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 ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
364 #define ISMULT2(s) (ISMULT1(*s) || ((*s) == '{' && regcurly(s)))
367 * Flags to be passed up and down.
369 #define HASWIDTH 0x01 /* Known to not match null strings, could match
371 #define SIMPLE 0x02 /* Exactly one character wide */
372 /* (or LNBREAK as a special case) */
373 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
374 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
375 #define RESTART_PARSE 0x20 /* Need to redo the parse */
376 #define NEED_UTF8 0x40 /* In conjunction with RESTART_PARSE, need to
377 calcuate sizes as UTF-8 */
379 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
381 /* whether trie related optimizations are enabled */
382 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
383 #define TRIE_STUDY_OPT
384 #define FULL_TRIE_STUDY
390 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
391 #define PBITVAL(paren) (1 << ((paren) & 7))
392 #define PAREN_OFFSET(depth) \
393 (RExC_study_chunk_recursed + (depth) * RExC_study_chunk_recursed_bytes)
394 #define PAREN_TEST(depth, paren) \
395 (PBYTE(PAREN_OFFSET(depth), paren) & PBITVAL(paren))
396 #define PAREN_SET(depth, paren) \
397 (PBYTE(PAREN_OFFSET(depth), paren) |= PBITVAL(paren))
398 #define PAREN_UNSET(depth, paren) \
399 (PBYTE(PAREN_OFFSET(depth), paren) &= ~PBITVAL(paren))
401 #define REQUIRE_UTF8(flagp) STMT_START { \
403 *flagp = RESTART_PARSE|NEED_UTF8; \
408 /* /u is to be chosen if we are supposed to use Unicode rules, or if the
409 * pattern is in UTF-8. This latter condition is in case the outermost rules
410 * are locale. See GH #17278 */
411 #define toUSE_UNI_CHARSET_NOT_DEPENDS (RExC_uni_semantics || UTF)
413 /* Change from /d into /u rules, and restart the parse. RExC_uni_semantics is
414 * a flag that indicates we need to override /d with /u as a result of
415 * something in the pattern. It should only be used in regards to calling
416 * set_regex_charset() or get_regex_charset() */
417 #define REQUIRE_UNI_RULES(flagp, restart_retval) \
419 if (DEPENDS_SEMANTICS) { \
420 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
421 RExC_uni_semantics = 1; \
422 if (RExC_seen_d_op && LIKELY(! IN_PARENS_PASS)) { \
423 /* No need to restart the parse if we haven't seen \
424 * anything that differs between /u and /d, and no need \
425 * to restart immediately if we're going to reparse \
426 * anyway to count parens */ \
427 *flagp |= RESTART_PARSE; \
428 return restart_retval; \
433 #define REQUIRE_BRANCHJ(flagp, restart_retval) \
435 RExC_use_BRANCHJ = 1; \
436 *flagp |= RESTART_PARSE; \
437 return restart_retval; \
440 /* Until we have completed the parse, we leave RExC_total_parens at 0 or
441 * less. After that, it must always be positive, because the whole re is
442 * considered to be surrounded by virtual parens. Setting it to negative
443 * indicates there is some construct that needs to know the actual number of
444 * parens to be properly handled. And that means an extra pass will be
445 * required after we've counted them all */
446 #define ALL_PARENS_COUNTED (RExC_total_parens > 0)
447 #define REQUIRE_PARENS_PASS \
448 STMT_START { /* No-op if have completed a pass */ \
449 if (! ALL_PARENS_COUNTED) RExC_total_parens = -1; \
451 #define IN_PARENS_PASS (RExC_total_parens < 0)
454 /* This is used to return failure (zero) early from the calling function if
455 * various flags in 'flags' are set. Two flags always cause a return:
456 * 'RESTART_PARSE' and 'NEED_UTF8'. 'extra' can be used to specify any
457 * additional flags that should cause a return; 0 if none. If the return will
458 * be done, '*flagp' is first set to be all of the flags that caused the
460 #define RETURN_FAIL_ON_RESTART_OR_FLAGS(flags,flagp,extra) \
462 if ((flags) & (RESTART_PARSE|NEED_UTF8|(extra))) { \
463 *(flagp) = (flags) & (RESTART_PARSE|NEED_UTF8|(extra)); \
468 #define MUST_RESTART(flags) ((flags) & (RESTART_PARSE))
470 #define RETURN_FAIL_ON_RESTART(flags,flagp) \
471 RETURN_FAIL_ON_RESTART_OR_FLAGS( flags, flagp, 0)
472 #define RETURN_FAIL_ON_RESTART_FLAGP(flagp) \
473 if (MUST_RESTART(*(flagp))) return 0
475 /* This converts the named class defined in regcomp.h to its equivalent class
476 * number defined in handy.h. */
477 #define namedclass_to_classnum(class) ((int) ((class) / 2))
478 #define classnum_to_namedclass(classnum) ((classnum) * 2)
480 #define _invlist_union_complement_2nd(a, b, output) \
481 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
482 #define _invlist_intersection_complement_2nd(a, b, output) \
483 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
485 /* We add a marker if we are deferring expansion of a property that is both
486 * 1) potentiallly user-defined; and
487 * 2) could also be an official Unicode property.
489 * Without this marker, any deferred expansion can only be for a user-defined
490 * one. This marker shouldn't conflict with any that could be in a legal name,
491 * and is appended to its name to indicate this. There is a string and
493 #define DEFERRED_COULD_BE_OFFICIAL_MARKERs "~"
494 #define DEFERRED_COULD_BE_OFFICIAL_MARKERc '~'
496 /* What is infinity for optimization purposes */
497 #define OPTIMIZE_INFTY SSize_t_MAX
499 /* About scan_data_t.
501 During optimisation we recurse through the regexp program performing
502 various inplace (keyhole style) optimisations. In addition study_chunk
503 and scan_commit populate this data structure with information about
504 what strings MUST appear in the pattern. We look for the longest
505 string that must appear at a fixed location, and we look for the
506 longest string that may appear at a floating location. So for instance
511 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
512 strings (because they follow a .* construct). study_chunk will identify
513 both FOO and BAR as being the longest fixed and floating strings respectively.
515 The strings can be composites, for instance
519 will result in a composite fixed substring 'foo'.
521 For each string some basic information is maintained:
524 This is the position the string must appear at, or not before.
525 It also implicitly (when combined with minlenp) tells us how many
526 characters must match before the string we are searching for.
527 Likewise when combined with minlenp and the length of the string it
528 tells us how many characters must appear after the string we have
532 Only used for floating strings. This is the rightmost point that
533 the string can appear at. If set to OPTIMIZE_INFTY it indicates that the
534 string can occur infinitely far to the right.
535 For fixed strings, it is equal to min_offset.
538 A pointer to the minimum number of characters of the pattern that the
539 string was found inside. This is important as in the case of positive
540 lookahead or positive lookbehind we can have multiple patterns
545 The minimum length of the pattern overall is 3, the minimum length
546 of the lookahead part is 3, but the minimum length of the part that
547 will actually match is 1. So 'FOO's minimum length is 3, but the
548 minimum length for the F is 1. This is important as the minimum length
549 is used to determine offsets in front of and behind the string being
550 looked for. Since strings can be composites this is the length of the
551 pattern at the time it was committed with a scan_commit. Note that
552 the length is calculated by study_chunk, so that the minimum lengths
553 are not known until the full pattern has been compiled, thus the
554 pointer to the value.
558 In the case of lookbehind the string being searched for can be
559 offset past the start point of the final matching string.
560 If this value was just blithely removed from the min_offset it would
561 invalidate some of the calculations for how many chars must match
562 before or after (as they are derived from min_offset and minlen and
563 the length of the string being searched for).
564 When the final pattern is compiled and the data is moved from the
565 scan_data_t structure into the regexp structure the information
566 about lookbehind is factored in, with the information that would
567 have been lost precalculated in the end_shift field for the
570 The fields pos_min and pos_delta are used to store the minimum offset
571 and the delta to the maximum offset at the current point in the pattern.
575 struct scan_data_substrs {
576 SV *str; /* longest substring found in pattern */
577 SSize_t min_offset; /* earliest point in string it can appear */
578 SSize_t max_offset; /* latest point in string it can appear */
579 SSize_t *minlenp; /* pointer to the minlen relevant to the string */
580 SSize_t lookbehind; /* is the pos of the string modified by LB */
581 I32 flags; /* per substring SF_* and SCF_* flags */
584 typedef struct scan_data_t {
585 /*I32 len_min; unused */
586 /*I32 len_delta; unused */
590 SSize_t last_end; /* min value, <0 unless valid. */
591 SSize_t last_start_min;
592 SSize_t last_start_max;
593 U8 cur_is_floating; /* whether the last_* values should be set as
594 * the next fixed (0) or floating (1)
597 /* [0] is longest fixed substring so far, [1] is longest float so far */
598 struct scan_data_substrs substrs[2];
600 I32 flags; /* common SF_* and SCF_* flags */
602 SSize_t *last_closep;
603 regnode_ssc *start_class;
607 * Forward declarations for pregcomp()'s friends.
610 static const scan_data_t zero_scan_data = {
611 0, 0, NULL, 0, 0, 0, 0,
613 { NULL, 0, 0, 0, 0, 0 },
614 { NULL, 0, 0, 0, 0, 0 },
621 #define SF_BEFORE_SEOL 0x0001
622 #define SF_BEFORE_MEOL 0x0002
623 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
625 #define SF_IS_INF 0x0040
626 #define SF_HAS_PAR 0x0080
627 #define SF_IN_PAR 0x0100
628 #define SF_HAS_EVAL 0x0200
631 /* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
632 * longest substring in the pattern. When it is not set the optimiser keeps
633 * track of position, but does not keep track of the actual strings seen,
635 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
638 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
639 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
640 * turned off because of the alternation (BRANCH). */
641 #define SCF_DO_SUBSTR 0x0400
643 #define SCF_DO_STCLASS_AND 0x0800
644 #define SCF_DO_STCLASS_OR 0x1000
645 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
646 #define SCF_WHILEM_VISITED_POS 0x2000
648 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
649 #define SCF_SEEN_ACCEPT 0x8000
650 #define SCF_TRIE_DOING_RESTUDY 0x10000
651 #define SCF_IN_DEFINE 0x20000
656 #define UTF cBOOL(RExC_utf8)
658 /* The enums for all these are ordered so things work out correctly */
659 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
660 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
661 == REGEX_DEPENDS_CHARSET)
662 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
663 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
664 >= REGEX_UNICODE_CHARSET)
665 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
666 == REGEX_ASCII_RESTRICTED_CHARSET)
667 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
668 >= REGEX_ASCII_RESTRICTED_CHARSET)
669 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
670 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
672 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
674 /* For programs that want to be strictly Unicode compatible by dying if any
675 * attempt is made to match a non-Unicode code point against a Unicode
677 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
679 #define OOB_NAMEDCLASS -1
681 /* There is no code point that is out-of-bounds, so this is problematic. But
682 * its only current use is to initialize a variable that is always set before
684 #define OOB_UNICODE 0xDEADBEEF
686 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
689 /* length of regex to show in messages that don't mark a position within */
690 #define RegexLengthToShowInErrorMessages 127
693 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
694 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
695 * op/pragma/warn/regcomp.
697 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
698 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
700 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
701 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
703 /* The code in this file in places uses one level of recursion with parsing
704 * rebased to an alternate string constructed by us in memory. This can take
705 * the form of something that is completely different from the input, or
706 * something that uses the input as part of the alternate. In the first case,
707 * there should be no possibility of an error, as we are in complete control of
708 * the alternate string. But in the second case we don't completely control
709 * the input portion, so there may be errors in that. Here's an example:
711 * is handled specially because \x{df} folds to a sequence of more than one
712 * character: 'ss'. What is done is to create and parse an alternate string,
713 * which looks like this:
714 * /(?:\x{DF}|[abc\x{DF}def])/ui
715 * where it uses the input unchanged in the middle of something it constructs,
716 * which is a branch for the DF outside the character class, and clustering
717 * parens around the whole thing. (It knows enough to skip the DF inside the
718 * class while in this substitute parse.) 'abc' and 'def' may have errors that
719 * need to be reported. The general situation looks like this:
721 * |<------- identical ------>|
723 * Input: ---------------------------------------------------------------
724 * Constructed: ---------------------------------------------------
726 * |<------- identical ------>|
728 * sI..eI is the portion of the input pattern we are concerned with here.
729 * sC..EC is the constructed substitute parse string.
730 * sC..tC is constructed by us
731 * tC..eC is an exact duplicate of the portion of the input pattern tI..eI.
732 * In the diagram, these are vertically aligned.
733 * eC..EC is also constructed by us.
734 * xC is the position in the substitute parse string where we found a
736 * xI is the position in the original pattern corresponding to xC.
738 * We want to display a message showing the real input string. Thus we need to
739 * translate from xC to xI. We know that xC >= tC, since the portion of the
740 * string sC..tC has been constructed by us, and so shouldn't have errors. We
742 * xI = tI + (xC - tC)
744 * When the substitute parse is constructed, the code needs to set:
747 * RExC_copy_start_in_input (tI)
748 * RExC_copy_start_in_constructed (tC)
749 * and restore them when done.
751 * During normal processing of the input pattern, both
752 * 'RExC_copy_start_in_input' and 'RExC_copy_start_in_constructed' are set to
753 * sI, so that xC equals xI.
756 #define sI RExC_precomp
757 #define eI RExC_precomp_end
758 #define sC RExC_start
760 #define tI RExC_copy_start_in_input
761 #define tC RExC_copy_start_in_constructed
762 #define xI(xC) (tI + (xC - tC))
763 #define xI_offset(xC) (xI(xC) - sI)
765 #define REPORT_LOCATION_ARGS(xC) \
767 (xI(xC) > eI) /* Don't run off end */ \
768 ? eI - sI /* Length before the <--HERE */ \
769 : ((xI_offset(xC) >= 0) \
771 : (Perl_croak(aTHX_ "panic: %s: %d: negative offset: %" \
772 IVdf " trying to output message for " \
774 __FILE__, __LINE__, (IV) xI_offset(xC), \
775 ((int) (eC - sC)), sC), 0)), \
776 sI), /* The input pattern printed up to the <--HERE */ \
778 (xI(xC) > eI) ? 0 : eI - xI(xC), /* Length after <--HERE */ \
779 (xI(xC) > eI) ? eI : xI(xC)) /* pattern after <--HERE */
781 /* Used to point after bad bytes for an error message, but avoid skipping
782 * past a nul byte. */
783 #define SKIP_IF_CHAR(s, e) (!*(s) ? 0 : UTF ? UTF8_SAFE_SKIP(s, e) : 1)
785 /* Set up to clean up after our imminent demise */
786 #define PREPARE_TO_DIE \
789 SAVEFREESV(RExC_rx_sv); \
790 if (RExC_open_parens) \
791 SAVEFREEPV(RExC_open_parens); \
792 if (RExC_close_parens) \
793 SAVEFREEPV(RExC_close_parens); \
797 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
798 * arg. Show regex, up to a maximum length. If it's too long, chop and add
801 #define _FAIL(code) STMT_START { \
802 const char *ellipses = ""; \
803 IV len = RExC_precomp_end - RExC_precomp; \
806 if (len > RegexLengthToShowInErrorMessages) { \
807 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
808 len = RegexLengthToShowInErrorMessages - 10; \
814 #define FAIL(msg) _FAIL( \
815 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
816 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
818 #define FAIL2(msg,arg) _FAIL( \
819 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
820 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
822 #define FAIL3(msg,arg1,arg2) _FAIL( \
823 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
824 arg1, arg2, UTF8fARG(UTF, len, RExC_precomp), ellipses))
827 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
829 #define Simple_vFAIL(m) STMT_START { \
830 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
831 m, REPORT_LOCATION_ARGS(RExC_parse)); \
835 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
837 #define vFAIL(m) STMT_START { \
843 * Like Simple_vFAIL(), but accepts two arguments.
845 #define Simple_vFAIL2(m,a1) STMT_START { \
846 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, \
847 REPORT_LOCATION_ARGS(RExC_parse)); \
851 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
853 #define vFAIL2(m,a1) STMT_START { \
855 Simple_vFAIL2(m, a1); \
860 * Like Simple_vFAIL(), but accepts three arguments.
862 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
863 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, a2, \
864 REPORT_LOCATION_ARGS(RExC_parse)); \
868 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
870 #define vFAIL3(m,a1,a2) STMT_START { \
872 Simple_vFAIL3(m, a1, a2); \
876 * Like Simple_vFAIL(), but accepts four arguments.
878 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
879 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, a2, a3, \
880 REPORT_LOCATION_ARGS(RExC_parse)); \
883 #define vFAIL4(m,a1,a2,a3) STMT_START { \
885 Simple_vFAIL4(m, a1, a2, a3); \
888 /* A specialized version of vFAIL2 that works with UTF8f */
889 #define vFAIL2utf8f(m, a1) STMT_START { \
891 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, \
892 REPORT_LOCATION_ARGS(RExC_parse)); \
895 #define vFAIL3utf8f(m, a1, a2) STMT_START { \
897 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, a2, \
898 REPORT_LOCATION_ARGS(RExC_parse)); \
901 /* Setting this to NULL is a signal to not output warnings */
902 #define TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE \
904 RExC_save_copy_start_in_constructed = RExC_copy_start_in_constructed;\
905 RExC_copy_start_in_constructed = NULL; \
907 #define RESTORE_WARNINGS \
908 RExC_copy_start_in_constructed = RExC_save_copy_start_in_constructed
910 /* Since a warning can be generated multiple times as the input is reparsed, we
911 * output it the first time we come to that point in the parse, but suppress it
912 * otherwise. 'RExC_copy_start_in_constructed' being NULL is a flag to not
913 * generate any warnings */
914 #define TO_OUTPUT_WARNINGS(loc) \
915 ( RExC_copy_start_in_constructed \
916 && ((xI(loc)) - RExC_precomp) > (Ptrdiff_t) RExC_latest_warn_offset)
918 /* After we've emitted a warning, we save the position in the input so we don't
920 #define UPDATE_WARNINGS_LOC(loc) \
922 if (TO_OUTPUT_WARNINGS(loc)) { \
923 RExC_latest_warn_offset = MAX(sI, MIN(eI, xI(loc))) \
928 /* 'warns' is the output of the packWARNx macro used in 'code' */
929 #define _WARN_HELPER(loc, warns, code) \
931 if (! RExC_copy_start_in_constructed) { \
932 Perl_croak( aTHX_ "panic! %s: %d: Tried to warn when none" \
933 " expected at '%s'", \
934 __FILE__, __LINE__, loc); \
936 if (TO_OUTPUT_WARNINGS(loc)) { \
940 UPDATE_WARNINGS_LOC(loc); \
944 /* m is not necessarily a "literal string", in this macro */
945 #define warn_non_literal_string(loc, packed_warn, m) \
946 _WARN_HELPER(loc, packed_warn, \
947 Perl_warner(aTHX_ packed_warn, \
948 "%s" REPORT_LOCATION, \
949 m, REPORT_LOCATION_ARGS(loc)))
950 #define reg_warn_non_literal_string(loc, m) \
951 warn_non_literal_string(loc, packWARN(WARN_REGEXP), m)
953 #define ckWARN2_non_literal_string(loc, packwarn, m, a1) \
956 Size_t format_size = strlen(m) + strlen(REPORT_LOCATION)+ 1;\
957 Newx(format, format_size, char); \
958 my_strlcpy(format, m, format_size); \
959 my_strlcat(format, REPORT_LOCATION, format_size); \
960 SAVEFREEPV(format); \
961 _WARN_HELPER(loc, packwarn, \
962 Perl_ck_warner(aTHX_ packwarn, \
964 a1, REPORT_LOCATION_ARGS(loc))); \
967 #define ckWARNreg(loc,m) \
968 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
969 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
971 REPORT_LOCATION_ARGS(loc)))
973 #define vWARN(loc, m) \
974 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
975 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
977 REPORT_LOCATION_ARGS(loc))) \
979 #define vWARN_dep(loc, m) \
980 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
981 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
983 REPORT_LOCATION_ARGS(loc)))
985 #define ckWARNdep(loc,m) \
986 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
987 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
989 REPORT_LOCATION_ARGS(loc)))
991 #define ckWARNregdep(loc,m) \
992 _WARN_HELPER(loc, packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
993 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
996 REPORT_LOCATION_ARGS(loc)))
998 #define ckWARN2reg_d(loc,m, a1) \
999 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1000 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
1001 m REPORT_LOCATION, \
1002 a1, REPORT_LOCATION_ARGS(loc)))
1004 #define ckWARN2reg(loc, m, a1) \
1005 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1006 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
1007 m REPORT_LOCATION, \
1008 a1, REPORT_LOCATION_ARGS(loc)))
1010 #define vWARN3(loc, m, a1, a2) \
1011 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1012 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
1013 m REPORT_LOCATION, \
1014 a1, a2, REPORT_LOCATION_ARGS(loc)))
1016 #define ckWARN3reg(loc, m, a1, a2) \
1017 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1018 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
1019 m REPORT_LOCATION, \
1021 REPORT_LOCATION_ARGS(loc)))
1023 #define vWARN4(loc, m, a1, a2, a3) \
1024 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1025 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
1026 m REPORT_LOCATION, \
1028 REPORT_LOCATION_ARGS(loc)))
1030 #define ckWARN4reg(loc, m, a1, a2, a3) \
1031 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1032 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
1033 m REPORT_LOCATION, \
1035 REPORT_LOCATION_ARGS(loc)))
1037 #define vWARN5(loc, m, a1, a2, a3, a4) \
1038 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1039 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
1040 m REPORT_LOCATION, \
1042 REPORT_LOCATION_ARGS(loc)))
1044 #define ckWARNexperimental(loc, class, m) \
1046 if (! RExC_warned_ ## class) { /* warn once per compilation */ \
1047 RExC_warned_ ## class = 1; \
1048 _WARN_HELPER(loc, packWARN(class), \
1049 Perl_ck_warner_d(aTHX_ packWARN(class), \
1050 m REPORT_LOCATION, \
1051 REPORT_LOCATION_ARGS(loc)));\
1055 /* Convert between a pointer to a node and its offset from the beginning of the
1057 #define REGNODE_p(offset) (RExC_emit_start + (offset))
1058 #define REGNODE_OFFSET(node) ((node) - RExC_emit_start)
1060 /* Macros for recording node offsets. 20001227 mjd@plover.com
1061 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
1062 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
1063 * Element 0 holds the number n.
1064 * Position is 1 indexed.
1066 #ifndef RE_TRACK_PATTERN_OFFSETS
1067 #define Set_Node_Offset_To_R(offset,byte)
1068 #define Set_Node_Offset(node,byte)
1069 #define Set_Cur_Node_Offset
1070 #define Set_Node_Length_To_R(node,len)
1071 #define Set_Node_Length(node,len)
1072 #define Set_Node_Cur_Length(node,start)
1073 #define Node_Offset(n)
1074 #define Node_Length(n)
1075 #define Set_Node_Offset_Length(node,offset,len)
1076 #define ProgLen(ri) ri->u.proglen
1077 #define SetProgLen(ri,x) ri->u.proglen = x
1078 #define Track_Code(code)
1080 #define ProgLen(ri) ri->u.offsets[0]
1081 #define SetProgLen(ri,x) ri->u.offsets[0] = x
1082 #define Set_Node_Offset_To_R(offset,byte) STMT_START { \
1083 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
1084 __LINE__, (int)(offset), (int)(byte))); \
1085 if((offset) < 0) { \
1086 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
1089 RExC_offsets[2*(offset)-1] = (byte); \
1093 #define Set_Node_Offset(node,byte) \
1094 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (byte)-RExC_start)
1095 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
1097 #define Set_Node_Length_To_R(node,len) STMT_START { \
1098 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
1099 __LINE__, (int)(node), (int)(len))); \
1101 Perl_croak(aTHX_ "value of node is %d in Length macro", \
1104 RExC_offsets[2*(node)] = (len); \
1108 #define Set_Node_Length(node,len) \
1109 Set_Node_Length_To_R(REGNODE_OFFSET(node), len)
1110 #define Set_Node_Cur_Length(node, start) \
1111 Set_Node_Length(node, RExC_parse - start)
1113 /* Get offsets and lengths */
1114 #define Node_Offset(n) (RExC_offsets[2*(REGNODE_OFFSET(n))-1])
1115 #define Node_Length(n) (RExC_offsets[2*(REGNODE_OFFSET(n))])
1117 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
1118 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (offset)); \
1119 Set_Node_Length_To_R(REGNODE_OFFSET(node), (len)); \
1122 #define Track_Code(code) STMT_START { code } STMT_END
1125 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
1126 #define EXPERIMENTAL_INPLACESCAN
1127 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
1131 Perl_re_printf(pTHX_ const char *fmt, ...)
1135 PerlIO *f= Perl_debug_log;
1136 PERL_ARGS_ASSERT_RE_PRINTF;
1138 result = PerlIO_vprintf(f, fmt, ap);
1144 Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
1148 PerlIO *f= Perl_debug_log;
1149 PERL_ARGS_ASSERT_RE_INDENTF;
1150 va_start(ap, depth);
1151 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
1152 result = PerlIO_vprintf(f, fmt, ap);
1156 #endif /* DEBUGGING */
1158 #define DEBUG_RExC_seen() \
1159 DEBUG_OPTIMISE_MORE_r({ \
1160 Perl_re_printf( aTHX_ "RExC_seen: "); \
1162 if (RExC_seen & REG_ZERO_LEN_SEEN) \
1163 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
1165 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
1166 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
1168 if (RExC_seen & REG_GPOS_SEEN) \
1169 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
1171 if (RExC_seen & REG_RECURSE_SEEN) \
1172 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
1174 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
1175 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
1177 if (RExC_seen & REG_VERBARG_SEEN) \
1178 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
1180 if (RExC_seen & REG_CUTGROUP_SEEN) \
1181 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
1183 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
1184 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
1186 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
1187 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
1189 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
1190 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
1192 Perl_re_printf( aTHX_ "\n"); \
1195 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
1196 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
1201 S_debug_show_study_flags(pTHX_ U32 flags, const char *open_str,
1202 const char *close_str)
1207 Perl_re_printf( aTHX_ "%s", open_str);
1208 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_SEOL);
1209 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_MEOL);
1210 DEBUG_SHOW_STUDY_FLAG(flags, SF_IS_INF);
1211 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_PAR);
1212 DEBUG_SHOW_STUDY_FLAG(flags, SF_IN_PAR);
1213 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_EVAL);
1214 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_SUBSTR);
1215 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_AND);
1216 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_OR);
1217 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS);
1218 DEBUG_SHOW_STUDY_FLAG(flags, SCF_WHILEM_VISITED_POS);
1219 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_RESTUDY);
1220 DEBUG_SHOW_STUDY_FLAG(flags, SCF_SEEN_ACCEPT);
1221 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_DOING_RESTUDY);
1222 DEBUG_SHOW_STUDY_FLAG(flags, SCF_IN_DEFINE);
1223 Perl_re_printf( aTHX_ "%s", close_str);
1228 S_debug_studydata(pTHX_ const char *where, scan_data_t *data,
1229 U32 depth, int is_inf)
1231 DECLARE_AND_GET_RE_DEBUG_FLAGS;
1233 DEBUG_OPTIMISE_MORE_r({
1236 Perl_re_indentf(aTHX_ "%s: Pos:%" IVdf "/%" IVdf " Flags: 0x%" UVXf,
1240 (IV)data->pos_delta,
1244 S_debug_show_study_flags(aTHX_ data->flags," [","]");
1246 Perl_re_printf( aTHX_
1247 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s",
1249 (IV)(data->last_closep ? *((data)->last_closep) : -1),
1250 is_inf ? "INF " : ""
1253 if (data->last_found) {
1255 Perl_re_printf(aTHX_
1256 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf,
1257 SvPVX_const(data->last_found),
1259 (IV)data->last_start_min,
1260 (IV)data->last_start_max
1263 for (i = 0; i < 2; i++) {
1264 Perl_re_printf(aTHX_
1265 " %s%s: '%s' @ %" IVdf "/%" IVdf,
1266 data->cur_is_floating == i ? "*" : "",
1267 i ? "Float" : "Fixed",
1268 SvPVX_const(data->substrs[i].str),
1269 (IV)data->substrs[i].min_offset,
1270 (IV)data->substrs[i].max_offset
1272 S_debug_show_study_flags(aTHX_ data->substrs[i].flags," [","]");
1276 Perl_re_printf( aTHX_ "\n");
1282 S_debug_peep(pTHX_ const char *str, const RExC_state_t *pRExC_state,
1283 regnode *scan, U32 depth, U32 flags)
1285 DECLARE_AND_GET_RE_DEBUG_FLAGS;
1292 Next = regnext(scan);
1293 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
1294 Perl_re_indentf( aTHX_ "%s>%3d: %s (%d)",
1297 REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),
1298 Next ? (REG_NODE_NUM(Next)) : 0 );
1299 S_debug_show_study_flags(aTHX_ flags," [ ","]");
1300 Perl_re_printf( aTHX_ "\n");
1305 # define DEBUG_STUDYDATA(where, data, depth, is_inf) \
1306 S_debug_studydata(aTHX_ where, data, depth, is_inf)
1308 # define DEBUG_PEEP(str, scan, depth, flags) \
1309 S_debug_peep(aTHX_ str, pRExC_state, scan, depth, flags)
1312 # define DEBUG_STUDYDATA(where, data, depth, is_inf) NOOP
1313 # define DEBUG_PEEP(str, scan, depth, flags) NOOP
1317 /* =========================================================
1318 * BEGIN edit_distance stuff.
1320 * This calculates how many single character changes of any type are needed to
1321 * transform a string into another one. It is taken from version 3.1 of
1323 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1326 /* Our unsorted dictionary linked list. */
1327 /* Note we use UVs, not chars. */
1332 struct dictionary* next;
1334 typedef struct dictionary item;
1337 PERL_STATIC_INLINE item*
1338 push(UV key, item* curr)
1341 Newx(head, 1, item);
1349 PERL_STATIC_INLINE item*
1350 find(item* head, UV key)
1352 item* iterator = head;
1354 if (iterator->key == key){
1357 iterator = iterator->next;
1363 PERL_STATIC_INLINE item*
1364 uniquePush(item* head, UV key)
1366 item* iterator = head;
1369 if (iterator->key == key) {
1372 iterator = iterator->next;
1375 return push(key, head);
1378 PERL_STATIC_INLINE void
1379 dict_free(item* head)
1381 item* iterator = head;
1384 item* temp = iterator;
1385 iterator = iterator->next;
1392 /* End of Dictionary Stuff */
1394 /* All calculations/work are done here */
1396 S_edit_distance(const UV* src,
1398 const STRLEN x, /* length of src[] */
1399 const STRLEN y, /* length of tgt[] */
1400 const SSize_t maxDistance
1404 UV swapCount, swapScore, targetCharCount, i, j;
1406 UV score_ceil = x + y;
1408 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1410 /* intialize matrix start values */
1411 Newx(scores, ( (x + 2) * (y + 2)), UV);
1412 scores[0] = score_ceil;
1413 scores[1 * (y + 2) + 0] = score_ceil;
1414 scores[0 * (y + 2) + 1] = score_ceil;
1415 scores[1 * (y + 2) + 1] = 0;
1416 head = uniquePush(uniquePush(head, src[0]), tgt[0]);
1421 for (i=1;i<=x;i++) {
1423 head = uniquePush(head, src[i]);
1424 scores[(i+1) * (y + 2) + 1] = i;
1425 scores[(i+1) * (y + 2) + 0] = score_ceil;
1428 for (j=1;j<=y;j++) {
1431 head = uniquePush(head, tgt[j]);
1432 scores[1 * (y + 2) + (j + 1)] = j;
1433 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1436 targetCharCount = find(head, tgt[j-1])->value;
1437 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1439 if (src[i-1] != tgt[j-1]){
1440 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));
1444 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1448 find(head, src[i-1])->value = i;
1452 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1455 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1459 /* END of edit_distance() stuff
1460 * ========================================================= */
1462 /* Mark that we cannot extend a found fixed substring at this point.
1463 Update the longest found anchored substring or the longest found
1464 floating substrings if needed. */
1467 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1468 SSize_t *minlenp, int is_inf)
1470 const STRLEN l = CHR_SVLEN(data->last_found);
1471 SV * const longest_sv = data->substrs[data->cur_is_floating].str;
1472 const STRLEN old_l = CHR_SVLEN(longest_sv);
1473 DECLARE_AND_GET_RE_DEBUG_FLAGS;
1475 PERL_ARGS_ASSERT_SCAN_COMMIT;
1477 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
1478 const U8 i = data->cur_is_floating;
1479 SvSetMagicSV(longest_sv, data->last_found);
1480 data->substrs[i].min_offset = l ? data->last_start_min : data->pos_min;
1483 data->substrs[0].max_offset = data->substrs[0].min_offset;
1485 data->substrs[1].max_offset =
1489 ? data->last_start_max
1490 /* temporary underflow guard for 5.32 */
1491 : data->pos_delta < 0 ? OPTIMIZE_INFTY
1492 : (data->pos_delta > OPTIMIZE_INFTY - data->pos_min
1494 : data->pos_min + data->pos_delta));
1497 data->substrs[i].flags &= ~SF_BEFORE_EOL;
1498 data->substrs[i].flags |= data->flags & SF_BEFORE_EOL;
1499 data->substrs[i].minlenp = minlenp;
1500 data->substrs[i].lookbehind = 0;
1503 SvCUR_set(data->last_found, 0);
1505 SV * const sv = data->last_found;
1506 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1507 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1512 data->last_end = -1;
1513 data->flags &= ~SF_BEFORE_EOL;
1514 DEBUG_STUDYDATA("commit", data, 0, is_inf);
1517 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1518 * list that describes which code points it matches */
1521 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1523 /* Set the SSC 'ssc' to match an empty string or any code point */
1525 PERL_ARGS_ASSERT_SSC_ANYTHING;
1527 assert(is_ANYOF_SYNTHETIC(ssc));
1529 /* mortalize so won't leak */
1530 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
1531 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1535 S_ssc_is_anything(const regnode_ssc *ssc)
1537 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1538 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1539 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1540 * in any way, so there's no point in using it */
1545 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1547 assert(is_ANYOF_SYNTHETIC(ssc));
1549 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1553 /* See if the list consists solely of the range 0 - Infinity */
1554 invlist_iterinit(ssc->invlist);
1555 ret = invlist_iternext(ssc->invlist, &start, &end)
1559 invlist_iterfinish(ssc->invlist);
1565 /* If e.g., both \w and \W are set, matches everything */
1566 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1568 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1569 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1579 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1581 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1582 * string, any code point, or any posix class under locale */
1584 PERL_ARGS_ASSERT_SSC_INIT;
1586 Zero(ssc, 1, regnode_ssc);
1587 set_ANYOF_SYNTHETIC(ssc);
1588 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1591 /* If any portion of the regex is to operate under locale rules that aren't
1592 * fully known at compile time, initialization includes it. The reason
1593 * this isn't done for all regexes is that the optimizer was written under
1594 * the assumption that locale was all-or-nothing. Given the complexity and
1595 * lack of documentation in the optimizer, and that there are inadequate
1596 * test cases for locale, many parts of it may not work properly, it is
1597 * safest to avoid locale unless necessary. */
1598 if (RExC_contains_locale) {
1599 ANYOF_POSIXL_SETALL(ssc);
1602 ANYOF_POSIXL_ZERO(ssc);
1607 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1608 const regnode_ssc *ssc)
1610 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1611 * to the list of code points matched, and locale posix classes; hence does
1612 * not check its flags) */
1617 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1619 assert(is_ANYOF_SYNTHETIC(ssc));
1621 invlist_iterinit(ssc->invlist);
1622 ret = invlist_iternext(ssc->invlist, &start, &end)
1626 invlist_iterfinish(ssc->invlist);
1632 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1639 #define INVLIST_INDEX 0
1640 #define ONLY_LOCALE_MATCHES_INDEX 1
1641 #define DEFERRED_USER_DEFINED_INDEX 2
1644 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1645 const regnode_charclass* const node)
1647 /* Returns a mortal inversion list defining which code points are matched
1648 * by 'node', which is of type ANYOF. Handles complementing the result if
1649 * appropriate. If some code points aren't knowable at this time, the
1650 * returned list must, and will, contain every code point that is a
1654 SV* only_utf8_locale_invlist = NULL;
1656 const U32 n = ARG(node);
1657 bool new_node_has_latin1 = FALSE;
1658 const U8 flags = (inRANGE(OP(node), ANYOFH, ANYOFRb))
1660 : ANYOF_FLAGS(node);
1662 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1664 /* Look at the data structure created by S_set_ANYOF_arg() */
1665 if (n != ANYOF_ONLY_HAS_BITMAP) {
1666 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1667 AV * const av = MUTABLE_AV(SvRV(rv));
1668 SV **const ary = AvARRAY(av);
1669 assert(RExC_rxi->data->what[n] == 's');
1671 if (av_tindex_skip_len_mg(av) >= DEFERRED_USER_DEFINED_INDEX) {
1673 /* Here there are things that won't be known until runtime -- we
1674 * have to assume it could be anything */
1675 invlist = sv_2mortal(_new_invlist(1));
1676 return _add_range_to_invlist(invlist, 0, UV_MAX);
1678 else if (ary[INVLIST_INDEX]) {
1680 /* Use the node's inversion list */
1681 invlist = sv_2mortal(invlist_clone(ary[INVLIST_INDEX], NULL));
1684 /* Get the code points valid only under UTF-8 locales */
1685 if ( (flags & ANYOFL_FOLD)
1686 && av_tindex_skip_len_mg(av) >= ONLY_LOCALE_MATCHES_INDEX)
1688 only_utf8_locale_invlist = ary[ONLY_LOCALE_MATCHES_INDEX];
1693 invlist = sv_2mortal(_new_invlist(0));
1696 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1697 * code points, and an inversion list for the others, but if there are code
1698 * points that should match only conditionally on the target string being
1699 * UTF-8, those are placed in the inversion list, and not the bitmap.
1700 * Since there are circumstances under which they could match, they are
1701 * included in the SSC. But if the ANYOF node is to be inverted, we have
1702 * to exclude them here, so that when we invert below, the end result
1703 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1704 * have to do this here before we add the unconditionally matched code
1706 if (flags & ANYOF_INVERT) {
1707 _invlist_intersection_complement_2nd(invlist,
1712 /* Add in the points from the bit map */
1713 if (! inRANGE(OP(node), ANYOFH, ANYOFRb)) {
1714 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1715 if (ANYOF_BITMAP_TEST(node, i)) {
1716 unsigned int start = i++;
1718 for (; i < NUM_ANYOF_CODE_POINTS
1719 && ANYOF_BITMAP_TEST(node, i); ++i)
1723 invlist = _add_range_to_invlist(invlist, start, i-1);
1724 new_node_has_latin1 = TRUE;
1729 /* If this can match all upper Latin1 code points, have to add them
1730 * as well. But don't add them if inverting, as when that gets done below,
1731 * it would exclude all these characters, including the ones it shouldn't
1732 * that were added just above */
1733 if (! (flags & ANYOF_INVERT) && OP(node) == ANYOFD
1734 && (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1736 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1739 /* Similarly for these */
1740 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1741 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1744 if (flags & ANYOF_INVERT) {
1745 _invlist_invert(invlist);
1747 else if (flags & ANYOFL_FOLD) {
1748 if (new_node_has_latin1) {
1750 /* Under /li, any 0-255 could fold to any other 0-255, depending on
1751 * the locale. We can skip this if there are no 0-255 at all. */
1752 _invlist_union(invlist, PL_Latin1, &invlist);
1754 invlist = add_cp_to_invlist(invlist, LATIN_SMALL_LETTER_DOTLESS_I);
1755 invlist = add_cp_to_invlist(invlist, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
1758 if (_invlist_contains_cp(invlist, LATIN_SMALL_LETTER_DOTLESS_I)) {
1759 invlist = add_cp_to_invlist(invlist, 'I');
1761 if (_invlist_contains_cp(invlist,
1762 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE))
1764 invlist = add_cp_to_invlist(invlist, 'i');
1769 /* Similarly add the UTF-8 locale possible matches. These have to be
1770 * deferred until after the non-UTF-8 locale ones are taken care of just
1771 * above, or it leads to wrong results under ANYOF_INVERT */
1772 if (only_utf8_locale_invlist) {
1773 _invlist_union_maybe_complement_2nd(invlist,
1774 only_utf8_locale_invlist,
1775 flags & ANYOF_INVERT,
1782 /* These two functions currently do the exact same thing */
1783 #define ssc_init_zero ssc_init
1785 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1786 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1788 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1789 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1790 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1793 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1794 const regnode_charclass *and_with)
1796 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1797 * another SSC or a regular ANYOF class. Can create false positives. */
1800 U8 and_with_flags = inRANGE(OP(and_with), ANYOFH, ANYOFRb)
1802 : ANYOF_FLAGS(and_with);
1805 PERL_ARGS_ASSERT_SSC_AND;
1807 assert(is_ANYOF_SYNTHETIC(ssc));
1809 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1810 * the code point inversion list and just the relevant flags */
1811 if (is_ANYOF_SYNTHETIC(and_with)) {
1812 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1813 anded_flags = and_with_flags;
1815 /* XXX This is a kludge around what appears to be deficiencies in the
1816 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1817 * there are paths through the optimizer where it doesn't get weeded
1818 * out when it should. And if we don't make some extra provision for
1819 * it like the code just below, it doesn't get added when it should.
1820 * This solution is to add it only when AND'ing, which is here, and
1821 * only when what is being AND'ed is the pristine, original node
1822 * matching anything. Thus it is like adding it to ssc_anything() but
1823 * only when the result is to be AND'ed. Probably the same solution
1824 * could be adopted for the same problem we have with /l matching,
1825 * which is solved differently in S_ssc_init(), and that would lead to
1826 * fewer false positives than that solution has. But if this solution
1827 * creates bugs, the consequences are only that a warning isn't raised
1828 * that should be; while the consequences for having /l bugs is
1829 * incorrect matches */
1830 if (ssc_is_anything((regnode_ssc *)and_with)) {
1831 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1835 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1836 if (OP(and_with) == ANYOFD) {
1837 anded_flags = and_with_flags & ANYOF_COMMON_FLAGS;
1840 anded_flags = and_with_flags
1841 &( ANYOF_COMMON_FLAGS
1842 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1843 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1844 if (ANYOFL_UTF8_LOCALE_REQD(and_with_flags)) {
1846 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1851 ANYOF_FLAGS(ssc) &= anded_flags;
1853 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1854 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1855 * 'and_with' may be inverted. When not inverted, we have the situation of
1857 * (C1 | P1) & (C2 | P2)
1858 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1859 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1860 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1861 * <= ((C1 & C2) | P1 | P2)
1862 * Alternatively, the last few steps could be:
1863 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1864 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1865 * <= (C1 | C2 | (P1 & P2))
1866 * We favor the second approach if either P1 or P2 is non-empty. This is
1867 * because these components are a barrier to doing optimizations, as what
1868 * they match cannot be known until the moment of matching as they are
1869 * dependent on the current locale, 'AND"ing them likely will reduce or
1871 * But we can do better if we know that C1,P1 are in their initial state (a
1872 * frequent occurrence), each matching everything:
1873 * (<everything>) & (C2 | P2) = C2 | P2
1874 * Similarly, if C2,P2 are in their initial state (again a frequent
1875 * occurrence), the result is a no-op
1876 * (C1 | P1) & (<everything>) = C1 | P1
1879 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1880 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1881 * <= (C1 & ~C2) | (P1 & ~P2)
1884 if ((and_with_flags & ANYOF_INVERT)
1885 && ! is_ANYOF_SYNTHETIC(and_with))
1889 ssc_intersection(ssc,
1891 FALSE /* Has already been inverted */
1894 /* If either P1 or P2 is empty, the intersection will be also; can skip
1896 if (! (and_with_flags & ANYOF_MATCHES_POSIXL)) {
1897 ANYOF_POSIXL_ZERO(ssc);
1899 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1901 /* Note that the Posix class component P from 'and_with' actually
1903 * P = Pa | Pb | ... | Pn
1904 * where each component is one posix class, such as in [\w\s].
1906 * ~P = ~(Pa | Pb | ... | Pn)
1907 * = ~Pa & ~Pb & ... & ~Pn
1908 * <= ~Pa | ~Pb | ... | ~Pn
1909 * The last is something we can easily calculate, but unfortunately
1910 * is likely to have many false positives. We could do better
1911 * in some (but certainly not all) instances if two classes in
1912 * P have known relationships. For example
1913 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1915 * :lower: & :print: = :lower:
1916 * And similarly for classes that must be disjoint. For example,
1917 * since \s and \w can have no elements in common based on rules in
1918 * the POSIX standard,
1919 * \w & ^\S = nothing
1920 * Unfortunately, some vendor locales do not meet the Posix
1921 * standard, in particular almost everything by Microsoft.
1922 * The loop below just changes e.g., \w into \W and vice versa */
1924 regnode_charclass_posixl temp;
1925 int add = 1; /* To calculate the index of the complement */
1927 Zero(&temp, 1, regnode_charclass_posixl);
1928 ANYOF_POSIXL_ZERO(&temp);
1929 for (i = 0; i < ANYOF_MAX; i++) {
1931 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1932 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1934 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1935 ANYOF_POSIXL_SET(&temp, i + add);
1937 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1939 ANYOF_POSIXL_AND(&temp, ssc);
1941 } /* else ssc already has no posixes */
1942 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1943 in its initial state */
1944 else if (! is_ANYOF_SYNTHETIC(and_with)
1945 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1947 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1948 * copy it over 'ssc' */
1949 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1950 if (is_ANYOF_SYNTHETIC(and_with)) {
1951 StructCopy(and_with, ssc, regnode_ssc);
1954 ssc->invlist = anded_cp_list;
1955 ANYOF_POSIXL_ZERO(ssc);
1956 if (and_with_flags & ANYOF_MATCHES_POSIXL) {
1957 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1961 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1962 || (and_with_flags & ANYOF_MATCHES_POSIXL))
1964 /* One or the other of P1, P2 is non-empty. */
1965 if (and_with_flags & ANYOF_MATCHES_POSIXL) {
1966 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1968 ssc_union(ssc, anded_cp_list, FALSE);
1970 else { /* P1 = P2 = empty */
1971 ssc_intersection(ssc, anded_cp_list, FALSE);
1977 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1978 const regnode_charclass *or_with)
1980 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1981 * another SSC or a regular ANYOF class. Can create false positives if
1982 * 'or_with' is to be inverted. */
1986 U8 or_with_flags = inRANGE(OP(or_with), ANYOFH, ANYOFRb)
1988 : ANYOF_FLAGS(or_with);
1990 PERL_ARGS_ASSERT_SSC_OR;
1992 assert(is_ANYOF_SYNTHETIC(ssc));
1994 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1995 * the code point inversion list and just the relevant flags */
1996 if (is_ANYOF_SYNTHETIC(or_with)) {
1997 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1998 ored_flags = or_with_flags;
2001 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
2002 ored_flags = or_with_flags & ANYOF_COMMON_FLAGS;
2003 if (OP(or_with) != ANYOFD) {
2006 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
2007 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
2008 if (ANYOFL_UTF8_LOCALE_REQD(or_with_flags)) {
2010 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
2015 ANYOF_FLAGS(ssc) |= ored_flags;
2017 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
2018 * C2 is the list of code points in 'or-with'; P2, its posix classes.
2019 * 'or_with' may be inverted. When not inverted, we have the simple
2020 * situation of computing:
2021 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
2022 * If P1|P2 yields a situation with both a class and its complement are
2023 * set, like having both \w and \W, this matches all code points, and we
2024 * can delete these from the P component of the ssc going forward. XXX We
2025 * might be able to delete all the P components, but I (khw) am not certain
2026 * about this, and it is better to be safe.
2029 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
2030 * <= (C1 | P1) | ~C2
2031 * <= (C1 | ~C2) | P1
2032 * (which results in actually simpler code than the non-inverted case)
2035 if ((or_with_flags & ANYOF_INVERT)
2036 && ! is_ANYOF_SYNTHETIC(or_with))
2038 /* We ignore P2, leaving P1 going forward */
2039 } /* else Not inverted */
2040 else if (or_with_flags & ANYOF_MATCHES_POSIXL) {
2041 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
2042 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
2044 for (i = 0; i < ANYOF_MAX; i += 2) {
2045 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
2047 ssc_match_all_cp(ssc);
2048 ANYOF_POSIXL_CLEAR(ssc, i);
2049 ANYOF_POSIXL_CLEAR(ssc, i+1);
2057 FALSE /* Already has been inverted */
2062 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
2064 PERL_ARGS_ASSERT_SSC_UNION;
2066 assert(is_ANYOF_SYNTHETIC(ssc));
2068 _invlist_union_maybe_complement_2nd(ssc->invlist,
2075 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
2077 const bool invert2nd)
2079 PERL_ARGS_ASSERT_SSC_INTERSECTION;
2081 assert(is_ANYOF_SYNTHETIC(ssc));
2083 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
2090 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
2092 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
2094 assert(is_ANYOF_SYNTHETIC(ssc));
2096 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
2100 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
2102 /* AND just the single code point 'cp' into the SSC 'ssc' */
2104 SV* cp_list = _new_invlist(2);
2106 PERL_ARGS_ASSERT_SSC_CP_AND;
2108 assert(is_ANYOF_SYNTHETIC(ssc));
2110 cp_list = add_cp_to_invlist(cp_list, cp);
2111 ssc_intersection(ssc, cp_list,
2112 FALSE /* Not inverted */
2114 SvREFCNT_dec_NN(cp_list);
2118 S_ssc_clear_locale(regnode_ssc *ssc)
2120 /* Set the SSC 'ssc' to not match any locale things */
2121 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
2123 assert(is_ANYOF_SYNTHETIC(ssc));
2125 ANYOF_POSIXL_ZERO(ssc);
2126 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
2130 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
2132 /* The synthetic start class is used to hopefully quickly winnow down
2133 * places where a pattern could start a match in the target string. If it
2134 * doesn't really narrow things down that much, there isn't much point to
2135 * having the overhead of using it. This function uses some very crude
2136 * heuristics to decide if to use the ssc or not.
2138 * It returns TRUE if 'ssc' rules out more than half what it considers to
2139 * be the "likely" possible matches, but of course it doesn't know what the
2140 * actual things being matched are going to be; these are only guesses
2142 * For /l matches, it assumes that the only likely matches are going to be
2143 * in the 0-255 range, uniformly distributed, so half of that is 127
2144 * For /a and /d matches, it assumes that the likely matches will be just
2145 * the ASCII range, so half of that is 63
2146 * For /u and there isn't anything matching above the Latin1 range, it
2147 * assumes that that is the only range likely to be matched, and uses
2148 * half that as the cut-off: 127. If anything matches above Latin1,
2149 * it assumes that all of Unicode could match (uniformly), except for
2150 * non-Unicode code points and things in the General Category "Other"
2151 * (unassigned, private use, surrogates, controls and formats). This
2152 * is a much large number. */
2154 U32 count = 0; /* Running total of number of code points matched by
2156 UV start, end; /* Start and end points of current range in inversion
2157 XXX outdated. UTF-8 locales are common, what about invert? list */
2158 const U32 max_code_points = (LOC)
2160 : (( ! UNI_SEMANTICS
2161 || invlist_highest(ssc->invlist) < 256)
2164 const U32 max_match = max_code_points / 2;
2166 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
2168 invlist_iterinit(ssc->invlist);
2169 while (invlist_iternext(ssc->invlist, &start, &end)) {
2170 if (start >= max_code_points) {
2173 end = MIN(end, max_code_points - 1);
2174 count += end - start + 1;
2175 if (count >= max_match) {
2176 invlist_iterfinish(ssc->invlist);
2186 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
2188 /* The inversion list in the SSC is marked mortal; now we need a more
2189 * permanent copy, which is stored the same way that is done in a regular
2190 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
2193 SV* invlist = invlist_clone(ssc->invlist, NULL);
2195 PERL_ARGS_ASSERT_SSC_FINALIZE;
2197 assert(is_ANYOF_SYNTHETIC(ssc));
2199 /* The code in this file assumes that all but these flags aren't relevant
2200 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
2201 * by the time we reach here */
2202 assert(! (ANYOF_FLAGS(ssc)
2203 & ~( ANYOF_COMMON_FLAGS
2204 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
2205 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
2207 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
2209 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist, NULL, NULL);
2210 SvREFCNT_dec(invlist);
2212 /* Make sure is clone-safe */
2213 ssc->invlist = NULL;
2215 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
2216 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
2217 OP(ssc) = ANYOFPOSIXL;
2219 else if (RExC_contains_locale) {
2223 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
2226 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
2227 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
2228 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
2229 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
2230 ? (TRIE_LIST_CUR( idx ) - 1) \
2236 dump_trie(trie,widecharmap,revcharmap)
2237 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
2238 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
2240 These routines dump out a trie in a somewhat readable format.
2241 The _interim_ variants are used for debugging the interim
2242 tables that are used to generate the final compressed
2243 representation which is what dump_trie expects.
2245 Part of the reason for their existence is to provide a form
2246 of documentation as to how the different representations function.
2251 Dumps the final compressed table form of the trie to Perl_debug_log.
2252 Used for debugging make_trie().
2256 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
2257 AV *revcharmap, U32 depth)
2260 SV *sv=sv_newmortal();
2261 int colwidth= widecharmap ? 6 : 4;
2263 DECLARE_AND_GET_RE_DEBUG_FLAGS;
2265 PERL_ARGS_ASSERT_DUMP_TRIE;
2267 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
2268 depth+1, "Match","Base","Ofs" );
2270 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2271 SV ** const tmp = av_fetch( revcharmap, state, 0);
2273 Perl_re_printf( aTHX_ "%*s",
2275 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2276 PL_colors[0], PL_colors[1],
2277 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2278 PERL_PV_ESCAPE_FIRSTCHAR
2283 Perl_re_printf( aTHX_ "\n");
2284 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
2286 for( state = 0 ; state < trie->uniquecharcount ; state++ )
2287 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2288 Perl_re_printf( aTHX_ "\n");
2290 for( state = 1 ; state < trie->statecount ; state++ ) {
2291 const U32 base = trie->states[ state ].trans.base;
2293 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
2295 if ( trie->states[ state ].wordnum ) {
2296 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
2298 Perl_re_printf( aTHX_ "%6s", "" );
2301 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
2306 while( ( base + ofs < trie->uniquecharcount ) ||
2307 ( base + ofs - trie->uniquecharcount < trie->lasttrans
2308 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2312 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
2314 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2315 if ( ( base + ofs >= trie->uniquecharcount )
2316 && ( base + ofs - trie->uniquecharcount
2318 && trie->trans[ base + ofs
2319 - trie->uniquecharcount ].check == state )
2321 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
2322 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2325 Perl_re_printf( aTHX_ "%*s", colwidth," ." );
2329 Perl_re_printf( aTHX_ "]");
2332 Perl_re_printf( aTHX_ "\n" );
2334 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
2336 for (word=1; word <= trie->wordcount; word++) {
2337 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2338 (int)word, (int)(trie->wordinfo[word].prev),
2339 (int)(trie->wordinfo[word].len));
2341 Perl_re_printf( aTHX_ "\n" );
2344 Dumps a fully constructed but uncompressed trie in list form.
2345 List tries normally only are used for construction when the number of
2346 possible chars (trie->uniquecharcount) is very high.
2347 Used for debugging make_trie().
2350 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2351 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2355 SV *sv=sv_newmortal();
2356 int colwidth= widecharmap ? 6 : 4;
2357 DECLARE_AND_GET_RE_DEBUG_FLAGS;
2359 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2361 /* print out the table precompression. */
2362 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
2364 Perl_re_indentf( aTHX_ "%s",
2365 depth+1, "------:-----+-----------------\n" );
2367 for( state=1 ; state < next_alloc ; state ++ ) {
2370 Perl_re_indentf( aTHX_ " %4" UVXf " :",
2371 depth+1, (UV)state );
2372 if ( ! trie->states[ state ].wordnum ) {
2373 Perl_re_printf( aTHX_ "%5s| ","");
2375 Perl_re_printf( aTHX_ "W%4x| ",
2376 trie->states[ state ].wordnum
2379 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2380 SV ** const tmp = av_fetch( revcharmap,
2381 TRIE_LIST_ITEM(state, charid).forid, 0);
2383 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
2385 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2387 PL_colors[0], PL_colors[1],
2388 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2389 | PERL_PV_ESCAPE_FIRSTCHAR
2391 TRIE_LIST_ITEM(state, charid).forid,
2392 (UV)TRIE_LIST_ITEM(state, charid).newstate
2395 Perl_re_printf( aTHX_ "\n%*s| ",
2396 (int)((depth * 2) + 14), "");
2399 Perl_re_printf( aTHX_ "\n");
2404 Dumps a fully constructed but uncompressed trie in table form.
2405 This is the normal DFA style state transition table, with a few
2406 twists to facilitate compression later.
2407 Used for debugging make_trie().
2410 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2411 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2416 SV *sv=sv_newmortal();
2417 int colwidth= widecharmap ? 6 : 4;
2418 DECLARE_AND_GET_RE_DEBUG_FLAGS;
2420 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2423 print out the table precompression so that we can do a visual check
2424 that they are identical.
2427 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
2429 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2430 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2432 Perl_re_printf( aTHX_ "%*s",
2434 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2435 PL_colors[0], PL_colors[1],
2436 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2437 PERL_PV_ESCAPE_FIRSTCHAR
2443 Perl_re_printf( aTHX_ "\n");
2444 Perl_re_indentf( aTHX_ "State+-", depth+1 );
2446 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2447 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
2450 Perl_re_printf( aTHX_ "\n" );
2452 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2454 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
2456 (UV)TRIE_NODENUM( state ) );
2458 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2459 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2461 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
2463 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
2465 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2466 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
2467 (UV)trie->trans[ state ].check );
2469 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
2470 (UV)trie->trans[ state ].check,
2471 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2479 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2480 startbranch: the first branch in the whole branch sequence
2481 first : start branch of sequence of branch-exact nodes.
2482 May be the same as startbranch
2483 last : Thing following the last branch.
2484 May be the same as tail.
2485 tail : item following the branch sequence
2486 count : words in the sequence
2487 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2488 depth : indent depth
2490 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2492 A trie is an N'ary tree where the branches are determined by digital
2493 decomposition of the key. IE, at the root node you look up the 1st character and
2494 follow that branch repeat until you find the end of the branches. Nodes can be
2495 marked as "accepting" meaning they represent a complete word. Eg:
2499 would convert into the following structure. Numbers represent states, letters
2500 following numbers represent valid transitions on the letter from that state, if
2501 the number is in square brackets it represents an accepting state, otherwise it
2502 will be in parenthesis.
2504 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2508 (1) +-i->(6)-+-s->[7]
2510 +-s->(3)-+-h->(4)-+-e->[5]
2512 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2514 This shows that when matching against the string 'hers' we will begin at state 1
2515 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2516 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2517 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2518 single traverse. We store a mapping from accepting to state to which word was
2519 matched, and then when we have multiple possibilities we try to complete the
2520 rest of the regex in the order in which they occurred in the alternation.
2522 The only prior NFA like behaviour that would be changed by the TRIE support is
2523 the silent ignoring of duplicate alternations which are of the form:
2525 / (DUPE|DUPE) X? (?{ ... }) Y /x
2527 Thus EVAL blocks following a trie may be called a different number of times with
2528 and without the optimisation. With the optimisations dupes will be silently
2529 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2530 the following demonstrates:
2532 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2534 which prints out 'word' three times, but
2536 'words'=~/(word|word|word)(?{ print $1 })S/
2538 which doesnt print it out at all. This is due to other optimisations kicking in.
2540 Example of what happens on a structural level:
2542 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2544 1: CURLYM[1] {1,32767}(18)
2555 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2556 and should turn into:
2558 1: CURLYM[1] {1,32767}(18)
2560 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2568 Cases where tail != last would be like /(?foo|bar)baz/:
2578 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2579 and would end up looking like:
2582 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2589 d = uvchr_to_utf8_flags(d, uv, 0);
2591 is the recommended Unicode-aware way of saying
2596 #define TRIE_STORE_REVCHAR(val) \
2599 SV *zlopp = newSV(UTF8_MAXBYTES); \
2600 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2601 unsigned char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2603 SvCUR_set(zlopp, kapow - flrbbbbb); \
2606 av_push(revcharmap, zlopp); \
2608 char ooooff = (char)val; \
2609 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2613 /* This gets the next character from the input, folding it if not already
2615 #define TRIE_READ_CHAR STMT_START { \
2618 /* if it is UTF then it is either already folded, or does not need \
2620 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2622 else if (folder == PL_fold_latin1) { \
2623 /* This folder implies Unicode rules, which in the range expressible \
2624 * by not UTF is the lower case, with the two exceptions, one of \
2625 * which should have been taken care of before calling this */ \
2626 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2627 uvc = toLOWER_L1(*uc); \
2628 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2631 /* raw data, will be folded later if needed */ \
2639 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2640 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2641 U32 ging = TRIE_LIST_LEN( state ) * 2; \
2642 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2643 TRIE_LIST_LEN( state ) = ging; \
2645 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2646 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2647 TRIE_LIST_CUR( state )++; \
2650 #define TRIE_LIST_NEW(state) STMT_START { \
2651 Newx( trie->states[ state ].trans.list, \
2652 4, reg_trie_trans_le ); \
2653 TRIE_LIST_CUR( state ) = 1; \
2654 TRIE_LIST_LEN( state ) = 4; \
2657 #define TRIE_HANDLE_WORD(state) STMT_START { \
2658 U16 dupe= trie->states[ state ].wordnum; \
2659 regnode * const noper_next = regnext( noper ); \
2662 /* store the word for dumping */ \
2664 if (OP(noper) != NOTHING) \
2665 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2667 tmp = newSVpvn_utf8( "", 0, UTF ); \
2668 av_push( trie_words, tmp ); \
2672 trie->wordinfo[curword].prev = 0; \
2673 trie->wordinfo[curword].len = wordlen; \
2674 trie->wordinfo[curword].accept = state; \
2676 if ( noper_next < tail ) { \
2678 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2680 trie->jump[curword] = (U16)(noper_next - convert); \
2682 jumper = noper_next; \
2684 nextbranch= regnext(cur); \
2688 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2689 /* chain, so that when the bits of chain are later */\
2690 /* linked together, the dups appear in the chain */\
2691 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2692 trie->wordinfo[dupe].prev = curword; \
2694 /* we haven't inserted this word yet. */ \
2695 trie->states[ state ].wordnum = curword; \
2700 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2701 ( ( base + charid >= ucharcount \
2702 && base + charid < ubound \
2703 && state == trie->trans[ base - ucharcount + charid ].check \
2704 && trie->trans[ base - ucharcount + charid ].next ) \
2705 ? trie->trans[ base - ucharcount + charid ].next \
2706 : ( state==1 ? special : 0 ) \
2709 #define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2711 TRIE_BITMAP_SET(trie, uvc); \
2712 /* store the folded codepoint */ \
2714 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2717 /* store first byte of utf8 representation of */ \
2718 /* variant codepoints */ \
2719 if (! UVCHR_IS_INVARIANT(uvc)) { \
2720 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2725 #define MADE_JUMP_TRIE 2
2726 #define MADE_EXACT_TRIE 4
2729 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2730 regnode *first, regnode *last, regnode *tail,
2731 U32 word_count, U32 flags, U32 depth)
2733 /* first pass, loop through and scan words */
2734 reg_trie_data *trie;
2735 HV *widecharmap = NULL;
2736 AV *revcharmap = newAV();
2742 regnode *jumper = NULL;
2743 regnode *nextbranch = NULL;
2744 regnode *convert = NULL;
2745 U32 *prev_states; /* temp array mapping each state to previous one */
2746 /* we just use folder as a flag in utf8 */
2747 const U8 * folder = NULL;
2749 /* in the below add_data call we are storing either 'tu' or 'tuaa'
2750 * which stands for one trie structure, one hash, optionally followed
2753 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuaa"));
2754 AV *trie_words = NULL;
2755 /* along with revcharmap, this only used during construction but both are
2756 * useful during debugging so we store them in the struct when debugging.
2759 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2760 STRLEN trie_charcount=0;
2762 SV *re_trie_maxbuff;
2763 DECLARE_AND_GET_RE_DEBUG_FLAGS;
2765 PERL_ARGS_ASSERT_MAKE_TRIE;
2767 PERL_UNUSED_ARG(depth);
2771 case EXACT: case EXACT_REQ8: case EXACTL: break;
2775 case EXACTFLU8: folder = PL_fold_latin1; break;
2776 case EXACTF: folder = PL_fold; break;
2777 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2780 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2782 trie->startstate = 1;
2783 trie->wordcount = word_count;
2784 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2785 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2786 if (flags == EXACT || flags == EXACT_REQ8 || flags == EXACTL)
2787 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2788 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2789 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2792 trie_words = newAV();
2795 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, GV_ADD);
2796 assert(re_trie_maxbuff);
2797 if (!SvIOK(re_trie_maxbuff)) {
2798 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2800 DEBUG_TRIE_COMPILE_r({
2801 Perl_re_indentf( aTHX_
2802 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2804 REG_NODE_NUM(startbranch), REG_NODE_NUM(first),
2805 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2808 /* Find the node we are going to overwrite */
2809 if ( first == startbranch && OP( last ) != BRANCH ) {
2810 /* whole branch chain */
2813 /* branch sub-chain */
2814 convert = NEXTOPER( first );
2817 /* -- First loop and Setup --
2819 We first traverse the branches and scan each word to determine if it
2820 contains widechars, and how many unique chars there are, this is
2821 important as we have to build a table with at least as many columns as we
2824 We use an array of integers to represent the character codes 0..255
2825 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2826 the native representation of the character value as the key and IV's for
2829 *TODO* If we keep track of how many times each character is used we can
2830 remap the columns so that the table compression later on is more
2831 efficient in terms of memory by ensuring the most common value is in the
2832 middle and the least common are on the outside. IMO this would be better
2833 than a most to least common mapping as theres a decent chance the most
2834 common letter will share a node with the least common, meaning the node
2835 will not be compressible. With a middle is most common approach the worst
2836 case is when we have the least common nodes twice.
2840 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2841 regnode *noper = NEXTOPER( cur );
2845 U32 wordlen = 0; /* required init */
2846 STRLEN minchars = 0;
2847 STRLEN maxchars = 0;
2848 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2851 if (OP(noper) == NOTHING) {
2852 /* skip past a NOTHING at the start of an alternation
2853 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2855 * If the next node is not something we are supposed to process
2856 * we will just ignore it due to the condition guarding the
2860 regnode *noper_next= regnext(noper);
2861 if (noper_next < tail)
2866 && ( OP(noper) == flags
2867 || (flags == EXACT && OP(noper) == EXACT_REQ8)
2868 || (flags == EXACTFU && ( OP(noper) == EXACTFU_REQ8
2869 || OP(noper) == EXACTFUP))))
2871 uc= (U8*)STRING(noper);
2872 e= uc + STR_LEN(noper);
2879 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2880 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2881 regardless of encoding */
2882 if (OP( noper ) == EXACTFUP) {
2883 /* false positives are ok, so just set this */
2884 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2888 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2890 TRIE_CHARCOUNT(trie)++;
2893 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2894 * is in effect. Under /i, this character can match itself, or
2895 * anything that folds to it. If not under /i, it can match just
2896 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2897 * all fold to k, and all are single characters. But some folds
2898 * expand to more than one character, so for example LATIN SMALL
2899 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2900 * the string beginning at 'uc' is 'ffi', it could be matched by
2901 * three characters, or just by the one ligature character. (It
2902 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2903 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2904 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2905 * match.) The trie needs to know the minimum and maximum number
2906 * of characters that could match so that it can use size alone to
2907 * quickly reject many match attempts. The max is simple: it is
2908 * the number of folded characters in this branch (since a fold is
2909 * never shorter than what folds to it. */
2913 /* And the min is equal to the max if not under /i (indicated by
2914 * 'folder' being NULL), or there are no multi-character folds. If
2915 * there is a multi-character fold, the min is incremented just
2916 * once, for the character that folds to the sequence. Each
2917 * character in the sequence needs to be added to the list below of
2918 * characters in the trie, but we count only the first towards the
2919 * min number of characters needed. This is done through the
2920 * variable 'foldlen', which is returned by the macros that look
2921 * for these sequences as the number of bytes the sequence
2922 * occupies. Each time through the loop, we decrement 'foldlen' by
2923 * how many bytes the current char occupies. Only when it reaches
2924 * 0 do we increment 'minchars' or look for another multi-character
2926 if (folder == NULL) {
2929 else if (foldlen > 0) {
2930 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2935 /* See if *uc is the beginning of a multi-character fold. If
2936 * so, we decrement the length remaining to look at, to account
2937 * for the current character this iteration. (We can use 'uc'
2938 * instead of the fold returned by TRIE_READ_CHAR because the
2939 * macro is smart enough to account for any unfolded
2942 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2943 foldlen -= UTF8SKIP(uc);
2946 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2951 /* The current character (and any potential folds) should be added
2952 * to the possible matching characters for this position in this
2956 U8 folded= folder[ (U8) uvc ];
2957 if ( !trie->charmap[ folded ] ) {
2958 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2959 TRIE_STORE_REVCHAR( folded );
2962 if ( !trie->charmap[ uvc ] ) {
2963 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2964 TRIE_STORE_REVCHAR( uvc );
2967 /* store the codepoint in the bitmap, and its folded
2969 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
2970 set_bit = 0; /* We've done our bit :-) */
2974 /* XXX We could come up with the list of code points that fold
2975 * to this using PL_utf8_foldclosures, except not for
2976 * multi-char folds, as there may be multiple combinations
2977 * there that could work, which needs to wait until runtime to
2978 * resolve (The comment about LIGATURE FFI above is such an
2983 widecharmap = newHV();
2985 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2988 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
2990 if ( !SvTRUE( *svpp ) ) {
2991 sv_setiv( *svpp, ++trie->uniquecharcount );
2992 TRIE_STORE_REVCHAR(uvc);
2995 } /* end loop through characters in this branch of the trie */
2997 /* We take the min and max for this branch and combine to find the min
2998 * and max for all branches processed so far */
2999 if( cur == first ) {
3000 trie->minlen = minchars;
3001 trie->maxlen = maxchars;
3002 } else if (minchars < trie->minlen) {
3003 trie->minlen = minchars;
3004 } else if (maxchars > trie->maxlen) {
3005 trie->maxlen = maxchars;
3007 } /* end first pass */
3008 DEBUG_TRIE_COMPILE_r(
3009 Perl_re_indentf( aTHX_
3010 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
3012 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
3013 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
3014 (int)trie->minlen, (int)trie->maxlen )
3018 We now know what we are dealing with in terms of unique chars and
3019 string sizes so we can calculate how much memory a naive
3020 representation using a flat table will take. If it's over a reasonable
3021 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
3022 conservative but potentially much slower representation using an array
3025 At the end we convert both representations into the same compressed
3026 form that will be used in regexec.c for matching with. The latter
3027 is a form that cannot be used to construct with but has memory
3028 properties similar to the list form and access properties similar
3029 to the table form making it both suitable for fast searches and
3030 small enough that its feasable to store for the duration of a program.
3032 See the comment in the code where the compressed table is produced
3033 inplace from the flat tabe representation for an explanation of how
3034 the compression works.
3039 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
3042 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
3043 > SvIV(re_trie_maxbuff) )
3046 Second Pass -- Array Of Lists Representation
3048 Each state will be represented by a list of charid:state records
3049 (reg_trie_trans_le) the first such element holds the CUR and LEN
3050 points of the allocated array. (See defines above).
3052 We build the initial structure using the lists, and then convert
3053 it into the compressed table form which allows faster lookups
3054 (but cant be modified once converted).
3057 STRLEN transcount = 1;
3059 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
3062 trie->states = (reg_trie_state *)
3063 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3064 sizeof(reg_trie_state) );
3068 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3070 regnode *noper = NEXTOPER( cur );
3071 U32 state = 1; /* required init */
3072 U16 charid = 0; /* sanity init */
3073 U32 wordlen = 0; /* required init */
3075 if (OP(noper) == NOTHING) {
3076 regnode *noper_next= regnext(noper);
3077 if (noper_next < tail)
3079 /* we will undo this assignment if noper does not
3080 * point at a trieable type in the else clause of
3081 * the following statement. */
3085 && ( OP(noper) == flags
3086 || (flags == EXACT && OP(noper) == EXACT_REQ8)
3087 || (flags == EXACTFU && ( OP(noper) == EXACTFU_REQ8
3088 || OP(noper) == EXACTFUP))))
3090 const U8 *uc= (U8*)STRING(noper);
3091 const U8 *e= uc + STR_LEN(noper);
3093 for ( ; uc < e ; uc += len ) {
3098 charid = trie->charmap[ uvc ];
3100 SV** const svpp = hv_fetch( widecharmap,
3107 charid=(U16)SvIV( *svpp );
3110 /* charid is now 0 if we dont know the char read, or
3111 * nonzero if we do */
3118 if ( !trie->states[ state ].trans.list ) {
3119 TRIE_LIST_NEW( state );
3122 check <= TRIE_LIST_USED( state );
3125 if ( TRIE_LIST_ITEM( state, check ).forid
3128 newstate = TRIE_LIST_ITEM( state, check ).newstate;
3133 newstate = next_alloc++;
3134 prev_states[newstate] = state;
3135 TRIE_LIST_PUSH( state, charid, newstate );
3140 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3144 /* If we end up here it is because we skipped past a NOTHING, but did not end up
3145 * on a trieable type. So we need to reset noper back to point at the first regop
3146 * in the branch before we call TRIE_HANDLE_WORD()
3148 noper= NEXTOPER(cur);
3150 TRIE_HANDLE_WORD(state);
3152 } /* end second pass */
3154 /* next alloc is the NEXT state to be allocated */
3155 trie->statecount = next_alloc;
3156 trie->states = (reg_trie_state *)
3157 PerlMemShared_realloc( trie->states,
3159 * sizeof(reg_trie_state) );
3161 /* and now dump it out before we compress it */
3162 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
3163 revcharmap, next_alloc,
3167 trie->trans = (reg_trie_trans *)
3168 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
3175 for( state=1 ; state < next_alloc ; state ++ ) {
3179 DEBUG_TRIE_COMPILE_MORE_r(
3180 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
3184 if (trie->states[state].trans.list) {
3185 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
3189 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3190 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
3191 if ( forid < minid ) {
3193 } else if ( forid > maxid ) {
3197 if ( transcount < tp + maxid - minid + 1) {
3199 trie->trans = (reg_trie_trans *)
3200 PerlMemShared_realloc( trie->trans,
3202 * sizeof(reg_trie_trans) );
3203 Zero( trie->trans + (transcount / 2),
3207 base = trie->uniquecharcount + tp - minid;
3208 if ( maxid == minid ) {
3210 for ( ; zp < tp ; zp++ ) {
3211 if ( ! trie->trans[ zp ].next ) {
3212 base = trie->uniquecharcount + zp - minid;
3213 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
3215 trie->trans[ zp ].check = state;
3221 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
3223 trie->trans[ tp ].check = state;
3228 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3229 const U32 tid = base
3230 - trie->uniquecharcount
3231 + TRIE_LIST_ITEM( state, idx ).forid;
3232 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
3234 trie->trans[ tid ].check = state;
3236 tp += ( maxid - minid + 1 );
3238 Safefree(trie->states[ state ].trans.list);
3241 DEBUG_TRIE_COMPILE_MORE_r(
3242 Perl_re_printf( aTHX_ " base: %d\n",base);
3245 trie->states[ state ].trans.base=base;
3247 trie->lasttrans = tp + 1;
3251 Second Pass -- Flat Table Representation.
3253 we dont use the 0 slot of either trans[] or states[] so we add 1 to
3254 each. We know that we will need Charcount+1 trans at most to store
3255 the data (one row per char at worst case) So we preallocate both
3256 structures assuming worst case.
3258 We then construct the trie using only the .next slots of the entry
3261 We use the .check field of the first entry of the node temporarily
3262 to make compression both faster and easier by keeping track of how
3263 many non zero fields are in the node.
3265 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
3268 There are two terms at use here: state as a TRIE_NODEIDX() which is
3269 a number representing the first entry of the node, and state as a
3270 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
3271 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
3272 if there are 2 entrys per node. eg:
3280 The table is internally in the right hand, idx form. However as we
3281 also have to deal with the states array which is indexed by nodenum
3282 we have to use TRIE_NODENUM() to convert.
3285 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
3288 trie->trans = (reg_trie_trans *)
3289 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
3290 * trie->uniquecharcount + 1,
3291 sizeof(reg_trie_trans) );
3292 trie->states = (reg_trie_state *)
3293 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3294 sizeof(reg_trie_state) );
3295 next_alloc = trie->uniquecharcount + 1;
3298 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3300 regnode *noper = NEXTOPER( cur );
3302 U32 state = 1; /* required init */
3304 U16 charid = 0; /* sanity init */
3305 U32 accept_state = 0; /* sanity init */
3307 U32 wordlen = 0; /* required init */
3309 if (OP(noper) == NOTHING) {
3310 regnode *noper_next= regnext(noper);
3311 if (noper_next < tail)
3313 /* we will undo this assignment if noper does not
3314 * point at a trieable type in the else clause of
3315 * the following statement. */
3319 && ( OP(noper) == flags
3320 || (flags == EXACT && OP(noper) == EXACT_REQ8)
3321 || (flags == EXACTFU && ( OP(noper) == EXACTFU_REQ8
3322 || OP(noper) == EXACTFUP))))
3324 const U8 *uc= (U8*)STRING(noper);
3325 const U8 *e= uc + STR_LEN(noper);
3327 for ( ; uc < e ; uc += len ) {
3332 charid = trie->charmap[ uvc ];
3334 SV* const * const svpp = hv_fetch( widecharmap,
3338 charid = svpp ? (U16)SvIV(*svpp) : 0;
3342 if ( !trie->trans[ state + charid ].next ) {
3343 trie->trans[ state + charid ].next = next_alloc;
3344 trie->trans[ state ].check++;
3345 prev_states[TRIE_NODENUM(next_alloc)]
3346 = TRIE_NODENUM(state);
3347 next_alloc += trie->uniquecharcount;
3349 state = trie->trans[ state + charid ].next;
3351 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3353 /* charid is now 0 if we dont know the char read, or
3354 * nonzero if we do */
3357 /* If we end up here it is because we skipped past a NOTHING, but did not end up
3358 * on a trieable type. So we need to reset noper back to point at the first regop
3359 * in the branch before we call TRIE_HANDLE_WORD().
3361 noper= NEXTOPER(cur);
3363 accept_state = TRIE_NODENUM( state );
3364 TRIE_HANDLE_WORD(accept_state);
3366 } /* end second pass */
3368 /* and now dump it out before we compress it */
3369 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3371 next_alloc, depth+1));
3375 * Inplace compress the table.*
3377 For sparse data sets the table constructed by the trie algorithm will
3378 be mostly 0/FAIL transitions or to put it another way mostly empty.
3379 (Note that leaf nodes will not contain any transitions.)
3381 This algorithm compresses the tables by eliminating most such
3382 transitions, at the cost of a modest bit of extra work during lookup:
3384 - Each states[] entry contains a .base field which indicates the
3385 index in the state[] array wheres its transition data is stored.
3387 - If .base is 0 there are no valid transitions from that node.
3389 - If .base is nonzero then charid is added to it to find an entry in
3392 -If trans[states[state].base+charid].check!=state then the
3393 transition is taken to be a 0/Fail transition. Thus if there are fail
3394 transitions at the front of the node then the .base offset will point
3395 somewhere inside the previous nodes data (or maybe even into a node
3396 even earlier), but the .check field determines if the transition is
3400 The following process inplace converts the table to the compressed
3401 table: We first do not compress the root node 1,and mark all its
3402 .check pointers as 1 and set its .base pointer as 1 as well. This
3403 allows us to do a DFA construction from the compressed table later,
3404 and ensures that any .base pointers we calculate later are greater
3407 - We set 'pos' to indicate the first entry of the second node.
3409 - We then iterate over the columns of the node, finding the first and
3410 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3411 and set the .check pointers accordingly, and advance pos
3412 appropriately and repreat for the next node. Note that when we copy
3413 the next pointers we have to convert them from the original
3414 NODEIDX form to NODENUM form as the former is not valid post
3417 - If a node has no transitions used we mark its base as 0 and do not
3418 advance the pos pointer.
3420 - If a node only has one transition we use a second pointer into the
3421 structure to fill in allocated fail transitions from other states.
3422 This pointer is independent of the main pointer and scans forward
3423 looking for null transitions that are allocated to a state. When it
3424 finds one it writes the single transition into the "hole". If the
3425 pointer doesnt find one the single transition is appended as normal.
3427 - Once compressed we can Renew/realloc the structures to release the
3430 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3431 specifically Fig 3.47 and the associated pseudocode.
3435 const U32 laststate = TRIE_NODENUM( next_alloc );
3438 trie->statecount = laststate;
3440 for ( state = 1 ; state < laststate ; state++ ) {
3442 const U32 stateidx = TRIE_NODEIDX( state );
3443 const U32 o_used = trie->trans[ stateidx ].check;
3444 U32 used = trie->trans[ stateidx ].check;
3445 trie->trans[ stateidx ].check = 0;
3448 used && charid < trie->uniquecharcount;
3451 if ( flag || trie->trans[ stateidx + charid ].next ) {
3452 if ( trie->trans[ stateidx + charid ].next ) {
3454 for ( ; zp < pos ; zp++ ) {
3455 if ( ! trie->trans[ zp ].next ) {
3459 trie->states[ state ].trans.base
3461 + trie->uniquecharcount
3463 trie->trans[ zp ].next
3464 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3466 trie->trans[ zp ].check = state;
3467 if ( ++zp > pos ) pos = zp;
3474 trie->states[ state ].trans.base
3475 = pos + trie->uniquecharcount - charid ;
3477 trie->trans[ pos ].next
3478 = SAFE_TRIE_NODENUM(
3479 trie->trans[ stateidx + charid ].next );
3480 trie->trans[ pos ].check = state;
3485 trie->lasttrans = pos + 1;
3486 trie->states = (reg_trie_state *)
3487 PerlMemShared_realloc( trie->states, laststate
3488 * sizeof(reg_trie_state) );
3489 DEBUG_TRIE_COMPILE_MORE_r(
3490 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
3492 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3496 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3499 } /* end table compress */
3501 DEBUG_TRIE_COMPILE_MORE_r(
3502 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
3504 (UV)trie->statecount,
3505 (UV)trie->lasttrans)
3507 /* resize the trans array to remove unused space */
3508 trie->trans = (reg_trie_trans *)
3509 PerlMemShared_realloc( trie->trans, trie->lasttrans
3510 * sizeof(reg_trie_trans) );
3512 { /* Modify the program and insert the new TRIE node */
3513 U8 nodetype =(U8)(flags & 0xFF);
3517 regnode *optimize = NULL;
3518 #ifdef RE_TRACK_PATTERN_OFFSETS
3521 U32 mjd_nodelen = 0;
3522 #endif /* RE_TRACK_PATTERN_OFFSETS */
3523 #endif /* DEBUGGING */
3525 This means we convert either the first branch or the first Exact,
3526 depending on whether the thing following (in 'last') is a branch
3527 or not and whther first is the startbranch (ie is it a sub part of
3528 the alternation or is it the whole thing.)
3529 Assuming its a sub part we convert the EXACT otherwise we convert
3530 the whole branch sequence, including the first.
3532 /* Find the node we are going to overwrite */
3533 if ( first != startbranch || OP( last ) == BRANCH ) {
3534 /* branch sub-chain */
3535 NEXT_OFF( first ) = (U16)(last - first);
3536 #ifdef RE_TRACK_PATTERN_OFFSETS
3538 mjd_offset= Node_Offset((convert));
3539 mjd_nodelen= Node_Length((convert));
3542 /* whole branch chain */
3544 #ifdef RE_TRACK_PATTERN_OFFSETS
3547 const regnode *nop = NEXTOPER( convert );
3548 mjd_offset= Node_Offset((nop));
3549 mjd_nodelen= Node_Length((nop));
3553 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
3555 (UV)mjd_offset, (UV)mjd_nodelen)
3558 /* But first we check to see if there is a common prefix we can
3559 split out as an EXACT and put in front of the TRIE node. */
3560 trie->startstate= 1;
3561 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3562 /* we want to find the first state that has more than
3563 * one transition, if that state is not the first state
3564 * then we have a common prefix which we can remove.
3567 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3569 I32 first_ofs = -1; /* keeps track of the ofs of the first
3570 transition, -1 means none */
3572 const U32 base = trie->states[ state ].trans.base;
3574 /* does this state terminate an alternation? */
3575 if ( trie->states[state].wordnum )
3578 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3579 if ( ( base + ofs >= trie->uniquecharcount ) &&
3580 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3581 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3583 if ( ++count > 1 ) {
3584 /* we have more than one transition */
3587 /* if this is the first state there is no common prefix
3588 * to extract, so we can exit */
3589 if ( state == 1 ) break;
3590 tmp = av_fetch( revcharmap, ofs, 0);
3591 ch = (U8*)SvPV_nolen_const( *tmp );
3593 /* if we are on count 2 then we need to initialize the
3594 * bitmap, and store the previous char if there was one
3597 /* clear the bitmap */
3598 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3600 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
3603 if (first_ofs >= 0) {
3604 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
3605 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3607 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3609 Perl_re_printf( aTHX_ "%s", (char*)ch)
3613 /* store the current firstchar in the bitmap */
3614 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3615 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
3621 /* This state has only one transition, its transition is part
3622 * of a common prefix - we need to concatenate the char it
3623 * represents to what we have so far. */
3624 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
3626 char *ch = SvPV( *tmp, len );
3628 SV *sv=sv_newmortal();
3629 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
3631 (UV)state, (UV)first_ofs,
3632 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3633 PL_colors[0], PL_colors[1],
3634 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3635 PERL_PV_ESCAPE_FIRSTCHAR
3640 OP( convert ) = nodetype;
3641 str=STRING(convert);
3642 setSTR_LEN(convert, 0);
3644 assert( ( STR_LEN(convert) + len ) < 256 );
3645 setSTR_LEN(convert, (U8)(STR_LEN(convert) + len));
3651 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
3656 trie->prefixlen = (state-1);
3658 regnode *n = convert+NODE_SZ_STR(convert);
3659 assert( NODE_SZ_STR(convert) <= U16_MAX );
3660 NEXT_OFF(convert) = (U16)(NODE_SZ_STR(convert));
3661 trie->startstate = state;
3662 trie->minlen -= (state - 1);
3663 trie->maxlen -= (state - 1);
3665 /* At least the UNICOS C compiler choked on this
3666 * being argument to DEBUG_r(), so let's just have
3669 #ifdef PERL_EXT_RE_BUILD
3675 regnode *fix = convert;
3676 U32 word = trie->wordcount;
3677 #ifdef RE_TRACK_PATTERN_OFFSETS
3680 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3681 while( ++fix < n ) {
3682 Set_Node_Offset_Length(fix, 0, 0);
3685 SV ** const tmp = av_fetch( trie_words, word, 0 );
3687 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3688 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3690 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3698 NEXT_OFF(convert) = (U16)(tail - convert);
3699 DEBUG_r(optimize= n);
3705 if ( trie->maxlen ) {
3706 NEXT_OFF( convert ) = (U16)(tail - convert);
3707 ARG_SET( convert, data_slot );
3708 /* Store the offset to the first unabsorbed branch in
3709 jump[0], which is otherwise unused by the jump logic.
