 # ifdef USE_ITHREADS
23742 # define DECLARATION_FOR_GLOBAL_CONTEXT \
23743 PerlInterpreter * save_aTHX = aTHX;
23744 # define SWITCH_TO_GLOBAL_CONTEXT \
23745 PERL_SET_CONTEXT((aTHX = PL_user_def_props_aTHX))
23746 # define RESTORE_CONTEXT PERL_SET_CONTEXT((aTHX = save_aTHX));
23747 # define CUR_CONTEXT aTHX
23748 # define ORIGINAL_CONTEXT save_aTHX
23750 # define DECLARATION_FOR_GLOBAL_CONTEXT dNOOP
23751 # define SWITCH_TO_GLOBAL_CONTEXT NOOP
23752 # define RESTORE_CONTEXT NOOP
23753 # define CUR_CONTEXT NULL
23754 # define ORIGINAL_CONTEXT NULL
23758 S_delete_recursion_entry(pTHX_ void *key)
23760 /* Deletes the entry used to detect recursion when expanding user-defined
23761 * properties. This is a function so it can be set up to be called even if
23762 * the program unexpectedly quits */
23764 SV ** current_entry;
23765 const STRLEN key_len = strlen((const char *) key);
23766 DECLARATION_FOR_GLOBAL_CONTEXT;
23768 SWITCH_TO_GLOBAL_CONTEXT;
23770 /* If the entry is one of these types, it is a permanent entry, and not the
23771 * one used to detect recursions. This function should delete only the
23772 * recursion entry */
23773 current_entry = hv_fetch(PL_user_def_props, (const char *) key, key_len, 0);
23775 && ! is_invlist(*current_entry)
23776 && ! SvPOK(*current_entry))
23778 (void) hv_delete(PL_user_def_props, (const char *) key, key_len,
23786 S_get_fq_name(pTHX_
23787 const char * const name, /* The first non-blank in the \p{}, \P{} */
23788 const Size_t name_len, /* Its length in bytes, not including any trailing space */
23789 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
23790 const bool has_colon_colon
23793 /* Returns a mortal SV containing the fully qualified version of the input
23798 fq_name = newSVpvs_flags("", SVs_TEMP);
23800 /* Use the current package if it wasn't included in our input */
23801 if (! has_colon_colon) {
23802 const HV * pkg = (IN_PERL_COMPILETIME)
23804 : CopSTASH(PL_curcop);
23805 const char* pkgname = HvNAME(pkg);
23807 Perl_sv_catpvf(aTHX_ fq_name, "%" UTF8f,
23808 UTF8fARG(is_utf8, strlen(pkgname), pkgname));
23809 sv_catpvs(fq_name, "::");
23812 Perl_sv_catpvf(aTHX_ fq_name, "%" UTF8f,
23813 UTF8fARG(is_utf8, name_len, name));
23818 S_parse_uniprop_string(pTHX_
23820 /* Parse the interior of a \p{}, \P{}. Returns its definition if knowable
23821 * now. If so, the return is an inversion list.
23823 * If the property is user-defined, it is a subroutine, which in turn
23824 * may call other subroutines. This function will call the whole nest of
23825 * them to get the definition they return; if some aren't known at the time
23826 * of the call to this function, the fully qualified name of the highest
23827 * level sub is returned. It is an error to call this function at runtime
23828 * without every sub defined.
23830 * If an error was found, NULL is returned, and 'msg' gets a suitable
23831 * message appended to it. (Appending allows the back trace of how we got
23832 * to the faulty definition to be displayed through nested calls of
23833 * user-defined subs.)
23835 * The caller should NOT try to free any returned inversion list.
23837 * Other parameters will be set on return as described below */
23839 const char * const name, /* The first non-blank in the \p{}, \P{} */
23840 Size_t name_len, /* Its length in bytes, not including any
23842 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
23843 const bool to_fold, /* ? Is this under /i */
23844 const bool runtime, /* TRUE if this is being called at run time */
23845 const bool deferrable, /* TRUE if it's ok for the definition to not be
23846 known at this call */
23847 AV ** strings, /* To return string property values, like named
23849 bool *user_defined_ptr, /* Upon return from this function it will be
23850 set to TRUE if any component is a
23851 user-defined property */
23852 SV * msg, /* Any error or warning msg(s) are appended to
23854 const STRLEN level) /* Recursion level of this call */
23856 char* lookup_name; /* normalized name for lookup in our tables */
23857 unsigned lookup_len; /* Its length */
23858 enum { Not_Strict = 0, /* Some properties have stricter name */
23859 Strict, /* normalization rules, which we decide */
23860 As_Is /* upon based on parsing */
23861 } stricter = Not_Strict;
23863 /* nv= or numeric_value=, or possibly one of the cjk numeric properties
23864 * (though it requires extra effort to download them from Unicode and
23865 * compile perl to know about them) */
23866 bool is_nv_type = FALSE;
23868 unsigned int i, j = 0;
23869 int equals_pos = -1; /* Where the '=' is found, or negative if none */
23870 int slash_pos = -1; /* Where the '/' is found, or negative if none */
23871 int table_index = 0; /* The entry number for this property in the table
23872 of all Unicode property names */
23873 bool starts_with_Is = FALSE; /* ? Does the name start with 'Is' */
23874 Size_t lookup_offset = 0; /* Used to ignore the first few characters of
23875 the normalized name in certain situations */
23876 Size_t non_pkg_begin = 0; /* Offset of first byte in 'name' that isn't
23877 part of a package name */
23878 Size_t lun_non_pkg_begin = 0; /* Similarly for 'lookup_name' */
23879 bool could_be_user_defined = TRUE; /* ? Could this be a user-defined
23880 property rather than a Unicode
23882 SV * prop_definition = NULL; /* The returned definition of 'name' or NULL
23883 if an error. If it is an inversion list,
23884 it is the definition. Otherwise it is a
23885 string containing the fully qualified sub
23887 SV * fq_name = NULL; /* For user-defined properties, the fully
23889 bool invert_return = FALSE; /* ? Do we need to complement the result before
23891 bool stripped_utf8_pkg = FALSE; /* Set TRUE if the input includes an
23892 explicit utf8:: package that we strip
23894 /* The expansion of properties that could be either user-defined or
23895 * official unicode ones is deferred until runtime, including a marker for
23896 * those that might be in the latter category. This boolean indicates if
23897 * we've seen that marker. If not, what we're parsing can't be such an
23898 * official Unicode property whose expansion was deferred */
23899 bool could_be_deferred_official = FALSE;
23901 PERL_ARGS_ASSERT_PARSE_UNIPROP_STRING;
23903 /* The input will be normalized into 'lookup_name' */
23904 Newx(lookup_name, name_len, char);
23905 SAVEFREEPV(lookup_name);
23907 /* Parse the input. */
23908 for (i = 0; i < name_len; i++) {
23909 char cur = name[i];
23911 /* Most of the characters in the input will be of this ilk, being parts
23913 if (isIDCONT_A(cur)) {
23915 /* Case differences are ignored. Our lookup routine assumes
23916 * everything is lowercase, so normalize to that */
23917 if (isUPPER_A(cur)) {
23918 lookup_name[j++] = toLOWER_A(cur);
23922 if (cur == '_') { /* Don't include these in the normalized name */
23926 lookup_name[j++] = cur;
23928 /* The first character in a user-defined name must be of this type.
23930 if (i - non_pkg_begin == 0 && ! isIDFIRST_A(cur)) {
23931 could_be_user_defined = FALSE;
23937 /* Here, the character is not something typically in a name, But these
23938 * two types of characters (and the '_' above) can be freely ignored in
23939 * most situations. Later it may turn out we shouldn't have ignored
23940 * them, and we have to reparse, but we don't have enough information
23941 * yet to make that decision */
23942 if (cur == '-' || isSPACE_A(cur)) {
23943 could_be_user_defined = FALSE;
23947 /* An equals sign or single colon mark the end of the first part of
23948 * the property name */
23950 || (cur == ':' && (i >= name_len - 1 || name[i+1] != ':')))
23952 lookup_name[j++] = '='; /* Treat the colon as an '=' */
23953 equals_pos = j; /* Note where it occurred in the input */
23954 could_be_user_defined = FALSE;
23958 /* If this looks like it is a marker we inserted at compile time,
23959 * set a flag and otherwise ignore it. If it isn't in the final
23960 * position, keep it as it would have been user input. */
23961 if ( UNLIKELY(cur == DEFERRED_COULD_BE_OFFICIAL_MARKERc)
23963 && could_be_user_defined
23964 && i == name_len - 1)
23967 could_be_deferred_official = TRUE;
23971 /* Otherwise, this character is part of the name. */
23972 lookup_name[j++] = cur;
23974 /* Here it isn't a single colon, so if it is a colon, it must be a
23978 /* A double colon should be a package qualifier. We note its
23979 * position and continue. Note that one could have
23980 * pkg1::pkg2::...::foo
23981 * so that the position at the end of the loop will be just after
23982 * the final qualifier */
23985 non_pkg_begin = i + 1;
23986 lookup_name[j++] = ':';
23987 lun_non_pkg_begin = j;
23989 else { /* Only word chars (and '::') can be in a user-defined name */
23990 could_be_user_defined = FALSE;
23992 } /* End of parsing through the lhs of the property name (or all of it if
23995 # define STRLENs(s) (sizeof("" s "") - 1)
23997 /* If there is a single package name 'utf8::', it is ambiguous. It could
23998 * be for a user-defined property, or it could be a Unicode property, as
23999 * all of them are considered to be for that package. For the purposes of
24000 * parsing the rest of the property, strip it off */
24001 if (non_pkg_begin == STRLENs("utf8::") && memBEGINPs(name, name_len, "utf8::")) {
24002 lookup_name += STRLENs("utf8::");
24003 j -= STRLENs("utf8::");
24004 equals_pos -= STRLENs("utf8::");
24005 stripped_utf8_pkg = TRUE;
24008 /* Here, we are either done with the whole property name, if it was simple;
24009 * or are positioned just after the '=' if it is compound. */
24011 if (equals_pos >= 0) {
24012 assert(stricter == Not_Strict); /* We shouldn't have set this yet */
24014 /* Space immediately after the '=' is ignored */
24016 for (; i < name_len; i++) {
24017 if (! isSPACE_A(name[i])) {
24022 /* Most punctuation after the equals indicates a subpattern, like
24024 if ( isPUNCT_A(name[i])
24029 /* A backslash means the real delimitter is the next character,
24030 * but it must be punctuation */
24031 && (name[i] != '\\' || (i < name_len && isPUNCT_A(name[i+1]))))
24033 bool special_property = memEQs(lookup_name, j - 1, "name")
24034 || memEQs(lookup_name, j - 1, "na");
24035 if (! special_property) {
24036 /* Find the property. The table includes the equals sign, so
24037 * we use 'j' as-is */
24038 table_index = do_uniprop_match(lookup_name, j);
24040 if (special_property || table_index) {
24041 REGEXP * subpattern_re;
24042 char open = name[i++];
24044 const char * pos_in_brackets;
24045 const char * const * prop_values;
24048 /* Backslash => delimitter is the character following. We
24049 * already checked that it is punctuation */
24050 if (open == '\\') {
24055 /* This data structure is constructed so that the matching
24056 * closing bracket is 3 past its matching opening. The second
24057 * set of closing is so that if the opening is something like
24058 * ']', the closing will be that as well. Something similar is
24059 * done in toke.c */
24060 pos_in_brackets = memCHRs("([<)]>)]>", open);
24061 close = (pos_in_brackets) ? pos_in_brackets[3] : open;
24064 || name[name_len-1] != close
24065 || (escaped && name[name_len-2] != '\\')
24066 /* Also make sure that there are enough characters.
24067 * e.g., '\\\' would show up incorrectly as legal even
24068 * though it is too short */
24069 || (SSize_t) (name_len - i - 1 - escaped) < 0)
24071 sv_catpvs(msg, "Unicode property wildcard not terminated");
24072 goto append_name_to_msg;
24075 Perl_ck_warner_d(aTHX_
24076 packWARN(WARN_EXPERIMENTAL__UNIPROP_WILDCARDS),
24077 "The Unicode property wildcards feature is experimental");
24079 if (special_property) {
24080 const char * error_msg;
24081 const char * revised_name = name + i;
24082 Size_t revised_name_len = name_len - (i + 1 + escaped);
24084 /* Currently, the only 'special_property' is name, which we
24085 * lookup in _charnames.pm */
24087 if (! load_charnames(newSVpvs("placeholder"),
24088 revised_name, revised_name_len,
24091 sv_catpv(msg, error_msg);
24092 goto append_name_to_msg;
24095 /* Farm this out to a function just to make the current
24096 * function less unwieldy */
24097 if (handle_names_wildcard(revised_name, revised_name_len,
24101 return prop_definition;
24107 prop_values = get_prop_values(table_index);
24109 /* Now create and compile the wildcard subpattern. Use /i
24110 * because the property values are supposed to match with case
24112 subpattern_re = compile_wildcard(name + i,
24113 name_len - i - 1 - escaped,
24117 /* For each legal property value, see if the supplied pattern
24119 while (*prop_values) {
24120 const char * const entry = *prop_values;
24121 const Size_t len = strlen(entry);
24122 SV* entry_sv = newSVpvn_flags(entry, len, SVs_TEMP);
24124 if (execute_wildcard(subpattern_re,
24126 (char *) entry + len,
24130 { /* Here, matched. Add to the returned list */
24131 Size_t total_len = j + len;
24132 SV * sub_invlist = NULL;
24133 char * this_string;
24135 /* We know this is a legal \p{property=value}. Call
24136 * the function to return the list of code points that
24138 Newxz(this_string, total_len + 1, char);
24139 Copy(lookup_name, this_string, j, char);
24140 my_strlcat(this_string, entry, total_len + 1);
24141 SAVEFREEPV(this_string);
24142 sub_invlist = parse_uniprop_string(this_string,
24152 _invlist_union(prop_definition, sub_invlist,
24156 prop_values++; /* Next iteration, look at next propvalue */
24157 } /* End of looking through property values; (the data
24158 structure is terminated by a NULL ptr) */
24160 SvREFCNT_dec_NN(subpattern_re);
24162 if (prop_definition) {
24163 return prop_definition;
24166 sv_catpvs(msg, "No Unicode property value wildcard matches:");
24167 goto append_name_to_msg;
24170 /* Here's how khw thinks we should proceed to handle the properties
24171 * not yet done: Bidi Mirroring Glyph can map to ""
24172 Bidi Paired Bracket can map to ""
24173 Case Folding (both full and simple)
24174 Shouldn't /i be good enough for Full
24175 Decomposition Mapping
24176 Equivalent Unified Ideograph can map to ""
24177 Lowercase Mapping (both full and simple)
24178 NFKC Case Fold can map to ""
24179 Titlecase Mapping (both full and simple)
24180 Uppercase Mapping (both full and simple)
24181 * Handle these the same way Name is done, using say, _wild.pm, but
24182 * having both loose and full, like in charclass_invlists.h.
24183 * Perhaps move block and script to that as they are somewhat large
24184 * in charclass_invlists.h.
24185 * For properties where the default is the code point itself, such
24186 * as any of the case changing mappings, the string would otherwise
24187 * consist of all Unicode code points in UTF-8 strung together.
24188 * This would be impractical. So instead, examine their compiled
24189 * pattern, looking at the ssc. If none, reject the pattern as an
24190 * error. Otherwise run the pattern against every code point in
24191 * the ssc. The ssc is kind of like tr18's 3.9 Possible Match Sets
24192 * And it might be good to create an API to return the ssc.