3710 We use this when dumping a trie and during optimisation. */
3712 trie->jump[0] = (U16)(nextbranch - convert);
3714 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3715 * and there is a bitmap
3716 * and the first "jump target" node we found leaves enough room
3717 * then convert the TRIE node into a TRIEC node, with the bitmap
3718 * embedded inline in the opcode - this is hypothetically faster.
3720 if ( !trie->states[trie->startstate].wordnum
3722 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3724 OP( convert ) = TRIEC;
3725 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3726 PerlMemShared_free(trie->bitmap);
3729 OP( convert ) = TRIE;
3731 /* store the type in the flags */
3732 convert->flags = nodetype;
3736 + regarglen[ OP( convert ) ];
3738 /* XXX We really should free up the resource in trie now,
3739 as we won't use them - (which resources?) dmq */
3741 /* needed for dumping*/
3742 DEBUG_r(if (optimize) {
3743 regnode *opt = convert;
3745 while ( ++opt < optimize) {
3746 Set_Node_Offset_Length(opt, 0, 0);
3749 Try to clean up some of the debris left after the
3752 while( optimize < jumper ) {
3753 Track_Code( mjd_nodelen += Node_Length((optimize)); );
3754 OP( optimize ) = OPTIMIZED;
3755 Set_Node_Offset_Length(optimize, 0, 0);
3758 Set_Node_Offset_Length(convert, mjd_offset, mjd_nodelen);
3760 } /* end node insert */
3762 /* Finish populating the prev field of the wordinfo array. Walk back
3763 * from each accept state until we find another accept state, and if
3764 * so, point the first word's .prev field at the second word. If the
3765 * second already has a .prev field set, stop now. This will be the
3766 * case either if we've already processed that word's accept state,
3767 * or that state had multiple words, and the overspill words were
3768 * already linked up earlier.
3775 for (word=1; word <= trie->wordcount; word++) {
3777 if (trie->wordinfo[word].prev)
3779 state = trie->wordinfo[word].accept;
3781 state = prev_states[state];
3784 prev = trie->states[state].wordnum;
3788 trie->wordinfo[word].prev = prev;
3790 Safefree(prev_states);
3794 /* and now dump out the compressed format */
3795 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3797 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3799 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3800 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3802 SvREFCNT_dec_NN(revcharmap);
3806 : trie->startstate>1
3812 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3814 /* The Trie is constructed and compressed now so we can build a fail array if
3817 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3819 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3823 We find the fail state for each state in the trie, this state is the longest
3824 proper suffix of the current state's 'word' that is also a proper prefix of
3825 another word in our trie. State 1 represents the word '' and is thus the
3826 default fail state. This allows the DFA not to have to restart after its
3827 tried and failed a word at a given point, it simply continues as though it
3828 had been matching the other word in the first place.
3830 'abcdgu'=~/abcdefg|cdgu/
3831 When we get to 'd' we are still matching the first word, we would encounter
3832 'g' which would fail, which would bring us to the state representing 'd' in
3833 the second word where we would try 'g' and succeed, proceeding to match
3836 /* add a fail transition */
3837 const U32 trie_offset = ARG(source);
3838 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3840 const U32 ucharcount = trie->uniquecharcount;
3841 const U32 numstates = trie->statecount;
3842 const U32 ubound = trie->lasttrans + ucharcount;
3846 U32 base = trie->states[ 1 ].trans.base;
3849 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3851 DECLARE_AND_GET_RE_DEBUG_FLAGS;
3853 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3854 PERL_UNUSED_CONTEXT;
3856 PERL_UNUSED_ARG(depth);
3859 if ( OP(source) == TRIE ) {
3860 struct regnode_1 *op = (struct regnode_1 *)
3861 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3862 StructCopy(source, op, struct regnode_1);
3863 stclass = (regnode *)op;
3865 struct regnode_charclass *op = (struct regnode_charclass *)
3866 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3867 StructCopy(source, op, struct regnode_charclass);
3868 stclass = (regnode *)op;
3870 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3872 ARG_SET( stclass, data_slot );
3873 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3874 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3875 aho->trie=trie_offset;
3876 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3877 Copy( trie->states, aho->states, numstates, reg_trie_state );
3878 Newx( q, numstates, U32);
3879 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3882 /* initialize fail[0..1] to be 1 so that we always have
3883 a valid final fail state */
3884 fail[ 0 ] = fail[ 1 ] = 1;
3886 for ( charid = 0; charid < ucharcount ; charid++ ) {
3887 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3889 q[ q_write ] = newstate;
3890 /* set to point at the root */
3891 fail[ q[ q_write++ ] ]=1;
3894 while ( q_read < q_write) {
3895 const U32 cur = q[ q_read++ % numstates ];
3896 base = trie->states[ cur ].trans.base;
3898 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3899 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3901 U32 fail_state = cur;
3904 fail_state = fail[ fail_state ];
3905 fail_base = aho->states[ fail_state ].trans.base;
3906 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3908 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3909 fail[ ch_state ] = fail_state;
3910 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3912 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3914 q[ q_write++ % numstates] = ch_state;
3918 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3919 when we fail in state 1, this allows us to use the
3920 charclass scan to find a valid start char. This is based on the principle
3921 that theres a good chance the string being searched contains lots of stuff
3922 that cant be a start char.
3924 fail[ 0 ] = fail[ 1 ] = 0;
3925 DEBUG_TRIE_COMPILE_r({
3926 Perl_re_indentf( aTHX_ "Stclass Failtable (%" UVuf " states): 0",
3927 depth, (UV)numstates
3929 for( q_read=1; q_read<numstates; q_read++ ) {
3930 Perl_re_printf( aTHX_ ", %" UVuf, (UV)fail[q_read]);
3932 Perl_re_printf( aTHX_ "\n");
3935 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3940 /* The below joins as many adjacent EXACTish nodes as possible into a single
3941 * one. The regop may be changed if the node(s) contain certain sequences that
3942 * require special handling. The joining is only done if:
3943 * 1) there is room in the current conglomerated node to entirely contain the
3945 * 2) they are compatible node types
3947 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3948 * these get optimized out
3950 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3951 * as possible, even if that means splitting an existing node so that its first
3952 * part is moved to the preceeding node. This would maximise the efficiency of
3953 * memEQ during matching.
3955 * If a node is to match under /i (folded), the number of characters it matches
3956 * can be different than its character length if it contains a multi-character
3957 * fold. *min_subtract is set to the total delta number of characters of the
3960 * And *unfolded_multi_char is set to indicate whether or not the node contains
3961 * an unfolded multi-char fold. This happens when it won't be known until
3962 * runtime whether the fold is valid or not; namely
3963 * 1) for EXACTF nodes that contain LATIN SMALL LETTER SHARP S, as only if the
3964 * target string being matched against turns out to be UTF-8 is that fold
3966 * 2) for EXACTFL nodes whose folding rules depend on the locale in force at
3968 * (Multi-char folds whose components are all above the Latin1 range are not
3969 * run-time locale dependent, and have already been folded by the time this
3970 * function is called.)
3972 * This is as good a place as any to discuss the design of handling these
3973 * multi-character fold sequences. It's been wrong in Perl for a very long
3974 * time. There are three code points in Unicode whose multi-character folds
3975 * were long ago discovered to mess things up. The previous designs for
3976 * dealing with these involved assigning a special node for them. This
3977 * approach doesn't always work, as evidenced by this example:
3978 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3979 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3980 * would match just the \xDF, it won't be able to handle the case where a
3981 * successful match would have to cross the node's boundary. The new approach
3982 * that hopefully generally solves the problem generates an EXACTFUP node
3983 * that is "sss" in this case.
3985 * It turns out that there are problems with all multi-character folds, and not
3986 * just these three. Now the code is general, for all such cases. The
3987 * approach taken is:
3988 * 1) This routine examines each EXACTFish node that could contain multi-
3989 * character folded sequences. Since a single character can fold into
3990 * such a sequence, the minimum match length for this node is less than
3991 * the number of characters in the node. This routine returns in
3992 * *min_subtract how many characters to subtract from the actual
3993 * length of the string to get a real minimum match length; it is 0 if
3994 * there are no multi-char foldeds. This delta is used by the caller to
3995 * adjust the min length of the match, and the delta between min and max,
3996 * so that the optimizer doesn't reject these possibilities based on size
3999 * 2) For the sequence involving the LATIN SMALL LETTER SHARP S (U+00DF)
4000 * under /u, we fold it to 'ss' in regatom(), and in this routine, after
4001 * joining, we scan for occurrences of the sequence 'ss' in non-UTF-8
4002 * EXACTFU nodes. The node type of such nodes is then changed to
4003 * EXACTFUP, indicating it is problematic, and needs careful handling.
4004 * (The procedures in step 1) above are sufficient to handle this case in
4005 * UTF-8 encoded nodes.) The reason this is problematic is that this is
4006 * the only case where there is a possible fold length change in non-UTF-8
4007 * patterns. By reserving a special node type for problematic cases, the
4008 * far more common regular EXACTFU nodes can be processed faster.
4009 * regexec.c takes advantage of this.
4011 * EXACTFUP has been created as a grab-bag for (hopefully uncommon)
4012 * problematic cases. These all only occur when the pattern is not
4013 * UTF-8. In addition to the 'ss' sequence where there is a possible fold
4014 * length change, it handles the situation where the string cannot be
4015 * entirely folded. The strings in an EXACTFish node are folded as much
4016 * as possible during compilation in regcomp.c. This saves effort in
4017 * regex matching. By using an EXACTFUP node when it is not possible to
4018 * fully fold at compile time, regexec.c can know that everything in an
4019 * EXACTFU node is folded, so folding can be skipped at runtime. The only
4020 * case where folding in EXACTFU nodes can't be done at compile time is
4021 * the presumably uncommon MICRO SIGN, when the pattern isn't UTF-8. This
4022 * is because its fold requires UTF-8 to represent. Thus EXACTFUP nodes
4023 * handle two very different cases. Alternatively, there could have been
4024 * a node type where there are length changes, one for unfolded, and one
4025 * for both. If yet another special case needed to be created, the number
4026 * of required node types would have to go to 7. khw figures that even
4027 * though there are plenty of node types to spare, that the maintenance
4028 * cost wasn't worth the small speedup of doing it that way, especially
4029 * since he thinks the MICRO SIGN is rarely encountered in practice.
4031 * There are other cases where folding isn't done at compile time, but
4032 * none of them are under /u, and hence not for EXACTFU nodes. The folds
4033 * in EXACTFL nodes aren't known until runtime, and vary as the locale
4034 * changes. Some folds in EXACTF depend on if the runtime target string
4035 * is UTF-8 or not. (regatom() will create an EXACTFU node even under /di
4036 * when no fold in it depends on the UTF-8ness of the target string.)
4038 * 3) A problem remains for unfolded multi-char folds. (These occur when the
4039 * validity of the fold won't be known until runtime, and so must remain
4040 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFAA
4041 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
4042 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
4043 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
4044 * The reason this is a problem is that the optimizer part of regexec.c
4045 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
4046 * that a character in the pattern corresponds to at most a single
4047 * character in the target string. (And I do mean character, and not byte
4048 * here, unlike other parts of the documentation that have never been
4049 * updated to account for multibyte Unicode.) Sharp s in EXACTF and
4050 * EXACTFL nodes can match the two character string 'ss'; in EXACTFAA
4051 * nodes it can match "\x{17F}\x{17F}". These, along with other ones in
4052 * EXACTFL nodes, violate the assumption, and they are the only instances
4053 * where it is violated. I'm reluctant to try to change the assumption,
4054 * as the code involved is impenetrable to me (khw), so instead the code
4055 * here punts. This routine examines EXACTFL nodes, and (when the pattern
4056 * isn't UTF-8) EXACTF and EXACTFAA for such unfolded folds, and returns a
4057 * boolean indicating whether or not the node contains such a fold. When
4058 * it is true, the caller sets a flag that later causes the optimizer in
4059 * this file to not set values for the floating and fixed string lengths,
4060 * and thus avoids the optimizer code in regexec.c that makes the invalid
4061 * assumption. Thus, there is no optimization based on string lengths for
4062 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
4063 * EXACTF and EXACTFAA nodes that contain the sharp s. (The reason the
4064 * assumption is wrong only in these cases is that all other non-UTF-8
4065 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
4066 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
4067 * EXACTF nodes because we don't know at compile time if it actually
4068 * matches 'ss' or not. For EXACTF nodes it will match iff the target
4069 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
4070 * always matches; and EXACTFAA where it never does. In an EXACTFAA node
4071 * in a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
4072 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
4073 * string would require the pattern to be forced into UTF-8, the overhead
4074 * of which we want to avoid. Similarly the unfolded multi-char folds in
4075 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
4078 * Similarly, the code that generates tries doesn't currently handle
4079 * not-already-folded multi-char folds, and it looks like a pain to change
4080 * that. Therefore, trie generation of EXACTFAA nodes with the sharp s
4081 * doesn't work. Instead, such an EXACTFAA is turned into a new regnode,
4082 * EXACTFAA_NO_TRIE, which the trie code knows not to handle. Most people
4083 * using /iaa matching will be doing so almost entirely with ASCII
4084 * strings, so this should rarely be encountered in practice */
4087 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
4088 UV *min_subtract, bool *unfolded_multi_char,
4089 U32 flags, regnode *val, U32 depth)
4091 /* Merge several consecutive EXACTish nodes into one. */
4093 regnode *n = regnext(scan);
4095 regnode *next = scan + NODE_SZ_STR(scan);
4099 regnode *stop = scan;
4100 DECLARE_AND_GET_RE_DEBUG_FLAGS;
4102 PERL_UNUSED_ARG(depth);
4105 PERL_ARGS_ASSERT_JOIN_EXACT;
4106 #ifndef EXPERIMENTAL_INPLACESCAN
4107 PERL_UNUSED_ARG(flags);
4108 PERL_UNUSED_ARG(val);
4110 DEBUG_PEEP("join", scan, depth, 0);
4112 assert(PL_regkind[OP(scan)] == EXACT);
4114 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
4115 * EXACT ones that are mergeable to the current one. */
4117 && ( PL_regkind[OP(n)] == NOTHING
4118 || (stringok && PL_regkind[OP(n)] == EXACT))
4120 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
4123 if (OP(n) == TAIL || n > next)
4125 if (PL_regkind[OP(n)] == NOTHING) {
4126 DEBUG_PEEP("skip:", n, depth, 0);
4127 NEXT_OFF(scan) += NEXT_OFF(n);
4128 next = n + NODE_STEP_REGNODE;
4135 else if (stringok) {
4136 const unsigned int oldl = STR_LEN(scan);
4137 regnode * const nnext = regnext(n);
4139 /* XXX I (khw) kind of doubt that this works on platforms (should
4140 * Perl ever run on one) where U8_MAX is above 255 because of lots
4141 * of other assumptions */
4142 /* Don't join if the sum can't fit into a single node */
4143 if (oldl + STR_LEN(n) > U8_MAX)
4146 /* Joining something that requires UTF-8 with something that
4147 * doesn't, means the result requires UTF-8. */
4148 if (OP(scan) == EXACT && (OP(n) == EXACT_REQ8)) {
4149 OP(scan) = EXACT_REQ8;
4151 else if (OP(scan) == EXACT_REQ8 && (OP(n) == EXACT)) {
4152 ; /* join is compatible, no need to change OP */
4154 else if ((OP(scan) == EXACTFU) && (OP(n) == EXACTFU_REQ8)) {
4155 OP(scan) = EXACTFU_REQ8;
4157 else if ((OP(scan) == EXACTFU_REQ8) && (OP(n) == EXACTFU)) {
4158 ; /* join is compatible, no need to change OP */
4160 else if (OP(scan) == EXACTFU && OP(n) == EXACTFU) {
4161 ; /* join is compatible, no need to change OP */
4163 else if (OP(scan) == EXACTFU && OP(n) == EXACTFU_S_EDGE) {
4165 /* Under /di, temporary EXACTFU_S_EDGE nodes are generated,
4166 * which can join with EXACTFU ones. We check for this case
4167 * here. These need to be resolved to either EXACTFU or
4168 * EXACTF at joining time. They have nothing in them that
4169 * would forbid them from being the more desirable EXACTFU
4170 * nodes except that they begin and/or end with a single [Ss].
4171 * The reason this is problematic is because they could be
4172 * joined in this loop with an adjacent node that ends and/or
4173 * begins with [Ss] which would then form the sequence 'ss',
4174 * which matches differently under /di than /ui, in which case
4175 * EXACTFU can't be used. If the 'ss' sequence doesn't get
4176 * formed, the nodes get absorbed into any adjacent EXACTFU
4177 * node. And if the only adjacent node is EXACTF, they get
4178 * absorbed into that, under the theory that a longer node is
4179 * better than two shorter ones, even if one is EXACTFU. Note
4180 * that EXACTFU_REQ8 is generated only for UTF-8 patterns,
4181 * and the EXACTFU_S_EDGE ones only for non-UTF-8. */
4183 if (STRING(n)[STR_LEN(n)-1] == 's') {
4185 /* Here the joined node would end with 's'. If the node
4186 * following the combination is an EXACTF one, it's better to
4187 * join this trailing edge 's' node with that one, leaving the
4188 * current one in 'scan' be the more desirable EXACTFU */
4189 if (OP(nnext) == EXACTF) {
4193 OP(scan) = EXACTFU_S_EDGE;
4195 } /* Otherwise, the beginning 's' of the 2nd node just
4196 becomes an interior 's' in 'scan' */
4198 else if (OP(scan) == EXACTF && OP(n) == EXACTF) {
4199 ; /* join is compatible, no need to change OP */
4201 else if (OP(scan) == EXACTF && OP(n) == EXACTFU_S_EDGE) {
4203 /* EXACTF nodes are compatible for joining with EXACTFU_S_EDGE
4204 * nodes. But the latter nodes can be also joined with EXACTFU
4205 * ones, and that is a better outcome, so if the node following
4206 * 'n' is EXACTFU, quit now so that those two can be joined
4208 if (OP(nnext) == EXACTFU) {
4212 /* The join is compatible, and the combined node will be
4213 * EXACTF. (These don't care if they begin or end with 's' */
4215 else if (OP(scan) == EXACTFU_S_EDGE && OP(n) == EXACTFU_S_EDGE) {
4216 if ( STRING(scan)[STR_LEN(scan)-1] == 's'
4217 && STRING(n)[0] == 's')
4219 /* When combined, we have the sequence 'ss', which means we
4220 * have to remain /di */
4224 else if (OP(scan) == EXACTFU_S_EDGE && OP(n) == EXACTFU) {
4225 if (STRING(n)[0] == 's') {
4226 ; /* Here the join is compatible and the combined node
4227 starts with 's', no need to change OP */
4229 else { /* Now the trailing 's' is in the interior */
4233 else if (OP(scan) == EXACTFU_S_EDGE && OP(n) == EXACTF) {
4235 /* The join is compatible, and the combined node will be
4236 * EXACTF. (These don't care if they begin or end with 's' */
4239 else if (OP(scan) != OP(n)) {
4241 /* The only other compatible joinings are the same node type */
4245 DEBUG_PEEP("merg", n, depth, 0);
4248 NEXT_OFF(scan) += NEXT_OFF(n);
4249 assert( ( STR_LEN(scan) + STR_LEN(n) ) < 256 );
4250 setSTR_LEN(scan, (U8)(STR_LEN(scan) + STR_LEN(n)));
4251 next = n + NODE_SZ_STR(n);
4252 /* Now we can overwrite *n : */
4253 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
4261 #ifdef EXPERIMENTAL_INPLACESCAN
4262 if (flags && !NEXT_OFF(n)) {
4263 DEBUG_PEEP("atch", val, depth, 0);
4264 if (reg_off_by_arg[OP(n)]) {
4265 ARG_SET(n, val - n);
4268 NEXT_OFF(n) = val - n;
4275 /* This temporary node can now be turned into EXACTFU, and must, as
4276 * regexec.c doesn't handle it */
4277 if (OP(scan) == EXACTFU_S_EDGE) {
4282 *unfolded_multi_char = FALSE;
4284 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
4285 * can now analyze for sequences of problematic code points. (Prior to
4286 * this final joining, sequences could have been split over boundaries, and
4287 * hence missed). The sequences only happen in folding, hence for any
4288 * non-EXACT EXACTish node */
4289 if (OP(scan) != EXACT && OP(scan) != EXACT_REQ8 && OP(scan) != EXACTL) {
4290 U8* s0 = (U8*) STRING(scan);
4292 U8* s_end = s0 + STR_LEN(scan);
4294 int total_count_delta = 0; /* Total delta number of characters that
4295 multi-char folds expand to */
4297 /* One pass is made over the node's string looking for all the
4298 * possibilities. To avoid some tests in the loop, there are two main
4299 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
4304 if (OP(scan) == EXACTFL) {
4307 /* An EXACTFL node would already have been changed to another
4308 * node type unless there is at least one character in it that
4309 * is problematic; likely a character whose fold definition
4310 * won't be known until runtime, and so has yet to be folded.
4311 * For all but the UTF-8 locale, folds are 1-1 in length, but
4312 * to handle the UTF-8 case, we need to create a temporary
4313 * folded copy using UTF-8 locale rules in order to analyze it.
4314 * This is because our macros that look to see if a sequence is
4315 * a multi-char fold assume everything is folded (otherwise the
4316 * tests in those macros would be too complicated and slow).
4317 * Note that here, the non-problematic folds will have already
4318 * been done, so we can just copy such characters. We actually
4319 * don't completely fold the EXACTFL string. We skip the
4320 * unfolded multi-char folds, as that would just create work
4321 * below to figure out the size they already are */
4323 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
4326 STRLEN s_len = UTF8SKIP(s);
4327 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
4328 Copy(s, d, s_len, U8);
4331 else if (is_FOLDS_TO_MULTI_utf8(s)) {
4332 *unfolded_multi_char = TRUE;
4333 Copy(s, d, s_len, U8);
4336 else if (isASCII(*s)) {
4337 *(d++) = toFOLD(*s);
4341 _toFOLD_utf8_flags(s, s_end, d, &len, FOLD_FLAGS_FULL);
4347 /* Point the remainder of the routine to look at our temporary
4351 } /* End of creating folded copy of EXACTFL string */
4353 /* Examine the string for a multi-character fold sequence. UTF-8
4354 * patterns have all characters pre-folded by the time this code is
4356 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
4357 length sequence we are looking for is 2 */
4359 int count = 0; /* How many characters in a multi-char fold */
4360 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
4361 if (! len) { /* Not a multi-char fold: get next char */
4366 { /* Here is a generic multi-char fold. */
4367 U8* multi_end = s + len;
4369 /* Count how many characters are in it. In the case of
4370 * /aa, no folds which contain ASCII code points are
4371 * allowed, so check for those, and skip if found. */
4372 if (OP(scan) != EXACTFAA && OP(scan) != EXACTFAA_NO_TRIE) {
4373 count = utf8_length(s, multi_end);
4377 while (s < multi_end) {
4380 goto next_iteration;
4390 /* The delta is how long the sequence is minus 1 (1 is how long
4391 * the character that folds to the sequence is) */
4392 total_count_delta += count - 1;
4396 /* We created a temporary folded copy of the string in EXACTFL
4397 * nodes. Therefore we need to be sure it doesn't go below zero,
4398 * as the real string could be shorter */
4399 if (OP(scan) == EXACTFL) {
4400 int total_chars = utf8_length((U8*) STRING(scan),
4401 (U8*) STRING(scan) + STR_LEN(scan));
4402 if (total_count_delta > total_chars) {
4403 total_count_delta = total_chars;
4407 *min_subtract += total_count_delta;
4410 else if (OP(scan) == EXACTFAA) {
4412 /* Non-UTF-8 pattern, EXACTFAA node. There can't be a multi-char
4413 * fold to the ASCII range (and there are no existing ones in the
4414 * upper latin1 range). But, as outlined in the comments preceding
4415 * this function, we need to flag any occurrences of the sharp s.
4416 * This character forbids trie formation (because of added
4418 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
4419 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
4420 || UNICODE_DOT_DOT_VERSION > 0)
4422 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
4423 OP(scan) = EXACTFAA_NO_TRIE;
4424 *unfolded_multi_char = TRUE;
4430 else if (OP(scan) != EXACTFAA_NO_TRIE) {
4432 /* Non-UTF-8 pattern, not EXACTFAA node. Look for the multi-char
4433 * folds that are all Latin1. As explained in the comments
4434 * preceding this function, we look also for the sharp s in EXACTF
4435 * and EXACTFL nodes; it can be in the final position. Otherwise
4436 * we can stop looking 1 byte earlier because have to find at least
4437 * two characters for a multi-fold */
4438 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
4443 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
4444 if (! len) { /* Not a multi-char fold. */
4445 if (*s == LATIN_SMALL_LETTER_SHARP_S
4446 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
4448 *unfolded_multi_char = TRUE;
4455 && isALPHA_FOLD_EQ(*s, 's')
4456 && isALPHA_FOLD_EQ(*(s+1), 's'))
4459 /* EXACTF nodes need to know that the minimum length
4460 * changed so that a sharp s in the string can match this
4461 * ss in the pattern, but they remain EXACTF nodes, as they
4462 * won't match this unless the target string is in UTF-8,
4463 * which we don't know until runtime. EXACTFL nodes can't
4464 * transform into EXACTFU nodes */
4465 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
4466 OP(scan) = EXACTFUP;
4470 *min_subtract += len - 1;
4478 /* Allow dumping but overwriting the collection of skipped
4479 * ops and/or strings with fake optimized ops */
4480 n = scan + NODE_SZ_STR(scan);
4488 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl", scan, depth, 0);});
4492 /* REx optimizer. Converts nodes into quicker variants "in place".
4493 Finds fixed substrings. */
4495 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
4496 to the position after last scanned or to NULL. */
4498 #define INIT_AND_WITHP \
4499 assert(!and_withp); \
4500 Newx(and_withp, 1, regnode_ssc); \
4501 SAVEFREEPV(and_withp)
4505 S_unwind_scan_frames(pTHX_ const void *p)
4507 scan_frame *f= (scan_frame *)p;
4509 scan_frame *n= f->next_frame;
4515 /* Follow the next-chain of the current node and optimize away
4516 all the NOTHINGs from it.
4519 S_rck_elide_nothing(pTHX_ regnode *node)
4521 PERL_ARGS_ASSERT_RCK_ELIDE_NOTHING;
4523 if (OP(node) != CURLYX) {
4524 const int max = (reg_off_by_arg[OP(node)]
4526 /* I32 may be smaller than U16 on CRAYs! */
4527 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
4528 int off = (reg_off_by_arg[OP(node)] ? ARG(node) : NEXT_OFF(node));
4532 /* Skip NOTHING and LONGJMP. */
4536 (PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
4537 || ((OP(n) == LONGJMP) && (noff = ARG(n)))
4543 if (reg_off_by_arg[OP(node)])
4546 NEXT_OFF(node) = off;
4551 /* the return from this sub is the minimum length that could possibly match */
4553 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
4554 SSize_t *minlenp, SSize_t *deltap,
4559 regnode_ssc *and_withp,
4560 U32 flags, U32 depth, bool was_mutate_ok)
4561 /* scanp: Start here (read-write). */
4562 /* deltap: Write maxlen-minlen here. */
4563 /* last: Stop before this one. */
4564 /* data: string data about the pattern */
4565 /* stopparen: treat close N as END */
4566 /* recursed: which subroutines have we recursed into */
4567 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
4569 SSize_t final_minlen;
4570 /* There must be at least this number of characters to match */
4573 regnode *scan = *scanp, *next;
4575 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
4576 int is_inf_internal = 0; /* The studied chunk is infinite */
4577 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
4578 scan_data_t data_fake;
4579 SV *re_trie_maxbuff = NULL;
4580 regnode *first_non_open = scan;
4581 SSize_t stopmin = OPTIMIZE_INFTY;
4582 scan_frame *frame = NULL;
4583 DECLARE_AND_GET_RE_DEBUG_FLAGS;
4585 PERL_ARGS_ASSERT_STUDY_CHUNK;
4586 RExC_study_started= 1;
4588 Zero(&data_fake, 1, scan_data_t);
4591 while (first_non_open && OP(first_non_open) == OPEN)
4592 first_non_open=regnext(first_non_open);
4598 RExC_study_chunk_recursed_count++;
4600 DEBUG_OPTIMISE_MORE_r(
4602 Perl_re_indentf( aTHX_ "study_chunk stopparen=%ld recursed_count=%lu depth=%lu recursed_depth=%lu scan=%p last=%p",
4603 depth, (long)stopparen,
4604 (unsigned long)RExC_study_chunk_recursed_count,
4605 (unsigned long)depth, (unsigned long)recursed_depth,
4608 if (recursed_depth) {
4611 for ( j = 0 ; j < recursed_depth ; j++ ) {
4612 for ( i = 0 ; i < (U32)RExC_total_parens ; i++ ) {
4613 if (PAREN_TEST(j, i) && (!j || !PAREN_TEST(j - 1, i))) {
4614 Perl_re_printf( aTHX_ " %d",(int)i);
4618 if ( j + 1 < recursed_depth ) {
4619 Perl_re_printf( aTHX_ ",");
4623 Perl_re_printf( aTHX_ "\n");
4626 while ( scan && OP(scan) != END && scan < last ){
4627 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
4628 node length to get a real minimum (because
4629 the folded version may be shorter) */
4630 bool unfolded_multi_char = FALSE;
4631 /* avoid mutating ops if we are anywhere within the recursed or
4632 * enframed handling for a GOSUB: the outermost level will handle it.
4634 bool mutate_ok = was_mutate_ok && !(frame && frame->in_gosub);
4635 /* Peephole optimizer: */
4636 DEBUG_STUDYDATA("Peep", data, depth, is_inf);
4637 DEBUG_PEEP("Peep", scan, depth, flags);
4640 /* The reason we do this here is that we need to deal with things like
4641 * /(?:f)(?:o)(?:o)/ which cant be dealt with by the normal EXACT
4642 * parsing code, as each (?:..) is handled by a different invocation of
4645 if (PL_regkind[OP(scan)] == EXACT
4646 && OP(scan) != LEXACT
4647 && OP(scan) != LEXACT_REQ8
4650 join_exact(pRExC_state, scan, &min_subtract, &unfolded_multi_char,
4651 0, NULL, depth + 1);
4654 /* Follow the next-chain of the current node and optimize
4655 away all the NOTHINGs from it.
4657 rck_elide_nothing(scan);
4659 /* The principal pseudo-switch. Cannot be a switch, since we look into
4660 * several different things. */
4661 if ( OP(scan) == DEFINEP ) {
4663 SSize_t deltanext = 0;
4664 SSize_t fake_last_close = 0;
4665 I32 f = SCF_IN_DEFINE;
4667 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4668 scan = regnext(scan);
4669 assert( OP(scan) == IFTHEN );
4670 DEBUG_PEEP("expect IFTHEN", scan, depth, flags);
4672 data_fake.last_closep= &fake_last_close;
4674 next = regnext(scan);
4675 scan = NEXTOPER(NEXTOPER(scan));
4676 DEBUG_PEEP("scan", scan, depth, flags);
4677 DEBUG_PEEP("next", next, depth, flags);
4679 /* we suppose the run is continuous, last=next...
4680 * NOTE we dont use the return here! */
4681 /* DEFINEP study_chunk() recursion */
4682 (void)study_chunk(pRExC_state, &scan, &minlen,
4683 &deltanext, next, &data_fake, stopparen,
4684 recursed_depth, NULL, f, depth+1, mutate_ok);
4689 OP(scan) == BRANCH ||
4690 OP(scan) == BRANCHJ ||
4693 next = regnext(scan);
4696 /* The op(next)==code check below is to see if we
4697 * have "BRANCH-BRANCH", "BRANCHJ-BRANCHJ", "IFTHEN-IFTHEN"
4698 * IFTHEN is special as it might not appear in pairs.
4699 * Not sure whether BRANCH-BRANCHJ is possible, regardless
4700 * we dont handle it cleanly. */
4701 if (OP(next) == code || code == IFTHEN) {
4702 /* NOTE - There is similar code to this block below for
4703 * handling TRIE nodes on a re-study. If you change stuff here
4704 * check there too. */
4705 SSize_t max1 = 0, min1 = OPTIMIZE_INFTY, num = 0;
4707 regnode * const startbranch=scan;
4709 if (flags & SCF_DO_SUBSTR) {
4710 /* Cannot merge strings after this. */
4711 scan_commit(pRExC_state, data, minlenp, is_inf);
4714 if (flags & SCF_DO_STCLASS)
4715 ssc_init_zero(pRExC_state, &accum);
4717 while (OP(scan) == code) {
4718 SSize_t deltanext, minnext, fake;
4720 regnode_ssc this_class;
4722 DEBUG_PEEP("Branch", scan, depth, flags);
4725 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4727 data_fake.whilem_c = data->whilem_c;
4728 data_fake.last_closep = data->last_closep;
4731 data_fake.last_closep = &fake;
4733 data_fake.pos_delta = delta;
4734 next = regnext(scan);
4736 scan = NEXTOPER(scan); /* everything */
4737 if (code != BRANCH) /* everything but BRANCH */
4738 scan = NEXTOPER(scan);
4740 if (flags & SCF_DO_STCLASS) {
4741 ssc_init(pRExC_state, &this_class);
4742 data_fake.start_class = &this_class;
4743 f = SCF_DO_STCLASS_AND;
4745 if (flags & SCF_WHILEM_VISITED_POS)
4746 f |= SCF_WHILEM_VISITED_POS;
4748 /* we suppose the run is continuous, last=next...*/
4749 /* recurse study_chunk() for each BRANCH in an alternation */
4750 minnext = study_chunk(pRExC_state, &scan, minlenp,
4751 &deltanext, next, &data_fake, stopparen,
4752 recursed_depth, NULL, f, depth+1,
4757 if (deltanext == OPTIMIZE_INFTY) {
4758 is_inf = is_inf_internal = 1;
4759 max1 = OPTIMIZE_INFTY;
4760 } else if (max1 < minnext + deltanext)
4761 max1 = minnext + deltanext;
4763 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4765 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4766 if ( stopmin > minnext)
4767 stopmin = min + min1;
4768 flags &= ~SCF_DO_SUBSTR;
4770 data->flags |= SCF_SEEN_ACCEPT;
4773 if (data_fake.flags & SF_HAS_EVAL)
4774 data->flags |= SF_HAS_EVAL;
4775 data->whilem_c = data_fake.whilem_c;
4777 if (flags & SCF_DO_STCLASS)
4778 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
4780 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
4782 if (flags & SCF_DO_SUBSTR) {
4783 data->pos_min += min1;
4784 if (data->pos_delta >= OPTIMIZE_INFTY - (max1 - min1))
4785 data->pos_delta = OPTIMIZE_INFTY;
4787 data->pos_delta += max1 - min1;
4788 if (max1 != min1 || is_inf)
4789 data->cur_is_floating = 1;
4792 if (delta == OPTIMIZE_INFTY
4793 || OPTIMIZE_INFTY - delta - (max1 - min1) < 0)
4794 delta = OPTIMIZE_INFTY;
4796 delta += max1 - min1;
4797 if (flags & SCF_DO_STCLASS_OR) {
4798 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
4800 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4801 flags &= ~SCF_DO_STCLASS;
4804 else if (flags & SCF_DO_STCLASS_AND) {
4806 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
4807 flags &= ~SCF_DO_STCLASS;
4810 /* Switch to OR mode: cache the old value of
4811 * data->start_class */
4813 StructCopy(data->start_class, and_withp, regnode_ssc);
4814 flags &= ~SCF_DO_STCLASS_AND;
4815 StructCopy(&accum, data->start_class, regnode_ssc);
4816 flags |= SCF_DO_STCLASS_OR;
4820 if (PERL_ENABLE_TRIE_OPTIMISATION
4821 && OP(startbranch) == BRANCH
4826 Assuming this was/is a branch we are dealing with: 'scan'
4827 now points at the item that follows the branch sequence,
4828 whatever it is. We now start at the beginning of the
4829 sequence and look for subsequences of
4835 which would be constructed from a pattern like
4838 If we can find such a subsequence we need to turn the first
4839 element into a trie and then add the subsequent branch exact
4840 strings to the trie.
4844 1. patterns where the whole set of branches can be
4847 2. patterns where only a subset can be converted.
4849 In case 1 we can replace the whole set with a single regop
4850 for the trie. In case 2 we need to keep the start and end
4853 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
4854 becomes BRANCH TRIE; BRANCH X;
4856 There is an additional case, that being where there is a
4857 common prefix, which gets split out into an EXACT like node
4858 preceding the TRIE node.
4860 If x(1..n)==tail then we can do a simple trie, if not we make
4861 a "jump" trie, such that when we match the appropriate word
4862 we "jump" to the appropriate tail node. Essentially we turn
4863 a nested if into a case structure of sorts.
4868 if (!re_trie_maxbuff) {
4869 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
4870 if (!SvIOK(re_trie_maxbuff))
4871 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
4873 if ( SvIV(re_trie_maxbuff)>=0 ) {
4875 regnode *first = (regnode *)NULL;
4876 regnode *prev = (regnode *)NULL;
4877 regnode *tail = scan;
4881 /* var tail is used because there may be a TAIL
4882 regop in the way. Ie, the exacts will point to the
4883 thing following the TAIL, but the last branch will
4884 point at the TAIL. So we advance tail. If we
4885 have nested (?:) we may have to move through several
4889 while ( OP( tail ) == TAIL ) {
4890 /* this is the TAIL generated by (?:) */
4891 tail = regnext( tail );
4895 DEBUG_TRIE_COMPILE_r({
4896 regprop(RExC_rx, RExC_mysv, tail, NULL, pRExC_state);
4897 Perl_re_indentf( aTHX_ "%s %" UVuf ":%s\n",
4899 "Looking for TRIE'able sequences. Tail node is ",
4900 (UV) REGNODE_OFFSET(tail),
4901 SvPV_nolen_const( RExC_mysv )
4907 Step through the branches
4908 cur represents each branch,
4909 noper is the first thing to be matched as part
4911 noper_next is the regnext() of that node.
4913 We normally handle a case like this
4914 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
4915 support building with NOJUMPTRIE, which restricts
4916 the trie logic to structures like /FOO|BAR/.
4918 If noper is a trieable nodetype then the branch is
4919 a possible optimization target. If we are building
4920 under NOJUMPTRIE then we require that noper_next is
4921 the same as scan (our current position in the regex
4924 Once we have two or more consecutive such branches
4925 we can create a trie of the EXACT's contents and
4926 stitch it in place into the program.
4928 If the sequence represents all of the branches in
4929 the alternation we replace the entire thing with a
4932 Otherwise when it is a subsequence we need to
4933 stitch it in place and replace only the relevant
4934 branches. This means the first branch has to remain
4935 as it is used by the alternation logic, and its
4936 next pointer, and needs to be repointed at the item
4937 on the branch chain following the last branch we
4938 have optimized away.
4940 This could be either a BRANCH, in which case the
4941 subsequence is internal, or it could be the item
4942 following the branch sequence in which case the
4943 subsequence is at the end (which does not
4944 necessarily mean the first node is the start of the
4947 TRIE_TYPE(X) is a define which maps the optype to a
4951 ----------------+-----------
4956 EXACTFU_REQ8 | EXACTFU
4960 EXACTFLU8 | EXACTFLU8
4964 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) \
4966 : ( EXACT == (X) || EXACT_REQ8 == (X) ) \
4968 : ( EXACTFU == (X) \
4969 || EXACTFU_REQ8 == (X) \
4970 || EXACTFUP == (X) ) \
4972 : ( EXACTFAA == (X) ) \
4974 : ( EXACTL == (X) ) \
4976 : ( EXACTFLU8 == (X) ) \
4980 /* dont use tail as the end marker for this traverse */
4981 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
4982 regnode * const noper = NEXTOPER( cur );
4983 U8 noper_type = OP( noper );
4984 U8 noper_trietype = TRIE_TYPE( noper_type );
4985 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
4986 regnode * const noper_next = regnext( noper );
4987 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4988 U8 noper_next_trietype = (noper_next && noper_next < tail) ? TRIE_TYPE( noper_next_type ) :0;
4991 DEBUG_TRIE_COMPILE_r({
4992 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4993 Perl_re_indentf( aTHX_ "- %d:%s (%d)",
4995 REG_NODE_NUM(cur), SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur) );
4997 regprop(RExC_rx, RExC_mysv, noper, NULL, pRExC_state);
4998 Perl_re_printf( aTHX_ " -> %d:%s",
4999 REG_NODE_NUM(noper), SvPV_nolen_const(RExC_mysv));
5002 regprop(RExC_rx, RExC_mysv, noper_next, NULL, pRExC_state);
5003 Perl_re_printf( aTHX_ "\t=> %d:%s\t",
5004 REG_NODE_NUM(noper_next), SvPV_nolen_const(RExC_mysv));
5006 Perl_re_printf( aTHX_ "(First==%d,Last==%d,Cur==%d,tt==%s,ntt==%s,nntt==%s)\n",
5007 REG_NODE_NUM(first), REG_NODE_NUM(prev), REG_NODE_NUM(cur),
5008 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
5012 /* Is noper a trieable nodetype that can be merged
5013 * with the current trie (if there is one)? */
5017 ( noper_trietype == NOTHING )
5018 || ( trietype == NOTHING )
5019 || ( trietype == noper_trietype )
5022 && noper_next >= tail
5026 /* Handle mergable triable node Either we are
5027 * the first node in a new trieable sequence,
5028 * in which case we do some bookkeeping,
5029 * otherwise we update the end pointer. */
5032 if ( noper_trietype == NOTHING ) {
5033 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
5034 regnode * const noper_next = regnext( noper );
5035 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
5036 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
5039 if ( noper_next_trietype ) {
5040 trietype = noper_next_trietype;
5041 } else if (noper_next_type) {
5042 /* a NOTHING regop is 1 regop wide.
5043 * We need at least two for a trie
5044 * so we can't merge this in */
5048 trietype = noper_trietype;
5051 if ( trietype == NOTHING )
5052 trietype = noper_trietype;
5057 } /* end handle mergable triable node */
5059 /* handle unmergable node -
5060 * noper may either be a triable node which can
5061 * not be tried together with the current trie,
5062 * or a non triable node */
5064 /* If last is set and trietype is not
5065 * NOTHING then we have found at least two
5066 * triable branch sequences in a row of a
5067 * similar trietype so we can turn them
5068 * into a trie. If/when we allow NOTHING to
5069 * start a trie sequence this condition
5070 * will be required, and it isn't expensive
5071 * so we leave it in for now. */
5072 if ( trietype && trietype != NOTHING )
5073 make_trie( pRExC_state,
5074 startbranch, first, cur, tail,
5075 count, trietype, depth+1 );
5076 prev = NULL; /* note: we clear/update
5077 first, trietype etc below,
5078 so we dont do it here */
5082 && noper_next >= tail
5085 /* noper is triable, so we can start a new
5089 trietype = noper_trietype;
5091 /* if we already saw a first but the
5092 * current node is not triable then we have
5093 * to reset the first information. */
5098 } /* end handle unmergable node */
5099 } /* loop over branches */
5100 DEBUG_TRIE_COMPILE_r({
5101 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
5102 Perl_re_indentf( aTHX_ "- %s (%d) <SCAN FINISHED> ",
5103 depth+1, SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur));
5104 Perl_re_printf( aTHX_ "(First==%d, Last==%d, Cur==%d, tt==%s)\n",
5105 REG_NODE_NUM(first), REG_NODE_NUM(prev), REG_NODE_NUM(cur),
5106 PL_reg_name[trietype]
5110 if ( prev && trietype ) {
5111 if ( trietype != NOTHING ) {
5112 /* the last branch of the sequence was part of
5113 * a trie, so we have to construct it here
5114 * outside of the loop */
5115 made= make_trie( pRExC_state, startbranch,
5116 first, scan, tail, count,
5117 trietype, depth+1 );
5118 #ifdef TRIE_STUDY_OPT
5119 if ( ((made == MADE_EXACT_TRIE &&
5120 startbranch == first)
5121 || ( first_non_open == first )) &&
5123 flags |= SCF_TRIE_RESTUDY;
5124 if ( startbranch == first
5127 RExC_seen &=~REG_TOP_LEVEL_BRANCHES_SEEN;
5132 /* at this point we know whatever we have is a
5133 * NOTHING sequence/branch AND if 'startbranch'
5134 * is 'first' then we can turn the whole thing
5137 if ( startbranch == first ) {
5139 /* the entire thing is a NOTHING sequence,
5140 * something like this: (?:|) So we can
5141 * turn it into a plain NOTHING op. */
5142 DEBUG_TRIE_COMPILE_r({
5143 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
5144 Perl_re_indentf( aTHX_ "- %s (%d) <NOTHING BRANCH SEQUENCE>\n",
5146 SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur));
5149 OP(startbranch)= NOTHING;
5150 NEXT_OFF(startbranch)= tail - startbranch;
5151 for ( opt= startbranch + 1; opt < tail ; opt++ )
5155 } /* end if ( prev) */
5156 } /* TRIE_MAXBUF is non zero */
5160 else if ( code == BRANCHJ ) { /* single branch is optimized. */
5161 scan = NEXTOPER(NEXTOPER(scan));
5162 } else /* single branch is optimized. */
5163 scan = NEXTOPER(scan);
5165 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB) {
5167 regnode *start = NULL;
5168 regnode *end = NULL;
5169 U32 my_recursed_depth= recursed_depth;
5171 if (OP(scan) != SUSPEND) { /* GOSUB */
5172 /* Do setup, note this code has side effects beyond
5173 * the rest of this block. Specifically setting
5174 * RExC_recurse[] must happen at least once during
5177 RExC_recurse[ARG2L(scan)] = scan;
5178 start = REGNODE_p(RExC_open_parens[paren]);
5179 end = REGNODE_p(RExC_close_parens[paren]);
5181 /* NOTE we MUST always execute the above code, even
5182 * if we do nothing with a GOSUB */
5184 ( flags & SCF_IN_DEFINE )
5187 (is_inf_internal || is_inf || (data && data->flags & SF_IS_INF))
5189 ( (flags & (SCF_DO_STCLASS | SCF_DO_SUBSTR)) == 0 )
5192 /* no need to do anything here if we are in a define. */
5193 /* or we are after some kind of infinite construct
5194 * so we can skip recursing into this item.
5195 * Since it is infinite we will not change the maxlen
5196 * or delta, and if we miss something that might raise
5197 * the minlen it will merely pessimise a little.
5199 * Iow /(?(DEFINE)(?<foo>foo|food))a+(?&foo)/
5200 * might result in a minlen of 1 and not of 4,
5201 * but this doesn't make us mismatch, just try a bit
5202 * harder than we should.
5204 * However we must assume this GOSUB is infinite, to
5205 * avoid wrongly applying other optimizations in the
5206 * enclosing scope - see GH 18096, for example.
5208 is_inf = is_inf_internal = 1;
5209 scan= regnext(scan);
5215 || !PAREN_TEST(recursed_depth - 1, paren)
5217 /* it is quite possible that there are more efficient ways
5218 * to do this. We maintain a bitmap per level of recursion
5219 * of which patterns we have entered so we can detect if a
5220 * pattern creates a possible infinite loop. When we
5221 * recurse down a level we copy the previous levels bitmap
5222 * down. When we are at recursion level 0 we zero the top
5223 * level bitmap. It would be nice to implement a different
5224 * more efficient way of doing this. In particular the top
5225 * level bitmap may be unnecessary.
5227 if (!recursed_depth) {
5228 Zero(RExC_study_chunk_recursed, RExC_study_chunk_recursed_bytes, U8);
5230 Copy(PAREN_OFFSET(recursed_depth - 1),
5231 PAREN_OFFSET(recursed_depth),
5232 RExC_study_chunk_recursed_bytes, U8);
5234 /* we havent recursed into this paren yet, so recurse into it */
5235 DEBUG_STUDYDATA("gosub-set", data, depth, is_inf);
5236 PAREN_SET(recursed_depth, paren);
5237 my_recursed_depth= recursed_depth + 1;
5239 DEBUG_STUDYDATA("gosub-inf", data, depth, is_inf);
5240 /* some form of infinite recursion, assume infinite length
5242 if (flags & SCF_DO_SUBSTR) {
5243 scan_commit(pRExC_state, data, minlenp, is_inf);
5244 data->cur_is_floating = 1;
5246 is_inf = is_inf_internal = 1;
5247 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5248 ssc_anything(data->start_class);
5249 flags &= ~SCF_DO_STCLASS;
5251 start= NULL; /* reset start so we dont recurse later on. */
5256 end = regnext(scan);
5259 scan_frame *newframe;
5261 if (!RExC_frame_last) {
5262 Newxz(newframe, 1, scan_frame);
5263 SAVEDESTRUCTOR_X(S_unwind_scan_frames, newframe);
5264 RExC_frame_head= newframe;
5266 } else if (!RExC_frame_last->next_frame) {
5267 Newxz(newframe, 1, scan_frame);
5268 RExC_frame_last->next_frame= newframe;
5269 newframe->prev_frame= RExC_frame_last;
5272 newframe= RExC_frame_last->next_frame;
5274 RExC_frame_last= newframe;
5276 newframe->next_regnode = regnext(scan);
5277 newframe->last_regnode = last;
5278 newframe->stopparen = stopparen;
5279 newframe->prev_recursed_depth = recursed_depth;
5280 newframe->this_prev_frame= frame;
5281 newframe->in_gosub = (
5282 (frame && frame->in_gosub) || OP(scan) == GOSUB
5285 DEBUG_STUDYDATA("frame-new", data, depth, is_inf);
5286 DEBUG_PEEP("fnew", scan, depth, flags);
5293 recursed_depth= my_recursed_depth;
5298 else if (PL_regkind[OP(scan)] == EXACT && ! isEXACTFish(OP(scan))) {
5299 SSize_t bytelen = STR_LEN(scan), charlen;
5303 const U8 * const s = (U8*)STRING(scan);
5304 uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
5305 charlen = utf8_length(s, s + bytelen);
5307 uc = *((U8*)STRING(scan));
5311 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
5312 /* The code below prefers earlier match for fixed
5313 offset, later match for variable offset. */
5314 if (data->last_end == -1) { /* Update the start info. */
5315 data->last_start_min = data->pos_min;
5316 data->last_start_max =
5317 is_inf ? OPTIMIZE_INFTY
5318 : (data->pos_delta > OPTIMIZE_INFTY - data->pos_min)
5319 ? OPTIMIZE_INFTY : data->pos_min + data->pos_delta;
5321 sv_catpvn(data->last_found, STRING(scan), bytelen);
5323 SvUTF8_on(data->last_found);
5325 SV * const sv = data->last_found;
5326 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
5327 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5328 if (mg && mg->mg_len >= 0)
5329 mg->mg_len += charlen;
5331 data->last_end = data->pos_min + charlen;
5332 data->pos_min += charlen; /* As in the first entry. */
5333 data->flags &= ~SF_BEFORE_EOL;
5336 /* ANDing the code point leaves at most it, and not in locale, and
5337 * can't match null string */
5338 if (flags & SCF_DO_STCLASS_AND) {
5339 ssc_cp_and(data->start_class, uc);
5340 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5341 ssc_clear_locale(data->start_class);
5343 else if (flags & SCF_DO_STCLASS_OR) {
5344 ssc_add_cp(data->start_class, uc);
5345 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5347 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5348 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5350 flags &= ~SCF_DO_STCLASS;
5352 else if (PL_regkind[OP(scan)] == EXACT) {
5353 /* But OP != EXACT!, so is EXACTFish */
5354 SSize_t bytelen = STR_LEN(scan), charlen;
5355 const U8 * s = (U8*)STRING(scan);
5357 /* Replace a length 1 ASCII fold pair node with an ANYOFM node,
5358 * with the mask set to the complement of the bit that differs
5359 * between upper and lower case, and the lowest code point of the
5360 * pair (which the '&' forces) */
5363 && ( OP(scan) == EXACTFAA
5364 || ( OP(scan) == EXACTFU
5365 && ! HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(*s)))
5368 U8 mask = ~ ('A' ^ 'a'); /* These differ in just one bit */
5371 ARG_SET(scan, *s & mask);
5373 /* we're not EXACTFish any more, so restudy */
5377 /* Search for fixed substrings supports EXACT only. */
5378 if (flags & SCF_DO_SUBSTR) {
5380 scan_commit(pRExC_state, data, minlenp, is_inf);
5382 charlen = UTF ? (SSize_t) utf8_length(s, s + bytelen) : bytelen;
5383 if (unfolded_multi_char) {
5384 RExC_seen |= REG_UNFOLDED_MULTI_SEEN;
5386 min += charlen - min_subtract;
5388 delta += min_subtract;
5389 if (flags & SCF_DO_SUBSTR) {
5390 data->pos_min += charlen - min_subtract;
5391 if (data->pos_min < 0) {
5394 data->pos_delta += min_subtract;
5396 data->cur_is_floating = 1; /* float */
5400 if (flags & SCF_DO_STCLASS) {
5401 SV* EXACTF_invlist = make_exactf_invlist(pRExC_state, scan);
5403 assert(EXACTF_invlist);
5404 if (flags & SCF_DO_STCLASS_AND) {
5405 if (OP(scan) != EXACTFL)
5406 ssc_clear_locale(data->start_class);
5407 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5408 ANYOF_POSIXL_ZERO(data->start_class);
5409 ssc_intersection(data->start_class, EXACTF_invlist, FALSE);
5411 else { /* SCF_DO_STCLASS_OR */
5412 ssc_union(data->start_class, EXACTF_invlist, FALSE);
5413 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5415 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5416 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5418 flags &= ~SCF_DO_STCLASS;
5419 SvREFCNT_dec(EXACTF_invlist);
5422 else if (REGNODE_VARIES(OP(scan))) {
5423 SSize_t mincount, maxcount, minnext, deltanext, pos_before = 0;
5424 I32 fl = 0, f = flags;
5425 regnode * const oscan = scan;
5426 regnode_ssc this_class;
5427 regnode_ssc *oclass = NULL;
5428 I32 next_is_eval = 0;
5430 switch (PL_regkind[OP(scan)]) {
5431 case WHILEM: /* End of (?:...)* . */
5432 scan = NEXTOPER(scan);
5435 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
5436 next = NEXTOPER(scan);
5437 if ( ( PL_regkind[OP(next)] == EXACT
5438 && ! isEXACTFish(OP(next)))
5439 || (flags & SCF_DO_STCLASS))
5442 maxcount = REG_INFTY;
5443 next = regnext(scan);
5444 scan = NEXTOPER(scan);
5448 if (flags & SCF_DO_SUBSTR)
5450 /* This will bypass the formal 'min += minnext * mincount'
5451 * calculation in the do_curly path, so assumes min width
5452 * of the PLUS payload is exactly one. */
5456 next = NEXTOPER(scan);
5458 /* This temporary node can now be turned into EXACTFU, and
5459 * must, as regexec.c doesn't handle it */
5460 if (OP(next) == EXACTFU_S_EDGE && mutate_ok) {
5464 if ( STR_LEN(next) == 1
5465 && isALPHA_A(* STRING(next))
5466 && ( OP(next) == EXACTFAA
5467 || ( OP(next) == EXACTFU
5468 && ! HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(* STRING(next))))
5471 /* These differ in just one bit */
5472 U8 mask = ~ ('A' ^ 'a');
5474 assert(isALPHA_A(* STRING(next)));
5476 /* Then replace it by an ANYOFM node, with
5477 * the mask set to the complement of the
5478 * bit that differs between upper and lower
5479 * case, and the lowest code point of the
5480 * pair (which the '&' forces) */
5482 ARG_SET(next, *STRING(next) & mask);
5486 if (flags & SCF_DO_STCLASS) {
5488 maxcount = REG_INFTY;
5489 next = regnext(scan);
5490 scan = NEXTOPER(scan);
5493 if (flags & SCF_DO_SUBSTR) {
5494 scan_commit(pRExC_state, data, minlenp, is_inf);
5495 /* Cannot extend fixed substrings */
5496 data->cur_is_floating = 1; /* float */
5498 is_inf = is_inf_internal = 1;
5499 scan = regnext(scan);
5500 goto optimize_curly_tail;
5502 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
5503 && (scan->flags == stopparen))
5508 mincount = ARG1(scan);
5509 maxcount = ARG2(scan);
5511 next = regnext(scan);
5512 if (OP(scan) == CURLYX) {
5513 I32 lp = (data ? *(data->last_closep) : 0);
5514 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
5516 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
5517 next_is_eval = (OP(scan) == EVAL);
5519 if (flags & SCF_DO_SUBSTR) {
5521 scan_commit(pRExC_state, data, minlenp, is_inf);
5522 /* Cannot extend fixed substrings */
5523 pos_before = data->pos_min;
5527 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
5529 data->flags |= SF_IS_INF;
5531 if (flags & SCF_DO_STCLASS) {
5532 ssc_init(pRExC_state, &this_class);
5533 oclass = data->start_class;
5534 data->start_class = &this_class;
5535 f |= SCF_DO_STCLASS_AND;
5536 f &= ~SCF_DO_STCLASS_OR;
5538 /* Exclude from super-linear cache processing any {n,m}
5539 regops for which the combination of input pos and regex
5540 pos is not enough information to determine if a match
5543 For example, in the regex /foo(bar\s*){4,8}baz/ with the
5544 regex pos at the \s*, the prospects for a match depend not
5545 only on the input position but also on how many (bar\s*)
5546 repeats into the {4,8} we are. */
5547 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
5548 f &= ~SCF_WHILEM_VISITED_POS;
5550 /* This will finish on WHILEM, setting scan, or on NULL: */
5551 /* recurse study_chunk() on loop bodies */
5552 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
5553 last, data, stopparen, recursed_depth, NULL,
5555 ? (f & ~SCF_DO_SUBSTR)
5557 , depth+1, mutate_ok);
5559 if (flags & SCF_DO_STCLASS)
5560 data->start_class = oclass;
5561 if (mincount == 0 || minnext == 0) {
5562 if (flags & SCF_DO_STCLASS_OR) {
5563 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5565 else if (flags & SCF_DO_STCLASS_AND) {
5566 /* Switch to OR mode: cache the old value of
5567 * data->start_class */
5569 StructCopy(data->start_class, and_withp, regnode_ssc);
5570 flags &= ~SCF_DO_STCLASS_AND;
5571 StructCopy(&this_class, data->start_class, regnode_ssc);
5572 flags |= SCF_DO_STCLASS_OR;
5573 ANYOF_FLAGS(data->start_class)
5574 |= SSC_MATCHES_EMPTY_STRING;
5576 } else { /* Non-zero len */
5577 if (flags & SCF_DO_STCLASS_OR) {
5578 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5579 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5581 else if (flags & SCF_DO_STCLASS_AND)
5582 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5583 flags &= ~SCF_DO_STCLASS;
5585 if (!scan) /* It was not CURLYX, but CURLY. */
5587 if (((flags & (SCF_TRIE_DOING_RESTUDY|SCF_DO_SUBSTR))==SCF_DO_SUBSTR)
5588 /* ? quantifier ok, except for (?{ ... }) */
5589 && (next_is_eval || !(mincount == 0 && maxcount == 1))
5590 && (minnext == 0) && (deltanext == 0)
5591 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
5592 && maxcount <= REG_INFTY/3) /* Complement check for big
5595 _WARN_HELPER(RExC_precomp_end, packWARN(WARN_REGEXP),
5596 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP),
5597 "Quantifier unexpected on zero-length expression "
5598 "in regex m/%" UTF8f "/",
5599 UTF8fARG(UTF, RExC_precomp_end - RExC_precomp,
5603 if ( ( minnext > 0 && mincount >= SSize_t_MAX / minnext )
5604 || min >= SSize_t_MAX - minnext * mincount )
5606 FAIL("Regexp out of space");
5609 min += minnext * mincount;
5610 is_inf_internal |= deltanext == OPTIMIZE_INFTY
5611 || (maxcount == REG_INFTY && minnext + deltanext > 0);
5612 is_inf |= is_inf_internal;
5614 delta = OPTIMIZE_INFTY;
5616 delta += (minnext + deltanext) * maxcount
5617 - minnext * mincount;
5619 /* Try powerful optimization CURLYX => CURLYN. */
5620 if ( OP(oscan) == CURLYX && data
5621 && data->flags & SF_IN_PAR
5622 && !(data->flags & SF_HAS_EVAL)
5623 && !deltanext && minnext == 1
5626 /* Try to optimize to CURLYN. */
5627 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
5628 regnode * const nxt1 = nxt;
5635 if (!REGNODE_SIMPLE(OP(nxt))
5636 && !(PL_regkind[OP(nxt)] == EXACT
5637 && STR_LEN(nxt) == 1))
5643 if (OP(nxt) != CLOSE)
5645 if (RExC_open_parens) {
5648 RExC_open_parens[ARG(nxt1)] = REGNODE_OFFSET(oscan);
5651 RExC_close_parens[ARG(nxt1)] = REGNODE_OFFSET(nxt) + 2;
5653 /* Now we know that nxt2 is the only contents: */
5654 oscan->flags = (U8)ARG(nxt);
5656 OP(nxt1) = NOTHING; /* was OPEN. */
5659 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5660 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
5661 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
5662 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5663 OP(nxt + 1) = OPTIMIZED; /* was count. */
5664 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
5669 /* Try optimization CURLYX => CURLYM. */
5670 if ( OP(oscan) == CURLYX && data
5671 && !(data->flags & SF_HAS_PAR)
5672 && !(data->flags & SF_HAS_EVAL)
5673 && !deltanext /* atom is fixed width */
5674 && minnext != 0 /* CURLYM can't handle zero width */
5675 /* Nor characters whose fold at run-time may be
5676 * multi-character */
5677 && ! (RExC_seen & REG_UNFOLDED_MULTI_SEEN)
5680 /* XXXX How to optimize if data == 0? */
5681 /* Optimize to a simpler form. */
5682 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
5686 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
5687 && (OP(nxt2) != WHILEM))
5689 OP(nxt2) = SUCCEED; /* Whas WHILEM */
5690 /* Need to optimize away parenths. */
5691 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
5692 /* Set the parenth number. */
5693 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
5695 oscan->flags = (U8)ARG(nxt);
5696 if (RExC_open_parens) {
5698 RExC_open_parens[ARG(nxt1)] = REGNODE_OFFSET(oscan);
5701 RExC_close_parens[ARG(nxt1)] = REGNODE_OFFSET(nxt2)
5704 OP(nxt1) = OPTIMIZED; /* was OPEN. */
5705 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5708 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5709 OP(nxt + 1) = OPTIMIZED; /* was count. */
5710 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
5711 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
5714 while ( nxt1 && (OP(nxt1) != WHILEM)) {
5715 regnode *nnxt = regnext(nxt1);
5717 if (reg_off_by_arg[OP(nxt1)])
5718 ARG_SET(nxt1, nxt2 - nxt1);
5719 else if (nxt2 - nxt1 < U16_MAX)
5720 NEXT_OFF(nxt1) = nxt2 - nxt1;
5722 OP(nxt) = NOTHING; /* Cannot beautify */
5727 /* Optimize again: */
5728 /* recurse study_chunk() on optimised CURLYX => CURLYM */
5729 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
5730 NULL, stopparen, recursed_depth, NULL, 0,
5731 depth+1, mutate_ok);
5736 else if ((OP(oscan) == CURLYX)
5737 && (flags & SCF_WHILEM_VISITED_POS)
5738 /* See the comment on a similar expression above.
5739 However, this time it's not a subexpression
5740 we care about, but the expression itself. */
5741 && (maxcount == REG_INFTY)
5743 /* This stays as CURLYX, we can put the count/of pair. */
5744 /* Find WHILEM (as in regexec.c) */
5745 regnode *nxt = oscan + NEXT_OFF(oscan);
5747 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
5749 nxt = PREVOPER(nxt);
5750 if (nxt->flags & 0xf) {
5751 /* we've already set whilem count on this node */
5752 } else if (++data->whilem_c < 16) {
5753 assert(data->whilem_c <= RExC_whilem_seen);
5754 nxt->flags = (U8)(data->whilem_c
5755 | (RExC_whilem_seen << 4)); /* On WHILEM */
5758 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
5760 if (flags & SCF_DO_SUBSTR) {
5761 SV *last_str = NULL;
5762 STRLEN last_chrs = 0;
5763 int counted = mincount != 0;
5765 if (data->last_end > 0 && mincount != 0) { /* Ends with a
5767 SSize_t b = pos_before >= data->last_start_min
5768 ? pos_before : data->last_start_min;
5770 const char * const s = SvPV_const(data->last_found, l);
5771 SSize_t old = b - data->last_start_min;
5775 old = utf8_hop_forward((U8*)s, old,
5776 (U8 *) SvEND(data->last_found))
5779 /* Get the added string: */
5780 last_str = newSVpvn_utf8(s + old, l, UTF);
5781 last_chrs = UTF ? utf8_length((U8*)(s + old),
5782 (U8*)(s + old + l)) : l;
5783 if (deltanext == 0 && pos_before == b) {
5784 /* What was added is a constant string */
5787 SvGROW(last_str, (mincount * l) + 1);
5788 repeatcpy(SvPVX(last_str) + l,
5789 SvPVX_const(last_str), l,
5791 SvCUR_set(last_str, SvCUR(last_str) * mincount);
5792 /* Add additional parts. */
5793 SvCUR_set(data->last_found,
5794 SvCUR(data->last_found) - l);
5795 sv_catsv(data->last_found, last_str);
5797 SV * sv = data->last_found;
5799 SvUTF8(sv) && SvMAGICAL(sv) ?
5800 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5801 if (mg && mg->mg_len >= 0)
5802 mg->mg_len += last_chrs * (mincount-1);
5804 last_chrs *= mincount;
5805 data->last_end += l * (mincount - 1);
5808 /* start offset must point into the last copy */
5809 data->last_start_min += minnext * (mincount - 1);
5810 data->last_start_max =
5813 : data->last_start_max +
5814 (maxcount - 1) * (minnext + data->pos_delta);
5817 /* It is counted once already... */
5818 data->pos_min += minnext * (mincount - counted);
5820 Perl_re_printf( aTHX_ "counted=%" UVuf " deltanext=%" UVuf
5821 " OPTIMIZE_INFTY=%" UVuf " minnext=%" UVuf
5822 " maxcount=%" UVuf " mincount=%" UVuf "\n",
5823 (UV)counted, (UV)deltanext, (UV)OPTIMIZE_INFTY, (UV)minnext, (UV)maxcount,
5825 if (deltanext != OPTIMIZE_INFTY)
5826 Perl_re_printf( aTHX_ "LHS=%" UVuf " RHS=%" UVuf "\n",
5827 (UV)(-counted * deltanext + (minnext + deltanext) * maxcount
5828 - minnext * mincount), (UV)(OPTIMIZE_INFTY - data->pos_delta));
5830 if (deltanext == OPTIMIZE_INFTY
5831 || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= OPTIMIZE_INFTY - data->pos_delta)
5832 data->pos_delta = OPTIMIZE_INFTY;
5834 data->pos_delta += - counted * deltanext +
5835 (minnext + deltanext) * maxcount - minnext * mincount;
5836 if (mincount != maxcount) {
5837 /* Cannot extend fixed substrings found inside
5839 scan_commit(pRExC_state, data, minlenp, is_inf);
5840 if (mincount && last_str) {
5841 SV * const sv = data->last_found;
5842 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
5843 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5847 sv_setsv(sv, last_str);
5848 data->last_end = data->pos_min;
5849 data->last_start_min = data->pos_min - last_chrs;
5850 data->last_start_max = is_inf
5852 : data->pos_min + data->pos_delta - last_chrs;
5854 data->cur_is_floating = 1; /* float */
5856 SvREFCNT_dec(last_str);
5858 if (data && (fl & SF_HAS_EVAL))
5859 data->flags |= SF_HAS_EVAL;
5860 optimize_curly_tail:
5861 rck_elide_nothing(oscan);
5865 Perl_croak(aTHX_ "panic: unexpected varying REx opcode %d",
5869 if (flags & SCF_DO_SUBSTR) {
5870 /* Cannot expect anything... */
5871 scan_commit(pRExC_state, data, minlenp, is_inf);
5872 data->cur_is_floating = 1; /* float */
5874 is_inf = is_inf_internal = 1;
5875 if (flags & SCF_DO_STCLASS_OR) {
5876 if (OP(scan) == CLUMP) {
5877 /* Actually is any start char, but very few code points
5878 * aren't start characters */
5879 ssc_match_all_cp(data->start_class);
5882 ssc_anything(data->start_class);
5885 flags &= ~SCF_DO_STCLASS;
5889 else if (OP(scan) == LNBREAK) {
5890 if (flags & SCF_DO_STCLASS) {
5891 if (flags & SCF_DO_STCLASS_AND) {
5892 ssc_intersection(data->start_class,
5893 PL_XPosix_ptrs[_CC_VERTSPACE], FALSE);
5894 ssc_clear_locale(data->start_class);
5895 ANYOF_FLAGS(data->start_class)
5896 &= ~SSC_MATCHES_EMPTY_STRING;
5898 else if (flags & SCF_DO_STCLASS_OR) {
5899 ssc_union(data->start_class,
5900 PL_XPosix_ptrs[_CC_VERTSPACE],
5902 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5904 /* See commit msg for
5905 * 749e076fceedeb708a624933726e7989f2302f6a */
5906 ANYOF_FLAGS(data->start_class)
5907 &= ~SSC_MATCHES_EMPTY_STRING;
5909 flags &= ~SCF_DO_STCLASS;
5912 if (delta != OPTIMIZE_INFTY)
5913 delta++; /* Because of the 2 char string cr-lf */
5914 if (flags & SCF_DO_SUBSTR) {
5915 /* Cannot expect anything... */
5916 scan_commit(pRExC_state, data, minlenp, is_inf);
5918 if (data->pos_delta != OPTIMIZE_INFTY) {
5919 data->pos_delta += 1;
5921 data->cur_is_floating = 1; /* float */
5924 else if (REGNODE_SIMPLE(OP(scan))) {
5926 if (flags & SCF_DO_SUBSTR) {
5927 scan_commit(pRExC_state, data, minlenp, is_inf);
5931 if (flags & SCF_DO_STCLASS) {
5933 SV* my_invlist = NULL;
5936 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5937 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5939 /* Some of the logic below assumes that switching
5940 locale on will only add false positives. */
5945 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d",
5949 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5950 ssc_match_all_cp(data->start_class);
5955 SV* REG_ANY_invlist = _new_invlist(2);
5956 REG_ANY_invlist = add_cp_to_invlist(REG_ANY_invlist,
5958 if (flags & SCF_DO_STCLASS_OR) {
5959 ssc_union(data->start_class,
5961 TRUE /* TRUE => invert, hence all but \n
5965 else if (flags & SCF_DO_STCLASS_AND) {
5966 ssc_intersection(data->start_class,
5968 TRUE /* TRUE => invert */
5970 ssc_clear_locale(data->start_class);
5972 SvREFCNT_dec_NN(REG_ANY_invlist);
5984 if (flags & SCF_DO_STCLASS_AND)
5985 ssc_and(pRExC_state, data->start_class,
5986 (regnode_charclass *) scan);
5988 ssc_or(pRExC_state, data->start_class,
5989 (regnode_charclass *) scan);
5992 case NANYOFM: /* NANYOFM already contains the inversion of the
5993 input ANYOF data, so, unlike things like
5994 NPOSIXA, don't change 'invert' to TRUE */
5998 SV* cp_list = get_ANYOFM_contents(scan);
6000 if (flags & SCF_DO_STCLASS_OR) {
6001 ssc_union(data->start_class, cp_list, invert);
6003 else if (flags & SCF_DO_STCLASS_AND) {
6004 ssc_intersection(data->start_class, cp_list, invert);
6007 SvREFCNT_dec_NN(cp_list);
6016 cp_list = _add_range_to_invlist(cp_list,
6018 ANYOFRbase(scan) + ANYOFRdelta(scan));
6020 if (flags & SCF_DO_STCLASS_OR) {
6021 ssc_union(data->start_class, cp_list, invert);
6023 else if (flags & SCF_DO_STCLASS_AND) {
6024 ssc_intersection(data->start_class, cp_list, invert);
6027 SvREFCNT_dec_NN(cp_list);
6036 namedclass = classnum_to_namedclass(FLAGS(scan)) + invert;
6037 if (flags & SCF_DO_STCLASS_AND) {
6038 bool was_there = cBOOL(
6039 ANYOF_POSIXL_TEST(data->start_class,
6041 ANYOF_POSIXL_ZERO(data->start_class);
6042 if (was_there) { /* Do an AND */
6043 ANYOF_POSIXL_SET(data->start_class, namedclass);
6045 /* No individual code points can now match */
6046 data->start_class->invlist
6047 = sv_2mortal(_new_invlist(0));
6050 int complement = namedclass + ((invert) ? -1 : 1);
6052 assert(flags & SCF_DO_STCLASS_OR);
6054 /* If the complement of this class was already there,
6055 * the result is that they match all code points,
6056 * (\d + \D == everything). Remove the classes from
6057 * future consideration. Locale is not relevant in
6059 if (ANYOF_POSIXL_TEST(data->start_class, complement)) {
6060 ssc_match_all_cp(data->start_class);
6061 ANYOF_POSIXL_CLEAR(data->start_class, namedclass);
6062 ANYOF_POSIXL_CLEAR(data->start_class, complement);
6064 else { /* The usual case; just add this class to the
6066 ANYOF_POSIXL_SET(data->start_class, namedclass);
6071 case NPOSIXA: /* For these, we always know the exact set of
6076 my_invlist = invlist_clone(PL_Posix_ptrs[FLAGS(scan)], NULL);
6077 goto join_posix_and_ascii;
6085 my_invlist = invlist_clone(PL_XPosix_ptrs[FLAGS(scan)], NULL);
6087 /* NPOSIXD matches all upper Latin1 code points unless the
6088 * target string being matched is UTF-8, which is
6089 * unknowable until match time. Since we are going to
6090 * invert, we want to get rid of all of them so that the
6091 * inversion will match all */
6092 if (OP(scan) == NPOSIXD) {
6093 _invlist_subtract(my_invlist, PL_UpperLatin1,
6097 join_posix_and_ascii:
6099 if (flags & SCF_DO_STCLASS_AND) {
6100 ssc_intersection(data->start_class, my_invlist, invert);
6101 ssc_clear_locale(data->start_class);
6104 assert(flags & SCF_DO_STCLASS_OR);
6105 ssc_union(data->start_class, my_invlist, invert);
6107 SvREFCNT_dec(my_invlist);
6109 if (flags & SCF_DO_STCLASS_OR)
6110 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6111 flags &= ~SCF_DO_STCLASS;
6114 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
6115 data->flags |= (OP(scan) == MEOL
6118 scan_commit(pRExC_state, data, minlenp, is_inf);
6121 else if ( PL_regkind[OP(scan)] == BRANCHJ
6122 /* Lookbehind, or need to calculate parens/evals/stclass: */
6123 && (scan->flags || data || (flags & SCF_DO_STCLASS))
6124 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM))
6126 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
6127 || OP(scan) == UNLESSM )
6129 /* Negative Lookahead/lookbehind
6130 In this case we can't do fixed string optimisation.
6133 SSize_t deltanext, minnext, fake = 0;
6138 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
6140 data_fake.whilem_c = data->whilem_c;
6141 data_fake.last_closep = data->last_closep;
6144 data_fake.last_closep = &fake;
6145 data_fake.pos_delta = delta;
6146 if ( flags & SCF_DO_STCLASS && !scan->flags
6147 && OP(scan) == IFMATCH ) { /* Lookahead */
6148 ssc_init(pRExC_state, &intrnl);
6149 data_fake.start_class = &intrnl;
6150 f |= SCF_DO_STCLASS_AND;
6152 if (flags & SCF_WHILEM_VISITED_POS)
6153 f |= SCF_WHILEM_VISITED_POS;
6154 next = regnext(scan);
6155 nscan = NEXTOPER(NEXTOPER(scan));
6157 /* recurse study_chunk() for lookahead body */
6158 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
6159 last, &data_fake, stopparen,
6160 recursed_depth, NULL, f, depth+1,
6164 || deltanext > (I32) U8_MAX
6165 || minnext > (I32)U8_MAX
6166 || minnext + deltanext > (I32)U8_MAX)
6168 FAIL2("Lookbehind longer than %" UVuf " not implemented",
6172 /* The 'next_off' field has been repurposed to count the
6173 * additional starting positions to try beyond the initial
6174 * one. (This leaves it at 0 for non-variable length
6175 * matches to avoid breakage for those not using this
6178 scan->next_off = deltanext;
6179 ckWARNexperimental(RExC_parse,
6180 WARN_EXPERIMENTAL__VLB,
6181 "Variable length lookbehind is experimental");
6183 scan->flags = (U8)minnext + deltanext;
6186 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
6188 if (data_fake.flags & SF_HAS_EVAL)
6189 data->flags |= SF_HAS_EVAL;
6190 data->whilem_c = data_fake.whilem_c;
6192 if (f & SCF_DO_STCLASS_AND) {
6193 if (flags & SCF_DO_STCLASS_OR) {
6194 /* OR before, AND after: ideally we would recurse with
6195 * data_fake to get the AND applied by study of the
6196 * remainder of the pattern, and then derecurse;
6197 * *** HACK *** for now just treat as "no information".
6198 * See [perl #56690].
6200 ssc_init(pRExC_state, data->start_class);
6202 /* AND before and after: combine and continue. These
6203 * assertions are zero-length, so can match an EMPTY
6205 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
6206 ANYOF_FLAGS(data->start_class)
6207 |= SSC_MATCHES_EMPTY_STRING;
6211 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
6213 /* Positive Lookahead/lookbehind
6214 In this case we can do fixed string optimisation,
6215 but we must be careful about it. Note in the case of
6216 lookbehind the positions will be offset by the minimum
6217 length of the pattern, something we won't know about
6218 until after the recurse.
6220 SSize_t deltanext, fake = 0;
6224 /* We use SAVEFREEPV so that when the full compile
6225 is finished perl will clean up the allocated
6226 minlens when it's all done. This way we don't
6227 have to worry about freeing them when we know
6228 they wont be used, which would be a pain.