24193 * Or handle them like the algorithmic names are done
24195 } /* End of is a wildcard subppattern */
24197 /* \p{name=...} is handled specially. Instead of using the normal
24198 * mechanism involving charclass_invlists.h, it uses _charnames.pm
24199 * which has the necessary (huge) data accessible to it, and which
24200 * doesn't get loaded unless necessary. The legal syntax for names is
24201 * somewhat different than other properties due both to the vagaries of
24202 * a few outlier official names, and the fact that only a few ASCII
24203 * characters are permitted in them */
24204 if ( memEQs(lookup_name, j - 1, "name")
24205 || memEQs(lookup_name, j - 1, "na"))
24210 const char * error_msg;
24212 SV * character_name;
24213 STRLEN character_len;
24218 /* Since the RHS (after skipping initial space) is passed unchanged
24219 * to charnames, and there are different criteria for what are
24220 * legal characters in the name, just parse it here. A character
24221 * name must begin with an ASCII alphabetic */
24222 if (! isALPHA(name[i])) {
24225 lookup_name[j++] = name[i];
24227 for (++i; i < name_len; i++) {
24228 /* Official names can only be in the ASCII range, and only
24229 * certain characters */
24230 if (! isASCII(name[i]) || ! isCHARNAME_CONT(name[i])) {
24233 lookup_name[j++] = name[i];
24236 /* Finished parsing, save the name into an SV */
24237 character_name = newSVpvn(lookup_name + equals_pos, j - equals_pos);
24239 /* Make sure _charnames is loaded. (The parameters give context
24240 * for any errors generated */
24241 table = load_charnames(character_name, name, name_len, &error_msg);
24242 if (table == NULL) {
24243 sv_catpv(msg, error_msg);
24244 goto append_name_to_msg;
24247 lookup_loose = get_cvs("_charnames::_loose_regcomp_lookup", 0);
24248 if (! lookup_loose) {
24250 "panic: Can't find '_charnames::_loose_regcomp_lookup");
24253 PUSHSTACKi(PERLSI_REGCOMP);
24259 XPUSHs(character_name);
24261 call_sv(MUTABLE_SV(lookup_loose), G_SCALAR);
24266 SvREFCNT_inc_simple_void_NN(character);
24273 if (! SvOK(character)) {
24277 cp = valid_utf8_to_uvchr((U8 *) SvPVX(character), &character_len);
24278 if (character_len == SvCUR(character)) {
24279 prop_definition = add_cp_to_invlist(NULL, cp);
24284 /* First of the remaining characters in the string. */
24285 char * remaining = SvPVX(character) + character_len;
24287 if (strings == NULL) {
24288 goto failed; /* XXX Perhaps a specific msg instead, like
24289 'not available here' */
24292 if (*strings == NULL) {
24293 *strings = newAV();
24296 this_string = newAV();
24297 av_push(this_string, newSVuv(cp));
24300 cp = valid_utf8_to_uvchr((U8 *) remaining, &character_len);
24301 av_push(this_string, newSVuv(cp));
24302 remaining += character_len;
24303 } while (remaining < SvEND(character));
24305 av_push(*strings, (SV *) this_string);
24308 return prop_definition;
24311 /* Certain properties whose values are numeric need special handling.
24312 * They may optionally be prefixed by 'is'. Ignore that prefix for the
24313 * purposes of checking if this is one of those properties */
24314 if (memBEGINPs(lookup_name, j, "is")) {
24318 /* Then check if it is one of these specially-handled properties. The
24319 * possibilities are hard-coded because easier this way, and the list
24320 * is unlikely to change.
24322 * All numeric value type properties are of this ilk, and are also
24323 * special in a different way later on. So find those first. There
24324 * are several numeric value type properties in the Unihan DB (which is
24325 * unlikely to be compiled with perl, but we handle it here in case it
24326 * does get compiled). They all end with 'numeric'. The interiors
24327 * aren't checked for the precise property. This would stop working if
24328 * a cjk property were to be created that ended with 'numeric' and
24329 * wasn't a numeric type */
24330 is_nv_type = memEQs(lookup_name + lookup_offset,
24331 j - 1 - lookup_offset, "numericvalue")
24332 || memEQs(lookup_name + lookup_offset,
24333 j - 1 - lookup_offset, "nv")
24334 || ( memENDPs(lookup_name + lookup_offset,
24335 j - 1 - lookup_offset, "numeric")
24336 && ( memBEGINPs(lookup_name + lookup_offset,
24337 j - 1 - lookup_offset, "cjk")
24338 || memBEGINPs(lookup_name + lookup_offset,
24339 j - 1 - lookup_offset, "k")));
24341 || memEQs(lookup_name + lookup_offset,
24342 j - 1 - lookup_offset, "canonicalcombiningclass")
24343 || memEQs(lookup_name + lookup_offset,
24344 j - 1 - lookup_offset, "ccc")
24345 || memEQs(lookup_name + lookup_offset,
24346 j - 1 - lookup_offset, "age")
24347 || memEQs(lookup_name + lookup_offset,
24348 j - 1 - lookup_offset, "in")
24349 || memEQs(lookup_name + lookup_offset,
24350 j - 1 - lookup_offset, "presentin"))
24354 /* Since the stuff after the '=' is a number, we can't throw away
24355 * '-' willy-nilly, as those could be a minus sign. Other stricter
24356 * rules also apply. However, these properties all can have the
24357 * rhs not be a number, in which case they contain at least one
24358 * alphabetic. In those cases, the stricter rules don't apply.
24359 * But the numeric type properties can have the alphas [Ee] to
24360 * signify an exponent, and it is still a number with stricter
24361 * rules. So look for an alpha that signifies not-strict */
24363 for (k = i; k < name_len; k++) {
24364 if ( isALPHA_A(name[k])
24365 && (! is_nv_type || ! isALPHA_FOLD_EQ(name[k], 'E')))
24367 stricter = Not_Strict;
24375 /* A number may have a leading '+' or '-'. The latter is retained
24377 if (name[i] == '+') {
24380 else if (name[i] == '-') {
24381 lookup_name[j++] = '-';
24385 /* Skip leading zeros including single underscores separating the
24386 * zeros, or between the final leading zero and the first other
24388 for (; i < name_len - 1; i++) {
24389 if ( name[i] != '0'
24390 && (name[i] != '_' || ! isDIGIT_A(name[i+1])))
24397 else { /* No '=' */
24399 /* Only a few properties without an '=' should be parsed with stricter
24400 * rules. The list is unlikely to change. */
24401 if ( memBEGINPs(lookup_name, j, "perl")
24402 && memNEs(lookup_name + 4, j - 4, "space")
24403 && memNEs(lookup_name + 4, j - 4, "word"))
24407 /* We set the inputs back to 0 and the code below will reparse,
24413 /* Here, we have either finished the property, or are positioned to parse
24414 * the remainder, and we know if stricter rules apply. Finish out, if not
24416 for (; i < name_len; i++) {
24417 char cur = name[i];
24419 /* In all instances, case differences are ignored, and we normalize to
24421 if (isUPPER_A(cur)) {
24422 lookup_name[j++] = toLOWER(cur);
24426 /* An underscore is skipped, but not under strict rules unless it
24427 * separates two digits */
24430 && ( i == 0 || (int) i == equals_pos || i == name_len- 1
24431 || ! isDIGIT_A(name[i-1]) || ! isDIGIT_A(name[i+1])))
24433 lookup_name[j++] = '_';
24438 /* Hyphens are skipped except under strict */
24439 if (cur == '-' && ! stricter) {
24443 /* XXX Bug in documentation. It says white space skipped adjacent to
24444 * non-word char. Maybe we should, but shouldn't skip it next to a dot
24446 if (isSPACE_A(cur) && ! stricter) {
24450 lookup_name[j++] = cur;
24452 /* Unless this is a non-trailing slash, we are done with it */
24453 if (i >= name_len - 1 || cur != '/') {
24459 /* A slash in the 'numeric value' property indicates that what follows
24460 * is a denominator. It can have a leading '+' and '0's that should be
24461 * skipped. But we have never allowed a negative denominator, so treat
24462 * a minus like every other character. (No need to rule out a second
24463 * '/', as that won't match anything anyway */
24466 if (i < name_len && name[i] == '+') {
24470 /* Skip leading zeros including underscores separating digits */
24471 for (; i < name_len - 1; i++) {
24472 if ( name[i] != '0'
24473 && (name[i] != '_' || ! isDIGIT_A(name[i+1])))
24479 /* Store the first real character in the denominator */
24480 if (i < name_len) {
24481 lookup_name[j++] = name[i];
24486 /* Here are completely done parsing the input 'name', and 'lookup_name'
24487 * contains a copy, normalized.
24489 * This special case is grandfathered in: 'L_' and 'GC=L_' are accepted and
24490 * different from without the underscores. */
24491 if ( ( UNLIKELY(memEQs(lookup_name, j, "l"))
24492 || UNLIKELY(memEQs(lookup_name, j, "gc=l")))
24493 && UNLIKELY(name[name_len-1] == '_'))
24495 lookup_name[j++] = '&';
24498 /* If the original input began with 'In' or 'Is', it could be a subroutine
24499 * call to a user-defined property instead of a Unicode property name. */
24500 if ( name_len - non_pkg_begin > 2
24501 && name[non_pkg_begin+0] == 'I'
24502 && (name[non_pkg_begin+1] == 'n' || name[non_pkg_begin+1] == 's'))
24504 /* Names that start with In have different characterstics than those
24505 * that start with Is */
24506 if (name[non_pkg_begin+1] == 's') {
24507 starts_with_Is = TRUE;
24511 could_be_user_defined = FALSE;
24514 if (could_be_user_defined) {
24517 /* If the user defined property returns the empty string, it could
24518 * easily be because the pattern is being compiled before the data it
24519 * actually needs to compile is available. This could be argued to be
24520 * a bug in the perl code, but this is a change of behavior for Perl,
24521 * so we handle it. This means that intentionally returning nothing
24522 * will not be resolved until runtime */
24523 bool empty_return = FALSE;
24525 /* Here, the name could be for a user defined property, which are
24526 * implemented as subs. */
24527 user_sub = get_cvn_flags(name, name_len, 0);
24530 /* Here, the property name could be a user-defined one, but there
24531 * is no subroutine to handle it (as of now). Defer handling it
24532 * until runtime. Otherwise, a block defined by Unicode in a later
24533 * release would get the synonym InFoo added for it, and existing
24534 * code that used that name would suddenly break if it referred to
24535 * the property before the sub was declared. See [perl #134146] */
24537 goto definition_deferred;
24540 /* Here, we are at runtime, and didn't find the user property. It
24541 * could be an official property, but only if no package was
24542 * specified, or just the utf8:: package. */
24543 if (could_be_deferred_official) {
24544 lookup_name += lun_non_pkg_begin;
24545 j -= lun_non_pkg_begin;
24547 else if (! stripped_utf8_pkg) {
24548 goto unknown_user_defined;
24551 /* Drop down to look up in the official properties */
24554 const char insecure[] = "Insecure user-defined property";
24556 /* Here, there is a sub by the correct name. Normally we call it
24557 * to get the property definition */
24559 SV * user_sub_sv = MUTABLE_SV(user_sub);
24560 SV * error; /* Any error returned by calling 'user_sub' */
24561 SV * key; /* The key into the hash of user defined sub names
24564 SV ** saved_user_prop_ptr; /* Hash entry for this property */
24566 /* How many times to retry when another thread is in the middle of
24567 * expanding the same definition we want */
24568 PERL_INT_FAST8_T retry_countdown = 10;
24570 DECLARATION_FOR_GLOBAL_CONTEXT;
24572 /* If we get here, we know this property is user-defined */
24573 *user_defined_ptr = TRUE;
24575 /* We refuse to call a potentially tainted subroutine; returning an
24578 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24579 sv_catpvn(msg, insecure, sizeof(insecure) - 1);
24580 goto append_name_to_msg;
24583 /* In principal, we only call each subroutine property definition
24584 * once during the life of the program. This guarantees that the
24585 * property definition never changes. The results of the single
24586 * sub call are stored in a hash, which is used instead for future
24587 * references to this property. The property definition is thus
24588 * immutable. But, to allow the user to have a /i-dependent
24589 * definition, we call the sub once for non-/i, and once for /i,
24590 * should the need arise, passing the /i status as a parameter.
24592 * We start by constructing the hash key name, consisting of the
24593 * fully qualified subroutine name, preceded by the /i status, so
24594 * that there is a key for /i and a different key for non-/i */
24595 key = newSVpvn(((to_fold) ? "1" : "0"), 1);
24596 fq_name = S_get_fq_name(aTHX_ name, name_len, is_utf8,
24597 non_pkg_begin != 0);
24598 sv_catsv(key, fq_name);
24601 /* We only call the sub once throughout the life of the program
24602 * (with the /i, non-/i exception noted above). That means the
24603 * hash must be global and accessible to all threads. It is
24604 * created at program start-up, before any threads are created, so
24605 * is accessible to all children. But this creates some
24608 * 1) The keys can't be shared, or else problems arise; sharing is
24609 * turned off at hash creation time
24610 * 2) All SVs in it are there for the remainder of the life of the
24611 * program, and must be created in the same interpreter context
24612 * as the hash, or else they will be freed from the wrong pool
24613 * at global destruction time. This is handled by switching to
24614 * the hash's context to create each SV going into it, and then
24615 * immediately switching back
24616 * 3) All accesses to the hash must be controlled by a mutex, to
24617 * prevent two threads from getting an unstable state should
24618 * they simultaneously be accessing it. The code below is
24619 * crafted so that the mutex is locked whenever there is an
24620 * access and unlocked only when the next stable state is
24623 * The hash stores either the definition of the property if it was
24624 * valid, or, if invalid, the error message that was raised. We
24625 * use the type of SV to distinguish.
24627 * There's also the need to guard against the definition expansion
24628 * from infinitely recursing. This is handled by storing the aTHX
24629 * of the expanding thread during the expansion. Again the SV type
24630 * is used to distinguish this from the other two cases. If we
24631 * come to here and the hash entry for this property is our aTHX,
24632 * it means we have recursed, and the code assumes that we would
24633 * infinitely recurse, so instead stops and raises an error.