6231 Newx( minnextp, 1, SSize_t );
6232 SAVEFREEPV(minnextp);
6235 StructCopy(data, &data_fake, scan_data_t);
6236 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
6239 scan_commit(pRExC_state, &data_fake, minlenp, is_inf);
6240 data_fake.last_found=newSVsv(data->last_found);
6244 data_fake.last_closep = &fake;
6245 data_fake.flags = 0;
6246 data_fake.substrs[0].flags = 0;
6247 data_fake.substrs[1].flags = 0;
6248 data_fake.pos_delta = delta;
6250 data_fake.flags |= SF_IS_INF;
6251 if ( flags & SCF_DO_STCLASS && !scan->flags
6252 && OP(scan) == IFMATCH ) { /* Lookahead */
6253 ssc_init(pRExC_state, &intrnl);
6254 data_fake.start_class = &intrnl;
6255 f |= SCF_DO_STCLASS_AND;
6257 if (flags & SCF_WHILEM_VISITED_POS)
6258 f |= SCF_WHILEM_VISITED_POS;
6259 next = regnext(scan);
6260 nscan = NEXTOPER(NEXTOPER(scan));
6262 /* positive lookahead study_chunk() recursion */
6263 *minnextp = study_chunk(pRExC_state, &nscan, minnextp,
6264 &deltanext, last, &data_fake,
6265 stopparen, recursed_depth, NULL,
6266 f, depth+1, mutate_ok);
6268 assert(0); /* This code has never been tested since this
6269 is normally not compiled */
6271 || deltanext > (I32) U8_MAX
6272 || *minnextp > (I32)U8_MAX
6273 || *minnextp + deltanext > (I32)U8_MAX)
6275 FAIL2("Lookbehind longer than %" UVuf " not implemented",
6280 scan->next_off = deltanext;
6282 scan->flags = (U8)*minnextp + deltanext;
6287 if (f & SCF_DO_STCLASS_AND) {
6288 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
6289 ANYOF_FLAGS(data->start_class) |= SSC_MATCHES_EMPTY_STRING;
6292 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
6294 if (data_fake.flags & SF_HAS_EVAL)
6295 data->flags |= SF_HAS_EVAL;
6296 data->whilem_c = data_fake.whilem_c;
6297 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
6299 if (RExC_rx->minlen<*minnextp)
6300 RExC_rx->minlen=*minnextp;
6301 scan_commit(pRExC_state, &data_fake, minnextp, is_inf);
6302 SvREFCNT_dec_NN(data_fake.last_found);
6304 for (i = 0; i < 2; i++) {
6305 if (data_fake.substrs[i].minlenp != minlenp) {
6306 data->substrs[i].min_offset =
6307 data_fake.substrs[i].min_offset;
6308 data->substrs[i].max_offset =
6309 data_fake.substrs[i].max_offset;
6310 data->substrs[i].minlenp =
6311 data_fake.substrs[i].minlenp;
6312 data->substrs[i].lookbehind += scan->flags;
6320 else if (OP(scan) == OPEN) {
6321 if (stopparen != (I32)ARG(scan))
6324 else if (OP(scan) == CLOSE) {
6325 if (stopparen == (I32)ARG(scan)) {
6328 if ((I32)ARG(scan) == is_par) {
6329 next = regnext(scan);
6331 if ( next && (OP(next) != WHILEM) && next < last)
6332 is_par = 0; /* Disable optimization */
6335 *(data->last_closep) = ARG(scan);
6337 else if (OP(scan) == EVAL) {
6339 data->flags |= SF_HAS_EVAL;
6341 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
6342 if (flags & SCF_DO_SUBSTR) {
6343 scan_commit(pRExC_state, data, minlenp, is_inf);
6344 flags &= ~SCF_DO_SUBSTR;
6346 if (data && OP(scan)==ACCEPT) {
6347 data->flags |= SCF_SEEN_ACCEPT;
6352 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
6354 if (flags & SCF_DO_SUBSTR) {
6355 scan_commit(pRExC_state, data, minlenp, is_inf);
6356 data->cur_is_floating = 1; /* float */
6358 is_inf = is_inf_internal = 1;
6359 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
6360 ssc_anything(data->start_class);
6361 flags &= ~SCF_DO_STCLASS;
6363 else if (OP(scan) == GPOS) {
6364 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
6365 !(delta || is_inf || (data && data->pos_delta)))
6367 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
6368 RExC_rx->intflags |= PREGf_ANCH_GPOS;
6369 if (RExC_rx->gofs < (STRLEN)min)
6370 RExC_rx->gofs = min;
6372 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
6376 #ifdef TRIE_STUDY_OPT
6377 #ifdef FULL_TRIE_STUDY
6378 else if (PL_regkind[OP(scan)] == TRIE) {
6379 /* NOTE - There is similar code to this block above for handling
6380 BRANCH nodes on the initial study. If you change stuff here
6382 regnode *trie_node= scan;
6383 regnode *tail= regnext(scan);
6384 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
6385 SSize_t max1 = 0, min1 = OPTIMIZE_INFTY;
6388 if (flags & SCF_DO_SUBSTR) { /* XXXX Add !SUSPEND? */
6389 /* Cannot merge strings after this. */
6390 scan_commit(pRExC_state, data, minlenp, is_inf);
6392 if (flags & SCF_DO_STCLASS)
6393 ssc_init_zero(pRExC_state, &accum);
6399 const regnode *nextbranch= NULL;
6402 for ( word=1 ; word <= trie->wordcount ; word++)
6404 SSize_t deltanext=0, minnext=0, f = 0, fake;
6405 regnode_ssc this_class;
6407 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
6409 data_fake.whilem_c = data->whilem_c;
6410 data_fake.last_closep = data->last_closep;
6413 data_fake.last_closep = &fake;
6414 data_fake.pos_delta = delta;
6415 if (flags & SCF_DO_STCLASS) {
6416 ssc_init(pRExC_state, &this_class);
6417 data_fake.start_class = &this_class;
6418 f = SCF_DO_STCLASS_AND;
6420 if (flags & SCF_WHILEM_VISITED_POS)
6421 f |= SCF_WHILEM_VISITED_POS;
6423 if (trie->jump[word]) {
6425 nextbranch = trie_node + trie->jump[0];
6426 scan= trie_node + trie->jump[word];
6427 /* We go from the jump point to the branch that follows
6428 it. Note this means we need the vestigal unused
6429 branches even though they arent otherwise used. */
6430 /* optimise study_chunk() for TRIE */
6431 minnext = study_chunk(pRExC_state, &scan, minlenp,
6432 &deltanext, (regnode *)nextbranch, &data_fake,
6433 stopparen, recursed_depth, NULL, f, depth+1,
6436 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
6437 nextbranch= regnext((regnode*)nextbranch);
6439 if (min1 > (SSize_t)(minnext + trie->minlen))
6440 min1 = minnext + trie->minlen;
6441 if (deltanext == OPTIMIZE_INFTY) {
6442 is_inf = is_inf_internal = 1;
6443 max1 = OPTIMIZE_INFTY;
6444 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
6445 max1 = minnext + deltanext + trie->maxlen;
6447 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
6449 if (data_fake.flags & SCF_SEEN_ACCEPT) {
6450 if ( stopmin > min + min1)
6451 stopmin = min + min1;
6452 flags &= ~SCF_DO_SUBSTR;
6454 data->flags |= SCF_SEEN_ACCEPT;
6457 if (data_fake.flags & SF_HAS_EVAL)
6458 data->flags |= SF_HAS_EVAL;
6459 data->whilem_c = data_fake.whilem_c;
6461 if (flags & SCF_DO_STCLASS)
6462 ssc_or(pRExC_state, &accum, (regnode_charclass *) &this_class);
6465 if (flags & SCF_DO_SUBSTR) {
6466 data->pos_min += min1;
6467 data->pos_delta += max1 - min1;
6468 if (max1 != min1 || is_inf)
6469 data->cur_is_floating = 1; /* float */
6472 if (delta != OPTIMIZE_INFTY) {
6473 if (OPTIMIZE_INFTY - (max1 - min1) >= delta)
6474 delta += max1 - min1;
6476 delta = OPTIMIZE_INFTY;
6478 if (flags & SCF_DO_STCLASS_OR) {
6479 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &accum);
6481 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6482 flags &= ~SCF_DO_STCLASS;
6485 else if (flags & SCF_DO_STCLASS_AND) {
6487 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
6488 flags &= ~SCF_DO_STCLASS;
6491 /* Switch to OR mode: cache the old value of
6492 * data->start_class */
6494 StructCopy(data->start_class, and_withp, regnode_ssc);
6495 flags &= ~SCF_DO_STCLASS_AND;
6496 StructCopy(&accum, data->start_class, regnode_ssc);
6497 flags |= SCF_DO_STCLASS_OR;
6504 else if (PL_regkind[OP(scan)] == TRIE) {
6505 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
6508 min += trie->minlen;
6509 delta += (trie->maxlen - trie->minlen);
6510 flags &= ~SCF_DO_STCLASS; /* xxx */
6511 if (flags & SCF_DO_SUBSTR) {
6512 /* Cannot expect anything... */
6513 scan_commit(pRExC_state, data, minlenp, is_inf);
6514 data->pos_min += trie->minlen;
6515 data->pos_delta += (trie->maxlen - trie->minlen);
6516 if (trie->maxlen != trie->minlen)
6517 data->cur_is_floating = 1; /* float */
6519 if (trie->jump) /* no more substrings -- for now /grr*/
6520 flags &= ~SCF_DO_SUBSTR;
6522 else if (OP(scan) == REGEX_SET) {
6523 Perl_croak(aTHX_ "panic: %s regnode should be resolved"
6524 " before optimization", reg_name[REGEX_SET]);
6527 #endif /* old or new */
6528 #endif /* TRIE_STUDY_OPT */
6530 /* Else: zero-length, ignore. */
6531 scan = regnext(scan);
6536 /* we need to unwind recursion. */
6539 DEBUG_STUDYDATA("frame-end", data, depth, is_inf);
6540 DEBUG_PEEP("fend", scan, depth, flags);
6542 /* restore previous context */
6543 last = frame->last_regnode;
6544 scan = frame->next_regnode;
6545 stopparen = frame->stopparen;
6546 recursed_depth = frame->prev_recursed_depth;
6548 RExC_frame_last = frame->prev_frame;
6549 frame = frame->this_prev_frame;
6550 goto fake_study_recurse;
6554 DEBUG_STUDYDATA("pre-fin", data, depth, is_inf);
6557 *deltap = is_inf_internal ? OPTIMIZE_INFTY : delta;
6559 if (flags & SCF_DO_SUBSTR && is_inf)
6560 data->pos_delta = OPTIMIZE_INFTY - data->pos_min;
6561 if (is_par > (I32)U8_MAX)
6563 if (is_par && pars==1 && data) {
6564 data->flags |= SF_IN_PAR;
6565 data->flags &= ~SF_HAS_PAR;
6567 else if (pars && data) {
6568 data->flags |= SF_HAS_PAR;
6569 data->flags &= ~SF_IN_PAR;
6571 if (flags & SCF_DO_STCLASS_OR)
6572 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6573 if (flags & SCF_TRIE_RESTUDY)
6574 data->flags |= SCF_TRIE_RESTUDY;
6576 DEBUG_STUDYDATA("post-fin", data, depth, is_inf);
6578 final_minlen = min < stopmin
6581 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN)) {
6582 if (final_minlen > OPTIMIZE_INFTY - delta)
6583 RExC_maxlen = OPTIMIZE_INFTY;
6584 else if (RExC_maxlen < final_minlen + delta)
6585 RExC_maxlen = final_minlen + delta;
6587 return final_minlen;
6591 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
6593 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
6595 PERL_ARGS_ASSERT_ADD_DATA;
6597 Renewc(RExC_rxi->data,
6598 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
6599 char, struct reg_data);
6601 Renew(RExC_rxi->data->what, count + n, U8);
6603 Newx(RExC_rxi->data->what, n, U8);
6604 RExC_rxi->data->count = count + n;
6605 Copy(s, RExC_rxi->data->what + count, n, U8);
6609 /*XXX: todo make this not included in a non debugging perl, but appears to be
6610 * used anyway there, in 'use re' */
6611 #ifndef PERL_IN_XSUB_RE
6613 Perl_reginitcolors(pTHX)
6615 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
6617 char *t = savepv(s);
6621 t = strchr(t, '\t');
6627 PL_colors[i] = t = (char *)"";
6632 PL_colors[i++] = (char *)"";
6639 #ifdef TRIE_STUDY_OPT
6640 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
6643 (data.flags & SCF_TRIE_RESTUDY) \
6651 #define CHECK_RESTUDY_GOTO_butfirst
6655 * pregcomp - compile a regular expression into internal code
6657 * Decides which engine's compiler to call based on the hint currently in
6661 #ifndef PERL_IN_XSUB_RE
6663 /* return the currently in-scope regex engine (or the default if none) */
6665 regexp_engine const *
6666 Perl_current_re_engine(pTHX)
6668 if (IN_PERL_COMPILETIME) {
6669 HV * const table = GvHV(PL_hintgv);
6672 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
6673 return &PL_core_reg_engine;
6674 ptr = hv_fetchs(table, "regcomp", FALSE);
6675 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
6676 return &PL_core_reg_engine;
6677 return INT2PTR(regexp_engine*, SvIV(*ptr));
6681 if (!PL_curcop->cop_hints_hash)
6682 return &PL_core_reg_engine;
6683 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
6684 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
6685 return &PL_core_reg_engine;
6686 return INT2PTR(regexp_engine*, SvIV(ptr));
6692 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
6694 regexp_engine const *eng = current_re_engine();
6695 DECLARE_AND_GET_RE_DEBUG_FLAGS;
6697 PERL_ARGS_ASSERT_PREGCOMP;
6699 /* Dispatch a request to compile a regexp to correct regexp engine. */
6701 Perl_re_printf( aTHX_ "Using engine %" UVxf "\n",
6704 return CALLREGCOMP_ENG(eng, pattern, flags);
6708 /* public(ish) entry point for the perl core's own regex compiling code.
6709 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
6710 * pattern rather than a list of OPs, and uses the internal engine rather
6711 * than the current one */
6714 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
6716 SV *pat = pattern; /* defeat constness! */
6718 PERL_ARGS_ASSERT_RE_COMPILE;
6720 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
6721 #ifdef PERL_IN_XSUB_RE
6724 &PL_core_reg_engine,
6726 NULL, NULL, rx_flags, 0);
6730 S_free_codeblocks(pTHX_ struct reg_code_blocks *cbs)
6734 if (--cbs->refcnt > 0)
6736 for (n = 0; n < cbs->count; n++) {
6737 REGEXP *rx = cbs->cb[n].src_regex;
6739 cbs->cb[n].src_regex = NULL;
6740 SvREFCNT_dec_NN(rx);
6748 static struct reg_code_blocks *
6749 S_alloc_code_blocks(pTHX_ int ncode)
6751 struct reg_code_blocks *cbs;
6752 Newx(cbs, 1, struct reg_code_blocks);
6755 SAVEDESTRUCTOR_X(S_free_codeblocks, cbs);
6757 Newx(cbs->cb, ncode, struct reg_code_block);
6764 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
6765 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
6766 * point to the realloced string and length.
6768 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
6772 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
6773 char **pat_p, STRLEN *plen_p, int num_code_blocks)
6775 U8 *const src = (U8*)*pat_p;
6780 DECLARE_AND_GET_RE_DEBUG_FLAGS;
6782 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6783 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
6785 /* 1 for each byte + 1 for each byte that expands to two, + trailing NUL */
6786 Newx(dst, *plen_p + variant_under_utf8_count(src, src + *plen_p) + 1, U8);
6789 while (s < *plen_p) {
6790 append_utf8_from_native_byte(src[s], &d);
6792 if (n < num_code_blocks) {
6793 assert(pRExC_state->code_blocks);
6794 if (!do_end && pRExC_state->code_blocks->cb[n].start == s) {
6795 pRExC_state->code_blocks->cb[n].start = d - dst - 1;
6796 assert(*(d - 1) == '(');
6799 else if (do_end && pRExC_state->code_blocks->cb[n].end == s) {
6800 pRExC_state->code_blocks->cb[n].end = d - dst - 1;
6801 assert(*(d - 1) == ')');
6810 *pat_p = (char*) dst;
6812 RExC_orig_utf8 = RExC_utf8 = 1;
6817 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
6818 * while recording any code block indices, and handling overloading,
6819 * nested qr// objects etc. If pat is null, it will allocate a new
6820 * string, or just return the first arg, if there's only one.
6822 * Returns the malloced/updated pat.
6823 * patternp and pat_count is the array of SVs to be concatted;
6824 * oplist is the optional list of ops that generated the SVs;
6825 * recompile_p is a pointer to a boolean that will be set if
6826 * the regex will need to be recompiled.
6827 * delim, if non-null is an SV that will be inserted between each element
6831 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
6832 SV *pat, SV ** const patternp, int pat_count,
6833 OP *oplist, bool *recompile_p, SV *delim)
6837 bool use_delim = FALSE;
6838 bool alloced = FALSE;
6840 /* if we know we have at least two args, create an empty string,
6841 * then concatenate args to that. For no args, return an empty string */
6842 if (!pat && pat_count != 1) {
6848 for (svp = patternp; svp < patternp + pat_count; svp++) {
6851 STRLEN orig_patlen = 0;
6853 SV *msv = use_delim ? delim : *svp;
6854 if (!msv) msv = &PL_sv_undef;
6856 /* if we've got a delimiter, we go round the loop twice for each
6857 * svp slot (except the last), using the delimiter the second
6866 if (SvTYPE(msv) == SVt_PVAV) {
6867 /* we've encountered an interpolated array within
6868 * the pattern, e.g. /...@a..../. Expand the list of elements,
6869 * then recursively append elements.
6870 * The code in this block is based on S_pushav() */
6872 AV *const av = (AV*)msv;
6873 const SSize_t maxarg = AvFILL(av) + 1;
6877 assert(oplist->op_type == OP_PADAV
6878 || oplist->op_type == OP_RV2AV);
6879 oplist = OpSIBLING(oplist);
6882 if (SvRMAGICAL(av)) {
6885 Newx(array, maxarg, SV*);
6887 for (i=0; i < maxarg; i++) {
6888 SV ** const svp = av_fetch(av, i, FALSE);
6889 array[i] = svp ? *svp : &PL_sv_undef;
6893 array = AvARRAY(av);
6895 pat = S_concat_pat(aTHX_ pRExC_state, pat,
6896 array, maxarg, NULL, recompile_p,
6898 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
6904 /* we make the assumption here that each op in the list of
6905 * op_siblings maps to one SV pushed onto the stack,
6906 * except for code blocks, with have both an OP_NULL and
6908 * This allows us to match up the list of SVs against the
6909 * list of OPs to find the next code block.
6911 * Note that PUSHMARK PADSV PADSV ..
6913 * PADRANGE PADSV PADSV ..
6914 * so the alignment still works. */
6917 if (oplist->op_type == OP_NULL
6918 && (oplist->op_flags & OPf_SPECIAL))
6920 assert(n < pRExC_state->code_blocks->count);
6921 pRExC_state->code_blocks->cb[n].start = pat ? SvCUR(pat) : 0;
6922 pRExC_state->code_blocks->cb[n].block = oplist;
6923 pRExC_state->code_blocks->cb[n].src_regex = NULL;
6926 oplist = OpSIBLING(oplist); /* skip CONST */
6929 oplist = OpSIBLING(oplist);;
6932 /* apply magic and QR overloading to arg */
6935 if (SvROK(msv) && SvAMAGIC(msv)) {
6936 SV *sv = AMG_CALLunary(msv, regexp_amg);
6940 if (SvTYPE(sv) != SVt_REGEXP)
6941 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
6946 /* try concatenation overload ... */
6947 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
6948 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
6951 /* overloading involved: all bets are off over literal
6952 * code. Pretend we haven't seen it */
6954 pRExC_state->code_blocks->count -= n;
6958 /* ... or failing that, try "" overload */
6959 while (SvAMAGIC(msv)
6960 && (sv = AMG_CALLunary(msv, string_amg))
6964 && SvRV(msv) == SvRV(sv))
6969 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
6973 /* this is a partially unrolled
6974 * sv_catsv_nomg(pat, msv);
6975 * that allows us to adjust code block indices if
6978 char *dst = SvPV_force_nomg(pat, dlen);
6980 if (SvUTF8(msv) && !SvUTF8(pat)) {
6981 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
6982 sv_setpvn(pat, dst, dlen);
6985 sv_catsv_nomg(pat, msv);
6989 /* We have only one SV to process, but we need to verify
6990 * it is properly null terminated or we will fail asserts
6991 * later. In theory we probably shouldn't get such SV's,
6992 * but if we do we should handle it gracefully. */
6993 if ( SvTYPE(msv) != SVt_PV || (SvLEN(msv) > SvCUR(msv) && *(SvEND(msv)) == 0) || SvIsCOW_shared_hash(msv) ) {
6994 /* not a string, or a string with a trailing null */
6997 /* a string with no trailing null, we need to copy it
6998 * so it has a trailing null */
6999 pat = sv_2mortal(newSVsv(msv));
7004 pRExC_state->code_blocks->cb[n-1].end = SvCUR(pat)-1;
7007 /* extract any code blocks within any embedded qr//'s */
7008 if (rx && SvTYPE(rx) == SVt_REGEXP
7009 && RX_ENGINE((REGEXP*)rx)->op_comp)
7012 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
7013 if (ri->code_blocks && ri->code_blocks->count) {
7015 /* the presence of an embedded qr// with code means
7016 * we should always recompile: the text of the
7017 * qr// may not have changed, but it may be a
7018 * different closure than last time */
7020 if (pRExC_state->code_blocks) {
7021 int new_count = pRExC_state->code_blocks->count
7022 + ri->code_blocks->count;
7023 Renew(pRExC_state->code_blocks->cb,
7024 new_count, struct reg_code_block);
7025 pRExC_state->code_blocks->count = new_count;
7028 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_
7029 ri->code_blocks->count);
7031 for (i=0; i < ri->code_blocks->count; i++) {
7032 struct reg_code_block *src, *dst;
7033 STRLEN offset = orig_patlen
7034 + ReANY((REGEXP *)rx)->pre_prefix;
7035 assert(n < pRExC_state->code_blocks->count);
7036 src = &ri->code_blocks->cb[i];
7037 dst = &pRExC_state->code_blocks->cb[n];
7038 dst->start = src->start + offset;
7039 dst->end = src->end + offset;
7040 dst->block = src->block;
7041 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
7050 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
7059 /* see if there are any run-time code blocks in the pattern.
7060 * False positives are allowed */
7063 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
7064 char *pat, STRLEN plen)
7069 PERL_UNUSED_CONTEXT;
7071 for (s = 0; s < plen; s++) {
7072 if ( pRExC_state->code_blocks
7073 && n < pRExC_state->code_blocks->count
7074 && s == pRExC_state->code_blocks->cb[n].start)
7076 s = pRExC_state->code_blocks->cb[n].end;
7080 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
7082 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
7084 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
7091 /* Handle run-time code blocks. We will already have compiled any direct
7092 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
7093 * copy of it, but with any literal code blocks blanked out and
7094 * appropriate chars escaped; then feed it into
7096 * eval "qr'modified_pattern'"
7100 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
7104 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
7106 * After eval_sv()-ing that, grab any new code blocks from the returned qr
7107 * and merge them with any code blocks of the original regexp.
7109 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
7110 * instead, just save the qr and return FALSE; this tells our caller that
7111 * the original pattern needs upgrading to utf8.
7115 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
7116 char *pat, STRLEN plen)
7120 DECLARE_AND_GET_RE_DEBUG_FLAGS;
7122 if (pRExC_state->runtime_code_qr) {
7123 /* this is the second time we've been called; this should
7124 * only happen if the main pattern got upgraded to utf8
7125 * during compilation; re-use the qr we compiled first time
7126 * round (which should be utf8 too)
7128 qr = pRExC_state->runtime_code_qr;
7129 pRExC_state->runtime_code_qr = NULL;
7130 assert(RExC_utf8 && SvUTF8(qr));
7136 int newlen = plen + 7; /* allow for "qr''xx\0" extra chars */
7140 /* determine how many extra chars we need for ' and \ escaping */
7141 for (s = 0; s < plen; s++) {
7142 if (pat[s] == '\'' || pat[s] == '\\')
7146 Newx(newpat, newlen, char);
7148 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
7150 for (s = 0; s < plen; s++) {
7151 if ( pRExC_state->code_blocks
7152 && n < pRExC_state->code_blocks->count
7153 && s == pRExC_state->code_blocks->cb[n].start)
7155 /* blank out literal code block so that they aren't
7156 * recompiled: eg change from/to:
7166 assert(pat[s] == '(');
7167 assert(pat[s+1] == '?');
7171 while (s < pRExC_state->code_blocks->cb[n].end) {
7179 if (pat[s] == '\'' || pat[s] == '\\')
7184 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED) {
7186 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED_MORE) {
7192 Perl_re_printf( aTHX_
7193 "%sre-parsing pattern for runtime code:%s %s\n",
7194 PL_colors[4], PL_colors[5], newpat);
7197 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
7203 PUSHSTACKi(PERLSI_REQUIRE);
7204 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
7205 * parsing qr''; normally only q'' does this. It also alters
7207 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
7208 SvREFCNT_dec_NN(sv);
7213 SV * const errsv = ERRSV;
7214 if (SvTRUE_NN(errsv))
7215 /* use croak_sv ? */
7216 Perl_croak_nocontext("%" SVf, SVfARG(errsv));
7218 assert(SvROK(qr_ref));
7220 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
7221 /* the leaving below frees the tmp qr_ref.
7222 * Give qr a life of its own */
7230 if (!RExC_utf8 && SvUTF8(qr)) {
7231 /* first time through; the pattern got upgraded; save the
7232 * qr for the next time through */
7233 assert(!pRExC_state->runtime_code_qr);
7234 pRExC_state->runtime_code_qr = qr;
7239 /* extract any code blocks within the returned qr// */
7242 /* merge the main (r1) and run-time (r2) code blocks into one */
7244 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
7245 struct reg_code_block *new_block, *dst;
7246 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
7250 if (!r2->code_blocks || !r2->code_blocks->count) /* we guessed wrong */
7252 SvREFCNT_dec_NN(qr);
7256 if (!r1->code_blocks)
7257 r1->code_blocks = S_alloc_code_blocks(aTHX_ 0);
7259 r1c = r1->code_blocks->count;
7260 r2c = r2->code_blocks->count;
7262 Newx(new_block, r1c + r2c, struct reg_code_block);
7266 while (i1 < r1c || i2 < r2c) {
7267 struct reg_code_block *src;
7271 src = &r2->code_blocks->cb[i2++];
7275 src = &r1->code_blocks->cb[i1++];
7276 else if ( r1->code_blocks->cb[i1].start
7277 < r2->code_blocks->cb[i2].start)
7279 src = &r1->code_blocks->cb[i1++];
7280 assert(src->end < r2->code_blocks->cb[i2].start);
7283 assert( r1->code_blocks->cb[i1].start
7284 > r2->code_blocks->cb[i2].start);
7285 src = &r2->code_blocks->cb[i2++];
7287 assert(src->end < r1->code_blocks->cb[i1].start);
7290 assert(pat[src->start] == '(');
7291 assert(pat[src->end] == ')');
7292 dst->start = src->start;
7293 dst->end = src->end;
7294 dst->block = src->block;
7295 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
7299 r1->code_blocks->count += r2c;
7300 Safefree(r1->code_blocks->cb);
7301 r1->code_blocks->cb = new_block;
7304 SvREFCNT_dec_NN(qr);
7310 S_setup_longest(pTHX_ RExC_state_t *pRExC_state,
7311 struct reg_substr_datum *rsd,
7312 struct scan_data_substrs *sub,
7313 STRLEN longest_length)
7315 /* This is the common code for setting up the floating and fixed length
7316 * string data extracted from Perl_re_op_compile() below. Returns a boolean
7317 * as to whether succeeded or not */
7321 bool eol = cBOOL(sub->flags & SF_BEFORE_EOL);
7322 bool meol = cBOOL(sub->flags & SF_BEFORE_MEOL);
7324 if (! (longest_length
7325 || (eol /* Can't have SEOL and MULTI */
7326 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
7328 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
7329 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
7334 /* copy the information about the longest from the reg_scan_data
7335 over to the program. */
7336 if (SvUTF8(sub->str)) {
7338 rsd->utf8_substr = sub->str;
7340 rsd->substr = sub->str;
7341 rsd->utf8_substr = NULL;
7343 /* end_shift is how many chars that must be matched that
7344 follow this item. We calculate it ahead of time as once the
7345 lookbehind offset is added in we lose the ability to correctly
7347 ml = sub->minlenp ? *(sub->minlenp) : (SSize_t)longest_length;
7348 rsd->end_shift = ml - sub->min_offset
7350 /* XXX SvTAIL is always false here - did you mean FBMcf_TAIL
7352 + (SvTAIL(sub->str) != 0)
7356 t = (eol/* Can't have SEOL and MULTI */
7357 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
7358 fbm_compile(sub->str, t ? FBMcf_TAIL : 0);
7364 S_set_regex_pv(pTHX_ RExC_state_t *pRExC_state, REGEXP *Rx)
7366 /* Calculates and sets in the compiled pattern 'Rx' the string to compile,
7367 * properly wrapped with the right modifiers */
7369 bool has_p = ((RExC_rx->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
7370 bool has_charset = RExC_utf8 || (get_regex_charset(RExC_rx->extflags)
7371 != REGEX_DEPENDS_CHARSET);
7373 /* The caret is output if there are any defaults: if not all the STD
7374 * flags are set, or if no character set specifier is needed */
7376 (((RExC_rx->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
7378 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
7379 == REG_RUN_ON_COMMENT_SEEN);
7380 U8 reganch = (U8)((RExC_rx->extflags & RXf_PMf_STD_PMMOD)
7381 >> RXf_PMf_STD_PMMOD_SHIFT);
7382 const char *fptr = STD_PAT_MODS; /*"msixxn"*/
7384 STRLEN pat_len = RExC_precomp_end - RExC_precomp;
7386 /* We output all the necessary flags; we never output a minus, as all
7387 * those are defaults, so are
7388 * covered by the caret */
7389 const STRLEN wraplen = pat_len + has_p + has_runon
7390 + has_default /* If needs a caret */
7391 + PL_bitcount[reganch] /* 1 char for each set standard flag */
7393 /* If needs a character set specifier */
7394 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
7395 + (sizeof("(?:)") - 1);
7397 PERL_ARGS_ASSERT_SET_REGEX_PV;
7399 /* make sure PL_bitcount bounds not exceeded */
7400 STATIC_ASSERT_STMT(sizeof(STD_PAT_MODS) <= 8);
7402 p = sv_grow(MUTABLE_SV(Rx), wraplen + 1); /* +1 for the ending NUL */
7405 SvFLAGS(Rx) |= SVf_UTF8;
7408 /* If a default, cover it using the caret */
7410 *p++= DEFAULT_PAT_MOD;
7416 name = get_regex_charset_name(RExC_rx->extflags, &len);
7417 if (strEQ(name, DEPENDS_PAT_MODS)) { /* /d under UTF-8 => /u */
7419 name = UNICODE_PAT_MODS;
7420 len = sizeof(UNICODE_PAT_MODS) - 1;
7422 Copy(name, p, len, char);
7426 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
7429 while((ch = *fptr++)) {
7437 Copy(RExC_precomp, p, pat_len, char);
7438 assert ((RX_WRAPPED(Rx) - p) < 16);
7439 RExC_rx->pre_prefix = p - RX_WRAPPED(Rx);
7442 /* Adding a trailing \n causes this to compile properly:
7443 my $R = qr / A B C # D E/x; /($R)/
7444 Otherwise the parens are considered part of the comment */
7449 SvCUR_set(Rx, p - RX_WRAPPED(Rx));
7453 * Perl_re_op_compile - the perl internal RE engine's function to compile a
7454 * regular expression into internal code.
7455 * The pattern may be passed either as:
7456 * a list of SVs (patternp plus pat_count)
7457 * a list of OPs (expr)
7458 * If both are passed, the SV list is used, but the OP list indicates
7459 * which SVs are actually pre-compiled code blocks
7461 * The SVs in the list have magic and qr overloading applied to them (and
7462 * the list may be modified in-place with replacement SVs in the latter
7465 * If the pattern hasn't changed from old_re, then old_re will be
7468 * eng is the current engine. If that engine has an op_comp method, then
7469 * handle directly (i.e. we assume that op_comp was us); otherwise, just
7470 * do the initial concatenation of arguments and pass on to the external
7473 * If is_bare_re is not null, set it to a boolean indicating whether the
7474 * arg list reduced (after overloading) to a single bare regex which has
7475 * been returned (i.e. /$qr/).
7477 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
7479 * pm_flags contains the PMf_* flags, typically based on those from the
7480 * pm_flags field of the related PMOP. Currently we're only interested in
7481 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL, PMf_WILDCARD.
7483 * For many years this code had an initial sizing pass that calculated
7484 * (sometimes incorrectly, leading to security holes) the size needed for the
7485 * compiled pattern. That was changed by commit
7486 * 7c932d07cab18751bfc7515b4320436273a459e2 in 5.29, which reallocs the size, a
7487 * node at a time, as parsing goes along. Patches welcome to fix any obsolete
7488 * references to this sizing pass.
7490 * Now, an initial crude guess as to the size needed is made, based on the
7491 * length of the pattern. Patches welcome to improve that guess. That amount
7492 * of space is malloc'd and then immediately freed, and then clawed back node
7493 * by node. This design is to minimze, to the extent possible, memory churn
7494 * when doing the reallocs.
7496 * A separate parentheses counting pass may be needed in some cases.
7497 * (Previously the sizing pass did this.) Patches welcome to reduce the number
7500 * The existence of a sizing pass necessitated design decisions that are no
7501 * longer needed. There are potential areas of simplification.
7503 * Beware that the optimization-preparation code in here knows about some
7504 * of the structure of the compiled regexp. [I'll say.]
7508 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
7509 OP *expr, const regexp_engine* eng, REGEXP *old_re,
7510 bool *is_bare_re, const U32 orig_rx_flags, const U32 pm_flags)
7512 REGEXP *Rx; /* Capital 'R' means points to a REGEXP */
7520 SV** new_patternp = patternp;
7522 /* these are all flags - maybe they should be turned
7523 * into a single int with different bit masks */
7524 I32 sawlookahead = 0;
7529 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
7531 bool runtime_code = 0;
7533 RExC_state_t RExC_state;
7534 RExC_state_t * const pRExC_state = &RExC_state;
7535 #ifdef TRIE_STUDY_OPT
7537 RExC_state_t copyRExC_state;
7539 DECLARE_AND_GET_RE_DEBUG_FLAGS;
7541 PERL_ARGS_ASSERT_RE_OP_COMPILE;
7543 DEBUG_r(if (!PL_colorset) reginitcolors());
7546 pRExC_state->warn_text = NULL;
7547 pRExC_state->unlexed_names = NULL;
7548 pRExC_state->code_blocks = NULL;
7551 *is_bare_re = FALSE;
7553 if (expr && (expr->op_type == OP_LIST ||
7554 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
7555 /* allocate code_blocks if needed */
7559 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o))
7560 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
7561 ncode++; /* count of DO blocks */
7564 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_ ncode);
7568 /* compile-time pattern with just OP_CONSTs and DO blocks */
7573 /* find how many CONSTs there are */
7576 if (expr->op_type == OP_CONST)
7579 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
7580 if (o->op_type == OP_CONST)
7584 /* fake up an SV array */
7586 assert(!new_patternp);
7587 Newx(new_patternp, n, SV*);
7588 SAVEFREEPV(new_patternp);
7592 if (expr->op_type == OP_CONST)
7593 new_patternp[n] = cSVOPx_sv(expr);
7595 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
7596 if (o->op_type == OP_CONST)
7597 new_patternp[n++] = cSVOPo_sv;
7602 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7603 "Assembling pattern from %d elements%s\n", pat_count,
7604 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
7606 /* set expr to the first arg op */
7608 if (pRExC_state->code_blocks && pRExC_state->code_blocks->count
7609 && expr->op_type != OP_CONST)
7611 expr = cLISTOPx(expr)->op_first;
7612 assert( expr->op_type == OP_PUSHMARK
7613 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
7614 || expr->op_type == OP_PADRANGE);
7615 expr = OpSIBLING(expr);
7618 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
7619 expr, &recompile, NULL);
7621 /* handle bare (possibly after overloading) regex: foo =~ $re */
7626 if (SvTYPE(re) == SVt_REGEXP) {
7630 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7631 "Precompiled pattern%s\n",
7632 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
7638 exp = SvPV_nomg(pat, plen);
7640 if (!eng->op_comp) {
7641 if ((SvUTF8(pat) && IN_BYTES)
7642 || SvGMAGICAL(pat) || SvAMAGIC(pat))
7644 /* make a temporary copy; either to convert to bytes,
7645 * or to avoid repeating get-magic / overloaded stringify */
7646 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
7647 (IN_BYTES ? 0 : SvUTF8(pat)));
7649 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
7652 /* ignore the utf8ness if the pattern is 0 length */
7653 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
7654 RExC_uni_semantics = 0;
7655 RExC_contains_locale = 0;
7656 RExC_strict = cBOOL(pm_flags & RXf_PMf_STRICT);
7657 RExC_in_script_run = 0;
7658 RExC_study_started = 0;
7659 pRExC_state->runtime_code_qr = NULL;
7660 RExC_frame_head= NULL;
7661 RExC_frame_last= NULL;
7662 RExC_frame_count= 0;
7663 RExC_latest_warn_offset = 0;
7664 RExC_use_BRANCHJ = 0;
7665 RExC_warned_WARN_EXPERIMENTAL__VLB = 0;
7666 RExC_warned_WARN_EXPERIMENTAL__REGEX_SETS = 0;
7667 RExC_total_parens = 0;
7668 RExC_open_parens = NULL;
7669 RExC_close_parens = NULL;
7670 RExC_paren_names = NULL;
7672 RExC_seen_d_op = FALSE;
7674 RExC_paren_name_list = NULL;
7678 RExC_mysv1= sv_newmortal();
7679 RExC_mysv2= sv_newmortal();
7683 SV *dsv= sv_newmortal();
7684 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, PL_dump_re_max_len);
7685 Perl_re_printf( aTHX_ "%sCompiling REx%s %s\n",
7686 PL_colors[4], PL_colors[5], s);
7689 /* we jump here if we have to recompile, e.g., from upgrading the pattern
7692 if ((pm_flags & PMf_USE_RE_EVAL)
7693 /* this second condition covers the non-regex literal case,
7694 * i.e. $foo =~ '(?{})'. */
7695 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
7697 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
7700 /* return old regex if pattern hasn't changed */
7701 /* XXX: note in the below we have to check the flags as well as the
7704 * Things get a touch tricky as we have to compare the utf8 flag
7705 * independently from the compile flags. */
7709 && !!RX_UTF8(old_re) == !!RExC_utf8
7710 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
7711 && RX_PRECOMP(old_re)
7712 && RX_PRELEN(old_re) == plen
7713 && memEQ(RX_PRECOMP(old_re), exp, plen)
7714 && !runtime_code /* with runtime code, always recompile */ )
7717 SV *dsv= sv_newmortal();
7718 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, PL_dump_re_max_len);
7719 Perl_re_printf( aTHX_ "%sSkipping recompilation of unchanged REx%s %s\n",
7720 PL_colors[4], PL_colors[5], s);
7725 /* Allocate the pattern's SV */
7726 RExC_rx_sv = Rx = (REGEXP*) newSV_type(SVt_REGEXP);
7727 RExC_rx = ReANY(Rx);
7728 if ( RExC_rx == NULL )
7729 FAIL("Regexp out of space");
7731 rx_flags = orig_rx_flags;
7733 if ( toUSE_UNI_CHARSET_NOT_DEPENDS
7734 && initial_charset == REGEX_DEPENDS_CHARSET)
7737 /* Set to use unicode semantics if the pattern is in utf8 and has the
7738 * 'depends' charset specified, as it means unicode when utf8 */
7739 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
7740 RExC_uni_semantics = 1;
7743 RExC_pm_flags = pm_flags;
7746 assert(TAINTING_get || !TAINT_get);
7748 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
7750 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
7751 /* whoops, we have a non-utf8 pattern, whilst run-time code
7752 * got compiled as utf8. Try again with a utf8 pattern */
7753 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7754 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7758 assert(!pRExC_state->runtime_code_qr);
7764 RExC_in_lookaround = 0;
7765 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
7766 RExC_recode_x_to_native = 0;
7767 RExC_in_multi_char_class = 0;
7769 RExC_start = RExC_copy_start_in_constructed = RExC_copy_start_in_input = RExC_precomp = exp;
7770 RExC_precomp_end = RExC_end = exp + plen;
7772 RExC_whilem_seen = 0;
7774 RExC_recurse = NULL;
7775 RExC_study_chunk_recursed = NULL;
7776 RExC_study_chunk_recursed_bytes= 0;
7777 RExC_recurse_count = 0;
7778 RExC_sets_depth = 0;
7779 pRExC_state->code_index = 0;
7781 /* Initialize the string in the compiled pattern. This is so that there is
7782 * something to output if necessary */
7783 set_regex_pv(pRExC_state, Rx);
7786 Perl_re_printf( aTHX_
7787 "Starting parse and generation\n");
7789 RExC_lastparse=NULL;
7792 /* Allocate space and zero-initialize. Note, the two step process
7793 of zeroing when in debug mode, thus anything assigned has to
7794 happen after that */
7797 /* On the first pass of the parse, we guess how big this will be. Then
7798 * we grow in one operation to that amount and then give it back. As
7799 * we go along, we re-allocate what we need.
7801 * XXX Currently the guess is essentially that the pattern will be an
7802 * EXACT node with one byte input, one byte output. This is crude, and
7803 * better heuristics are welcome.
7805 * On any subsequent passes, we guess what we actually computed in the
7806 * latest earlier pass. Such a pass probably didn't complete so is
7807 * missing stuff. We could improve those guesses by knowing where the
7808 * parse stopped, and use the length so far plus apply the above
7809 * assumption to what's left. */
7810 RExC_size = STR_SZ(RExC_end - RExC_start);
7813 Newxc(RExC_rxi, sizeof(regexp_internal) + RExC_size, char, regexp_internal);
7814 if ( RExC_rxi == NULL )
7815 FAIL("Regexp out of space");
7817 Zero(RExC_rxi, sizeof(regexp_internal) + RExC_size, char);
7818 RXi_SET( RExC_rx, RExC_rxi );
7820 /* We start from 0 (over from 0 in the case this is a reparse. The first
7821 * node parsed will give back any excess memory we have allocated so far).
7825 /* non-zero initialization begins here */
7826 RExC_rx->engine= eng;
7827 RExC_rx->extflags = rx_flags;
7828 RXp_COMPFLAGS(RExC_rx) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
7830 if (pm_flags & PMf_IS_QR) {
7831 RExC_rxi->code_blocks = pRExC_state->code_blocks;
7832 if (RExC_rxi->code_blocks) {
7833 RExC_rxi->code_blocks->refcnt++;
7837 RExC_rx->intflags = 0;
7839 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
7842 /* This NUL is guaranteed because the pattern comes from an SV*, and the sv
7843 * code makes sure the final byte is an uncounted NUL. But should this
7844 * ever not be the case, lots of things could read beyond the end of the
7845 * buffer: loops like
7846 * while(isFOO(*RExC_parse)) RExC_parse++;
7847 * strchr(RExC_parse, "foo");
7848 * etc. So it is worth noting. */
7849 assert(*RExC_end == '\0');
7853 RExC_parens_buf_size = 0;
7854 RExC_emit_start = RExC_rxi->program;
7855 pRExC_state->code_index = 0;
7857 *((char*) RExC_emit_start) = (char) REG_MAGIC;
7861 if (reg(pRExC_state, 0, &flags, 1)) {
7863 /* Success!, But we may need to redo the parse knowing how many parens
7864 * there actually are */
7865 if (IN_PARENS_PASS) {
7866 flags |= RESTART_PARSE;
7869 /* We have that number in RExC_npar */
7870 RExC_total_parens = RExC_npar;
7872 else if (! MUST_RESTART(flags)) {
7874 Perl_croak(aTHX_ "panic: reg returned failure to re_op_compile, flags=%#" UVxf, (UV) flags);
7877 /* Here, we either have success, or we have to redo the parse for some reason */
7878 if (MUST_RESTART(flags)) {
7880 /* It's possible to write a regexp in ascii that represents Unicode
7881 codepoints outside of the byte range, such as via \x{100}. If we
7882 detect such a sequence we have to convert the entire pattern to utf8
7883 and then recompile, as our sizing calculation will have been based
7884 on 1 byte == 1 character, but we will need to use utf8 to encode
7885 at least some part of the pattern, and therefore must convert the whole
7888 if (flags & NEED_UTF8) {
7890 /* We have stored the offset of the final warning output so far.
7891 * That must be adjusted. Any variant characters between the start
7892 * of the pattern and this warning count for 2 bytes in the final,
7893 * so just add them again */
7894 if (UNLIKELY(RExC_latest_warn_offset > 0)) {
7895 RExC_latest_warn_offset +=
7896 variant_under_utf8_count((U8 *) exp, (U8 *) exp
7897 + RExC_latest_warn_offset);
7899 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7900 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7901 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "Need to redo parse after upgrade\n"));
7904 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "Need to redo parse\n"));
7907 if (ALL_PARENS_COUNTED) {
7908 /* Make enough room for all the known parens, and zero it */
7909 Renew(RExC_open_parens, RExC_total_parens, regnode_offset);
7910 Zero(RExC_open_parens, RExC_total_parens, regnode_offset);
7911 RExC_open_parens[0] = 1; /* +1 for REG_MAGIC */
7913 Renew(RExC_close_parens, RExC_total_parens, regnode_offset);
7914 Zero(RExC_close_parens, RExC_total_parens, regnode_offset);
7916 else { /* Parse did not complete. Reinitialize the parentheses
7918 RExC_total_parens = 0;
7919 if (RExC_open_parens) {
7920 Safefree(RExC_open_parens);
7921 RExC_open_parens = NULL;
7923 if (RExC_close_parens) {
7924 Safefree(RExC_close_parens);
7925 RExC_close_parens = NULL;
7929 /* Clean up what we did in this parse */
7930 SvREFCNT_dec_NN(RExC_rx_sv);
7935 /* Here, we have successfully parsed and generated the pattern's program
7936 * for the regex engine. We are ready to finish things up and look for
7939 /* Update the string to compile, with correct modifiers, etc */
7940 set_regex_pv(pRExC_state, Rx);
7942 RExC_rx->nparens = RExC_total_parens - 1;
7944 /* Uses the upper 4 bits of the FLAGS field, so keep within that size */
7945 if (RExC_whilem_seen > 15)
7946 RExC_whilem_seen = 15;
7949 Perl_re_printf( aTHX_
7950 "Required size %" IVdf " nodes\n", (IV)RExC_size);
7952 RExC_lastparse=NULL;
7955 #ifdef RE_TRACK_PATTERN_OFFSETS
7956 DEBUG_OFFSETS_r(Perl_re_printf( aTHX_
7957 "%s %" UVuf " bytes for offset annotations.\n",
7958 RExC_offsets ? "Got" : "Couldn't get",
7959 (UV)((RExC_offsets[0] * 2 + 1))));
7960 DEBUG_OFFSETS_r(if (RExC_offsets) {
7961 const STRLEN len = RExC_offsets[0];
7963 DECLARE_AND_GET_RE_DEBUG_FLAGS;
7964 Perl_re_printf( aTHX_
7965 "Offsets: [%" UVuf "]\n\t", (UV)RExC_offsets[0]);
7966 for (i = 1; i <= len; i++) {
7967 if (RExC_offsets[i*2-1] || RExC_offsets[i*2])
7968 Perl_re_printf( aTHX_ "%" UVuf ":%" UVuf "[%" UVuf "] ",
7969 (UV)i, (UV)RExC_offsets[i*2-1], (UV)RExC_offsets[i*2]);
7971 Perl_re_printf( aTHX_ "\n");
7975 SetProgLen(RExC_rxi,RExC_size);
7978 DEBUG_DUMP_PRE_OPTIMIZE_r({
7979 SV * const sv = sv_newmortal();
7980 RXi_GET_DECL(RExC_rx, ri);
7982 Perl_re_printf( aTHX_ "Program before optimization:\n");
7984 (void)dumpuntil(RExC_rx, ri->program, ri->program + 1, NULL, NULL,
7989 Perl_re_printf( aTHX_ "Starting post parse optimization\n");
7992 /* XXXX To minimize changes to RE engine we always allocate
7993 3-units-long substrs field. */
7994 Newx(RExC_rx->substrs, 1, struct reg_substr_data);
7995 if (RExC_recurse_count) {
7996 Newx(RExC_recurse, RExC_recurse_count, regnode *);
7997 SAVEFREEPV(RExC_recurse);
8000 if (RExC_seen & REG_RECURSE_SEEN) {
8001 /* Note, RExC_total_parens is 1 + the number of parens in a pattern.
8002 * So its 1 if there are no parens. */
8003 RExC_study_chunk_recursed_bytes= (RExC_total_parens >> 3) +
8004 ((RExC_total_parens & 0x07) != 0);
8005 Newx(RExC_study_chunk_recursed,
8006 RExC_study_chunk_recursed_bytes * RExC_total_parens, U8);
8007 SAVEFREEPV(RExC_study_chunk_recursed);
8011 RExC_rx->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
8013 RExC_study_chunk_recursed_count= 0;
8015 Zero(RExC_rx->substrs, 1, struct reg_substr_data);
8016 if (RExC_study_chunk_recursed) {
8017 Zero(RExC_study_chunk_recursed,
8018 RExC_study_chunk_recursed_bytes * RExC_total_parens, U8);
8022 #ifdef TRIE_STUDY_OPT
8024 StructCopy(&zero_scan_data, &data, scan_data_t);
8025 copyRExC_state = RExC_state;
8028 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "Restudying\n"));
8030 RExC_state = copyRExC_state;
8031 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
8032 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
8034 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
8035 StructCopy(&zero_scan_data, &data, scan_data_t);
8038 StructCopy(&zero_scan_data, &data, scan_data_t);
8041 /* Dig out information for optimizations. */
8042 RExC_rx->extflags = RExC_flags; /* was pm_op */
8043 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
8046 SvUTF8_on(Rx); /* Unicode in it? */
8047 RExC_rxi->regstclass = NULL;
8048 if (RExC_naughty >= TOO_NAUGHTY) /* Probably an expensive pattern. */
8049 RExC_rx->intflags |= PREGf_NAUGHTY;
8050 scan = RExC_rxi->program + 1; /* First BRANCH. */
8052 /* testing for BRANCH here tells us whether there is "must appear"
8053 data in the pattern. If there is then we can use it for optimisations */
8054 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
8057 STRLEN longest_length[2];
8058 regnode_ssc ch_class; /* pointed to by data */
8060 SSize_t last_close = 0; /* pointed to by data */
8061 regnode *first= scan;
8062 regnode *first_next= regnext(first);
8066 * Skip introductions and multiplicators >= 1
8067 * so that we can extract the 'meat' of the pattern that must
8068 * match in the large if() sequence following.
8069 * NOTE that EXACT is NOT covered here, as it is normally
8070 * picked up by the optimiser separately.
8072 * This is unfortunate as the optimiser isnt handling lookahead
8073 * properly currently.
8076 while ((OP(first) == OPEN && (sawopen = 1)) ||
8077 /* An OR of *one* alternative - should not happen now. */
8078 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
8079 /* for now we can't handle lookbehind IFMATCH*/
8080 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
8081 (OP(first) == PLUS) ||
8082 (OP(first) == MINMOD) ||
8083 /* An {n,m} with n>0 */
8084 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
8085 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
8088 * the only op that could be a regnode is PLUS, all the rest
8089 * will be regnode_1 or regnode_2.
8091 * (yves doesn't think this is true)
8093 if (OP(first) == PLUS)
8096 if (OP(first) == MINMOD)
8098 first += regarglen[OP(first)];
8100 first = NEXTOPER(first);
8101 first_next= regnext(first);
8104 /* Starting-point info. */
8106 DEBUG_PEEP("first:", first, 0, 0);
8107 /* Ignore EXACT as we deal with it later. */
8108 if (PL_regkind[OP(first)] == EXACT) {
8109 if (! isEXACTFish(OP(first))) {
8110 NOOP; /* Empty, get anchored substr later. */
8113 RExC_rxi->regstclass = first;
8116 else if (PL_regkind[OP(first)] == TRIE &&
8117 ((reg_trie_data *)RExC_rxi->data->data[ ARG(first) ])->minlen>0)
8119 /* this can happen only on restudy */
8120 RExC_rxi->regstclass = construct_ahocorasick_from_trie(pRExC_state, (regnode *)first, 0);
8123 else if (REGNODE_SIMPLE(OP(first)))
8124 RExC_rxi->regstclass = first;
8125 else if (PL_regkind[OP(first)] == BOUND ||
8126 PL_regkind[OP(first)] == NBOUND)
8127 RExC_rxi->regstclass = first;
8128 else if (PL_regkind[OP(first)] == BOL) {
8129 RExC_rx->intflags |= (OP(first) == MBOL
8132 first = NEXTOPER(first);
8135 else if (OP(first) == GPOS) {
8136 RExC_rx->intflags |= PREGf_ANCH_GPOS;
8137 first = NEXTOPER(first);
8140 else if ((!sawopen || !RExC_sawback) &&
8142 (OP(first) == STAR &&
8143 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
8144 !(RExC_rx->intflags & PREGf_ANCH) && !pRExC_state->code_blocks)
8146 /* turn .* into ^.* with an implied $*=1 */
8148 (OP(NEXTOPER(first)) == REG_ANY)
8151 RExC_rx->intflags |= (type | PREGf_IMPLICIT);
8152 first = NEXTOPER(first);
8155 if (sawplus && !sawminmod && !sawlookahead
8156 && (!sawopen || !RExC_sawback)
8157 && !pRExC_state->code_blocks) /* May examine pos and $& */
8158 /* x+ must match at the 1st pos of run of x's */
8159 RExC_rx->intflags |= PREGf_SKIP;
8161 /* Scan is after the zeroth branch, first is atomic matcher. */
8162 #ifdef TRIE_STUDY_OPT
8165 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
8166 (IV)(first - scan + 1))
8170 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
8171 (IV)(first - scan + 1))
8177 * If there's something expensive in the r.e., find the
8178 * longest literal string that must appear and make it the
8179 * regmust. Resolve ties in favor of later strings, since
8180 * the regstart check works with the beginning of the r.e.
8181 * and avoiding duplication strengthens checking. Not a
8182 * strong reason, but sufficient in the absence of others.
8183 * [Now we resolve ties in favor of the earlier string if
8184 * it happens that c_offset_min has been invalidated, since the
8185 * earlier string may buy us something the later one won't.]
8188 data.substrs[0].str = newSVpvs("");
8189 data.substrs[1].str = newSVpvs("");
8190 data.last_found = newSVpvs("");
8191 data.cur_is_floating = 0; /* initially any found substring is fixed */
8192 ENTER_with_name("study_chunk");
8193 SAVEFREESV(data.substrs[0].str);
8194 SAVEFREESV(data.substrs[1].str);
8195 SAVEFREESV(data.last_found);
8197 if (!RExC_rxi->regstclass) {
8198 ssc_init(pRExC_state, &ch_class);
8199 data.start_class = &ch_class;
8200 stclass_flag = SCF_DO_STCLASS_AND;
8201 } else /* XXXX Check for BOUND? */
8203 data.last_closep = &last_close;
8207 * MAIN ENTRY FOR study_chunk() FOR m/PATTERN/
8208 * (NO top level branches)
8210 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
8211 scan + RExC_size, /* Up to end */
8213 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
8214 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
8218 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
8221 if ( RExC_total_parens == 1 && !data.cur_is_floating
8222 && data.last_start_min == 0 && data.last_end > 0
8223 && !RExC_seen_zerolen
8224 && !(RExC_seen & REG_VERBARG_SEEN)
8225 && !(RExC_seen & REG_GPOS_SEEN)
8227 RExC_rx->extflags |= RXf_CHECK_ALL;
8229 scan_commit(pRExC_state, &data,&minlen, 0);
8232 /* XXX this is done in reverse order because that's the way the
8233 * code was before it was parameterised. Don't know whether it
8234 * actually needs doing in reverse order. DAPM */
8235 for (i = 1; i >= 0; i--) {
8236 longest_length[i] = CHR_SVLEN(data.substrs[i].str);
8239 && SvCUR(data.substrs[0].str) /* ok to leave SvCUR */
8240 && data.substrs[0].min_offset
8241 == data.substrs[1].min_offset
8242 && SvCUR(data.substrs[0].str)
8243 == SvCUR(data.substrs[1].str)
8245 && S_setup_longest (aTHX_ pRExC_state,
8246 &(RExC_rx->substrs->data[i]),
8250 RExC_rx->substrs->data[i].min_offset =
8251 data.substrs[i].min_offset - data.substrs[i].lookbehind;
8253 RExC_rx->substrs->data[i].max_offset = data.substrs[i].max_offset;
8254 /* Don't offset infinity */
8255 if (data.substrs[i].max_offset < OPTIMIZE_INFTY)
8256 RExC_rx->substrs->data[i].max_offset -= data.substrs[i].lookbehind;
8257 SvREFCNT_inc_simple_void_NN(data.substrs[i].str);
8260 RExC_rx->substrs->data[i].substr = NULL;
8261 RExC_rx->substrs->data[i].utf8_substr = NULL;
8262 longest_length[i] = 0;
8266 LEAVE_with_name("study_chunk");
8268 if (RExC_rxi->regstclass
8269 && (OP(RExC_rxi->regstclass) == REG_ANY || OP(RExC_rxi->regstclass) == SANY))
8270 RExC_rxi->regstclass = NULL;
8272 if ((!(RExC_rx->substrs->data[0].substr || RExC_rx->substrs->data[0].utf8_substr)
8273 || RExC_rx->substrs->data[0].min_offset)
8275 && ! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
8276 && is_ssc_worth_it(pRExC_state, data.start_class))
8278 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
8280 ssc_finalize(pRExC_state, data.start_class);
8282 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
8283 StructCopy(data.start_class,
8284 (regnode_ssc*)RExC_rxi->data->data[n],
8286 RExC_rxi->regstclass = (regnode*)RExC_rxi->data->data[n];
8287 RExC_rx->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
8288 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
8289 regprop(RExC_rx, sv, (regnode*)data.start_class, NULL, pRExC_state);
8290 Perl_re_printf( aTHX_
8291 "synthetic stclass \"%s\".\n",
8292 SvPVX_const(sv));});
8293 data.start_class = NULL;
8296 /* A temporary algorithm prefers floated substr to fixed one of
8297 * same length to dig more info. */
8298 i = (longest_length[0] <= longest_length[1]);
8299 RExC_rx->substrs->check_ix = i;
8300 RExC_rx->check_end_shift = RExC_rx->substrs->data[i].end_shift;
8301 RExC_rx->check_substr = RExC_rx->substrs->data[i].substr;
8302 RExC_rx->check_utf8 = RExC_rx->substrs->data[i].utf8_substr;
8303 RExC_rx->check_offset_min = RExC_rx->substrs->data[i].min_offset;
8304 RExC_rx->check_offset_max = RExC_rx->substrs->data[i].max_offset;
8305 if (!i && (RExC_rx->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS)))
8306 RExC_rx->intflags |= PREGf_NOSCAN;
8308 if ((RExC_rx->check_substr || RExC_rx->check_utf8) ) {
8309 RExC_rx->extflags |= RXf_USE_INTUIT;
8310 if (SvTAIL(RExC_rx->check_substr ? RExC_rx->check_substr : RExC_rx->check_utf8))
8311 RExC_rx->extflags |= RXf_INTUIT_TAIL;
8314 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
8315 if ( (STRLEN)minlen < longest_length[1] )
8316 minlen= longest_length[1];
8317 if ( (STRLEN)minlen < longest_length[0] )
8318 minlen= longest_length[0];
8322 /* Several toplevels. Best we can is to set minlen. */
8324 regnode_ssc ch_class;
8325 SSize_t last_close = 0;
8327 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "\nMulti Top Level\n"));
8329 scan = RExC_rxi->program + 1;
8330 ssc_init(pRExC_state, &ch_class);
8331 data.start_class = &ch_class;
8332 data.last_closep = &last_close;
8336 * MAIN ENTRY FOR study_chunk() FOR m/P1|P2|.../
8337 * (patterns WITH top level branches)
8339 minlen = study_chunk(pRExC_state,
8340 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
8341 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
8342 ? SCF_TRIE_DOING_RESTUDY
8346 CHECK_RESTUDY_GOTO_butfirst(NOOP);
8348 RExC_rx->check_substr = NULL;
8349 RExC_rx->check_utf8 = NULL;
8350 RExC_rx->substrs->data[0].substr = NULL;
8351 RExC_rx->substrs->data[0].utf8_substr = NULL;
8352 RExC_rx->substrs->data[1].substr = NULL;
8353 RExC_rx->substrs->data[1].utf8_substr = NULL;
8355 if (! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
8356 && is_ssc_worth_it(pRExC_state, data.start_class))
8358 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
8360 ssc_finalize(pRExC_state, data.start_class);
8362 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
8363 StructCopy(data.start_class,
8364 (regnode_ssc*)RExC_rxi->data->data[n],
8366 RExC_rxi->regstclass = (regnode*)RExC_rxi->data->data[n];
8367 RExC_rx->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
8368 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
8369 regprop(RExC_rx, sv, (regnode*)data.start_class, NULL, pRExC_state);
8370 Perl_re_printf( aTHX_
8371 "synthetic stclass \"%s\".\n",
8372 SvPVX_const(sv));});
8373 data.start_class = NULL;
8377 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
8378 RExC_rx->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
8379 RExC_rx->maxlen = REG_INFTY;
8382 RExC_rx->maxlen = RExC_maxlen;
8385 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
8386 the "real" pattern. */
8388 Perl_re_printf( aTHX_ "minlen: %" IVdf " RExC_rx->minlen:%" IVdf " maxlen:%" IVdf "\n",
8389 (IV)minlen, (IV)RExC_rx->minlen, (IV)RExC_maxlen);
8391 RExC_rx->minlenret = minlen;
8392 if (RExC_rx->minlen < minlen)
8393 RExC_rx->minlen = minlen;
8395 if (RExC_seen & REG_RECURSE_SEEN ) {
8396 RExC_rx->intflags |= PREGf_RECURSE_SEEN;
8397 Newx(RExC_rx->recurse_locinput, RExC_rx->nparens + 1, char *);
8399 if (RExC_seen & REG_GPOS_SEEN)
8400 RExC_rx->intflags |= PREGf_GPOS_SEEN;
8401 if (RExC_seen & REG_LOOKBEHIND_SEEN)
8402 RExC_rx->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
8404 if (pRExC_state->code_blocks)
8405 RExC_rx->extflags |= RXf_EVAL_SEEN;
8406 if (RExC_seen & REG_VERBARG_SEEN)
8408 RExC_rx->intflags |= PREGf_VERBARG_SEEN;
8409 RExC_rx->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
8411 if (RExC_seen & REG_CUTGROUP_SEEN)
8412 RExC_rx->intflags |= PREGf_CUTGROUP_SEEN;
8413 if (pm_flags & PMf_USE_RE_EVAL)
8414 RExC_rx->intflags |= PREGf_USE_RE_EVAL;
8415 if (RExC_paren_names)
8416 RXp_PAREN_NAMES(RExC_rx) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
8418 RXp_PAREN_NAMES(RExC_rx) = NULL;
8420 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
8421 * so it can be used in pp.c */
8422 if (RExC_rx->intflags & PREGf_ANCH)
8423 RExC_rx->extflags |= RXf_IS_ANCHORED;
8427 /* this is used to identify "special" patterns that might result
8428 * in Perl NOT calling the regex engine and instead doing the match "itself",
8429 * particularly special cases in split//. By having the regex compiler
8430 * do this pattern matching at a regop level (instead of by inspecting the pattern)
8431 * we avoid weird issues with equivalent patterns resulting in different behavior,
8432 * AND we allow non Perl engines to get the same optimizations by the setting the
8433 * flags appropriately - Yves */
8434 regnode *first = RExC_rxi->program + 1;
8436 regnode *next = regnext(first);
8439 if (PL_regkind[fop] == NOTHING && nop == END)
8440 RExC_rx->extflags |= RXf_NULL;
8441 else if ((fop == MBOL || (fop == SBOL && !first->flags)) && nop == END)
8442 /* when fop is SBOL first->flags will be true only when it was
8443 * produced by parsing /\A/, and not when parsing /^/. This is
8444 * very important for the split code as there we want to
8445 * treat /^/ as /^/m, but we do not want to treat /\A/ as /^/m.
8446 * See rt #122761 for more details. -- Yves */
8447 RExC_rx->extflags |= RXf_START_ONLY;
8448 else if (fop == PLUS
8449 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
8451 RExC_rx->extflags |= RXf_WHITE;
8452 else if ( RExC_rx->extflags & RXf_SPLIT
8453 && (PL_regkind[fop] == EXACT && ! isEXACTFish(fop))
8454 && STR_LEN(first) == 1
8455 && *(STRING(first)) == ' '
8457 RExC_rx->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
8461 if (RExC_contains_locale) {
8462 RXp_EXTFLAGS(RExC_rx) |= RXf_TAINTED;
8466 if (RExC_paren_names) {
8467 RExC_rxi->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
8468 RExC_rxi->data->data[RExC_rxi->name_list_idx]
8469 = (void*)SvREFCNT_inc(RExC_paren_name_list);
8472 RExC_rxi->name_list_idx = 0;
8474 while ( RExC_recurse_count > 0 ) {
8475 const regnode *scan = RExC_recurse[ --RExC_recurse_count ];
8477 * This data structure is set up in study_chunk() and is used
8478 * to calculate the distance between a GOSUB regopcode and
8479 * the OPEN/CURLYM (CURLYM's are special and can act like OPEN's)
8482 * If for some reason someone writes code that optimises
8483 * away a GOSUB opcode then the assert should be changed to
8484 * an if(scan) to guard the ARG2L_SET() - Yves
8487 assert(scan && OP(scan) == GOSUB);
8488 ARG2L_SET( scan, RExC_open_parens[ARG(scan)] - REGNODE_OFFSET(scan));
8491 Newxz(RExC_rx->offs, RExC_total_parens, regexp_paren_pair);
8492 /* assume we don't need to swap parens around before we match */
8494 Perl_re_printf( aTHX_ "study_chunk_recursed_count: %lu\n",
8495 (unsigned long)RExC_study_chunk_recursed_count);
8499 Perl_re_printf( aTHX_ "Final program:\n");
8503 if (RExC_open_parens) {
8504 Safefree(RExC_open_parens);
8505 RExC_open_parens = NULL;
8507 if (RExC_close_parens) {
8508 Safefree(RExC_close_parens);
8509 RExC_close_parens = NULL;
8513 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
8514 * by setting the regexp SV to readonly-only instead. If the
8515 * pattern's been recompiled, the USEDness should remain. */
8516 if (old_re && SvREADONLY(old_re))
8524 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
8527 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
8529 PERL_UNUSED_ARG(value);
8531 if (flags & RXapif_FETCH) {
8532 return reg_named_buff_fetch(rx, key, flags);
8533 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
8534 Perl_croak_no_modify();
8536 } else if (flags & RXapif_EXISTS) {
8537 return reg_named_buff_exists(rx, key, flags)
8540 } else if (flags & RXapif_REGNAMES) {
8541 return reg_named_buff_all(rx, flags);
8542 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
8543 return reg_named_buff_scalar(rx, flags);
8545 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
8551 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
8554 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
8555 PERL_UNUSED_ARG(lastkey);
8557 if (flags & RXapif_FIRSTKEY)
8558 return reg_named_buff_firstkey(rx, flags);
8559 else if (flags & RXapif_NEXTKEY)
8560 return reg_named_buff_nextkey(rx, flags);
8562 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
8569 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
8573 struct regexp *const rx = ReANY(r);
8575 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
8577 if (rx && RXp_PAREN_NAMES(rx)) {
8578 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
8581 SV* sv_dat=HeVAL(he_str);
8582 I32 *nums=(I32*)SvPVX(sv_dat);
8583 AV * const retarray = (flags & RXapif_ALL) ? newAV() : NULL;
8584 for ( i=0; i<SvIVX(sv_dat); i++ ) {
8585 if ((I32)(rx->nparens) >= nums[i]
8586 && rx->offs[nums[i]].start != -1
8587 && rx->offs[nums[i]].end != -1)
8590 CALLREG_NUMBUF_FETCH(r, nums[i], ret);
8595 ret = newSVsv(&PL_sv_undef);
8598 av_push(retarray, ret);
8601 return newRV_noinc(MUTABLE_SV(retarray));
8608 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
8611 struct regexp *const rx = ReANY(r);
8613 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
8615 if (rx && RXp_PAREN_NAMES(rx)) {
8616 if (flags & RXapif_ALL) {
8617 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
8619 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
8621 SvREFCNT_dec_NN(sv);
8633 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
8635 struct regexp *const rx = ReANY(r);
8637 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
8639 if ( rx && RXp_PAREN_NAMES(rx) ) {
8640 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
8642 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
8649 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
8651 struct regexp *const rx = ReANY(r);
8652 DECLARE_AND_GET_RE_DEBUG_FLAGS;
8654 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
8656 if (rx && RXp_PAREN_NAMES(rx)) {
8657 HV *hv = RXp_PAREN_NAMES(rx);
8659 while ( (temphe = hv_iternext_flags(hv, 0)) ) {
8662 SV* sv_dat = HeVAL(temphe);
8663 I32 *nums = (I32*)SvPVX(sv_dat);
8664 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8665 if ((I32)(rx->lastparen) >= nums[i] &&
8666 rx->offs[nums[i]].start != -1 &&
8667 rx->offs[nums[i]].end != -1)
8673 if (parno || flags & RXapif_ALL) {
8674 return newSVhek(HeKEY_hek(temphe));
8682 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
8687 struct regexp *const rx = ReANY(r);
8689 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
8691 if (rx && RXp_PAREN_NAMES(rx)) {
8692 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
8693 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
8694 } else if (flags & RXapif_ONE) {
8695 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
8696 av = MUTABLE_AV(SvRV(ret));
8697 length = av_count(av);
8698 SvREFCNT_dec_NN(ret);
8699 return newSViv(length);
8701 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
8706 return &PL_sv_undef;
8710 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
8712 struct regexp *const rx = ReANY(r);
8715 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
8717 if (rx && RXp_PAREN_NAMES(rx)) {
8718 HV *hv= RXp_PAREN_NAMES(rx);
8720 (void)hv_iterinit(hv);
8721 while ( (temphe = hv_iternext_flags(hv, 0)) ) {
8724 SV* sv_dat = HeVAL(temphe);
8725 I32 *nums = (I32*)SvPVX(sv_dat);
8726 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8727 if ((I32)(rx->lastparen) >= nums[i] &&
8728 rx->offs[nums[i]].start != -1 &&
8729 rx->offs[nums[i]].end != -1)
8735 if (parno || flags & RXapif_ALL) {
8736 av_push(av, newSVhek(HeKEY_hek(temphe)));
8741 return newRV_noinc(MUTABLE_SV(av));
8745 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
8748 struct regexp *const rx = ReANY(r);
8754 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
8756 if ( n == RX_BUFF_IDX_CARET_PREMATCH
8757 || n == RX_BUFF_IDX_CARET_FULLMATCH
8758 || n == RX_BUFF_IDX_CARET_POSTMATCH
8761 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8763 /* on something like
8766 * the KEEPCOPY is set on the PMOP rather than the regex */
8767 if (PL_curpm && r == PM_GETRE(PL_curpm))
8768 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8777 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
8778 /* no need to distinguish between them any more */
8779 n = RX_BUFF_IDX_FULLMATCH;
8781 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
8782 && rx->offs[0].start != -1)
8784 /* $`, ${^PREMATCH} */
8785 i = rx->offs[0].start;
8789 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
8790 && rx->offs[0].end != -1)
8792 /* $', ${^POSTMATCH} */
8793 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
8794 i = rx->sublen + rx->suboffset - rx->offs[0].end;
8797 if (inRANGE(n, 0, (I32)rx->nparens) &&
8798 (s1 = rx->offs[n].start) != -1 &&
8799 (t1 = rx->offs[n].end) != -1)
8801 /* $&, ${^MATCH}, $1 ... */
8803 s = rx->subbeg + s1 - rx->suboffset;
8808 assert(s >= rx->subbeg);
8809 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
8811 #ifdef NO_TAINT_SUPPORT
8812 sv_setpvn(sv, s, i);
8814 const int oldtainted = TAINT_get;
8816 sv_setpvn(sv, s, i);
8817 TAINT_set(oldtainted);
8819 if (RXp_MATCH_UTF8(rx))
8824 if (RXp_MATCH_TAINTED(rx)) {
8825 if (SvTYPE(sv) >= SVt_PVMG) {
8826 MAGIC* const mg = SvMAGIC(sv);
8829 SvMAGIC_set(sv, mg->mg_moremagic);
8831 if ((mgt = SvMAGIC(sv))) {
8832 mg->mg_moremagic = mgt;
8833 SvMAGIC_set(sv, mg);
8850 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
8851 SV const * const value)
8853 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
8855 PERL_UNUSED_ARG(rx);
8856 PERL_UNUSED_ARG(paren);
8857 PERL_UNUSED_ARG(value);
8860 Perl_croak_no_modify();
8864 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
8867 struct regexp *const rx = ReANY(r);
8871 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
8873 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
8874 || paren == RX_BUFF_IDX_CARET_FULLMATCH
8875 || paren == RX_BUFF_IDX_CARET_POSTMATCH
8878 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8880 /* on something like
8883 * the KEEPCOPY is set on the PMOP rather than the regex */
8884 if (PL_curpm && r == PM_GETRE(PL_curpm))
8885 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8891 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
8893 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
8894 case RX_BUFF_IDX_PREMATCH: /* $` */
8895 if (rx->offs[0].start != -1) {
8896 i = rx->offs[0].start;
8905 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
8906 case RX_BUFF_IDX_POSTMATCH: /* $' */
8907 if (rx->offs[0].end != -1) {
8908 i = rx->sublen - rx->offs[0].end;
8910 s1 = rx->offs[0].end;
8917 default: /* $& / ${^MATCH}, $1, $2, ... */
8918 if (paren <= (I32)rx->nparens &&
8919 (s1 = rx->offs[paren].start) != -1 &&
8920 (t1 = rx->offs[paren].end) != -1)
8926 if (ckWARN(WARN_UNINITIALIZED))
8927 report_uninit((const SV *)sv);
8932 if (i > 0 && RXp_MATCH_UTF8(rx)) {
8933 const char * const s = rx->subbeg - rx->suboffset + s1;
8938 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
8945 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
8947 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
8948 PERL_UNUSED_ARG(rx);
8952 return newSVpvs("Regexp");
8955 /* Scans the name of a named buffer from the pattern.
8956 * If flags is REG_RSN_RETURN_NULL returns null.
8957 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
8958 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
8959 * to the parsed name as looked up in the RExC_paren_names hash.
8960 * If there is an error throws a vFAIL().. type exception.
8963 #define REG_RSN_RETURN_NULL 0
8964 #define REG_RSN_RETURN_NAME 1
8965 #define REG_RSN_RETURN_DATA 2
8968 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
8970 char *name_start = RExC_parse;
8973 PERL_ARGS_ASSERT_REG_SCAN_NAME;
8975 assert (RExC_parse <= RExC_end);
8976 if (RExC_parse == RExC_end) NOOP;
8977 else if (isIDFIRST_lazy_if_safe(RExC_parse, RExC_end, UTF)) {
8978 /* Note that the code here assumes well-formed UTF-8. Skip IDFIRST by
8979 * using do...while */
8982 RExC_parse += UTF8SKIP(RExC_parse);
8983 } while ( RExC_parse < RExC_end
8984 && isWORDCHAR_utf8_safe((U8*)RExC_parse, (U8*) RExC_end));
8988 } while (RExC_parse < RExC_end && isWORDCHAR(*RExC_parse));
8990 RExC_parse++; /* so the <- from the vFAIL is after the offending
8992 vFAIL("Group name must start with a non-digit word character");
8994 sv_name = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
8995 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
8996 if ( flags == REG_RSN_RETURN_NAME)
8998 else if (flags==REG_RSN_RETURN_DATA) {
9001 if ( ! sv_name ) /* should not happen*/
9002 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
9003 if (RExC_paren_names)
9004 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
9006 sv_dat = HeVAL(he_str);
9007 if ( ! sv_dat ) { /* Didn't find group */
9009 /* It might be a forward reference; we can't fail until we
9010 * know, by completing the parse to get all the groups, and
9012 if (ALL_PARENS_COUNTED) {
9013 vFAIL("Reference to nonexistent named group");
9016 REQUIRE_PARENS_PASS;
9022 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
9023 (unsigned long) flags);
9026 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
9027 if (RExC_lastparse!=RExC_parse) { \
9028 Perl_re_printf( aTHX_ "%s", \
9029 Perl_pv_pretty(aTHX_ RExC_mysv1, RExC_parse, \
9030 RExC_end - RExC_parse, 16, \
9032 PERL_PV_ESCAPE_UNI_DETECT | \
9033 PERL_PV_PRETTY_ELLIPSES | \
9034 PERL_PV_PRETTY_LTGT | \
9035 PERL_PV_ESCAPE_RE | \
9036 PERL_PV_PRETTY_EXACTSIZE \
9040 Perl_re_printf( aTHX_ "%16s",""); \
9042 if (RExC_lastnum!=RExC_emit) \
9043 Perl_re_printf( aTHX_ "|%4zu", RExC_emit); \
9045 Perl_re_printf( aTHX_ "|%4s",""); \
9046 Perl_re_printf( aTHX_ "|%*s%-4s", \
9047 (int)((depth*2)), "", \
9050 RExC_lastnum=RExC_emit; \
9051 RExC_lastparse=RExC_parse; \
9056 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
9057 DEBUG_PARSE_MSG((funcname)); \
9058 Perl_re_printf( aTHX_ "%4s","\n"); \
9060 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({\
9061 DEBUG_PARSE_MSG((funcname)); \
9062 Perl_re_printf( aTHX_ fmt "\n",args); \
9065 /* This section of code defines the inversion list object and its methods. The
9066 * interfaces are highly subject to change, so as much as possible is static to
9067 * this file. An inversion list is here implemented as a malloc'd C UV array
9068 * as an SVt_INVLIST scalar.
9070 * An inversion list for Unicode is an array of code points, sorted by ordinal
9071 * number. Each element gives the code point that begins a range that extends
9072 * up-to but not including the code point given by the next element. The final
9073 * element gives the first code point of a range that extends to the platform's
9074 * infinity. The even-numbered elements (invlist[0], invlist[2], invlist[4],
9075 * ...) give ranges whose code points are all in the inversion list. We say
9076 * that those ranges are in the set. The odd-numbered elements give ranges
9077 * whose code points are not in the inversion list, and hence not in the set.
9078 * Thus, element [0] is the first code point in the list. Element [1]
9079 * is the first code point beyond that not in the list; and element [2] is the
9080 * first code point beyond that that is in the list. In other words, the first
9081 * range is invlist[0]..(invlist[1]-1), and all code points in that range are
9082 * in the inversion list. The second range is invlist[1]..(invlist[2]-1), and
9083 * all code points in that range are not in the inversion list. The third
9084 * range invlist[2]..(invlist[3]-1) gives code points that are in the inversion
9085 * list, and so forth. Thus every element whose index is divisible by two
9086 * gives the beginning of a range that is in the list, and every element whose
9087 * index is not divisible by two gives the beginning of a range not in the
9088 * list. If the final element's index is divisible by two, the inversion list
9089 * extends to the platform's infinity; otherwise the highest code point in the
9090 * inversion list is the contents of that element minus 1.
9092 * A range that contains just a single code point N will look like
9094 * invlist[i+1] == N+1
9096 * If N is UV_MAX (the highest representable code point on the machine), N+1 is
9097 * impossible to represent, so element [i+1] is omitted. The single element
9099 * invlist[0] == UV_MAX
9100 * contains just UV_MAX, but is interpreted as matching to infinity.
9102 * Taking the complement (inverting) an inversion list is quite simple, if the
9103 * first element is 0, remove it; otherwise add a 0 element at the beginning.
9104 * This implementation reserves an element at the beginning of each inversion
9105 * list to always contain 0; there is an additional flag in the header which
9106 * indicates if the list begins at the 0, or is offset to begin at the next
9107 * element. This means that the inversion list can be inverted without any
9108 * copying; just flip the flag.
9110 * More about inversion lists can be found in "Unicode Demystified"
9111 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
9113 * The inversion list data structure is currently implemented as an SV pointing
9114 * to an array of UVs that the SV thinks are bytes. This allows us to have an
9115 * array of UV whose memory management is automatically handled by the existing
9116 * facilities for SV's.
9118 * Some of the methods should always be private to the implementation, and some
9119 * should eventually be made public */
9121 /* The header definitions are in F<invlist_inline.h> */
9123 #ifndef PERL_IN_XSUB_RE
9125 PERL_STATIC_INLINE UV*
9126 S__invlist_array_init(SV* const invlist, const bool will_have_0)
9128 /* Returns a pointer to the first element in the inversion list's array.
9129 * This is called upon initialization of an inversion list. Where the
9130 * array begins depends on whether the list has the code point U+0000 in it
9131 * or not. The other parameter tells it whether the code that follows this
9132 * call is about to put a 0 in the inversion list or not. The first
9133 * element is either the element reserved for 0, if TRUE, or the element
9134 * after it, if FALSE */
9136 bool* offset = get_invlist_offset_addr(invlist);
9137 UV* zero_addr = (UV *) SvPVX(invlist);
9139 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
9142 assert(! _invlist_len(invlist));
9146 /* 1^1 = 0; 1^0 = 1 */
9147 *offset = 1 ^ will_have_0;
9148 return zero_addr + *offset;
9152 S_invlist_replace_list_destroys_src(pTHX_ SV * dest, SV * src)
9154 /* Replaces the inversion list in 'dest' with the one from 'src'. It
9155 * steals the list from 'src', so 'src' is made to have a NULL list. This
9156 * is similar to what SvSetMagicSV() would do, if it were implemented on
9157 * inversion lists, though this routine avoids a copy */
9159 const UV src_len = _invlist_len(src);
9160 const bool src_offset = *get_invlist_offset_addr(src);
9161 const STRLEN src_byte_len = SvLEN(src);
9162 char * array = SvPVX(src);
9164 const int oldtainted = TAINT_get;
9166 PERL_ARGS_ASSERT_INVLIST_REPLACE_LIST_DESTROYS_SRC;
9168 assert(is_invlist(src));
9169 assert(is_invlist(dest));
9170 assert(! invlist_is_iterating(src));
9171 assert(SvCUR(src) == 0 || SvCUR(src) < SvLEN(src));
9173 /* Make sure it ends in the right place with a NUL, as our inversion list
9174 * manipulations aren't careful to keep this true, but sv_usepvn_flags()
9176 array[src_byte_len - 1] = '\0';
9178 TAINT_NOT; /* Otherwise it breaks */
9179 sv_usepvn_flags(dest,
9183 /* This flag is documented to cause a copy to be avoided */
9184 SV_HAS_TRAILING_NUL);
9185 TAINT_set(oldtainted);
9190 /* Finish up copying over the other fields in an inversion list */
9191 *get_invlist_offset_addr(dest) = src_offset;
9192 invlist_set_len(dest, src_len, src_offset);
9193 *get_invlist_previous_index_addr(dest) = 0;
9194 invlist_iterfinish(dest);
9197 PERL_STATIC_INLINE IV*
9198 S_get_invlist_previous_index_addr(SV* invlist)
9200 /* Return the address of the IV that is reserved to hold the cached index
9202 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
9204 assert(is_invlist(invlist));
9206 return &(((XINVLIST*) SvANY(invlist))->prev_index);
9209 PERL_STATIC_INLINE IV
9210 S_invlist_previous_index(SV* const invlist)
9212 /* Returns cached index of previous search */
9214 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
9216 return *get_invlist_previous_index_addr(invlist);
9219 PERL_STATIC_INLINE void
9220 S_invlist_set_previous_index(SV* const invlist, const IV index)
9222 /* Caches <index> for later retrieval */
9224 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
9226 assert(index == 0 || index < (int) _invlist_len(invlist));
9228 *get_invlist_previous_index_addr(invlist) = index;
9231 PERL_STATIC_INLINE void
9232 S_invlist_trim(SV* invlist)
9234 /* Free the not currently-being-used space in an inversion list */
9236 /* But don't free up the space needed for the 0 UV that is always at the
9237 * beginning of the list, nor the trailing NUL */
9238 const UV min_size = TO_INTERNAL_SIZE(1) + 1;
9240 PERL_ARGS_ASSERT_INVLIST_TRIM;
9242 assert(is_invlist(invlist));
9244 SvPV_renew(invlist, MAX(min_size, SvCUR(invlist) + 1));
9247 PERL_STATIC_INLINE void
9248 S_invlist_clear(pTHX_ SV* invlist) /* Empty the inversion list */
9250 PERL_ARGS_ASSERT_INVLIST_CLEAR;
9252 assert(is_invlist(invlist));
9254 invlist_set_len(invlist, 0, 0);
9255 invlist_trim(invlist);
9258 #endif /* ifndef PERL_IN_XSUB_RE */
9260 PERL_STATIC_INLINE bool
9261 S_invlist_is_iterating(SV* const invlist)
9263 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
9265 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
9268 #ifndef PERL_IN_XSUB_RE
9270 PERL_STATIC_INLINE UV
9271 S_invlist_max(SV* const invlist)
9273 /* Returns the maximum number of elements storable in the inversion list's
9274 * array, without having to realloc() */
9276 PERL_ARGS_ASSERT_INVLIST_MAX;
9278 assert(is_invlist(invlist));
9280 /* Assumes worst case, in which the 0 element is not counted in the
9281 * inversion list, so subtracts 1 for that */
9282 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
9283 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
9284 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
9288 S_initialize_invlist_guts(pTHX_ SV* invlist, const Size_t initial_size)
9290 PERL_ARGS_ASSERT_INITIALIZE_INVLIST_GUTS;
9292 /* First 1 is in case the zero element isn't in the list; second 1 is for
9294 SvGROW(invlist, TO_INTERNAL_SIZE(initial_size + 1) + 1);
9295 invlist_set_len(invlist, 0, 0);
9297 /* Force iterinit() to be used to get iteration to work */
9298 invlist_iterfinish(invlist);
9300 *get_invlist_previous_index_addr(invlist) = 0;
9301 SvPOK_on(invlist); /* This allows B to extract the PV */
9305 Perl__new_invlist(pTHX_ IV initial_size)
9308 /* Return a pointer to a newly constructed inversion list, with enough
9309 * space to store 'initial_size' elements. If that number is negative, a
9310 * system default is used instead */
9314 if (initial_size < 0) {
9318 new_list = newSV_type(SVt_INVLIST);
9319 initialize_invlist_guts(new_list, initial_size);
9325 Perl__new_invlist_C_array(pTHX_ const UV* const list)
9327 /* Return a pointer to a newly constructed inversion list, initialized to
9328 * point to <list>, which has to be in the exact correct inversion list
9329 * form, including internal fields. Thus this is a dangerous routine that
9330 * should not be used in the wrong hands. The passed in 'list' contains
9331 * several header fields at the beginning that are not part of the
9332 * inversion list body proper */
9334 const STRLEN length = (STRLEN) list[0];
9335 const UV version_id = list[1];
9336 const bool offset = cBOOL(list[2]);
9337 #define HEADER_LENGTH 3
9338 /* If any of the above changes in any way, you must change HEADER_LENGTH
9339 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
9340 * perl -E 'say int(rand 2**31-1)'
9342 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
9343 data structure type, so that one being
9344 passed in can be validated to be an
9345 inversion list of the correct vintage.
9348 SV* invlist = newSV_type(SVt_INVLIST);
9350 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
9352 if (version_id != INVLIST_VERSION_ID) {
9353 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
9356 /* The generated array passed in includes header elements that aren't part
9357 * of the list proper, so start it just after them */
9358 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
9360 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
9361 shouldn't touch it */
9363 *(get_invlist_offset_addr(invlist)) = offset;
9365 /* The 'length' passed to us is the physical number of elements in the
9366 * inversion list. But if there is an offset the logical number is one
9368 invlist_set_len(invlist, length - offset, offset);
9370 invlist_set_previous_index(invlist, 0);
9372 /* Initialize the iteration pointer. */
9373 invlist_iterfinish(invlist);
9375 SvREADONLY_on(invlist);
9382 S__append_range_to_invlist(pTHX_ SV* const invlist,
9383 const UV start, const UV end)
9385 /* Subject to change or removal. Append the range from 'start' to 'end' at
9386 * the end of the inversion list. The range must be above any existing
9390 UV max = invlist_max(invlist);
9391 UV len = _invlist_len(invlist);
9394 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
9396 if (len == 0) { /* Empty lists must be initialized */
9397 offset = start != 0;
9398 array = _invlist_array_init(invlist, ! offset);
9401 /* Here, the existing list is non-empty. The current max entry in the
9402 * list is generally the first value not in the set, except when the
9403 * set extends to the end of permissible values, in which case it is
9404 * the first entry in that final set, and so this call is an attempt to
9405 * append out-of-order */
9407 UV final_element = len - 1;
9408 array = invlist_array(invlist);
9409 if ( array[final_element] > start
9410 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
9412 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",
9413 array[final_element], start,
9414 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
9417 /* Here, it is a legal append. If the new range begins 1 above the end
9418 * of the range below it, it is extending the range below it, so the
9419 * new first value not in the set is one greater than the newly
9420 * extended range. */
9421 offset = *get_invlist_offset_addr(invlist);
9422 if (array[final_element] == start) {
9423 if (end != UV_MAX) {
9424 array[final_element] = end + 1;
9427 /* But if the end is the maximum representable on the machine,
9428 * assume that infinity was actually what was meant. Just let
9429 * the range that this would extend to have no end */
9430 invlist_set_len(invlist, len - 1, offset);
9436 /* Here the new range doesn't extend any existing set. Add it */
9438 len += 2; /* Includes an element each for the start and end of range */
9440 /* If wll overflow the existing space, extend, which may cause the array to
9443 invlist_extend(invlist, len);
9445 /* Have to set len here to avoid assert failure in invlist_array() */
9446 invlist_set_len(invlist, len, offset);
9448 array = invlist_array(invlist);
9451 invlist_set_len(invlist, len, offset);
9454 /* The next item on the list starts the range, the one after that is
9455 * one past the new range. */
9456 array[len - 2] = start;
9457 if (end != UV_MAX) {
9458 array[len - 1] = end + 1;
9461 /* But if the end is the maximum representable on the machine, just let
9462 * the range have no end */
9463 invlist_set_len(invlist, len - 1, offset);
9468 Perl__invlist_search(SV* const invlist, const UV cp)
9470 /* Searches the inversion list for the entry that contains the input code
9471 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
9472 * return value is the index into the list's array of the range that
9473 * contains <cp>, that is, 'i' such that
9474 * array[i] <= cp < array[i+1]
9479 IV high = _invlist_len(invlist);
9480 const IV highest_element = high - 1;
9483 PERL_ARGS_ASSERT__INVLIST_SEARCH;
9485 /* If list is empty, return failure. */
9490 /* (We can't get the array unless we know the list is non-empty) */
9491 array = invlist_array(invlist);
9493 mid = invlist_previous_index(invlist);
9495 if (mid > highest_element) {
9496 mid = highest_element;
9499 /* <mid> contains the cache of the result of the previous call to this
9500 * function (0 the first time). See if this call is for the same result,
9501 * or if it is for mid-1. This is under the theory that calls to this
9502 * function will often be for related code points that are near each other.
9503 * And benchmarks show that caching gives better results. We also test
9504 * here if the code point is within the bounds of the list. These tests
9505 * replace others that would have had to be made anyway to make sure that
9506 * the array bounds were not exceeded, and these give us extra information
9507 * at the same time */
9508 if (cp >= array[mid]) {
9509 if (cp >= array[highest_element]) {
9510 return highest_element;
9513 /* Here, array[mid] <= cp < array[highest_element]. This means that
9514 * the final element is not the answer, so can exclude it; it also
9515 * means that <mid> is not the final element, so can refer to 'mid + 1'
9517 if (cp < array[mid + 1]) {
9523 else { /* cp < aray[mid] */
9524 if (cp < array[0]) { /* Fail if outside the array */
9528 if (cp >= array[mid - 1]) {
9533 /* Binary search. What we are looking for is <i> such that
9534 * array[i] <= cp < array[i+1]
9535 * The loop below converges on the i+1. Note that there may not be an
9536 * (i+1)th element in the array, and things work nonetheless */
9537 while (low < high) {
9538 mid = (low + high) / 2;
9539 assert(mid <= highest_element);
9540 if (array[mid] <= cp) { /* cp >= array[mid] */
9543 /* We could do this extra test to exit the loop early.