24634 * (Any recursion has always been treated as infinite recursion in
24637 * If instead, the entry is for a different aTHX, it means that
24638 * that thread has gotten here first, and hasn't finished expanding
24639 * the definition yet. We just have to wait until it is done. We
24640 * sleep and retry a few times, returning an error if the other
24641 * thread doesn't complete. */
24644 USER_PROP_MUTEX_LOCK;
24646 /* If we have an entry for this key, the subroutine has already
24647 * been called once with this /i status. */
24648 saved_user_prop_ptr = hv_fetch(PL_user_def_props,
24649 SvPVX(key), SvCUR(key), 0);
24650 if (saved_user_prop_ptr) {
24652 /* If the saved result is an inversion list, it is the valid
24653 * definition of this property */
24654 if (is_invlist(*saved_user_prop_ptr)) {
24655 prop_definition = *saved_user_prop_ptr;
24657 /* The SV in the hash won't be removed until global
24658 * destruction, so it is stable and we can unlock */
24659 USER_PROP_MUTEX_UNLOCK;
24661 /* The caller shouldn't try to free this SV */
24662 return prop_definition;
24665 /* Otherwise, if it is a string, it is the error message
24666 * that was returned when we first tried to evaluate this
24667 * property. Fail, and append the message */
24668 if (SvPOK(*saved_user_prop_ptr)) {
24669 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24670 sv_catsv(msg, *saved_user_prop_ptr);
24672 /* The SV in the hash won't be removed until global
24673 * destruction, so it is stable and we can unlock */
24674 USER_PROP_MUTEX_UNLOCK;
24679 assert(SvIOK(*saved_user_prop_ptr));
24681 /* Here, we have an unstable entry in the hash. Either another
24682 * thread is in the middle of expanding the property's
24683 * definition, or we are ourselves recursing. We use the aTHX
24684 * in it to distinguish */
24685 if (SvIV(*saved_user_prop_ptr) != PTR2IV(CUR_CONTEXT)) {
24687 /* Here, it's another thread doing the expanding. We've
24688 * looked as much as we are going to at the contents of the
24689 * hash entry. It's safe to unlock. */
24690 USER_PROP_MUTEX_UNLOCK;
24692 /* Retry a few times */
24693 if (retry_countdown-- > 0) {
24698 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24699 sv_catpvs(msg, "Timeout waiting for another thread to "
24701 goto append_name_to_msg;
24704 /* Here, we are recursing; don't dig any deeper */
24705 USER_PROP_MUTEX_UNLOCK;
24707 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24709 "Infinite recursion in user-defined property");
24710 goto append_name_to_msg;
24713 /* Here, this thread has exclusive control, and there is no entry
24714 * for this property in the hash. So we have the go ahead to
24715 * expand the definition ourselves. */
24717 PUSHSTACKi(PERLSI_REGCOMP);
24720 /* Create a temporary placeholder in the hash to detect recursion
24722 SWITCH_TO_GLOBAL_CONTEXT;
24723 placeholder= newSVuv(PTR2IV(ORIGINAL_CONTEXT));
24724 (void) hv_store_ent(PL_user_def_props, key, placeholder, 0);
24727 /* Now that we have a placeholder, we can let other threads
24729 USER_PROP_MUTEX_UNLOCK;
24731 /* Make sure the placeholder always gets destroyed */
24732 SAVEDESTRUCTOR_X(S_delete_recursion_entry, SvPVX(key));
24737 /* Call the user's function, with the /i status as a parameter.
24738 * Note that we have gone to a lot of trouble to keep this call
24739 * from being within the locked mutex region. */
24740 XPUSHs(boolSV(to_fold));
24743 /* The following block was taken from swash_init(). Presumably
24744 * they apply to here as well, though we no longer use a swash --
24748 /* We might get here via a subroutine signature which uses a utf8
24749 * parameter name, at which point PL_subname will have been set
24750 * but not yet used. */
24751 save_item(PL_subname);
24753 /* G_SCALAR guarantees a single return value */
24754 (void) call_sv(user_sub_sv, G_EVAL|G_SCALAR);
24759 if (TAINT_get || SvTRUE(error)) {
24760 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24761 if (SvTRUE(error)) {
24762 sv_catpvs(msg, "Error \"");
24763 sv_catsv(msg, error);
24764 sv_catpvs(msg, "\"");
24767 if (SvTRUE(error)) sv_catpvs(msg, "; ");
24768 sv_catpvn(msg, insecure, sizeof(insecure) - 1);
24771 if (name_len > 0) {
24772 sv_catpvs(msg, " in expansion of ");
24773 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8,
24779 prop_definition = NULL;
24782 SV * contents = POPs;
24784 /* The contents is supposed to be the expansion of the property
24785 * definition. If the definition is deferrable, and we got an
24786 * empty string back, set a flag to later defer it (after clean
24789 && (! SvPOK(contents) || SvCUR(contents) == 0))
24791 empty_return = TRUE;
24793 else { /* Otherwise, call a function to check for valid syntax,
24796 prop_definition = handle_user_defined_property(
24798 is_utf8, to_fold, runtime,
24800 contents, user_defined_ptr,
24806 /* Here, we have the results of the expansion. Delete the
24807 * placeholder, and if the definition is now known, replace it with
24808 * that definition. We need exclusive access to the hash, and we
24809 * can't let anyone else in, between when we delete the placeholder
24810 * and add the permanent entry */
24811 USER_PROP_MUTEX_LOCK;
24813 S_delete_recursion_entry(aTHX_ SvPVX(key));
24815 if ( ! empty_return
24816 && (! prop_definition || is_invlist(prop_definition)))
24818 /* If we got success we use the inversion list defining the
24819 * property; otherwise use the error message */
24820 SWITCH_TO_GLOBAL_CONTEXT;
24821 (void) hv_store_ent(PL_user_def_props,
24824 ? newSVsv(prop_definition)
24830 /* All done, and the hash now has a permanent entry for this
24831 * property. Give up exclusive control */
24832 USER_PROP_MUTEX_UNLOCK;
24838 if (empty_return) {
24839 goto definition_deferred;
24842 if (prop_definition) {
24844 /* If the definition is for something not known at this time,
24845 * we toss it, and go return the main property name, as that's
24846 * the one the user will be aware of */
24847 if (! is_invlist(prop_definition)) {
24848 SvREFCNT_dec_NN(prop_definition);
24849 goto definition_deferred;
24852 sv_2mortal(prop_definition);
24856 return prop_definition;
24858 } /* End of calling the subroutine for the user-defined property */
24859 } /* End of it could be a user-defined property */
24861 /* Here it wasn't a user-defined property that is known at this time. See
24862 * if it is a Unicode property */
24864 lookup_len = j; /* This is a more mnemonic name than 'j' */
24866 /* Get the index into our pointer table of the inversion list corresponding
24867 * to the property */
24868 table_index = do_uniprop_match(lookup_name, lookup_len);
24870 /* If it didn't find the property ... */
24871 if (table_index == 0) {
24873 /* Try again stripping off any initial 'Is'. This is because we
24874 * promise that an initial Is is optional. The same isn't true of
24875 * names that start with 'In'. Those can match only blocks, and the
24876 * lookup table already has those accounted for. The lookup table also
24877 * has already accounted for Perl extensions (without and = sign)
24878 * starting with 'i's'. */
24879 if (starts_with_Is && equals_pos >= 0) {
24885 table_index = do_uniprop_match(lookup_name, lookup_len);
24888 if (table_index == 0) {
24891 /* Here, we didn't find it. If not a numeric type property, and
24892 * can't be a user-defined one, it isn't a legal property */
24893 if (! is_nv_type) {
24894 if (! could_be_user_defined) {
24898 /* Here, the property name is legal as a user-defined one. At
24899 * compile time, it might just be that the subroutine for that
24900 * property hasn't been encountered yet, but at runtime, it's
24901 * an error to try to use an undefined one */
24902 if (! deferrable) {
24903 goto unknown_user_defined;;
24906 goto definition_deferred;
24907 } /* End of isn't a numeric type property */
24909 /* The numeric type properties need more work to decide. What we
24910 * do is make sure we have the number in canonical form and look
24913 if (slash_pos < 0) { /* No slash */
24915 /* When it isn't a rational, take the input, convert it to a
24916 * NV, then create a canonical string representation of that
24920 SSize_t value_len = lookup_len - equals_pos;
24922 /* Get the value */
24923 if ( value_len <= 0
24924 || my_atof3(lookup_name + equals_pos, &value,
24926 != lookup_name + lookup_len)
24931 /* If the value is an integer, the canonical value is integral
24933 if (Perl_ceil(value) == value) {
24934 canonical = Perl_form(aTHX_ "%.*s%.0" NVff,
24935 equals_pos, lookup_name, value);
24937 else { /* Otherwise, it is %e with a known precision */
24940 canonical = Perl_form(aTHX_ "%.*s%.*" NVef,
24941 equals_pos, lookup_name,
24942 PL_E_FORMAT_PRECISION, value);
24944 /* The exponent generated is expecting two digits, whereas
24945 * %e on some systems will generate three. Remove leading
24946 * zeros in excess of 2 from the exponent. We start
24947 * looking for them after the '=' */
24948 exp_ptr = strchr(canonical + equals_pos, 'e');
24950 char * cur_ptr = exp_ptr + 2; /* past the 'e[+-]' */
24951 SSize_t excess_exponent_len = strlen(cur_ptr) - 2;
24953 assert(*(cur_ptr - 1) == '-' || *(cur_ptr - 1) == '+');
24955 if (excess_exponent_len > 0) {
24956 SSize_t leading_zeros = strspn(cur_ptr, "0");
24957 SSize_t excess_leading_zeros
24958 = MIN(leading_zeros, excess_exponent_len);
24959 if (excess_leading_zeros > 0) {
24960 Move(cur_ptr + excess_leading_zeros,
24962 strlen(cur_ptr) - excess_leading_zeros
24963 + 1, /* Copy the NUL as well */
24970 else { /* Has a slash. Create a rational in canonical form */
24971 UV numerator, denominator, gcd, trial;
24972 const char * end_ptr;
24973 const char * sign = "";
24975 /* We can't just find the numerator, denominator, and do the
24976 * division, then use the method above, because that is
24977 * inexact. And the input could be a rational that is within
24978 * epsilon (given our precision) of a valid rational, and would
24979 * then incorrectly compare valid.
24981 * We're only interested in the part after the '=' */
24982 const char * this_lookup_name = lookup_name + equals_pos;
24983 lookup_len -= equals_pos;
24984 slash_pos -= equals_pos;
24986 /* Handle any leading minus */
24987 if (this_lookup_name[0] == '-') {
24989 this_lookup_name++;
24994 /* Convert the numerator to numeric */
24995 end_ptr = this_lookup_name + slash_pos;
24996 if (! grok_atoUV(this_lookup_name, &numerator, &end_ptr)) {
25000 /* It better have included all characters before the slash */
25001 if (*end_ptr != '/') {
25005 /* Set to look at just the denominator */
25006 this_lookup_name += slash_pos;
25007 lookup_len -= slash_pos;
25008 end_ptr = this_lookup_name + lookup_len;
25010 /* Convert the denominator to numeric */
25011 if (! grok_atoUV(this_lookup_name, &denominator, &end_ptr)) {
25015 /* It better be the rest of the characters, and don't divide by
25017 if ( end_ptr != this_lookup_name + lookup_len
25018 || denominator == 0)
25023 /* Get the greatest common denominator using
25024 http://en.wikipedia.org/wiki/Euclidean_algorithm */
25026 trial = denominator;
25027 while (trial != 0) {
25029 trial = gcd % trial;
25033 /* If already in lowest possible terms, we have already tried
25034 * looking this up */
25039 /* Reduce the rational, which should put it in canonical form
25042 denominator /= gcd;
25044 canonical = Perl_form(aTHX_ "%.*s%s%" UVuf "/%" UVuf,
25045 equals_pos, lookup_name, sign, numerator, denominator);
25048 /* Here, we have the number in canonical form. Try that */
25049 table_index = do_uniprop_match(canonical, strlen(canonical));
25050 if (table_index == 0) {
25053 } /* End of still didn't find the property in our table */
25054 } /* End of didn't find the property in our table */
25056 /* Here, we have a non-zero return, which is an index into a table of ptrs.
25057 * A negative return signifies that the real index is the absolute value,
25058 * but the result needs to be inverted */
25059 if (table_index < 0) {
25060 invert_return = TRUE;
25061 table_index = -table_index;
25064 /* Out-of band indices indicate a deprecated property. The proper index is
25065 * modulo it with the table size. And dividing by the table size yields
25066 * an offset into a table constructed by regen/mk_invlists.pl to contain
25067 * the corresponding warning message */
25068 if (table_index > MAX_UNI_KEYWORD_INDEX) {
25069 Size_t warning_offset = table_index / MAX_UNI_KEYWORD_INDEX;
25070 table_index %= MAX_UNI_KEYWORD_INDEX;
25071 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED),
25072 "Use of '%.*s' in \\p{} or \\P{} is deprecated because: %s",
25073 (int) name_len, name,
25074 get_deprecated_property_msg(warning_offset));
25077 /* In a few properties, a different property is used under /i. These are
25078 * unlikely to change, so are hard-coded here. */
25080 if ( table_index == UNI_XPOSIXUPPER
25081 || table_index == UNI_XPOSIXLOWER
25082 || table_index == UNI_TITLE)
25084 table_index = UNI_CASED;
25086 else if ( table_index == UNI_UPPERCASELETTER
25087 || table_index == UNI_LOWERCASELETTER
25088 # ifdef UNI_TITLECASELETTER /* Missing from early Unicodes */
25089 || table_index == UNI_TITLECASELETTER
25092 table_index = UNI_CASEDLETTER;
25094 else if ( table_index == UNI_POSIXUPPER
25095 || table_index == UNI_POSIXLOWER)
25097 table_index = UNI_POSIXALPHA;
25101 /* Create and return the inversion list */
25102 prop_definition = get_prop_definition(table_index);
25103 sv_2mortal(prop_definition);
25105 /* See if there is a private use override to add to this definition */
25107 COPHH * hinthash = (IN_PERL_COMPILETIME)
25108 ? CopHINTHASH_get(&PL_compiling)
25109 : CopHINTHASH_get(PL_curcop);
25110 SV * pu_overrides = cophh_fetch_pv(hinthash, "private_use", 0, 0);
25112 if (UNLIKELY(pu_overrides && SvPOK(pu_overrides))) {
25114 /* See if there is an element in the hints hash for this table */
25115 SV * pu_lookup = Perl_newSVpvf(aTHX_ "%d=", table_index);
25116 const char * pos = strstr(SvPVX(pu_overrides), SvPVX(pu_lookup));
25120 SV * pu_definition;
25122 SV * expanded_prop_definition =
25123 sv_2mortal(invlist_clone(prop_definition, NULL));
25125 /* If so, it's definition is the string from here to the next
25126 * \a character. And its format is the same as a user-defined
25128 pos += SvCUR(pu_lookup);
25129 pu_definition = newSVpvn(pos, strchr(pos, '\a') - pos);
25130 pu_invlist = handle_user_defined_property(lookup_name,
25133 0, /* Not folded */
25141 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
25142 sv_catpvs(msg, "Insecure private-use override");
25143 goto append_name_to_msg;
25146 /* For now, as a safety measure, make sure that it doesn't
25147 * override non-private use code points */
25148 _invlist_intersection(pu_invlist, PL_Private_Use, &pu_invlist);
25150 /* Add it to the list to be returned */
25151 _invlist_union(prop_definition, pu_invlist,
25152 &expanded_prop_definition);
25153 prop_definition = expanded_prop_definition;
25154 Perl_ck_warner_d(aTHX_ packWARN(WARN_EXPERIMENTAL__PRIVATE_USE), "The private_use feature is experimental");
25159 if (invert_return) {
25160 _invlist_invert(prop_definition);
25162 return prop_definition;
25164 unknown_user_defined:
25165 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
25166 sv_catpvs(msg, "Unknown user-defined property name");
25167 goto append_name_to_msg;
25170 if (non_pkg_begin != 0) {
25171 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
25172 sv_catpvs(msg, "Illegal user-defined property name");
25175 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
25176 sv_catpvs(msg, "Can't find Unicode property definition");
25180 append_name_to_msg:
25182 const char * prefix = (runtime && level == 0) ? " \\p{" : " \"";
25183 const char * suffix = (runtime && level == 0) ? "}" : "\"";
25185 sv_catpv(msg, prefix);
25186 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8, name_len, name));
25187 sv_catpv(msg, suffix);
25192 definition_deferred:
25195 bool is_qualified = non_pkg_begin != 0; /* If has "::" */
25197 /* Here it could yet to be defined, so defer evaluation of this until
25198 * its needed at runtime. We need the fully qualified property name to
25199 * avoid ambiguity */
25201 fq_name = S_get_fq_name(aTHX_ name, name_len, is_utf8,
25205 /* If it didn't come with a package, or the package is utf8::, this
25206 * actually could be an official Unicode property whose inclusion we
25207 * are deferring until runtime to make sure that it isn't overridden by
25208 * a user-defined property of the same name (which we haven't
25209 * encountered yet). Add a marker to indicate this possibility, for
25210 * use at such time when we first need the definition during pattern
25211 * matching execution */
25212 if (! is_qualified || memBEGINPs(name, non_pkg_begin, "utf8::")) {
25213 sv_catpvs(fq_name, DEFERRED_COULD_BE_OFFICIAL_MARKERs);
25216 /* We also need a trailing newline */
25217 sv_catpvs(fq_name, "\n");
25219 *user_defined_ptr = TRUE;
25225 S_handle_names_wildcard(pTHX_ const char * wname, /* wildcard name to match */
25226 const STRLEN wname_len, /* Its length */
25227 SV ** prop_definition,