9544 if (cp < array[low]) {
9549 else { /* cp < array[mid] */
9556 invlist_set_previous_index(invlist, high);
9561 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9562 const bool complement_b, SV** output)
9564 /* Take the union of two inversion lists and point '*output' to it. On
9565 * input, '*output' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9566 * even 'a' or 'b'). If to an inversion list, the contents of the original
9567 * list will be replaced by the union. The first list, 'a', may be
9568 * NULL, in which case a copy of the second list is placed in '*output'.
9569 * If 'complement_b' is TRUE, the union is taken of the complement
9570 * (inversion) of 'b' instead of b itself.
9572 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9573 * Richard Gillam, published by Addison-Wesley, and explained at some
9574 * length there. The preface says to incorporate its examples into your
9575 * code at your own risk.
9577 * The algorithm is like a merge sort. */
9579 const UV* array_a; /* a's array */
9581 UV len_a; /* length of a's array */
9584 SV* u; /* the resulting union */
9588 UV i_a = 0; /* current index into a's array */
9592 /* running count, as explained in the algorithm source book; items are
9593 * stopped accumulating and are output when the count changes to/from 0.
9594 * The count is incremented when we start a range that's in an input's set,
9595 * and decremented when we start a range that's not in a set. So this
9596 * variable can be 0, 1, or 2. When it is 0 neither input is in their set,
9597 * and hence nothing goes into the union; 1, just one of the inputs is in
9598 * its set (and its current range gets added to the union); and 2 when both
9599 * inputs are in their sets. */
9602 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
9604 assert(*output == NULL || is_invlist(*output));
9606 len_b = _invlist_len(b);
9609 /* Here, 'b' is empty, hence it's complement is all possible code
9610 * points. So if the union includes the complement of 'b', it includes
9611 * everything, and we need not even look at 'a'. It's easiest to
9612 * create a new inversion list that matches everything. */
9614 SV* everything = _add_range_to_invlist(NULL, 0, UV_MAX);
9616 if (*output == NULL) { /* If the output didn't exist, just point it
9618 *output = everything;
9620 else { /* Otherwise, replace its contents with the new list */
9621 invlist_replace_list_destroys_src(*output, everything);
9622 SvREFCNT_dec_NN(everything);
9628 /* Here, we don't want the complement of 'b', and since 'b' is empty,
9629 * the union will come entirely from 'a'. If 'a' is NULL or empty, the
9630 * output will be empty */
9632 if (a == NULL || _invlist_len(a) == 0) {
9633 if (*output == NULL) {
9634 *output = _new_invlist(0);
9637 invlist_clear(*output);
9642 /* Here, 'a' is not empty, but 'b' is, so 'a' entirely determines the
9643 * union. We can just return a copy of 'a' if '*output' doesn't point
9644 * to an existing list */
9645 if (*output == NULL) {
9646 *output = invlist_clone(a, NULL);
9650 /* If the output is to overwrite 'a', we have a no-op, as it's
9656 /* Here, '*output' is to be overwritten by 'a' */
9657 u = invlist_clone(a, NULL);
9658 invlist_replace_list_destroys_src(*output, u);
9664 /* Here 'b' is not empty. See about 'a' */
9666 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
9668 /* Here, 'a' is empty (and b is not). That means the union will come
9669 * entirely from 'b'. If '*output' is NULL, we can directly return a
9670 * clone of 'b'. Otherwise, we replace the contents of '*output' with
9673 SV ** dest = (*output == NULL) ? output : &u;
9674 *dest = invlist_clone(b, NULL);
9676 _invlist_invert(*dest);
9680 invlist_replace_list_destroys_src(*output, u);
9687 /* Here both lists exist and are non-empty */
9688 array_a = invlist_array(a);
9689 array_b = invlist_array(b);
9691 /* If are to take the union of 'a' with the complement of b, set it
9692 * up so are looking at b's complement. */
9695 /* To complement, we invert: if the first element is 0, remove it. To
9696 * do this, we just pretend the array starts one later */
9697 if (array_b[0] == 0) {
9703 /* But if the first element is not zero, we pretend the list starts
9704 * at the 0 that is always stored immediately before the array. */
9710 /* Size the union for the worst case: that the sets are completely
9712 u = _new_invlist(len_a + len_b);
9714 /* Will contain U+0000 if either component does */
9715 array_u = _invlist_array_init(u, ( len_a > 0 && array_a[0] == 0)
9716 || (len_b > 0 && array_b[0] == 0));
9718 /* Go through each input list item by item, stopping when have exhausted
9720 while (i_a < len_a && i_b < len_b) {
9721 UV cp; /* The element to potentially add to the union's array */
9722 bool cp_in_set; /* is it in the input list's set or not */
9724 /* We need to take one or the other of the two inputs for the union.
9725 * Since we are merging two sorted lists, we take the smaller of the
9726 * next items. In case of a tie, we take first the one that is in its
9727 * set. If we first took the one not in its set, it would decrement
9728 * the count, possibly to 0 which would cause it to be output as ending
9729 * the range, and the next time through we would take the same number,
9730 * and output it again as beginning the next range. By doing it the
9731 * opposite way, there is no possibility that the count will be
9732 * momentarily decremented to 0, and thus the two adjoining ranges will
9733 * be seamlessly merged. (In a tie and both are in the set or both not
9734 * in the set, it doesn't matter which we take first.) */
9735 if ( array_a[i_a] < array_b[i_b]
9736 || ( array_a[i_a] == array_b[i_b]
9737 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9739 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9740 cp = array_a[i_a++];
9743 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9744 cp = array_b[i_b++];
9747 /* Here, have chosen which of the two inputs to look at. Only output
9748 * if the running count changes to/from 0, which marks the
9749 * beginning/end of a range that's in the set */
9752 array_u[i_u++] = cp;
9759 array_u[i_u++] = cp;
9765 /* The loop above increments the index into exactly one of the input lists
9766 * each iteration, and ends when either index gets to its list end. That
9767 * means the other index is lower than its end, and so something is
9768 * remaining in that one. We decrement 'count', as explained below, if
9769 * that list is in its set. (i_a and i_b each currently index the element
9770 * beyond the one we care about.) */
9771 if ( (i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9772 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9777 /* Above we decremented 'count' if the list that had unexamined elements in
9778 * it was in its set. This has made it so that 'count' being non-zero
9779 * means there isn't anything left to output; and 'count' equal to 0 means
9780 * that what is left to output is precisely that which is left in the
9781 * non-exhausted input list.
9783 * To see why, note first that the exhausted input obviously has nothing
9784 * left to add to the union. If it was in its set at its end, that means
9785 * the set extends from here to the platform's infinity, and hence so does
9786 * the union and the non-exhausted set is irrelevant. The exhausted set
9787 * also contributed 1 to 'count'. If 'count' was 2, it got decremented to
9788 * 1, but if it was 1, the non-exhausted set wasn't in its set, and so
9789 * 'count' remains at 1. This is consistent with the decremented 'count'
9790 * != 0 meaning there's nothing left to add to the union.
9792 * But if the exhausted input wasn't in its set, it contributed 0 to
9793 * 'count', and the rest of the union will be whatever the other input is.
9794 * If 'count' was 0, neither list was in its set, and 'count' remains 0;
9795 * otherwise it gets decremented to 0. This is consistent with 'count'
9796 * == 0 meaning the remainder of the union is whatever is left in the
9797 * non-exhausted list. */
9802 IV copy_count = len_a - i_a;
9803 if (copy_count > 0) { /* The non-exhausted input is 'a' */
9804 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
9806 else { /* The non-exhausted input is b */
9807 copy_count = len_b - i_b;
9808 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
9810 len_u = i_u + copy_count;
9813 /* Set the result to the final length, which can change the pointer to
9814 * array_u, so re-find it. (Note that it is unlikely that this will
9815 * change, as we are shrinking the space, not enlarging it) */
9816 if (len_u != _invlist_len(u)) {
9817 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
9819 array_u = invlist_array(u);
9822 if (*output == NULL) { /* Simply return the new inversion list */
9826 /* Otherwise, overwrite the inversion list that was in '*output'. We
9827 * could instead free '*output', and then set it to 'u', but experience
9828 * has shown [perl #127392] that if the input is a mortal, we can get a
9829 * huge build-up of these during regex compilation before they get
9831 invlist_replace_list_destroys_src(*output, u);
9839 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9840 const bool complement_b, SV** i)
9842 /* Take the intersection of two inversion lists and point '*i' to it. On
9843 * input, '*i' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9844 * even 'a' or 'b'). If to an inversion list, the contents of the original
9845 * list will be replaced by the intersection. The first list, 'a', may be
9846 * NULL, in which case '*i' will be an empty list. If 'complement_b' is
9847 * TRUE, the result will be the intersection of 'a' and the complement (or
9848 * inversion) of 'b' instead of 'b' directly.
9850 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9851 * Richard Gillam, published by Addison-Wesley, and explained at some
9852 * length there. The preface says to incorporate its examples into your
9853 * code at your own risk. In fact, it had bugs
9855 * The algorithm is like a merge sort, and is essentially the same as the
9859 const UV* array_a; /* a's array */
9861 UV len_a; /* length of a's array */
9864 SV* r; /* the resulting intersection */
9868 UV i_a = 0; /* current index into a's array */
9872 /* running count of how many of the two inputs are postitioned at ranges
9873 * that are in their sets. As explained in the algorithm source book,
9874 * items are stopped accumulating and are output when the count changes
9875 * to/from 2. The count is incremented when we start a range that's in an
9876 * input's set, and decremented when we start a range that's not in a set.
9877 * Only when it is 2 are we in the intersection. */
9880 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
9882 assert(*i == NULL || is_invlist(*i));
9884 /* Special case if either one is empty */
9885 len_a = (a == NULL) ? 0 : _invlist_len(a);
9886 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
9887 if (len_a != 0 && complement_b) {
9889 /* Here, 'a' is not empty, therefore from the enclosing 'if', 'b'
9890 * must be empty. Here, also we are using 'b's complement, which
9891 * hence must be every possible code point. Thus the intersection
9894 if (*i == a) { /* No-op */
9899 *i = invlist_clone(a, NULL);
9903 r = invlist_clone(a, NULL);
9904 invlist_replace_list_destroys_src(*i, r);
9909 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
9910 * intersection must be empty */
9912 *i = _new_invlist(0);
9920 /* Here both lists exist and are non-empty */
9921 array_a = invlist_array(a);
9922 array_b = invlist_array(b);
9924 /* If are to take the intersection of 'a' with the complement of b, set it
9925 * up so are looking at b's complement. */
9928 /* To complement, we invert: if the first element is 0, remove it. To
9929 * do this, we just pretend the array starts one later */
9930 if (array_b[0] == 0) {
9936 /* But if the first element is not zero, we pretend the list starts
9937 * at the 0 that is always stored immediately before the array. */
9943 /* Size the intersection for the worst case: that the intersection ends up
9944 * fragmenting everything to be completely disjoint */
9945 r= _new_invlist(len_a + len_b);
9947 /* Will contain U+0000 iff both components do */
9948 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
9949 && len_b > 0 && array_b[0] == 0);
9951 /* Go through each list item by item, stopping when have exhausted one of
9953 while (i_a < len_a && i_b < len_b) {
9954 UV cp; /* The element to potentially add to the intersection's
9956 bool cp_in_set; /* Is it in the input list's set or not */
9958 /* We need to take one or the other of the two inputs for the
9959 * intersection. Since we are merging two sorted lists, we take the
9960 * smaller of the next items. In case of a tie, we take first the one
9961 * that is not in its set (a difference from the union algorithm). If
9962 * we first took the one in its set, it would increment the count,
9963 * possibly to 2 which would cause it to be output as starting a range
9964 * in the intersection, and the next time through we would take that
9965 * same number, and output it again as ending the set. By doing the
9966 * opposite of this, there is no possibility that the count will be
9967 * momentarily incremented to 2. (In a tie and both are in the set or
9968 * both not in the set, it doesn't matter which we take first.) */
9969 if ( array_a[i_a] < array_b[i_b]
9970 || ( array_a[i_a] == array_b[i_b]
9971 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9973 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9974 cp = array_a[i_a++];
9977 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9981 /* Here, have chosen which of the two inputs to look at. Only output
9982 * if the running count changes to/from 2, which marks the
9983 * beginning/end of a range that's in the intersection */
9987 array_r[i_r++] = cp;
9992 array_r[i_r++] = cp;
9999 /* The loop above increments the index into exactly one of the input lists
10000 * each iteration, and ends when either index gets to its list end. That
10001 * means the other index is lower than its end, and so something is
10002 * remaining in that one. We increment 'count', as explained below, if the
10003 * exhausted list was in its set. (i_a and i_b each currently index the
10004 * element beyond the one we care about.) */
10005 if ( (i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
10006 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
10011 /* Above we incremented 'count' if the exhausted list was in its set. This
10012 * has made it so that 'count' being below 2 means there is nothing left to
10013 * output; otheriwse what's left to add to the intersection is precisely
10014 * that which is left in the non-exhausted input list.
10016 * To see why, note first that the exhausted input obviously has nothing
10017 * left to affect the intersection. If it was in its set at its end, that
10018 * means the set extends from here to the platform's infinity, and hence
10019 * anything in the non-exhausted's list will be in the intersection, and
10020 * anything not in it won't be. Hence, the rest of the intersection is
10021 * precisely what's in the non-exhausted list The exhausted set also
10022 * contributed 1 to 'count', meaning 'count' was at least 1. Incrementing
10023 * it means 'count' is now at least 2. This is consistent with the
10024 * incremented 'count' being >= 2 means to add the non-exhausted list to
10025 * the intersection.
10027 * But if the exhausted input wasn't in its set, it contributed 0 to
10028 * 'count', and the intersection can't include anything further; the
10029 * non-exhausted set is irrelevant. 'count' was at most 1, and doesn't get
10030 * incremented. This is consistent with 'count' being < 2 meaning nothing
10031 * further to add to the intersection. */
10032 if (count < 2) { /* Nothing left to put in the intersection. */
10035 else { /* copy the non-exhausted list, unchanged. */
10036 IV copy_count = len_a - i_a;
10037 if (copy_count > 0) { /* a is the one with stuff left */
10038 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
10040 else { /* b is the one with stuff left */
10041 copy_count = len_b - i_b;
10042 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
10044 len_r = i_r + copy_count;
10047 /* Set the result to the final length, which can change the pointer to
10048 * array_r, so re-find it. (Note that it is unlikely that this will
10049 * change, as we are shrinking the space, not enlarging it) */
10050 if (len_r != _invlist_len(r)) {
10051 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
10053 array_r = invlist_array(r);
10056 if (*i == NULL) { /* Simply return the calculated intersection */
10059 else { /* Otherwise, replace the existing inversion list in '*i'. We could
10060 instead free '*i', and then set it to 'r', but experience has
10061 shown [perl #127392] that if the input is a mortal, we can get a
10062 huge build-up of these during regex compilation before they get
10065 invlist_replace_list_destroys_src(*i, r);
10070 SvREFCNT_dec_NN(r);
10077 Perl__add_range_to_invlist(pTHX_ SV* invlist, UV start, UV end)
10079 /* Add the range from 'start' to 'end' inclusive to the inversion list's
10080 * set. A pointer to the inversion list is returned. This may actually be
10081 * a new list, in which case the passed in one has been destroyed. The
10082 * passed-in inversion list can be NULL, in which case a new one is created
10083 * with just the one range in it. The new list is not necessarily
10084 * NUL-terminated. Space is not freed if the inversion list shrinks as a
10085 * result of this function. The gain would not be large, and in many
10086 * cases, this is called multiple times on a single inversion list, so
10087 * anything freed may almost immediately be needed again.
10089 * This used to mostly call the 'union' routine, but that is much more
10090 * heavyweight than really needed for a single range addition */
10092 UV* array; /* The array implementing the inversion list */
10093 UV len; /* How many elements in 'array' */
10094 SSize_t i_s; /* index into the invlist array where 'start'
10096 SSize_t i_e = 0; /* And the index where 'end' should go */
10097 UV cur_highest; /* The highest code point in the inversion list
10098 upon entry to this function */
10100 /* This range becomes the whole inversion list if none already existed */
10101 if (invlist == NULL) {
10102 invlist = _new_invlist(2);
10103 _append_range_to_invlist(invlist, start, end);
10107 /* Likewise, if the inversion list is currently empty */
10108 len = _invlist_len(invlist);
10110 _append_range_to_invlist(invlist, start, end);
10114 /* Starting here, we have to know the internals of the list */
10115 array = invlist_array(invlist);
10117 /* If the new range ends higher than the current highest ... */
10118 cur_highest = invlist_highest(invlist);
10119 if (end > cur_highest) {
10121 /* If the whole range is higher, we can just append it */
10122 if (start > cur_highest) {
10123 _append_range_to_invlist(invlist, start, end);
10127 /* Otherwise, add the portion that is higher ... */
10128 _append_range_to_invlist(invlist, cur_highest + 1, end);
10130 /* ... and continue on below to handle the rest. As a result of the
10131 * above append, we know that the index of the end of the range is the
10132 * final even numbered one of the array. Recall that the final element
10133 * always starts a range that extends to infinity. If that range is in
10134 * the set (meaning the set goes from here to infinity), it will be an
10135 * even index, but if it isn't in the set, it's odd, and the final
10136 * range in the set is one less, which is even. */
10137 if (end == UV_MAX) {
10145 /* We have dealt with appending, now see about prepending. If the new
10146 * range starts lower than the current lowest ... */
10147 if (start < array[0]) {
10149 /* Adding something which has 0 in it is somewhat tricky, and uncommon.
10150 * Let the union code handle it, rather than having to know the
10151 * trickiness in two code places. */
10152 if (UNLIKELY(start == 0)) {
10155 range_invlist = _new_invlist(2);
10156 _append_range_to_invlist(range_invlist, start, end);
10158 _invlist_union(invlist, range_invlist, &invlist);
10160 SvREFCNT_dec_NN(range_invlist);
10165 /* If the whole new range comes before the first entry, and doesn't
10166 * extend it, we have to insert it as an additional range */
10167 if (end < array[0] - 1) {
10169 goto splice_in_new_range;
10172 /* Here the new range adjoins the existing first range, extending it
10176 /* And continue on below to handle the rest. We know that the index of
10177 * the beginning of the range is the first one of the array */
10180 else { /* Not prepending any part of the new range to the existing list.
10181 * Find where in the list it should go. This finds i_s, such that:
10182 * invlist[i_s] <= start < array[i_s+1]
10184 i_s = _invlist_search(invlist, start);
10187 /* At this point, any extending before the beginning of the inversion list
10188 * and/or after the end has been done. This has made it so that, in the
10189 * code below, each endpoint of the new range is either in a range that is
10190 * in the set, or is in a gap between two ranges that are. This means we
10191 * don't have to worry about exceeding the array bounds.
10193 * Find where in the list the new range ends (but we can skip this if we
10194 * have already determined what it is, or if it will be the same as i_s,
10195 * which we already have computed) */
10197 i_e = (start == end)
10199 : _invlist_search(invlist, end);
10202 /* Here generally invlist[i_e] <= end < array[i_e+1]. But if invlist[i_e]
10203 * is a range that goes to infinity there is no element at invlist[i_e+1],
10204 * so only the first relation holds. */
10206 if ( ! ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
10208 /* Here, the ranges on either side of the beginning of the new range
10209 * are in the set, and this range starts in the gap between them.
10211 * The new range extends the range above it downwards if the new range
10212 * ends at or above that range's start */
10213 const bool extends_the_range_above = ( end == UV_MAX
10214 || end + 1 >= array[i_s+1]);
10216 /* The new range extends the range below it upwards if it begins just
10217 * after where that range ends */
10218 if (start == array[i_s]) {
10220 /* If the new range fills the entire gap between the other ranges,
10221 * they will get merged together. Other ranges may also get
10222 * merged, depending on how many of them the new range spans. In
10223 * the general case, we do the merge later, just once, after we
10224 * figure out how many to merge. But in the case where the new
10225 * range exactly spans just this one gap (possibly extending into
10226 * the one above), we do the merge here, and an early exit. This
10227 * is done here to avoid having to special case later. */
10228 if (i_e - i_s <= 1) {
10230 /* If i_e - i_s == 1, it means that the new range terminates
10231 * within the range above, and hence 'extends_the_range_above'
10232 * must be true. (If the range above it extends to infinity,
10233 * 'i_s+2' will be above the array's limit, but 'len-i_s-2'
10234 * will be 0, so no harm done.) */
10235 if (extends_the_range_above) {
10236 Move(array + i_s + 2, array + i_s, len - i_s - 2, UV);
10237 invlist_set_len(invlist,
10239 *(get_invlist_offset_addr(invlist)));
10243 /* Here, i_e must == i_s. We keep them in sync, as they apply
10244 * to the same range, and below we are about to decrement i_s
10249 /* Here, the new range is adjacent to the one below. (It may also
10250 * span beyond the range above, but that will get resolved later.)
10251 * Extend the range below to include this one. */
10252 array[i_s] = (end == UV_MAX) ? UV_MAX : end + 1;
10254 start = array[i_s];
10256 else if (extends_the_range_above) {
10258 /* Here the new range only extends the range above it, but not the
10259 * one below. It merges with the one above. Again, we keep i_e
10260 * and i_s in sync if they point to the same range */
10265 array[i_s] = start;
10269 /* Here, we've dealt with the new range start extending any adjoining
10272 * If the new range extends to infinity, it is now the final one,
10273 * regardless of what was there before */
10274 if (UNLIKELY(end == UV_MAX)) {
10275 invlist_set_len(invlist, i_s + 1, *(get_invlist_offset_addr(invlist)));
10279 /* If i_e started as == i_s, it has also been dealt with,
10280 * and been updated to the new i_s, which will fail the following if */
10281 if (! ELEMENT_RANGE_MATCHES_INVLIST(i_e)) {
10283 /* Here, the ranges on either side of the end of the new range are in
10284 * the set, and this range ends in the gap between them.
10286 * If this range is adjacent to (hence extends) the range above it, it
10287 * becomes part of that range; likewise if it extends the range below,
10288 * it becomes part of that range */
10289 if (end + 1 == array[i_e+1]) {
10291 array[i_e] = start;
10293 else if (start <= array[i_e]) {
10294 array[i_e] = end + 1;
10301 /* If the range fits entirely in an existing range (as possibly already
10302 * extended above), it doesn't add anything new */
10303 if (ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
10307 /* Here, no part of the range is in the list. Must add it. It will
10308 * occupy 2 more slots */
10309 splice_in_new_range:
10311 invlist_extend(invlist, len + 2);
10312 array = invlist_array(invlist);
10313 /* Move the rest of the array down two slots. Don't include any
10315 Move(array + i_e + 1, array + i_e + 3, len - i_e - 1, UV);
10317 /* Do the actual splice */
10318 array[i_e+1] = start;
10319 array[i_e+2] = end + 1;
10320 invlist_set_len(invlist, len + 2, *(get_invlist_offset_addr(invlist)));
10324 /* Here the new range crossed the boundaries of a pre-existing range. The
10325 * code above has adjusted things so that both ends are in ranges that are
10326 * in the set. This means everything in between must also be in the set.
10327 * Just squash things together */
10328 Move(array + i_e + 1, array + i_s + 1, len - i_e - 1, UV);
10329 invlist_set_len(invlist,
10331 *(get_invlist_offset_addr(invlist)));
10337 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
10338 UV** other_elements_ptr)
10340 /* Create and return an inversion list whose contents are to be populated
10341 * by the caller. The caller gives the number of elements (in 'size') and
10342 * the very first element ('element0'). This function will set
10343 * '*other_elements_ptr' to an array of UVs, where the remaining elements
10344 * are to be placed.
10346 * Obviously there is some trust involved that the caller will properly
10347 * fill in the other elements of the array.
10349 * (The first element needs to be passed in, as the underlying code does
10350 * things differently depending on whether it is zero or non-zero) */
10352 SV* invlist = _new_invlist(size);
10355 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
10357 invlist = add_cp_to_invlist(invlist, element0);
10358 offset = *get_invlist_offset_addr(invlist);
10360 invlist_set_len(invlist, size, offset);
10361 *other_elements_ptr = invlist_array(invlist) + 1;
10367 #ifndef PERL_IN_XSUB_RE
10369 Perl__invlist_invert(pTHX_ SV* const invlist)
10371 /* Complement the input inversion list. This adds a 0 if the list didn't
10372 * have a zero; removes it otherwise. As described above, the data
10373 * structure is set up so that this is very efficient */
10375 PERL_ARGS_ASSERT__INVLIST_INVERT;
10377 assert(! invlist_is_iterating(invlist));
10379 /* The inverse of matching nothing is matching everything */
10380 if (_invlist_len(invlist) == 0) {
10381 _append_range_to_invlist(invlist, 0, UV_MAX);
10385 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
10389 Perl_invlist_clone(pTHX_ SV* const invlist, SV* new_invlist)
10391 /* Return a new inversion list that is a copy of the input one, which is
10392 * unchanged. The new list will not be mortal even if the old one was. */
10394 const STRLEN nominal_length = _invlist_len(invlist);
10395 const STRLEN physical_length = SvCUR(invlist);
10396 const bool offset = *(get_invlist_offset_addr(invlist));
10398 PERL_ARGS_ASSERT_INVLIST_CLONE;
10400 if (new_invlist == NULL) {
10401 new_invlist = _new_invlist(nominal_length);
10404 sv_upgrade(new_invlist, SVt_INVLIST);
10405 initialize_invlist_guts(new_invlist, nominal_length);
10408 *(get_invlist_offset_addr(new_invlist)) = offset;
10409 invlist_set_len(new_invlist, nominal_length, offset);
10410 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
10412 return new_invlist;
10417 PERL_STATIC_INLINE UV
10418 S_invlist_lowest(SV* const invlist)
10420 /* Returns the lowest code point that matches an inversion list. This API
10421 * has an ambiguity, as it returns 0 under either the lowest is actually
10422 * 0, or if the list is empty. If this distinction matters to you, check
10423 * for emptiness before calling this function */
10425 UV len = _invlist_len(invlist);
10428 PERL_ARGS_ASSERT_INVLIST_LOWEST;
10434 array = invlist_array(invlist);
10440 S_invlist_contents(pTHX_ SV* const invlist, const bool traditional_style)
10442 /* Get the contents of an inversion list into a string SV so that they can
10443 * be printed out. If 'traditional_style' is TRUE, it uses the format
10444 * traditionally done for debug tracing; otherwise it uses a format
10445 * suitable for just copying to the output, with blanks between ranges and
10446 * a dash between range components */
10450 const char intra_range_delimiter = (traditional_style ? '\t' : '-');
10451 const char inter_range_delimiter = (traditional_style ? '\n' : ' ');
10453 if (traditional_style) {
10454 output = newSVpvs("\n");
10457 output = newSVpvs("");
10460 PERL_ARGS_ASSERT_INVLIST_CONTENTS;
10462 assert(! invlist_is_iterating(invlist));
10464 invlist_iterinit(invlist);
10465 while (invlist_iternext(invlist, &start, &end)) {
10466 if (end == UV_MAX) {
10467 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%cINFTY%c",
10468 start, intra_range_delimiter,
10469 inter_range_delimiter);
10471 else if (end != start) {
10472 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c%04" UVXf "%c",
10474 intra_range_delimiter,
10475 end, inter_range_delimiter);
10478 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c",
10479 start, inter_range_delimiter);
10483 if (SvCUR(output) && ! traditional_style) {/* Get rid of trailing blank */
10484 SvCUR_set(output, SvCUR(output) - 1);
10490 #ifndef PERL_IN_XSUB_RE
10492 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
10493 const char * const indent, SV* const invlist)
10495 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
10496 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
10497 * the string 'indent'. The output looks like this:
10498 [0] 0x000A .. 0x000D
10500 [4] 0x2028 .. 0x2029
10501 [6] 0x3104 .. INFTY
10502 * This means that the first range of code points matched by the list are
10503 * 0xA through 0xD; the second range contains only the single code point
10504 * 0x85, etc. An inversion list is an array of UVs. Two array elements
10505 * are used to define each range (except if the final range extends to
10506 * infinity, only a single element is needed). The array index of the
10507 * first element for the corresponding range is given in brackets. */
10512 PERL_ARGS_ASSERT__INVLIST_DUMP;
10514 if (invlist_is_iterating(invlist)) {
10515 Perl_dump_indent(aTHX_ level, file,
10516 "%sCan't dump inversion list because is in middle of iterating\n",
10521 invlist_iterinit(invlist);
10522 while (invlist_iternext(invlist, &start, &end)) {
10523 if (end == UV_MAX) {
10524 Perl_dump_indent(aTHX_ level, file,
10525 "%s[%" UVuf "] 0x%04" UVXf " .. INFTY\n",
10526 indent, (UV)count, start);
10528 else if (end != start) {
10529 Perl_dump_indent(aTHX_ level, file,
10530 "%s[%" UVuf "] 0x%04" UVXf " .. 0x%04" UVXf "\n",
10531 indent, (UV)count, start, end);
10534 Perl_dump_indent(aTHX_ level, file, "%s[%" UVuf "] 0x%04" UVXf "\n",
10535 indent, (UV)count, start);
10543 #if defined(PERL_ARGS_ASSERT__INVLISTEQ) && !defined(PERL_IN_XSUB_RE)
10545 Perl__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
10547 /* Return a boolean as to if the two passed in inversion lists are
10548 * identical. The final argument, if TRUE, says to take the complement of
10549 * the second inversion list before doing the comparison */
10551 const UV len_a = _invlist_len(a);
10552 UV len_b = _invlist_len(b);
10554 const UV* array_a = NULL;
10555 const UV* array_b = NULL;
10557 PERL_ARGS_ASSERT__INVLISTEQ;
10559 /* This code avoids accessing the arrays unless it knows the length is
10564 return ! complement_b;
10568 array_a = invlist_array(a);
10572 array_b = invlist_array(b);
10575 /* If are to compare 'a' with the complement of b, set it
10576 * up so are looking at b's complement. */
10577 if (complement_b) {
10579 /* The complement of nothing is everything, so <a> would have to have
10580 * just one element, starting at zero (ending at infinity) */
10582 return (len_a == 1 && array_a[0] == 0);
10584 if (array_b[0] == 0) {
10586 /* Otherwise, to complement, we invert. Here, the first element is
10587 * 0, just remove it. To do this, we just pretend the array starts
10595 /* But if the first element is not zero, we pretend the list starts
10596 * at the 0 that is always stored immediately before the array. */
10602 return len_a == len_b
10603 && memEQ(array_a, array_b, len_a * sizeof(array_a[0]));
10609 * As best we can, determine the characters that can match the start of
10610 * the given EXACTF-ish node. This is for use in creating ssc nodes, so there
10611 * can be false positive matches
10613 * Returns the invlist as a new SV*; it is the caller's responsibility to
10614 * call SvREFCNT_dec() when done with it.
10617 S_make_exactf_invlist(pTHX_ RExC_state_t *pRExC_state, regnode *node)
10619 const U8 * s = (U8*)STRING(node);
10620 SSize_t bytelen = STR_LEN(node);
10622 /* Start out big enough for 2 separate code points */
10623 SV* invlist = _new_invlist(4);
10625 PERL_ARGS_ASSERT_MAKE_EXACTF_INVLIST;
10630 /* We punt and assume can match anything if the node begins
10631 * with a multi-character fold. Things are complicated. For
10632 * example, /ffi/i could match any of:
10633 * "\N{LATIN SMALL LIGATURE FFI}"
10634 * "\N{LATIN SMALL LIGATURE FF}I"
10635 * "F\N{LATIN SMALL LIGATURE FI}"
10636 * plus several other things; and making sure we have all the
10637 * possibilities is hard. */
10638 if (is_MULTI_CHAR_FOLD_latin1_safe(s, s + bytelen)) {
10639 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10642 /* Any Latin1 range character can potentially match any
10643 * other depending on the locale, and in Turkic locales, U+130 and
10645 if (OP(node) == EXACTFL) {
10646 _invlist_union(invlist, PL_Latin1, &invlist);
10647 invlist = add_cp_to_invlist(invlist,
10648 LATIN_SMALL_LETTER_DOTLESS_I);
10649 invlist = add_cp_to_invlist(invlist,
10650 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
10653 /* But otherwise, it matches at least itself. We can
10654 * quickly tell if it has a distinct fold, and if so,
10655 * it matches that as well */
10656 invlist = add_cp_to_invlist(invlist, uc);
10657 if (IS_IN_SOME_FOLD_L1(uc))
10658 invlist = add_cp_to_invlist(invlist, PL_fold_latin1[uc]);
10661 /* Some characters match above-Latin1 ones under /i. This
10662 * is true of EXACTFL ones when the locale is UTF-8 */
10663 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(uc)
10664 && (! isASCII(uc) || ! inRANGE(OP(node), EXACTFAA,
10665 EXACTFAA_NO_TRIE)))
10667 add_above_Latin1_folds(pRExC_state, (U8) uc, &invlist);
10671 else { /* Pattern is UTF-8 */
10672 U8 folded[UTF8_MAX_FOLD_CHAR_EXPAND * UTF8_MAXBYTES_CASE + 1] = { '\0' };
10673 const U8* e = s + bytelen;
10676 fc = uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
10678 /* The only code points that aren't folded in a UTF EXACTFish
10679 * node are the problematic ones in EXACTFL nodes */
10680 if (OP(node) == EXACTFL && is_PROBLEMATIC_LOCALE_FOLDEDS_START_cp(uc)) {
10681 /* We need to check for the possibility that this EXACTFL
10682 * node begins with a multi-char fold. Therefore we fold
10683 * the first few characters of it so that we can make that
10689 for (i = 0; i < UTF8_MAX_FOLD_CHAR_EXPAND && s < e; i++) {
10691 *(d++) = (U8) toFOLD(*s);
10692 if (fc < 0) { /* Save the first fold */
10699 UV fold = toFOLD_utf8_safe(s, e, d, &len);
10700 if (fc < 0) { /* Save the first fold */
10708 /* And set up so the code below that looks in this folded
10709 * buffer instead of the node's string */
10714 /* When we reach here 's' points to the fold of the first
10715 * character(s) of the node; and 'e' points to far enough along
10716 * the folded string to be just past any possible multi-char
10719 * Like the non-UTF case above, we punt if the node begins with a
10720 * multi-char fold */
10722 if (is_MULTI_CHAR_FOLD_utf8_safe(s, e)) {
10723 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10725 else { /* Single char fold */
10728 const U32 * remaining_folds;
10729 Size_t folds_count;
10731 /* It matches itself */
10732 invlist = add_cp_to_invlist(invlist, fc);
10734 /* ... plus all the things that fold to it, which are found in
10735 * PL_utf8_foldclosures */
10736 folds_count = _inverse_folds(fc, &first_fold,
10738 for (k = 0; k < folds_count; k++) {
10739 UV c = (k == 0) ? first_fold : remaining_folds[k-1];
10741 /* /aa doesn't allow folds between ASCII and non- */
10742 if ( inRANGE(OP(node), EXACTFAA, EXACTFAA_NO_TRIE)
10743 && isASCII(c) != isASCII(fc))
10748 invlist = add_cp_to_invlist(invlist, c);
10751 if (OP(node) == EXACTFL) {
10753 /* If either [iI] are present in an EXACTFL node the above code
10754 * should have added its normal case pair, but under a Turkish
10755 * locale they could match instead the case pairs from it. Add
10756 * those as potential matches as well */
10757 if (isALPHA_FOLD_EQ(fc, 'I')) {
10758 invlist = add_cp_to_invlist(invlist,
10759 LATIN_SMALL_LETTER_DOTLESS_I);
10760 invlist = add_cp_to_invlist(invlist,
10761 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
10763 else if (fc == LATIN_SMALL_LETTER_DOTLESS_I) {
10764 invlist = add_cp_to_invlist(invlist, 'I');
10766 else if (fc == LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE) {
10767 invlist = add_cp_to_invlist(invlist, 'i');
10776 #undef HEADER_LENGTH
10777 #undef TO_INTERNAL_SIZE
10778 #undef FROM_INTERNAL_SIZE
10779 #undef INVLIST_VERSION_ID
10781 /* End of inversion list object */
10784 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
10786 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
10787 * constructs, and updates RExC_flags with them. On input, RExC_parse
10788 * should point to the first flag; it is updated on output to point to the
10789 * final ')' or ':'. There needs to be at least one flag, or this will
10792 /* for (?g), (?gc), and (?o) warnings; warning
10793 about (?c) will warn about (?g) -- japhy */
10795 #define WASTED_O 0x01
10796 #define WASTED_G 0x02
10797 #define WASTED_C 0x04
10798 #define WASTED_GC (WASTED_G|WASTED_C)
10799 I32 wastedflags = 0x00;
10800 U32 posflags = 0, negflags = 0;
10801 U32 *flagsp = &posflags;
10802 char has_charset_modifier = '\0';
10804 bool has_use_defaults = FALSE;
10805 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
10806 int x_mod_count = 0;
10808 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
10810 /* '^' as an initial flag sets certain defaults */
10811 if (UCHARAT(RExC_parse) == '^') {
10813 has_use_defaults = TRUE;
10814 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
10815 cs = (toUSE_UNI_CHARSET_NOT_DEPENDS)
10816 ? REGEX_UNICODE_CHARSET
10817 : REGEX_DEPENDS_CHARSET;
10818 set_regex_charset(&RExC_flags, cs);
10821 cs = get_regex_charset(RExC_flags);
10822 if ( cs == REGEX_DEPENDS_CHARSET
10823 && (toUSE_UNI_CHARSET_NOT_DEPENDS))
10825 cs = REGEX_UNICODE_CHARSET;
10829 while (RExC_parse < RExC_end) {
10830 /* && memCHRs("iogcmsx", *RExC_parse) */
10831 /* (?g), (?gc) and (?o) are useless here
10832 and must be globally applied -- japhy */
10833 if ((RExC_pm_flags & PMf_WILDCARD)) {
10834 if (flagsp == & negflags) {
10835 if (*RExC_parse == 'm') {
10837 /* diag_listed_as: Use of %s is not allowed in Unicode
10838 property wildcard subpatterns in regex; marked by <--
10840 vFAIL("Use of modifier '-m' is not allowed in Unicode"
10841 " property wildcard subpatterns");
10845 if (*RExC_parse == 's') {
10846 goto modifier_illegal_in_wildcard;
10851 switch (*RExC_parse) {
10853 /* Code for the imsxn flags */
10854 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp, x_mod_count);
10856 case LOCALE_PAT_MOD:
10857 if (has_charset_modifier) {
10858 goto excess_modifier;
10860 else if (flagsp == &negflags) {
10863 cs = REGEX_LOCALE_CHARSET;
10864 has_charset_modifier = LOCALE_PAT_MOD;
10866 case UNICODE_PAT_MOD:
10867 if (has_charset_modifier) {
10868 goto excess_modifier;
10870 else if (flagsp == &negflags) {
10873 cs = REGEX_UNICODE_CHARSET;
10874 has_charset_modifier = UNICODE_PAT_MOD;
10876 case ASCII_RESTRICT_PAT_MOD:
10877 if (flagsp == &negflags) {
10880 if (has_charset_modifier) {
10881 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
10882 goto excess_modifier;
10884 /* Doubled modifier implies more restricted */
10885 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
10888 cs = REGEX_ASCII_RESTRICTED_CHARSET;
10890 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
10892 case DEPENDS_PAT_MOD:
10893 if (has_use_defaults) {
10894 goto fail_modifiers;
10896 else if (flagsp == &negflags) {
10899 else if (has_charset_modifier) {
10900 goto excess_modifier;
10903 /* The dual charset means unicode semantics if the
10904 * pattern (or target, not known until runtime) are
10905 * utf8, or something in the pattern indicates unicode
10907 cs = (toUSE_UNI_CHARSET_NOT_DEPENDS)
10908 ? REGEX_UNICODE_CHARSET
10909 : REGEX_DEPENDS_CHARSET;
10910 has_charset_modifier = DEPENDS_PAT_MOD;
10914 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
10915 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
10917 else if (has_charset_modifier == *(RExC_parse - 1)) {
10918 vFAIL2("Regexp modifier \"%c\" may not appear twice",
10919 *(RExC_parse - 1));
10922 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
10924 NOT_REACHED; /*NOTREACHED*/
10927 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
10928 *(RExC_parse - 1));
10929 NOT_REACHED; /*NOTREACHED*/
10930 case GLOBAL_PAT_MOD: /* 'g' */
10931 if (RExC_pm_flags & PMf_WILDCARD) {
10932 goto modifier_illegal_in_wildcard;
10935 case ONCE_PAT_MOD: /* 'o' */
10936 if (ckWARN(WARN_REGEXP)) {
10937 const I32 wflagbit = *RExC_parse == 'o'
10940 if (! (wastedflags & wflagbit) ) {
10941 wastedflags |= wflagbit;
10942 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10945 "Useless (%s%c) - %suse /%c modifier",
10946 flagsp == &negflags ? "?-" : "?",
10948 flagsp == &negflags ? "don't " : "",
10955 case CONTINUE_PAT_MOD: /* 'c' */
10956 if (RExC_pm_flags & PMf_WILDCARD) {
10957 goto modifier_illegal_in_wildcard;
10959 if (ckWARN(WARN_REGEXP)) {
10960 if (! (wastedflags & WASTED_C) ) {
10961 wastedflags |= WASTED_GC;
10962 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10965 "Useless (%sc) - %suse /gc modifier",
10966 flagsp == &negflags ? "?-" : "?",
10967 flagsp == &negflags ? "don't " : ""
10972 case KEEPCOPY_PAT_MOD: /* 'p' */
10973 if (RExC_pm_flags & PMf_WILDCARD) {
10974 goto modifier_illegal_in_wildcard;
10976 if (flagsp == &negflags) {
10977 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
10979 *flagsp |= RXf_PMf_KEEPCOPY;
10983 /* A flag is a default iff it is following a minus, so
10984 * if there is a minus, it means will be trying to
10985 * re-specify a default which is an error */
10986 if (has_use_defaults || flagsp == &negflags) {
10987 goto fail_modifiers;
10989 flagsp = &negflags;
10990 wastedflags = 0; /* reset so (?g-c) warns twice */
10996 if ( (RExC_pm_flags & PMf_WILDCARD)
10997 && cs != REGEX_ASCII_MORE_RESTRICTED_CHARSET)
11000 /* diag_listed_as: Use of %s is not allowed in Unicode
11001 property wildcard subpatterns in regex; marked by <--
11003 vFAIL2("Use of modifier '%c' is not allowed in Unicode"
11004 " property wildcard subpatterns",
11005 has_charset_modifier);
11008 if ((posflags & (RXf_PMf_EXTENDED|RXf_PMf_EXTENDED_MORE)) == RXf_PMf_EXTENDED) {
11009 negflags |= RXf_PMf_EXTENDED_MORE;
11011 RExC_flags |= posflags;
11013 if (negflags & RXf_PMf_EXTENDED) {
11014 negflags |= RXf_PMf_EXTENDED_MORE;
11016 RExC_flags &= ~negflags;
11017 set_regex_charset(&RExC_flags, cs);
11022 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
11023 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11024 vFAIL2utf8f("Sequence (%" UTF8f "...) not recognized",
11025 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
11026 NOT_REACHED; /*NOTREACHED*/
11029 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11032 vFAIL("Sequence (?... not terminated");
11034 modifier_illegal_in_wildcard:
11036 /* diag_listed_as: Use of %s is not allowed in Unicode property wildcard
11037 subpatterns in regex; marked by <-- HERE in m/%s/ */
11038 vFAIL2("Use of modifier '%c' is not allowed in Unicode property wildcard"
11039 " subpatterns", *(RExC_parse - 1));
11043 - reg - regular expression, i.e. main body or parenthesized thing
11045 * Caller must absorb opening parenthesis.
11047 * Combining parenthesis handling with the base level of regular expression
11048 * is a trifle forced, but the need to tie the tails of the branches to what
11049 * follows makes it hard to avoid.
11051 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
11053 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
11055 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
11058 STATIC regnode_offset
11059 S_handle_named_backref(pTHX_ RExC_state_t *pRExC_state,
11061 char * parse_start,
11065 regnode_offset ret;
11066 char* name_start = RExC_parse;
11068 SV *sv_dat = reg_scan_name(pRExC_state, REG_RSN_RETURN_DATA);
11069 DECLARE_AND_GET_RE_DEBUG_FLAGS;
11071 PERL_ARGS_ASSERT_HANDLE_NAMED_BACKREF;
11073 if (RExC_parse == name_start || *RExC_parse != ch) {
11074 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
11075 vFAIL2("Sequence %.3s... not terminated", parse_start);
11079 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11080 RExC_rxi->data->data[num]=(void*)sv_dat;
11081 SvREFCNT_inc_simple_void_NN(sv_dat);
11084 ret = reganode(pRExC_state,
11087 : (ASCII_FOLD_RESTRICTED)
11089 : (AT_LEAST_UNI_SEMANTICS)
11095 *flagp |= HASWIDTH;
11097 Set_Node_Offset(REGNODE_p(ret), parse_start+1);
11098 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
11100 nextchar(pRExC_state);
11104 /* On success, returns the offset at which any next node should be placed into
11105 * the regex engine program being compiled.
11107 * Returns 0 otherwise, with *flagp set to indicate why:
11108 * TRYAGAIN at the end of (?) that only sets flags.
11109 * RESTART_PARSE if the parse needs to be restarted, or'd with
11110 * NEED_UTF8 if the pattern needs to be upgraded to UTF-8.
11111 * Otherwise would only return 0 if regbranch() returns 0, which cannot
11113 STATIC regnode_offset
11114 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp, U32 depth)
11115 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
11116 * 2 is like 1, but indicates that nextchar() has been called to advance
11117 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
11118 * this flag alerts us to the need to check for that */
11120 regnode_offset ret = 0; /* Will be the head of the group. */
11122 regnode_offset lastbr;
11123 regnode_offset ender = 0;
11126 U32 oregflags = RExC_flags;
11127 bool have_branch = 0;
11129 I32 freeze_paren = 0;
11130 I32 after_freeze = 0;
11131 I32 num; /* numeric backreferences */
11132 SV * max_open; /* Max number of unclosed parens */
11133 I32 was_in_lookaround = RExC_in_lookaround;
11135 char * parse_start = RExC_parse; /* MJD */
11136 char * const oregcomp_parse = RExC_parse;
11138 DECLARE_AND_GET_RE_DEBUG_FLAGS;
11140 PERL_ARGS_ASSERT_REG;
11141 DEBUG_PARSE("reg ");
11143 max_open = get_sv(RE_COMPILE_RECURSION_LIMIT, GV_ADD);
11145 if (!SvIOK(max_open)) {
11146 sv_setiv(max_open, RE_COMPILE_RECURSION_INIT);
11148 if (depth > 4 * (UV) SvIV(max_open)) { /* We increase depth by 4 for each
11150 vFAIL("Too many nested open parens");
11153 *flagp = 0; /* Initialize. */
11155 /* Having this true makes it feasible to have a lot fewer tests for the
11156 * parse pointer being in scope. For example, we can write
11157 * while(isFOO(*RExC_parse)) RExC_parse++;
11159 * while(RExC_parse < RExC_end && isFOO(*RExC_parse)) RExC_parse++;
11161 assert(*RExC_end == '\0');
11163 /* Make an OPEN node, if parenthesized. */
11166 /* Under /x, space and comments can be gobbled up between the '(' and
11167 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
11168 * intervening space, as the sequence is a token, and a token should be
11170 bool has_intervening_patws = (paren == 2)
11171 && *(RExC_parse - 1) != '(';
11173 if (RExC_parse >= RExC_end) {
11174 vFAIL("Unmatched (");
11177 if (paren == 'r') { /* Atomic script run */
11181 else if ( *RExC_parse == '*') { /* (*VERB:ARG), (*construct:...) */
11182 char *start_verb = RExC_parse + 1;
11184 char *start_arg = NULL;
11185 unsigned char op = 0;
11186 int arg_required = 0;
11187 int internal_argval = -1; /* if >-1 we are not allowed an argument*/
11188 bool has_upper = FALSE;
11190 if (has_intervening_patws) {
11191 RExC_parse++; /* past the '*' */
11193 /* For strict backwards compatibility, don't change the message
11194 * now that we also have lowercase operands */
11195 if (isUPPER(*RExC_parse)) {
11196 vFAIL("In '(*VERB...)', the '(' and '*' must be adjacent");
11199 vFAIL("In '(*...)', the '(' and '*' must be adjacent");
11202 while (RExC_parse < RExC_end && *RExC_parse != ')' ) {
11203 if ( *RExC_parse == ':' ) {
11204 start_arg = RExC_parse + 1;
11208 if (isUPPER(*RExC_parse)) {
11214 RExC_parse += UTF8SKIP(RExC_parse);
11217 verb_len = RExC_parse - start_verb;
11219 if (RExC_parse >= RExC_end) {
11220 goto unterminated_verb_pattern;
11223 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11224 while ( RExC_parse < RExC_end && *RExC_parse != ')' ) {
11225 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11227 if ( RExC_parse >= RExC_end || *RExC_parse != ')' ) {
11228 unterminated_verb_pattern:
11230 vFAIL("Unterminated verb pattern argument");
11233 vFAIL("Unterminated '(*...' argument");
11237 if ( RExC_parse >= RExC_end || *RExC_parse != ')' ) {
11239 vFAIL("Unterminated verb pattern");
11242 vFAIL("Unterminated '(*...' construct");
11247 /* Here, we know that RExC_parse < RExC_end */
11249 switch ( *start_verb ) {
11250 case 'A': /* (*ACCEPT) */
11251 if ( memEQs(start_verb, verb_len,"ACCEPT") ) {
11253 internal_argval = RExC_nestroot;
11256 case 'C': /* (*COMMIT) */
11257 if ( memEQs(start_verb, verb_len,"COMMIT") )
11260 case 'F': /* (*FAIL) */
11261 if ( verb_len==1 || memEQs(start_verb, verb_len,"FAIL") ) {
11265 case ':': /* (*:NAME) */
11266 case 'M': /* (*MARK:NAME) */
11267 if ( verb_len==0 || memEQs(start_verb, verb_len,"MARK") ) {
11272 case 'P': /* (*PRUNE) */
11273 if ( memEQs(start_verb, verb_len,"PRUNE") )
11276 case 'S': /* (*SKIP) */
11277 if ( memEQs(start_verb, verb_len,"SKIP") )
11280 case 'T': /* (*THEN) */
11281 /* [19:06] <TimToady> :: is then */
11282 if ( memEQs(start_verb, verb_len,"THEN") ) {
11284 RExC_seen |= REG_CUTGROUP_SEEN;
11288 if ( memEQs(start_verb, verb_len, "asr")
11289 || memEQs(start_verb, verb_len, "atomic_script_run"))
11291 paren = 'r'; /* Mnemonic: recursed run */
11294 else if (memEQs(start_verb, verb_len, "atomic")) {
11295 paren = 't'; /* AtOMIC */
11296 goto alpha_assertions;
11300 if ( memEQs(start_verb, verb_len, "plb")
11301 || memEQs(start_verb, verb_len, "positive_lookbehind"))
11304 goto lookbehind_alpha_assertions;
11306 else if ( memEQs(start_verb, verb_len, "pla")
11307 || memEQs(start_verb, verb_len, "positive_lookahead"))
11310 goto alpha_assertions;
11314 if ( memEQs(start_verb, verb_len, "nlb")
11315 || memEQs(start_verb, verb_len, "negative_lookbehind"))
11318 goto lookbehind_alpha_assertions;
11320 else if ( memEQs(start_verb, verb_len, "nla")
11321 || memEQs(start_verb, verb_len, "negative_lookahead"))
11324 goto alpha_assertions;
11328 if ( memEQs(start_verb, verb_len, "sr")
11329 || memEQs(start_verb, verb_len, "script_run"))
11331 regnode_offset atomic;
11337 /* This indicates Unicode rules. */
11338 REQUIRE_UNI_RULES(flagp, 0);
11344 RExC_parse = start_arg;
11346 if (RExC_in_script_run) {
11348 /* Nested script runs are treated as no-ops, because
11349 * if the nested one fails, the outer one must as
11350 * well. It could fail sooner, and avoid (??{} with
11351 * side effects, but that is explicitly documented as
11352 * undefined behavior. */
11356 if (paren == 's') {
11361 /* But, the atomic part of a nested atomic script run
11362 * isn't a no-op, but can be treated just like a '(?>'
11368 if (paren == 's') {
11369 /* Here, we're starting a new regular script run */
11370 ret = reg_node(pRExC_state, SROPEN);
11371 RExC_in_script_run = 1;
11376 /* Here, we are starting an atomic script run. This is
11377 * handled by recursing to deal with the atomic portion
11378 * separately, enclosed in SROPEN ... SRCLOSE nodes */
11380 ret = reg_node(pRExC_state, SROPEN);
11382 RExC_in_script_run = 1;
11384 atomic = reg(pRExC_state, 'r', &flags, depth);
11385 if (flags & (RESTART_PARSE|NEED_UTF8)) {
11386 *flagp = flags & (RESTART_PARSE|NEED_UTF8);
11390 if (! REGTAIL(pRExC_state, ret, atomic)) {
11391 REQUIRE_BRANCHJ(flagp, 0);
11394 if (! REGTAIL(pRExC_state, atomic, reg_node(pRExC_state,
11397 REQUIRE_BRANCHJ(flagp, 0);
11400 RExC_in_script_run = 0;
11406 lookbehind_alpha_assertions:
11407 RExC_seen |= REG_LOOKBEHIND_SEEN;
11412 RExC_in_lookaround++;
11413 RExC_seen_zerolen++;
11419 /* An empty negative lookahead assertion simply is failure */
11420 if (paren == 'A' && RExC_parse == start_arg) {
11421 ret=reganode(pRExC_state, OPFAIL, 0);
11422 nextchar(pRExC_state);
11426 RExC_parse = start_arg;
11431 "'(*%" UTF8f "' requires a terminating ':'",
11432 UTF8fARG(UTF, verb_len, start_verb));
11433 NOT_REACHED; /*NOTREACHED*/
11435 } /* End of switch */
11438 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
11440 if (has_upper || verb_len == 0) {
11442 "Unknown verb pattern '%" UTF8f "'",
11443 UTF8fARG(UTF, verb_len, start_verb));
11447 "Unknown '(*...)' construct '%" UTF8f "'",
11448 UTF8fARG(UTF, verb_len, start_verb));
11451 if ( RExC_parse == start_arg ) {
11454 if ( arg_required && !start_arg ) {
11455 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
11456 (int) verb_len, start_verb);
11458 if (internal_argval == -1) {
11459 ret = reganode(pRExC_state, op, 0);
11461 ret = reg2Lanode(pRExC_state, op, 0, internal_argval);
11463 RExC_seen |= REG_VERBARG_SEEN;
11465 SV *sv = newSVpvn( start_arg,
11466 RExC_parse - start_arg);
11467 ARG(REGNODE_p(ret)) = add_data( pRExC_state,
11468 STR_WITH_LEN("S"));
11469 RExC_rxi->data->data[ARG(REGNODE_p(ret))]=(void*)sv;
11470 FLAGS(REGNODE_p(ret)) = 1;
11472 FLAGS(REGNODE_p(ret)) = 0;
11474 if ( internal_argval != -1 )
11475 ARG2L_SET(REGNODE_p(ret), internal_argval);
11476 nextchar(pRExC_state);
11479 else if (*RExC_parse == '?') { /* (?...) */
11480 bool is_logical = 0;
11481 const char * const seqstart = RExC_parse;
11482 const char * endptr;
11483 const char non_existent_group_msg[]
11484 = "Reference to nonexistent group";
11485 const char impossible_group[] = "Invalid reference to group";
11487 if (has_intervening_patws) {
11489 vFAIL("In '(?...)', the '(' and '?' must be adjacent");
11492 RExC_parse++; /* past the '?' */
11493 paren = *RExC_parse; /* might be a trailing NUL, if not
11495 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11496 if (RExC_parse > RExC_end) {
11499 ret = 0; /* For look-ahead/behind. */
11502 case 'P': /* (?P...) variants for those used to PCRE/Python */
11503 paren = *RExC_parse;
11504 if ( paren == '<') { /* (?P<...>) named capture */
11506 if (RExC_parse >= RExC_end) {
11507 vFAIL("Sequence (?P<... not terminated");
11509 goto named_capture;
11511 else if (paren == '>') { /* (?P>name) named recursion */
11513 if (RExC_parse >= RExC_end) {
11514 vFAIL("Sequence (?P>... not terminated");
11516 goto named_recursion;
11518 else if (paren == '=') { /* (?P=...) named backref */
11520 return handle_named_backref(pRExC_state, flagp,
11523 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
11524 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11525 vFAIL3("Sequence (%.*s...) not recognized",
11526 (int) (RExC_parse - seqstart), seqstart);
11527 NOT_REACHED; /*NOTREACHED*/
11528 case '<': /* (?<...) */
11529 /* If you want to support (?<*...), first reconcile with GH #17363 */
11530 if (*RExC_parse == '!')
11532 else if (*RExC_parse != '=')
11539 case '\'': /* (?'...') */
11540 name_start = RExC_parse;
11541 svname = reg_scan_name(pRExC_state, REG_RSN_RETURN_NAME);
11542 if ( RExC_parse == name_start
11543 || RExC_parse >= RExC_end
11544 || *RExC_parse != paren)
11546 vFAIL2("Sequence (?%c... not terminated",
11547 paren=='>' ? '<' : (char) paren);
11552 if (!svname) /* shouldn't happen */
11554 "panic: reg_scan_name returned NULL");
11555 if (!RExC_paren_names) {
11556 RExC_paren_names= newHV();
11557 sv_2mortal(MUTABLE_SV(RExC_paren_names));
11559 RExC_paren_name_list= newAV();
11560 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
11563 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
11565 sv_dat = HeVAL(he_str);
11567 /* croak baby croak */
11569 "panic: paren_name hash element allocation failed");
11570 } else if ( SvPOK(sv_dat) ) {
11571 /* (?|...) can mean we have dupes so scan to check
11572 its already been stored. Maybe a flag indicating
11573 we are inside such a construct would be useful,
11574 but the arrays are likely to be quite small, so
11575 for now we punt -- dmq */
11576 IV count = SvIV(sv_dat);
11577 I32 *pv = (I32*)SvPVX(sv_dat);
11579 for ( i = 0 ; i < count ; i++ ) {
11580 if ( pv[i] == RExC_npar ) {
11586 pv = (I32*)SvGROW(sv_dat,
11587 SvCUR(sv_dat) + sizeof(I32)+1);
11588 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
11589 pv[count] = RExC_npar;
11590 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
11593 (void)SvUPGRADE(sv_dat, SVt_PVNV);
11594 sv_setpvn(sv_dat, (char *)&(RExC_npar),
11597 SvIV_set(sv_dat, 1);
11600 /* Yes this does cause a memory leak in debugging Perls
11602 if (!av_store(RExC_paren_name_list,
11603 RExC_npar, SvREFCNT_inc_NN(svname)))
11604 SvREFCNT_dec_NN(svname);
11607 /*sv_dump(sv_dat);*/
11609 nextchar(pRExC_state);
11611 goto capturing_parens;
11614 RExC_seen |= REG_LOOKBEHIND_SEEN;
11615 RExC_in_lookaround++;
11617 if (RExC_parse >= RExC_end) {
11618 vFAIL("Sequence (?... not terminated");
11620 RExC_seen_zerolen++;
11622 case '=': /* (?=...) */
11623 RExC_seen_zerolen++;
11624 RExC_in_lookaround++;
11626 case '!': /* (?!...) */
11627 RExC_seen_zerolen++;
11628 /* check if we're really just a "FAIL" assertion */
11629 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
11630 FALSE /* Don't force to /x */ );
11631 if (*RExC_parse == ')') {
11632 ret=reganode(pRExC_state, OPFAIL, 0);
11633 nextchar(pRExC_state);
11636 RExC_in_lookaround++;
11638 case '|': /* (?|...) */
11639 /* branch reset, behave like a (?:...) except that
11640 buffers in alternations share the same numbers */
11642 after_freeze = freeze_paren = RExC_npar;
11644 /* XXX This construct currently requires an extra pass.
11645 * Investigation would be required to see if that could be
11647 REQUIRE_PARENS_PASS;
11649 case ':': /* (?:...) */
11650 case '>': /* (?>...) */
11652 case '$': /* (?$...) */
11653 case '@': /* (?@...) */
11654 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
11656 case '0' : /* (?0) */
11657 case 'R' : /* (?R) */
11658 if (RExC_parse == RExC_end || *RExC_parse != ')')
11659 FAIL("Sequence (?R) not terminated");
11661 RExC_seen |= REG_RECURSE_SEEN;
11663 /* XXX These constructs currently require an extra pass.
11664 * It probably could be changed */
11665 REQUIRE_PARENS_PASS;
11667 *flagp |= POSTPONED;
11668 goto gen_recurse_regop;
11670 /* named and numeric backreferences */
11671 case '&': /* (?&NAME) */
11672 parse_start = RExC_parse - 1;
11675 SV *sv_dat = reg_scan_name(pRExC_state,
11676 REG_RSN_RETURN_DATA);
11677 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
11679 if (RExC_parse >= RExC_end || *RExC_parse != ')')
11680 vFAIL("Sequence (?&... not terminated");
11681 goto gen_recurse_regop;
11684 if (! inRANGE(RExC_parse[0], '1', '9')) {
11686 vFAIL("Illegal pattern");
11688 goto parse_recursion;
11690 case '-': /* (?-1) */
11691 if (! inRANGE(RExC_parse[0], '1', '9')) {
11692 RExC_parse--; /* rewind to let it be handled later */
11696 case '1': case '2': case '3': case '4': /* (?1) */
11697 case '5': case '6': case '7': case '8': case '9':
11698 RExC_parse = (char *) seqstart + 1; /* Point to the digit */
11701 bool is_neg = FALSE;
11703 parse_start = RExC_parse - 1; /* MJD */
11704 if (*RExC_parse == '-') {
11709 if (grok_atoUV(RExC_parse, &unum, &endptr)
11713 RExC_parse = (char*)endptr;
11715 else { /* Overflow, or something like that. Position
11716 beyond all digits for the message */
11717 while (RExC_parse < RExC_end && isDIGIT(*RExC_parse)) {
11720 vFAIL(impossible_group);
11723 /* -num is always representable on 1 and 2's complement
11728 if (*RExC_parse!=')')
11729 vFAIL("Expecting close bracket");
11732 if (paren == '-' || paren == '+') {
11734 /* Don't overflow */
11735 if (UNLIKELY(I32_MAX - RExC_npar < num)) {
11737 vFAIL(impossible_group);
11741 Diagram of capture buffer numbering.
11742 Top line is the normal capture buffer numbers
11743 Bottom line is the negative indexing as from
11747 /(a(x)y)(a(b(c(?+2)d)e)f)(g(h))/
11748 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
11749 - 5 4 3 2 1 X Y x x
11751 Resolve to absolute group. Recall that RExC_npar is +1 of
11752 the actual parenthesis group number. For lookahead, we
11753 have to compensate for that. Using the above example, when
11754 we get to Y in the parse, num is 2 and RExC_npar is 6. We
11755 want 7 for +2, and 4 for -2.
11757 if ( paren == '+' ) {
11763 if (paren == '-' && num < 1) {
11765 vFAIL(non_existent_group_msg);
11769 if (num >= RExC_npar) {
11771 /* It might be a forward reference; we can't fail until we
11772 * know, by completing the parse to get all the groups, and
11773 * then reparsing */
11774 if (ALL_PARENS_COUNTED) {
11775 if (num >= RExC_total_parens) {
11777 vFAIL(non_existent_group_msg);
11781 REQUIRE_PARENS_PASS;
11785 /* We keep track how many GOSUB items we have produced.
11786 To start off the ARG2L() of the GOSUB holds its "id",
11787 which is used later in conjunction with RExC_recurse
11788 to calculate the offset we need to jump for the GOSUB,
11789 which it will store in the final representation.
11790 We have to defer the actual calculation until much later
11791 as the regop may move.
11793 ret = reg2Lanode(pRExC_state, GOSUB, num, RExC_recurse_count);
11794 RExC_recurse_count++;
11795 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11796 "%*s%*s Recurse #%" UVuf " to %" IVdf "\n",
11797 22, "| |", (int)(depth * 2 + 1), "",
11798 (UV)ARG(REGNODE_p(ret)),
11799 (IV)ARG2L(REGNODE_p(ret))));
11800 RExC_seen |= REG_RECURSE_SEEN;
11802 Set_Node_Length(REGNODE_p(ret),
11803 1 + regarglen[OP(REGNODE_p(ret))]); /* MJD */
11804 Set_Node_Offset(REGNODE_p(ret), parse_start); /* MJD */
11806 *flagp |= POSTPONED;
11807 assert(*RExC_parse == ')');
11808 nextchar(pRExC_state);
11813 case '?': /* (??...) */
11815 if (*RExC_parse != '{') {
11816 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
11817 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11819 "Sequence (%" UTF8f "...) not recognized",
11820 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
11821 NOT_REACHED; /*NOTREACHED*/
11823 *flagp |= POSTPONED;
11827 case '{': /* (?{...}) */
11830 struct reg_code_block *cb;
11833 RExC_seen_zerolen++;
11835 if ( !pRExC_state->code_blocks
11836 || pRExC_state->code_index
11837 >= pRExC_state->code_blocks->count
11838 || pRExC_state->code_blocks->cb[pRExC_state->code_index].start
11839 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
11842 if (RExC_pm_flags & PMf_USE_RE_EVAL)
11843 FAIL("panic: Sequence (?{...}): no code block found\n");
11844 FAIL("Eval-group not allowed at runtime, use re 'eval'");
11846 /* this is a pre-compiled code block (?{...}) */
11847 cb = &pRExC_state->code_blocks->cb[pRExC_state->code_index];
11848 RExC_parse = RExC_start + cb->end;
11850 if (cb->src_regex) {
11851 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
11852 RExC_rxi->data->data[n] =
11853 (void*)SvREFCNT_inc((SV*)cb->src_regex);
11854 RExC_rxi->data->data[n+1] = (void*)o;
11857 n = add_data(pRExC_state,
11858 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
11859 RExC_rxi->data->data[n] = (void*)o;
11861 pRExC_state->code_index++;
11862 nextchar(pRExC_state);
11865 regnode_offset eval;
11866 ret = reg_node(pRExC_state, LOGICAL);
11868 eval = reg2Lanode(pRExC_state, EVAL,
11871 /* for later propagation into (??{})
11873 RExC_flags & RXf_PMf_COMPILETIME
11875 FLAGS(REGNODE_p(ret)) = 2;
11876 if (! REGTAIL(pRExC_state, ret, eval)) {
11877 REQUIRE_BRANCHJ(flagp, 0);
11879 /* deal with the length of this later - MJD */
11882 ret = reg2Lanode(pRExC_state, EVAL, n, 0);
11883 Set_Node_Length(REGNODE_p(ret), RExC_parse - parse_start + 1);
11884 Set_Node_Offset(REGNODE_p(ret), parse_start);
11887 case '(': /* (?(?{...})...) and (?(?=...)...) */
11890 const int DEFINE_len = sizeof("DEFINE") - 1;
11891 if ( RExC_parse < RExC_end - 1
11892 && ( ( RExC_parse[0] == '?' /* (?(?...)) */
11893 && ( RExC_parse[1] == '='
11894 || RExC_parse[1] == '!'
11895 || RExC_parse[1] == '<'
11896 || RExC_parse[1] == '{'))
11897 || ( RExC_parse[0] == '*' /* (?(*...)) */
11898 && ( memBEGINs(RExC_parse + 1,
11899 (Size_t) (RExC_end - (RExC_parse + 1)),
11901 || memBEGINs(RExC_parse + 1,
11902 (Size_t) (RExC_end - (RExC_parse + 1)),
11904 || memBEGINs(RExC_parse + 1,
11905 (Size_t) (RExC_end - (RExC_parse + 1)),
11907 || memBEGINs(RExC_parse + 1,
11908 (Size_t) (RExC_end - (RExC_parse + 1)),
11910 || memBEGINs(RExC_parse + 1,
11911 (Size_t) (RExC_end - (RExC_parse + 1)),
11912 "positive_lookahead:")
11913 || memBEGINs(RExC_parse + 1,
11914 (Size_t) (RExC_end - (RExC_parse + 1)),
11915 "positive_lookbehind:")
11916 || memBEGINs(RExC_parse + 1,
11917 (Size_t) (RExC_end - (RExC_parse + 1)),
11918 "negative_lookahead:")
11919 || memBEGINs(RExC_parse + 1,
11920 (Size_t) (RExC_end - (RExC_parse + 1)),
11921 "negative_lookbehind:"))))
11922 ) { /* Lookahead or eval. */
11924 regnode_offset tail;
11926 ret = reg_node(pRExC_state, LOGICAL);
11927 FLAGS(REGNODE_p(ret)) = 1;
11929 tail = reg(pRExC_state, 1, &flag, depth+1);
11930 RETURN_FAIL_ON_RESTART(flag, flagp);
11931 if (! REGTAIL(pRExC_state, ret, tail)) {
11932 REQUIRE_BRANCHJ(flagp, 0);
11936 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
11937 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
11939 char ch = RExC_parse[0] == '<' ? '>' : '\'';
11940 char *name_start= RExC_parse++;
11942 SV *sv_dat=reg_scan_name(pRExC_state, REG_RSN_RETURN_DATA);
11943 if ( RExC_parse == name_start
11944 || RExC_parse >= RExC_end
11945 || *RExC_parse != ch)
11947 vFAIL2("Sequence (?(%c... not terminated",
11948 (ch == '>' ? '<' : ch));
11952 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11953 RExC_rxi->data->data[num]=(void*)sv_dat;
11954 SvREFCNT_inc_simple_void_NN(sv_dat);
11956 ret = reganode(pRExC_state, GROUPPN, num);
11957 goto insert_if_check_paren;
11959 else if (memBEGINs(RExC_parse,
11960 (STRLEN) (RExC_end - RExC_parse),
11963 ret = reganode(pRExC_state, DEFINEP, 0);
11964 RExC_parse += DEFINE_len;
11966 goto insert_if_check_paren;
11968 else if (RExC_parse[0] == 'R') {
11970 /* parno == 0 => /(?(R)YES|NO)/ "in any form of recursion OR eval"
11971 * parno == 1 => /(?(R0)YES|NO)/ "in GOSUB (?0) / (?R)"
11972 * parno == 2 => /(?(R1)YES|NO)/ "in GOSUB (?1) (parno-1)"
11975 if (RExC_parse[0] == '0') {
11979 else if (inRANGE(RExC_parse[0], '1', '9')) {
11982 if (grok_atoUV(RExC_parse, &uv, &endptr)
11985 parno = (I32)uv + 1;
11986 RExC_parse = (char*)endptr;
11988 /* else "Switch condition not recognized" below */
11989 } else if (RExC_parse[0] == '&') {
11992 sv_dat = reg_scan_name(pRExC_state,
11993 REG_RSN_RETURN_DATA);
11995 parno = 1 + *((I32 *)SvPVX(sv_dat));
11997 ret = reganode(pRExC_state, INSUBP, parno);
11998 goto insert_if_check_paren;
12000 else if (inRANGE(RExC_parse[0], '1', '9')) {
12005 if (grok_atoUV(RExC_parse, &uv, &endptr)
12009 RExC_parse = (char*)endptr;
12012 vFAIL("panic: grok_atoUV returned FALSE");
12014 ret = reganode(pRExC_state, GROUPP, parno);
12016 insert_if_check_paren:
12017 if (UCHARAT(RExC_parse) != ')') {
12019 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
12021 vFAIL("Switch condition not recognized");
12023 nextchar(pRExC_state);
12025 if (! REGTAIL(pRExC_state, ret, reganode(pRExC_state,
12028 REQUIRE_BRANCHJ(flagp, 0);
12030 br = regbranch(pRExC_state, &flags, 1, depth+1);
12032 RETURN_FAIL_ON_RESTART(flags,flagp);
12033 FAIL2("panic: regbranch returned failure, flags=%#" UVxf,
12036 if (! REGTAIL(pRExC_state, br, reganode(pRExC_state,
12039 REQUIRE_BRANCHJ(flagp, 0);
12041 c = UCHARAT(RExC_parse);
12042 nextchar(pRExC_state);
12043 if (flags&HASWIDTH)
12044 *flagp |= HASWIDTH;
12047 vFAIL("(?(DEFINE)....) does not allow branches");
12049 /* Fake one for optimizer. */
12050 lastbr = reganode(pRExC_state, IFTHEN, 0);
12052 if (!regbranch(pRExC_state, &flags, 1, depth+1)) {
12053 RETURN_FAIL_ON_RESTART(flags, flagp);
12054 FAIL2("panic: regbranch returned failure, flags=%#" UVxf,
12057 if (! REGTAIL(pRExC_state, ret, lastbr)) {
12058 REQUIRE_BRANCHJ(flagp, 0);
12060 if (flags&HASWIDTH)
12061 *flagp |= HASWIDTH;
12062 c = UCHARAT(RExC_parse);
12063 nextchar(pRExC_state);
12068 if (RExC_parse >= RExC_end)
12069 vFAIL("Switch (?(condition)... not terminated");
12071 vFAIL("Switch (?(condition)... contains too many branches");
12073 ender = reg_node(pRExC_state, TAIL);
12074 if (! REGTAIL(pRExC_state, br, ender)) {
12075 REQUIRE_BRANCHJ(flagp, 0);
12078 if (! REGTAIL(pRExC_state, lastbr, ender)) {
12079 REQUIRE_BRANCHJ(flagp, 0);
12081 if (! REGTAIL(pRExC_state,
12084 NEXTOPER(REGNODE_p(lastbr)))),
12087 REQUIRE_BRANCHJ(flagp, 0);
12091 if (! REGTAIL(pRExC_state, ret, ender)) {
12092 REQUIRE_BRANCHJ(flagp, 0);
12094 #if 0 /* Removing this doesn't cause failures in the test suite -- khw */
12095 RExC_size++; /* XXX WHY do we need this?!!
12096 For large programs it seems to be required
12097 but I can't figure out why. -- dmq*/
12102 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
12104 vFAIL("Unknown switch condition (?(...))");
12106 case '[': /* (?[ ... ]) */
12107 return handle_regex_sets(pRExC_state, NULL, flagp, depth+1,
12109 case 0: /* A NUL */
12110 RExC_parse--; /* for vFAIL to print correctly */
12111 vFAIL("Sequence (? incomplete");
12115 if (RExC_strict) { /* [perl #132851] */
12116 ckWARNreg(RExC_parse, "Empty (?) without any modifiers");
12119 case '*': /* If you want to support (?*...), first reconcile with GH #17363 */
12121 default: /* e.g., (?i) */
12122 RExC_parse = (char *) seqstart + 1;
12124 parse_lparen_question_flags(pRExC_state);
12125 if (UCHARAT(RExC_parse) != ':') {
12126 if (RExC_parse < RExC_end)
12127 nextchar(pRExC_state);
12132 nextchar(pRExC_state);
12137 else if (!(RExC_flags & RXf_PMf_NOCAPTURE)) { /* (...) */
12141 if (! ALL_PARENS_COUNTED) {
12142 /* If we are in our first pass through (and maybe only pass),
12143 * we need to allocate memory for the capturing parentheses
12147 if (!RExC_parens_buf_size) {
12148 /* first guess at number of parens we might encounter */
12149 RExC_parens_buf_size = 10;
12151 /* setup RExC_open_parens, which holds the address of each
12152 * OPEN tag, and to make things simpler for the 0 index the
12153 * start of the program - this is used later for offsets */
12154 Newxz(RExC_open_parens, RExC_parens_buf_size,
12156 RExC_open_parens[0] = 1; /* +1 for REG_MAGIC */
12158 /* setup RExC_close_parens, which holds the address of each
12159 * CLOSE tag, and to make things simpler for the 0 index
12160 * the end of the program - this is used later for offsets
12162 Newxz(RExC_close_parens, RExC_parens_buf_size,
12164 /* we dont know where end op starts yet, so we dont need to
12165 * set RExC_close_parens[0] like we do RExC_open_parens[0]
12168 else if (RExC_npar > RExC_parens_buf_size) {
12169 I32 old_size = RExC_parens_buf_size;
12171 RExC_parens_buf_size *= 2;
12173 Renew(RExC_open_parens, RExC_parens_buf_size,
12175 Zero(RExC_open_parens + old_size,
12176 RExC_parens_buf_size - old_size, regnode_offset);
12178 Renew(RExC_close_parens, RExC_parens_buf_size,
12180 Zero(RExC_close_parens + old_size,
12181 RExC_parens_buf_size - old_size, regnode_offset);
12185 ret = reganode(pRExC_state, OPEN, parno);
12186 if (!RExC_nestroot)
12187 RExC_nestroot = parno;
12188 if (RExC_open_parens && !RExC_open_parens[parno])
12190 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12191 "%*s%*s Setting open paren #%" IVdf " to %zu\n",
12192 22, "| |", (int)(depth * 2 + 1), "",
12194 RExC_open_parens[parno]= ret;
12197 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
12198 Set_Node_Offset(REGNODE_p(ret), RExC_parse); /* MJD */
12201 /* with RXf_PMf_NOCAPTURE treat (...) as (?:...) */
12210 /* Pick up the branches, linking them together. */
12211 parse_start = RExC_parse; /* MJD */
12212 br = regbranch(pRExC_state, &flags, 1, depth+1);
12214 /* branch_len = (paren != 0); */
12217 RETURN_FAIL_ON_RESTART(flags, flagp);
12218 FAIL2("panic: regbranch returned failure, flags=%#" UVxf, (UV) flags);
12220 if (*RExC_parse == '|') {
12221 if (RExC_use_BRANCHJ) {
12222 reginsert(pRExC_state, BRANCHJ, br, depth+1);
12225 reginsert(pRExC_state, BRANCH, br, depth+1);
12226 Set_Node_Length(REGNODE_p(br), paren != 0);
12227 Set_Node_Offset_To_R(br, parse_start-RExC_start);
12231 else if (paren == ':') {
12232 *flagp |= flags&SIMPLE;
12234 if (is_open) { /* Starts with OPEN. */
12235 if (! REGTAIL(pRExC_state, ret, br)) { /* OPEN -> first. */
12236 REQUIRE_BRANCHJ(flagp, 0);
12239 else if (paren != '?') /* Not Conditional */
12241 *flagp |= flags & (HASWIDTH | POSTPONED);
12243 while (*RExC_parse == '|') {
12244 if (RExC_use_BRANCHJ) {
12247 ender = reganode(pRExC_state, LONGJMP, 0);
12249 /* Append to the previous. */
12250 shut_gcc_up = REGTAIL(pRExC_state,
12251 REGNODE_OFFSET(NEXTOPER(NEXTOPER(REGNODE_p(lastbr)))),
12253 PERL_UNUSED_VAR(shut_gcc_up);
12255 nextchar(pRExC_state);
12256 if (freeze_paren) {
12257 if (RExC_npar > after_freeze)
12258 after_freeze = RExC_npar;
12259 RExC_npar = freeze_paren;
12261 br = regbranch(pRExC_state, &flags, 0, depth+1);
12264 RETURN_FAIL_ON_RESTART(flags, flagp);
12265 FAIL2("panic: regbranch returned failure, flags=%#" UVxf, (UV) flags);
12267 if (! REGTAIL(pRExC_state, lastbr, br)) { /* BRANCH -> BRANCH. */
12268 REQUIRE_BRANCHJ(flagp, 0);
12271 *flagp |= flags & (HASWIDTH | POSTPONED);
12274 if (have_branch || paren != ':') {
12277 /* Make a closing node, and hook it on the end. */
12280 ender = reg_node(pRExC_state, TAIL);
12283 ender = reganode(pRExC_state, CLOSE, parno);
12284 if ( RExC_close_parens ) {
12285 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12286 "%*s%*s Setting close paren #%" IVdf " to %zu\n",
12287 22, "| |", (int)(depth * 2 + 1), "",
12288 (IV)parno, ender));
12289 RExC_close_parens[parno]= ender;
12290 if (RExC_nestroot == parno)
12293 Set_Node_Offset(REGNODE_p(ender), RExC_parse+1); /* MJD */
12294 Set_Node_Length(REGNODE_p(ender), 1); /* MJD */
12297 ender = reg_node(pRExC_state, SRCLOSE);
12298 RExC_in_script_run = 0;
12308 *flagp &= ~HASWIDTH;
12310 case 't': /* aTomic */
12312 ender = reg_node(pRExC_state, SUCCEED);
12315 ender = reg_node(pRExC_state, END);
12316 assert(!RExC_end_op); /* there can only be one! */
12317 RExC_end_op = REGNODE_p(ender);
12318 if (RExC_close_parens) {
12319 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12320 "%*s%*s Setting close paren #0 (END) to %zu\n",
12321 22, "| |", (int)(depth * 2 + 1), "",
12324 RExC_close_parens[0]= ender;
12329 DEBUG_PARSE_MSG("lsbr");
12330 regprop(RExC_rx, RExC_mysv1, REGNODE_p(lastbr), NULL, pRExC_state);
12331 regprop(RExC_rx, RExC_mysv2, REGNODE_p(ender), NULL, pRExC_state);
12332 Perl_re_printf( aTHX_ "~ tying lastbr %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
12333 SvPV_nolen_const(RExC_mysv1),
12335 SvPV_nolen_const(RExC_mysv2),
12337 (IV)(ender - lastbr)
12340 if (! REGTAIL(pRExC_state, lastbr, ender)) {
12341 REQUIRE_BRANCHJ(flagp, 0);
12345 char is_nothing= 1;
12347 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
12349 /* Hook the tails of the branches to the closing node. */
12350 for (br = REGNODE_p(ret); br; br = regnext(br)) {
12351 const U8 op = PL_regkind[OP(br)];
12352 if (op == BRANCH) {
12353 if (! REGTAIL_STUDY(pRExC_state,
12354 REGNODE_OFFSET(NEXTOPER(br)),
12357 REQUIRE_BRANCHJ(flagp, 0);
12359 if ( OP(NEXTOPER(br)) != NOTHING
12360 || regnext(NEXTOPER(br)) != REGNODE_p(ender))
12363 else if (op == BRANCHJ) {
12364 bool shut_gcc_up = REGTAIL_STUDY(pRExC_state,
12365 REGNODE_OFFSET(NEXTOPER(NEXTOPER(br))),
12367 PERL_UNUSED_VAR(shut_gcc_up);
12368 /* for now we always disable this optimisation * /
12369 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
12370 || regnext(NEXTOPER(NEXTOPER(br))) != REGNODE_p(ender))
12376 regnode * ret_as_regnode = REGNODE_p(ret);
12377 br= PL_regkind[OP(ret_as_regnode)] != BRANCH
12378 ? regnext(ret_as_regnode)
12381 DEBUG_PARSE_MSG("NADA");
12382 regprop(RExC_rx, RExC_mysv1, ret_as_regnode,
12383 NULL, pRExC_state);
12384 regprop(RExC_rx, RExC_mysv2, REGNODE_p(ender),
12385 NULL, pRExC_state);
12386 Perl_re_printf( aTHX_ "~ converting ret %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
12387 SvPV_nolen_const(RExC_mysv1),
12388 (IV)REG_NODE_NUM(ret_as_regnode),
12389 SvPV_nolen_const(RExC_mysv2),
12395 if (OP(REGNODE_p(ender)) == TAIL) {
12397 RExC_emit= REGNODE_OFFSET(br) + 1;
12400 for ( opt= br + 1; opt < REGNODE_p(ender) ; opt++ )
12401 OP(opt)= OPTIMIZED;
12402 NEXT_OFF(br)= REGNODE_p(ender) - br;
12410 /* Even/odd or x=don't care: 010101x10x */
12411 static const char parens[] = "=!aA<,>Bbt";
12412 /* flag below is set to 0 up through 'A'; 1 for larger */
12414 if (paren && (p = strchr(parens, paren))) {
12415 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
12416 int flag = (p - parens) > 3;
12418 if (paren == '>' || paren == 't') {
12419 node = SUSPEND, flag = 0;
12422 reginsert(pRExC_state, node, ret, depth+1);
12423 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
12424 Set_Node_Offset(REGNODE_p(ret), parse_start + 1);
12425 FLAGS(REGNODE_p(ret)) = flag;
12426 if (! REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL)))
12428 REQUIRE_BRANCHJ(flagp, 0);
12433 /* Check for proper termination. */
12435 /* restore original flags, but keep (?p) and, if we've encountered
12436 * something in the parse that changes /d rules into /u, keep the /u */
12437 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
12438 if (DEPENDS_SEMANTICS && toUSE_UNI_CHARSET_NOT_DEPENDS) {
12439 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
12441 if (RExC_parse >= RExC_end || UCHARAT(RExC_parse) != ')') {
12442 RExC_parse = oregcomp_parse;
12443 vFAIL("Unmatched (");
12445 nextchar(pRExC_state);
12447 else if (!paren && RExC_parse < RExC_end) {
12448 if (*RExC_parse == ')') {
12450 vFAIL("Unmatched )");
12453 FAIL("Junk on end of regexp"); /* "Can't happen". */
12454 NOT_REACHED; /* NOTREACHED */
12457 if (after_freeze > RExC_npar)
12458 RExC_npar = after_freeze;
12460 RExC_in_lookaround = was_in_lookaround;
12466 - regbranch - one alternative of an | operator
12468 * Implements the concatenation operator.
12470 * On success, returns the offset at which any next node should be placed into
12471 * the regex engine program being compiled.
12473 * Returns 0 otherwise, setting flagp to RESTART_PARSE if the parse needs
12474 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to
12477 STATIC regnode_offset
12478 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
12480 regnode_offset ret;
12481 regnode_offset chain = 0;
12482 regnode_offset latest;
12483 I32 flags = 0, c = 0;
12484 DECLARE_AND_GET_RE_DEBUG_FLAGS;
12486 PERL_ARGS_ASSERT_REGBRANCH;
12488 DEBUG_PARSE("brnc");
12493 if (RExC_use_BRANCHJ)
12494 ret = reganode(pRExC_state, BRANCHJ, 0);
12496 ret = reg_node(pRExC_state, BRANCH);
12497 Set_Node_Length(REGNODE_p(ret), 1);
12501 *flagp = 0; /* Initialize. */
12503 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
12504 FALSE /* Don't force to /x */ );
12505 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
12506 flags &= ~TRYAGAIN;
12507 latest = regpiece(pRExC_state, &flags, depth+1);
12509 if (flags & TRYAGAIN)
12511 RETURN_FAIL_ON_RESTART(flags, flagp);
12512 FAIL2("panic: regpiece returned failure, flags=%#" UVxf, (UV) flags);
12516 *flagp |= flags&(HASWIDTH|POSTPONED);
12518 /* FIXME adding one for every branch after the first is probably
12519 * excessive now we have TRIE support. (hv) */
12521 if (! REGTAIL(pRExC_state, chain, latest)) {
12522 /* XXX We could just redo this branch, but figuring out what
12523 * bookkeeping needs to be reset is a pain, and it's likely
12524 * that other branches that goto END will also be too large */
12525 REQUIRE_BRANCHJ(flagp, 0);
12531 if (chain == 0) { /* Loop ran zero times. */
12532 chain = reg_node(pRExC_state, NOTHING);
12537 *flagp |= flags&SIMPLE;
12544 - regcurly - a little FSA that accepts {\d+,?\d*}
12547 #ifndef PERL_IN_XSUB_RE
12549 Perl_regcurly(const char *s)
12551 PERL_ARGS_ASSERT_REGCURLY;
12557 while (isDIGIT(*s))
12561 while (isDIGIT(*s))
12569 - regpiece - something followed by possible quantifier * + ? {n,m}
12571 * Note that the branching code sequences used for ? and the general cases
12572 * of * and + are somewhat optimized: they use the same NOTHING node as
12573 * both the endmarker for their branch list and the body of the last branch.
12574 * It might seem that this node could be dispensed with entirely, but the
12575 * endmarker role is not redundant.
12577 * On success, returns the offset at which any next node should be placed into
12578 * the regex engine program being compiled.
12580 * Returns 0 otherwise, with *flagp set to indicate why:
12581 * TRYAGAIN if regatom() returns 0 with TRYAGAIN.
12582 * RESTART_PARSE if the parse needs to be restarted, or'd with
12583 * NEED_UTF8 if the pattern needs to be upgraded to UTF-8.
12585 STATIC regnode_offset
12586 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
12588 regnode_offset ret;
12592 const char * const origparse = RExC_parse;
12594 I32 max = REG_INFTY;
12595 #ifdef RE_TRACK_PATTERN_OFFSETS
12598 const char *maxpos = NULL;
12601 /* Save the original in case we change the emitted regop to a FAIL. */
12602 const regnode_offset orig_emit = RExC_emit;
12604 DECLARE_AND_GET_RE_DEBUG_FLAGS;
12606 PERL_ARGS_ASSERT_REGPIECE;
12608 DEBUG_PARSE("piec");
12610 ret = regatom(pRExC_state, &flags, depth+1);
12612 RETURN_FAIL_ON_RESTART_OR_FLAGS(flags, flagp, TRYAGAIN);
12613 FAIL2("panic: regatom returned failure, flags=%#" UVxf, (UV) flags);
12616 #ifdef RE_TRACK_PATTERN_OFFSETS
12617 parse_start = RExC_parse;
12624 nextchar(pRExC_state);
12629 nextchar(pRExC_state);
12634 nextchar(pRExC_state);
12638 case '{': /* A '{' may or may not indicate a quantifier; call regcurly()
12639 to determine which */
12640 if (regcurly(RExC_parse)) {
12641 const char* endptr;
12643 /* Here is a quantifier, parse for min and max values */
12645 next = RExC_parse + 1;
12646 while (isDIGIT(*next) || *next == ',') {
12647 if (*next == ',') {
12656 assert(*next == '}');
12661 if (isDIGIT(*RExC_parse)) {
12663 if (!grok_atoUV(RExC_parse, &uv, &endptr))
12664 vFAIL("Invalid quantifier in {,}");
12665 if (uv >= REG_INFTY)
12666 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
12671 if (*maxpos == ',')
12674 maxpos = RExC_parse;
12675 if (isDIGIT(*maxpos)) {
12677 if (!grok_atoUV(maxpos, &uv, &endptr))
12678 vFAIL("Invalid quantifier in {,}");
12679 if (uv >= REG_INFTY)
12680 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
12683 max = REG_INFTY; /* meaning "infinity" */
12687 nextchar(pRExC_state);
12688 if (max < min) { /* If can't match, warn and optimize to fail
12690 reginsert(pRExC_state, OPFAIL, orig_emit, depth+1);
12691 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
12692 NEXT_OFF(REGNODE_p(orig_emit)) =
12693 regarglen[OPFAIL] + NODE_STEP_REGNODE;
12696 else if (min == max && *RExC_parse == '?')
12698 ckWARN2reg(RExC_parse + 1,
12699 "Useless use of greediness modifier '%c'",
12704 } /* End of is regcurly() */
12706 /* Here was a '{', but what followed it didn't form a quantifier. */
12712 NOT_REACHED; /*NOTREACHED*/
12715 /* Here we have a quantifier, and have calculated 'min' and 'max'.
12717 * Check and possibly adjust a zero width operand */
12718 if (! (flags & (HASWIDTH|POSTPONED))) {
12719 if (max > REG_INFTY/3) {
12720 if (origparse[0] == '\\' && origparse[1] == 'K') {
12722 "%" UTF8f " is forbidden - matches null string"
12724 UTF8fARG(UTF, (RExC_parse >= origparse
12725 ? RExC_parse - origparse
12729 ckWARN2reg(RExC_parse,
12730 "%" UTF8f " matches null string many times",
12731 UTF8fARG(UTF, (RExC_parse >= origparse
12732 ? RExC_parse - origparse
12738 /* There's no point in trying to match something 0 length more than
12739 * once except for extra side effects, which we don't have here since
12749 /* If this is a code block pass it up */
12750 *flagp |= (flags & POSTPONED);
12753 *flagp |= (flags & HASWIDTH);
12754 if (max == REG_INFTY)
12755 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
12758 /* 'SIMPLE' operands don't require full generality */
12759 if ((flags&SIMPLE)) {
12760 if (max == REG_INFTY) {
12762 if (UNLIKELY(RExC_pm_flags & PMf_WILDCARD)) {
12763 goto min0_maxINF_wildcard_forbidden;
12766 reginsert(pRExC_state, STAR, ret, depth+1);
12770 else if (min == 1) {
12771 reginsert(pRExC_state, PLUS, ret, depth+1);
12777 /* Here, SIMPLE, but not the '*' and '+' special cases */
12779 MARK_NAUGHTY_EXP(2, 2);
12780 reginsert(pRExC_state, CURLY, ret, depth+1);
12781 Set_Node_Offset(REGNODE_p(ret), parse_start+1); /* MJD */
12782 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
12784 else { /* not SIMPLE */
12785 const regnode_offset w = reg_node(pRExC_state, WHILEM);
12787 FLAGS(REGNODE_p(w)) = 0;
12788 if (! REGTAIL(pRExC_state, ret, w)) {
12789 REQUIRE_BRANCHJ(flagp, 0);
12791 if (RExC_use_BRANCHJ) {
12792 reginsert(pRExC_state, LONGJMP, ret, depth+1);
12793 reginsert(pRExC_state, NOTHING, ret, depth+1);
12794 NEXT_OFF(REGNODE_p(ret)) = 3; /* Go over LONGJMP. */
12796 reginsert(pRExC_state, CURLYX, ret, depth+1);
12798 Set_Node_Offset(REGNODE_p(ret), parse_start+1);
12799 Set_Node_Length(REGNODE_p(ret),
12800 op == '{' ? (RExC_parse - parse_start) : 1);
12802 if (RExC_use_BRANCHJ)
12803 NEXT_OFF(REGNODE_p(ret)) = 3; /* Go over NOTHING to
12805 if (! REGTAIL(pRExC_state, ret, reg_node(pRExC_state,
12808 REQUIRE_BRANCHJ(flagp, 0);
12810 RExC_whilem_seen++;
12811 MARK_NAUGHTY_EXP(1, 4); /* compound interest */
12814 /* Finish up the CURLY/CURLYX case */
12815 FLAGS(REGNODE_p(ret)) = 0;
12817 ARG1_SET(REGNODE_p(ret), (U16)min);
12818 ARG2_SET(REGNODE_p(ret), (U16)max);
12822 /* Process any greediness modifiers */
12823 if (*RExC_parse == '?') {
12824 nextchar(pRExC_state);
12825 reginsert(pRExC_state, MINMOD, ret, depth+1);
12826 if (! REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE)) {
12827 REQUIRE_BRANCHJ(flagp, 0);
12830 else if (*RExC_parse == '+') {
12831 regnode_offset ender;
12832 nextchar(pRExC_state);
12833 ender = reg_node(pRExC_state, SUCCEED);
12834 if (! REGTAIL(pRExC_state, ret, ender)) {
12835 REQUIRE_BRANCHJ(flagp, 0);
12837 reginsert(pRExC_state, SUSPEND, ret, depth+1);
12838 ender = reg_node(pRExC_state, TAIL);
12839 if (! REGTAIL(pRExC_state, ret, ender)) {
12840 REQUIRE_BRANCHJ(flagp, 0);
12844 /* Forbid extra quantifiers */
12845 if (ISMULT2(RExC_parse)) {
12847 vFAIL("Nested quantifiers");
12852 min0_maxINF_wildcard_forbidden:
12854 /* Here we are in a wildcard match, and the minimum match length is 0, and
12855 * the max could be infinity. This is currently forbidden. The only
12856 * reason is to make it harder to write patterns that take a long long time
12857 * to halt, and because the use of this construct isn't necessary in
12858 * matching Unicode property values */
12860 /* diag_listed_as: Use of %s is not allowed in Unicode property wildcard
12861 subpatterns in regex; marked by <-- HERE in m/%s/
12863 vFAIL("Use of quantifier '*' is not allowed in Unicode property wildcard"
12866 /* Note, don't need to worry about the input being '{0,}', as a '}' isn't
12867 * legal at all in wildcards, so can't get this far */
12869 NOT_REACHED; /*NOTREACHED*/
12873 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state,
12874 regnode_offset * node_p,
12882 /* This routine teases apart the various meanings of \N and returns
12883 * accordingly. The input parameters constrain which meaning(s) is/are valid
12884 * in the current context.
12886 * Exactly one of <node_p> and <code_point_p> must be non-NULL.
12888 * If <code_point_p> is not NULL, the context is expecting the result to be a
12889 * single code point. If this \N instance turns out to a single code point,
12890 * the function returns TRUE and sets *code_point_p to that code point.
12892 * If <node_p> is not NULL, the context is expecting the result to be one of
12893 * the things representable by a regnode. If this \N instance turns out to be
12894 * one such, the function generates the regnode, returns TRUE and sets *node_p
12895 * to point to the offset of that regnode into the regex engine program being
12898 * If this instance of \N isn't legal in any context, this function will
12899 * generate a fatal error and not return.
12901 * On input, RExC_parse should point to the first char following the \N at the
12902 * time of the call. On successful return, RExC_parse will have been updated
12903 * to point to just after the sequence identified by this routine. Also
12904 * *flagp has been updated as needed.
12906 * When there is some problem with the current context and this \N instance,
12907 * the function returns FALSE, without advancing RExC_parse, nor setting
12908 * *node_p, nor *code_point_p, nor *flagp.
12910 * If <cp_count> is not NULL, the caller wants to know the length (in code
12911 * points) that this \N sequence matches. This is set, and the input is
12912 * parsed for errors, even if the function returns FALSE, as detailed below.
12914 * There are 6 possibilities here, as detailed in the next 6 paragraphs.
12916 * Probably the most common case is for the \N to specify a single code point.
12917 * *cp_count will be set to 1, and *code_point_p will be set to that code
12920 * Another possibility is for the input to be an empty \N{}. This is no
12921 * longer accepted, and will generate a fatal error.
12923 * Another possibility is for a custom charnames handler to be in effect which
12924 * translates the input name to an empty string. *cp_count will be set to 0.
12925 * *node_p will be set to a generated NOTHING node.
12927 * Still another possibility is for the \N to mean [^\n]. *cp_count will be
12928 * set to 0. *node_p will be set to a generated REG_ANY node.
12930 * The fifth possibility is that \N resolves to a sequence of more than one
12931 * code points. *cp_count will be set to the number of code points in the
12932 * sequence. *node_p will be set to a generated node returned by this
12933 * function calling S_reg().
12935 * The final possibility is that it is premature to be calling this function;
12936 * the parse needs to be restarted. This can happen when this changes from
12937 * /d to /u rules, or when the pattern needs to be upgraded to UTF-8. The
12938 * latter occurs only when the fifth possibility would otherwise be in
12939 * effect, and is because one of those code points requires the pattern to be
12940 * recompiled as UTF-8. The function returns FALSE, and sets the
12941 * RESTART_PARSE and NEED_UTF8 flags in *flagp, as appropriate. When this
12942 * happens, the caller needs to desist from continuing parsing, and return
12943 * this information to its caller. This is not set for when there is only one
12944 * code point, as this can be called as part of an ANYOF node, and they can
12945 * store above-Latin1 code points without the pattern having to be in UTF-8.
12947 * For non-single-quoted regexes, the tokenizer has resolved character and
12948 * sequence names inside \N{...} into their Unicode values, normalizing the
12949 * result into what we should see here: '\N{U+c1.c2...}', where c1... are the
12950 * hex-represented code points in the sequence. This is done there because
12951 * the names can vary based on what charnames pragma is in scope at the time,
12952 * so we need a way to take a snapshot of what they resolve to at the time of
12953 * the original parse. [perl #56444].
12955 * That parsing is skipped for single-quoted regexes, so here we may get
12956 * '\N{NAME}', which is parsed now. If the single-quoted regex is something
12957 * like '\N{U+41}', that code point is Unicode, and has to be translated into
12958 * the native character set for non-ASCII platforms. The other possibilities
12959 * are already native, so no translation is done. */
12961 char * endbrace; /* points to '}' following the name */
12962 char* p = RExC_parse; /* Temporary */
12964 SV * substitute_parse = NULL;
12969 DECLARE_AND_GET_RE_DEBUG_FLAGS;
12971 PERL_ARGS_ASSERT_GROK_BSLASH_N;
12973 assert(cBOOL(node_p) ^ cBOOL(code_point_p)); /* Exactly one should be set */
12974 assert(! (node_p && cp_count)); /* At most 1 should be set */
12976 if (cp_count) { /* Initialize return for the most common case */
12980 /* The [^\n] meaning of \N ignores spaces and comments under the /x
12981 * modifier. The other meanings do not, so use a temporary until we find
12982 * out which we are being called with */
12983 skip_to_be_ignored_text(pRExC_state, &p,
12984 FALSE /* Don't force to /x */ );
12986 /* Disambiguate between \N meaning a named character versus \N meaning
12987 * [^\n]. The latter is assumed when the {...} following the \N is a legal
12988 * quantifier, or if there is no '{' at all */
12989 if (*p != '{' || regcurly(p)) {
12999 *node_p = reg_node(pRExC_state, REG_ANY);
13000 *flagp |= HASWIDTH|SIMPLE;
13002 Set_Node_Length(REGNODE_p(*(node_p)), 1); /* MJD */
13006 /* The test above made sure that the next real character is a '{', but
13007 * under the /x modifier, it could be separated by space (or a comment and
13008 * \n) and this is not allowed (for consistency with \x{...} and the
13009 * tokenizer handling of \N{NAME}). */
13010 if (*RExC_parse != '{') {
13011 vFAIL("Missing braces on \\N{}");
13014 RExC_parse++; /* Skip past the '{' */
13016 endbrace = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
13017 if (! endbrace) { /* no trailing brace */
13018 vFAIL2("Missing right brace on \\%c{}", 'N');
13021 /* Here, we have decided it should be a named character or sequence. These
13022 * imply Unicode semantics */
13023 REQUIRE_UNI_RULES(flagp, FALSE);
13025 /* \N{_} is what toke.c returns to us to indicate a name that evaluates to
13026 * nothing at all (not allowed under strict) */
13027 if (endbrace - RExC_parse == 1 && *RExC_parse == '_') {
13028 RExC_parse = endbrace;
13030 RExC_parse++; /* Position after the "}" */
13031 vFAIL("Zero length \\N{}");
13037 nextchar(pRExC_state);
13042 *node_p = reg_node(pRExC_state, NOTHING);
13046 if (endbrace - RExC_parse < 2 || ! strBEGINs(RExC_parse, "U+")) {
13048 /* Here, the name isn't of the form U+.... This can happen if the
13049 * pattern is single-quoted, so didn't get evaluated in toke.c. Now
13050 * is the time to find out what the name means */
13052 const STRLEN name_len = endbrace - RExC_parse;
13053 SV * value_sv; /* What does this name evaluate to */
13055 const U8 * value; /* string of name's value */
13056 STRLEN value_len; /* and its length */
13058 /* RExC_unlexed_names is a hash of names that weren't evaluated by
13059 * toke.c, and their values. Make sure is initialized */
13060 if (! RExC_unlexed_names) {
13061 RExC_unlexed_names = newHV();
13064 /* If we have already seen this name in this pattern, use that. This
13065 * allows us to only call the charnames handler once per name per
13066 * pattern. A broken or malicious handler could return something
13067 * different each time, which could cause the results to vary depending
13068 * on if something gets added or subtracted from the pattern that
13069 * causes the number of passes to change, for example */
13070 if ((value_svp = hv_fetch(RExC_unlexed_names, RExC_parse,
13073 value_sv = *value_svp;
13075 else { /* Otherwise we have to go out and get the name */
13076 const char * error_msg = NULL;
13077 value_sv = get_and_check_backslash_N_name(RExC_parse, endbrace,
13081 RExC_parse = endbrace;
13085 /* If no error message, should have gotten a valid return */
13088 /* Save the name's meaning for later use */
13089 if (! hv_store(RExC_unlexed_names, RExC_parse, name_len,
13092 Perl_croak(aTHX_ "panic: hv_store() unexpectedly failed");
13096 /* Here, we have the value the name evaluates to in 'value_sv' */
13097 value = (U8 *) SvPV(value_sv, value_len);
13099 /* See if the result is one code point vs 0 or multiple */
13100 if (inRANGE(value_len, 1, ((UV) SvUTF8(value_sv)
13104 /* Here, exactly one code point. If that isn't what is wanted,
13106 if (! code_point_p) {
13111 /* Convert from string to numeric code point */
13112 *code_point_p = (SvUTF8(value_sv))
13113 ? valid_utf8_to_uvchr(value, NULL)
13116 /* Have parsed this entire single code point \N{...}. *cp_count
13117 * has already been set to 1, so don't do it again. */
13118 RExC_parse = endbrace;
13119 nextchar(pRExC_state);
13121 } /* End of is a single code point */
13123 /* Count the code points, if caller desires. The API says to do this
13124 * even if we will later return FALSE */
13128 *cp_count = (SvUTF8(value_sv))
13129 ? utf8_length(value, value + value_len)
13133 /* Fail if caller doesn't want to handle a multi-code-point sequence.
13134 * But don't back the pointer up if the caller wants to know how many
13135 * code points there are (they need to handle it themselves in this
13144 /* Convert this to a sub-pattern of the form "(?: ... )", and then call
13145 * reg recursively to parse it. That way, it retains its atomicness,
13146 * while not having to worry about any special handling that some code
13147 * points may have. */
13149 substitute_parse = newSVpvs("?:");
13150 sv_catsv(substitute_parse, value_sv);
13151 sv_catpv(substitute_parse, ")");
13153 /* The value should already be native, so no need to convert on EBCDIC
13155 assert(! RExC_recode_x_to_native);
13158 else { /* \N{U+...} */
13159 Size_t count = 0; /* code point count kept internally */
13161 /* We can get to here when the input is \N{U+...} or when toke.c has
13162 * converted a name to the \N{U+...} form. This include changing a
13163 * name that evaluates to multiple code points to \N{U+c1.c2.c3 ...} */
13165 RExC_parse += 2; /* Skip past the 'U+' */
13167 /* Code points are separated by dots. The '}' terminates the whole
13170 do { /* Loop until the ending brace */
13171 I32 flags = PERL_SCAN_SILENT_OVERFLOW
13172 | PERL_SCAN_SILENT_ILLDIGIT
13173 | PERL_SCAN_NOTIFY_ILLDIGIT
13174 | PERL_SCAN_ALLOW_MEDIAL_UNDERSCORES
13175 | PERL_SCAN_DISALLOW_PREFIX;
13176 STRLEN len = endbrace - RExC_parse;
13178 char * start_digit = RExC_parse;
13179 UV cp = grok_hex(RExC_parse, &len, &flags, &overflow_value);
13184 vFAIL("Invalid hexadecimal number in \\N{U+...}");
13189 if (cp > MAX_LEGAL_CP) {
13190 vFAIL(form_cp_too_large_msg(16, start_digit, len, 0));
13193 if (RExC_parse >= endbrace) { /* Got to the closing '}' */
13198 /* Here, is a single code point; fail if doesn't want that */
13199 if (! code_point_p) {
13204 /* A single code point is easy to handle; just return it */
13205 *code_point_p = UNI_TO_NATIVE(cp);
13206 RExC_parse = endbrace;
13207 nextchar(pRExC_state);
13211 /* Here, the parse stopped bfore the ending brace. This is legal
13212 * only if that character is a dot separating code points, like a
13213 * multiple character sequence (of the form "\N{U+c1.c2. ... }".
13214 * So the next character must be a dot (and the one after that
13215 * can't be the endbrace, or we'd have something like \N{U+100.} )
13217 if (*RExC_parse != '.' || RExC_parse + 1 >= endbrace) {
13218 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
13219 ? UTF8SKIP(RExC_parse)
13221 RExC_parse = MIN(endbrace, RExC_parse);/* Guard against
13226 /* Here, looks like its really a multiple character sequence. Fail
13227 * if that's not what the caller wants. But continue with counting
13228 * and error checking if they still want a count */
13229 if (! node_p && ! cp_count) {
13233 /* What is done here is to convert this to a sub-pattern of the
13234 * form \x{char1}\x{char2}... and then call reg recursively to
13235 * parse it (enclosing in "(?: ... )" ). That way, it retains its
13236 * atomicness, while not having to worry about special handling
13237 * that some code points may have. We don't create a subpattern,
13238 * but go through the motions of code point counting and error
13239 * checking, if the caller doesn't want a node returned. */
13241 if (node_p && ! substitute_parse) {
13242 substitute_parse = newSVpvs("?:");
13248 /* Convert to notation the rest of the code understands */
13249 sv_catpvs(substitute_parse, "\\x{");
13250 sv_catpvn(substitute_parse, start_digit,
13251 RExC_parse - start_digit);
13252 sv_catpvs(substitute_parse, "}");
13255 /* Move to after the dot (or ending brace the final time through.)
13260 } while (RExC_parse < endbrace);
13262 if (! node_p) { /* Doesn't want the node */
13269 sv_catpvs(substitute_parse, ")");
13271 /* The values are Unicode, and therefore have to be converted to native
13272 * on a non-Unicode (meaning non-ASCII) platform. */
13273 SET_recode_x_to_native(1);
13276 /* Here, we have the string the name evaluates to, ready to be parsed,
13277 * stored in 'substitute_parse' as a series of valid "\x{...}\x{...}"
13278 * constructs. This can be called from within a substitute parse already.
13279 * The error reporting mechanism doesn't work for 2 levels of this, but the
13280 * code above has validated this new construct, so there should be no
13281 * errors generated by the below. And this isn' an exact copy, so the
13282 * mechanism to seamlessly deal with this won't work, so turn off warnings
13284 save_start = RExC_start;
13285 orig_end = RExC_end;
13287 RExC_parse = RExC_start = SvPVX(substitute_parse);
13288 RExC_end = RExC_parse + SvCUR(substitute_parse);
13289 TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE;
13291 *node_p = reg(pRExC_state, 1, &flags, depth+1);
13293 /* Restore the saved values */
13295 RExC_start = save_start;
13296 RExC_parse = endbrace;
13297 RExC_end = orig_end;
13298 SET_recode_x_to_native(0);
13300 SvREFCNT_dec_NN(substitute_parse);
13303 RETURN_FAIL_ON_RESTART(flags, flagp);
13304 FAIL2("panic: reg returned failure to grok_bslash_N, flags=%#" UVxf,
13307 *flagp |= flags&(HASWIDTH|SIMPLE|POSTPONED);
13309 nextchar(pRExC_state);
13316 S_compute_EXACTish(RExC_state_t *pRExC_state)
13320 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
13328 op = get_regex_charset(RExC_flags);
13329 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
13330 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
13331 been, so there is no hole */
13334 return op + EXACTF;
13338 S_new_regcurly(const char *s, const char *e)
13340 /* This is a temporary function designed to match the most lenient form of
13341 * a {m,n} quantifier we ever envision, with either number omitted, and
13342 * spaces anywhere between/before/after them.
13344 * If this function fails, then the string it matches is very unlikely to
13345 * ever be considered a valid quantifier, so we can allow the '{' that
13346 * begins it to be considered as a literal */
13348 bool has_min = FALSE;
13349 bool has_max = FALSE;
13351 PERL_ARGS_ASSERT_NEW_REGCURLY;
13353 if (s >= e || *s++ != '{')
13356 while (s < e && isSPACE(*s)) {
13359 while (s < e && isDIGIT(*s)) {
13363 while (s < e && isSPACE(*s)) {
13369 while (s < e && isSPACE(*s)) {
13372 while (s < e && isDIGIT(*s)) {
13376 while (s < e && isSPACE(*s)) {
13381 return s < e && *s == '}' && (has_min || has_max);
13384 /* Parse backref decimal value, unless it's too big to sensibly be a backref,
13385 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
13388 S_backref_value(char *p, char *e)
13390 const char* endptr = e;
13392 if (grok_atoUV(p, &val, &endptr) && val <= I32_MAX)
13399 - regatom - the lowest level
13401 Try to identify anything special at the start of the current parse position.
13402 If there is, then handle it as required. This may involve generating a
13403 single regop, such as for an assertion; or it may involve recursing, such as
13404 to handle a () structure.
13406 If the string doesn't start with something special then we gobble up
13407 as much literal text as we can. If we encounter a quantifier, we have to
13408 back off the final literal character, as that quantifier applies to just it
13409 and not to the whole string of literals.
13411 Once we have been able to handle whatever type of thing started the
13412 sequence, we return the offset into the regex engine program being compiled
13413 at which any next regnode should be placed.
13415 Returns 0, setting *flagp to TRYAGAIN if reg() returns 0 with TRYAGAIN.
13416 Returns 0, setting *flagp to RESTART_PARSE if the parse needs to be
13417 restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
13418 Otherwise does not return 0.
13420 Note: we have to be careful with escapes, as they can be both literal
13421 and special, and in the case of \10 and friends, context determines which.
13423 A summary of the code structure is:
13425 switch (first_byte) {
13426 cases for each special:
13427 handle this special;
13430 switch (2nd byte) {
13431 cases for each unambiguous special:
13432 handle this special;
13434 cases for each ambigous special/literal:
13436 if (special) handle here
13438 default: // unambiguously literal:
13441 default: // is a literal char
13444 create EXACTish node for literal;
13445 while (more input and node isn't full) {
13446 switch (input_byte) {
13447 cases for each special;
13448 make sure parse pointer is set so that the next call to
13449 regatom will see this special first
13450 goto loopdone; // EXACTish node terminated by prev. char
13452 append char to EXACTISH node;
13454 get next input byte;
13458 return the generated node;
13460 Specifically there are two separate switches for handling
13461 escape sequences, with the one for handling literal escapes requiring
13462 a dummy entry for all of the special escapes that are actually handled
13467 STATIC regnode_offset
13468 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
13470 regnode_offset ret = 0;
13476 DECLARE_AND_GET_RE_DEBUG_FLAGS;
13478 *flagp = 0; /* Initialize. */
13480 DEBUG_PARSE("atom");
13482 PERL_ARGS_ASSERT_REGATOM;
13485 parse_start = RExC_parse;
13486 assert(RExC_parse < RExC_end);
13487 switch ((U8)*RExC_parse) {
13489 RExC_seen_zerolen++;
13490 nextchar(pRExC_state);
13491 if (RExC_flags & RXf_PMf_MULTILINE)
13492 ret = reg_node(pRExC_state, MBOL);
13494 ret = reg_node(pRExC_state, SBOL);
13495 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13498 nextchar(pRExC_state);
13500 RExC_seen_zerolen++;
13501 if (RExC_flags & RXf_PMf_MULTILINE)
13502 ret = reg_node(pRExC_state, MEOL);
13504 ret = reg_node(pRExC_state, SEOL);
13505 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13508 nextchar(pRExC_state);
13509 if (RExC_flags & RXf_PMf_SINGLELINE)
13510 ret = reg_node(pRExC_state, SANY);
13512 ret = reg_node(pRExC_state, REG_ANY);
13513 *flagp |= HASWIDTH|SIMPLE;
13515 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13519 char * const oregcomp_parse = ++RExC_parse;
13520 ret = regclass(pRExC_state, flagp, depth+1,
13521 FALSE, /* means parse the whole char class */
13522 TRUE, /* allow multi-char folds */
13523 FALSE, /* don't silence non-portable warnings. */
13524 (bool) RExC_strict,
13525 TRUE, /* Allow an optimized regnode result */
13528 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13529 FAIL2("panic: regclass returned failure to regatom, flags=%#" UVxf,
13532 if (*RExC_parse != ']') {
13533 RExC_parse = oregcomp_parse;
13534 vFAIL("Unmatched [");
13536 nextchar(pRExC_state);
13537 Set_Node_Length(REGNODE_p(ret), RExC_parse - oregcomp_parse + 1); /* MJD */
13541 nextchar(pRExC_state);
13542 ret = reg(pRExC_state, 2, &flags, depth+1);
13544 if (flags & TRYAGAIN) {
13545 if (RExC_parse >= RExC_end) {
13546 /* Make parent create an empty node if needed. */
13547 *flagp |= TRYAGAIN;
13552 RETURN_FAIL_ON_RESTART(flags, flagp);
13553 FAIL2("panic: reg returned failure to regatom, flags=%#" UVxf,
13556 *flagp |= flags&(HASWIDTH|SIMPLE|POSTPONED);
13560 if (flags & TRYAGAIN) {
13561 *flagp |= TRYAGAIN;
13564 vFAIL("Internal urp");
13565 /* Supposed to be caught earlier. */
13571 vFAIL("Quantifier follows nothing");
13576 This switch handles escape sequences that resolve to some kind
13577 of special regop and not to literal text. Escape sequences that
13578 resolve to literal text are handled below in the switch marked
13581 Every entry in this switch *must* have a corresponding entry
13582 in the literal escape switch. However, the opposite is not
13583 required, as the default for this switch is to jump to the
13584 literal text handling code.
13587 switch ((U8)*RExC_parse) {
13588 /* Special Escapes */
13590 RExC_seen_zerolen++;
13591 /* Under wildcards, this is changed to match \n; should be
13592 * invisible to the user, as they have to compile under /m */
13593 if (RExC_pm_flags & PMf_WILDCARD) {
13594 ret = reg_node(pRExC_state, MBOL);
13597 ret = reg_node(pRExC_state, SBOL);
13598 /* SBOL is shared with /^/ so we set the flags so we can tell
13599 * /\A/ from /^/ in split. */
13600 FLAGS(REGNODE_p(ret)) = 1;
13602 goto finish_meta_pat;
13604 if (RExC_pm_flags & PMf_WILDCARD) {
13606 /* diag_listed_as: Use of %s is not allowed in Unicode property
13607 wildcard subpatterns in regex; marked by <-- HERE in m/%s/
13609 vFAIL("Use of '\\G' is not allowed in Unicode property"
13610 " wildcard subpatterns");
13612 ret = reg_node(pRExC_state, GPOS);
13613 RExC_seen |= REG_GPOS_SEEN;
13614 goto finish_meta_pat;
13616 if (!RExC_in_lookaround) {
13617 RExC_seen_zerolen++;
13618 ret = reg_node(pRExC_state, KEEPS);
13619 /* XXX:dmq : disabling in-place substitution seems to
13620 * be necessary here to avoid cases of memory corruption, as
13621 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
13623 RExC_seen |= REG_LOOKBEHIND_SEEN;
13624 goto finish_meta_pat;
13627 ++RExC_parse; /* advance past the 'K' */
13628 vFAIL("\\K not permitted in lookahead/lookbehind");
13631 if (RExC_pm_flags & PMf_WILDCARD) {
13632 /* See comment under \A above */
13633 ret = reg_node(pRExC_state, MEOL);
13636 ret = reg_node(pRExC_state, SEOL);
13638 RExC_seen_zerolen++; /* Do not optimize RE away */
13639 goto finish_meta_pat;
13641 if (RExC_pm_flags & PMf_WILDCARD) {
13642 /* See comment under \A above */
13643 ret = reg_node(pRExC_state, MEOL);
13646 ret = reg_node(pRExC_state, EOS);
13648 RExC_seen_zerolen++; /* Do not optimize RE away */
13649 goto finish_meta_pat;
13651 vFAIL("\\C no longer supported");
13653 ret = reg_node(pRExC_state, CLUMP);
13654 *flagp |= HASWIDTH;
13655 goto finish_meta_pat;
13663 regex_charset charset = get_regex_charset(RExC_flags);
13665 RExC_seen_zerolen++;
13666 RExC_seen |= REG_LOOKBEHIND_SEEN;
13667 op = BOUND + charset;
13669 if (RExC_parse >= RExC_end || *(RExC_parse + 1) != '{') {
13670 flags = TRADITIONAL_BOUND;
13671 if (op > BOUNDA) { /* /aa is same as /a */
13677 char name = *RExC_parse;
13678 char * endbrace = NULL;
13680 endbrace = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
13683 vFAIL2("Missing right brace on \\%c{}", name);
13685 /* XXX Need to decide whether to take spaces or not. Should be
13686 * consistent with \p{}, but that currently is SPACE, which
13687 * means vertical too, which seems wrong
13688 * while (isBLANK(*RExC_parse)) {
13691 if (endbrace == RExC_parse) {
13692 RExC_parse++; /* After the '}' */
13693 vFAIL2("Empty \\%c{}", name);
13695 length = endbrace - RExC_parse;
13696 /*while (isBLANK(*(RExC_parse + length - 1))) {
13699 switch (*RExC_parse) {
13702 && (memNEs(RExC_parse + 1, length - 1, "cb")))
13704 goto bad_bound_type;
13709 if (length != 2 || *(RExC_parse + 1) != 'b') {
13710 goto bad_bound_type;
13715 if (length != 2 || *(RExC_parse + 1) != 'b') {
13716 goto bad_bound_type;
13721 if (length != 2 || *(RExC_parse + 1) != 'b') {
13722 goto bad_bound_type;
13728 RExC_parse = endbrace;
13730 "'%" UTF8f "' is an unknown bound type",
13731 UTF8fARG(UTF, length, endbrace - length));
13732 NOT_REACHED; /*NOTREACHED*/
13734 RExC_parse = endbrace;
13735 REQUIRE_UNI_RULES(flagp, 0);
13740 else if (op >= BOUNDA) { /* /aa is same as /a */
13744 /* Don't have to worry about UTF-8, in this message because
13745 * to get here the contents of the \b must be ASCII */
13746 ckWARN4reg(RExC_parse + 1, /* Include the '}' in msg */
13747 "Using /u for '%.*s' instead of /%s",
13749 endbrace - length + 1,
13750 (charset == REGEX_ASCII_RESTRICTED_CHARSET)
13751 ? ASCII_RESTRICT_PAT_MODS
13752 : ASCII_MORE_RESTRICT_PAT_MODS);
13757 RExC_seen_d_op = TRUE;
13759 else if (op == BOUNDL) {
13760 RExC_contains_locale = 1;
13764 op += NBOUND - BOUND;
13767 ret = reg_node(pRExC_state, op);
13768 FLAGS(REGNODE_p(ret)) = flags;
13770 goto finish_meta_pat;
13774 ret = reg_node(pRExC_state, LNBREAK);
13775 *flagp |= HASWIDTH|SIMPLE;
13776 goto finish_meta_pat;
13790 /* These all have the same meaning inside [brackets], and it knows
13791 * how to do the best optimizations for them. So, pretend we found
13792 * these within brackets, and let it do the work */
13795 ret = regclass(pRExC_state, flagp, depth+1,
13796 TRUE, /* means just parse this element */
13797 FALSE, /* don't allow multi-char folds */
13798 FALSE, /* don't silence non-portable warnings. It
13799 would be a bug if these returned
13801 (bool) RExC_strict,
13802 TRUE, /* Allow an optimized regnode result */
13804 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13805 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
13806 * multi-char folds are allowed. */
13808 FAIL2("panic: regclass returned failure to regatom, flags=%#" UVxf,
13811 RExC_parse--; /* regclass() leaves this one too far ahead */
13814 /* The escapes above that don't take a parameter can't be
13815 * followed by a '{'. But 'pX', 'p{foo}' and
13816 * correspondingly 'P' can be */
13817 if ( RExC_parse - parse_start == 1
13818 && UCHARAT(RExC_parse + 1) == '{'
13819 && UNLIKELY(! new_regcurly(RExC_parse + 1, RExC_end)))
13822 vFAIL("Unescaped left brace in regex is illegal here");
13824 Set_Node_Offset(REGNODE_p(ret), parse_start);
13825 Set_Node_Length(REGNODE_p(ret), RExC_parse - parse_start + 1); /* MJD */
13826 nextchar(pRExC_state);
13829 /* Handle \N, \N{} and \N{NAMED SEQUENCE} (the latter meaning the
13830 * \N{...} evaluates to a sequence of more than one code points).
13831 * The function call below returns a regnode, which is our result.
13832 * The parameters cause it to fail if the \N{} evaluates to a
13833 * single code point; we handle those like any other literal. The
13834 * reason that the multicharacter case is handled here and not as
13835 * part of the EXACtish code is because of quantifiers. In
13836 * /\N{BLAH}+/, the '+' applies to the whole thing, and doing it
13837 * this way makes that Just Happen. dmq.
13838 * join_exact() will join this up with adjacent EXACTish nodes
13839 * later on, if appropriate. */
13841 if (grok_bslash_N(pRExC_state,
13842 &ret, /* Want a regnode returned */
13843 NULL, /* Fail if evaluates to a single code
13845 NULL, /* Don't need a count of how many code
13854 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13856 /* Here, evaluates to a single code point. Go get that */
13857 RExC_parse = parse_start;
13860 case 'k': /* Handle \k<NAME> and \k'NAME' */
13864 if ( RExC_parse >= RExC_end - 1
13865 || (( ch = RExC_parse[1]) != '<'
13870 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
13871 vFAIL2("Sequence %.2s... not terminated", parse_start);
13874 ret = handle_named_backref(pRExC_state,
13886 case '1': case '2': case '3': case '4':
13887 case '5': case '6': case '7': case '8': case '9':
13892 if (*RExC_parse == 'g') {
13896 if (*RExC_parse == '{') {
13900 if (*RExC_parse == '-') {
13904 if (hasbrace && !isDIGIT(*RExC_parse)) {
13905 if (isrel) RExC_parse--;
13907 goto parse_named_seq;
13910 if (RExC_parse >= RExC_end) {
13911 goto unterminated_g;
13913 num = S_backref_value(RExC_parse, RExC_end);
13915 vFAIL("Reference to invalid group 0");
13916 else if (num == I32_MAX) {
13917 if (isDIGIT(*RExC_parse))
13918 vFAIL("Reference to nonexistent group");
13921 vFAIL("Unterminated \\g... pattern");
13925 num = RExC_npar - num;
13927 vFAIL("Reference to nonexistent or unclosed group");
13931 num = S_backref_value(RExC_parse, RExC_end);
13932 /* bare \NNN might be backref or octal - if it is larger
13933 * than or equal RExC_npar then it is assumed to be an
13934 * octal escape. Note RExC_npar is +1 from the actual
13935 * number of parens. */
13936 /* Note we do NOT check if num == I32_MAX here, as that is
13937 * handled by the RExC_npar check */
13940 /* any numeric escape < 10 is always a backref */
13942 /* any numeric escape < RExC_npar is a backref */
13943 && num >= RExC_npar
13944 /* cannot be an octal escape if it starts with [89] */
13945 && ! inRANGE(*RExC_parse, '8', '9')
13947 /* Probably not meant to be a backref, instead likely
13948 * to be an octal character escape, e.g. \35 or \777.
13949 * The above logic should make it obvious why using
13950 * octal escapes in patterns is problematic. - Yves */
13951 RExC_parse = parse_start;
13956 /* At this point RExC_parse points at a numeric escape like
13957 * \12 or \88 or something similar, which we should NOT treat
13958 * as an octal escape. It may or may not be a valid backref
13959 * escape. For instance \88888888 is unlikely to be a valid
13961 while (isDIGIT(*RExC_parse))
13964 if (*RExC_parse != '}')
13965 vFAIL("Unterminated \\g{...} pattern");
13968 if (num >= (I32)RExC_npar) {
13970 /* It might be a forward reference; we can't fail until we
13971 * know, by completing the parse to get all the groups, and
13972 * then reparsing */
13973 if (ALL_PARENS_COUNTED) {
13974 if (num >= RExC_total_parens) {
13975 vFAIL("Reference to nonexistent group");
13979 REQUIRE_PARENS_PASS;
13983 ret = reganode(pRExC_state,
13986 : (ASCII_FOLD_RESTRICTED)
13988 : (AT_LEAST_UNI_SEMANTICS)
13994 if (OP(REGNODE_p(ret)) == REFF) {
13995 RExC_seen_d_op = TRUE;
13997 *flagp |= HASWIDTH;
13999 /* override incorrect value set in reganode MJD */
14000 Set_Node_Offset(REGNODE_p(ret), parse_start);
14001 Set_Node_Cur_Length(REGNODE_p(ret), parse_start-1);
14002 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14003 FALSE /* Don't force to /x */ );
14007 if (RExC_parse >= RExC_end)
14008 FAIL("Trailing \\");
14011 /* Do not generate "unrecognized" warnings here, we fall
14012 back into the quick-grab loop below */
14013 RExC_parse = parse_start;
14015 } /* end of switch on a \foo sequence */
14020 /* '#' comments should have been spaced over before this function was
14022 assert((RExC_flags & RXf_PMf_EXTENDED) == 0);
14024 if (RExC_flags & RXf_PMf_EXTENDED) {
14025 RExC_parse = reg_skipcomment( pRExC_state, RExC_parse );
14026 if (RExC_parse < RExC_end)
14036 /* Here, we have determined that the next thing is probably a
14037 * literal character. RExC_parse points to the first byte of its
14038 * definition. (It still may be an escape sequence that evaluates
14039 * to a single character) */
14044 char *s, *old_s = NULL, *old_old_s = NULL;
14046 U32 max_string_len = 255;
14048 /* We may have to reparse the node, artificially stopping filling
14049 * it early, based on info gleaned in the first parse. This
14050 * variable gives where we stop. Make it above the normal stopping
14051 * place first time through; otherwise it would stop too early */
14052 U32 upper_fill = max_string_len + 1;
14054 /* We start out as an EXACT node, even if under /i, until we find a
14055 * character which is in a fold. The algorithm now segregates into
14056 * separate nodes, characters that fold from those that don't under
14057 * /i. (This hopefully will create nodes that are fixed strings
14058 * even under /i, giving the optimizer something to grab on to.)
14059 * So, if a node has something in it and the next character is in
14060 * the opposite category, that node is closed up, and the function
14061 * returns. Then regatom is called again, and a new node is
14062 * created for the new category. */
14063 U8 node_type = EXACT;
14065 /* Assume the node will be fully used; the excess is given back at
14066 * the end. Under /i, we may need to temporarily add the fold of
14067 * an extra character or two at the end to check for splitting
14068 * multi-char folds, so allocate extra space for that. We can't
14069 * make any other length assumptions, as a byte input sequence
14070 * could shrink down. */
14071 Ptrdiff_t current_string_nodes = STR_SZ(max_string_len
14075 ? UTF8_MAXBYTES_CASE
14076 /* Max non-UTF-8 expansion is 2 */ : 2)));
14078 bool next_is_quantifier;
14079 char * oldp = NULL;
14081 /* We can convert EXACTF nodes to EXACTFU if they contain only
14082 * characters that match identically regardless of the target
14083 * string's UTF8ness. The reason to do this is that EXACTF is not
14084 * trie-able, EXACTFU is, and EXACTFU requires fewer operations at
14087 * Similarly, we can convert EXACTFL nodes to EXACTFLU8 if they
14088 * contain only above-Latin1 characters (hence must be in UTF8),
14089 * which don't participate in folds with Latin1-range characters,
14090 * as the latter's folds aren't known until runtime. */
14091 bool maybe_exactfu = FOLD && (DEPENDS_SEMANTICS || LOC);
14093 /* Single-character EXACTish nodes are almost always SIMPLE. This
14094 * allows us to override this as encountered */
14095 U8 maybe_SIMPLE = SIMPLE;
14097 /* Does this node contain something that can't match unless the
14098 * target string is (also) in UTF-8 */
14099 bool requires_utf8_target = FALSE;
14101 /* The sequence 'ss' is problematic in non-UTF-8 patterns. */
14102 bool has_ss = FALSE;
14104 /* So is the MICRO SIGN */
14105 bool has_micro_sign = FALSE;
14107 /* Set when we fill up the current node and there is still more
14108 * text to process */
14111 /* Allocate an EXACT node. The node_type may change below to
14112 * another EXACTish node, but since the size of the node doesn't
14113 * change, it works */
14114 ret = regnode_guts(pRExC_state, node_type, current_string_nodes,
14116 FILL_NODE(ret, node_type);
14119 s = STRING(REGNODE_p(ret));
14130 maybe_exactfu = FOLD && (DEPENDS_SEMANTICS || LOC);
14131 maybe_SIMPLE = SIMPLE;
14132 requires_utf8_target = FALSE;
14134 has_micro_sign = FALSE;
14138 /* This breaks under rare circumstances. If folding, we do not
14139 * want to split a node at a character that is a non-final in a
14140 * multi-char fold, as an input string could just happen to want to
14141 * match across the node boundary. The code at the end of the loop
14142 * looks for this, and backs off until it finds not such a
14143 * character, but it is possible (though extremely, extremely
14144 * unlikely) for all characters in the node to be non-final fold
14145 * ones, in which case we just leave the node fully filled, and
14146 * hope that it doesn't match the string in just the wrong place */
14148 assert( ! UTF /* Is at the beginning of a character */
14149 || UTF8_IS_INVARIANT(UCHARAT(RExC_parse))
14150 || UTF8_IS_START(UCHARAT(RExC_parse)));
14152 overflowed = FALSE;
14154 /* Here, we have a literal character. Find the maximal string of
14155 * them in the input that we can fit into a single EXACTish node.
14156 * We quit at the first non-literal or when the node gets full, or
14157 * under /i the categorization of folding/non-folding character
14159 while (p < RExC_end && len < upper_fill) {
14161 /* In most cases each iteration adds one byte to the output.
14162 * The exceptions override this */
14163 Size_t added_len = 1;
14169 /* White space has already been ignored */
14170 assert( (RExC_flags & RXf_PMf_EXTENDED) == 0
14171 || ! is_PATWS_safe((p), RExC_end, UTF));
14174 const char* message;
14187 /* Literal Escapes Switch
14189 This switch is meant to handle escape sequences that
14190 resolve to a literal character.
14192 Every escape sequence that represents something
14193 else, like an assertion or a char class, is handled
14194 in the switch marked 'Special Escapes' above in this
14195 routine, but also has an entry here as anything that
14196 isn't explicitly mentioned here will be treated as
14197 an unescaped equivalent literal.
14200 switch ((U8)*++p) {
14202 /* These are all the special escapes. */
14203 case 'A': /* Start assertion */
14204 case 'b': case 'B': /* Word-boundary assertion*/
14205 case 'C': /* Single char !DANGEROUS! */
14206 case 'd': case 'D': /* digit class */
14207 case 'g': case 'G': /* generic-backref, pos assertion */
14208 case 'h': case 'H': /* HORIZWS */
14209 case 'k': case 'K': /* named backref, keep marker */
14210 case 'p': case 'P': /* Unicode property */
14211 case 'R': /* LNBREAK */
14212 case 's': case 'S': /* space class */
14213 case 'v': case 'V': /* VERTWS */
14214 case 'w': case 'W': /* word class */
14215 case 'X': /* eXtended Unicode "combining
14216 character sequence" */
14217 case 'z': case 'Z': /* End of line/string assertion */
14221 /* Anything after here is an escape that resolves to a
14222 literal. (Except digits, which may or may not)
14228 case 'N': /* Handle a single-code point named character. */
14229 RExC_parse = p + 1;
14230 if (! grok_bslash_N(pRExC_state,
14231 NULL, /* Fail if evaluates to
14232 anything other than a
14233 single code point */
14234 &ender, /* The returned single code
14236 NULL, /* Don't need a count of
14237 how many code points */
14242 if (*flagp & NEED_UTF8)
14243 FAIL("panic: grok_bslash_N set NEED_UTF8");
14244 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
14246 /* Here, it wasn't a single code point. Go close
14247 * up this EXACTish node. The switch() prior to
14248 * this switch handles the other cases */
14249 RExC_parse = p = oldp;
14253 RExC_parse = parse_start;
14255 /* The \N{} means the pattern, if previously /d,
14256 * becomes /u. That means it can't be an EXACTF node,
14257 * but an EXACTFU */
14258 if (node_type == EXACTF) {
14259 node_type = EXACTFU;
14261 /* If the node already contains something that
14262 * differs between EXACTF and EXACTFU, reparse it
14264 if (! maybe_exactfu) {
14285 ender = ESC_NATIVE;
14293 if (! grok_bslash_o(&p,
14298 (bool) RExC_strict,
14299 FALSE, /* No illegal cp's */
14302 RExC_parse = p; /* going to die anyway; point to
14303 exact spot of failure */
14307 if (message && TO_OUTPUT_WARNINGS(p)) {
14308 warn_non_literal_string(p, packed_warn, message);
14312 if (! grok_bslash_x(&p,
14317 (bool) RExC_strict,
14318 FALSE, /* No illegal cp's */
14321 RExC_parse = p; /* going to die anyway; point
14322 to exact spot of failure */
14326 if (message && TO_OUTPUT_WARNINGS(p)) {
14327 warn_non_literal_string(p, packed_warn, message);
14331 if (ender < 0x100) {
14332 if (RExC_recode_x_to_native) {
14333 ender = LATIN1_TO_NATIVE(ender);
14340 if (! grok_bslash_c(*p, &grok_c_char,
14341 &message, &packed_warn))
14343 /* going to die anyway; point to exact spot of
14345 RExC_parse = p + ((UTF)
14346 ? UTF8_SAFE_SKIP(p, RExC_end)
14351 ender = grok_c_char;
14353 if (message && TO_OUTPUT_WARNINGS(p)) {
14354 warn_non_literal_string(p, packed_warn, message);
14358 case '8': case '9': /* must be a backreference */
14360 /* we have an escape like \8 which cannot be an octal escape
14361 * so we exit the loop, and let the outer loop handle this
14362 * escape which may or may not be a legitimate backref. */
14364 case '1': case '2': case '3':case '4':
14365 case '5': case '6': case '7':
14366 /* When we parse backslash escapes there is ambiguity
14367 * between backreferences and octal escapes. Any escape
14368 * from \1 - \9 is a backreference, any multi-digit
14369 * escape which does not start with 0 and which when
14370 * evaluated as decimal could refer to an already
14371 * parsed capture buffer is a back reference. Anything
14374 * Note this implies that \118 could be interpreted as
14375 * 118 OR as "\11" . "8" depending on whether there
14376 * were 118 capture buffers defined already in the
14379 /* NOTE, RExC_npar is 1 more than the actual number of
14380 * parens we have seen so far, hence the "<" as opposed
14382 if ( !isDIGIT(p[1]) || S_backref_value(p, RExC_end) < RExC_npar)
14383 { /* Not to be treated as an octal constant, go
14391 I32 flags = PERL_SCAN_SILENT_ILLDIGIT
14392 | PERL_SCAN_NOTIFY_ILLDIGIT;
14394 ender = grok_oct(p, &numlen, &flags, NULL);
14396 if ( (flags & PERL_SCAN_NOTIFY_ILLDIGIT)
14397 && isDIGIT(*p) /* like \08, \178 */
14398 && ckWARN(WARN_REGEXP))
14400 reg_warn_non_literal_string(
14402 form_alien_digit_msg(8, numlen, p,
14403 RExC_end, UTF, FALSE));
14409 FAIL("Trailing \\");
14412 if (isALPHANUMERIC(*p)) {
14413 /* An alpha followed by '{' is going to fail next
14414 * iteration, so don't output this warning in that
14416 if (! isALPHA(*p) || *(p + 1) != '{') {
14417 ckWARN2reg(p + 1, "Unrecognized escape \\%.1s"
14418 " passed through", p);
14421 goto normal_default;
14422 } /* End of switch on '\' */
14425 /* Trying to gain new uses for '{' without breaking too
14426 * much existing code is hard. The solution currently
14428 * 1) If there is no ambiguity that a '{' should always
14429 * be taken literally, at the start of a construct, we
14431 * 2) If the literal '{' conflicts with our desired use
14432 * of it as a metacharacter, we die. The deprecation
14433 * cycles for this have come and gone.
14434 * 3) If there is ambiguity, we raise a simple warning.
14435 * This could happen, for example, if the user
14436 * intended it to introduce a quantifier, but slightly
14437 * misspelled the quantifier. Without this warning,
14438 * the quantifier would silently be taken as a literal
14439 * string of characters instead of a meta construct */
14440 if (len || (p > RExC_start && isALPHA_A(*(p - 1)))) {
14442 || ( p > parse_start + 1
14443 && isALPHA_A(*(p - 1))
14444 && *(p - 2) == '\\')
14445 || new_regcurly(p, RExC_end))
14447 RExC_parse = p + 1;
14448 vFAIL("Unescaped left brace in regex is "
14451 ckWARNreg(p + 1, "Unescaped left brace in regex is"
14452 " passed through");
14454 goto normal_default;
14457 if (p > RExC_parse && RExC_strict) {
14458 ckWARN2reg(p + 1, "Unescaped literal '%c'", *p);
14461 default: /* A literal character */
14463 if (! UTF8_IS_INVARIANT(*p) && UTF) {
14465 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
14466 &numlen, UTF8_ALLOW_DEFAULT);
14472 } /* End of switch on the literal */
14474 /* Here, have looked at the literal character, and <ender>
14475 * contains its ordinal; <p> points to the character after it.
14479 REQUIRE_UTF8(flagp);
14480 if ( UNICODE_IS_PERL_EXTENDED(ender)
14481 && TO_OUTPUT_WARNINGS(p))
14483 ckWARN2_non_literal_string(p,
14484 packWARN(WARN_PORTABLE),
14485 PL_extended_cp_format,
14490 /* We need to check if the next non-ignored thing is a
14491 * quantifier. Move <p> to after anything that should be
14492 * ignored, which, as a side effect, positions <p> for the next
14493 * loop iteration */
14494 skip_to_be_ignored_text(pRExC_state, &p,
14495 FALSE /* Don't force to /x */ );
14497 /* If the next thing is a quantifier, it applies to this
14498 * character only, which means that this character has to be in
14499 * its own node and can't just be appended to the string in an
14500 * existing node, so if there are already other characters in
14501 * the node, close the node with just them, and set up to do
14502 * this character again next time through, when it will be the
14503 * only thing in its new node */
14505 next_is_quantifier = LIKELY(p < RExC_end)
14506 && UNLIKELY(ISMULT2(p));
14508 if (next_is_quantifier && LIKELY(len)) {
14513 /* Ready to add 'ender' to the node */
14515 if (! FOLD) { /* The simple case, just append the literal */
14518 /* Don't output if it would overflow */
14519 if (UNLIKELY(len > max_string_len - ((UTF)
14520 ? UVCHR_SKIP(ender)
14527 if (UVCHR_IS_INVARIANT(ender) || ! UTF) {
14528 *(s++) = (char) ender;
14531 U8 * new_s = uvchr_to_utf8((U8*)s, ender);
14532 added_len = (char *) new_s - s;
14533 s = (char *) new_s;
14536 requires_utf8_target = TRUE;
14540 else if (LOC && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)) {
14542 /* Here are folding under /l, and the code point is
14543 * problematic. If this is the first character in the
14544 * node, change the node type to folding. Otherwise, if
14545 * this is the first problematic character, close up the
14546 * existing node, so can start a new node with this one */
14548 node_type = EXACTFL;
14549 RExC_contains_locale = 1;
14551 else if (node_type == EXACT) {
14556 /* This problematic code point means we can't simplify
14558 maybe_exactfu = FALSE;
14560 /* Although these two characters have folds that are
14561 * locale-problematic, they also have folds to above Latin1
14562 * that aren't a problem. Doing these now helps at
14564 if (UNLIKELY( ender == GREEK_CAPITAL_LETTER_MU
14565 || ender == LATIN_CAPITAL_LETTER_SHARP_S))
14570 /* Here, we are adding a problematic fold character.
14571 * "Problematic" in this context means that its fold isn't
14572 * known until runtime. (The non-problematic code points
14573 * are the above-Latin1 ones that fold to also all
14574 * above-Latin1. Their folds don't vary no matter what the
14575 * locale is.) But here we have characters whose fold
14576 * depends on the locale. We just add in the unfolded
14577 * character, and wait until runtime to fold it */
14578 goto not_fold_common;
14580 else /* regular fold; see if actually is in a fold */
14581 if ( (ender < 256 && ! IS_IN_SOME_FOLD_L1(ender))
14583 && ! _invlist_contains_cp(PL_in_some_fold, ender)))
14585 /* Here, folding, but the character isn't in a fold.
14587 * Start a new node if previous characters in the node were
14589 if (len && node_type != EXACT) {
14594 /* Here, continuing a node with non-folded characters. Add
14596 goto not_fold_common;
14598 else { /* Here, does participate in some fold */
14600 /* If this is the first character in the node, change its
14601 * type to folding. Otherwise, if this is the first
14602 * folding character in the node, close up the existing
14603 * node, so can start a new node with this one. */
14605 node_type = compute_EXACTish(pRExC_state);
14607 else if (node_type == EXACT) {
14612 if (UTF) { /* Alway use the folded value for UTF-8
14614 if (UVCHR_IS_INVARIANT(ender)) {
14615 if (UNLIKELY(len + 1 > max_string_len)) {
14620 *(s)++ = (U8) toFOLD(ender);
14626 folded = _to_uni_fold_flags(
14628 (U8 *) s, /* We have allocated extra space
14629 in 's' so can't run off the
14633 | (( ASCII_FOLD_RESTRICTED
14634 || node_type == EXACTFL)
14635 ? FOLD_FLAGS_NOMIX_ASCII
14637 if (UNLIKELY(len + added_len > max_string_len)) {
14645 && LIKELY(folded != GREEK_SMALL_LETTER_MU))
14647 /* U+B5 folds to the MU, so its possible for a
14648 * non-UTF-8 target to match it */
14649 requires_utf8_target = TRUE;
14653 else { /* Here is non-UTF8. */
14655 /* The fold will be one or (rarely) two characters.
14656 * Check that there's room for at least a single one
14657 * before setting any flags, etc. Because otherwise an
14658 * overflowing character could cause a flag to be set
14659 * even though it doesn't end up in this node. (For
14660 * the two character fold, we check again, before
14661 * setting any flags) */
14662 if (UNLIKELY(len + 1 > max_string_len)) {
14667 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
14668 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
14669 || UNICODE_DOT_DOT_VERSION > 0)
14671 /* On non-ancient Unicodes, check for the only possible
14672 * multi-char fold */
14673 if (UNLIKELY(ender == LATIN_SMALL_LETTER_SHARP_S)) {
14675 /* This potential multi-char fold means the node
14676 * can't be simple (because it could match more
14677 * than a single char). And in some cases it will
14678 * match 'ss', so set that flag */
14682 /* It can't change to be an EXACTFU (unless already
14683 * is one). We fold it iff under /u rules. */
14684 if (node_type != EXACTFU) {
14685 maybe_exactfu = FALSE;
14688 if (UNLIKELY(len + 2 > max_string_len)) {
14697 goto done_with_this_char;
14700 else if ( UNLIKELY(isALPHA_FOLD_EQ(ender, 's'))
14702 && UNLIKELY(isALPHA_FOLD_EQ(*(s-1), 's')))
14704 /* Also, the sequence 'ss' is special when not
14705 * under /u. If the target string is UTF-8, it
14706 * should match SHARP S; otherwise it won't. So,
14707 * here we have to exclude the possibility of this
14708 * node moving to /u.*/
14710 maybe_exactfu = FALSE;
14713 /* Here, the fold will be a single character */
14715 if (UNLIKELY(ender == MICRO_SIGN)) {
14716 has_micro_sign = TRUE;
14718 else if (PL_fold[ender] != PL_fold_latin1[ender]) {
14720 /* If the character's fold differs between /d and
14721 * /u, this can't change to be an EXACTFU node */
14722 maybe_exactfu = FALSE;
14725 *(s++) = (DEPENDS_SEMANTICS)
14726 ? (char) toFOLD(ender)
14728 /* Under /u, the fold of any character in
14729 * the 0-255 range happens to be its
14730 * lowercase equivalent, except for LATIN
14731 * SMALL LETTER SHARP S, which was handled
14732 * above, and the MICRO SIGN, whose fold
14733 * requires UTF-8 to represent. */
14734 : (char) toLOWER_L1(ender);
14736 } /* End of adding current character to the node */
14738 done_with_this_char:
14742 if (next_is_quantifier) {
14744 /* Here, the next input is a quantifier, and to get here,
14745 * the current character is the only one in the node. */
14749 } /* End of loop through literal characters */
14751 /* Here we have either exhausted the input or run out of room in
14752 * the node. If the former, we are done. (If we encountered a
14753 * character that can't be in the node, transfer is made directly
14754 * to <loopdone>, and so we wouldn't have fallen off the end of the
14756 if (LIKELY(! overflowed)) {
14760 /* Here we have run out of room. We can grow plain EXACT and
14761 * LEXACT nodes. If the pattern is gigantic enough, though,
14762 * eventually we'll have to artificially chunk the pattern into
14763 * multiple nodes. */
14764 if (! LOC && (node_type == EXACT || node_type == LEXACT)) {
14765 Size_t overhead = 1 + regarglen[OP(REGNODE_p(ret))];
14766 Size_t overhead_expansion = 0;
14768 Size_t max_nodes_for_string;
14772 /* Here we couldn't fit the final character in the current
14773 * node, so it will have to be reparsed, no matter what else we
14777 /* If would have overflowed a regular EXACT node, switch
14778 * instead to an LEXACT. The code below is structured so that
14779 * the actual growing code is common to changing from an EXACT
14780 * or just increasing the LEXACT size. This means that we have
14781 * to save the string in the EXACT case before growing, and
14782 * then copy it afterwards to its new location */
14783 if (node_type == EXACT) {
14784 overhead_expansion = regarglen[LEXACT] - regarglen[EXACT];
14785 RExC_emit += overhead_expansion;
14786 Copy(s0, temp, len, char);
14789 /* Ready to grow. If it was a plain EXACT, the string was
14790 * saved, and the first few bytes of it overwritten by adding
14791 * an argument field. We assume, as we do elsewhere in this
14792 * file, that one byte of remaining input will translate into
14793 * one byte of output, and if that's too small, we grow again,
14794 * if too large the excess memory is freed at the end */
14796 max_nodes_for_string = U16_MAX - overhead - overhead_expansion;
14797 achievable = MIN(max_nodes_for_string,
14798 current_string_nodes + STR_SZ(RExC_end - p));
14799 delta = achievable - current_string_nodes;
14801 /* If there is just no more room, go finish up this chunk of
14807 change_engine_size(pRExC_state, delta + overhead_expansion);
14808 current_string_nodes += delta;
14810 = sizeof(struct regnode) * current_string_nodes;
14811 upper_fill = max_string_len + 1;
14813 /* If the length was small, we know this was originally an
14814 * EXACT node now converted to LEXACT, and the string has to be
14815 * restored. Otherwise the string was untouched. 260 is just
14816 * a number safely above 255 so don't have to worry about
14817 * getting it precise */
14819 node_type = LEXACT;
14820 FILL_NODE(ret, node_type);
14821 s0 = STRING(REGNODE_p(ret));
14822 Copy(temp, s0, len, char);
14826 goto continue_parse;
14829 bool splittable = FALSE;
14830 bool backed_up = FALSE;
14831 char * e; /* should this be U8? */
14832 char * s_start; /* should this be U8? */
14834 /* Here is /i. Running out of room creates a problem if we are
14835 * folding, and the split happens in the middle of a
14836 * multi-character fold, as a match that should have occurred,
14837 * won't, due to the way nodes are matched, and our artificial
14838 * boundary. So back off until we aren't splitting such a
14839 * fold. If there is no such place to back off to, we end up
14840 * taking the entire node as-is. This can happen if the node
14841 * consists entirely of 'f' or entirely of 's' characters (or
14842 * things that fold to them) as 'ff' and 'ss' are
14843 * multi-character folds.
14845 * The Unicode standard says that multi character folds consist
14846 * of either two or three characters. That means we would be
14847 * splitting one if the final character in the node is at the
14848 * beginning of either type, or is the second of a three
14852 * ender is the code point of the character that won't fit
14854 * s points to just beyond the final byte in the node.
14855 * It's where we would place ender if there were
14856 * room, and where in fact we do place ender's fold
14857 * in the code below, as we've over-allocated space
14858 * for s0 (hence s) to allow for this
14859 * e starts at 's' and advances as we append things.
14860 * old_s is the same as 's'. (If ender had fit, 's' would
14861 * have been advanced to beyond it).
14862 * old_old_s points to the beginning byte of the final
14863 * character in the node
14864 * p points to the beginning byte in the input of the
14865 * character beyond 'ender'.
14866 * oldp points to the beginning byte in the input of
14869 * In the case of /il, we haven't folded anything that could be
14870 * affected by the locale. That means only above-Latin1
14871 * characters that fold to other above-latin1 characters get
14872 * folded at compile time. To check where a good place to
14873 * split nodes is, everything in it will have to be folded.
14874 * The boolean 'maybe_exactfu' keeps track in /il if there are
14875 * any unfolded characters in the node. */
14876 bool need_to_fold_loc = LOC && ! maybe_exactfu;
14878 /* If we do need to fold the node, we need a place to store the
14879 * folded copy, and a way to map back to the unfolded original
14881 char * locfold_buf = NULL;
14882 Size_t * loc_correspondence = NULL;
14884 if (! need_to_fold_loc) { /* The normal case. Just
14885 initialize to the actual node */
14888 s = old_old_s; /* Point to the beginning of the final char
14889 that fits in the node */
14893 /* Here, we have filled a /il node, and there are unfolded
14894 * characters in it. If the runtime locale turns out to be
14895 * UTF-8, there are possible multi-character folds, just
14896 * like when not under /l. The node hence can't terminate
14897 * in the middle of such a fold. To determine this, we
14898 * have to create a folded copy of this node. That means
14899 * reparsing the node, folding everything assuming a UTF-8
14900 * locale. (If at runtime it isn't such a locale, the
14901 * actions here wouldn't have been necessary, but we have
14902 * to assume the worst case.) If we find we need to back
14903 * off the folded string, we do so, and then map that
14904 * position back to the original unfolded node, which then
14905 * gets output, truncated at that spot */
14907 char * redo_p = RExC_parse;
14911 /* Allow enough space assuming a single byte input folds to
14912 * a single byte output, plus assume that the two unparsed
14913 * characters (that we may need) fold to the largest number
14914 * of bytes possible, plus extra for one more worst case
14915 * scenario. In the loop below, if we start eating into
14916 * that final spare space, we enlarge this initial space */
14917 Size_t size = max_string_len + (3 * UTF8_MAXBYTES_CASE) + 1;
14919 Newxz(locfold_buf, size, char);
14920 Newxz(loc_correspondence, size, Size_t);
14922 /* Redo this node's parse, folding into 'locfold_buf' */
14923 redo_p = RExC_parse;
14924 old_redo_e = redo_e = locfold_buf;
14925 while (redo_p <= oldp) {
14927 old_redo_e = redo_e;
14928 loc_correspondence[redo_e - locfold_buf]
14929 = redo_p - RExC_parse;
14934 (void) _to_utf8_fold_flags((U8 *) redo_p,
14939 redo_e += added_len;
14940 redo_p += UTF8SKIP(redo_p);
14944 /* Note that if this code is run on some ancient
14945 * Unicode versions, SHARP S doesn't fold to 'ss',
14946 * but rather than clutter the code with #ifdef's,
14947 * as is done above, we ignore that possibility.
14948 * This is ok because this code doesn't affect what
14949 * gets matched, but merely where the node gets
14951 if (UCHARAT(redo_p) != LATIN_SMALL_LETTER_SHARP_S) {
14952 *redo_e++ = toLOWER_L1(UCHARAT(redo_p));
14962 /* If we're getting so close to the end that a
14963 * worst-case fold in the next character would cause us
14964 * to overflow, increase, assuming one byte output byte
14965 * per one byte input one, plus room for another worst
14967 if ( redo_p <= oldp
14968 && redo_e > locfold_buf + size
14969 - (UTF8_MAXBYTES_CASE + 1))
14971 Size_t new_size = size
14973 + UTF8_MAXBYTES_CASE + 1;
14974 Ptrdiff_t e_offset = redo_e - locfold_buf;
14976 Renew(locfold_buf, new_size, char);
14977 Renew(loc_correspondence, new_size, Size_t);
14980 redo_e = locfold_buf + e_offset;
14984 /* Set so that things are in terms of the folded, temporary
14987 s_start = locfold_buf;
14992 /* Here, we have 's', 's_start' and 'e' set up to point to the
14993 * input that goes into the node, folded.
14995 * If the final character of the node and the fold of ender
14996 * form the first two characters of a three character fold, we
14997 * need to peek ahead at the next (unparsed) character in the
14998 * input to determine if the three actually do form such a
14999 * fold. Just looking at that character is not generally
15000 * sufficient, as it could be, for example, an escape sequence
15001 * that evaluates to something else, and it needs to be folded.
15003 * khw originally thought to just go through the parse loop one
15004 * extra time, but that doesn't work easily as that iteration
15005 * could cause things to think that the parse is over and to
15006 * goto loopdone. The character could be a '$' for example, or
15007 * the character beyond could be a quantifier, and other
15008 * glitches as well.
15010 * The solution used here for peeking ahead is to look at that
15011 * next character. If it isn't ASCII punctuation, then it will
15012 * be something that would continue on in an EXACTish node if
15013 * there were space. We append the fold of it to s, having
15014 * reserved enough room in s0 for the purpose. If we can't
15015 * reasonably peek ahead, we instead assume the worst case:
15016 * that it is something that would form the completion of a
15019 * If we can't split between s and ender, we work backwards
15020 * character-by-character down to s0. At each current point
15021 * see if we are at the beginning of a multi-char fold. If so,
15022 * that means we would be splitting the fold across nodes, and
15023 * so we back up one and try again.
15025 * If we're not at the beginning, we still could be at the
15026 * final two characters of a (rare) three character fold. We
15027 * check if the sequence starting at the character before the
15028 * current position (and including the current and next
15029 * characters) is a three character fold. If not, the node can
15030 * be split here. If it is, we have to backup two characters
15033 * Otherwise, the node can be split at the current position.
15035 * The same logic is used for UTF-8 patterns and not */
15039 /* Append the fold of ender */
15040 (void) _to_uni_fold_flags(
15044 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
15045 ? FOLD_FLAGS_NOMIX_ASCII
15049 /* 's' and the character folded to by ender may be the
15050 * first two of a three-character fold, in which case the
15051 * node should not be split here. That may mean examining
15052 * the so-far unparsed character starting at 'p'. But if
15053 * ender folded to more than one character, we already have
15054 * three characters to look at. Also, we first check if
15055 * the sequence consisting of s and the next character form
15056 * the first two of some three character fold. If not,
15057 * there's no need to peek ahead. */
15058 if ( added_len <= UTF8SKIP(e - added_len)
15059 && UNLIKELY(is_THREE_CHAR_FOLD_HEAD_utf8_safe(s, e)))
15061 /* Here, the two do form the beginning of a potential
15062 * three character fold. The unexamined character may
15063 * or may not complete it. Peek at it. It might be
15064 * something that ends the node or an escape sequence,
15065 * in which case we don't know without a lot of work
15066 * what it evaluates to, so we have to assume the worst
15067 * case: that it does complete the fold, and so we
15068 * can't split here. All such instances will have
15069 * that character be an ASCII punctuation character,
15070 * like a backslash. So, for that case, backup one and
15071 * drop down to try at that position */
15073 s = (char *) utf8_hop_back((U8 *) s, -1,
15078 /* Here, since it's not punctuation, it must be a
15079 * real character, and we can append its fold to
15080 * 'e' (having deliberately reserved enough space
15081 * for this eventuality) and drop down to check if
15082 * the three actually do form a folded sequence */
15083 (void) _to_utf8_fold_flags(
15084 (U8 *) p, (U8 *) RExC_end,
15087 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
15088 ? FOLD_FLAGS_NOMIX_ASCII
15094 /* Here, we either have three characters available in
15095 * sequence starting at 's', or we have two characters and
15096 * know that the following one can't possibly be part of a
15097 * three character fold. We go through the node backwards
15098 * until we find a place where we can split it without
15099 * breaking apart a multi-character fold. At any given
15100 * point we have to worry about if such a fold begins at
15101 * the current 's', and also if a three-character fold
15102 * begins at s-1, (containing s and s+1). Splitting in
15103 * either case would break apart a fold */
15105 char *prev_s = (char *) utf8_hop_back((U8 *) s, -1,
15108 /* If is a multi-char fold, can't split here. Backup
15109 * one char and try again */
15110 if (UNLIKELY(is_MULTI_CHAR_FOLD_utf8_safe(s, e))) {
15116 /* If the two characters beginning at 's' are part of a
15117 * three character fold starting at the character
15118 * before s, we can't split either before or after s.
15119 * Backup two chars and try again */
15120 if ( LIKELY(s > s_start)
15121 && UNLIKELY(is_THREE_CHAR_FOLD_utf8_safe(prev_s, e)))
15124 s = (char *) utf8_hop_back((U8 *) s, -1, (U8 *) s_start);
15129 /* Here there's no multi-char fold between s and the
15130 * next character following it. We can split */
15134 } while (s > s_start); /* End of loops backing up through the node */
15136 /* Here we either couldn't find a place to split the node,
15137 * or else we broke out of the loop setting 'splittable' to
15138 * true. In the latter case, the place to split is between
15139 * the first and second characters in the sequence starting
15145 else { /* Pattern not UTF-8 */
15146 if ( ender != LATIN_SMALL_LETTER_SHARP_S
15147 || ASCII_FOLD_RESTRICTED)
15149 assert( toLOWER_L1(ender) < 256 );
15150 *e++ = (char)(toLOWER_L1(ender)); /* should e and the cast be U8? */
15158 && UNLIKELY(is_THREE_CHAR_FOLD_HEAD_latin1_safe(s, e)))
15165 if ( UCHARAT(p) != LATIN_SMALL_LETTER_SHARP_S
15166 || ASCII_FOLD_RESTRICTED)
15168 assert( toLOWER_L1(ender) < 256 );
15169 *e++ = (char)(toLOWER_L1(ender)); /* should e and the cast be U8? */
15179 if (UNLIKELY(is_MULTI_CHAR_FOLD_latin1_safe(s, e))) {
15185 if ( LIKELY(s > s_start)
15186 && UNLIKELY(is_THREE_CHAR_FOLD_latin1_safe(s - 1, e)))
15196 } while (s > s_start);
15203 /* Here, we are done backing up. If we didn't backup at all
15204 * (the likely case), just proceed */
15207 /* If we did find a place to split, reparse the entire node
15208 * stopping where we have calculated. */
15211 /* If we created a temporary folded string under /l, we
15212 * have to map that back to the original */
15213 if (need_to_fold_loc) {
15214 upper_fill = loc_correspondence[s - s_start];
15215 if (upper_fill == 0) {
15216 FAIL2("panic: loc_correspondence[%d] is 0",
15217 (int) (s - s_start));
15219 Safefree(locfold_buf);
15220 Safefree(loc_correspondence);
15223 upper_fill = s - s0;
15228 /* Here the node consists entirely of non-final multi-char
15229 * folds. (Likely it is all 'f's or all 's's.) There's no
15230 * decent place to split it, so give up and just take the
15235 if (need_to_fold_loc) {
15236 Safefree(locfold_buf);
15237 Safefree(loc_correspondence);
15239 } /* End of verifying node ends with an appropriate char */
15241 /* We need to start the next node at the character that didn't fit
15245 loopdone: /* Jumped to when encounters something that shouldn't be
15248 /* Free up any over-allocated space; cast is to silence bogus
15249 * warning in MS VC */
15250 change_engine_size(pRExC_state,
15251 - (Ptrdiff_t) (current_string_nodes - STR_SZ(len)));
15253 /* I (khw) don't know if you can get here with zero length, but the
15254 * old code handled this situation by creating a zero-length EXACT
15255 * node. Might as well be NOTHING instead */
15257 OP(REGNODE_p(ret)) = NOTHING;
15261 /* If the node type is EXACT here, check to see if it
15262 * should be EXACTL, or EXACT_REQ8. */
15263 if (node_type == EXACT) {
15265 node_type = EXACTL;
15267 else if (requires_utf8_target) {
15268 node_type = EXACT_REQ8;
15271 else if (node_type == LEXACT) {
15272 if (requires_utf8_target) {
15273 node_type = LEXACT_REQ8;
15277 if ( UNLIKELY(has_micro_sign || has_ss)
15278 && (node_type == EXACTFU || ( node_type == EXACTF
15279 && maybe_exactfu)))
15280 { /* These two conditions are problematic in non-UTF-8
15283 node_type = EXACTFUP;
15285 else if (node_type == EXACTFL) {
15287 /* 'maybe_exactfu' is deliberately set above to
15288 * indicate this node type, where all code points in it
15290 if (maybe_exactfu) {
15291 node_type = EXACTFLU8;
15294 _invlist_contains_cp(PL_HasMultiCharFold, ender)))
15296 /* A character that folds to more than one will
15297 * match multiple characters, so can't be SIMPLE.
15298 * We don't have to worry about this with EXACTFLU8
15299 * nodes just above, as they have already been
15300 * folded (since the fold doesn't vary at run
15301 * time). Here, if the final character in the node
15302 * folds to multiple, it can't be simple. (This
15303 * only has an effect if the node has only a single
15304 * character, hence the final one, as elsewhere we
15305 * turn off simple for nodes whose length > 1 */
15309 else if (node_type == EXACTF) { /* Means is /di */
15311 /* This intermediate variable is needed solely because
15312 * the asserts in the macro where used exceed Win32's
15313 * literal string capacity */
15314 char first_char = * STRING(REGNODE_p(ret));
15316 /* If 'maybe_exactfu' is clear, then we need to stay
15317 * /di. If it is set, it means there are no code
15318 * points that match differently depending on UTF8ness
15319 * of the target string, so it can become an EXACTFU
15321 if (! maybe_exactfu) {
15322 RExC_seen_d_op = TRUE;
15324 else if ( isALPHA_FOLD_EQ(first_char, 's')
15325 || isALPHA_FOLD_EQ(ender, 's'))
15327 /* But, if the node begins or ends in an 's' we
15328 * have to defer changing it into an EXACTFU, as
15329 * the node could later get joined with another one
15330 * that ends or begins with 's' creating an 'ss'
15331 * sequence which would then wrongly match the
15332 * sharp s without the target being UTF-8. We
15333 * create a special node that we resolve later when
15334 * we join nodes together */
15336 node_type = EXACTFU_S_EDGE;
15339 node_type = EXACTFU;
15343 if (requires_utf8_target && node_type == EXACTFU) {
15344 node_type = EXACTFU_REQ8;
15348 OP(REGNODE_p(ret)) = node_type;
15349 setSTR_LEN(REGNODE_p(ret), len);
15350 RExC_emit += STR_SZ(len);
15352 /* If the node isn't a single character, it can't be SIMPLE */
15353 if (len > (Size_t) ((UTF) ? UTF8SKIP(STRING(REGNODE_p(ret))) : 1)) {
15357 *flagp |= HASWIDTH | maybe_SIMPLE;
15360 Set_Node_Length(REGNODE_p(ret), p - parse_start - 1);
15364 /* len is STRLEN which is unsigned, need to copy to signed */
15367 vFAIL("Internal disaster");
15370 } /* End of label 'defchar:' */
15372 } /* End of giant switch on input character */
15374 /* Position parse to next real character */
15375 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
15376 FALSE /* Don't force to /x */ );
15377 if ( *RExC_parse == '{'
15378 && OP(REGNODE_p(ret)) != SBOL && ! regcurly(RExC_parse))
15380 if (RExC_strict || new_regcurly(RExC_parse, RExC_end)) {
15382 vFAIL("Unescaped left brace in regex is illegal here");
15384 ckWARNreg(RExC_parse + 1, "Unescaped left brace in regex is"
15385 " passed through");
15393 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
15395 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
15396 * sets up the bitmap and any flags, removing those code points from the
15397 * inversion list, setting it to NULL should it become completely empty */
15400 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
15401 assert(PL_regkind[OP(node)] == ANYOF);
15403 /* There is no bitmap for this node type */
15404 if (inRANGE(OP(node), ANYOFH, ANYOFRb)) {
15408 ANYOF_BITMAP_ZERO(node);
15409 if (*invlist_ptr) {
15411 /* This gets set if we actually need to modify things */
15412 bool change_invlist = FALSE;
15416 /* Start looking through *invlist_ptr */
15417 invlist_iterinit(*invlist_ptr);
15418 while (invlist_iternext(*invlist_ptr, &start, &end)) {
15422 if (end == UV_MAX && start <= NUM_ANYOF_CODE_POINTS) {
15423 ANYOF_FLAGS(node) |= ANYOF_MATCHES_ALL_ABOVE_BITMAP;
15426 /* Quit if are above what we should change */
15427 if (start >= NUM_ANYOF_CODE_POINTS) {
15431 change_invlist = TRUE;
15433 /* Set all the bits in the range, up to the max that we are doing */
15434 high = (end < NUM_ANYOF_CODE_POINTS - 1)
15436 : NUM_ANYOF_CODE_POINTS - 1;
15437 for (i = start; i <= (int) high; i++) {
15438 ANYOF_BITMAP_SET(node, i);
15441 invlist_iterfinish(*invlist_ptr);
15443 /* Done with loop; remove any code points that are in the bitmap from
15444 * *invlist_ptr; similarly for code points above the bitmap if we have
15445 * a flag to match all of them anyways */
15446 if (change_invlist) {
15447 _invlist_subtract(*invlist_ptr, PL_InBitmap, invlist_ptr);
15449 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
15450 _invlist_intersection(*invlist_ptr, PL_InBitmap, invlist_ptr);
15453 /* If have completely emptied it, remove it completely */
15454 if (_invlist_len(*invlist_ptr) == 0) {
15455 SvREFCNT_dec_NN(*invlist_ptr);
15456 *invlist_ptr = NULL;
15461 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
15462 Character classes ([:foo:]) can also be negated ([:^foo:]).
15463 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
15464 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
15465 but trigger failures because they are currently unimplemented. */
15467 #define POSIXCC_DONE(c) ((c) == ':')
15468 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
15469 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
15470 #define MAYBE_POSIXCC(c) (POSIXCC(c) || (c) == '^' || (c) == ';')
15472 #define WARNING_PREFIX "Assuming NOT a POSIX class since "
15473 #define NO_BLANKS_POSIX_WARNING "no blanks are allowed in one"
15474 #define SEMI_COLON_POSIX_WARNING "a semi-colon was found instead of a colon"
15476 #define NOT_MEANT_TO_BE_A_POSIX_CLASS (OOB_NAMEDCLASS - 1)
15478 /* 'posix_warnings' and 'warn_text' are names of variables in the following
15480 #define ADD_POSIX_WARNING(p, text) STMT_START { \
15481 if (posix_warnings) { \
15482 if (! RExC_warn_text ) RExC_warn_text = \
15483 (AV *) sv_2mortal((SV *) newAV()); \
15484 av_push(RExC_warn_text, Perl_newSVpvf(aTHX_ \
15488 REPORT_LOCATION_ARGS(p))); \
15491 #define CLEAR_POSIX_WARNINGS() \
15493 if (posix_warnings && RExC_warn_text) \
15494 av_clear(RExC_warn_text); \
15497 #define CLEAR_POSIX_WARNINGS_AND_RETURN(ret) \
15499 CLEAR_POSIX_WARNINGS(); \
15504 S_handle_possible_posix(pTHX_ RExC_state_t *pRExC_state,
15506 const char * const s, /* Where the putative posix class begins.
15507 Normally, this is one past the '['. This
15508 parameter exists so it can be somewhere
15509 besides RExC_parse. */
15510 char ** updated_parse_ptr, /* Where to set the updated parse pointer, or
15512 AV ** posix_warnings, /* Where to place any generated warnings, or
15514 const bool check_only /* Don't die if error */
15517 /* This parses what the caller thinks may be one of the three POSIX
15519 * 1) a character class, like [:blank:]
15520 * 2) a collating symbol, like [. .]
15521 * 3) an equivalence class, like [= =]
15522 * In the latter two cases, it croaks if it finds a syntactically legal
15523 * one, as these are not handled by Perl.
15525 * The main purpose is to look for a POSIX character class. It returns:
15526 * a) the class number
15527 * if it is a completely syntactically and semantically legal class.
15528 * 'updated_parse_ptr', if not NULL, is set to point to just after the
15529 * closing ']' of the class
15530 * b) OOB_NAMEDCLASS
15531 * if it appears that one of the three POSIX constructs was meant, but
15532 * its specification was somehow defective. 'updated_parse_ptr', if
15533 * not NULL, is set to point to the character just after the end
15534 * character of the class. See below for handling of warnings.
15535 * c) NOT_MEANT_TO_BE_A_POSIX_CLASS
15536 * if it doesn't appear that a POSIX construct was intended.
15537 * 'updated_parse_ptr' is not changed. No warnings nor errors are
15540 * In b) there may be errors or warnings generated. If 'check_only' is
15541 * TRUE, then any errors are discarded. Warnings are returned to the
15542 * caller via an AV* created into '*posix_warnings' if it is not NULL. If
15543 * instead it is NULL, warnings are suppressed.
15545 * The reason for this function, and its complexity is that a bracketed
15546 * character class can contain just about anything. But it's easy to
15547 * mistype the very specific posix class syntax but yielding a valid
15548 * regular bracketed class, so it silently gets compiled into something
15549 * quite unintended.
15551 * The solution adopted here maintains backward compatibility except that
15552 * it adds a warning if it looks like a posix class was intended but
15553 * improperly specified. The warning is not raised unless what is input
15554 * very closely resembles one of the 14 legal posix classes. To do this,
15555 * it uses fuzzy parsing. It calculates how many single-character edits it
15556 * would take to transform what was input into a legal posix class. Only
15557 * if that number is quite small does it think that the intention was a
15558 * posix class. Obviously these are heuristics, and there will be cases
15559 * where it errs on one side or another, and they can be tweaked as
15560 * experience informs.
15562 * The syntax for a legal posix class is:
15564 * qr/(?xa: \[ : \^? [[:lower:]]{4,6} : \] )/
15566 * What this routine considers syntactically to be an intended posix class
15567 * is this (the comments indicate some restrictions that the pattern
15570 * qr/(?x: \[? # The left bracket, possibly
15572 * \h* # possibly followed by blanks
15573 * (?: \^ \h* )? # possibly a misplaced caret
15574 * [:;]? # The opening class character,
15575 * # possibly omitted. A typo
15576 * # semi-colon can also be used.
15578 * \^? # possibly a correctly placed
15579 * # caret, but not if there was also
15580 * # a misplaced one
15582 * .{3,15} # The class name. If there are
15583 * # deviations from the legal syntax,
15584 * # its edit distance must be close
15585 * # to a real class name in order
15586 * # for it to be considered to be
15587 * # an intended posix class.
15589 * [[:punct:]]? # The closing class character,
15590 * # possibly omitted. If not a colon
15591 * # nor semi colon, the class name
15592 * # must be even closer to a valid
15595 * \]? # The right bracket, possibly
15599 * In the above, \h must be ASCII-only.
15601 * These are heuristics, and can be tweaked as field experience dictates.
15602 * There will be cases when someone didn't intend to specify a posix class
15603 * that this warns as being so. The goal is to minimize these, while
15604 * maximizing the catching of things intended to be a posix class that
15605 * aren't parsed as such.
15609 const char * const e = RExC_end;
15610 unsigned complement = 0; /* If to complement the class */
15611 bool found_problem = FALSE; /* Assume OK until proven otherwise */
15612 bool has_opening_bracket = FALSE;
15613 bool has_opening_colon = FALSE;
15614 int class_number = OOB_NAMEDCLASS; /* Out-of-bounds until find
15616 const char * possible_end = NULL; /* used for a 2nd parse pass */
15617 const char* name_start; /* ptr to class name first char */
15619 /* If the number of single-character typos the input name is away from a
15620 * legal name is no more than this number, it is considered to have meant
15621 * the legal name */
15622 int max_distance = 2;
15624 /* to store the name. The size determines the maximum length before we
15625 * decide that no posix class was intended. Should be at least
15626 * sizeof("alphanumeric") */
15628 STATIC_ASSERT_DECL(C_ARRAY_LENGTH(input_text) >= sizeof "alphanumeric");
15630 PERL_ARGS_ASSERT_HANDLE_POSSIBLE_POSIX;
15632 CLEAR_POSIX_WARNINGS();
15635 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
15638 if (*(p - 1) != '[') {
15639 ADD_POSIX_WARNING(p, "it doesn't start with a '['");
15640 found_problem = TRUE;
15643 has_opening_bracket = TRUE;
15646 /* They could be confused and think you can put spaces between the
15649 found_problem = TRUE;
15653 } while (p < e && isBLANK(*p));
15655 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15658 /* For [. .] and [= =]. These are quite different internally from [: :],
15659 * so they are handled separately. */
15660 if (POSIXCC_NOTYET(*p) && p < e - 3) /* 1 for the close, and 1 for the ']'
15661 and 1 for at least one char in it
15664 const char open_char = *p;
15665 const char * temp_ptr = p + 1;
15667 /* These two constructs are not handled by perl, and if we find a
15668 * syntactically valid one, we croak. khw, who wrote this code, finds
15669 * this explanation of them very unclear:
15670 * http://pubs.opengroup.org/onlinepubs/009696899/basedefs/xbd_chap09.html
15671 * And searching the rest of the internet wasn't very helpful either.
15672 * It looks like just about any byte can be in these constructs,
15673 * depending on the locale. But unless the pattern is being compiled
15674 * under /l, which is very rare, Perl runs under the C or POSIX locale.
15675 * In that case, it looks like [= =] isn't allowed at all, and that
15676 * [. .] could be any single code point, but for longer strings the
15677 * constituent characters would have to be the ASCII alphabetics plus
15678 * the minus-hyphen. Any sensible locale definition would limit itself
15679 * to these. And any portable one definitely should. Trying to parse
15680 * the general case is a nightmare (see [perl #127604]). So, this code
15681 * looks only for interiors of these constructs that match:
15683 * Using \w relaxes the apparent rules a little, without adding much
15684 * danger of mistaking something else for one of these constructs.
15686 * [. .] in some implementations described on the internet is usable to
15687 * escape a character that otherwise is special in bracketed character
15688 * classes. For example [.].] means a literal right bracket instead of
15689 * the ending of the class
15691 * [= =] can legitimately contain a [. .] construct, but we don't
15692 * handle this case, as that [. .] construct will later get parsed
15693 * itself and croak then. And [= =] is checked for even when not under
15694 * /l, as Perl has long done so.
15696 * The code below relies on there being a trailing NUL, so it doesn't
15697 * have to keep checking if the parse ptr < e.
15699 if (temp_ptr[1] == open_char) {
15702 else while ( temp_ptr < e
15703 && (isWORDCHAR(*temp_ptr) || *temp_ptr == '-'))
15708 if (*temp_ptr == open_char) {
15710 if (*temp_ptr == ']') {
15712 if (! found_problem && ! check_only) {
15713 RExC_parse = (char *) temp_ptr;
15714 vFAIL3("POSIX syntax [%c %c] is reserved for future "
15715 "extensions", open_char, open_char);
15718 /* Here, the syntax wasn't completely valid, or else the call
15719 * is to check-only */
15720 if (updated_parse_ptr) {
15721 *updated_parse_ptr = (char *) temp_ptr;
15724 CLEAR_POSIX_WARNINGS_AND_RETURN(OOB_NAMEDCLASS);
15728 /* If we find something that started out to look like one of these
15729 * constructs, but isn't, we continue below so that it can be checked
15730 * for being a class name with a typo of '.' or '=' instead of a colon.
15734 /* Here, we think there is a possibility that a [: :] class was meant, and
15735 * we have the first real character. It could be they think the '^' comes
15738 found_problem = TRUE;
15739 ADD_POSIX_WARNING(p + 1, "the '^' must come after the colon");
15744 found_problem = TRUE;
15748 } while (p < e && isBLANK(*p));
15750 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15754 /* But the first character should be a colon, which they could have easily
15755 * mistyped on a qwerty keyboard as a semi-colon (and which may be hard to
15756 * distinguish from a colon, so treat that as a colon). */
15759 has_opening_colon = TRUE;
15761 else if (*p == ';') {
15762 found_problem = TRUE;
15764 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
15765 has_opening_colon = TRUE;
15768 found_problem = TRUE;
15769 ADD_POSIX_WARNING(p, "there must be a starting ':'");
15771 /* Consider an initial punctuation (not one of the recognized ones) to
15772 * be a left terminator */
15773 if (*p != '^' && *p != ']' && isPUNCT(*p)) {
15778 /* They may think that you can put spaces between the components */
15780 found_problem = TRUE;
15784 } while (p < e && isBLANK(*p));
15786 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15791 /* We consider something like [^:^alnum:]] to not have been intended to
15792 * be a posix class, but XXX maybe we should */
15794 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15801 /* Again, they may think that you can put spaces between the components */
15803 found_problem = TRUE;
15807 } while (p < e && isBLANK(*p));
15809 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15814 /* XXX This ']' may be a typo, and something else was meant. But
15815 * treating it as such creates enough complications, that that
15816 * possibility isn't currently considered here. So we assume that the
15817 * ']' is what is intended, and if we've already found an initial '[',
15818 * this leaves this construct looking like [:] or [:^], which almost
15819 * certainly weren't intended to be posix classes */
15820 if (has_opening_bracket) {
15821 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15824 /* But this function can be called when we parse the colon for
15825 * something like qr/[alpha:]]/, so we back up to look for the
15830 found_problem = TRUE;
15831 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
15833 else if (*p != ':') {
15835 /* XXX We are currently very restrictive here, so this code doesn't
15836 * consider the possibility that, say, /[alpha.]]/ was intended to
15837 * be a posix class. */
15838 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15841 /* Here we have something like 'foo:]'. There was no initial colon,
15842 * and we back up over 'foo. XXX Unlike the going forward case, we
15843 * don't handle typos of non-word chars in the middle */
15844 has_opening_colon = FALSE;
15847 while (p > RExC_start && isWORDCHAR(*p)) {
15852 /* Here, we have positioned ourselves to where we think the first
15853 * character in the potential class is */
15856 /* Now the interior really starts. There are certain key characters that
15857 * can end the interior, or these could just be typos. To catch both
15858 * cases, we may have to do two passes. In the first pass, we keep on
15859 * going unless we come to a sequence that matches
15860 * qr/ [[:punct:]] [[:blank:]]* \] /xa
15861 * This means it takes a sequence to end the pass, so two typos in a row if
15862 * that wasn't what was intended. If the class is perfectly formed, just
15863 * this one pass is needed. We also stop if there are too many characters
15864 * being accumulated, but this number is deliberately set higher than any
15865 * real class. It is set high enough so that someone who thinks that
15866 * 'alphanumeric' is a correct name would get warned that it wasn't.
15867 * While doing the pass, we keep track of where the key characters were in
15868 * it. If we don't find an end to the class, and one of the key characters
15869 * was found, we redo the pass, but stop when we get to that character.
15870 * Thus the key character was considered a typo in the first pass, but a
15871 * terminator in the second. If two key characters are found, we stop at
15872 * the second one in the first pass. Again this can miss two typos, but
15873 * catches a single one
15875 * In the first pass, 'possible_end' starts as NULL, and then gets set to
15876 * point to the first key character. For the second pass, it starts as -1.
15882 bool has_blank = FALSE;
15883 bool has_upper = FALSE;
15884 bool has_terminating_colon = FALSE;
15885 bool has_terminating_bracket = FALSE;
15886 bool has_semi_colon = FALSE;
15887 unsigned int name_len = 0;
15888 int punct_count = 0;
15892 /* Squeeze out blanks when looking up the class name below */
15893 if (isBLANK(*p) ) {
15895 found_problem = TRUE;
15900 /* The name will end with a punctuation */
15902 const char * peek = p + 1;
15904 /* Treat any non-']' punctuation followed by a ']' (possibly
15905 * with intervening blanks) as trying to terminate the class.
15906 * ']]' is very likely to mean a class was intended (but
15907 * missing the colon), but the warning message that gets
15908 * generated shows the error position better if we exit the
15909 * loop at the bottom (eventually), so skip it here. */
15911 if (peek < e && isBLANK(*peek)) {
15913 found_problem = TRUE;
15916 } while (peek < e && isBLANK(*peek));
15919 if (peek < e && *peek == ']') {
15920 has_terminating_bracket = TRUE;
15922 has_terminating_colon = TRUE;
15924 else if (*p == ';') {
15925 has_semi_colon = TRUE;
15926 has_terminating_colon = TRUE;
15929 found_problem = TRUE;
15936 /* Here we have punctuation we thought didn't end the class.
15937 * Keep track of the position of the key characters that are
15938 * more likely to have been class-enders */
15939 if (*p == ']' || *p == '[' || *p == ':' || *p == ';') {
15941 /* Allow just one such possible class-ender not actually
15942 * ending the class. */
15943 if (possible_end) {
15949 /* If we have too many punctuation characters, no use in
15951 if (++punct_count > max_distance) {
15955 /* Treat the punctuation as a typo. */
15956 input_text[name_len++] = *p;
15959 else if (isUPPER(*p)) { /* Use lowercase for lookup */
15960 input_text[name_len++] = toLOWER(*p);
15962 found_problem = TRUE;
15964 } else if (! UTF || UTF8_IS_INVARIANT(*p)) {
15965 input_text[name_len++] = *p;
15969 input_text[name_len++] = utf8_to_uvchr_buf((U8 *) p, e, NULL);
15973 /* The declaration of 'input_text' is how long we allow a potential
15974 * class name to be, before saying they didn't mean a class name at
15976 if (name_len >= C_ARRAY_LENGTH(input_text)) {
15981 /* We get to here when the possible class name hasn't been properly
15982 * terminated before:
15983 * 1) we ran off the end of the pattern; or
15984 * 2) found two characters, each of which might have been intended to
15985 * be the name's terminator
15986 * 3) found so many punctuation characters in the purported name,
15987 * that the edit distance to a valid one is exceeded
15988 * 4) we decided it was more characters than anyone could have
15989 * intended to be one. */
15991 found_problem = TRUE;
15993 /* In the final two cases, we know that looking up what we've
15994 * accumulated won't lead to a match, even a fuzzy one. */
15995 if ( name_len >= C_ARRAY_LENGTH(input_text)
15996 || punct_count > max_distance)
15998 /* If there was an intermediate key character that could have been
15999 * an intended end, redo the parse, but stop there */
16000 if (possible_end && possible_end != (char *) -1) {
16001 possible_end = (char *) -1; /* Special signal value to say
16002 we've done a first pass */
16007 /* Otherwise, it can't have meant to have been a class */
16008 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
16011 /* If we ran off the end, and the final character was a punctuation
16012 * one, back up one, to look at that final one just below. Later, we
16013 * will restore the parse pointer if appropriate */
16014 if (name_len && p == e && isPUNCT(*(p-1))) {
16019 if (p < e && isPUNCT(*p)) {
16021 has_terminating_bracket = TRUE;
16023 /* If this is a 2nd ']', and the first one is just below this
16024 * one, consider that to be the real terminator. This gives a
16025 * uniform and better positioning for the warning message */
16027 && possible_end != (char *) -1
16028 && *possible_end == ']'
16029 && name_len && input_text[name_len - 1] == ']')
16034 /* And this is actually equivalent to having done the 2nd
16035 * pass now, so set it to not try again */
16036 possible_end = (char *) -1;
16041 has_terminating_colon = TRUE;
16043 else if (*p == ';') {
16044 has_semi_colon = TRUE;
16045 has_terminating_colon = TRUE;
16053 /* Here, we have a class name to look up. We can short circuit the
16054 * stuff below for short names that can't possibly be meant to be a
16055 * class name. (We can do this on the first pass, as any second pass
16056 * will yield an even shorter name) */
16057 if (name_len < 3) {
16058 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
16061 /* Find which class it is. Initially switch on the length of the name.
16063 switch (name_len) {
16065 if (memEQs(name_start, 4, "word")) {
16066 /* this is not POSIX, this is the Perl \w */
16067 class_number = ANYOF_WORDCHAR;
16071 /* Names all of length 5: alnum alpha ascii blank cntrl digit
16072 * graph lower print punct space upper
16073 * Offset 4 gives the best switch position. */
16074 switch (name_start[4]) {
16076 if (memBEGINs(name_start, 5, "alph")) /* alpha */
16077 class_number = ANYOF_ALPHA;
16080 if (memBEGINs(name_start, 5, "spac")) /* space */
16081 class_number = ANYOF_SPACE;
16084 if (memBEGINs(name_start, 5, "grap")) /* graph */
16085 class_number = ANYOF_GRAPH;
16088 if (memBEGINs(name_start, 5, "asci")) /* ascii */
16089 class_number = ANYOF_ASCII;
16092 if (memBEGINs(name_start, 5, "blan")) /* blank */
16093 class_number = ANYOF_BLANK;
16096 if (memBEGINs(name_start, 5, "cntr")) /* cntrl */
16097 class_number = ANYOF_CNTRL;
16100 if (memBEGINs(name_start, 5, "alnu")) /* alnum */
16101 class_number = ANYOF_ALPHANUMERIC;
16104 if (memBEGINs(name_start, 5, "lowe")) /* lower */
16105 class_number = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
16106 else if (memBEGINs(name_start, 5, "uppe")) /* upper */
16107 class_number = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
16110 if (memBEGINs(name_start, 5, "digi")) /* digit */
16111 class_number = ANYOF_DIGIT;
16112 else if (memBEGINs(name_start, 5, "prin")) /* print */
16113 class_number = ANYOF_PRINT;
16114 else if (memBEGINs(name_start, 5, "punc")) /* punct */
16115 class_number = ANYOF_PUNCT;
16120 if (memEQs(name_start, 6, "xdigit"))
16121 class_number = ANYOF_XDIGIT;
16125 /* If the name exactly matches a posix class name the class number will
16126 * here be set to it, and the input almost certainly was meant to be a
16127 * posix class, so we can skip further checking. If instead the syntax
16128 * is exactly correct, but the name isn't one of the legal ones, we
16129 * will return that as an error below. But if neither of these apply,
16130 * it could be that no posix class was intended at all, or that one
16131 * was, but there was a typo. We tease these apart by doing fuzzy
16132 * matching on the name */
16133 if (class_number == OOB_NAMEDCLASS && found_problem) {
16134 const UV posix_names[][6] = {
16135 { 'a', 'l', 'n', 'u', 'm' },
16136 { 'a', 'l', 'p', 'h', 'a' },
16137 { 'a', 's', 'c', 'i', 'i' },
16138 { 'b', 'l', 'a', 'n', 'k' },
16139 { 'c', 'n', 't', 'r', 'l' },
16140 { 'd', 'i', 'g', 'i', 't' },
16141 { 'g', 'r', 'a', 'p', 'h' },
16142 { 'l', 'o', 'w', 'e', 'r' },
16143 { 'p', 'r', 'i', 'n', 't' },
16144 { 'p', 'u', 'n', 'c', 't' },
16145 { 's', 'p', 'a', 'c', 'e' },
16146 { 'u', 'p', 'p', 'e', 'r' },
16147 { 'w', 'o', 'r', 'd' },
16148 { 'x', 'd', 'i', 'g', 'i', 't' }
16150 /* The names of the above all have added NULs to make them the same
16151 * size, so we need to also have the real lengths */
16152 const UV posix_name_lengths[] = {
16153 sizeof("alnum") - 1,
16154 sizeof("alpha") - 1,
16155 sizeof("ascii") - 1,
16156 sizeof("blank") - 1,
16157 sizeof("cntrl") - 1,
16158 sizeof("digit") - 1,
16159 sizeof("graph") - 1,
16160 sizeof("lower") - 1,
16161 sizeof("print") - 1,
16162 sizeof("punct") - 1,
16163 sizeof("space") - 1,
16164 sizeof("upper") - 1,
16165 sizeof("word") - 1,
16166 sizeof("xdigit")- 1
16169 int temp_max = max_distance; /* Use a temporary, so if we
16170 reparse, we haven't changed the
16173 /* Use a smaller max edit distance if we are missing one of the
16175 if ( has_opening_bracket + has_opening_colon < 2
16176 || has_terminating_bracket + has_terminating_colon < 2)
16181 /* See if the input name is close to a legal one */
16182 for (i = 0; i < C_ARRAY_LENGTH(posix_names); i++) {
16184 /* Short circuit call if the lengths are too far apart to be
16186 if (abs( (int) (name_len - posix_name_lengths[i]))
16192 if (edit_distance(input_text,
16195 posix_name_lengths[i],
16199 { /* If it is close, it probably was intended to be a class */
16200 goto probably_meant_to_be;
16204 /* Here the input name is not close enough to a valid class name
16205 * for us to consider it to be intended to be a posix class. If
16206 * we haven't already done so, and the parse found a character that
16207 * could have been terminators for the name, but which we absorbed
16208 * as typos during the first pass, repeat the parse, signalling it
16209 * to stop at that character */
16210 if (possible_end && possible_end != (char *) -1) {
16211 possible_end = (char *) -1;
16216 /* Here neither pass found a close-enough class name */
16217 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
16220 probably_meant_to_be:
16222 /* Here we think that a posix specification was intended. Update any
16224 if (updated_parse_ptr) {
16225 *updated_parse_ptr = (char *) p;
16228 /* If a posix class name was intended but incorrectly specified, we
16229 * output or return the warnings */
16230 if (found_problem) {
16232 /* We set flags for these issues in the parse loop above instead of
16233 * adding them to the list of warnings, because we can parse it
16234 * twice, and we only want one warning instance */
16236 ADD_POSIX_WARNING(p, "the name must be all lowercase letters");
16239 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
16241 if (has_semi_colon) {
16242 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
16244 else if (! has_terminating_colon) {
16245 ADD_POSIX_WARNING(p, "there is no terminating ':'");
16247 if (! has_terminating_bracket) {
16248 ADD_POSIX_WARNING(p, "there is no terminating ']'");
16251 if ( posix_warnings
16253 && av_count(RExC_warn_text) > 0)
16255 *posix_warnings = RExC_warn_text;
16258 else if (class_number != OOB_NAMEDCLASS) {
16259 /* If it is a known class, return the class. The class number
16260 * #defines are structured so each complement is +1 to the normal
16262 CLEAR_POSIX_WARNINGS_AND_RETURN(class_number + complement);
16264 else if (! check_only) {
16266 /* Here, it is an unrecognized class. This is an error (unless the
16267 * call is to check only, which we've already handled above) */
16268 const char * const complement_string = (complement)
16271 RExC_parse = (char *) p;
16272 vFAIL3utf8f("POSIX class [:%s%" UTF8f ":] unknown",
16274 UTF8fARG(UTF, RExC_parse - name_start - 2, name_start));
16278 return OOB_NAMEDCLASS;
16280 #undef ADD_POSIX_WARNING
16282 STATIC unsigned int
16283 S_regex_set_precedence(const U8 my_operator) {
16285 /* Returns the precedence in the (?[...]) construct of the input operator,
16286 * specified by its character representation. The precedence follows
16287 * general Perl rules, but it extends this so that ')' and ']' have (low)
16288 * precedence even though they aren't really operators */
16290 switch (my_operator) {
16306 NOT_REACHED; /* NOTREACHED */
16307 return 0; /* Silence compiler warning */
16310 STATIC regnode_offset
16311 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
16312 I32 *flagp, U32 depth,
16313 char * const oregcomp_parse)
16315 /* Handle the (?[...]) construct to do set operations */
16317 U8 curchar; /* Current character being parsed */
16318 UV start, end; /* End points of code point ranges */
16319 SV* final = NULL; /* The end result inversion list */
16320 SV* result_string; /* 'final' stringified */
16321 AV* stack; /* stack of operators and operands not yet
16323 AV* fence_stack = NULL; /* A stack containing the positions in
16324 'stack' of where the undealt-with left
16325 parens would be if they were actually
16327 /* The 'volatile' is a workaround for an optimiser bug
16328 * in Solaris Studio 12.3. See RT #127455 */
16329 volatile IV fence = 0; /* Position of where most recent undealt-
16330 with left paren in stack is; -1 if none.
16332 STRLEN len; /* Temporary */
16333 regnode_offset node; /* Temporary, and final regnode returned by
16335 const bool save_fold = FOLD; /* Temporary */
16336 char *save_end, *save_parse; /* Temporaries */
16337 const bool in_locale = LOC; /* we turn off /l during processing */
16339 DECLARE_AND_GET_RE_DEBUG_FLAGS;
16341 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
16342 PERL_UNUSED_ARG(oregcomp_parse); /* Only for Set_Node_Length */
16344 DEBUG_PARSE("xcls");
16347 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
16350 /* The use of this operator implies /u. This is required so that the
16351 * compile time values are valid in all runtime cases */
16352 REQUIRE_UNI_RULES(flagp, 0);
16354 ckWARNexperimental(RExC_parse,
16355 WARN_EXPERIMENTAL__REGEX_SETS,
16356 "The regex_sets feature is experimental");
16358 /* Everything in this construct is a metacharacter. Operands begin with
16359 * either a '\' (for an escape sequence), or a '[' for a bracketed
16360 * character class. Any other character should be an operator, or
16361 * parenthesis for grouping. Both types of operands are handled by calling
16362 * regclass() to parse them. It is called with a parameter to indicate to
16363 * return the computed inversion list. The parsing here is implemented via
16364 * a stack. Each entry on the stack is a single character representing one
16365 * of the operators; or else a pointer to an operand inversion list. */
16367 #define IS_OPERATOR(a) SvIOK(a)
16368 #define IS_OPERAND(a) (! IS_OPERATOR(a))
16370 /* The stack is kept in Łukasiewicz order. (That's pronounced similar
16371 * to luke-a-shave-itch (or -itz), but people who didn't want to bother
16372 * with pronouncing it called it Reverse Polish instead, but now that YOU
16373 * know how to pronounce it you can use the correct term, thus giving due
16374 * credit to the person who invented it, and impressing your geek friends.
16375 * Wikipedia says that the pronounciation of "Ł" has been changing so that
16376 * it is now more like an English initial W (as in wonk) than an L.)
16378 * This means that, for example, 'a | b & c' is stored on the stack as
16386 * where the numbers in brackets give the stack [array] element number.
16387 * In this implementation, parentheses are not stored on the stack.
16388 * Instead a '(' creates a "fence" so that the part of the stack below the
16389 * fence is invisible except to the corresponding ')' (this allows us to
16390 * replace testing for parens, by using instead subtraction of the fence
16391 * position). As new operands are processed they are pushed onto the stack
16392 * (except as noted in the next paragraph). New operators of higher
16393 * precedence than the current final one are inserted on the stack before
16394 * the lhs operand (so that when the rhs is pushed next, everything will be
16395 * in the correct positions shown above. When an operator of equal or
16396 * lower precedence is encountered in parsing, all the stacked operations
16397 * of equal or higher precedence are evaluated, leaving the result as the
16398 * top entry on the stack. This makes higher precedence operations
16399 * evaluate before lower precedence ones, and causes operations of equal
16400 * precedence to left associate.
16402 * The only unary operator '!' is immediately pushed onto the stack when
16403 * encountered. When an operand is encountered, if the top of the stack is
16404 * a '!", the complement is immediately performed, and the '!' popped. The
16405 * resulting value is treated as a new operand, and the logic in the
16406 * previous paragraph is executed. Thus in the expression
16408 * the stack looks like
16414 * as 'b' gets parsed, the latter gets evaluated to '!b', and the stack
16421 * A ')' is treated as an operator with lower precedence than all the
16422 * aforementioned ones, which causes all operations on the stack above the
16423 * corresponding '(' to be evaluated down to a single resultant operand.
16424 * Then the fence for the '(' is removed, and the operand goes through the
16425 * algorithm above, without the fence.
16427 * A separate stack is kept of the fence positions, so that the position of
16428 * the latest so-far unbalanced '(' is at the top of it.
16430 * The ']' ending the construct is treated as the lowest operator of all,
16431 * so that everything gets evaluated down to a single operand, which is the
16434 sv_2mortal((SV *)(stack = newAV()));
16435 sv_2mortal((SV *)(fence_stack = newAV()));
16437 while (RExC_parse < RExC_end) {
16438 I32 top_index; /* Index of top-most element in 'stack' */
16439 SV** top_ptr; /* Pointer to top 'stack' element */
16440 SV* current = NULL; /* To contain the current inversion list
16442 SV* only_to_avoid_leaks;
16444 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
16445 TRUE /* Force /x */ );
16446 if (RExC_parse >= RExC_end) { /* Fail */
16450 curchar = UCHARAT(RExC_parse);
16454 #ifdef ENABLE_REGEX_SETS_DEBUGGING
16455 /* Enable with -Accflags=-DENABLE_REGEX_SETS_DEBUGGING */
16456 DEBUG_U(dump_regex_sets_structures(pRExC_state,
16457 stack, fence, fence_stack));
16460 top_index = av_tindex_skip_len_mg(stack);
16463 SV** stacked_ptr; /* Ptr to something already on 'stack' */
16464 char stacked_operator; /* The topmost operator on the 'stack'. */
16465 SV* lhs; /* Operand to the left of the operator */
16466 SV* rhs; /* Operand to the right of the operator */
16467 SV* fence_ptr; /* Pointer to top element of the fence
16471 if ( RExC_parse < RExC_end - 2
16472 && UCHARAT(RExC_parse + 1) == '?'
16473 && UCHARAT(RExC_parse + 2) == '^')
16475 const regnode_offset orig_emit = RExC_emit;
16476 SV * resultant_invlist;
16478 /* If is a '(?^', could be an embedded '(?^flags:(?[...])'.
16479 * This happens when we have some thing like
16481 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
16483 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
16485 * Here we would be handling the interpolated
16486 * '$thai_or_lao'. We handle this by a recursive call to
16487 * reg which returns the inversion list the
16488 * interpolated expression evaluates to. Actually, the
16489 * return is a special regnode containing a pointer to that
16490 * inversion list. If the return isn't that regnode alone,
16491 * we know that this wasn't such an interpolation, which is
16492 * an error: we need to get a single inversion list back
16493 * from the recursion */
16498 node = reg(pRExC_state, 2, flagp, depth+1);
16499 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16501 if ( OP(REGNODE_p(node)) != REGEX_SET
16502 /* If more than a single node returned, the nested
16503 * parens evaluated to more than just a (?[...]),
16504 * which isn't legal */
16505 || RExC_emit != orig_emit
16506 + NODE_STEP_REGNODE
16507 + regarglen[REGEX_SET])
16509 vFAIL("Expecting interpolated extended charclass");
16511 resultant_invlist = (SV *) ARGp(REGNODE_p(node));
16512 current = invlist_clone(resultant_invlist, NULL);
16513 SvREFCNT_dec(resultant_invlist);
16516 RExC_emit = orig_emit;
16517 goto handle_operand;
16520 /* A regular '('. Look behind for illegal syntax */
16521 if (top_index - fence >= 0) {
16522 /* If the top entry on the stack is an operator, it had
16523 * better be a '!', otherwise the entry below the top
16524 * operand should be an operator */
16525 if ( ! (top_ptr = av_fetch(stack, top_index, FALSE))
16526 || (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) != '!')
16527 || ( IS_OPERAND(*top_ptr)
16528 && ( top_index - fence < 1
16529 || ! (stacked_ptr = av_fetch(stack,
16532 || ! IS_OPERATOR(*stacked_ptr))))
16535 vFAIL("Unexpected '(' with no preceding operator");
16539 /* Stack the position of this undealt-with left paren */
16540 av_push(fence_stack, newSViv(fence));
16541 fence = top_index + 1;
16545 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
16546 * multi-char folds are allowed. */
16547 if (!regclass(pRExC_state, flagp, depth+1,
16548 TRUE, /* means parse just the next thing */
16549 FALSE, /* don't allow multi-char folds */
16550 FALSE, /* don't silence non-portable warnings. */
16552 FALSE, /* Require return to be an ANYOF */
16555 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16556 goto regclass_failed;
16561 /* regclass() will return with parsing just the \ sequence,
16562 * leaving the parse pointer at the next thing to parse */
16564 goto handle_operand;
16566 case '[': /* Is a bracketed character class */
16568 /* See if this is a [:posix:] class. */
16569 bool is_posix_class = (OOB_NAMEDCLASS
16570 < handle_possible_posix(pRExC_state,
16574 TRUE /* checking only */));
16575 /* If it is a posix class, leave the parse pointer at the '['
16576 * to fool regclass() into thinking it is part of a
16577 * '[[:posix:]]'. */
16578 if (! is_posix_class) {
16582 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
16583 * multi-char folds are allowed. */
16584 if (!regclass(pRExC_state, flagp, depth+1,
16585 is_posix_class, /* parse the whole char
16586 class only if not a
16588 FALSE, /* don't allow multi-char folds */
16589 TRUE, /* silence non-portable warnings. */
16591 FALSE, /* Require return to be an ANYOF */
16594 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16595 goto regclass_failed;
16600 /* function call leaves parse pointing to the ']', except if we
16602 if (is_posix_class) {
16606 goto handle_operand;
16610 if (top_index >= 1) {
16611 goto join_operators;
16614 /* Only a single operand on the stack: are done */
16618 if (av_tindex_skip_len_mg(fence_stack) < 0) {
16619 if (UCHARAT(RExC_parse - 1) == ']') {
16623 vFAIL("Unexpected ')'");
16626 /* If nothing after the fence, is missing an operand */
16627 if (top_index - fence < 0) {
16631 /* If at least two things on the stack, treat this as an
16633 if (top_index - fence >= 1) {
16634 goto join_operators;
16637 /* Here only a single thing on the fenced stack, and there is a
16638 * fence. Get rid of it */
16639 fence_ptr = av_pop(fence_stack);
16641 fence = SvIV(fence_ptr);
16642 SvREFCNT_dec_NN(fence_ptr);
16649 /* Having gotten rid of the fence, we pop the operand at the
16650 * stack top and process it as a newly encountered operand */
16651 current = av_pop(stack);
16652 if (IS_OPERAND(current)) {
16653 goto handle_operand;
16665 /* These binary operators should have a left operand already
16667 if ( top_index - fence < 0
16668 || top_index - fence == 1
16669 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
16670 || ! IS_OPERAND(*top_ptr))
16672 goto unexpected_binary;
16675 /* If only the one operand is on the part of the stack visible
16676 * to us, we just place this operator in the proper position */
16677 if (top_index - fence < 2) {
16679 /* Place the operator before the operand */
16681 SV* lhs = av_pop(stack);
16682 av_push(stack, newSVuv(curchar));
16683 av_push(stack, lhs);
16687 /* But if there is something else on the stack, we need to
16688 * process it before this new operator if and only if the
16689 * stacked operation has equal or higher precedence than the
16694 /* The operator on the stack is supposed to be below both its
16696 if ( ! (stacked_ptr = av_fetch(stack, top_index - 2, FALSE))
16697 || IS_OPERAND(*stacked_ptr))
16699 /* But if not, it's legal and indicates we are completely
16700 * done if and only if we're currently processing a ']',
16701 * which should be the final thing in the expression */
16702 if (curchar == ']') {
16708 vFAIL2("Unexpected binary operator '%c' with no "
16709 "preceding operand", curchar);
16711 stacked_operator = (char) SvUV(*stacked_ptr);
16713 if (regex_set_precedence(curchar)
16714 > regex_set_precedence(stacked_operator))
16716 /* Here, the new operator has higher precedence than the
16717 * stacked one. This means we need to add the new one to
16718 * the stack to await its rhs operand (and maybe more
16719 * stuff). We put it before the lhs operand, leaving
16720 * untouched the stacked operator and everything below it
16722 lhs = av_pop(stack);
16723 assert(IS_OPERAND(lhs));
16725 av_push(stack, newSVuv(curchar));
16726 av_push(stack, lhs);
16730 /* Here, the new operator has equal or lower precedence than
16731 * what's already there. This means the operation already
16732 * there should be performed now, before the new one. */
16734 rhs = av_pop(stack);
16735 if (! IS_OPERAND(rhs)) {
16737 /* This can happen when a ! is not followed by an operand,
16738 * like in /(?[\t &!])/ */
16742 lhs = av_pop(stack);
16744 if (! IS_OPERAND(lhs)) {
16746 /* This can happen when there is an empty (), like in
16747 * /(?[[0]+()+])/ */
16751 switch (stacked_operator) {
16753 _invlist_intersection(lhs, rhs, &rhs);
16758 _invlist_union(lhs, rhs, &rhs);
16762 _invlist_subtract(lhs, rhs, &rhs);
16765 case '^': /* The union minus the intersection */
16770 _invlist_union(lhs, rhs, &u);
16771 _invlist_intersection(lhs, rhs, &i);
16772 _invlist_subtract(u, i, &rhs);
16773 SvREFCNT_dec_NN(i);
16774 SvREFCNT_dec_NN(u);
16780 /* Here, the higher precedence operation has been done, and the
16781 * result is in 'rhs'. We overwrite the stacked operator with
16782 * the result. Then we redo this code to either push the new
16783 * operator onto the stack or perform any higher precedence
16784 * stacked operation */
16785 only_to_avoid_leaks = av_pop(stack);
16786 SvREFCNT_dec(only_to_avoid_leaks);
16787 av_push(stack, rhs);
16790 case '!': /* Highest priority, right associative */
16792 /* If what's already at the top of the stack is another '!",
16793 * they just cancel each other out */
16794 if ( (top_ptr = av_fetch(stack, top_index, FALSE))
16795 && (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) == '!'))
16797 only_to_avoid_leaks = av_pop(stack);
16798 SvREFCNT_dec(only_to_avoid_leaks);
16800 else { /* Otherwise, since it's right associative, just push
16802 av_push(stack, newSVuv(curchar));
16807 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16808 if (RExC_parse >= RExC_end) {
16811 vFAIL("Unexpected character");
16815 /* Here 'current' is the operand. If something is already on the
16816 * stack, we have to check if it is a !. But first, the code above
16817 * may have altered the stack in the time since we earlier set
16820 top_index = av_tindex_skip_len_mg(stack);
16821 if (top_index - fence >= 0) {
16822 /* If the top entry on the stack is an operator, it had better
16823 * be a '!', otherwise the entry below the top operand should
16824 * be an operator */
16825 top_ptr = av_fetch(stack, top_index, FALSE);
16827 if (IS_OPERATOR(*top_ptr)) {
16829 /* The only permissible operator at the top of the stack is
16830 * '!', which is applied immediately to this operand. */
16831 curchar = (char) SvUV(*top_ptr);
16832 if (curchar != '!') {
16833 SvREFCNT_dec(current);
16834 vFAIL2("Unexpected binary operator '%c' with no "
16835 "preceding operand", curchar);
16838 _invlist_invert(current);
16840 only_to_avoid_leaks = av_pop(stack);
16841 SvREFCNT_dec(only_to_avoid_leaks);
16843 /* And we redo with the inverted operand. This allows
16844 * handling multiple ! in a row */
16845 goto handle_operand;
16847 /* Single operand is ok only for the non-binary ')'
16849 else if ((top_index - fence == 0 && curchar != ')')
16850 || (top_index - fence > 0
16851 && (! (stacked_ptr = av_fetch(stack,
16854 || IS_OPERAND(*stacked_ptr))))
16856 SvREFCNT_dec(current);
16857 vFAIL("Operand with no preceding operator");
16861 /* Here there was nothing on the stack or the top element was
16862 * another operand. Just add this new one */
16863 av_push(stack, current);
16865 } /* End of switch on next parse token */
16867 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16868 } /* End of loop parsing through the construct */
16870 vFAIL("Syntax error in (?[...])");
16874 if (RExC_parse >= RExC_end || RExC_parse[1] != ')') {
16875 if (RExC_parse < RExC_end) {
16879 vFAIL("Unexpected ']' with no following ')' in (?[...");
16882 if (av_tindex_skip_len_mg(fence_stack) >= 0) {
16883 vFAIL("Unmatched (");
16886 if (av_tindex_skip_len_mg(stack) < 0 /* Was empty */
16887 || ((final = av_pop(stack)) == NULL)
16888 || ! IS_OPERAND(final)
16889 || ! is_invlist(final)
16890 || av_tindex_skip_len_mg(stack) >= 0) /* More left on stack */
16893 SvREFCNT_dec(final);
16894 vFAIL("Incomplete expression within '(?[ ])'");
16897 /* Here, 'final' is the resultant inversion list from evaluating the
16898 * expression. Return it if so requested */
16899 if (return_invlist) {
16900 *return_invlist = final;
16904 if (RExC_sets_depth) { /* If within a recursive call, return in a special
16907 node = regpnode(pRExC_state, REGEX_SET, final);
16911 /* Otherwise generate a resultant node, based on 'final'. regclass()
16912 * is expecting a string of ranges and individual code points */
16913 invlist_iterinit(final);
16914 result_string = newSVpvs("");
16915 while (invlist_iternext(final, &start, &end)) {
16916 if (start == end) {
16917 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}", start);
16920 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}-\\x{%"
16921 UVXf "}", start, end);
16925 /* About to generate an ANYOF (or similar) node from the inversion list
16926 * we have calculated */
16927 save_parse = RExC_parse;
16928 RExC_parse = SvPV(result_string, len);
16929 save_end = RExC_end;
16930 RExC_end = RExC_parse + len;
16931 TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE;
16933 /* We turn off folding around the call, as the class we have
16934 * constructed already has all folding taken into consideration, and we
16935 * don't want regclass() to add to that */
16936 RExC_flags &= ~RXf_PMf_FOLD;
16937 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if multi-char
16938 * folds are allowed. */
16939 node = regclass(pRExC_state, flagp, depth+1,
16940 FALSE, /* means parse the whole char class */
16941 FALSE, /* don't allow multi-char folds */
16942 TRUE, /* silence non-portable warnings. The above may
16943 very well have generated non-portable code
16944 points, but they're valid on this machine */
16945 FALSE, /* similarly, no need for strict */
16947 /* We can optimize into something besides an ANYOF,
16948 * except under /l, which needs to be ANYOF because of
16949 * runtime checks for locale sanity, etc */
16955 RExC_parse = save_parse + 1;
16956 RExC_end = save_end;
16957 SvREFCNT_dec_NN(final);
16958 SvREFCNT_dec_NN(result_string);
16961 RExC_flags |= RXf_PMf_FOLD;
16965 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16966 goto regclass_failed;
16969 /* Fix up the node type if we are in locale. (We have pretended we are
16970 * under /u for the purposes of regclass(), as this construct will only
16971 * work under UTF-8 locales. But now we change the opcode to be ANYOFL
16972 * (so as to cause any warnings about bad locales to be output in
16973 * regexec.c), and add the flag that indicates to check if not in a
16974 * UTF-8 locale. The reason we above forbid optimization into
16975 * something other than an ANYOF node is simply to minimize the number
16976 * of code changes in regexec.c. Otherwise we would have to create new
16977 * EXACTish node types and deal with them. This decision could be
16978 * revisited should this construct become popular.
16980 * (One might think we could look at the resulting ANYOF node and
16981 * suppress the flag if everything is above 255, as those would be
16982 * UTF-8 only, but this isn't true, as the components that led to that
16983 * result could have been locale-affected, and just happen to cancel
16984 * each other out under UTF-8 locales.) */
16986 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
16988 assert(OP(REGNODE_p(node)) == ANYOF);
16990 OP(REGNODE_p(node)) = ANYOFL;
16991 ANYOF_FLAGS(REGNODE_p(node))
16992 |= ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
16996 nextchar(pRExC_state);
16997 Set_Node_Length(REGNODE_p(node), RExC_parse - oregcomp_parse + 1); /* MJD */
17001 FAIL2("panic: regclass returned failure to handle_sets, " "flags=%#" UVxf,
17005 #ifdef ENABLE_REGEX_SETS_DEBUGGING
17008 S_dump_regex_sets_structures(pTHX_ RExC_state_t *pRExC_state,
17009 AV * stack, const IV fence, AV * fence_stack)
17010 { /* Dumps the stacks in handle_regex_sets() */
17012 const SSize_t stack_top = av_tindex_skip_len_mg(stack);
17013 const SSize_t fence_stack_top = av_tindex_skip_len_mg(fence_stack);
17016 PERL_ARGS_ASSERT_DUMP_REGEX_SETS_STRUCTURES;
17018 PerlIO_printf(Perl_debug_log, "\nParse position is:%s\n", RExC_parse);
17020 if (stack_top < 0) {
17021 PerlIO_printf(Perl_debug_log, "Nothing on stack\n");
17024 PerlIO_printf(Perl_debug_log, "Stack: (fence=%d)\n", (int) fence);
17025 for (i = stack_top; i >= 0; i--) {
17026 SV ** element_ptr = av_fetch(stack, i, FALSE);
17027 if (! element_ptr) {
17030 if (IS_OPERATOR(*element_ptr)) {
17031 PerlIO_printf(Perl_debug_log, "[%d]: %c\n",
17032 (int) i, (int) SvIV(*element_ptr));
17035 PerlIO_printf(Perl_debug_log, "[%d] ", (int) i);
17036 sv_dump(*element_ptr);
17041 if (fence_stack_top < 0) {
17042 PerlIO_printf(Perl_debug_log, "Nothing on fence_stack\n");
17045 PerlIO_printf(Perl_debug_log, "Fence_stack: \n");
17046 for (i = fence_stack_top; i >= 0; i--) {
17047 SV ** element_ptr = av_fetch(fence_stack, i, FALSE);
17048 if (! element_ptr) {
17051 PerlIO_printf(Perl_debug_log, "[%d]: %d\n",
17052 (int) i, (int) SvIV(*element_ptr));
17063 S_add_above_Latin1_folds(pTHX_ RExC_state_t *pRExC_state, const U8 cp, SV** invlist)
17065 /* This adds the Latin1/above-Latin1 folding rules.
17067 * This should be called only for a Latin1-range code points, cp, which is
17068 * known to be involved in a simple fold with other code points above
17069 * Latin1. It would give false results if /aa has been specified.
17070 * Multi-char folds are outside the scope of this, and must be handled
17073 PERL_ARGS_ASSERT_ADD_ABOVE_LATIN1_FOLDS;
17075 assert(HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(cp));
17077 /* The rules that are valid for all Unicode versions are hard-coded in */
17082 add_cp_to_invlist(*invlist, KELVIN_SIGN);
17086 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_LONG_S);
17089 *invlist = add_cp_to_invlist(*invlist, GREEK_CAPITAL_LETTER_MU);
17090 *invlist = add_cp_to_invlist(*invlist, GREEK_SMALL_LETTER_MU);
17092 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
17093 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
17094 *invlist = add_cp_to_invlist(*invlist, ANGSTROM_SIGN);
17096 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
17097 *invlist = add_cp_to_invlist(*invlist,
17098 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
17101 default: /* Other code points are checked against the data for the
17102 current Unicode version */
17104 Size_t folds_count;
17106 const U32 * remaining_folds;
17110 folded_cp = toFOLD(cp);
17113 U8 dummy_fold[UTF8_MAXBYTES_CASE+1];
17115 folded_cp = _to_fold_latin1(cp, dummy_fold, &dummy_len, 0);
17118 if (folded_cp > 255) {
17119 *invlist = add_cp_to_invlist(*invlist, folded_cp);
17122 folds_count = _inverse_folds(folded_cp, &first_fold,
17124 if (folds_count == 0) {
17126 /* Use deprecated warning to increase the chances of this being
17128 ckWARN2reg_d(RExC_parse,
17129 "Perl folding rules are not up-to-date for 0x%02X;"
17130 " please use the perlbug utility to report;", cp);
17135 if (first_fold > 255) {
17136 *invlist = add_cp_to_invlist(*invlist, first_fold);
17138 for (i = 0; i < folds_count - 1; i++) {
17139 if (remaining_folds[i] > 255) {
17140 *invlist = add_cp_to_invlist(*invlist,
17141 remaining_folds[i]);
17151 S_output_posix_warnings(pTHX_ RExC_state_t *pRExC_state, AV* posix_warnings)
17153 /* Output the elements of the array given by '*posix_warnings' as REGEXP
17157 const bool first_is_fatal = ckDEAD(packWARN(WARN_REGEXP));
17159 PERL_ARGS_ASSERT_OUTPUT_POSIX_WARNINGS;
17161 if (! TO_OUTPUT_WARNINGS(RExC_parse)) {
17162 CLEAR_POSIX_WARNINGS();
17166 while ((msg = av_shift(posix_warnings)) != &PL_sv_undef) {
17167 if (first_is_fatal) { /* Avoid leaking this */
17168 av_undef(posix_warnings); /* This isn't necessary if the
17169 array is mortal, but is a
17171 (void) sv_2mortal(msg);
17174 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s", SvPVX(msg));
17175 SvREFCNT_dec_NN(msg);
17178 UPDATE_WARNINGS_LOC(RExC_parse);
17181 PERL_STATIC_INLINE Size_t
17182 S_find_first_differing_byte_pos(const U8 * s1, const U8 * s2, const Size_t max)
17184 const U8 * const start = s1;
17185 const U8 * const send = start + max;
17187 PERL_ARGS_ASSERT_FIND_FIRST_DIFFERING_BYTE_POS;
17189 while (s1 < send && *s1 == *s2) {
17198 S_add_multi_match(pTHX_ AV* multi_char_matches, SV* multi_string, const STRLEN cp_count)
17200 /* This adds the string scalar <multi_string> to the array
17201 * <multi_char_matches>. <multi_string> is known to have exactly
17202 * <cp_count> code points in it. This is used when constructing a
17203 * bracketed character class and we find something that needs to match more
17204 * than a single character.
17206 * <multi_char_matches> is actually an array of arrays. Each top-level
17207 * element is an array that contains all the strings known so far that are
17208 * the same length. And that length (in number of code points) is the same
17209 * as the index of the top-level array. Hence, the [2] element is an
17210 * array, each element thereof is a string containing TWO code points;
17211 * while element [3] is for strings of THREE characters, and so on. Since
17212 * this is for multi-char strings there can never be a [0] nor [1] element.
17214 * When we rewrite the character class below, we will do so such that the
17215 * longest strings are written first, so that it prefers the longest
17216 * matching strings first. This is done even if it turns out that any
17217 * quantifier is non-greedy, out of this programmer's (khw) laziness. Tom
17218 * Christiansen has agreed that this is ok. This makes the test for the
17219 * ligature 'ffi' come before the test for 'ff', for example */
17222 AV** this_array_ptr;
17224 PERL_ARGS_ASSERT_ADD_MULTI_MATCH;
17226 if (! multi_char_matches) {
17227 multi_char_matches = newAV();
17230 if (av_exists(multi_char_matches, cp_count)) {
17231 this_array_ptr = (AV**) av_fetch(multi_char_matches, cp_count, FALSE);
17232 this_array = *this_array_ptr;
17235 this_array = newAV();
17236 av_store(multi_char_matches, cp_count,
17239 av_push(this_array, multi_string);
17241 return multi_char_matches;
17244 /* The names of properties whose definitions are not known at compile time are
17245 * stored in this SV, after a constant heading. So if the length has been
17246 * changed since initialization, then there is a run-time definition. */
17247 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
17248 (SvCUR(listsv) != initial_listsv_len)
17250 /* There is a restricted set of white space characters that are legal when
17251 * ignoring white space in a bracketed character class. This generates the
17252 * code to skip them.
17254 * There is a line below that uses the same white space criteria but is outside
17255 * this macro. Both here and there must use the same definition */
17256 #define SKIP_BRACKETED_WHITE_SPACE(do_skip, p, stop_p) \
17259 while (p < stop_p && isBLANK_A(UCHARAT(p))) \
17266 STATIC regnode_offset
17267 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
17268 const bool stop_at_1, /* Just parse the next thing, don't
17269 look for a full character class */
17270 bool allow_mutiple_chars,
17271 const bool silence_non_portable, /* Don't output warnings
17275 bool optimizable, /* ? Allow a non-ANYOF return
17277 SV** ret_invlist /* Return an inversion list, not a node */
17280 /* parse a bracketed class specification. Most of these will produce an
17281 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
17282 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
17283 * under /i with multi-character folds: it will be rewritten following the
17284 * paradigm of this example, where the <multi-fold>s are characters which
17285 * fold to multiple character sequences:
17286 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
17287 * gets effectively rewritten as:
17288 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
17289 * reg() gets called (recursively) on the rewritten version, and this
17290 * function will return what it constructs. (Actually the <multi-fold>s
17291 * aren't physically removed from the [abcdefghi], it's just that they are
17292 * ignored in the recursion by means of a flag:
17293 * <RExC_in_multi_char_class>.)
17295 * ANYOF nodes contain a bit map for the first NUM_ANYOF_CODE_POINTS
17296 * characters, with the corresponding bit set if that character is in the
17297 * list. For characters above this, an inversion list is used. There
17298 * are extra bits for \w, etc. in locale ANYOFs, as what these match is not
17299 * determinable at compile time
17301 * On success, returns the offset at which any next node should be placed
17302 * into the regex engine program being compiled.
17304 * Returns 0 otherwise, setting flagp to RESTART_PARSE if the parse needs
17305 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to
17309 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
17311 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
17312 regnode_offset ret = -1; /* Initialized to an illegal value */
17314 int namedclass = OOB_NAMEDCLASS;
17315 char *rangebegin = NULL;
17316 SV *listsv = NULL; /* List of \p{user-defined} whose definitions
17317 aren't available at the time this was called */
17318 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
17319 than just initialized. */
17320 SV* properties = NULL; /* Code points that match \p{} \P{} */
17321 SV* posixes = NULL; /* Code points that match classes like [:word:],
17322 extended beyond the Latin1 range. These have to
17323 be kept separate from other code points for much
17324 of this function because their handling is
17325 different under /i, and for most classes under
17327 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
17328 separate for a while from the non-complemented
17329 versions because of complications with /d
17331 SV* simple_posixes = NULL; /* But under some conditions, the classes can be
17332 treated more simply than the general case,
17333 leading to less compilation and execution
17335 UV element_count = 0; /* Number of distinct elements in the class.
17336 Optimizations may be possible if this is tiny */
17337 AV * multi_char_matches = NULL; /* Code points that fold to more than one
17338 character; used under /i */
17340 char * stop_ptr = RExC_end; /* where to stop parsing */
17342 /* ignore unescaped whitespace? */
17343 const bool skip_white = cBOOL( ret_invlist
17344 || (RExC_flags & RXf_PMf_EXTENDED_MORE));
17346 /* inversion list of code points this node matches only when the target
17347 * string is in UTF-8. These are all non-ASCII, < 256. (Because is under
17349 SV* upper_latin1_only_utf8_matches = NULL;
17351 /* Inversion list of code points this node matches regardless of things
17352 * like locale, folding, utf8ness of the target string */
17353 SV* cp_list = NULL;
17355 /* Like cp_list, but code points on this list need to be checked for things
17356 * that fold to/from them under /i */
17357 SV* cp_foldable_list = NULL;
17359 /* Like cp_list, but code points on this list are valid only when the
17360 * runtime locale is UTF-8 */
17361 SV* only_utf8_locale_list = NULL;
17363 /* In a range, if one of the endpoints is non-character-set portable,
17364 * meaning that it hard-codes a code point that may mean a different
17365 * charactger in ASCII vs. EBCDIC, as opposed to, say, a literal 'A' or a
17366 * mnemonic '\t' which each mean the same character no matter which
17367 * character set the platform is on. */
17368 unsigned int non_portable_endpoint = 0;
17370 /* Is the range unicode? which means on a platform that isn't 1-1 native
17371 * to Unicode (i.e. non-ASCII), each code point in it should be considered
17372 * to be a Unicode value. */
17373 bool unicode_range = FALSE;
17374 bool invert = FALSE; /* Is this class to be complemented */
17376 bool warn_super = ALWAYS_WARN_SUPER;
17378 const char * orig_parse = RExC_parse;
17380 /* This variable is used to mark where the end in the input is of something
17381 * that looks like a POSIX construct but isn't. During the parse, when
17382 * something looks like it could be such a construct is encountered, it is
17383 * checked for being one, but not if we've already checked this area of the
17384 * input. Only after this position is reached do we check again */
17385 char *not_posix_region_end = RExC_parse - 1;
17387 AV* posix_warnings = NULL;
17388 const bool do_posix_warnings = ckWARN(WARN_REGEXP);
17389 U8 op = END; /* The returned node-type, initialized to an impossible
17391 U8 anyof_flags = 0; /* flag bits if the node is an ANYOF-type */
17392 U32 posixl = 0; /* bit field of posix classes matched under /l */
17395 /* Flags as to what things aren't knowable until runtime. (Note that these are
17396 * mutually exclusive.) */
17397 #define HAS_USER_DEFINED_PROPERTY 0x01 /* /u any user-defined properties that
17398 haven't been defined as of yet */
17399 #define HAS_D_RUNTIME_DEPENDENCY 0x02 /* /d if the target being matched is
17401 #define HAS_L_RUNTIME_DEPENDENCY 0x04 /* /l what the posix classes match and
17402 what gets folded */
17403 U32 has_runtime_dependency = 0; /* OR of the above flags */
17405 DECLARE_AND_GET_RE_DEBUG_FLAGS;
17407 PERL_ARGS_ASSERT_REGCLASS;
17409 PERL_UNUSED_ARG(depth);
17412 assert(! (ret_invlist && allow_mutiple_chars));
17414 /* If wants an inversion list returned, we can't optimize to something
17417 optimizable = FALSE;
17420 DEBUG_PARSE("clas");
17422 #if UNICODE_MAJOR_VERSION < 3 /* no multifolds in early Unicode */ \
17423 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0 \
17424 && UNICODE_DOT_DOT_VERSION == 0)
17425 allow_mutiple_chars = FALSE;
17428 /* We include the /i status at the beginning of this so that we can
17429 * know it at runtime */
17430 listsv = sv_2mortal(Perl_newSVpvf(aTHX_ "#%d\n", cBOOL(FOLD)));
17431 initial_listsv_len = SvCUR(listsv);
17432 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
17434 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse, RExC_end);
17436 assert(RExC_parse <= RExC_end);
17438 if (UCHARAT(RExC_parse) == '^') { /* Complement the class */
17441 allow_mutiple_chars = FALSE;
17443 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse, RExC_end);
17446 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
17447 if (! ret_invlist && MAYBE_POSIXCC(UCHARAT(RExC_parse))) {
17448 int maybe_class = handle_possible_posix(pRExC_state,
17450 ¬_posix_region_end,
17452 TRUE /* checking only */);
17453 if (maybe_class >= OOB_NAMEDCLASS && do_posix_warnings) {
17454 ckWARN4reg(not_posix_region_end,
17455 "POSIX syntax [%c %c] belongs inside character classes%s",
17456 *RExC_parse, *RExC_parse,
17457 (maybe_class == OOB_NAMEDCLASS)
17458 ? ((POSIXCC_NOTYET(*RExC_parse))
17459 ? " (but this one isn't implemented)"
17460 : " (but this one isn't fully valid)")
17466 /* If the caller wants us to just parse a single element, accomplish this
17467 * by faking the loop ending condition */
17468 if (stop_at_1 && RExC_end > RExC_parse) {
17469 stop_ptr = RExC_parse + 1;
17472 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
17473 if (UCHARAT(RExC_parse) == ']')
17474 goto charclassloop;
17478 if ( posix_warnings
17479 && av_tindex_skip_len_mg(posix_warnings) >= 0
17480 && RExC_parse > not_posix_region_end)
17482 /* Warnings about posix class issues are considered tentative until
17483 * we are far enough along in the parse that we can no longer
17484 * change our mind, at which point we output them. This is done
17485 * each time through the loop so that a later class won't zap them
17486 * before they have been dealt with. */
17487 output_posix_warnings(pRExC_state, posix_warnings);
17490 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse, RExC_end);
17492 if (RExC_parse >= stop_ptr) {
17496 if (UCHARAT(RExC_parse) == ']') {
17502 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
17503 save_value = value;
17504 save_prevvalue = prevvalue;
17507 rangebegin = RExC_parse;
17509 non_portable_endpoint = 0;
17511 if (UTF && ! UTF8_IS_INVARIANT(* RExC_parse)) {
17512 value = utf8n_to_uvchr((U8*)RExC_parse,
17513 RExC_end - RExC_parse,
17514 &numlen, UTF8_ALLOW_DEFAULT);
17515 RExC_parse += numlen;
17518 value = UCHARAT(RExC_parse++);
17520 if (value == '[') {
17521 char * posix_class_end;
17522 namedclass = handle_possible_posix(pRExC_state,
17525 do_posix_warnings ? &posix_warnings : NULL,
17526 FALSE /* die if error */);
17527 if (namedclass > OOB_NAMEDCLASS) {
17529 /* If there was an earlier attempt to parse this particular
17530 * posix class, and it failed, it was a false alarm, as this
17531 * successful one proves */
17532 if ( posix_warnings
17533 && av_tindex_skip_len_mg(posix_warnings) >= 0
17534 && not_posix_region_end >= RExC_parse
17535 && not_posix_region_end <= posix_class_end)
17537 av_undef(posix_warnings);
17540 RExC_parse = posix_class_end;
17542 else if (namedclass == OOB_NAMEDCLASS) {
17543 not_posix_region_end = posix_class_end;
17546 namedclass = OOB_NAMEDCLASS;
17549 else if ( RExC_parse - 1 > not_posix_region_end
17550 && MAYBE_POSIXCC(value))
17552 (void) handle_possible_posix(
17554 RExC_parse - 1, /* -1 because parse has already been
17556 ¬_posix_region_end,
17557 do_posix_warnings ? &posix_warnings : NULL,
17558 TRUE /* checking only */);
17560 else if ( strict && ! skip_white
17561 && ( _generic_isCC(value, _CC_VERTSPACE)
17562 || is_VERTWS_cp_high(value)))
17564 vFAIL("Literal vertical space in [] is illegal except under /x");
17566 else if (value == '\\') {
17567 /* Is a backslash; get the code point of the char after it */
17569 if (RExC_parse >= RExC_end) {
17570 vFAIL("Unmatched [");
17573 if (UTF && ! UTF8_IS_INVARIANT(UCHARAT(RExC_parse))) {
17574 value = utf8n_to_uvchr((U8*)RExC_parse,
17575 RExC_end - RExC_parse,
17576 &numlen, UTF8_ALLOW_DEFAULT);
17577 RExC_parse += numlen;
17580 value = UCHARAT(RExC_parse++);
17582 /* Some compilers cannot handle switching on 64-bit integer
17583 * values, therefore value cannot be an UV. Yes, this will
17584 * be a problem later if we want switch on Unicode.
17585 * A similar issue a little bit later when switching on
17586 * namedclass. --jhi */
17588 /* If the \ is escaping white space when white space is being
17589 * skipped, it means that that white space is wanted literally, and
17590 * is already in 'value'. Otherwise, need to translate the escape
17591 * into what it signifies. */
17592 if (! skip_white || ! isBLANK_A(value)) switch ((I32)value) {
17593 const char * message;
17597 case 'w': namedclass = ANYOF_WORDCHAR; break;
17598 case 'W': namedclass = ANYOF_NWORDCHAR; break;
17599 case 's': namedclass = ANYOF_SPACE; break;
17600 case 'S': namedclass = ANYOF_NSPACE; break;
17601 case 'd': namedclass = ANYOF_DIGIT; break;
17602 case 'D': namedclass = ANYOF_NDIGIT; break;
17603 case 'v': namedclass = ANYOF_VERTWS; break;
17604 case 'V': namedclass = ANYOF_NVERTWS; break;
17605 case 'h': namedclass = ANYOF_HORIZWS; break;
17606 case 'H': namedclass = ANYOF_NHORIZWS; break;
17607 case 'N': /* Handle \N{NAME} in class */
17609 const char * const backslash_N_beg = RExC_parse - 2;
17612 if (! grok_bslash_N(pRExC_state,
17613 NULL, /* No regnode */
17614 &value, /* Yes single value */
17615 &cp_count, /* Multiple code pt count */
17621 if (*flagp & NEED_UTF8)
17622 FAIL("panic: grok_bslash_N set NEED_UTF8");
17624 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
17626 if (cp_count < 0) {
17627 vFAIL("\\N in a character class must be a named character: \\N{...}");
17629 else if (cp_count == 0) {
17630 ckWARNreg(RExC_parse,
17631 "Ignoring zero length \\N{} in character class");
17633 else { /* cp_count > 1 */
17634 assert(cp_count > 1);
17635 if (! RExC_in_multi_char_class) {
17636 if ( ! allow_mutiple_chars
17639 || *RExC_parse == '-')
17643 vFAIL("\\N{} here is restricted to one character");
17645 ckWARNreg(RExC_parse, "Using just the first character returned by \\N{} in character class");
17646 break; /* <value> contains the first code
17647 point. Drop out of the switch to
17651 SV * multi_char_N = newSVpvn(backslash_N_beg,
17652 RExC_parse - backslash_N_beg);
17654 = add_multi_match(multi_char_matches,
17659 } /* End of cp_count != 1 */
17661 /* This element should not be processed further in this
17664 value = save_value;
17665 prevvalue = save_prevvalue;
17666 continue; /* Back to top of loop to get next char */
17669 /* Here, is a single code point, and <value> contains it */
17670 unicode_range = TRUE; /* \N{} are Unicode */
17678 if (RExC_pm_flags & PMf_WILDCARD) {
17680 /* diag_listed_as: Use of %s is not allowed in Unicode
17681 property wildcard subpatterns in regex; marked by <--
17683 vFAIL3("Use of '\\%c%c' is not allowed in Unicode property"
17684 " wildcard subpatterns", (char) value, *(RExC_parse - 1));
17687 /* \p means they want Unicode semantics */
17688 REQUIRE_UNI_RULES(flagp, 0);
17690 if (RExC_parse >= RExC_end)
17691 vFAIL2("Empty \\%c", (U8)value);
17692 if (*RExC_parse == '{') {
17693 const U8 c = (U8)value;
17694 e = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
17697 vFAIL2("Missing right brace on \\%c{}", c);
17702 /* White space is allowed adjacent to the braces and after
17703 * any '^', even when not under /x */
17704 while (isSPACE(*RExC_parse)) {
17708 if (UCHARAT(RExC_parse) == '^') {
17710 /* toggle. (The rhs xor gets the single bit that
17711 * differs between P and p; the other xor inverts just
17713 value ^= 'P' ^ 'p';
17716 while (isSPACE(*RExC_parse)) {
17721 if (e == RExC_parse)
17722 vFAIL2("Empty \\%c{}", c);
17724 n = e - RExC_parse;
17725 while (isSPACE(*(RExC_parse + n - 1)))
17728 } /* The \p isn't immediately followed by a '{' */
17729 else if (! isALPHA(*RExC_parse)) {
17730 RExC_parse += (UTF)
17731 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
17733 vFAIL2("Character following \\%c must be '{' or a "
17734 "single-character Unicode property name",
17742 char* name = RExC_parse;
17744 /* Any message returned about expanding the definition */
17745 SV* msg = newSVpvs_flags("", SVs_TEMP);
17747 /* If set TRUE, the property is user-defined as opposed to
17748 * official Unicode */
17749 bool user_defined = FALSE;
17750 AV * strings = NULL;
17752 SV * prop_definition = parse_uniprop_string(
17753 name, n, UTF, FOLD,
17754 FALSE, /* This is compile-time */
17756 /* We can't defer this defn when
17757 * the full result is required in
17759 ! cBOOL(ret_invlist),
17766 if (SvCUR(msg)) { /* Assumes any error causes a msg */
17767 assert(prop_definition == NULL);
17768 RExC_parse = e + 1;
17769 if (SvUTF8(msg)) { /* msg being UTF-8 makes the whole
17770 thing so, or else the display is
17774 /* diag_listed_as: Can't find Unicode property definition "%s" in regex; marked by <-- HERE in m/%s/ */
17775 vFAIL2utf8f("%" UTF8f, UTF8fARG(SvUTF8(msg),
17776 SvCUR(msg), SvPVX(msg)));
17779 assert(prop_definition || strings);
17783 if (! prop_definition) {
17784 RExC_parse = e + 1;
17785 vFAIL("Unicode string properties are not implemented in (?[...])");
17789 "Using just the single character results"
17790 " returned by \\p{} in (?[...])");
17793 else if (! RExC_in_multi_char_class) {
17794 if (invert ^ (value == 'P')) {
17795 RExC_parse = e + 1;
17796 vFAIL("Inverting a character class which contains"
17797 " a multi-character sequence is illegal");
17800 /* For each multi-character string ... */
17801 while (av_count(strings) > 0) {
17802 /* ... Each entry is itself an array of code
17804 AV * this_string = (AV *) av_shift( strings);
17805 STRLEN cp_count = av_count(this_string);
17806 SV * final = newSV(cp_count * 4);
17809 /* Create another string of sequences of \x{...} */
17810 while (av_count(this_string) > 0) {
17811 SV * character = av_shift(this_string);
17812 UV cp = SvUV(character);
17815 REQUIRE_UTF8(flagp);
17817 Perl_sv_catpvf(aTHX_ final, "\\x{%" UVXf "}",
17819 SvREFCNT_dec_NN(character);
17821 SvREFCNT_dec_NN(this_string);
17823 /* And add that to the list of such things */
17825 = add_multi_match(multi_char_matches,
17830 SvREFCNT_dec_NN(strings);
17833 if (! prop_definition) { /* If we got only a string,
17834 this iteration didn't really
17835 find a character */
17838 else if (! is_invlist(prop_definition)) {
17840 /* Here, the definition isn't known, so we have gotten
17841 * returned a string that will be evaluated if and when
17842 * encountered at runtime. We add it to the list of
17843 * such properties, along with whether it should be
17844 * complemented or not */
17845 if (value == 'P') {
17846 sv_catpvs(listsv, "!");
17849 sv_catpvs(listsv, "+");
17851 sv_catsv(listsv, prop_definition);
17853 has_runtime_dependency |= HAS_USER_DEFINED_PROPERTY;
17855 /* We don't know yet what this matches, so have to flag
17857 anyof_flags |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
17860 assert (prop_definition && is_invlist(prop_definition));
17862 /* Here we do have the complete property definition
17864 * Temporary workaround for [perl #133136]. For this
17865 * precise input that is in the .t that is failing,
17866 * load utf8.pm, which is what the test wants, so that
17867 * that .t passes */
17868 if ( memEQs(RExC_start, e + 1 - RExC_start,
17870 && ! hv_common(GvHVn(PL_incgv),
17872 "utf8.pm", sizeof("utf8.pm") - 1,
17873 0, HV_FETCH_ISEXISTS, NULL, 0))
17875 require_pv("utf8.pm");
17878 if (! user_defined &&
17879 /* We warn on matching an above-Unicode code point
17880 * if the match would return true, except don't
17881 * warn for \p{All}, which has exactly one element
17883 (_invlist_contains_cp(prop_definition, 0x110000)
17884 && (! (_invlist_len(prop_definition) == 1
17885 && *invlist_array(prop_definition) == 0))))
17890 /* Invert if asking for the complement */
17891 if (value == 'P') {
17892 _invlist_union_complement_2nd(properties,
17897 _invlist_union(properties, prop_definition, &properties);
17902 RExC_parse = e + 1;
17903 namedclass = ANYOF_UNIPROP; /* no official name, but it's
17907 case 'n': value = '\n'; break;
17908 case 'r': value = '\r'; break;
17909 case 't': value = '\t'; break;
17910 case 'f': value = '\f'; break;
17911 case 'b': value = '\b'; break;
17912 case 'e': value = ESC_NATIVE; break;
17913 case 'a': value = '\a'; break;
17915 RExC_parse--; /* function expects to be pointed at the 'o' */
17916 if (! grok_bslash_o(&RExC_parse,
17922 cBOOL(range), /* MAX_UV allowed for range
17928 else if (message && TO_OUTPUT_WARNINGS(RExC_parse)) {
17929 warn_non_literal_string(RExC_parse, packed_warn, message);
17933 non_portable_endpoint++;
17937 RExC_parse--; /* function expects to be pointed at the 'x' */
17938 if (! grok_bslash_x(&RExC_parse,
17944 cBOOL(range), /* MAX_UV allowed for range
17950 else if (message && TO_OUTPUT_WARNINGS(RExC_parse)) {
17951 warn_non_literal_string(RExC_parse, packed_warn, message);
17955 non_portable_endpoint++;
17959 if (! grok_bslash_c(*RExC_parse, &grok_c_char, &message,
17962 /* going to die anyway; point to exact spot of
17964 RExC_parse += (UTF)
17965 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
17970 value = grok_c_char;
17972 if (message && TO_OUTPUT_WARNINGS(RExC_parse)) {
17973 warn_non_literal_string(RExC_parse, packed_warn, message);
17976 non_portable_endpoint++;
17978 case '0': case '1': case '2': case '3': case '4':
17979 case '5': case '6': case '7':
17981 /* Take 1-3 octal digits */
17982 I32 flags = PERL_SCAN_SILENT_ILLDIGIT
17983 | PERL_SCAN_NOTIFY_ILLDIGIT;
17984 numlen = (strict) ? 4 : 3;
17985 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
17986 RExC_parse += numlen;
17989 RExC_parse += (UTF)
17990 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
17992 vFAIL("Need exactly 3 octal digits");
17994 else if ( (flags & PERL_SCAN_NOTIFY_ILLDIGIT)
17995 && RExC_parse < RExC_end
17996 && isDIGIT(*RExC_parse)
17997 && ckWARN(WARN_REGEXP))
17999 reg_warn_non_literal_string(
18001 form_alien_digit_msg(8, numlen, RExC_parse,
18002 RExC_end, UTF, FALSE));
18006 non_portable_endpoint++;
18011 /* Allow \_ to not give an error */
18012 if (isWORDCHAR(value) && value != '_') {
18014 vFAIL2("Unrecognized escape \\%c in character class",
18018 ckWARN2reg(RExC_parse,
18019 "Unrecognized escape \\%c in character class passed through",
18024 } /* End of switch on char following backslash */
18025 } /* end of handling backslash escape sequences */
18027 /* Here, we have the current token in 'value' */
18029 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
18032 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
18033 * literal, as is the character that began the false range, i.e.
18034 * the 'a' in the examples */
18036 const int w = (RExC_parse >= rangebegin)
18037 ? RExC_parse - rangebegin
18041 "False [] range \"%" UTF8f "\"",
18042 UTF8fARG(UTF, w, rangebegin));
18045 ckWARN2reg(RExC_parse,
18046 "False [] range \"%" UTF8f "\"",
18047 UTF8fARG(UTF, w, rangebegin));
18048 cp_list = add_cp_to_invlist(cp_list, '-');
18049 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
18053 range = 0; /* this was not a true range */
18054 element_count += 2; /* So counts for three values */
18057 classnum = namedclass_to_classnum(namedclass);
18059 if (LOC && namedclass < ANYOF_POSIXL_MAX
18060 #ifndef HAS_ISASCII
18061 && classnum != _CC_ASCII
18064 SV* scratch_list = NULL;
18066 /* What the Posix classes (like \w, [:space:]) match isn't
18067 * generally knowable under locale until actual match time. A
18068 * special node is used for these which has extra space for a
18069 * bitmap, with a bit reserved for each named class that is to
18070 * be matched against. (This isn't needed for \p{} and
18071 * pseudo-classes, as they are not affected by locale, and
18072 * hence are dealt with separately.) However, if a named class
18073 * and its complement are both present, then it matches
18074 * everything, and there is no runtime dependency. Odd numbers
18075 * are the complements of the next lower number, so xor works.
18076 * (Note that something like [\w\D] should match everything,
18077 * because \d should be a proper subset of \w. But rather than
18078 * trust that the locale is well behaved, we leave this to
18079 * runtime to sort out) */
18080 if (POSIXL_TEST(posixl, namedclass ^ 1)) {
18081 cp_list = _add_range_to_invlist(cp_list, 0, UV_MAX);
18082 POSIXL_ZERO(posixl);
18083 has_runtime_dependency &= ~HAS_L_RUNTIME_DEPENDENCY;
18084 anyof_flags &= ~ANYOF_MATCHES_POSIXL;
18085 continue; /* We could ignore the rest of the class, but
18086 best to parse it for any errors */
18088 else { /* Here, isn't the complement of any already parsed
18090 POSIXL_SET(posixl, namedclass);
18091 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
18092 anyof_flags |= ANYOF_MATCHES_POSIXL;
18094 /* The above-Latin1 characters are not subject to locale
18095 * rules. Just add them to the unconditionally-matched
18098 /* Get the list of the above-Latin1 code points this
18100 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
18101 PL_XPosix_ptrs[classnum],
18103 /* Odd numbers are complements,
18104 * like NDIGIT, NASCII, ... */
18105 namedclass % 2 != 0,
18107 /* Checking if 'cp_list' is NULL first saves an extra
18108 * clone. Its reference count will be decremented at the
18109 * next union, etc, or if this is the only instance, at the
18110 * end of the routine */
18112 cp_list = scratch_list;
18115 _invlist_union(cp_list, scratch_list, &cp_list);
18116 SvREFCNT_dec_NN(scratch_list);
18118 continue; /* Go get next character */
18123 /* Here, is not /l, or is a POSIX class for which /l doesn't
18124 * matter (or is a Unicode property, which is skipped here). */
18125 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
18126 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
18128 /* Here, should be \h, \H, \v, or \V. None of /d, /i
18129 * nor /l make a difference in what these match,
18130 * therefore we just add what they match to cp_list. */
18131 if (classnum != _CC_VERTSPACE) {
18132 assert( namedclass == ANYOF_HORIZWS
18133 || namedclass == ANYOF_NHORIZWS);
18135 /* It turns out that \h is just a synonym for
18137 classnum = _CC_BLANK;
18140 _invlist_union_maybe_complement_2nd(
18142 PL_XPosix_ptrs[classnum],
18143 namedclass % 2 != 0, /* Complement if odd
18144 (NHORIZWS, NVERTWS)
18149 else if ( AT_LEAST_UNI_SEMANTICS
18150 || classnum == _CC_ASCII
18151 || (DEPENDS_SEMANTICS && ( classnum == _CC_DIGIT
18152 || classnum == _CC_XDIGIT)))
18154 /* We usually have to worry about /d affecting what POSIX
18155 * classes match, with special code needed because we won't
18156 * know until runtime what all matches. But there is no
18157 * extra work needed under /u and /a; and [:ascii:] is
18158 * unaffected by /d; and :digit: and :xdigit: don't have
18159 * runtime differences under /d. So we can special case
18160 * these, and avoid some extra work below, and at runtime.
18162 _invlist_union_maybe_complement_2nd(
18164 ((AT_LEAST_ASCII_RESTRICTED)
18165 ? PL_Posix_ptrs[classnum]
18166 : PL_XPosix_ptrs[classnum]),
18167 namedclass % 2 != 0,
18170 else { /* Garden variety class. If is NUPPER, NALPHA, ...
18171 complement and use nposixes */
18172 SV** posixes_ptr = namedclass % 2 == 0
18175 _invlist_union_maybe_complement_2nd(
18177 PL_XPosix_ptrs[classnum],
18178 namedclass % 2 != 0,
18182 } /* end of namedclass \blah */
18184 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse, RExC_end);
18186 /* If 'range' is set, 'value' is the ending of a range--check its
18187 * validity. (If value isn't a single code point in the case of a
18188 * range, we should have figured that out above in the code that
18189 * catches false ranges). Later, we will handle each individual code
18190 * point in the range. If 'range' isn't set, this could be the
18191 * beginning of a range, so check for that by looking ahead to see if
18192 * the next real character to be processed is the range indicator--the
18197 /* For unicode ranges, we have to test that the Unicode as opposed
18198 * to the native values are not decreasing. (Above 255, there is
18199 * no difference between native and Unicode) */
18200 if (unicode_range && prevvalue < 255 && value < 255) {
18201 if (NATIVE_TO_LATIN1(prevvalue) > NATIVE_TO_LATIN1(value)) {
18202 goto backwards_range;
18207 if (prevvalue > value) /* b-a */ {
18212 w = RExC_parse - rangebegin;
18214 "Invalid [] range \"%" UTF8f "\"",
18215 UTF8fARG(UTF, w, rangebegin));
18216 NOT_REACHED; /* NOTREACHED */
18220 prevvalue = value; /* save the beginning of the potential range */
18221 if (! stop_at_1 /* Can't be a range if parsing just one thing */
18222 && *RExC_parse == '-')
18224 char* next_char_ptr = RExC_parse + 1;
18226 /* Get the next real char after the '-' */
18227 SKIP_BRACKETED_WHITE_SPACE(skip_white, next_char_ptr, RExC_end);
18229 /* If the '-' is at the end of the class (just before the ']',
18230 * it is a literal minus; otherwise it is a range */
18231 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
18232 RExC_parse = next_char_ptr;
18234 /* a bad range like \w-, [:word:]- ? */
18235 if (namedclass > OOB_NAMEDCLASS) {
18236 if (strict || ckWARN(WARN_REGEXP)) {
18237 const int w = RExC_parse >= rangebegin
18238 ? RExC_parse - rangebegin
18241 vFAIL4("False [] range \"%*.*s\"",
18246 "False [] range \"%*.*s\"",
18250 cp_list = add_cp_to_invlist(cp_list, '-');
18253 range = 1; /* yeah, it's a range! */
18254 continue; /* but do it the next time */
18259 if (namedclass > OOB_NAMEDCLASS) {
18263 /* Here, we have a single value this time through the loop, and
18264 * <prevvalue> is the beginning of the range, if any; or <value> if
18267 /* non-Latin1 code point implies unicode semantics. */
18269 if (value > MAX_LEGAL_CP && ( value != UV_MAX
18270 || prevvalue > MAX_LEGAL_CP))
18272 vFAIL(form_cp_too_large_msg(16, NULL, 0, value));
18274 REQUIRE_UNI_RULES(flagp, 0);
18275 if ( ! silence_non_portable
18276 && UNICODE_IS_PERL_EXTENDED(value)
18277 && TO_OUTPUT_WARNINGS(RExC_parse))
18279 ckWARN2_non_literal_string(RExC_parse,
18280 packWARN(WARN_PORTABLE),
18281 PL_extended_cp_format,
18286 /* Ready to process either the single value, or the completed range.
18287 * For single-valued non-inverted ranges, we consider the possibility
18288 * of multi-char folds. (We made a conscious decision to not do this
18289 * for the other cases because it can often lead to non-intuitive
18290 * results. For example, you have the peculiar case that:
18291 * "s s" =~ /^[^\xDF]+$/i => Y
18292 * "ss" =~ /^[^\xDF]+$/i => N
18294 * See [perl #89750] */
18295 if (FOLD && allow_mutiple_chars && value == prevvalue) {
18296 if ( value == LATIN_SMALL_LETTER_SHARP_S
18297 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
18300 /* Here <value> is indeed a multi-char fold. Get what it is */
18302 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
18305 UV folded = _to_uni_fold_flags(
18309 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
18310 ? FOLD_FLAGS_NOMIX_ASCII
18314 /* Here, <folded> should be the first character of the
18315 * multi-char fold of <value>, with <foldbuf> containing the
18316 * whole thing. But, if this fold is not allowed (because of
18317 * the flags), <fold> will be the same as <value>, and should
18318 * be processed like any other character, so skip the special
18320 if (folded != value) {
18322 /* Skip if we are recursed, currently parsing the class
18323 * again. Otherwise add this character to the list of
18324 * multi-char folds. */
18325 if (! RExC_in_multi_char_class) {
18326 STRLEN cp_count = utf8_length(foldbuf,
18327 foldbuf + foldlen);
18328 SV* multi_fold = sv_2mortal(newSVpvs(""));
18330 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%" UVXf "}", value);
18333 = add_multi_match(multi_char_matches,
18339 /* This element should not be processed further in this
18342 value = save_value;
18343 prevvalue = save_prevvalue;
18349 if (strict && ckWARN(WARN_REGEXP)) {
18352 /* If the range starts above 255, everything is portable and
18353 * likely to be so for any forseeable character set, so don't
18355 if (unicode_range && non_portable_endpoint && prevvalue < 256) {
18356 vWARN(RExC_parse, "Both or neither range ends should be Unicode");
18358 else if (prevvalue != value) {
18360 /* Under strict, ranges that stop and/or end in an ASCII
18361 * printable should have each end point be a portable value
18362 * for it (preferably like 'A', but we don't warn if it is
18363 * a (portable) Unicode name or code point), and the range
18364 * must be all digits or all letters of the same case.
18365 * Otherwise, the range is non-portable and unclear as to
18366 * what it contains */
18367 if ( (isPRINT_A(prevvalue) || isPRINT_A(value))
18368 && ( non_portable_endpoint
18369 || ! ( (isDIGIT_A(prevvalue) && isDIGIT_A(value))
18370 || (isLOWER_A(prevvalue) && isLOWER_A(value))
18371 || (isUPPER_A(prevvalue) && isUPPER_A(value))
18373 vWARN(RExC_parse, "Ranges of ASCII printables should"
18374 " be some subset of \"0-9\","
18375 " \"A-Z\", or \"a-z\"");
18377 else if (prevvalue >= FIRST_NON_ASCII_DECIMAL_DIGIT) {
18378 SSize_t index_start;
18379 SSize_t index_final;
18381 /* But the nature of Unicode and languages mean we
18382 * can't do the same checks for above-ASCII ranges,
18383 * except in the case of digit ones. These should
18384 * contain only digits from the same group of 10. The
18385 * ASCII case is handled just above. Hence here, the
18386 * range could be a range of digits. First some
18387 * unlikely special cases. Grandfather in that a range
18388 * ending in 19DA (NEW TAI LUE THAM DIGIT ONE) is bad
18389 * if its starting value is one of the 10 digits prior
18390 * to it. This is because it is an alternate way of
18391 * writing 19D1, and some people may expect it to be in
18392 * that group. But it is bad, because it won't give
18393 * the expected results. In Unicode 5.2 it was
18394 * considered to be in that group (of 11, hence), but
18395 * this was fixed in the next version */
18397 if (UNLIKELY(value == 0x19DA && prevvalue >= 0x19D0)) {
18398 goto warn_bad_digit_range;
18400 else if (UNLIKELY( prevvalue >= 0x1D7CE
18401 && value <= 0x1D7FF))
18403 /* This is the only other case currently in Unicode
18404 * where the algorithm below fails. The code
18405 * points just above are the end points of a single
18406 * range containing only decimal digits. It is 5
18407 * different series of 0-9. All other ranges of
18408 * digits currently in Unicode are just a single
18409 * series. (And mktables will notify us if a later
18410 * Unicode version breaks this.)
18412 * If the range being checked is at most 9 long,
18413 * and the digit values represented are in
18414 * numerical order, they are from the same series.
18416 if ( value - prevvalue > 9
18417 || ((( value - 0x1D7CE) % 10)
18418 <= (prevvalue - 0x1D7CE) % 10))
18420 goto warn_bad_digit_range;
18425 /* For all other ranges of digits in Unicode, the
18426 * algorithm is just to check if both end points
18427 * are in the same series, which is the same range.
18429 index_start = _invlist_search(
18430 PL_XPosix_ptrs[_CC_DIGIT],
18433 /* Warn if the range starts and ends with a digit,
18434 * and they are not in the same group of 10. */
18435 if ( index_start >= 0
18436 && ELEMENT_RANGE_MATCHES_INVLIST(index_start)
18438 _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
18439 value)) != index_start
18440 && index_final >= 0
18441 && ELEMENT_RANGE_MATCHES_INVLIST(index_final))
18443 warn_bad_digit_range:
18444 vWARN(RExC_parse, "Ranges of digits should be"
18445 " from the same group of"
18452 if ((! range || prevvalue == value) && non_portable_endpoint) {
18453 if (isPRINT_A(value)) {
18456 if (isBACKSLASHED_PUNCT(value)) {
18457 literal[d++] = '\\';
18459 literal[d++] = (char) value;
18460 literal[d++] = '\0';
18463 "\"%.*s\" is more clearly written simply as \"%s\"",
18464 (int) (RExC_parse - rangebegin),
18469 else if (isMNEMONIC_CNTRL(value)) {
18471 "\"%.*s\" is more clearly written simply as \"%s\"",
18472 (int) (RExC_parse - rangebegin),
18474 cntrl_to_mnemonic((U8) value)
18480 /* Deal with this element of the class */
18483 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
18486 /* On non-ASCII platforms, for ranges that span all of 0..255, and ones
18487 * that don't require special handling, we can just add the range like
18488 * we do for ASCII platforms */
18489 if ((UNLIKELY(prevvalue == 0) && value >= 255)
18490 || ! (prevvalue < 256
18492 || (! non_portable_endpoint
18493 && ((isLOWER_A(prevvalue) && isLOWER_A(value))
18494 || (isUPPER_A(prevvalue)
18495 && isUPPER_A(value)))))))
18497 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
18501 /* Here, requires special handling. This can be because it is a
18502 * range whose code points are considered to be Unicode, and so
18503 * must be individually translated into native, or because its a
18504 * subrange of 'A-Z' or 'a-z' which each aren't contiguous in
18505 * EBCDIC, but we have defined them to include only the "expected"
18506 * upper or lower case ASCII alphabetics. Subranges above 255 are
18507 * the same in native and Unicode, so can be added as a range */
18508 U8 start = NATIVE_TO_LATIN1(prevvalue);
18510 U8 end = (value < 256) ? NATIVE_TO_LATIN1(value) : 255;
18511 for (j = start; j <= end; j++) {
18512 cp_foldable_list = add_cp_to_invlist(cp_foldable_list, LATIN1_TO_NATIVE(j));
18515 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
18521 range = 0; /* this range (if it was one) is done now */
18522 } /* End of loop through all the text within the brackets */
18524 if ( posix_warnings && av_tindex_skip_len_mg(posix_warnings) >= 0) {
18525 output_posix_warnings(pRExC_state, posix_warnings);
18528 /* If anything in the class expands to more than one character, we have to
18529 * deal with them by building up a substitute parse string, and recursively
18530 * calling reg() on it, instead of proceeding */
18531 if (multi_char_matches) {
18532 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
18535 char *save_end = RExC_end;
18536 char *save_parse = RExC_parse;
18537 char *save_start = RExC_start;
18538 Size_t constructed_prefix_len = 0; /* This gives the length of the
18539 constructed portion of the
18540 substitute parse. */
18541 bool first_time = TRUE; /* First multi-char occurrence doesn't get
18546 /* Only one level of recursion allowed */
18547 assert(RExC_copy_start_in_constructed == RExC_precomp);
18549 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
18550 because too confusing */
18552 sv_catpvs(substitute_parse, "(?:");
18556 /* Look at the longest strings first */
18557 for (cp_count = av_tindex_skip_len_mg(multi_char_matches);
18562 if (av_exists(multi_char_matches, cp_count)) {
18563 AV** this_array_ptr;
18566 this_array_ptr = (AV**) av_fetch(multi_char_matches,
18568 while ((this_sequence = av_pop(*this_array_ptr)) !=
18571 if (! first_time) {
18572 sv_catpvs(substitute_parse, "|");
18574 first_time = FALSE;
18576 sv_catpv(substitute_parse, SvPVX(this_sequence));
18581 /* If the character class contains anything else besides these
18582 * multi-character strings, have to include it in recursive parsing */
18583 if (element_count) {
18584 bool has_l_bracket = orig_parse > RExC_start && *(orig_parse - 1) == '[';
18586 sv_catpvs(substitute_parse, "|");
18587 if (has_l_bracket) { /* Add an [ if the original had one */
18588 sv_catpvs(substitute_parse, "[");
18590 constructed_prefix_len = SvCUR(substitute_parse);
18591 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
18593 /* Put in a closing ']' to match any opening one, but not if going
18594 * off the end, as otherwise we are adding something that really
18596 if (has_l_bracket && RExC_parse < RExC_end) {
18597 sv_catpvs(substitute_parse, "]");
18601 sv_catpvs(substitute_parse, ")");
18604 /* This is a way to get the parse to skip forward a whole named
18605 * sequence instead of matching the 2nd character when it fails the
18607 sv_catpvs(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
18611 /* Set up the data structure so that any errors will be properly
18612 * reported. See the comments at the definition of
18613 * REPORT_LOCATION_ARGS for details */
18614 RExC_copy_start_in_input = (char *) orig_parse;
18615 RExC_start = RExC_parse = SvPV(substitute_parse, len);
18616 RExC_copy_start_in_constructed = RExC_start + constructed_prefix_len;
18617 RExC_end = RExC_parse + len;
18618 RExC_in_multi_char_class = 1;
18620 ret = reg(pRExC_state, 1, ®_flags, depth+1);
18622 *flagp |= reg_flags & (HASWIDTH|SIMPLE|POSTPONED|RESTART_PARSE|NEED_UTF8);
18624 /* And restore so can parse the rest of the pattern */
18625 RExC_parse = save_parse;
18626 RExC_start = RExC_copy_start_in_constructed = RExC_copy_start_in_input = save_start;
18627 RExC_end = save_end;
18628 RExC_in_multi_char_class = 0;
18629 SvREFCNT_dec_NN(multi_char_matches);
18633 /* If folding, we calculate all characters that could fold to or from the
18634 * ones already on the list */
18635 if (cp_foldable_list) {
18637 UV start, end; /* End points of code point ranges */
18639 SV* fold_intersection = NULL;
18642 /* Our calculated list will be for Unicode rules. For locale
18643 * matching, we have to keep a separate list that is consulted at
18644 * runtime only when the locale indicates Unicode rules (and we
18645 * don't include potential matches in the ASCII/Latin1 range, as
18646 * any code point could fold to any other, based on the run-time
18647 * locale). For non-locale, we just use the general list */
18649 use_list = &only_utf8_locale_list;
18652 use_list = &cp_list;
18655 /* Only the characters in this class that participate in folds need
18656 * be checked. Get the intersection of this class and all the
18657 * possible characters that are foldable. This can quickly narrow
18658 * down a large class */
18659 _invlist_intersection(PL_in_some_fold, cp_foldable_list,
18660 &fold_intersection);
18662 /* Now look at the foldable characters in this class individually */
18663 invlist_iterinit(fold_intersection);
18664 while (invlist_iternext(fold_intersection, &start, &end)) {
18668 /* Look at every character in the range */
18669 for (j = start; j <= end; j++) {
18670 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
18673 Size_t folds_count;
18675 const U32 * remaining_folds;
18679 /* Under /l, we don't know what code points below 256
18680 * fold to, except we do know the MICRO SIGN folds to
18681 * an above-255 character if the locale is UTF-8, so we
18682 * add it to the special list (in *use_list) Otherwise
18683 * we know now what things can match, though some folds
18684 * are valid under /d only if the target is UTF-8.
18685 * Those go in a separate list */
18686 if ( IS_IN_SOME_FOLD_L1(j)
18687 && ! (LOC && j != MICRO_SIGN))
18690 /* ASCII is always matched; non-ASCII is matched
18691 * only under Unicode rules (which could happen
18692 * under /l if the locale is a UTF-8 one */
18693 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
18694 *use_list = add_cp_to_invlist(*use_list,
18695 PL_fold_latin1[j]);
18697 else if (j != PL_fold_latin1[j]) {
18698 upper_latin1_only_utf8_matches
18699 = add_cp_to_invlist(
18700 upper_latin1_only_utf8_matches,
18701 PL_fold_latin1[j]);
18705 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(j)
18706 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
18708 add_above_Latin1_folds(pRExC_state,
18715 /* Here is an above Latin1 character. We don't have the
18716 * rules hard-coded for it. First, get its fold. This is
18717 * the simple fold, as the multi-character folds have been
18718 * handled earlier and separated out */
18719 folded = _to_uni_fold_flags(j, foldbuf, &foldlen,
18720 (ASCII_FOLD_RESTRICTED)
18721 ? FOLD_FLAGS_NOMIX_ASCII
18724 /* Single character fold of above Latin1. Add everything
18725 * in its fold closure to the list that this node should
18727 folds_count = _inverse_folds(folded, &first_fold,
18729 for (k = 0; k <= folds_count; k++) {
18730 UV c = (k == 0) /* First time through use itself */
18732 : (k == 1) /* 2nd time use, the first fold */
18735 /* Then the remaining ones */
18736 : remaining_folds[k-2];
18738 /* /aa doesn't allow folds between ASCII and non- */
18739 if (( ASCII_FOLD_RESTRICTED
18740 && (isASCII(c) != isASCII(j))))
18745 /* Folds under /l which cross the 255/256 boundary are
18746 * added to a separate list. (These are valid only
18747 * when the locale is UTF-8.) */
18748 if (c < 256 && LOC) {
18749 *use_list = add_cp_to_invlist(*use_list, c);
18753 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
18755 cp_list = add_cp_to_invlist(cp_list, c);
18758 /* Similarly folds involving non-ascii Latin1
18759 * characters under /d are added to their list */
18760 upper_latin1_only_utf8_matches
18761 = add_cp_to_invlist(
18762 upper_latin1_only_utf8_matches,
18768 SvREFCNT_dec_NN(fold_intersection);
18771 /* Now that we have finished adding all the folds, there is no reason
18772 * to keep the foldable list separate */
18773 _invlist_union(cp_list, cp_foldable_list, &cp_list);
18774 SvREFCNT_dec_NN(cp_foldable_list);
18777 /* And combine the result (if any) with any inversion lists from posix
18778 * classes. The lists are kept separate up to now because we don't want to
18779 * fold the classes */
18780 if (simple_posixes) { /* These are the classes known to be unaffected by
18783 _invlist_union(cp_list, simple_posixes, &cp_list);
18784 SvREFCNT_dec_NN(simple_posixes);
18787 cp_list = simple_posixes;
18790 if (posixes || nposixes) {
18791 if (! DEPENDS_SEMANTICS) {
18793 /* For everything but /d, we can just add the current 'posixes' and
18794 * 'nposixes' to the main list */
18797 _invlist_union(cp_list, posixes, &cp_list);
18798 SvREFCNT_dec_NN(posixes);
18806 _invlist_union(cp_list, nposixes, &cp_list);
18807 SvREFCNT_dec_NN(nposixes);
18810 cp_list = nposixes;
18815 /* Under /d, things like \w match upper Latin1 characters only if
18816 * the target string is in UTF-8. But things like \W match all the
18817 * upper Latin1 characters if the target string is not in UTF-8.
18819 * Handle the case with something like \W separately */
18821 SV* only_non_utf8_list = invlist_clone(PL_UpperLatin1, NULL);
18823 /* A complemented posix class matches all upper Latin1
18824 * characters if not in UTF-8. And it matches just certain
18825 * ones when in UTF-8. That means those certain ones are
18826 * matched regardless, so can just be added to the
18827 * unconditional list */
18829 _invlist_union(cp_list, nposixes, &cp_list);
18830 SvREFCNT_dec_NN(nposixes);
18834 cp_list = nposixes;
18837 /* Likewise for 'posixes' */
18838 _invlist_union(posixes, cp_list, &cp_list);
18839 SvREFCNT_dec(posixes);
18841 /* Likewise for anything else in the range that matched only
18843 if (upper_latin1_only_utf8_matches) {
18844 _invlist_union(cp_list,
18845 upper_latin1_only_utf8_matches,
18847 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
18848 upper_latin1_only_utf8_matches = NULL;
18851 /* If we don't match all the upper Latin1 characters regardless
18852 * of UTF-8ness, we have to set a flag to match the rest when
18854 _invlist_subtract(only_non_utf8_list, cp_list,
18855 &only_non_utf8_list);
18856 if (_invlist_len(only_non_utf8_list) != 0) {
18857 anyof_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
18859 SvREFCNT_dec_NN(only_non_utf8_list);
18862 /* Here there were no complemented posix classes. That means
18863 * the upper Latin1 characters in 'posixes' match only when the
18864 * target string is in UTF-8. So we have to add them to the
18865 * list of those types of code points, while adding the
18866 * remainder to the unconditional list.
18868 * First calculate what they are */
18869 SV* nonascii_but_latin1_properties = NULL;
18870 _invlist_intersection(posixes, PL_UpperLatin1,
18871 &nonascii_but_latin1_properties);
18873 /* And add them to the final list of such characters. */
18874 _invlist_union(upper_latin1_only_utf8_matches,
18875 nonascii_but_latin1_properties,
18876 &upper_latin1_only_utf8_matches);
18878 /* Remove them from what now becomes the unconditional list */
18879 _invlist_subtract(posixes, nonascii_but_latin1_properties,
18882 /* And add those unconditional ones to the final list */
18884 _invlist_union(cp_list, posixes, &cp_list);
18885 SvREFCNT_dec_NN(posixes);
18892 SvREFCNT_dec(nonascii_but_latin1_properties);
18894 /* Get rid of any characters from the conditional list that we
18895 * now know are matched unconditionally, which may make that
18897 _invlist_subtract(upper_latin1_only_utf8_matches,
18899 &upper_latin1_only_utf8_matches);
18900 if (_invlist_len(upper_latin1_only_utf8_matches) == 0) {
18901 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
18902 upper_latin1_only_utf8_matches = NULL;
18908 /* And combine the result (if any) with any inversion list from properties.
18909 * The lists are kept separate up to now so that we can distinguish the two
18910 * in regards to matching above-Unicode. A run-time warning is generated
18911 * if a Unicode property is matched against a non-Unicode code point. But,
18912 * we allow user-defined properties to match anything, without any warning,
18913 * and we also suppress the warning if there is a portion of the character
18914 * class that isn't a Unicode property, and which matches above Unicode, \W
18915 * or [\x{110000}] for example.
18916 * (Note that in this case, unlike the Posix one above, there is no
18917 * <upper_latin1_only_utf8_matches>, because having a Unicode property
18918 * forces Unicode semantics */
18922 /* If it matters to the final outcome, see if a non-property
18923 * component of the class matches above Unicode. If so, the
18924 * warning gets suppressed. This is true even if just a single
18925 * such code point is specified, as, though not strictly correct if
18926 * another such code point is matched against, the fact that they
18927 * are using above-Unicode code points indicates they should know
18928 * the issues involved */
18930 warn_super = ! (invert
18931 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
18934 _invlist_union(properties, cp_list, &cp_list);
18935 SvREFCNT_dec_NN(properties);
18938 cp_list = properties;
18943 |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
18945 /* Because an ANYOF node is the only one that warns, this node
18946 * can't be optimized into something else */
18947 optimizable = FALSE;
18951 /* Here, we have calculated what code points should be in the character
18954 * Now we can see about various optimizations. Fold calculation (which we
18955 * did above) needs to take place before inversion. Otherwise /[^k]/i
18956 * would invert to include K, which under /i would match k, which it
18957 * shouldn't. Therefore we can't invert folded locale now, as it won't be
18958 * folded until runtime */
18960 /* If we didn't do folding, it's because some information isn't available
18961 * until runtime; set the run-time fold flag for these We know to set the
18962 * flag if we have a non-NULL list for UTF-8 locales, or the class matches
18963 * at least one 0-255 range code point */
18966 /* Some things on the list might be unconditionally included because of
18967 * other components. Remove them, and clean up the list if it goes to
18969 if (only_utf8_locale_list && cp_list) {
18970 _invlist_subtract(only_utf8_locale_list, cp_list,
18971 &only_utf8_locale_list);
18973 if (_invlist_len(only_utf8_locale_list) == 0) {
18974 SvREFCNT_dec_NN(only_utf8_locale_list);
18975 only_utf8_locale_list = NULL;
18978 if ( only_utf8_locale_list
18979 || (cp_list && ( _invlist_contains_cp(cp_list, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE)
18980 || _invlist_contains_cp(cp_list, LATIN_SMALL_LETTER_DOTLESS_I))))
18982 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
18985 | ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
18987 else if (cp_list && invlist_lowest(cp_list) < 256) {
18988 /* If nothing is below 256, has no locale dependency; otherwise it
18990 anyof_flags |= ANYOFL_FOLD;
18991 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
18994 else if ( DEPENDS_SEMANTICS
18995 && ( upper_latin1_only_utf8_matches
18996 || (anyof_flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)))
18998 RExC_seen_d_op = TRUE;
18999 has_runtime_dependency |= HAS_D_RUNTIME_DEPENDENCY;
19002 /* Optimize inverted patterns (e.g. [^a-z]) when everything is known at
19006 && ! has_runtime_dependency)
19008 _invlist_invert(cp_list);
19010 /* Clear the invert flag since have just done it here */
19014 /* All possible optimizations below still have these characteristics.
19015 * (Multi-char folds aren't SIMPLE, but they don't get this far in this
19017 *flagp |= HASWIDTH|SIMPLE;
19020 *ret_invlist = cp_list;
19022 return (cp_list) ? RExC_emit : 0;
19025 if (anyof_flags & ANYOF_LOCALE_FLAGS) {
19026 RExC_contains_locale = 1;
19029 /* Some character classes are equivalent to other nodes. Such nodes take
19030 * up less room, and some nodes require fewer operations to execute, than
19031 * ANYOF nodes. EXACTish nodes may be joinable with adjacent nodes to
19032 * improve efficiency. */
19035 PERL_UINT_FAST8_T i;
19036 UV partial_cp_count = 0;
19037 UV start[MAX_FOLD_FROMS+1] = { 0 }; /* +1 for the folded-to char */
19038 UV end[MAX_FOLD_FROMS+1] = { 0 };
19039 bool single_range = FALSE;
19041 if (cp_list) { /* Count the code points in enough ranges that we would
19042 see all the ones possible in any fold in this version
19045 invlist_iterinit(cp_list);
19046 for (i = 0; i <= MAX_FOLD_FROMS; i++) {
19047 if (! invlist_iternext(cp_list, &start[i], &end[i])) {
19050 partial_cp_count += end[i] - start[i] + 1;
19054 single_range = TRUE;
19056 invlist_iterfinish(cp_list);
19059 /* If we know at compile time that this matches every possible code
19060 * point, any run-time dependencies don't matter */
19061 if (start[0] == 0 && end[0] == UV_MAX) {
19063 ret = reganode(pRExC_state, OPFAIL, 0);
19066 ret = reg_node(pRExC_state, SANY);
19072 /* Similarly, for /l posix classes, if both a class and its
19073 * complement match, any run-time dependencies don't matter */
19075 for (namedclass = 0; namedclass < ANYOF_POSIXL_MAX;
19078 if ( POSIXL_TEST(posixl, namedclass) /* class */
19079 && POSIXL_TEST(posixl, namedclass + 1)) /* its complement */
19082 ret = reganode(pRExC_state, OPFAIL, 0);
19085 ret = reg_node(pRExC_state, SANY);
19092 /* For well-behaved locales, some classes are subsets of others,
19093 * so complementing the subset and including the non-complemented
19094 * superset should match everything, like [\D[:alnum:]], and
19095 * [[:^alpha:][:alnum:]], but some implementations of locales are
19096 * buggy, and khw thinks its a bad idea to have optimization change
19097 * behavior, even if it avoids an OS bug in a given case */
19099 #define isSINGLE_BIT_SET(n) isPOWER_OF_2(n)
19101 /* If is a single posix /l class, can optimize to just that op.
19102 * Such a node will not match anything in the Latin1 range, as that
19103 * is not determinable until runtime, but will match whatever the
19104 * class does outside that range. (Note that some classes won't
19105 * match anything outside the range, like [:ascii:]) */
19106 if ( isSINGLE_BIT_SET(posixl)
19107 && (partial_cp_count == 0 || start[0] > 255))
19110 SV * class_above_latin1 = NULL;
19111 bool already_inverted;
19112 bool are_equivalent;
19114 /* Compute which bit is set, which is the same thing as, e.g.,
19115 * ANYOF_CNTRL. From
19116 * https://graphics.stanford.edu/~seander/bithacks.html#IntegerLogDeBruijn
19118 static const int MultiplyDeBruijnBitPosition2[32] =
19120 0, 1, 28, 2, 29, 14, 24, 3, 30, 22, 20, 15, 25, 17, 4, 8,
19121 31, 27, 13, 23, 21, 19, 16, 7, 26, 12, 18, 6, 11, 5, 10, 9
19124 namedclass = MultiplyDeBruijnBitPosition2[(posixl
19125 * 0x077CB531U) >> 27];
19126 classnum = namedclass_to_classnum(namedclass);
19128 /* The named classes are such that the inverted number is one
19129 * larger than the non-inverted one */
19130 already_inverted = namedclass
19131 - classnum_to_namedclass(classnum);
19133 /* Create an inversion list of the official property, inverted
19134 * if the constructed node list is inverted, and restricted to
19135 * only the above latin1 code points, which are the only ones
19136 * known at compile time */
19137 _invlist_intersection_maybe_complement_2nd(
19139 PL_XPosix_ptrs[classnum],
19141 &class_above_latin1);
19142 are_equivalent = _invlistEQ(class_above_latin1, cp_list,
19144 SvREFCNT_dec_NN(class_above_latin1);
19146 if (are_equivalent) {
19148 /* Resolve the run-time inversion flag with this possibly
19149 * inverted class */
19150 invert = invert ^ already_inverted;
19152 ret = reg_node(pRExC_state,
19153 POSIXL + invert * (NPOSIXL - POSIXL));
19154 FLAGS(REGNODE_p(ret)) = classnum;
19160 /* khw can't think of any other possible transformation involving
19162 if (has_runtime_dependency & HAS_USER_DEFINED_PROPERTY) {
19166 if (! has_runtime_dependency) {
19168 /* If the list is empty, nothing matches. This happens, for
19169 * example, when a Unicode property that doesn't match anything is
19170 * the only element in the character class (perluniprops.pod notes
19171 * such properties). */
19172 if (partial_cp_count == 0) {
19174 ret = reg_node(pRExC_state, SANY);
19177 ret = reganode(pRExC_state, OPFAIL, 0);
19183 /* If matches everything but \n */
19184 if ( start[0] == 0 && end[0] == '\n' - 1
19185 && start[1] == '\n' + 1 && end[1] == UV_MAX)
19188 ret = reg_node(pRExC_state, REG_ANY);
19194 /* Next see if can optimize classes that contain just a few code points
19195 * into an EXACTish node. The reason to do this is to let the
19196 * optimizer join this node with adjacent EXACTish ones, and ANYOF
19197 * nodes require conversion to code point from UTF-8.
19199 * An EXACTFish node can be generated even if not under /i, and vice
19200 * versa. But care must be taken. An EXACTFish node has to be such
19201 * that it only matches precisely the code points in the class, but we
19202 * want to generate the least restrictive one that does that, to
19203 * increase the odds of being able to join with an adjacent node. For
19204 * example, if the class contains [kK], we have to make it an EXACTFAA
19205 * node to prevent the KELVIN SIGN from matching. Whether we are under
19206 * /i or not is irrelevant in this case. Less obvious is the pattern
19207 * qr/[\x{02BC}]n/i. U+02BC is MODIFIER LETTER APOSTROPHE. That is
19208 * supposed to match the single character U+0149 LATIN SMALL LETTER N
19209 * PRECEDED BY APOSTROPHE. And so even though there is no simple fold
19210 * that includes \X{02BC}, there is a multi-char fold that does, and so
19211 * the node generated for it must be an EXACTFish one. On the other
19212 * hand qr/:/i should generate a plain EXACT node since the colon
19213 * participates in no fold whatsoever, and having it EXACT tells the
19214 * optimizer the target string cannot match unless it has a colon in
19220 /* Only try if there are no more code points in the class than
19221 * in the max possible fold */
19222 && inRANGE(partial_cp_count, 1, MAX_FOLD_FROMS + 1))
19224 if (partial_cp_count == 1 && ! upper_latin1_only_utf8_matches)
19226 /* We can always make a single code point class into an
19227 * EXACTish node. */
19231 /* Here is /l: Use EXACTL, except if there is a fold not
19232 * known until runtime so shows as only a single code point
19233 * here. For code points above 255, we know which can
19234 * cause problems by having a potential fold to the Latin1
19237 || ( start[0] > 255
19238 && ! is_PROBLEMATIC_LOCALE_FOLD_cp(start[0])))
19246 else if (! FOLD) { /* Not /l and not /i */
19247 op = (start[0] < 256) ? EXACT : EXACT_REQ8;
19249 else if (start[0] < 256) { /* /i, not /l, and the code point is
19252 /* Under /i, it gets a little tricky. A code point that
19253 * doesn't participate in a fold should be an EXACT node.
19254 * We know this one isn't the result of a simple fold, or
19255 * there'd be more than one code point in the list, but it
19256 * could be part of a multi- character fold. In that case
19257 * we better not create an EXACT node, as we would wrongly
19258 * be telling the optimizer that this code point must be in
19259 * the target string, and that is wrong. This is because
19260 * if the sequence around this code point forms a
19261 * multi-char fold, what needs to be in the string could be
19262 * the code point that folds to the sequence.
19264 * This handles the case of below-255 code points, as we
19265 * have an easy look up for those. The next clause handles
19266 * the above-256 one */
19267 op = IS_IN_SOME_FOLD_L1(start[0])
19271 else { /* /i, larger code point. Since we are under /i, and
19272 have just this code point, we know that it can't
19273 fold to something else, so PL_InMultiCharFold
19275 op = _invlist_contains_cp(PL_InMultiCharFold,
19283 else if ( ! (has_runtime_dependency & ~HAS_D_RUNTIME_DEPENDENCY)
19284 && _invlist_contains_cp(PL_in_some_fold, start[0]))
19286 /* Here, the only runtime dependency, if any, is from /d, and
19287 * the class matches more than one code point, and the lowest
19288 * code point participates in some fold. It might be that the
19289 * other code points are /i equivalent to this one, and hence
19290 * they would representable by an EXACTFish node. Above, we
19291 * eliminated classes that contain too many code points to be
19292 * EXACTFish, with the test for MAX_FOLD_FROMS
19294 * First, special case the ASCII fold pairs, like 'B' and 'b'.
19295 * We do this because we have EXACTFAA at our disposal for the
19297 if (partial_cp_count == 2 && isASCII(start[0])) {
19299 /* The only ASCII characters that participate in folds are
19301 assert(isALPHA(start[0]));
19302 if ( end[0] == start[0] /* First range is a single
19303 character, so 2nd exists */
19304 && isALPHA_FOLD_EQ(start[0], start[1]))
19307 /* Here, is part of an ASCII fold pair */
19309 if ( ASCII_FOLD_RESTRICTED
19310 || HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(start[0]))
19312 /* If the second clause just above was true, it
19313 * means we can't be under /i, or else the list
19314 * would have included more than this fold pair.
19315 * Therefore we have to exclude the possibility of
19316 * whatever else it is that folds to these, by
19317 * using EXACTFAA */
19320 else if (HAS_NONLATIN1_FOLD_CLOSURE(start[0])) {
19322 /* Here, there's no simple fold that start[0] is part
19323 * of, but there is a multi-character one. If we
19324 * are not under /i, we want to exclude that
19325 * possibility; if under /i, we want to include it
19327 op = (FOLD) ? EXACTFU : EXACTFAA;
19331 /* Here, the only possible fold start[0] particpates in
19332 * is with start[1]. /i or not isn't relevant */
19336 value = toFOLD(start[0]);
19339 else if ( ! upper_latin1_only_utf8_matches
19340 || ( _invlist_len(upper_latin1_only_utf8_matches)
19343 invlist_highest(upper_latin1_only_utf8_matches)]
19346 /* Here, the smallest character is non-ascii or there are
19347 * more than 2 code points matched by this node. Also, we
19348 * either don't have /d UTF-8 dependent matches, or if we
19349 * do, they look like they could be a single character that
19350 * is the fold of the lowest one in the always-match list.
19351 * This test quickly excludes most of the false positives
19352 * when there are /d UTF-8 depdendent matches. These are
19353 * like LATIN CAPITAL LETTER A WITH GRAVE matching LATIN
19354 * SMALL LETTER A WITH GRAVE iff the target string is
19355 * UTF-8. (We don't have to worry above about exceeding
19356 * the array bounds of PL_fold_latin1[] because any code
19357 * point in 'upper_latin1_only_utf8_matches' is below 256.)
19359 * EXACTFAA would apply only to pairs (hence exactly 2 code
19360 * points) in the ASCII range, so we can't use it here to
19361 * artificially restrict the fold domain, so we check if
19362 * the class does or does not match some EXACTFish node.
19363 * Further, if we aren't under /i, and the folded-to
19364 * character is part of a multi-character fold, we can't do
19365 * this optimization, as the sequence around it could be
19366 * that multi-character fold, and we don't here know the
19367 * context, so we have to assume it is that multi-char
19368 * fold, to prevent potential bugs.
19370 * To do the general case, we first find the fold of the
19371 * lowest code point (which may be higher than the lowest
19372 * one), then find everything that folds to it. (The data
19373 * structure we have only maps from the folded code points,
19374 * so we have to do the earlier step.) */
19377 U8 foldbuf[UTF8_MAXBYTES_CASE];
19378 UV folded = _to_uni_fold_flags(start[0],
19379 foldbuf, &foldlen, 0);
19381 const U32 * remaining_folds;
19382 Size_t folds_to_this_cp_count = _inverse_folds(
19386 Size_t folds_count = folds_to_this_cp_count + 1;
19387 SV * fold_list = _new_invlist(folds_count);
19390 /* If there are UTF-8 dependent matches, create a temporary
19391 * list of what this node matches, including them. */
19392 SV * all_cp_list = NULL;
19393 SV ** use_this_list = &cp_list;
19395 if (upper_latin1_only_utf8_matches) {
19396 all_cp_list = _new_invlist(0);
19397 use_this_list = &all_cp_list;
19398 _invlist_union(cp_list,
19399 upper_latin1_only_utf8_matches,
19403 /* Having gotten everything that participates in the fold
19404 * containing the lowest code point, we turn that into an
19405 * inversion list, making sure everything is included. */
19406 fold_list = add_cp_to_invlist(fold_list, start[0]);
19407 fold_list = add_cp_to_invlist(fold_list, folded);
19408 if (folds_to_this_cp_count > 0) {
19409 fold_list = add_cp_to_invlist(fold_list, first_fold);
19410 for (i = 0; i + 1 < folds_to_this_cp_count; i++) {
19411 fold_list = add_cp_to_invlist(fold_list,
19412 remaining_folds[i]);
19416 /* If the fold list is identical to what's in this ANYOF
19417 * node, the node can be represented by an EXACTFish one
19419 if (_invlistEQ(*use_this_list, fold_list,
19420 0 /* Don't complement */ )
19423 /* But, we have to be careful, as mentioned above.
19424 * Just the right sequence of characters could match
19425 * this if it is part of a multi-character fold. That
19426 * IS what we want if we are under /i. But it ISN'T
19427 * what we want if not under /i, as it could match when
19428 * it shouldn't. So, when we aren't under /i and this
19429 * character participates in a multi-char fold, we
19430 * don't optimize into an EXACTFish node. So, for each
19431 * case below we have to check if we are folding
19432 * and if not, if it is not part of a multi-char fold.
19434 if (start[0] > 255) { /* Highish code point */
19435 if (FOLD || ! _invlist_contains_cp(
19436 PL_InMultiCharFold, folded))
19440 : (ASCII_FOLD_RESTRICTED)
19445 } /* Below, the lowest code point < 256 */
19448 && DEPENDS_SEMANTICS)
19449 { /* An EXACTF node containing a single character
19450 's', can be an EXACTFU if it doesn't get
19451 joined with an adjacent 's' */
19452 op = EXACTFU_S_EDGE;
19456 || ! HAS_NONLATIN1_FOLD_CLOSURE(start[0]))
19458 if (upper_latin1_only_utf8_matches) {
19461 /* We can't use the fold, as that only matches
19465 else if ( UNLIKELY(start[0] == MICRO_SIGN)
19467 { /* EXACTFUP is a special node for this
19469 op = (ASCII_FOLD_RESTRICTED)
19472 value = MICRO_SIGN;
19474 else if ( ASCII_FOLD_RESTRICTED
19475 && ! isASCII(start[0]))
19476 { /* For ASCII under /iaa, we can use EXACTFU
19488 SvREFCNT_dec_NN(fold_list);
19489 SvREFCNT_dec(all_cp_list);
19496 /* Here, we have calculated what EXACTish node to use. Have to
19497 * convert to UTF-8 if not already there */
19500 SvREFCNT_dec(cp_list);;
19501 REQUIRE_UTF8(flagp);
19504 /* This is a kludge to the special casing issues with this
19505 * ligature under /aa. FB05 should fold to FB06, but the
19506 * call above to _to_uni_fold_flags() didn't find this, as
19507 * it didn't use the /aa restriction in order to not miss
19508 * other folds that would be affected. This is the only
19509 * instance likely to ever be a problem in all of Unicode.
19510 * So special case it. */
19511 if ( value == LATIN_SMALL_LIGATURE_LONG_S_T
19512 && ASCII_FOLD_RESTRICTED)
19514 value = LATIN_SMALL_LIGATURE_ST;
19518 len = (UTF) ? UVCHR_SKIP(value) : 1;
19520 ret = regnode_guts(pRExC_state, op, len, "exact");
19521 FILL_NODE(ret, op);
19522 RExC_emit += 1 + STR_SZ(len);
19523 setSTR_LEN(REGNODE_p(ret), len);
19525 *STRINGs(REGNODE_p(ret)) = (U8) value;
19528 uvchr_to_utf8((U8 *) STRINGs(REGNODE_p(ret)), value);
19534 if (! has_runtime_dependency) {
19536 /* See if this can be turned into an ANYOFM node. Think about the
19537 * bit patterns in two different bytes. In some positions, the
19538 * bits in each will be 1; and in other positions both will be 0;
19539 * and in some positions the bit will be 1 in one byte, and 0 in
19540 * the other. Let 'n' be the number of positions where the bits
19541 * differ. We create a mask which has exactly 'n' 0 bits, each in
19542 * a position where the two bytes differ. Now take the set of all
19543 * bytes that when ANDed with the mask yield the same result. That
19544 * set has 2**n elements, and is representable by just two 8 bit
19545 * numbers: the result and the mask. Importantly, matching the set
19546 * can be vectorized by creating a word full of the result bytes,
19547 * and a word full of the mask bytes, yielding a significant speed
19548 * up. Here, see if this node matches such a set. As a concrete
19549 * example consider [01], and the byte representing '0' which is
19550 * 0x30 on ASCII machines. It has the bits 0011 0000. Take the
19551 * mask 1111 1110. If we AND 0x31 and 0x30 with that mask we get
19552 * 0x30. Any other bytes ANDed yield something else. So [01],
19553 * which is a common usage, is optimizable into ANYOFM, and can
19554 * benefit from the speed up. We can only do this on UTF-8
19555 * invariant bytes, because they have the same bit patterns under
19557 PERL_UINT_FAST8_T inverted = 0;
19559 const PERL_UINT_FAST8_T max_permissible = 0xFF;
19561 const PERL_UINT_FAST8_T max_permissible = 0x7F;
19563 /* If doesn't fit the criteria for ANYOFM, invert and try again.
19564 * If that works we will instead later generate an NANYOFM, and
19565 * invert back when through */
19566 if (invlist_highest(cp_list) > max_permissible) {
19567 _invlist_invert(cp_list);
19571 if (invlist_highest(cp_list) <= max_permissible) {
19572 UV this_start, this_end;
19573 UV lowest_cp = UV_MAX; /* init'ed to suppress compiler warn */
19574 U8 bits_differing = 0;
19575 Size_t full_cp_count = 0;
19576 bool first_time = TRUE;
19578 /* Go through the bytes and find the bit positions that differ
19580 invlist_iterinit(cp_list);
19581 while (invlist_iternext(cp_list, &this_start, &this_end)) {
19582 unsigned int i = this_start;
19585 if (! UVCHR_IS_INVARIANT(i)) {
19589 first_time = FALSE;
19590 lowest_cp = this_start;
19592 /* We have set up the code point to compare with.
19593 * Don't compare it with itself */
19597 /* Find the bit positions that differ from the lowest code
19598 * point in the node. Keep track of all such positions by
19600 for (; i <= this_end; i++) {
19601 if (! UVCHR_IS_INVARIANT(i)) {
19605 bits_differing |= i ^ lowest_cp;
19608 full_cp_count += this_end - this_start + 1;
19611 /* At the end of the loop, we count how many bits differ from
19612 * the bits in lowest code point, call the count 'd'. If the
19613 * set we found contains 2**d elements, it is the closure of
19614 * all code points that differ only in those bit positions. To
19615 * convince yourself of that, first note that the number in the
19616 * closure must be a power of 2, which we test for. The only
19617 * way we could have that count and it be some differing set,
19618 * is if we got some code points that don't differ from the
19619 * lowest code point in any position, but do differ from each
19620 * other in some other position. That means one code point has
19621 * a 1 in that position, and another has a 0. But that would
19622 * mean that one of them differs from the lowest code point in
19623 * that position, which possibility we've already excluded. */
19624 if ( (inverted || full_cp_count > 1)
19625 && full_cp_count == 1U << PL_bitcount[bits_differing])
19629 op = ANYOFM + inverted;;
19631 /* We need to make the bits that differ be 0's */
19632 ANYOFM_mask = ~ bits_differing; /* This goes into FLAGS */
19634 /* The argument is the lowest code point */
19635 ret = reganode(pRExC_state, op, lowest_cp);
19636 FLAGS(REGNODE_p(ret)) = ANYOFM_mask;
19640 invlist_iterfinish(cp_list);
19644 _invlist_invert(cp_list);
19651 /* XXX We could create an ANYOFR_LOW node here if we saved above if
19652 * all were invariants, it wasn't inverted, and there is a single
19653 * range. This would be faster than some of the posix nodes we
19654 * create below like /\d/a, but would be twice the size. Without
19655 * having actually measured the gain, khw doesn't think the
19656 * tradeoff is really worth it */
19659 if (! (anyof_flags & ANYOF_LOCALE_FLAGS)) {
19660 PERL_UINT_FAST8_T type;
19661 SV * intersection = NULL;
19662 SV* d_invlist = NULL;
19664 /* See if this matches any of the POSIX classes. The POSIXA and
19665 * POSIXD ones are about the same speed as ANYOF ops, but take less
19666 * room; the ones that have above-Latin1 code point matches are
19667 * somewhat faster than ANYOF. */
19669 for (type = POSIXA; type >= POSIXD; type--) {
19672 if (type == POSIXL) { /* But not /l posix classes */
19676 for (posix_class = 0;
19677 posix_class <= _HIGHEST_REGCOMP_DOT_H_SYNC;
19680 SV** our_code_points = &cp_list;
19681 SV** official_code_points;
19684 if (type == POSIXA) {
19685 official_code_points = &PL_Posix_ptrs[posix_class];
19688 official_code_points = &PL_XPosix_ptrs[posix_class];
19691 /* Skip non-existent classes of this type. e.g. \v only
19692 * has an entry in PL_XPosix_ptrs */
19693 if (! *official_code_points) {
19697 /* Try both the regular class, and its inversion */
19698 for (try_inverted = 0; try_inverted < 2; try_inverted++) {
19699 bool this_inverted = invert ^ try_inverted;
19701 if (type != POSIXD) {
19703 /* This class that isn't /d can't match if we have
19704 * /d dependencies */
19705 if (has_runtime_dependency
19706 & HAS_D_RUNTIME_DEPENDENCY)
19711 else /* is /d */ if (! this_inverted) {
19713 /* /d classes don't match anything non-ASCII below
19714 * 256 unconditionally (which cp_list contains) */
19715 _invlist_intersection(cp_list, PL_UpperLatin1,
19717 if (_invlist_len(intersection) != 0) {
19721 SvREFCNT_dec(d_invlist);
19722 d_invlist = invlist_clone(cp_list, NULL);
19724 /* But under UTF-8 it turns into using /u rules.
19725 * Add the things it matches under these conditions
19726 * so that we check below that these are identical
19727 * to what the tested class should match */
19728 if (upper_latin1_only_utf8_matches) {
19731 upper_latin1_only_utf8_matches,
19734 our_code_points = &d_invlist;
19736 else { /* POSIXD, inverted. If this doesn't have this
19737 flag set, it isn't /d. */
19738 if (! (anyof_flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
19742 our_code_points = &cp_list;
19745 /* Here, have weeded out some things. We want to see
19746 * if the list of characters this node contains
19747 * ('*our_code_points') precisely matches those of the
19748 * class we are currently checking against
19749 * ('*official_code_points'). */
19750 if (_invlistEQ(*our_code_points,
19751 *official_code_points,
19754 /* Here, they precisely match. Optimize this ANYOF
19755 * node into its equivalent POSIX one of the
19756 * correct type, possibly inverted */
19757 ret = reg_node(pRExC_state, (try_inverted)
19761 FLAGS(REGNODE_p(ret)) = posix_class;
19762 SvREFCNT_dec(d_invlist);
19763 SvREFCNT_dec(intersection);
19769 SvREFCNT_dec(d_invlist);
19770 SvREFCNT_dec(intersection);
19773 /* If it is a single contiguous range, ANYOFR is an efficient regnode,
19774 * both in size and speed. Currently, a 20 bit range base (smallest
19775 * code point in the range), and a 12 bit maximum delta are packed into
19776 * a 32 bit word. This allows for using it on all of the Unicode code
19777 * points except for the highest plane, which is only for private use
19778 * code points. khw doubts that a bigger delta is likely in real world
19781 && ! has_runtime_dependency
19782 && anyof_flags == 0
19783 && start[0] < (1 << ANYOFR_BASE_BITS)
19784 && end[0] - start[0]
19785 < ((1U << (sizeof(((struct regnode_1 *)NULL)->arg1)
19786 * CHARBITS - ANYOFR_BASE_BITS))))
19789 U8 low_utf8[UTF8_MAXBYTES+1];
19790 U8 high_utf8[UTF8_MAXBYTES+1];
19792 ret = reganode(pRExC_state, ANYOFR,
19793 (start[0] | (end[0] - start[0]) << ANYOFR_BASE_BITS));
19795 /* Place the lowest UTF-8 start byte in the flags field, so as to
19796 * allow efficient ruling out at run time of many possible inputs.
19798 (void) uvchr_to_utf8(low_utf8, start[0]);
19799 (void) uvchr_to_utf8(high_utf8, end[0]);
19801 /* If all code points share the same first byte, this can be an
19802 * ANYOFRb. Otherwise store the lowest UTF-8 start byte which can
19803 * quickly rule out many inputs at run-time without having to
19804 * compute the code point from UTF-8. For EBCDIC, we use I8, as
19805 * not doing that transformation would not rule out nearly so many
19807 if (low_utf8[0] == high_utf8[0]) {
19808 OP(REGNODE_p(ret)) = ANYOFRb;
19809 ANYOF_FLAGS(REGNODE_p(ret)) = low_utf8[0];
19812 ANYOF_FLAGS(REGNODE_p(ret))
19813 = NATIVE_UTF8_TO_I8(low_utf8[0]);
19819 /* If didn't find an optimization and there is no need for a bitmap,
19820 * optimize to indicate that */
19821 if ( start[0] >= NUM_ANYOF_CODE_POINTS
19823 && ! upper_latin1_only_utf8_matches
19824 && anyof_flags == 0)
19826 U8 low_utf8[UTF8_MAXBYTES+1];
19827 UV highest_cp = invlist_highest(cp_list);
19829 /* Currently the maximum allowed code point by the system is
19830 * IV_MAX. Higher ones are reserved for future internal use. This
19831 * particular regnode can be used for higher ones, but we can't
19832 * calculate the code point of those. IV_MAX suffices though, as
19833 * it will be a large first byte */
19834 Size_t low_len = uvchr_to_utf8(low_utf8, MIN(start[0], IV_MAX))
19837 /* We store the lowest possible first byte of the UTF-8
19838 * representation, using the flags field. This allows for quick
19839 * ruling out of some inputs without having to convert from UTF-8
19840 * to code point. For EBCDIC, we use I8, as not doing that
19841 * transformation would not rule out nearly so many things */
19842 anyof_flags = NATIVE_UTF8_TO_I8(low_utf8[0]);
19846 /* If the first UTF-8 start byte for the highest code point in the
19847 * range is suitably small, we may be able to get an upper bound as
19849 if (highest_cp <= IV_MAX) {
19850 U8 high_utf8[UTF8_MAXBYTES+1];
19851 Size_t high_len = uvchr_to_utf8(high_utf8, highest_cp)
19854 /* If the lowest and highest are the same, we can get an exact
19855 * first byte instead of a just minimum or even a sequence of
19856 * exact leading bytes. We signal these with different
19858 if (low_utf8[0] == high_utf8[0]) {
19859 Size_t len = find_first_differing_byte_pos(low_utf8,
19861 MIN(low_len, high_len));
19865 /* No need to convert to I8 for EBCDIC as this is an
19867 anyof_flags = low_utf8[0];
19872 ret = regnode_guts(pRExC_state, op,
19873 regarglen[op] + STR_SZ(len),
19875 FILL_NODE(ret, op);
19876 ((struct regnode_anyofhs *) REGNODE_p(ret))->str_len
19878 Copy(low_utf8, /* Add the common bytes */
19879 ((struct regnode_anyofhs *) REGNODE_p(ret))->string,
19881 RExC_emit += NODE_SZ_STR(REGNODE_p(ret));
19882 set_ANYOF_arg(pRExC_state, REGNODE_p(ret), cp_list,
19883 NULL, only_utf8_locale_list);
19887 else if (NATIVE_UTF8_TO_I8(high_utf8[0]) <= MAX_ANYOF_HRx_BYTE)
19890 /* Here, the high byte is not the same as the low, but is
19891 * small enough that its reasonable to have a loose upper
19892 * bound, which is packed in with the strict lower bound.
19893 * See comments at the definition of MAX_ANYOF_HRx_BYTE.
19894 * On EBCDIC platforms, I8 is used. On ASCII platforms I8
19895 * is the same thing as UTF-8 */
19898 U8 max_range_diff = MAX_ANYOF_HRx_BYTE - anyof_flags;
19899 U8 range_diff = NATIVE_UTF8_TO_I8(high_utf8[0])
19902 if (range_diff <= max_range_diff / 8) {
19905 else if (range_diff <= max_range_diff / 4) {
19908 else if (range_diff <= max_range_diff / 2) {
19911 anyof_flags = (anyof_flags - 0xC0) << 2 | bits;
19916 goto done_finding_op;
19918 } /* End of seeing if can optimize it into a different node */
19920 is_anyof: /* It's going to be an ANYOF node. */
19921 op = (has_runtime_dependency & HAS_D_RUNTIME_DEPENDENCY)
19931 ret = regnode_guts(pRExC_state, op, regarglen[op], "anyof");
19932 FILL_NODE(ret, op); /* We set the argument later */
19933 RExC_emit += 1 + regarglen[op];
19934 ANYOF_FLAGS(REGNODE_p(ret)) = anyof_flags;
19936 /* Here, <cp_list> contains all the code points we can determine at
19937 * compile time that match under all conditions. Go through it, and
19938 * for things that belong in the bitmap, put them there, and delete from
19939 * <cp_list>. While we are at it, see if everything above 255 is in the
19940 * list, and if so, set a flag to speed up execution */
19942 populate_ANYOF_from_invlist(REGNODE_p(ret), &cp_list);
19945 ANYOF_POSIXL_SET_TO_BITMAP(REGNODE_p(ret), posixl);
19949 ANYOF_FLAGS(REGNODE_p(ret)) |= ANYOF_INVERT;
19952 /* Here, the bitmap has been populated with all the Latin1 code points that
19953 * always match. Can now add to the overall list those that match only
19954 * when the target string is UTF-8 (<upper_latin1_only_utf8_matches>).
19956 if (upper_latin1_only_utf8_matches) {
19958 _invlist_union(cp_list,
19959 upper_latin1_only_utf8_matches,
19961 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
19964 cp_list = upper_latin1_only_utf8_matches;
19966 ANYOF_FLAGS(REGNODE_p(ret)) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
19969 set_ANYOF_arg(pRExC_state, REGNODE_p(ret), cp_list,
19970 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
19973 only_utf8_locale_list);
19974 SvREFCNT_dec(cp_list);;
19975 SvREFCNT_dec(only_utf8_locale_list);
19980 /* Here, the node is getting optimized into something that's not an ANYOF
19981 * one. Finish up. */
19983 Set_Node_Offset_Length(REGNODE_p(ret), orig_parse - RExC_start,
19984 RExC_parse - orig_parse);;
19985 SvREFCNT_dec(cp_list);;
19986 SvREFCNT_dec(only_utf8_locale_list);
19990 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
19993 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
19994 regnode* const node,
19996 SV* const runtime_defns,
19997 SV* const only_utf8_locale_list)
19999 /* Sets the arg field of an ANYOF-type node 'node', using information about
20000 * the node passed-in. If there is nothing outside the node's bitmap, the
20001 * arg is set to ANYOF_ONLY_HAS_BITMAP. Otherwise, it sets the argument to
20002 * the count returned by add_data(), having allocated and stored an array,
20005 * av[0] stores the inversion list defining this class as far as known at
20006 * this time, or PL_sv_undef if nothing definite is now known.
20007 * av[1] stores the inversion list of code points that match only if the
20008 * current locale is UTF-8, or if none, PL_sv_undef if there is an
20009 * av[2], or no entry otherwise.
20010 * av[2] stores the list of user-defined properties whose subroutine
20011 * definitions aren't known at this time, or no entry if none. */
20015 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
20017 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
20018 assert(! (ANYOF_FLAGS(node)
20019 & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP));
20020 ARG_SET(node, ANYOF_ONLY_HAS_BITMAP);
20023 AV * const av = newAV();
20027 av_store(av, INVLIST_INDEX, SvREFCNT_inc_NN(cp_list));
20030 /* (Note that if any of this changes, the size calculations in
20031 * S_optimize_regclass() might need to be updated.) */
20033 if (only_utf8_locale_list) {
20034 av_store(av, ONLY_LOCALE_MATCHES_INDEX,
20035 SvREFCNT_inc_NN(only_utf8_locale_list));
20038 if (runtime_defns) {
20039 av_store(av, DEFERRED_USER_DEFINED_INDEX,
20040 SvREFCNT_inc_NN(runtime_defns));
20043 rv = newRV_noinc(MUTABLE_SV(av));
20044 n = add_data(pRExC_state, STR_WITH_LEN("s"));
20045 RExC_rxi->data->data[n] = (void*)rv;
20052 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
20053 Perl_get_regclass_nonbitmap_data(pTHX_ const regexp *prog, const regnode* node, bool doinit, SV** listsvp, SV** only_utf8_locale_ptr, SV** output_invlist)
20055 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)
20059 /* For internal core use only.
20060 * Returns the inversion list for the input 'node' in the regex 'prog'.
20061 * If <doinit> is 'true', will attempt to create the inversion list if not
20063 * If <listsvp> is non-null, will return the printable contents of the
20064 * property definition. This can be used to get debugging information
20065 * even before the inversion list exists, by calling this function with
20066 * 'doinit' set to false, in which case the components that will be used
20067 * to eventually create the inversion list are returned (in a printable
20069 * If <only_utf8_locale_ptr> is not NULL, it is where this routine is to
20070 * store an inversion list of code points that should match only if the
20071 * execution-time locale is a UTF-8 one.
20072 * If <output_invlist> is not NULL, it is where this routine is to store an
20073 * inversion list of the code points that would be instead returned in
20074 * <listsvp> if this were NULL. Thus, what gets output in <listsvp>
20075 * when this parameter is used, is just the non-code point data that
20076 * will go into creating the inversion list. This currently should be just
20077 * user-defined properties whose definitions were not known at compile
20078 * time. Using this parameter allows for easier manipulation of the
20079 * inversion list's data by the caller. It is illegal to call this
20080 * function with this parameter set, but not <listsvp>
20082 * Tied intimately to how S_set_ANYOF_arg sets up the data structure. Note
20083 * that, in spite of this function's name, the inversion list it returns
20084 * may include the bitmap data as well */
20086 SV *si = NULL; /* Input initialization string */
20087 SV* invlist = NULL;
20089 RXi_GET_DECL(prog, progi);
20090 const struct reg_data * const data = prog ? progi->data : NULL;
20092 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
20093 PERL_ARGS_ASSERT_GET_REGCLASS_NONBITMAP_DATA;
20095 PERL_ARGS_ASSERT_GET_RE_GCLASS_NONBITMAP_DATA;
20097 assert(! output_invlist || listsvp);
20099 if (data && data->count) {
20100 const U32 n = ARG(node);
20102 if (data->what[n] == 's') {
20103 SV * const rv = MUTABLE_SV(data->data[n]);
20104 AV * const av = MUTABLE_AV(SvRV(rv));
20105 SV **const ary = AvARRAY(av);
20107 invlist = ary[INVLIST_INDEX];
20109 if (av_tindex_skip_len_mg(av) >= ONLY_LOCALE_MATCHES_INDEX) {
20110 *only_utf8_locale_ptr = ary[ONLY_LOCALE_MATCHES_INDEX];
20113 if (av_tindex_skip_len_mg(av) >= DEFERRED_USER_DEFINED_INDEX) {
20114 si = ary[DEFERRED_USER_DEFINED_INDEX];
20117 if (doinit && (si || invlist)) {
20120 SV * msg = newSVpvs_flags("", SVs_TEMP);
20122 SV * prop_definition = handle_user_defined_property(
20123 "", 0, FALSE, /* There is no \p{}, \P{} */
20124 SvPVX_const(si)[1] - '0', /* /i or not has been
20125 stored here for just
20127 TRUE, /* run time */
20128 FALSE, /* This call must find the defn */
20129 si, /* The property definition */
20132 0 /* base level call */
20136 assert(prop_definition == NULL);
20138 Perl_croak(aTHX_ "%" UTF8f,
20139 UTF8fARG(SvUTF8(msg), SvCUR(msg), SvPVX(msg)));
20143 _invlist_union(invlist, prop_definition, &invlist);
20144 SvREFCNT_dec_NN(prop_definition);
20147 invlist = prop_definition;
20150 STATIC_ASSERT_STMT(ONLY_LOCALE_MATCHES_INDEX == 1 + INVLIST_INDEX);
20151 STATIC_ASSERT_STMT(DEFERRED_USER_DEFINED_INDEX == 1 + ONLY_LOCALE_MATCHES_INDEX);
20153 ary[INVLIST_INDEX] = invlist;
20154 av_fill(av, (ary[ONLY_LOCALE_MATCHES_INDEX])
20155 ? ONLY_LOCALE_MATCHES_INDEX
20163 /* If requested, return a printable version of what this ANYOF node matches
20166 SV* matches_string = NULL;
20168 /* This function can be called at compile-time, before everything gets
20169 * resolved, in which case we return the currently best available
20170 * information, which is the string that will eventually be used to do
20171 * that resolving, 'si' */
20173 /* Here, we only have 'si' (and possibly some passed-in data in
20174 * 'invlist', which is handled below) If the caller only wants
20175 * 'si', use that. */
20176 if (! output_invlist) {
20177 matches_string = newSVsv(si);
20180 /* But if the caller wants an inversion list of the node, we
20181 * need to parse 'si' and place as much as possible in the
20182 * desired output inversion list, making 'matches_string' only
20183 * contain the currently unresolvable things */
20184 const char *si_string = SvPVX(si);
20185 STRLEN remaining = SvCUR(si);
20189 /* Ignore everything before and including the first new-line */
20190 si_string = (const char *) memchr(si_string, '\n', SvCUR(si));
20191 assert (si_string != NULL);
20193 remaining = SvPVX(si) + SvCUR(si) - si_string;
20195 while (remaining > 0) {
20197 /* The data consists of just strings defining user-defined
20198 * property names, but in prior incarnations, and perhaps
20199 * somehow from pluggable regex engines, it could still
20200 * hold hex code point definitions, all of which should be
20201 * legal (or it wouldn't have gotten this far). Each
20202 * component of a range would be separated by a tab, and
20203 * each range by a new-line. If these are found, instead
20204 * add them to the inversion list */
20205 I32 grok_flags = PERL_SCAN_SILENT_ILLDIGIT
20206 |PERL_SCAN_SILENT_NON_PORTABLE;
20207 STRLEN len = remaining;
20208 UV cp = grok_hex(si_string, &len, &grok_flags, NULL);
20210 /* If the hex decode routine found something, it should go
20211 * up to the next \n */
20212 if ( *(si_string + len) == '\n') {
20213 if (count) { /* 2nd code point on line */
20214 *output_invlist = _add_range_to_invlist(*output_invlist, prev_cp, cp);
20217 *output_invlist = add_cp_to_invlist(*output_invlist, cp);
20220 goto prepare_for_next_iteration;
20223 /* If the hex decode was instead for the lower range limit,
20224 * save it, and go parse the upper range limit */
20225 if (*(si_string + len) == '\t') {
20226 assert(count == 0);
20230 prepare_for_next_iteration:
20231 si_string += len + 1;
20232 remaining -= len + 1;
20236 /* Here, didn't find a legal hex number. Just add the text
20237 * from here up to the next \n, omitting any trailing
20241 len = strcspn(si_string,
20242 DEFERRED_COULD_BE_OFFICIAL_MARKERs "\n");
20244 if (matches_string) {
20245 sv_catpvn(matches_string, si_string, len);
20248 matches_string = newSVpvn(si_string, len);
20250 sv_catpvs(matches_string, " ");
20254 && UCHARAT(si_string)
20255 == DEFERRED_COULD_BE_OFFICIAL_MARKERc)
20260 if (remaining && UCHARAT(si_string) == '\n') {
20264 } /* end of loop through the text */
20266 assert(matches_string);
20267 if (SvCUR(matches_string)) { /* Get rid of trailing blank */
20268 SvCUR_set(matches_string, SvCUR(matches_string) - 1);
20270 } /* end of has an 'si' */
20273 /* Add the stuff that's already known */
20276 /* Again, if the caller doesn't want the output inversion list, put
20277 * everything in 'matches-string' */
20278 if (! output_invlist) {
20279 if ( ! matches_string) {
20280 matches_string = newSVpvs("\n");
20282 sv_catsv(matches_string, invlist_contents(invlist,
20283 TRUE /* traditional style */
20286 else if (! *output_invlist) {
20287 *output_invlist = invlist_clone(invlist, NULL);
20290 _invlist_union(*output_invlist, invlist, output_invlist);
20294 *listsvp = matches_string;
20300 /* reg_skipcomment()
20302 Absorbs an /x style # comment from the input stream,
20303 returning a pointer to the first character beyond the comment, or if the
20304 comment terminates the pattern without anything following it, this returns
20305 one past the final character of the pattern (in other words, RExC_end) and
20306 sets the REG_RUN_ON_COMMENT_SEEN flag.
20308 Note it's the callers responsibility to ensure that we are
20309 actually in /x mode
20313 PERL_STATIC_INLINE char*
20314 S_reg_skipcomment(RExC_state_t *pRExC_state, char* p)
20316 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
20320 while (p < RExC_end) {
20321 if (*(++p) == '\n') {
20326 /* we ran off the end of the pattern without ending the comment, so we have
20327 * to add an \n when wrapping */
20328 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
20333 S_skip_to_be_ignored_text(pTHX_ RExC_state_t *pRExC_state,
20335 const bool force_to_xmod
20338 /* If the text at the current parse position '*p' is a '(?#...)' comment,
20339 * or if we are under /x or 'force_to_xmod' is TRUE, and the text at '*p'
20340 * is /x whitespace, advance '*p' so that on exit it points to the first
20341 * byte past all such white space and comments */
20343 const bool use_xmod = force_to_xmod || (RExC_flags & RXf_PMf_EXTENDED);
20345 PERL_ARGS_ASSERT_SKIP_TO_BE_IGNORED_TEXT;
20347 assert( ! UTF || UTF8_IS_INVARIANT(**p) || UTF8_IS_START(**p));
20350 if (RExC_end - (*p) >= 3
20352 && *(*p + 1) == '?'
20353 && *(*p + 2) == '#')
20355 while (*(*p) != ')') {
20356 if ((*p) == RExC_end)
20357 FAIL("Sequence (?#... not terminated");
20365 const char * save_p = *p;
20366 while ((*p) < RExC_end) {
20368 if ((len = is_PATWS_safe((*p), RExC_end, UTF))) {
20371 else if (*(*p) == '#') {
20372 (*p) = reg_skipcomment(pRExC_state, (*p));
20378 if (*p != save_p) {
20391 Advances the parse position by one byte, unless that byte is the beginning
20392 of a '(?#...)' style comment, or is /x whitespace and /x is in effect. In
20393 those two cases, the parse position is advanced beyond all such comments and
20396 This is the UTF, (?#...), and /x friendly way of saying RExC_parse++.
20400 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
20402 PERL_ARGS_ASSERT_NEXTCHAR;
20404 if (RExC_parse < RExC_end) {
20406 || UTF8_IS_INVARIANT(*RExC_parse)
20407 || UTF8_IS_START(*RExC_parse));
20409 RExC_parse += (UTF)
20410 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
20413 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
20414 FALSE /* Don't force /x */ );
20419 S_change_engine_size(pTHX_ RExC_state_t *pRExC_state, const Ptrdiff_t size)
20421 /* 'size' is the delta number of smallest regnode equivalents to add or
20422 * subtract from the current memory allocated to the regex engine being
20425 PERL_ARGS_ASSERT_CHANGE_ENGINE_SIZE;
20430 sizeof(regexp_internal) + (RExC_size + 1) * sizeof(regnode),
20431 /* +1 for REG_MAGIC */
20434 if ( RExC_rxi == NULL )
20435 FAIL("Regexp out of space");
20436 RXi_SET(RExC_rx, RExC_rxi);
20438 RExC_emit_start = RExC_rxi->program;
20440 Zero(REGNODE_p(RExC_emit), size, regnode);
20443 #ifdef RE_TRACK_PATTERN_OFFSETS
20444 Renew(RExC_offsets, 2*RExC_size+1, U32);
20446 Zero(RExC_offsets + 2*(RExC_size - size) + 1, 2 * size, U32);
20448 RExC_offsets[0] = RExC_size;
20452 STATIC regnode_offset
20453 S_regnode_guts(pTHX_ RExC_state_t *pRExC_state, const U8 op, const STRLEN extra_size, const char* const name)
20455 /* Allocate a regnode for 'op', with 'extra_size' extra (smallest) regnode
20456 * equivalents space. It aligns and increments RExC_size
20458 * It returns the regnode's offset into the regex engine program */
20460 const regnode_offset ret = RExC_emit;
20462 DECLARE_AND_GET_RE_DEBUG_FLAGS;
20464 PERL_ARGS_ASSERT_REGNODE_GUTS;
20466 SIZE_ALIGN(RExC_size);
20467 change_engine_size(pRExC_state, (Ptrdiff_t) 1 + extra_size);
20468 NODE_ALIGN_FILL(REGNODE_p(ret));
20469 #ifndef RE_TRACK_PATTERN_OFFSETS
20470 PERL_UNUSED_ARG(name);
20471 PERL_UNUSED_ARG(op);
20473 assert(extra_size >= regarglen[op] || PL_regkind[op] == ANYOF);
20475 if (RExC_offsets) { /* MJD */
20477 ("%s:%d: (op %s) %s %" UVuf " (len %" UVuf ") (max %" UVuf ").\n",
20480 (UV)(RExC_emit) > RExC_offsets[0]
20481 ? "Overwriting end of array!\n" : "OK",
20483 (UV)(RExC_parse - RExC_start),
20484 (UV)RExC_offsets[0]));
20485 Set_Node_Offset(REGNODE_p(RExC_emit), RExC_parse + (op == END));
20492 - reg_node - emit a node
20494 STATIC regnode_offset /* Location. */
20495 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
20497 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reg_node");
20498 regnode_offset ptr = ret;
20500 PERL_ARGS_ASSERT_REG_NODE;
20502 assert(regarglen[op] == 0);
20504 FILL_ADVANCE_NODE(ptr, op);
20510 - reganode - emit a node with an argument
20512 STATIC regnode_offset /* Location. */
20513 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
20515 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reganode");
20516 regnode_offset ptr = ret;
20518 PERL_ARGS_ASSERT_REGANODE;
20520 /* ANYOF are special cased to allow non-length 1 args */
20521 assert(regarglen[op] == 1);
20523 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
20529 - regpnode - emit a temporary node with a SV* argument
20531 STATIC regnode_offset /* Location. */
20532 S_regpnode(pTHX_ RExC_state_t *pRExC_state, U8 op, SV * arg)
20534 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "regpnode");
20535 regnode_offset ptr = ret;
20537 PERL_ARGS_ASSERT_REGPNODE;
20539 FILL_ADVANCE_NODE_ARGp(ptr, op, arg);
20544 STATIC regnode_offset
20545 S_reg2Lanode(pTHX_ RExC_state_t *pRExC_state, const U8 op, const U32 arg1, const I32 arg2)
20547 /* emit a node with U32 and I32 arguments */
20549 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reg2Lanode");
20550 regnode_offset ptr = ret;
20552 PERL_ARGS_ASSERT_REG2LANODE;
20554 assert(regarglen[op] == 2);
20556 FILL_ADVANCE_NODE_2L_ARG(ptr, op, arg1, arg2);
20562 - reginsert - insert an operator in front of already-emitted operand
20564 * That means that on exit 'operand' is the offset of the newly inserted
20565 * operator, and the original operand has been relocated.
20567 * IMPORTANT NOTE - it is the *callers* responsibility to correctly
20568 * set up NEXT_OFF() of the inserted node if needed. Something like this:
20570 * reginsert(pRExC, OPFAIL, orig_emit, depth+1);
20571 * NEXT_OFF(orig_emit) = regarglen[OPFAIL] + NODE_STEP_REGNODE;
20573 * ALSO NOTE - FLAGS(newly-inserted-operator) will be set to 0 as well.
20576 S_reginsert(pTHX_ RExC_state_t *pRExC_state, const U8 op,
20577 const regnode_offset operand, const U32 depth)
20582 const int offset = regarglen[(U8)op];
20583 const int size = NODE_STEP_REGNODE + offset;
20584 DECLARE_AND_GET_RE_DEBUG_FLAGS;
20586 PERL_ARGS_ASSERT_REGINSERT;
20587 PERL_UNUSED_CONTEXT;
20588 PERL_UNUSED_ARG(depth);
20589 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
20590 DEBUG_PARSE_FMT("inst"," - %s", PL_reg_name[op]);
20591 assert(!RExC_study_started); /* I believe we should never use reginsert once we have started
20592 studying. If this is wrong then we need to adjust RExC_recurse
20593 below like we do with RExC_open_parens/RExC_close_parens. */
20594 change_engine_size(pRExC_state, (Ptrdiff_t) size);
20595 src = REGNODE_p(RExC_emit);
20597 dst = REGNODE_p(RExC_emit);
20599 /* If we are in a "count the parentheses" pass, the numbers are unreliable,
20600 * and [perl #133871] shows this can lead to problems, so skip this
20601 * realignment of parens until a later pass when they are reliable */
20602 if (! IN_PARENS_PASS && RExC_open_parens) {
20604 /*DEBUG_PARSE_FMT("inst"," - %" IVdf, (IV)RExC_npar);*/
20605 /* remember that RExC_npar is rex->nparens + 1,
20606 * iow it is 1 more than the number of parens seen in
20607 * the pattern so far. */
20608 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
20609 /* note, RExC_open_parens[0] is the start of the
20610 * regex, it can't move. RExC_close_parens[0] is the end
20611 * of the regex, it *can* move. */
20612 if ( paren && RExC_open_parens[paren] >= operand ) {
20613 /*DEBUG_PARSE_FMT("open"," - %d", size);*/
20614 RExC_open_parens[paren] += size;
20616 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
20618 if ( RExC_close_parens[paren] >= operand ) {
20619 /*DEBUG_PARSE_FMT("close"," - %d", size);*/
20620 RExC_close_parens[paren] += size;
20622 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
20627 RExC_end_op += size;
20629 while (src > REGNODE_p(operand)) {
20630 StructCopy(--src, --dst, regnode);
20631 #ifdef RE_TRACK_PATTERN_OFFSETS
20632 if (RExC_offsets) { /* MJD 20010112 */
20634 ("%s(%d): (op %s) %s copy %" UVuf " -> %" UVuf " (max %" UVuf ").\n",
20638 (UV)(REGNODE_OFFSET(dst)) > RExC_offsets[0]
20639 ? "Overwriting end of array!\n" : "OK",
20640 (UV)REGNODE_OFFSET(src),
20641 (UV)REGNODE_OFFSET(dst),
20642 (UV)RExC_offsets[0]));
20643 Set_Node_Offset_To_R(REGNODE_OFFSET(dst), Node_Offset(src));
20644 Set_Node_Length_To_R(REGNODE_OFFSET(dst), Node_Length(src));
20649 place = REGNODE_p(operand); /* Op node, where operand used to be. */
20650 #ifdef RE_TRACK_PATTERN_OFFSETS
20651 if (RExC_offsets) { /* MJD */
20653 ("%s(%d): (op %s) %s %" UVuf " <- %" UVuf " (max %" UVuf ").\n",
20657 (UV)REGNODE_OFFSET(place) > RExC_offsets[0]
20658 ? "Overwriting end of array!\n" : "OK",
20659 (UV)REGNODE_OFFSET(place),
20660 (UV)(RExC_parse - RExC_start),
20661 (UV)RExC_offsets[0]));
20662 Set_Node_Offset(place, RExC_parse);
20663 Set_Node_Length(place, 1);
20666 src = NEXTOPER(place);
20668 FILL_NODE(operand, op);
20670 /* Zero out any arguments in the new node */
20671 Zero(src, offset, regnode);
20675 - regtail - set the next-pointer at the end of a node chain of p to val. If
20676 that value won't fit in the space available, instead returns FALSE.
20677 (Except asserts if we can't fit in the largest space the regex
20678 engine is designed for.)
20679 - SEE ALSO: regtail_study
20682 S_regtail(pTHX_ RExC_state_t * pRExC_state,
20683 const regnode_offset p,
20684 const regnode_offset val,
20687 regnode_offset scan;
20688 DECLARE_AND_GET_RE_DEBUG_FLAGS;
20690 PERL_ARGS_ASSERT_REGTAIL;
20692 PERL_UNUSED_ARG(depth);
20695 /* The final node in the chain is the first one with a nonzero next pointer
20697 scan = (regnode_offset) p;
20699 regnode * const temp = regnext(REGNODE_p(scan));
20701 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
20702 regprop(RExC_rx, RExC_mysv, REGNODE_p(scan), NULL, pRExC_state);
20703 Perl_re_printf( aTHX_ "~ %s (%zu) %s %s\n",
20704 SvPV_nolen_const(RExC_mysv), scan,
20705 (temp == NULL ? "->" : ""),
20706 (temp == NULL ? PL_reg_name[OP(REGNODE_p(val))] : "")
20711 scan = REGNODE_OFFSET(temp);
20714 /* Populate this node's next pointer */
20715 assert(val >= scan);
20716 if (reg_off_by_arg[OP(REGNODE_p(scan))]) {
20717 assert((UV) (val - scan) <= U32_MAX);
20718 ARG_SET(REGNODE_p(scan), val - scan);
20721 if (val - scan > U16_MAX) {
20722 /* Populate this with something that won't loop and will likely
20723 * lead to a crash if the caller ignores the failure return, and
20724 * execution continues */
20725 NEXT_OFF(REGNODE_p(scan)) = U16_MAX;
20728 NEXT_OFF(REGNODE_p(scan)) = val - scan;
20736 - regtail_study - set the next-pointer at the end of a node chain of p to val.
20737 - Look for optimizable sequences at the same time.
20738 - currently only looks for EXACT chains.
20740 This is experimental code. The idea is to use this routine to perform
20741 in place optimizations on branches and groups as they are constructed,
20742 with the long term intention of removing optimization from study_chunk so
20743 that it is purely analytical.
20745 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
20746 to control which is which.
20748 This used to return a value that was ignored. It was a problem that it is
20749 #ifdef'd to be another function that didn't return a value. khw has changed it
20750 so both currently return a pass/fail return.
20753 /* TODO: All four parms should be const */
20756 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode_offset p,
20757 const regnode_offset val, U32 depth)
20759 regnode_offset scan;
20761 #ifdef EXPERIMENTAL_INPLACESCAN
20764 DECLARE_AND_GET_RE_DEBUG_FLAGS;
20766 PERL_ARGS_ASSERT_REGTAIL_STUDY;
20769 /* Find last node. */
20773 regnode * const temp = regnext(REGNODE_p(scan));
20774 #ifdef EXPERIMENTAL_INPLACESCAN
20775 if (PL_regkind[OP(REGNODE_p(scan))] == EXACT) {
20776 bool unfolded_multi_char; /* Unexamined in this routine */
20777 if (join_exact(pRExC_state, scan, &min,
20778 &unfolded_multi_char, 1, REGNODE_p(val), depth+1))
20779 return TRUE; /* Was return EXACT */
20783 if (PL_regkind[OP(REGNODE_p(scan))] == EXACT) {
20784 if (exact == PSEUDO )
20785 exact= OP(REGNODE_p(scan));
20786 else if (exact != OP(REGNODE_p(scan)) )
20789 else if (OP(REGNODE_p(scan)) != NOTHING) {
20794 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
20795 regprop(RExC_rx, RExC_mysv, REGNODE_p(scan), NULL, pRExC_state);
20796 Perl_re_printf( aTHX_ "~ %s (%zu) -> %s\n",
20797 SvPV_nolen_const(RExC_mysv),
20799 PL_reg_name[exact]);
20803 scan = REGNODE_OFFSET(temp);
20806 DEBUG_PARSE_MSG("");
20807 regprop(RExC_rx, RExC_mysv, REGNODE_p(val), NULL, pRExC_state);
20808 Perl_re_printf( aTHX_
20809 "~ attach to %s (%" IVdf ") offset to %" IVdf "\n",
20810 SvPV_nolen_const(RExC_mysv),
20815 if (reg_off_by_arg[OP(REGNODE_p(scan))]) {
20816 assert((UV) (val - scan) <= U32_MAX);
20817 ARG_SET(REGNODE_p(scan), val - scan);
20820 if (val - scan > U16_MAX) {
20821 /* Populate this with something that won't loop and will likely
20822 * lead to a crash if the caller ignores the failure return, and
20823 * execution continues */
20824 NEXT_OFF(REGNODE_p(scan)) = U16_MAX;
20827 NEXT_OFF(REGNODE_p(scan)) = val - scan;
20830 return TRUE; /* Was 'return exact' */
20835 S_get_ANYOFM_contents(pTHX_ const regnode * n) {
20837 /* Returns an inversion list of all the code points matched by the
20838 * ANYOFM/NANYOFM node 'n' */
20840 SV * cp_list = _new_invlist(-1);
20841 const U8 lowest = (U8) ARG(n);
20844 U8 needed = 1U << PL_bitcount[ (U8) ~ FLAGS(n)];
20846 PERL_ARGS_ASSERT_GET_ANYOFM_CONTENTS;
20848 /* Starting with the lowest code point, any code point that ANDed with the
20849 * mask yields the lowest code point is in the set */
20850 for (i = lowest; i <= 0xFF; i++) {
20851 if ((i & FLAGS(n)) == ARG(n)) {
20852 cp_list = add_cp_to_invlist(cp_list, i);
20855 /* We know how many code points (a power of two) that are in the
20856 * set. No use looking once we've got that number */
20857 if (count >= needed) break;
20861 if (OP(n) == NANYOFM) {
20862 _invlist_invert(cp_list);
20868 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
20873 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
20878 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
20880 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
20881 if (flags & (1<<bit)) {
20882 if (!set++ && lead)
20883 Perl_re_printf( aTHX_ "%s", lead);
20884 Perl_re_printf( aTHX_ "%s ", PL_reg_intflags_name[bit]);
20889 Perl_re_printf( aTHX_ "\n");
20891 Perl_re_printf( aTHX_ "%s[none-set]\n", lead);
20896 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
20902 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
20904 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
20905 if (flags & (1<<bit)) {
20906 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
20909 if (!set++ && lead)
20910 Perl_re_printf( aTHX_ "%s", lead);
20911 Perl_re_printf( aTHX_ "%s ", PL_reg_extflags_name[bit]);
20914 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
20915 if (!set++ && lead) {
20916 Perl_re_printf( aTHX_ "%s", lead);
20919 case REGEX_UNICODE_CHARSET:
20920 Perl_re_printf( aTHX_ "UNICODE");
20922 case REGEX_LOCALE_CHARSET:
20923 Perl_re_printf( aTHX_ "LOCALE");
20925 case REGEX_ASCII_RESTRICTED_CHARSET:
20926 Perl_re_printf( aTHX_ "ASCII-RESTRICTED");
20928 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
20929 Perl_re_printf( aTHX_ "ASCII-MORE_RESTRICTED");
20932 Perl_re_printf( aTHX_ "UNKNOWN CHARACTER SET");
20938 Perl_re_printf( aTHX_ "\n");
20940 Perl_re_printf( aTHX_ "%s[none-set]\n", lead);
20946 Perl_regdump(pTHX_ const regexp *r)
20950 SV * const sv = sv_newmortal();
20951 SV *dsv= sv_newmortal();
20952 RXi_GET_DECL(r, ri);
20953 DECLARE_AND_GET_RE_DEBUG_FLAGS;
20955 PERL_ARGS_ASSERT_REGDUMP;
20957 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
20959 /* Header fields of interest. */
20960 for (i = 0; i < 2; i++) {
20961 if (r->substrs->data[i].substr) {
20962 RE_PV_QUOTED_DECL(s, 0, dsv,
20963 SvPVX_const(r->substrs->data[i].substr),
20964 RE_SV_DUMPLEN(r->substrs->data[i].substr),
20965 PL_dump_re_max_len);
20966 Perl_re_printf( aTHX_
20967 "%s %s%s at %" IVdf "..%" UVuf " ",
20968 i ? "floating" : "anchored",
20970 RE_SV_TAIL(r->substrs->data[i].substr),
20971 (IV)r->substrs->data[i].min_offset,
20972 (UV)r->substrs->data[i].max_offset);
20974 else if (r->substrs->data[i].utf8_substr) {
20975 RE_PV_QUOTED_DECL(s, 1, dsv,
20976 SvPVX_const(r->substrs->data[i].utf8_substr),
20977 RE_SV_DUMPLEN(r->substrs->data[i].utf8_substr),
20979 Perl_re_printf( aTHX_
20980 "%s utf8 %s%s at %" IVdf "..%" UVuf " ",
20981 i ? "floating" : "anchored",
20983 RE_SV_TAIL(r->substrs->data[i].utf8_substr),
20984 (IV)r->substrs->data[i].min_offset,
20985 (UV)r->substrs->data[i].max_offset);
20989 if (r->check_substr || r->check_utf8)
20990 Perl_re_printf( aTHX_
20992 ( r->check_substr == r->substrs->data[1].substr
20993 && r->check_utf8 == r->substrs->data[1].utf8_substr
20994 ? "(checking floating" : "(checking anchored"));
20995 if (r->intflags & PREGf_NOSCAN)
20996 Perl_re_printf( aTHX_ " noscan");
20997 if (r->extflags & RXf_CHECK_ALL)
20998 Perl_re_printf( aTHX_ " isall");
20999 if (r->check_substr || r->check_utf8)
21000 Perl_re_printf( aTHX_ ") ");
21002 if (ri->regstclass) {
21003 regprop(r, sv, ri->regstclass, NULL, NULL);
21004 Perl_re_printf( aTHX_ "stclass %s ", SvPVX_const(sv));
21006 if (r->intflags & PREGf_ANCH) {
21007 Perl_re_printf( aTHX_ "anchored");
21008 if (r->intflags & PREGf_ANCH_MBOL)
21009 Perl_re_printf( aTHX_ "(MBOL)");
21010 if (r->intflags & PREGf_ANCH_SBOL)
21011 Perl_re_printf( aTHX_ "(SBOL)");
21012 if (r->intflags & PREGf_ANCH_GPOS)
21013 Perl_re_printf( aTHX_ "(GPOS)");
21014 Perl_re_printf( aTHX_ " ");
21016 if (r->intflags & PREGf_GPOS_SEEN)
21017 Perl_re_printf( aTHX_ "GPOS:%" UVuf " ", (UV)r->gofs);
21018 if (r->intflags & PREGf_SKIP)
21019 Perl_re_printf( aTHX_ "plus ");
21020 if (r->intflags & PREGf_IMPLICIT)
21021 Perl_re_printf( aTHX_ "implicit ");
21022 Perl_re_printf( aTHX_ "minlen %" IVdf " ", (IV)r->minlen);
21023 if (r->extflags & RXf_EVAL_SEEN)
21024 Perl_re_printf( aTHX_ "with eval ");
21025 Perl_re_printf( aTHX_ "\n");
21027 regdump_extflags("r->extflags: ", r->extflags);
21028 regdump_intflags("r->intflags: ", r->intflags);
21031 PERL_ARGS_ASSERT_REGDUMP;
21032 PERL_UNUSED_CONTEXT;
21033 PERL_UNUSED_ARG(r);
21034 #endif /* DEBUGGING */
21037 /* Should be synchronized with ANYOF_ #defines in regcomp.h */
21040 # if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 \
21041 || _CC_LOWER != 3 || _CC_UPPER != 4 || _CC_PUNCT != 5 \
21042 || _CC_PRINT != 6 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 \
21043 || _CC_CASED != 9 || _CC_SPACE != 10 || _CC_BLANK != 11 \
21044 || _CC_XDIGIT != 12 || _CC_CNTRL != 13 || _CC_ASCII != 14 \
21045 || _CC_VERTSPACE != 15
21046 # error Need to adjust order of anyofs[]
21048 static const char * const anyofs[] = {
21085 - regprop - printable representation of opcode, with run time support
21089 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo, const RExC_state_t *pRExC_state)
21093 RXi_GET_DECL(prog, progi);
21094 DECLARE_AND_GET_RE_DEBUG_FLAGS;
21096 PERL_ARGS_ASSERT_REGPROP;
21100 if (OP(o) > REGNODE_MAX) { /* regnode.type is unsigned */
21101 if (pRExC_state) { /* This gives more info, if we have it */
21102 FAIL3("panic: corrupted regexp opcode %d > %d",
21103 (int)OP(o), (int)REGNODE_MAX);
21106 Perl_croak(aTHX_ "panic: corrupted regexp opcode %d > %d",
21107 (int)OP(o), (int)REGNODE_MAX);
21110 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
21112 k = PL_regkind[OP(o)];
21115 sv_catpvs(sv, " ");
21116 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
21117 * is a crude hack but it may be the best for now since
21118 * we have no flag "this EXACTish node was UTF-8"
21120 pv_pretty(sv, STRING(o), STR_LEN(o), PL_dump_re_max_len,
21121 PL_colors[0], PL_colors[1],
21122 PERL_PV_ESCAPE_UNI_DETECT |
21123 PERL_PV_ESCAPE_NONASCII |
21124 PERL_PV_PRETTY_ELLIPSES |
21125 PERL_PV_PRETTY_LTGT |
21126 PERL_PV_PRETTY_NOCLEAR
21128 } else if (k == TRIE) {
21129 /* print the details of the trie in dumpuntil instead, as
21130 * progi->data isn't available here */
21131 const char op = OP(o);
21132 const U32 n = ARG(o);
21133 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
21134 (reg_ac_data *)progi->data->data[n] :
21136 const reg_trie_data * const trie
21137 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
21139 Perl_sv_catpvf(aTHX_ sv, "-%s", PL_reg_name[o->flags]);
21140 DEBUG_TRIE_COMPILE_r({
21142 sv_catpvs(sv, "(JUMP)");
21143 Perl_sv_catpvf(aTHX_ sv,
21144 "<S:%" UVuf "/%" IVdf " W:%" UVuf " L:%" UVuf "/%" UVuf " C:%" UVuf "/%" UVuf ">",
21145 (UV)trie->startstate,
21146 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
21147 (UV)trie->wordcount,
21150 (UV)TRIE_CHARCOUNT(trie),
21151 (UV)trie->uniquecharcount
21154 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
21155 sv_catpvs(sv, "[");
21156 (void) put_charclass_bitmap_innards(sv,
21157 ((IS_ANYOF_TRIE(op))
21159 : TRIE_BITMAP(trie)),
21166 sv_catpvs(sv, "]");
21168 } else if (k == CURLY) {
21169 U32 lo = ARG1(o), hi = ARG2(o);
21170 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
21171 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
21172 Perl_sv_catpvf(aTHX_ sv, "{%u,", (unsigned) lo);
21173 if (hi == REG_INFTY)
21174 sv_catpvs(sv, "INFTY");
21176 Perl_sv_catpvf(aTHX_ sv, "%u", (unsigned) hi);
21177 sv_catpvs(sv, "}");
21179 else if (k == WHILEM && o->flags) /* Ordinal/of */
21180 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
21181 else if (k == REF || k == OPEN || k == CLOSE
21182 || k == GROUPP || OP(o)==ACCEPT)
21184 AV *name_list= NULL;
21185 U32 parno= OP(o) == ACCEPT ? (U32)ARG2L(o) : ARG(o);
21186 Perl_sv_catpvf(aTHX_ sv, "%" UVuf, (UV)parno); /* Parenth number */
21187 if ( RXp_PAREN_NAMES(prog) ) {
21188 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
21189 } else if ( pRExC_state ) {
21190 name_list= RExC_paren_name_list;
21193 if ( k != REF || (OP(o) < REFN)) {
21194 SV **name= av_fetch(name_list, parno, 0 );
21196 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
21199 SV *sv_dat= MUTABLE_SV(progi->data->data[ parno ]);
21200 I32 *nums=(I32*)SvPVX(sv_dat);
21201 SV **name= av_fetch(name_list, nums[0], 0 );
21204 for ( n=0; n<SvIVX(sv_dat); n++ ) {
21205 Perl_sv_catpvf(aTHX_ sv, "%s%" IVdf,
21206 (n ? "," : ""), (IV)nums[n]);
21208 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
21212 if ( k == REF && reginfo) {
21213 U32 n = ARG(o); /* which paren pair */
21214 I32 ln = prog->offs[n].start;
21215 if (prog->lastparen < n || ln == -1 || prog->offs[n].end == -1)
21216 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
21217 else if (ln == prog->offs[n].end)
21218 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
21220 const char *s = reginfo->strbeg + ln;
21221 Perl_sv_catpvf(aTHX_ sv, ": ");
21222 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
21223 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
21226 } else if (k == GOSUB) {
21227 AV *name_list= NULL;
21228 if ( RXp_PAREN_NAMES(prog) ) {
21229 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
21230 } else if ( pRExC_state ) {
21231 name_list= RExC_paren_name_list;
21234 /* Paren and offset */
21235 Perl_sv_catpvf(aTHX_ sv, "%d[%+d:%d]", (int)ARG(o),(int)ARG2L(o),
21236 (int)((o + (int)ARG2L(o)) - progi->program) );
21238 SV **name= av_fetch(name_list, ARG(o), 0 );
21240 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
21243 else if (k == LOGICAL)
21244 /* 2: embedded, otherwise 1 */
21245 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
21246 else if (k == ANYOF || k == ANYOFR) {
21250 bool do_sep = FALSE; /* Do we need to separate various components of
21252 /* Set if there is still an unresolved user-defined property */
21253 SV *unresolved = NULL;
21255 /* Things that are ignored except when the runtime locale is UTF-8 */
21256 SV *only_utf8_locale_invlist = NULL;
21258 /* Code points that don't fit in the bitmap */
21259 SV *nonbitmap_invlist = NULL;
21261 /* And things that aren't in the bitmap, but are small enough to be */
21262 SV* bitmap_range_not_in_bitmap = NULL;
21266 if (inRANGE(OP(o), ANYOFH, ANYOFRb)) {
21272 flags = ANYOF_FLAGS(o);
21273 bitmap = ANYOF_BITMAP(o);
21277 if (OP(o) == ANYOFL || OP(o) == ANYOFPOSIXL) {
21278 if (ANYOFL_UTF8_LOCALE_REQD(flags)) {
21279 sv_catpvs(sv, "{utf8-locale-reqd}");
21281 if (flags & ANYOFL_FOLD) {
21282 sv_catpvs(sv, "{i}");
21286 inverted = flags & ANYOF_INVERT;
21288 /* If there is stuff outside the bitmap, get it */
21289 if (arg != ANYOF_ONLY_HAS_BITMAP) {
21290 if (inRANGE(OP(o), ANYOFR, ANYOFRb)) {
21291 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
21293 ANYOFRbase(o) + ANYOFRdelta(o));
21296 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
21297 (void) get_regclass_nonbitmap_data(prog, o, FALSE,
21299 &only_utf8_locale_invlist,
21300 &nonbitmap_invlist);
21302 (void) get_re_gclass_nonbitmap_data(prog, o, FALSE,
21304 &only_utf8_locale_invlist,
21305 &nonbitmap_invlist);
21309 /* The non-bitmap data may contain stuff that could fit in the
21310 * bitmap. This could come from a user-defined property being
21311 * finally resolved when this call was done; or much more likely
21312 * because there are matches that require UTF-8 to be valid, and so
21313 * aren't in the bitmap (or ANYOFR). This is teased apart later */
21314 _invlist_intersection(nonbitmap_invlist,
21316 &bitmap_range_not_in_bitmap);
21317 /* Leave just the things that don't fit into the bitmap */
21318 _invlist_subtract(nonbitmap_invlist,
21320 &nonbitmap_invlist);
21323 /* Obey this flag to add all above-the-bitmap code points */
21324 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
21325 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
21326 NUM_ANYOF_CODE_POINTS,
21330 /* Ready to start outputting. First, the initial left bracket */
21331 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
21333 /* ANYOFH by definition doesn't have anything that will fit inside the
21334 * bitmap; ANYOFR may or may not. */
21335 if ( ! inRANGE(OP(o), ANYOFH, ANYOFHr)
21336 && ( ! inRANGE(OP(o), ANYOFR, ANYOFRb)
21337 || ANYOFRbase(o) < NUM_ANYOF_CODE_POINTS))
21339 /* Then all the things that could fit in the bitmap */
21340 do_sep = put_charclass_bitmap_innards(sv,
21342 bitmap_range_not_in_bitmap,
21343 only_utf8_locale_invlist,
21347 /* Can't try inverting for a
21348 * better display if there
21349 * are things that haven't
21352 || inRANGE(OP(o), ANYOFR, ANYOFRb));
21353 SvREFCNT_dec(bitmap_range_not_in_bitmap);
21355 /* If there are user-defined properties which haven't been defined
21356 * yet, output them. If the result is not to be inverted, it is
21357 * clearest to output them in a separate [] from the bitmap range
21358 * stuff. If the result is to be complemented, we have to show
21359 * everything in one [], as the inversion applies to the whole
21360 * thing. Use {braces} to separate them from anything in the
21361 * bitmap and anything above the bitmap. */
21364 if (! do_sep) { /* If didn't output anything in the bitmap
21366 sv_catpvs(sv, "^");
21368 sv_catpvs(sv, "{");
21371 Perl_sv_catpvf(aTHX_ sv,"%s][%s", PL_colors[1],
21374 sv_catsv(sv, unresolved);
21376 sv_catpvs(sv, "}");
21378 do_sep = ! inverted;
21382 /* And, finally, add the above-the-bitmap stuff */
21383 if (nonbitmap_invlist && _invlist_len(nonbitmap_invlist)) {
21386 /* See if truncation size is overridden */
21387 const STRLEN dump_len = (PL_dump_re_max_len > 256)
21388 ? PL_dump_re_max_len
21391 /* This is output in a separate [] */
21393 Perl_sv_catpvf(aTHX_ sv,"%s][%s", PL_colors[1], PL_colors[0]);
21396 /* And, for easy of understanding, it is shown in the
21397 * uncomplemented form if possible. The one exception being if
21398 * there are unresolved items, where the inversion has to be
21399 * delayed until runtime */
21400 if (inverted && ! unresolved) {
21401 _invlist_invert(nonbitmap_invlist);
21402 _invlist_subtract(nonbitmap_invlist, PL_InBitmap, &nonbitmap_invlist);
21405 contents = invlist_contents(nonbitmap_invlist,
21406 FALSE /* output suitable for catsv */
21409 /* If the output is shorter than the permissible maximum, just do it. */
21410 if (SvCUR(contents) <= dump_len) {
21411 sv_catsv(sv, contents);
21414 const char * contents_string = SvPVX(contents);
21415 STRLEN i = dump_len;
21417 /* Otherwise, start at the permissible max and work back to the
21418 * first break possibility */
21419 while (i > 0 && contents_string[i] != ' ') {
21422 if (i == 0) { /* Fail-safe. Use the max if we couldn't
21423 find a legal break */
21427 sv_catpvn(sv, contents_string, i);
21428 sv_catpvs(sv, "...");
21431 SvREFCNT_dec_NN(contents);
21432 SvREFCNT_dec_NN(nonbitmap_invlist);
21435 /* And finally the matching, closing ']' */
21436 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
21438 if (OP(o) == ANYOFHs) {
21439 Perl_sv_catpvf(aTHX_ sv, " (Leading UTF-8 bytes=%s", _byte_dump_string((U8 *) ((struct regnode_anyofhs *) o)->string, FLAGS(o), 1));
21441 else if (inRANGE(OP(o), ANYOFH, ANYOFRb)) {
21442 U8 lowest = (OP(o) != ANYOFHr)
21444 : LOWEST_ANYOF_HRx_BYTE(FLAGS(o));
21445 U8 highest = (OP(o) == ANYOFHr)
21446 ? HIGHEST_ANYOF_HRx_BYTE(FLAGS(o))
21447 : (OP(o) == ANYOFH || OP(o) == ANYOFR)
21451 if (OP(o) != ANYOFR || ! isASCII(ANYOFRbase(o) + ANYOFRdelta(o)))
21454 Perl_sv_catpvf(aTHX_ sv, " (First UTF-8 byte=%02X", lowest);
21455 if (lowest != highest) {
21456 Perl_sv_catpvf(aTHX_ sv, "-%02X", highest);
21458 Perl_sv_catpvf(aTHX_ sv, ")");
21462 SvREFCNT_dec(unresolved);
21464 else if (k == ANYOFM) {
21465 SV * cp_list = get_ANYOFM_contents(o);
21467 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
21468 if (OP(o) == NANYOFM) {
21469 _invlist_invert(cp_list);
21472 put_charclass_bitmap_innards(sv, NULL, cp_list, NULL, NULL, 0, TRUE);
21473 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
21475 SvREFCNT_dec(cp_list);
21477 else if (k == POSIXD || k == NPOSIXD) {
21478 U8 index = FLAGS(o) * 2;
21479 if (index < C_ARRAY_LENGTH(anyofs)) {
21480 if (*anyofs[index] != '[') {
21481 sv_catpvs(sv, "[");
21483 sv_catpv(sv, anyofs[index]);
21484 if (*anyofs[index] != '[') {
21485 sv_catpvs(sv, "]");
21489 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
21492 else if (k == BOUND || k == NBOUND) {
21493 /* Must be synced with order of 'bound_type' in regcomp.h */
21494 const char * const bounds[] = {
21495 "", /* Traditional */
21501 assert(FLAGS(o) < C_ARRAY_LENGTH(bounds));
21502 sv_catpv(sv, bounds[FLAGS(o)]);
21504 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH)) {
21505 Perl_sv_catpvf(aTHX_ sv, "[%d", -(o->flags));
21507 Perl_sv_catpvf(aTHX_ sv, "..-%d", o->flags - o->next_off);
21509 Perl_sv_catpvf(aTHX_ sv, "]");
21511 else if (OP(o) == SBOL)
21512 Perl_sv_catpvf(aTHX_ sv, " /%s/", o->flags ? "\\A" : "^");
21514 /* add on the verb argument if there is one */
21515 if ( ( k == VERB || OP(o) == ACCEPT || OP(o) == OPFAIL ) && o->flags) {
21517 Perl_sv_catpvf(aTHX_ sv, ":%" SVf,
21518 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
21520 sv_catpvs(sv, ":NULL");
21523 PERL_UNUSED_CONTEXT;
21524 PERL_UNUSED_ARG(sv);
21525 PERL_UNUSED_ARG(o);
21526 PERL_UNUSED_ARG(prog);
21527 PERL_UNUSED_ARG(reginfo);
21528 PERL_UNUSED_ARG(pRExC_state);
21529 #endif /* DEBUGGING */
21535 Perl_re_intuit_string(pTHX_ REGEXP * const r)
21536 { /* Assume that RE_INTUIT is set */
21537 /* Returns an SV containing a string that must appear in the target for it
21538 * to match, or NULL if nothing is known that must match.
21540 * CAUTION: the SV can be freed during execution of the regex engine */
21542 struct regexp *const prog = ReANY(r);
21543 DECLARE_AND_GET_RE_DEBUG_FLAGS;
21545 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
21546 PERL_UNUSED_CONTEXT;
21550 if (prog->maxlen > 0) {
21551 const char * const s = SvPV_nolen_const(RX_UTF8(r)
21552 ? prog->check_utf8 : prog->check_substr);
21554 if (!PL_colorset) reginitcolors();
21555 Perl_re_printf( aTHX_
21556 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
21558 RX_UTF8(r) ? "utf8 " : "",
21559 PL_colors[5], PL_colors[0],
21562 (strlen(s) > PL_dump_re_max_len ? "..." : ""));
21566 /* use UTF8 check substring if regexp pattern itself is in UTF8 */
21567 return RX_UTF8(r) ? prog->check_utf8 : prog->check_substr;
21573 handles refcounting and freeing the perl core regexp structure. When
21574 it is necessary to actually free the structure the first thing it
21575 does is call the 'free' method of the regexp_engine associated to
21576 the regexp, allowing the handling of the void *pprivate; member
21577 first. (This routine is not overridable by extensions, which is why
21578 the extensions free is called first.)
21580 See regdupe and regdupe_internal if you change anything here.
21582 #ifndef PERL_IN_XSUB_RE
21584 Perl_pregfree(pTHX_ REGEXP *r)
21590 Perl_pregfree2(pTHX_ REGEXP *rx)
21592 struct regexp *const r = ReANY(rx);
21593 DECLARE_AND_GET_RE_DEBUG_FLAGS;
21595 PERL_ARGS_ASSERT_PREGFREE2;
21600 if (r->mother_re) {
21601 ReREFCNT_dec(r->mother_re);
21603 CALLREGFREE_PVT(rx); /* free the private data */
21604 SvREFCNT_dec(RXp_PAREN_NAMES(r));
21608 for (i = 0; i < 2; i++) {
21609 SvREFCNT_dec(r->substrs->data[i].substr);
21610 SvREFCNT_dec(r->substrs->data[i].utf8_substr);
21612 Safefree(r->substrs);
21614 RX_MATCH_COPY_FREE(rx);
21615 #ifdef PERL_ANY_COW
21616 SvREFCNT_dec(r->saved_copy);
21619 SvREFCNT_dec(r->qr_anoncv);
21620 if (r->recurse_locinput)
21621 Safefree(r->recurse_locinput);
21627 Copy ssv to dsv, both of which should of type SVt_REGEXP or SVt_PVLV,
21628 except that dsv will be created if NULL.
21630 This function is used in two main ways. First to implement
21631 $r = qr/....; $s = $$r;
21633 Secondly, it is used as a hacky workaround to the structural issue of
21635 being stored in the regexp structure which is in turn stored in
21636 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
21637 could be PL_curpm in multiple contexts, and could require multiple
21638 result sets being associated with the pattern simultaneously, such
21639 as when doing a recursive match with (??{$qr})
21641 The solution is to make a lightweight copy of the regexp structure
21642 when a qr// is returned from the code executed by (??{$qr}) this
21643 lightweight copy doesn't actually own any of its data except for
21644 the starp/end and the actual regexp structure itself.
21650 Perl_reg_temp_copy(pTHX_ REGEXP *dsv, REGEXP *ssv)
21652 struct regexp *drx;
21653 struct regexp *const srx = ReANY(ssv);
21654 const bool islv = dsv && SvTYPE(dsv) == SVt_PVLV;
21656 PERL_ARGS_ASSERT_REG_TEMP_COPY;
21659 dsv = (REGEXP*) newSV_type(SVt_REGEXP);
21661 assert(SvTYPE(dsv) == SVt_REGEXP || (SvTYPE(dsv) == SVt_PVLV));
21663 /* our only valid caller, sv_setsv_flags(), should have done
21664 * a SV_CHECK_THINKFIRST_COW_DROP() by now */
21665 assert(!SvOOK(dsv));
21666 assert(!SvIsCOW(dsv));
21667 assert(!SvROK(dsv));
21669 if (SvPVX_const(dsv)) {
21671 Safefree(SvPVX(dsv));
21676 SvOK_off((SV *)dsv);
21679 /* For PVLVs, the head (sv_any) points to an XPVLV, while
21680 * the LV's xpvlenu_rx will point to a regexp body, which
21681 * we allocate here */
21682 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
21683 assert(!SvPVX(dsv));
21684 ((XPV*)SvANY(dsv))->xpv_len_u.xpvlenu_rx = temp->sv_any;
21685 temp->sv_any = NULL;
21686 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
21687 SvREFCNT_dec_NN(temp);
21688 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
21689 ing below will not set it. */
21690 SvCUR_set(dsv, SvCUR(ssv));
21693 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
21694 sv_force_normal(sv) is called. */
21698 SvFLAGS(dsv) |= SvFLAGS(ssv) & (SVf_POK|SVp_POK|SVf_UTF8);
21699 SvPV_set(dsv, RX_WRAPPED(ssv));
21700 /* We share the same string buffer as the original regexp, on which we
21701 hold a reference count, incremented when mother_re is set below.
21702 The string pointer is copied here, being part of the regexp struct.
21704 memcpy(&(drx->xpv_cur), &(srx->xpv_cur),
21705 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
21709 const I32 npar = srx->nparens+1;
21710 Newx(drx->offs, npar, regexp_paren_pair);
21711 Copy(srx->offs, drx->offs, npar, regexp_paren_pair);
21713 if (srx->substrs) {
21715 Newx(drx->substrs, 1, struct reg_substr_data);
21716 StructCopy(srx->substrs, drx->substrs, struct reg_substr_data);
21718 for (i = 0; i < 2; i++) {
21719 SvREFCNT_inc_void(drx->substrs->data[i].substr);
21720 SvREFCNT_inc_void(drx->substrs->data[i].utf8_substr);
21723 /* check_substr and check_utf8, if non-NULL, point to either their
21724 anchored or float namesakes, and don't hold a second reference. */
21726 RX_MATCH_COPIED_off(dsv);
21727 #ifdef PERL_ANY_COW
21728 drx->saved_copy = NULL;
21730 drx->mother_re = ReREFCNT_inc(srx->mother_re ? srx->mother_re : ssv);
21731 SvREFCNT_inc_void(drx->qr_anoncv);
21732 if (srx->recurse_locinput)
21733 Newx(drx->recurse_locinput, srx->nparens + 1, char *);
21740 /* regfree_internal()
21742 Free the private data in a regexp. This is overloadable by
21743 extensions. Perl takes care of the regexp structure in pregfree(),
21744 this covers the *pprivate pointer which technically perl doesn't
21745 know about, however of course we have to handle the
21746 regexp_internal structure when no extension is in use.
21748 Note this is called before freeing anything in the regexp
21753 Perl_regfree_internal(pTHX_ REGEXP * const rx)
21755 struct regexp *const r = ReANY(rx);
21756 RXi_GET_DECL(r, ri);
21757 DECLARE_AND_GET_RE_DEBUG_FLAGS;
21759 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
21769 SV *dsv= sv_newmortal();
21770 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
21771 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), PL_dump_re_max_len);
21772 Perl_re_printf( aTHX_ "%sFreeing REx:%s %s\n",
21773 PL_colors[4], PL_colors[5], s);
21777 #ifdef RE_TRACK_PATTERN_OFFSETS
21779 Safefree(ri->u.offsets); /* 20010421 MJD */
21781 if (ri->code_blocks)
21782 S_free_codeblocks(aTHX_ ri->code_blocks);
21785 int n = ri->data->count;
21788 /* If you add a ->what type here, update the comment in regcomp.h */
21789 switch (ri->data->what[n]) {
21795 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
21798 Safefree(ri->data->data[n]);
21804 { /* Aho Corasick add-on structure for a trie node.
21805 Used in stclass optimization only */
21807 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
21809 refcount = --aho->refcount;
21812 PerlMemShared_free(aho->states);
21813 PerlMemShared_free(aho->fail);
21814 /* do this last!!!! */
21815 PerlMemShared_free(ri->data->data[n]);
21816 /* we should only ever get called once, so
21817 * assert as much, and also guard the free
21818 * which /might/ happen twice. At the least
21819 * it will make code anlyzers happy and it
21820 * doesn't cost much. - Yves */
21821 assert(ri->regstclass);
21822 if (ri->regstclass) {
21823 PerlMemShared_free(ri->regstclass);
21824 ri->regstclass = 0;
21831 /* trie structure. */
21833 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
21835 refcount = --trie->refcount;
21838 PerlMemShared_free(trie->charmap);
21839 PerlMemShared_free(trie->states);
21840 PerlMemShared_free(trie->trans);
21842 PerlMemShared_free(trie->bitmap);
21844 PerlMemShared_free(trie->jump);
21845 PerlMemShared_free(trie->wordinfo);
21846 /* do this last!!!! */
21847 PerlMemShared_free(ri->data->data[n]);
21852 Perl_croak(aTHX_ "panic: regfree data code '%c'",
21853 ri->data->what[n]);
21856 Safefree(ri->data->what);
21857 Safefree(ri->data);
21863 #define av_dup_inc(s, t) MUTABLE_AV(sv_dup_inc((const SV *)s, t))
21864 #define hv_dup_inc(s, t) MUTABLE_HV(sv_dup_inc((const SV *)s, t))
21865 #define SAVEPVN(p, n) ((p) ? savepvn(p, n) : NULL)
21868 =for apidoc re_dup_guts
21869 Duplicate a regexp.
21871 This routine is expected to clone a given regexp structure. It is only
21872 compiled under USE_ITHREADS.
21874 After all of the core data stored in struct regexp is duplicated
21875 the C<regexp_engine.dupe> method is used to copy any private data
21876 stored in the *pprivate pointer. This allows extensions to handle
21877 any duplication they need to do.
21881 See pregfree() and regfree_internal() if you change anything here.
21883 #if defined(USE_ITHREADS)
21884 #ifndef PERL_IN_XSUB_RE
21886 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
21889 const struct regexp *r = ReANY(sstr);
21890 struct regexp *ret = ReANY(dstr);
21892 PERL_ARGS_ASSERT_RE_DUP_GUTS;
21894 npar = r->nparens+1;
21895 Newx(ret->offs, npar, regexp_paren_pair);
21896 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
21898 if (ret->substrs) {
21899 /* Do it this way to avoid reading from *r after the StructCopy().
21900 That way, if any of the sv_dup_inc()s dislodge *r from the L1
21901 cache, it doesn't matter. */
21903 const bool anchored = r->check_substr
21904 ? r->check_substr == r->substrs->data[0].substr
21905 : r->check_utf8 == r->substrs->data[0].utf8_substr;
21906 Newx(ret->substrs, 1, struct reg_substr_data);
21907 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
21909 for (i = 0; i < 2; i++) {
21910 ret->substrs->data[i].substr =
21911 sv_dup_inc(ret->substrs->data[i].substr, param);
21912 ret->substrs->data[i].utf8_substr =
21913 sv_dup_inc(ret->substrs->data[i].utf8_substr, param);
21916 /* check_substr and check_utf8, if non-NULL, point to either their
21917 anchored or float namesakes, and don't hold a second reference. */
21919 if (ret->check_substr) {
21921 assert(r->check_utf8 == r->substrs->data[0].utf8_substr);
21923 ret->check_substr = ret->substrs->data[0].substr;
21924 ret->check_utf8 = ret->substrs->data[0].utf8_substr;
21926 assert(r->check_substr == r->substrs->data[1].substr);
21927 assert(r->check_utf8 == r->substrs->data[1].utf8_substr);
21929 ret->check_substr = ret->substrs->data[1].substr;
21930 ret->check_utf8 = ret->substrs->data[1].utf8_substr;
21932 } else if (ret->check_utf8) {
21934 ret->check_utf8 = ret->substrs->data[0].utf8_substr;
21936 ret->check_utf8 = ret->substrs->data[1].utf8_substr;
21941 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
21942 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
21943 if (r->recurse_locinput)
21944 Newx(ret->recurse_locinput, r->nparens + 1, char *);
21947 RXi_SET(ret, CALLREGDUPE_PVT(dstr, param));
21949 if (RX_MATCH_COPIED(dstr))
21950 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
21952 ret->subbeg = NULL;
21953 #ifdef PERL_ANY_COW
21954 ret->saved_copy = NULL;
21957 /* Whether mother_re be set or no, we need to copy the string. We
21958 cannot refrain from copying it when the storage points directly to
21959 our mother regexp, because that's
21960 1: a buffer in a different thread
21961 2: something we no longer hold a reference on
21962 so we need to copy it locally. */
21963 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED_const(sstr), SvCUR(sstr)+1);
21964 /* set malloced length to a non-zero value so it will be freed
21965 * (otherwise in combination with SVf_FAKE it looks like an alien
21966 * buffer). It doesn't have to be the actual malloced size, since it
21967 * should never be grown */
21968 SvLEN_set(dstr, SvCUR(sstr)+1);
21969 ret->mother_re = NULL;
21971 #endif /* PERL_IN_XSUB_RE */
21976 This is the internal complement to regdupe() which is used to copy
21977 the structure pointed to by the *pprivate pointer in the regexp.
21978 This is the core version of the extension overridable cloning hook.
21979 The regexp structure being duplicated will be copied by perl prior
21980 to this and will be provided as the regexp *r argument, however
21981 with the /old/ structures pprivate pointer value. Thus this routine
21982 may override any copying normally done by perl.
21984 It returns a pointer to the new regexp_internal structure.
21988 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
21990 struct regexp *const r = ReANY(rx);
21991 regexp_internal *reti;
21993 RXi_GET_DECL(r, ri);
21995 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
21999 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
22000 char, regexp_internal);
22001 Copy(ri->program, reti->program, len+1, regnode);
22004 if (ri->code_blocks) {
22006 Newx(reti->code_blocks, 1, struct reg_code_blocks);
22007 Newx(reti->code_blocks->cb, ri->code_blocks->count,
22008 struct reg_code_block);
22009 Copy(ri->code_blocks->cb, reti->code_blocks->cb,
22010 ri->code_blocks->count, struct reg_code_block);
22011 for (n = 0; n < ri->code_blocks->count; n++)
22012 reti->code_blocks->cb[n].src_regex = (REGEXP*)
22013 sv_dup_inc((SV*)(ri->code_blocks->cb[n].src_regex), param);
22014 reti->code_blocks->count = ri->code_blocks->count;
22015 reti->code_blocks->refcnt = 1;
22018 reti->code_blocks = NULL;
22020 reti->regstclass = NULL;
22023 struct reg_data *d;
22024 const int count = ri->data->count;
22027 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
22028 char, struct reg_data);
22029 Newx(d->what, count, U8);
22032 for (i = 0; i < count; i++) {
22033 d->what[i] = ri->data->what[i];
22034 switch (d->what[i]) {
22035 /* see also regcomp.h and regfree_internal() */
22036 case 'a': /* actually an AV, but the dup function is identical.
22037 values seem to be "plain sv's" generally. */
22038 case 'r': /* a compiled regex (but still just another SV) */
22039 case 's': /* an RV (currently only used for an RV to an AV by the ANYOF code)
22040 this use case should go away, the code could have used
22041 'a' instead - see S_set_ANYOF_arg() for array contents. */
22042 case 'S': /* actually an SV, but the dup function is identical. */
22043 case 'u': /* actually an HV, but the dup function is identical.
22044 values are "plain sv's" */
22045 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
22048 /* Synthetic Start Class - "Fake" charclass we generate to optimize
22049 * patterns which could start with several different things. Pre-TRIE
22050 * this was more important than it is now, however this still helps
22051 * in some places, for instance /x?a+/ might produce a SSC equivalent
22052 * to [xa]. This is used by Perl_re_intuit_start() and S_find_byclass()
22055 /* This is cheating. */
22056 Newx(d->data[i], 1, regnode_ssc);
22057 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
22058 reti->regstclass = (regnode*)d->data[i];
22061 /* AHO-CORASICK fail table */
22062 /* Trie stclasses are readonly and can thus be shared
22063 * without duplication. We free the stclass in pregfree
22064 * when the corresponding reg_ac_data struct is freed.
22066 reti->regstclass= ri->regstclass;
22069 /* TRIE transition table */
22071 ((reg_trie_data*)ri->data->data[i])->refcount++;
22074 case 'l': /* (?{...}) or (??{ ... }) code (cb->block) */
22075 case 'L': /* same when RExC_pm_flags & PMf_HAS_CV and code
22076 is not from another regexp */
22077 d->data[i] = ri->data->data[i];
22080 Perl_croak(aTHX_ "panic: re_dup_guts unknown data code '%c'",
22081 ri->data->what[i]);
22090 reti->name_list_idx = ri->name_list_idx;
22092 #ifdef RE_TRACK_PATTERN_OFFSETS
22093 if (ri->u.offsets) {
22094 Newx(reti->u.offsets, 2*len+1, U32);
22095 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
22098 SetProgLen(reti, len);
22101 return (void*)reti;
22104 #endif /* USE_ITHREADS */
22106 #ifndef PERL_IN_XSUB_RE
22109 - regnext - dig the "next" pointer out of a node
22112 Perl_regnext(pTHX_ regnode *p)
22119 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
22120 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
22121 (int)OP(p), (int)REGNODE_MAX);
22124 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
22134 S_re_croak(pTHX_ bool utf8, const char* pat,...)
22137 STRLEN len = strlen(pat);
22140 const char *message;
22142 PERL_ARGS_ASSERT_RE_CROAK;
22146 Copy(pat, buf, len , char);
22148 buf[len + 1] = '\0';
22149 va_start(args, pat);
22150 msv = vmess(buf, &args);
22152 message = SvPV_const(msv, len);
22155 Copy(message, buf, len , char);
22156 /* len-1 to avoid \n */
22157 Perl_croak(aTHX_ "%" UTF8f, UTF8fARG(utf8, len-1, buf));
22160 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
22162 #ifndef PERL_IN_XSUB_RE
22164 Perl_save_re_context(pTHX)
22169 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
22172 const REGEXP * const rx = PM_GETRE(PL_curpm);
22174 nparens = RX_NPARENS(rx);
22177 /* RT #124109. This is a complete hack; in the SWASHNEW case we know
22178 * that PL_curpm will be null, but that utf8.pm and the modules it
22179 * loads will only use $1..$3.
22180 * The t/porting/re_context.t test file checks this assumption.
22185 for (i = 1; i <= nparens; i++) {
22186 char digits[TYPE_CHARS(long)];
22187 const STRLEN len = my_snprintf(digits, sizeof(digits),
22189 GV *const *const gvp
22190 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
22193 GV * const gv = *gvp;
22194 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
22204 S_put_code_point(pTHX_ SV *sv, UV c)
22206 PERL_ARGS_ASSERT_PUT_CODE_POINT;
22209 Perl_sv_catpvf(aTHX_ sv, "\\x{%04" UVXf "}", c);
22211 else if (isPRINT(c)) {
22212 const char string = (char) c;
22214 /* We use {phrase} as metanotation in the class, so also escape literal
22216 if (isBACKSLASHED_PUNCT(c) || c == '{' || c == '}')
22217 sv_catpvs(sv, "\\");
22218 sv_catpvn(sv, &string, 1);
22220 else if (isMNEMONIC_CNTRL(c)) {
22221 Perl_sv_catpvf(aTHX_ sv, "%s", cntrl_to_mnemonic((U8) c));
22224 Perl_sv_catpvf(aTHX_ sv, "\\x%02X", (U8) c);
22229 S_put_range(pTHX_ SV *sv, UV start, const UV end, const bool allow_literals)
22231 /* Appends to 'sv' a displayable version of the range of code points from
22232 * 'start' to 'end'. Mnemonics (like '\r') are used for the few controls
22233 * that have them, when they occur at the beginning or end of the range.
22234 * It uses hex to output the remaining code points, unless 'allow_literals'
22235 * is true, in which case the printable ASCII ones are output as-is (though
22236 * some of these will be escaped by put_code_point()).
22238 * NOTE: This is designed only for printing ranges of code points that fit
22239 * inside an ANYOF bitmap. Higher code points are simply suppressed
22242 const unsigned int min_range_count = 3;
22244 assert(start <= end);
22246 PERL_ARGS_ASSERT_PUT_RANGE;
22248 while (start <= end) {
22250 const char * format;
22252 if ( end - start < min_range_count
22253 && (end - start <= 2 || (isPRINT_A(start) && isPRINT_A(end))))
22255 /* Output a range of 1 or 2 chars individually, or longer ranges
22256 * when printable */
22257 for (; start <= end; start++) {
22258 put_code_point(sv, start);
22263 /* If permitted by the input options, and there is a possibility that
22264 * this range contains a printable literal, look to see if there is
22266 if (allow_literals && start <= MAX_PRINT_A) {
22268 /* If the character at the beginning of the range isn't an ASCII
22269 * printable, effectively split the range into two parts:
22270 * 1) the portion before the first such printable,
22272 * and output them separately. */
22273 if (! isPRINT_A(start)) {
22274 UV temp_end = start + 1;
22276 /* There is no point looking beyond the final possible
22277 * printable, in MAX_PRINT_A */
22278 UV max = MIN(end, MAX_PRINT_A);
22280 while (temp_end <= max && ! isPRINT_A(temp_end)) {
22284 /* Here, temp_end points to one beyond the first printable if
22285 * found, or to one beyond 'max' if not. If none found, make
22286 * sure that we use the entire range */
22287 if (temp_end > MAX_PRINT_A) {
22288 temp_end = end + 1;
22291 /* Output the first part of the split range: the part that
22292 * doesn't have printables, with the parameter set to not look
22293 * for literals (otherwise we would infinitely recurse) */
22294 put_range(sv, start, temp_end - 1, FALSE);
22296 /* The 2nd part of the range (if any) starts here. */
22299 /* We do a continue, instead of dropping down, because even if
22300 * the 2nd part is non-empty, it could be so short that we want
22301 * to output it as individual characters, as tested for at the
22302 * top of this loop. */
22306 /* Here, 'start' is a printable ASCII. If it is an alphanumeric,
22307 * output a sub-range of just the digits or letters, then process
22308 * the remaining portion as usual. */
22309 if (isALPHANUMERIC_A(start)) {
22310 UV mask = (isDIGIT_A(start))
22315 UV temp_end = start + 1;
22317 /* Find the end of the sub-range that includes just the
22318 * characters in the same class as the first character in it */
22319 while (temp_end <= end && _generic_isCC_A(temp_end, mask)) {
22324 /* For short ranges, don't duplicate the code above to output
22325 * them; just call recursively */
22326 if (temp_end - start < min_range_count) {
22327 put_range(sv, start, temp_end, FALSE);
22329 else { /* Output as a range */
22330 put_code_point(sv, start);
22331 sv_catpvs(sv, "-");
22332 put_code_point(sv, temp_end);
22334 start = temp_end + 1;
22338 /* We output any other printables as individual characters */
22339 if (isPUNCT_A(start) || isSPACE_A(start)) {
22340 while (start <= end && (isPUNCT_A(start)
22341 || isSPACE_A(start)))
22343 put_code_point(sv, start);
22348 } /* End of looking for literals */
22350 /* Here is not to output as a literal. Some control characters have
22351 * mnemonic names. Split off any of those at the beginning and end of
22352 * the range to print mnemonically. It isn't possible for many of
22353 * these to be in a row, so this won't overwhelm with output */
22355 && (isMNEMONIC_CNTRL(start) || isMNEMONIC_CNTRL(end)))
22357 while (isMNEMONIC_CNTRL(start) && start <= end) {
22358 put_code_point(sv, start);
22362 /* If this didn't take care of the whole range ... */
22363 if (start <= end) {
22365 /* Look backwards from the end to find the final non-mnemonic
22368 while (isMNEMONIC_CNTRL(temp_end)) {
22372 /* And separately output the interior range that doesn't start
22373 * or end with mnemonics */
22374 put_range(sv, start, temp_end, FALSE);
22376 /* Then output the mnemonic trailing controls */
22377 start = temp_end + 1;
22378 while (start <= end) {
22379 put_code_point(sv, start);
22386 /* As a final resort, output the range or subrange as hex. */
22388 if (start >= NUM_ANYOF_CODE_POINTS) {
22391 else { /* Have to split range at the bitmap boundary */
22392 this_end = (end < NUM_ANYOF_CODE_POINTS)
22394 : NUM_ANYOF_CODE_POINTS - 1;
22396 #if NUM_ANYOF_CODE_POINTS > 256
22397 format = (this_end < 256)
22398 ? "\\x%02" UVXf "-\\x%02" UVXf
22399 : "\\x{%04" UVXf "}-\\x{%04" UVXf "}";
22401 format = "\\x%02" UVXf "-\\x%02" UVXf;
22403 GCC_DIAG_IGNORE_STMT(-Wformat-nonliteral);
22404 Perl_sv_catpvf(aTHX_ sv, format, start, this_end);
22405 GCC_DIAG_RESTORE_STMT;
22411 S_put_charclass_bitmap_innards_invlist(pTHX_ SV *sv, SV* invlist)
22413 /* Concatenate onto the PV in 'sv' a displayable form of the inversion list
22417 bool allow_literals = TRUE;
22419 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_INVLIST;
22421 /* Generally, it is more readable if printable characters are output as
22422 * literals, but if a range (nearly) spans all of them, it's best to output
22423 * it as a single range. This code will use a single range if all but 2
22424 * ASCII printables are in it */
22425 invlist_iterinit(invlist);
22426 while (invlist_iternext(invlist, &start, &end)) {
22428 /* If the range starts beyond the final printable, it doesn't have any
22430 if (start > MAX_PRINT_A) {
22434 /* In both ASCII and EBCDIC, a SPACE is the lowest printable. To span
22435 * all but two, the range must start and end no later than 2 from
22437 if (start < ' ' + 2 && end > MAX_PRINT_A - 2) {
22438 if (end > MAX_PRINT_A) {
22444 if (end - start >= MAX_PRINT_A - ' ' - 2) {
22445 allow_literals = FALSE;
22450 invlist_iterfinish(invlist);
22452 /* Here we have figured things out. Output each range */
22453 invlist_iterinit(invlist);
22454 while (invlist_iternext(invlist, &start, &end)) {
22455 if (start >= NUM_ANYOF_CODE_POINTS) {
22458 put_range(sv, start, end, allow_literals);
22460 invlist_iterfinish(invlist);
22466 S_put_charclass_bitmap_innards_common(pTHX_
22467 SV* invlist, /* The bitmap */
22468 SV* posixes, /* Under /l, things like [:word:], \S */
22469 SV* only_utf8, /* Under /d, matches iff the target is UTF-8 */
22470 SV* not_utf8, /* /d, matches iff the target isn't UTF-8 */
22471 SV* only_utf8_locale, /* Under /l, matches if the locale is UTF-8 */
22472 const bool invert /* Is the result to be inverted? */
22475 /* Create and return an SV containing a displayable version of the bitmap
22476 * and associated information determined by the input parameters. If the
22477 * output would have been only the inversion indicator '^', NULL is instead
22482 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_COMMON;
22485 output = newSVpvs("^");
22488 output = newSVpvs("");
22491 /* First, the code points in the bitmap that are unconditionally there */
22492 put_charclass_bitmap_innards_invlist(output, invlist);
22494 /* Traditionally, these have been placed after the main code points */
22496 sv_catsv(output, posixes);
22499 if (only_utf8 && _invlist_len(only_utf8)) {
22500 Perl_sv_catpvf(aTHX_ output, "%s{utf8}%s", PL_colors[1], PL_colors[0]);
22501 put_charclass_bitmap_innards_invlist(output, only_utf8);
22504 if (not_utf8 && _invlist_len(not_utf8)) {
22505 Perl_sv_catpvf(aTHX_ output, "%s{not utf8}%s", PL_colors[1], PL_colors[0]);
22506 put_charclass_bitmap_innards_invlist(output, not_utf8);
22509 if (only_utf8_locale && _invlist_len(only_utf8_locale)) {
22510 Perl_sv_catpvf(aTHX_ output, "%s{utf8 locale}%s", PL_colors[1], PL_colors[0]);
22511 put_charclass_bitmap_innards_invlist(output, only_utf8_locale);
22513 /* This is the only list in this routine that can legally contain code
22514 * points outside the bitmap range. The call just above to
22515 * 'put_charclass_bitmap_innards_invlist' will simply suppress them, so
22516 * output them here. There's about a half-dozen possible, and none in
22517 * contiguous ranges longer than 2 */
22518 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
22520 SV* above_bitmap = NULL;
22522 _invlist_subtract(only_utf8_locale, PL_InBitmap, &above_bitmap);
22524 invlist_iterinit(above_bitmap);
22525 while (invlist_iternext(above_bitmap, &start, &end)) {
22528 for (i = start; i <= end; i++) {
22529 put_code_point(output, i);
22532 invlist_iterfinish(above_bitmap);
22533 SvREFCNT_dec_NN(above_bitmap);
22537 if (invert && SvCUR(output) == 1) {
22545 S_put_charclass_bitmap_innards(pTHX_ SV *sv,
22547 SV *nonbitmap_invlist,
22548 SV *only_utf8_locale_invlist,
22549 const regnode * const node,
22551 const bool force_as_is_display)
22553 /* Appends to 'sv' a displayable version of the innards of the bracketed
22554 * character class defined by the other arguments:
22555 * 'bitmap' points to the bitmap, or NULL if to ignore that.
22556 * 'nonbitmap_invlist' is an inversion list of the code points that are in
22557 * the bitmap range, but for some reason aren't in the bitmap; NULL if
22558 * none. The reasons for this could be that they require some
22559 * condition such as the target string being or not being in UTF-8
22560 * (under /d), or because they came from a user-defined property that
22561 * was not resolved at the time of the regex compilation (under /u)
22562 * 'only_utf8_locale_invlist' is an inversion list of the code points that
22563 * are valid only if the runtime locale is a UTF-8 one; NULL if none
22564 * 'node' is the regex pattern ANYOF node. It is needed only when the
22565 * above two parameters are not null, and is passed so that this
22566 * routine can tease apart the various reasons for them.
22567 * 'flags' is the flags field of 'node'
22568 * 'force_as_is_display' is TRUE if this routine should definitely NOT try
22569 * to invert things to see if that leads to a cleaner display. If
22570 * FALSE, this routine is free to use its judgment about doing this.
22572 * It returns TRUE if there was actually something output. (It may be that
22573 * the bitmap, etc is empty.)
22575 * When called for outputting the bitmap of a non-ANYOF node, just pass the
22576 * bitmap, with the succeeding parameters set to NULL, and the final one to
22580 /* In general, it tries to display the 'cleanest' representation of the
22581 * innards, choosing whether to display them inverted or not, regardless of
22582 * whether the class itself is to be inverted. However, there are some
22583 * cases where it can't try inverting, as what actually matches isn't known
22584 * until runtime, and hence the inversion isn't either. */
22586 bool inverting_allowed = ! force_as_is_display;
22589 STRLEN orig_sv_cur = SvCUR(sv);
22591 SV* invlist; /* Inversion list we accumulate of code points that
22592 are unconditionally matched */
22593 SV* only_utf8 = NULL; /* Under /d, list of matches iff the target is
22595 SV* not_utf8 = NULL; /* /d, list of matches iff the target isn't UTF-8
22597 SV* posixes = NULL; /* Under /l, string of things like [:word:], \D */
22598 SV* only_utf8_locale = NULL; /* Under /l, list of matches if the locale
22601 SV* as_is_display; /* The output string when we take the inputs
22603 SV* inverted_display; /* The output string when we invert the inputs */
22605 bool invert = cBOOL(flags & ANYOF_INVERT); /* Is the input to be inverted
22607 /* We are biased in favor of displaying things without them being inverted,
22608 * as that is generally easier to understand */
22609 const int bias = 5;
22611 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS;
22613 /* Start off with whatever code points are passed in. (We clone, so we
22614 * don't change the caller's list) */
22615 if (nonbitmap_invlist) {
22616 assert(invlist_highest(nonbitmap_invlist) < NUM_ANYOF_CODE_POINTS);
22617 invlist = invlist_clone(nonbitmap_invlist, NULL);
22619 else { /* Worst case size is every other code point is matched */
22620 invlist = _new_invlist(NUM_ANYOF_CODE_POINTS / 2);
22624 if (OP(node) == ANYOFD) {
22626 /* This flag indicates that the code points below 0x100 in the
22627 * nonbitmap list are precisely the ones that match only when the
22628 * target is UTF-8 (they should all be non-ASCII). */
22629 if (flags & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)
22631 _invlist_intersection(invlist, PL_UpperLatin1, &only_utf8);
22632 _invlist_subtract(invlist, only_utf8, &invlist);
22635 /* And this flag for matching all non-ASCII 0xFF and below */
22636 if (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)
22638 not_utf8 = invlist_clone(PL_UpperLatin1, NULL);
22641 else if (OP(node) == ANYOFL || OP(node) == ANYOFPOSIXL) {
22643 /* If either of these flags are set, what matches isn't
22644 * determinable except during execution, so don't know enough here
22646 if (flags & (ANYOFL_FOLD|ANYOF_MATCHES_POSIXL)) {
22647 inverting_allowed = FALSE;
22650 /* What the posix classes match also varies at runtime, so these
22651 * will be output symbolically. */
22652 if (ANYOF_POSIXL_TEST_ANY_SET(node)) {
22655 posixes = newSVpvs("");
22656 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
22657 if (ANYOF_POSIXL_TEST(node, i)) {
22658 sv_catpv(posixes, anyofs[i]);
22665 /* Accumulate the bit map into the unconditional match list */
22667 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
22668 if (BITMAP_TEST(bitmap, i)) {
22671 i < NUM_ANYOF_CODE_POINTS && BITMAP_TEST(bitmap, i);
22674 invlist = _add_range_to_invlist(invlist, start, i-1);
22679 /* Make sure that the conditional match lists don't have anything in them
22680 * that match unconditionally; otherwise the output is quite confusing.
22681 * This could happen if the code that populates these misses some
22684 _invlist_subtract(only_utf8, invlist, &only_utf8);
22687 _invlist_subtract(not_utf8, invlist, ¬_utf8);
22690 if (only_utf8_locale_invlist) {
22692 /* Since this list is passed in, we have to make a copy before
22694 only_utf8_locale = invlist_clone(only_utf8_locale_invlist, NULL);
22696 _invlist_subtract(only_utf8_locale, invlist, &only_utf8_locale);
22698 /* And, it can get really weird for us to try outputting an inverted
22699 * form of this list when it has things above the bitmap, so don't even
22701 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
22702 inverting_allowed = FALSE;
22706 /* Calculate what the output would be if we take the input as-is */
22707 as_is_display = put_charclass_bitmap_innards_common(invlist,
22714 /* If have to take the output as-is, just do that */
22715 if (! inverting_allowed) {
22716 if (as_is_display) {
22717 sv_catsv(sv, as_is_display);
22718 SvREFCNT_dec_NN(as_is_display);
22721 else { /* But otherwise, create the output again on the inverted input, and
22722 use whichever version is shorter */
22724 int inverted_bias, as_is_bias;
22726 /* We will apply our bias to whichever of the results doesn't have
22736 inverted_bias = bias;
22739 /* Now invert each of the lists that contribute to the output,
22740 * excluding from the result things outside the possible range */
22742 /* For the unconditional inversion list, we have to add in all the
22743 * conditional code points, so that when inverted, they will be gone
22745 _invlist_union(only_utf8, invlist, &invlist);
22746 _invlist_union(not_utf8, invlist, &invlist);
22747 _invlist_union(only_utf8_locale, invlist, &invlist);
22748 _invlist_invert(invlist);
22749 _invlist_intersection(invlist, PL_InBitmap, &invlist);
22752 _invlist_invert(only_utf8);
22753 _invlist_intersection(only_utf8, PL_UpperLatin1, &only_utf8);
22755 else if (not_utf8) {
22757 /* If a code point matches iff the target string is not in UTF-8,
22758 * then complementing the result has it not match iff not in UTF-8,
22759 * which is the same thing as matching iff it is UTF-8. */
22760 only_utf8 = not_utf8;
22764 if (only_utf8_locale) {
22765 _invlist_invert(only_utf8_locale);
22766 _invlist_intersection(only_utf8_locale,
22768 &only_utf8_locale);
22771 inverted_display = put_charclass_bitmap_innards_common(
22776 only_utf8_locale, invert);
22778 /* Use the shortest representation, taking into account our bias
22779 * against showing it inverted */
22780 if ( inverted_display
22781 && ( ! as_is_display
22782 || ( SvCUR(inverted_display) + inverted_bias
22783 < SvCUR(as_is_display) + as_is_bias)))
22785 sv_catsv(sv, inverted_display);
22787 else if (as_is_display) {
22788 sv_catsv(sv, as_is_display);
22791 SvREFCNT_dec(as_is_display);
22792 SvREFCNT_dec(inverted_display);
22795 SvREFCNT_dec_NN(invlist);
22796 SvREFCNT_dec(only_utf8);
22797 SvREFCNT_dec(not_utf8);
22798 SvREFCNT_dec(posixes);
22799 SvREFCNT_dec(only_utf8_locale);
22801 return SvCUR(sv) > orig_sv_cur;
22804 #define CLEAR_OPTSTART \
22805 if (optstart) STMT_START { \
22806 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ \
22807 " (%" IVdf " nodes)\n", (IV)(node - optstart))); \
22811 #define DUMPUNTIL(b,e) \
22813 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
22815 STATIC const regnode *
22816 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
22817 const regnode *last, const regnode *plast,
22818 SV* sv, I32 indent, U32 depth)
22820 U8 op = PSEUDO; /* Arbitrary non-END op. */
22821 const regnode *next;
22822 const regnode *optstart= NULL;
22824 RXi_GET_DECL(r, ri);
22825 DECLARE_AND_GET_RE_DEBUG_FLAGS;
22827 PERL_ARGS_ASSERT_DUMPUNTIL;
22829 #ifdef DEBUG_DUMPUNTIL
22830 Perl_re_printf( aTHX_ "--- %d : %d - %d - %d\n", indent, node-start,
22831 last ? last-start : 0, plast ? plast-start : 0);
22834 if (plast && plast < last)
22837 while (PL_regkind[op] != END && (!last || node < last)) {
22839 /* While that wasn't END last time... */
22842 if (op == CLOSE || op == SRCLOSE || op == WHILEM)
22844 next = regnext((regnode *)node);
22847 if (OP(node) == OPTIMIZED) {
22848 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
22855 regprop(r, sv, node, NULL, NULL);
22856 Perl_re_printf( aTHX_ "%4" IVdf ":%*s%s", (IV)(node - start),
22857 (int)(2*indent + 1), "", SvPVX_const(sv));
22859 if (OP(node) != OPTIMIZED) {
22860 if (next == NULL) /* Next ptr. */
22861 Perl_re_printf( aTHX_ " (0)");
22862 else if (PL_regkind[(U8)op] == BRANCH
22863 && PL_regkind[OP(next)] != BRANCH )
22864 Perl_re_printf( aTHX_ " (FAIL)");
22866 Perl_re_printf( aTHX_ " (%" IVdf ")", (IV)(next - start));
22867 Perl_re_printf( aTHX_ "\n");
22871 if (PL_regkind[(U8)op] == BRANCHJ) {
22874 const regnode *nnode = (OP(next) == LONGJMP
22875 ? regnext((regnode *)next)
22877 if (last && nnode > last)
22879 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
22882 else if (PL_regkind[(U8)op] == BRANCH) {
22884 DUMPUNTIL(NEXTOPER(node), next);
22886 else if ( PL_regkind[(U8)op] == TRIE ) {
22887 const regnode *this_trie = node;
22888 const char op = OP(node);
22889 const U32 n = ARG(node);
22890 const reg_ac_data * const ac = op>=AHOCORASICK ?
22891 (reg_ac_data *)ri->data->data[n] :
22893 const reg_trie_data * const trie =
22894 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
22896 AV *const trie_words
22897 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
22899 const regnode *nextbranch= NULL;
22902 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
22903 SV ** const elem_ptr = av_fetch(trie_words, word_idx, 0);
22905 Perl_re_indentf( aTHX_ "%s ",
22908 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
22909 SvCUR(*elem_ptr), PL_dump_re_max_len,
22910 PL_colors[0], PL_colors[1],
22912 ? PERL_PV_ESCAPE_UNI
22914 | PERL_PV_PRETTY_ELLIPSES
22915 | PERL_PV_PRETTY_LTGT
22920 U16 dist= trie->jump[word_idx+1];
22921 Perl_re_printf( aTHX_ "(%" UVuf ")\n",
22922 (UV)((dist ? this_trie + dist : next) - start));
22925 nextbranch= this_trie + trie->jump[0];
22926 DUMPUNTIL(this_trie + dist, nextbranch);
22928 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
22929 nextbranch= regnext((regnode *)nextbranch);
22931 Perl_re_printf( aTHX_ "\n");
22934 if (last && next > last)
22939 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
22940 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
22941 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
22943 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
22945 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
22947 else if ( op == PLUS || op == STAR) {
22948 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
22950 else if (PL_regkind[(U8)op] == EXACT || op == ANYOFHs) {
22951 /* Literal string, where present. */
22952 node += NODE_SZ_STR(node) - 1;
22953 node = NEXTOPER(node);
22956 node = NEXTOPER(node);
22957 node += regarglen[(U8)op];
22959 if (op == CURLYX || op == OPEN || op == SROPEN)
22963 #ifdef DEBUG_DUMPUNTIL
22964 Perl_re_printf( aTHX_ "--- %d\n", (int)indent);
22969 #endif /* DEBUGGING */
22971 #ifndef PERL_IN_XSUB_RE
22973 # include "uni_keywords.h"
22976 Perl_init_uniprops(pTHX)
22980 char * dump_len_string;
22982 dump_len_string = PerlEnv_getenv("PERL_DUMP_RE_MAX_LEN");
22983 if ( ! dump_len_string
22984 || ! grok_atoUV(dump_len_string, (UV *)&PL_dump_re_max_len, NULL))
22986 PL_dump_re_max_len = 60; /* A reasonable default */
22990 PL_user_def_props = newHV();
22992 # ifdef USE_ITHREADS
22994 HvSHAREKEYS_off(PL_user_def_props);
22995 PL_user_def_props_aTHX = aTHX;
22999 /* Set up the inversion list interpreter-level variables */
23001 PL_XPosix_ptrs[_CC_ASCII] = _new_invlist_C_array(uni_prop_ptrs[UNI_ASCII]);
23002 PL_XPosix_ptrs[_CC_ALPHANUMERIC] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXALNUM]);
23003 PL_XPosix_ptrs[_CC_ALPHA] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXALPHA]);
23004 PL_XPosix_ptrs[_CC_BLANK] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXBLANK]);
23005 PL_XPosix_ptrs[_CC_CASED] = _new_invlist_C_array(uni_prop_ptrs[UNI_CASED]);
23006 PL_XPosix_ptrs[_CC_CNTRL] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXCNTRL]);
23007 PL_XPosix_ptrs[_CC_DIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXDIGIT]);
23008 PL_XPosix_ptrs[_CC_GRAPH] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXGRAPH]);
23009 PL_XPosix_ptrs[_CC_LOWER] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXLOWER]);
23010 PL_XPosix_ptrs[_CC_PRINT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXPRINT]);
23011 PL_XPosix_ptrs[_CC_PUNCT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXPUNCT]);
23012 PL_XPosix_ptrs[_CC_SPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXSPACE]);
23013 PL_XPosix_ptrs[_CC_UPPER] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXUPPER]);
23014 PL_XPosix_ptrs[_CC_VERTSPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_VERTSPACE]);
23015 PL_XPosix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXWORD]);
23016 PL_XPosix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXXDIGIT]);
23018 PL_Posix_ptrs[_CC_ASCII] = _new_invlist_C_array(uni_prop_ptrs[UNI_ASCII]);
23019 PL_Posix_ptrs[_CC_ALPHANUMERIC] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXALNUM]);
23020 PL_Posix_ptrs[_CC_ALPHA] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXALPHA]);
23021 PL_Posix_ptrs[_CC_BLANK] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXBLANK]);
23022 PL_Posix_ptrs[_CC_CASED] = PL_Posix_ptrs[_CC_ALPHA];
23023 PL_Posix_ptrs[_CC_CNTRL] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXCNTRL]);
23024 PL_Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXDIGIT]);
23025 PL_Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXGRAPH]);
23026 PL_Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXLOWER]);
23027 PL_Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXPRINT]);
23028 PL_Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXPUNCT]);
23029 PL_Posix_ptrs[_CC_SPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXSPACE]);
23030 PL_Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXUPPER]);
23031 PL_Posix_ptrs[_CC_VERTSPACE] = NULL;
23032 PL_Posix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXWORD]);
23033 PL_Posix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXXDIGIT]);
23035 PL_GCB_invlist = _new_invlist_C_array(_Perl_GCB_invlist);
23036 PL_SB_invlist = _new_invlist_C_array(_Perl_SB_invlist);
23037 PL_WB_invlist = _new_invlist_C_array(_Perl_WB_invlist);
23038 PL_LB_invlist = _new_invlist_C_array(_Perl_LB_invlist);
23039 PL_SCX_invlist = _new_invlist_C_array(_Perl_SCX_invlist);
23041 PL_InBitmap = _new_invlist_C_array(InBitmap_invlist);
23042 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
23043 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
23044 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
23046 PL_Assigned_invlist = _new_invlist_C_array(uni_prop_ptrs[UNI_ASSIGNED]);
23048 PL_utf8_perl_idstart = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_IDSTART]);
23049 PL_utf8_perl_idcont = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_IDCONT]);
23051 PL_utf8_charname_begin = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_CHARNAME_BEGIN]);
23052 PL_utf8_charname_continue = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_CHARNAME_CONTINUE]);
23054 PL_in_some_fold = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_ANY_FOLDS]);
23055 PL_HasMultiCharFold = _new_invlist_C_array(uni_prop_ptrs[
23056 UNI__PERL_FOLDS_TO_MULTI_CHAR]);
23057 PL_InMultiCharFold = _new_invlist_C_array(uni_prop_ptrs[
23058 UNI__PERL_IS_IN_MULTI_CHAR_FOLD]);
23059 PL_utf8_toupper = _new_invlist_C_array(Uppercase_Mapping_invlist);
23060 PL_utf8_tolower = _new_invlist_C_array(Lowercase_Mapping_invlist);
23061 PL_utf8_totitle = _new_invlist_C_array(Titlecase_Mapping_invlist);
23062 PL_utf8_tofold = _new_invlist_C_array(Case_Folding_invlist);
23063 PL_utf8_tosimplefold = _new_invlist_C_array(Simple_Case_Folding_invlist);
23064 PL_utf8_foldclosures = _new_invlist_C_array(_Perl_IVCF_invlist);
23065 PL_utf8_mark = _new_invlist_C_array(uni_prop_ptrs[UNI_M]);
23066 PL_CCC_non0_non230 = _new_invlist_C_array(_Perl_CCC_non0_non230_invlist);
23067 PL_Private_Use = _new_invlist_C_array(uni_prop_ptrs[UNI_CO]);
23070 /* The below are used only by deprecated functions. They could be removed */
23071 PL_utf8_xidcont = _new_invlist_C_array(uni_prop_ptrs[UNI_XIDC]);
23072 PL_utf8_idcont = _new_invlist_C_array(uni_prop_ptrs[UNI_IDC]);
23073 PL_utf8_xidstart = _new_invlist_C_array(uni_prop_ptrs[UNI_XIDS]);
23077 /* These four functions are compiled only in regcomp.c, where they have access
23078 * to the data they return. They are a way for re_comp.c to get access to that
23079 * data without having to compile the whole data structures. */
23082 Perl_do_uniprop_match(const char * const key, const U16 key_len)
23084 PERL_ARGS_ASSERT_DO_UNIPROP_MATCH;
23086 return match_uniprop((U8 *) key, key_len);
23090 Perl_get_prop_definition(pTHX_ const int table_index)
23092 PERL_ARGS_ASSERT_GET_PROP_DEFINITION;
23094 /* Create and return the inversion list */
23095 return _new_invlist_C_array(uni_prop_ptrs[table_index]);
23098 const char * const *
23099 Perl_get_prop_values(const int table_index)
23101 PERL_ARGS_ASSERT_GET_PROP_VALUES;
23103 return UNI_prop_value_ptrs[table_index];
23107 Perl_get_deprecated_property_msg(const Size_t warning_offset)
23109 PERL_ARGS_ASSERT_GET_DEPRECATED_PROPERTY_MSG;
23111 return deprecated_property_msgs[warning_offset];
23116 This code was mainly added for backcompat to give a warning for non-portable
23117 code points in user-defined properties. But experiments showed that the
23118 warning in earlier perls were only omitted on overflow, which should be an
23119 error, so there really isnt a backcompat issue, and actually adding the
23120 warning when none was present before might cause breakage, for little gain. So
23121 khw left this code in, but not enabled. Tests were never added.
23124 Ei |const char *|get_extended_utf8_msg|const UV cp
23126 PERL_STATIC_INLINE const char *
23127 S_get_extended_utf8_msg(pTHX_ const UV cp)
23129 U8 dummy[UTF8_MAXBYTES + 1];
23133 uvchr_to_utf8_flags_msgs(dummy, cp, UNICODE_WARN_PERL_EXTENDED,
23136 msg = hv_fetchs(msgs, "text", 0);
23139 (void) sv_2mortal((SV *) msgs);
23141 return SvPVX(*msg);
23145 #endif /* end of ! PERL_IN_XSUB_RE */
23148 S_compile_wildcard(pTHX_ const char * subpattern, const STRLEN len,
23149 const bool ignore_case)
23151 /* Pretends that the input subpattern is qr/subpattern/aam, compiling it
23152 * possibly with /i if the 'ignore_case' parameter is true. Use /aa
23153 * because nothing outside of ASCII will match. Use /m because the input
23154 * string may be a bunch of lines strung together.
23156 * Also sets up the debugging info */
23158 U32 flags = PMf_MULTILINE|PMf_WILDCARD;
23160 SV * subpattern_sv = sv_2mortal(newSVpvn(subpattern, len));
23161 REGEXP * subpattern_re;
23162 DECLARE_AND_GET_RE_DEBUG_FLAGS;
23164 PERL_ARGS_ASSERT_COMPILE_WILDCARD;
23169 set_regex_charset(&flags, REGEX_ASCII_MORE_RESTRICTED_CHARSET);
23171 /* Like in op.c, we copy the compile time pm flags to the rx ones */
23172 rx_flags = flags & RXf_PMf_COMPILETIME;
23174 #ifndef PERL_IN_XSUB_RE
23175 /* Use the core engine if this file is regcomp.c. That means no
23176 * 'use re "Debug ..." is in effect, so the core engine is sufficient */
23177 subpattern_re = Perl_re_op_compile(aTHX_ &subpattern_sv, 1, NULL,
23178 &PL_core_reg_engine,
23182 if (isDEBUG_WILDCARD) {
23183 /* Use the special debugging engine if this file is re_comp.c and wants
23184 * to output the wildcard matching. This uses whatever
23185 * 'use re "Debug ..." is in effect */
23186 subpattern_re = Perl_re_op_compile(aTHX_ &subpattern_sv, 1, NULL,
23192 /* Use the special wildcard engine if this file is re_comp.c and
23193 * doesn't want to output the wildcard matching. This uses whatever
23194 * 'use re "Debug ..." is in effect for compilation, but this engine
23195 * structure has been set up so that it uses the core engine for
23196 * execution, so no execution debugging as a result of re.pm will be
23198 subpattern_re = Perl_re_op_compile(aTHX_ &subpattern_sv, 1, NULL,
23202 /* XXX The above has the effect that any user-supplied regex engine
23203 * won't be called for matching wildcards. That might be good, or bad.
23204 * It could be changed in several ways. The reason it is done the
23205 * current way is to avoid having to save and restore
23206 * ^{^RE_DEBUG_FLAGS} around the execution. save_scalar() perhaps
23207 * could be used. Another suggestion is to keep the authoritative
23208 * value of the debug flags in a thread-local variable and add set/get
23209 * magic to ${^RE_DEBUG_FLAGS} to keep the C level variable up to date.
23210 * Still another is to pass a flag, say in the engine's intflags that
23211 * would be checked each time before doing the debug output */
23215 assert(subpattern_re); /* Should have died if didn't compile successfully */
23216 return subpattern_re;
23220 S_execute_wildcard(pTHX_ REGEXP * const prog, char* stringarg, char *strend,
23221 char *strbeg, SSize_t minend, SV *screamer, U32 nosave)
23224 DECLARE_AND_GET_RE_DEBUG_FLAGS;
23226 PERL_ARGS_ASSERT_EXECUTE_WILDCARD;
23230 /* The compilation has set things up so that if the program doesn't want to
23231 * see the wildcard matching procedure, it will get the core execution
23232 * engine, which is subject only to -Dr. So we have to turn that off
23233 * around this procedure */
23234 if (! isDEBUG_WILDCARD) {
23235 /* Note! Casts away 'volatile' */
23237 PL_debug &= ~ DEBUG_r_FLAG;
23240 result = CALLREGEXEC(prog, stringarg, strend, strbeg, minend, screamer,
23248 S_handle_user_defined_property(pTHX_
23250 /* Parses the contents of a user-defined property definition; returning the
23251 * expanded definition if possible. If so, the return is an inversion
23254 * If there are subroutines that are part of the expansion and which aren't
23255 * known at the time of the call to this function, this returns what
23256 * parse_uniprop_string() returned for the first one encountered.
23258 * If an error was found, NULL is returned, and 'msg' gets a suitable
23259 * message appended to it. (Appending allows the back trace of how we got
23260 * to the faulty definition to be displayed through nested calls of
23261 * user-defined subs.)
23263 * The caller IS responsible for freeing any returned SV.
23265 * The syntax of the contents is pretty much described in perlunicode.pod,
23266 * but we also allow comments on each line */
23268 const char * name, /* Name of property */
23269 const STRLEN name_len, /* The name's length in bytes */
23270 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
23271 const bool to_fold, /* ? Is this under /i */
23272 const bool runtime, /* ? Are we in compile- or run-time */
23273 const bool deferrable, /* Is it ok for this property's full definition
23274 to be deferred until later? */
23275 SV* contents, /* The property's definition */
23276 bool *user_defined_ptr, /* This will be set TRUE as we wouldn't be
23277 getting called unless this is thought to be
23278 a user-defined property */
23279 SV * msg, /* Any error or warning msg(s) are appended to
23281 const STRLEN level) /* Recursion level of this call */
23284 const char * string = SvPV_const(contents, len);
23285 const char * const e = string + len;
23286 const bool is_contents_utf8 = cBOOL(SvUTF8(contents));
23287 const STRLEN msgs_length_on_entry = SvCUR(msg);
23289 const char * s0 = string; /* Points to first byte in the current line
23290 being parsed in 'string' */
23291 const char overflow_msg[] = "Code point too large in \"";
23292 SV* running_definition = NULL;
23294 PERL_ARGS_ASSERT_HANDLE_USER_DEFINED_PROPERTY;
23296 *user_defined_ptr = TRUE;
23298 /* Look at each line */
23300 const char * s; /* Current byte */
23301 char op = '+'; /* Default operation is 'union' */
23302 IV min = 0; /* range begin code point */
23303 IV max = -1; /* and range end */
23304 SV* this_definition;
23306 /* Skip comment lines */
23308 s0 = strchr(s0, '\n');
23316 /* For backcompat, allow an empty first line */
23322 /* First character in the line may optionally be the operation */
23331 /* If the line is one or two hex digits separated by blank space, its
23332 * a range; otherwise it is either another user-defined property or an
23337 if (! isXDIGIT(*s)) {
23338 goto check_if_property;
23341 do { /* Each new hex digit will add 4 bits. */
23342 if (min > ( (IV) MAX_LEGAL_CP >> 4)) {
23343 s = strchr(s, '\n');
23347 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23348 sv_catpv(msg, overflow_msg);
23349 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
23350 UTF8fARG(is_contents_utf8, s - s0, s0));
23351 sv_catpvs(msg, "\"");
23352 goto return_failure;
23355 /* Accumulate this digit into the value */
23356 min = (min << 4) + READ_XDIGIT(s);
23357 } while (isXDIGIT(*s));
23359 while (isBLANK(*s)) { s++; }
23361 /* We allow comments at the end of the line */
23363 s = strchr(s, '\n');
23369 else if (s < e && *s != '\n') {
23370 if (! isXDIGIT(*s)) {
23371 goto check_if_property;
23374 /* Look for the high point of the range */
23377 if (max > ( (IV) MAX_LEGAL_CP >> 4)) {
23378 s = strchr(s, '\n');
23382 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23383 sv_catpv(msg, overflow_msg);
23384 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
23385 UTF8fARG(is_contents_utf8, s - s0, s0));
23386 sv_catpvs(msg, "\"");
23387 goto return_failure;
23390 max = (max << 4) + READ_XDIGIT(s);
23391 } while (isXDIGIT(*s));
23393 while (isBLANK(*s)) { s++; }
23396 s = strchr(s, '\n');
23401 else if (s < e && *s != '\n') {
23402 goto check_if_property;
23406 if (max == -1) { /* The line only had one entry */
23409 else if (max < min) {
23410 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23411 sv_catpvs(msg, "Illegal range in \"");
23412 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
23413 UTF8fARG(is_contents_utf8, s - s0, s0));
23414 sv_catpvs(msg, "\"");
23415 goto return_failure;
23418 # if 0 /* See explanation at definition above of get_extended_utf8_msg() */
23420 if ( UNICODE_IS_PERL_EXTENDED(min)
23421 || UNICODE_IS_PERL_EXTENDED(max))
23423 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23425 /* If both code points are non-portable, warn only on the lower
23427 sv_catpv(msg, get_extended_utf8_msg(
23428 (UNICODE_IS_PERL_EXTENDED(min))
23430 sv_catpvs(msg, " in \"");
23431 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
23432 UTF8fARG(is_contents_utf8, s - s0, s0));
23433 sv_catpvs(msg, "\"");
23438 /* Here, this line contains a legal range */
23439 this_definition = sv_2mortal(_new_invlist(2));
23440 this_definition = _add_range_to_invlist(this_definition, min, max);
23445 /* Here it isn't a legal range line. See if it is a legal property
23446 * line. First find the end of the meat of the line */
23447 s = strpbrk(s, "#\n");
23452 /* Ignore trailing blanks in keeping with the requirements of
23453 * parse_uniprop_string() */
23455 while (s > s0 && isBLANK_A(*s)) {
23460 this_definition = parse_uniprop_string(s0, s - s0,
23461 is_utf8, to_fold, runtime,
23464 user_defined_ptr, msg,
23466 ? level /* Don't increase level
23467 if input is empty */
23470 if (this_definition == NULL) {
23471 goto return_failure; /* 'msg' should have had the reason
23472 appended to it by the above call */
23475 if (! is_invlist(this_definition)) { /* Unknown at this time */
23476 return newSVsv(this_definition);
23480 s = strchr(s, '\n');
23490 _invlist_union(running_definition, this_definition,
23491 &running_definition);
23494 _invlist_subtract(running_definition, this_definition,
23495 &running_definition);
23498 _invlist_intersection(running_definition, this_definition,
23499 &running_definition);
23502 _invlist_union_complement_2nd(running_definition,
23503 this_definition, &running_definition);
23506 Perl_croak(aTHX_ "panic: %s: %d: Unexpected operation %d",
23507 __FILE__, __LINE__, op);
23511 /* Position past the '\n' */
23513 } /* End of loop through the lines of 'contents' */
23515 /* Here, we processed all the lines in 'contents' without error. If we
23516 * didn't add any warnings, simply return success */
23517 if (msgs_length_on_entry == SvCUR(msg)) {
23519 /* If the expansion was empty, the answer isn't nothing: its an empty
23520 * inversion list */
23521 if (running_definition == NULL) {
23522 running_definition = _new_invlist(1);
23525 return running_definition;
23528 /* Otherwise, add some explanatory text, but we will return success */
23532 running_definition = NULL;
23536 if (name_len > 0) {
23537 sv_catpvs(msg, " in expansion of ");
23538 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8, name_len, name));
23541 return running_definition;
23544 /* As explained below, certain operations need to take place in the first
23545 * thread created. These macros switch contexts */
23546 # ifdef USE_ITHREADS
23547 # define DECLARATION_FOR_GLOBAL_CONTEXT \
23548 PerlInterpreter * save_aTHX = aTHX;
23549 # define SWITCH_TO_GLOBAL_CONTEXT \
23550 PERL_SET_CONTEXT((aTHX = PL_user_def_props_aTHX))
23551 # define RESTORE_CONTEXT PERL_SET_CONTEXT((aTHX = save_aTHX));
23552 # define CUR_CONTEXT aTHX
23553 # define ORIGINAL_CONTEXT save_aTHX
23555 # define DECLARATION_FOR_GLOBAL_CONTEXT dNOOP
23556 # define SWITCH_TO_GLOBAL_CONTEXT NOOP
23557 # define RESTORE_CONTEXT NOOP
23558 # define CUR_CONTEXT NULL
23559 # define ORIGINAL_CONTEXT NULL
23563 S_delete_recursion_entry(pTHX_ void *key)
23565 /* Deletes the entry used to detect recursion when expanding user-defined
23566 * properties. This is a function so it can be set up to be called even if
23567 * the program unexpectedly quits */
23569 SV ** current_entry;
23570 const STRLEN key_len = strlen((const char *) key);
23571 DECLARATION_FOR_GLOBAL_CONTEXT;
23573 SWITCH_TO_GLOBAL_CONTEXT;
23575 /* If the entry is one of these types, it is a permanent entry, and not the
23576 * one used to detect recursions. This function should delete only the
23577 * recursion entry */
23578 current_entry = hv_fetch(PL_user_def_props, (const char *) key, key_len, 0);
23580 && ! is_invlist(*current_entry)
23581 && ! SvPOK(*current_entry))
23583 (void) hv_delete(PL_user_def_props, (const char *) key, key_len,
23591 S_get_fq_name(pTHX_
23592 const char * const name, /* The first non-blank in the \p{}, \P{} */
23593 const Size_t name_len, /* Its length in bytes, not including any trailing space */
23594 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
23595 const bool has_colon_colon
23598 /* Returns a mortal SV containing the fully qualified version of the input
23603 fq_name = newSVpvs_flags("", SVs_TEMP);
23605 /* Use the current package if it wasn't included in our input */
23606 if (! has_colon_colon) {
23607 const HV * pkg = (IN_PERL_COMPILETIME)
23609 : CopSTASH(PL_curcop);
23610 const char* pkgname = HvNAME(pkg);
23612 Perl_sv_catpvf(aTHX_ fq_name, "%" UTF8f,
23613 UTF8fARG(is_utf8, strlen(pkgname), pkgname));
23614 sv_catpvs(fq_name, "::");
23617 Perl_sv_catpvf(aTHX_ fq_name, "%" UTF8f,
23618 UTF8fARG(is_utf8, name_len, name));
23623 S_parse_uniprop_string(pTHX_
23625 /* Parse the interior of a \p{}, \P{}. Returns its definition if knowable
23626 * now. If so, the return is an inversion list.
23628 * If the property is user-defined, it is a subroutine, which in turn
23629 * may call other subroutines. This function will call the whole nest of
23630 * them to get the definition they return; if some aren't known at the time
23631 * of the call to this function, the fully qualified name of the highest
23632 * level sub is returned. It is an error to call this function at runtime
23633 * without every sub defined.
23635 * If an error was found, NULL is returned, and 'msg' gets a suitable
23636 * message appended to it. (Appending allows the back trace of how we got
23637 * to the faulty definition to be displayed through nested calls of
23638 * user-defined subs.)
23640 * The caller should NOT try to free any returned inversion list.
23642 * Other parameters will be set on return as described below */
23644 const char * const name, /* The first non-blank in the \p{}, \P{} */
23645 Size_t name_len, /* Its length in bytes, not including any
23647 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
23648 const bool to_fold, /* ? Is this under /i */
23649 const bool runtime, /* TRUE if this is being called at run time */
23650 const bool deferrable, /* TRUE if it's ok for the definition to not be
23651 known at this call */
23652 AV ** strings, /* To return string property values, like named
23654 bool *user_defined_ptr, /* Upon return from this function it will be
23655 set to TRUE if any component is a
23656 user-defined property */
23657 SV * msg, /* Any error or warning msg(s) are appended to
23659 const STRLEN level) /* Recursion level of this call */
23661 char* lookup_name; /* normalized name for lookup in our tables */
23662 unsigned lookup_len; /* Its length */
23663 enum { Not_Strict = 0, /* Some properties have stricter name */
23664 Strict, /* normalization rules, which we decide */
23665 As_Is /* upon based on parsing */
23666 } stricter = Not_Strict;
23668 /* nv= or numeric_value=, or possibly one of the cjk numeric properties
23669 * (though it requires extra effort to download them from Unicode and
23670 * compile perl to know about them) */
23671 bool is_nv_type = FALSE;
23673 unsigned int i, j = 0;
23674 int equals_pos = -1; /* Where the '=' is found, or negative if none */
23675 int slash_pos = -1; /* Where the '/' is found, or negative if none */
23676 int table_index = 0; /* The entry number for this property in the table
23677 of all Unicode property names */
23678 bool starts_with_Is = FALSE; /* ? Does the name start with 'Is' */
23679 Size_t lookup_offset = 0; /* Used to ignore the first few characters of
23680 the normalized name in certain situations */
23681 Size_t non_pkg_begin = 0; /* Offset of first byte in 'name' that isn't
23682 part of a package name */
23683 Size_t lun_non_pkg_begin = 0; /* Similarly for 'lookup_name' */
23684 bool could_be_user_defined = TRUE; /* ? Could this be a user-defined
23685 property rather than a Unicode
23687 SV * prop_definition = NULL; /* The returned definition of 'name' or NULL
23688 if an error. If it is an inversion list,
23689 it is the definition. Otherwise it is a
23690 string containing the fully qualified sub
23692 SV * fq_name = NULL; /* For user-defined properties, the fully
23694 bool invert_return = FALSE; /* ? Do we need to complement the result before
23696 bool stripped_utf8_pkg = FALSE; /* Set TRUE if the input includes an
23697 explicit utf8:: package that we strip
23699 /* The expansion of properties that could be either user-defined or
23700 * official unicode ones is deferred until runtime, including a marker for
23701 * those that might be in the latter category. This boolean indicates if
23702 * we've seen that marker. If not, what we're parsing can't be such an
23703 * official Unicode property whose expansion was deferred */
23704 bool could_be_deferred_official = FALSE;
23706 PERL_ARGS_ASSERT_PARSE_UNIPROP_STRING;
23708 /* The input will be normalized into 'lookup_name' */
23709 Newx(lookup_name, name_len, char);
23710 SAVEFREEPV(lookup_name);
23712 /* Parse the input. */
23713 for (i = 0; i < name_len; i++) {
23714 char cur = name[i];
23716 /* Most of the characters in the input will be of this ilk, being parts
23718 if (isIDCONT_A(cur)) {
23720 /* Case differences are ignored. Our lookup routine assumes
23721 * everything is lowercase, so normalize to that */
23722 if (isUPPER_A(cur)) {
23723 lookup_name[j++] = toLOWER_A(cur);
23727 if (cur == '_') { /* Don't include these in the normalized name */
23731 lookup_name[j++] = cur;
23733 /* The first character in a user-defined name must be of this type.
23735 if (i - non_pkg_begin == 0 && ! isIDFIRST_A(cur)) {
23736 could_be_user_defined = FALSE;
23742 /* Here, the character is not something typically in a name, But these
23743 * two types of characters (and the '_' above) can be freely ignored in
23744 * most situations. Later it may turn out we shouldn't have ignored
23745 * them, and we have to reparse, but we don't have enough information
23746 * yet to make that decision */
23747 if (cur == '-' || isSPACE_A(cur)) {
23748 could_be_user_defined = FALSE;
23752 /* An equals sign or single colon mark the end of the first part of
23753 * the property name */
23755 || (cur == ':' && (i >= name_len - 1 || name[i+1] != ':')))
23757 lookup_name[j++] = '='; /* Treat the colon as an '=' */
23758 equals_pos = j; /* Note where it occurred in the input */
23759 could_be_user_defined = FALSE;
23763 /* If this looks like it is a marker we inserted at compile time,
23764 * set a flag and otherwise ignore it. If it isn't in the final
23765 * position, keep it as it would have been user input. */
23766 if ( UNLIKELY(cur == DEFERRED_COULD_BE_OFFICIAL_MARKERc)
23768 && could_be_user_defined
23769 && i == name_len - 1)
23772 could_be_deferred_official = TRUE;
23776 /* Otherwise, this character is part of the name. */
23777 lookup_name[j++] = cur;
23779 /* Here it isn't a single colon, so if it is a colon, it must be a
23783 /* A double colon should be a package qualifier. We note its
23784 * position and continue. Note that one could have
23785 * pkg1::pkg2::...::foo
23786 * so that the position at the end of the loop will be just after
23787 * the final qualifier */
23790 non_pkg_begin = i + 1;
23791 lookup_name[j++] = ':';
23792 lun_non_pkg_begin = j;
23794 else { /* Only word chars (and '::') can be in a user-defined name */
23795 could_be_user_defined = FALSE;
23797 } /* End of parsing through the lhs of the property name (or all of it if
23800 # define STRLENs(s) (sizeof("" s "") - 1)
23802 /* If there is a single package name 'utf8::', it is ambiguous. It could
23803 * be for a user-defined property, or it could be a Unicode property, as
23804 * all of them are considered to be for that package. For the purposes of
23805 * parsing the rest of the property, strip it off */
23806 if (non_pkg_begin == STRLENs("utf8::") && memBEGINPs(name, name_len, "utf8::")) {
23807 lookup_name += STRLENs("utf8::");
23808 j -= STRLENs("utf8::");
23809 equals_pos -= STRLENs("utf8::");
23810 stripped_utf8_pkg = TRUE;
23813 /* Here, we are either done with the whole property name, if it was simple;
23814 * or are positioned just after the '=' if it is compound. */
23816 if (equals_pos >= 0) {
23817 assert(stricter == Not_Strict); /* We shouldn't have set this yet */
23819 /* Space immediately after the '=' is ignored */
23821 for (; i < name_len; i++) {
23822 if (! isSPACE_A(name[i])) {
23827 /* Most punctuation after the equals indicates a subpattern, like
23829 if ( isPUNCT_A(name[i])
23834 /* A backslash means the real delimitter is the next character,
23835 * but it must be punctuation */
23836 && (name[i] != '\\' || (i < name_len && isPUNCT_A(name[i+1]))))
23838 bool special_property = memEQs(lookup_name, j - 1, "name")
23839 || memEQs(lookup_name, j - 1, "na");
23840 if (! special_property) {
23841 /* Find the property. The table includes the equals sign, so
23842 * we use 'j' as-is */
23843 table_index = do_uniprop_match(lookup_name, j);
23845 if (special_property || table_index) {
23846 REGEXP * subpattern_re;
23847 char open = name[i++];
23849 const char * pos_in_brackets;
23850 const char * const * prop_values;
23853 /* Backslash => delimitter is the character following. We
23854 * already checked that it is punctuation */
23855 if (open == '\\') {
23860 /* This data structure is constructed so that the matching
23861 * closing bracket is 3 past its matching opening. The second
23862 * set of closing is so that if the opening is something like
23863 * ']', the closing will be that as well. Something similar is
23864 * done in toke.c */
23865 pos_in_brackets = memCHRs("([<)]>)]>", open);
23866 close = (pos_in_brackets) ? pos_in_brackets[3] : open;
23869 || name[name_len-1] != close
23870 || (escaped && name[name_len-2] != '\\')
23871 /* Also make sure that there are enough characters.
23872 * e.g., '\\\' would show up incorrectly as legal even
23873 * though it is too short */
23874 || (SSize_t) (name_len - i - 1 - escaped) < 0)
23876 sv_catpvs(msg, "Unicode property wildcard not terminated");
23877 goto append_name_to_msg;
23880 Perl_ck_warner_d(aTHX_
23881 packWARN(WARN_EXPERIMENTAL__UNIPROP_WILDCARDS),
23882 "The Unicode property wildcards feature is experimental");
23884 if (special_property) {
23885 const char * error_msg;
23886 const char * revised_name = name + i;
23887 Size_t revised_name_len = name_len - (i + 1 + escaped);
23889 /* Currently, the only 'special_property' is name, which we
23890 * lookup in _charnames.pm */
23892 if (! load_charnames(newSVpvs("placeholder"),
23893 revised_name, revised_name_len,
23896 sv_catpv(msg, error_msg);
23897 goto append_name_to_msg;
23900 /* Farm this out to a function just to make the current
23901 * function less unwieldy */
23902 if (handle_names_wildcard(revised_name, revised_name_len,
23906 return prop_definition;
23912 prop_values = get_prop_values(table_index);
23914 /* Now create and compile the wildcard subpattern. Use /i
23915 * because the property values are supposed to match with case
23917 subpattern_re = compile_wildcard(name + i,
23918 name_len - i - 1 - escaped,
23922 /* For each legal property value, see if the supplied pattern
23924 while (*prop_values) {
23925 const char * const entry = *prop_values;
23926 const Size_t len = strlen(entry);
23927 SV* entry_sv = newSVpvn_flags(entry, len, SVs_TEMP);
23929 if (execute_wildcard(subpattern_re,
23931 (char *) entry + len,
23935 { /* Here, matched. Add to the returned list */
23936 Size_t total_len = j + len;
23937 SV * sub_invlist = NULL;
23938 char * this_string;
23940 /* We know this is a legal \p{property=value}. Call
23941 * the function to return the list of code points that
23943 Newxz(this_string, total_len + 1, char);
23944 Copy(lookup_name, this_string, j, char);
23945 my_strlcat(this_string, entry, total_len + 1);
23946 SAVEFREEPV(this_string);
23947 sub_invlist = parse_uniprop_string(this_string,
23957 _invlist_union(prop_definition, sub_invlist,
23961 prop_values++; /* Next iteration, look at next propvalue */
23962 } /* End of looking through property values; (the data
23963 structure is terminated by a NULL ptr) */
23965 SvREFCNT_dec_NN(subpattern_re);
23967 if (prop_definition) {
23968 return prop_definition;
23971 sv_catpvs(msg, "No Unicode property value wildcard matches:");
23972 goto append_name_to_msg;
23975 /* Here's how khw thinks we should proceed to handle the properties
23976 * not yet done: Bidi Mirroring Glyph can map to ""
23977 Bidi Paired