25230 /* Deal with Name property wildcard subpatterns; returns TRUE if there were
25231 * any matches, adding them to prop_definition */
25235 CV * get_names_info; /* entry to charnames.pm to get info we need */
25236 SV * names_string; /* Contains all character names, except algo */
25237 SV * algorithmic_names; /* Contains info about algorithmically
25238 generated character names */
25239 REGEXP * subpattern_re; /* The user's pattern to match with */
25240 struct regexp * prog; /* The compiled pattern */
25241 char * all_names_start; /* lib/unicore/Name.pl string of every
25242 (non-algorithmic) character name */
25243 char * cur_pos; /* We match, effectively using /gc; this is
25244 where we are now */
25245 bool found_matches = FALSE; /* Did any name match so far? */
25246 SV * empty; /* For matching zero length names */
25247 SV * must_sv; /* Contains the substring, if any, that must be
25248 in a name for the subpattern to match */
25249 const char * must; /* The PV of 'must' */
25250 STRLEN must_len; /* And its length */
25251 SV * syllable_name = NULL; /* For Hangul syllables */
25252 const char hangul_prefix[] = "HANGUL SYLLABLE ";
25253 const STRLEN hangul_prefix_len = sizeof(hangul_prefix) - 1;
25255 /* By inspection, there are a maximum of 7 bytes in the suffix of a hangul
25256 * syllable name, and these are immutable and guaranteed by the Unicode
25257 * standard to never be extended */
25258 const STRLEN syl_max_len = hangul_prefix_len + 7;
25262 PERL_ARGS_ASSERT_HANDLE_NAMES_WILDCARD;
25264 /* Make sure _charnames is loaded. (The parameters give context
25265 * for any errors generated */
25266 get_names_info = get_cv("_charnames::_get_names_info", 0);
25267 if (! get_names_info) {
25268 Perl_croak(aTHX_ "panic: Can't find '_charnames::_get_names_info");
25271 /* Get the charnames data */
25272 PUSHSTACKi(PERLSI_REGCOMP);
25280 /* Special _charnames entry point that returns the info this routine
25282 call_sv(MUTABLE_SV(get_names_info), G_LIST);
25286 /* Data structure for names which end in their very own code points */
25287 algorithmic_names = POPs;
25288 SvREFCNT_inc_simple_void_NN(algorithmic_names);
25290 /* The lib/unicore/Name.pl string */
25291 names_string = POPs;
25292 SvREFCNT_inc_simple_void_NN(names_string);
25299 if ( ! SvROK(names_string)
25300 || ! SvROK(algorithmic_names))
25301 { /* Perhaps should panic instead XXX */
25302 SvREFCNT_dec(names_string);
25303 SvREFCNT_dec(algorithmic_names);
25307 names_string = sv_2mortal(SvRV(names_string));
25308 all_names_start = SvPVX(names_string);
25309 cur_pos = all_names_start;
25311 algorithmic_names= sv_2mortal(SvRV(algorithmic_names));
25313 /* Compile the subpattern consisting of the name being looked for */
25314 subpattern_re = compile_wildcard(wname, wname_len, FALSE /* /-i */ );
25316 must_sv = re_intuit_string(subpattern_re);
25318 /* regexec.c can free the re_intuit_string() return. GH #17734 */
25319 must_sv = sv_2mortal(newSVsv(must_sv));
25320 must = SvPV(must_sv, must_len);
25327 /* (Note: 'must' could contain a NUL. And yet we use strspn() below on it.
25328 * This works because the NUL causes the function to return early, thus
25329 * showing that there are characters in it other than the acceptable ones,
25330 * which is our desired result.) */
25332 prog = ReANY(subpattern_re);
25334 /* If only nothing is matched, skip to where empty names are looked for */
25335 if (prog->maxlen == 0) {
25339 /* And match against the string of all names /gc. Don't even try if it
25340 * must match a character not found in any name. */
25341 if (strspn(must, "\n -0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ()") == must_len)
25343 while (execute_wildcard(subpattern_re,
25345 SvEND(names_string),
25346 all_names_start, 0,
25349 { /* Here, matched. */
25351 /* Note the string entries look like
25352 * 00001\nSTART OF HEADING\n\n
25353 * so we could match anywhere in that string. We have to rule out
25354 * matching a code point line */
25355 char * this_name_start = all_names_start
25356 + RX_OFFS(subpattern_re)->start;
25357 char * this_name_end = all_names_start
25358 + RX_OFFS(subpattern_re)->end;
25361 UV cp = 0; /* Silences some compilers */
25362 AV * this_string = NULL;
25363 bool is_multi = FALSE;
25365 /* If matched nothing, advance to next possible match */
25366 if (this_name_start == this_name_end) {
25367 cur_pos = (char *) memchr(this_name_end + 1, '\n',
25368 SvEND(names_string) - this_name_end);
25369 if (cur_pos == NULL) {
25374 /* Position the next match to start beyond the current returned
25376 cur_pos = (char *) memchr(this_name_end, '\n',
25377 SvEND(names_string) - this_name_end);
25380 /* Back up to the \n just before the beginning of the character. */
25381 cp_end = (char *) my_memrchr(all_names_start,
25383 this_name_start - all_names_start);
25385 /* If we didn't find a \n, it means it matched somewhere in the
25386 * initial '00000' in the string, so isn't a real match */
25387 if (cp_end == NULL) {
25391 this_name_start = cp_end + 1; /* The name starts just after */
25392 cp_end--; /* the \n, and the code point */
25393 /* ends just before it */
25395 /* All code points are 5 digits long */
25396 cp_start = cp_end - 4;
25398 /* This shouldn't happen, as we found a \n, and the first \n is
25399 * further along than what we subtracted */
25400 assert(cp_start >= all_names_start);
25402 if (cp_start == all_names_start) {
25403 *prop_definition = add_cp_to_invlist(*prop_definition, 0);
25407 /* If the character is a blank, we either have a named sequence, or
25408 * something is wrong */
25409 if (*(cp_start - 1) == ' ') {
25410 cp_start = (char *) my_memrchr(all_names_start,
25412 cp_start - all_names_start);
25416 assert(cp_start != NULL && cp_start >= all_names_start + 2);
25418 /* Except for the first line in the string, the sequence before the
25419 * code point is \n\n. If that isn't the case here, we didn't
25420 * match the name of a character. (We could have matched a named
25421 * sequence, not currently handled */
25422 if (*(cp_start - 1) != '\n' || *(cp_start - 2) != '\n') {
25426 /* We matched! Add this to the list */
25427 found_matches = TRUE;
25429 /* Loop through all the code points in the sequence */
25430 while (cp_start < cp_end) {
25432 /* Calculate this code point from its 5 digits */
25433 cp = (XDIGIT_VALUE(cp_start[0]) << 16)
25434 + (XDIGIT_VALUE(cp_start[1]) << 12)
25435 + (XDIGIT_VALUE(cp_start[2]) << 8)
25436 + (XDIGIT_VALUE(cp_start[3]) << 4)
25437 + XDIGIT_VALUE(cp_start[4]);
25439 cp_start += 6; /* Go past any blank */
25441 if (cp_start < cp_end || is_multi) {
25442 if (this_string == NULL) {
25443 this_string = newAV();
25447 av_push(this_string, newSVuv(cp));
25451 if (is_multi) { /* Was more than one code point */
25452 if (*strings == NULL) {
25453 *strings = newAV();
25456 av_push(*strings, (SV *) this_string);
25458 else { /* Only a single code point */
25459 *prop_definition = add_cp_to_invlist(*prop_definition, cp);
25461 } /* End of loop through the non-algorithmic names string */
25464 /* There are also character names not in 'names_string'. These are
25465 * algorithmically generatable. Try this pattern on each possible one.
25466 * (khw originally planned to leave this out given the large number of
25467 * matches attempted; but the speed turned out to be quite acceptable
25469 * There are plenty of opportunities to optimize to skip many of the tests.
25470 * beyond the rudimentary ones already here */
25472 /* First see if the subpattern matches any of the algorithmic generatable
25473 * Hangul syllable names.
25475 * We know none of these syllable names will match if the input pattern
25476 * requires more bytes than any syllable has, or if the input pattern only
25477 * matches an empty name, or if the pattern has something it must match and
25478 * one of the characters in that isn't in any Hangul syllable. */
25479 if ( prog->minlen <= (SSize_t) syl_max_len
25480 && prog->maxlen > 0
25481 && (strspn(must, "\n ABCDEGHIJKLMNOPRSTUWY") == must_len))
25483 /* These constants, names, values, and algorithm are adapted from the
25484 * Unicode standard, version 5.1, section 3.12, and should never
25486 const char * JamoL[] = {
25487 "G", "GG", "N", "D", "DD", "R", "M", "B", "BB",
25488 "S", "SS", "", "J", "JJ", "C", "K", "T", "P", "H"
25490 const int LCount = C_ARRAY_LENGTH(JamoL);
25492 const char * JamoV[] = {
25493 "A", "AE", "YA", "YAE", "EO", "E", "YEO", "YE", "O", "WA",
25494 "WAE", "OE", "YO", "U", "WEO", "WE", "WI", "YU", "EU", "YI",
25497 const int VCount = C_ARRAY_LENGTH(JamoV);
25499 const char * JamoT[] = {
25500 "", "G", "GG", "GS", "N", "NJ", "NH", "D", "L",
25501 "LG", "LM", "LB", "LS", "LT", "LP", "LH", "M", "B",
25502 "BS", "S", "SS", "NG", "J", "C", "K", "T", "P", "H"
25504 const int TCount = C_ARRAY_LENGTH(JamoT);
25508 /* This is the initial Hangul syllable code point; each time through the
25509 * inner loop, it maps to the next higher code point. For more info,
25510 * see the Hangul syllable section of the Unicode standard. */
25513 syllable_name = sv_2mortal(newSV(syl_max_len));
25514 sv_setpvn(syllable_name, hangul_prefix, hangul_prefix_len);
25516 for (L = 0; L < LCount; L++) {
25517 for (V = 0; V < VCount; V++) {
25518 for (T = 0; T < TCount; T++) {
25520 /* Truncate back to the prefix, which is unvarying */
25521 SvCUR_set(syllable_name, hangul_prefix_len);
25523 sv_catpv(syllable_name, JamoL[L]);
25524 sv_catpv(syllable_name, JamoV[V]);
25525 sv_catpv(syllable_name, JamoT[T]);
25527 if (execute_wildcard(subpattern_re,
25528 SvPVX(syllable_name),
25529 SvEND(syllable_name),
25530 SvPVX(syllable_name), 0,
25534 *prop_definition = add_cp_to_invlist(*prop_definition,
25536 found_matches = TRUE;
25545 /* The rest of the algorithmically generatable names are of the form
25546 * "PREFIX-code_point". The prefixes and the code point limits of each
25547 * were returned to us in the array 'algorithmic_names' from data in
25548 * lib/unicore/Name.pm. 'code_point' in the name is expressed in hex. */
25549 for (i = 0; i <= av_top_index((AV *) algorithmic_names); i++) {
25552 /* Each element of the array is a hash, giving the details for the
25553 * series of names it covers. There is the base name of the characters
25554 * in the series, and the low and high code points in the series. And,
25555 * for optimization purposes a string containing all the legal
25556 * characters that could possibly be in a name in this series. */
25557 HV * this_series = (HV *) SvRV(* av_fetch((AV *) algorithmic_names, i, 0));
25558 SV * prefix = * hv_fetchs(this_series, "name", 0);
25559 IV low = SvIV(* hv_fetchs(this_series, "low", 0));
25560 IV high = SvIV(* hv_fetchs(this_series, "high", 0));
25561 char * legal = SvPVX(* hv_fetchs(this_series, "legal", 0));
25563 /* Pre-allocate an SV with enough space */
25564 SV * algo_name = sv_2mortal(Perl_newSVpvf(aTHX_ "%s-0000",
25566 if (high >= 0x10000) {
25567 sv_catpvs(algo_name, "0");
25570 /* This series can be skipped entirely if the pattern requires
25571 * something longer than any name in the series, or can only match an
25572 * empty name, or contains a character not found in any name in the
25574 if ( prog->minlen <= (SSize_t) SvCUR(algo_name)
25575 && prog->maxlen > 0
25576 && (strspn(must, legal) == must_len))
25578 for (j = low; j <= high; j++) { /* For each code point in the series */
25580 /* Get its name, and see if it matches the subpattern */
25581 Perl_sv_setpvf(aTHX_ algo_name, "%s-%X", SvPVX(prefix),
25584 if (execute_wildcard(subpattern_re,
25587 SvPVX(algo_name), 0,
25591 *prop_definition = add_cp_to_invlist(*prop_definition, j);
25592 found_matches = TRUE;
25599 /* Finally, see if the subpattern matches an empty string */
25600 empty = newSVpvs("");
25601 if (execute_wildcard(subpattern_re,
25608 /* Many code points have empty names. Currently these are the \p{GC=C}
25609 * ones, minus CC and CF */
25611 SV * empty_names_ref = get_prop_definition(UNI_C);
25612 SV * empty_names = invlist_clone(empty_names_ref, NULL);
25614 SV * subtract = get_prop_definition(UNI_CC);
25616 _invlist_subtract(empty_names, subtract, &empty_names);
25617 SvREFCNT_dec_NN(empty_names_ref);
25618 SvREFCNT_dec_NN(subtract);
25620 subtract = get_prop_definition(UNI_CF);
25621 _invlist_subtract(empty_names, subtract, &empty_names);
25622 SvREFCNT_dec_NN(subtract);
25624 _invlist_union(*prop_definition, empty_names, prop_definition);
25625 found_matches = TRUE;
25626 SvREFCNT_dec_NN(empty_names);
25628 SvREFCNT_dec_NN(empty);
25631 /* If we ever were to accept aliases for, say private use names, we would
25632 * need to do something fancier to find empty names. The code below works
25633 * (at the time it was written), and is slower than the above */
25634 const char empties_pat[] = "^.";
25635 if (strNE(name, empties_pat)) {
25636 SV * empty = newSVpvs("");
25637 if (execute_wildcard(subpattern_re,
25644 SV * empties = NULL;
25646 (void) handle_names_wildcard(empties_pat, strlen(empties_pat), &empties);
25648 _invlist_union_complement_2nd(*prop_definition, empties, prop_definition);
25649 SvREFCNT_dec_NN(empties);
25651 found_matches = TRUE;
25653 SvREFCNT_dec_NN(empty);
25657 SvREFCNT_dec_NN(subpattern_re);
25658 return found_matches;
25662 * ex: set ts=8 sts=4 sw=4 et: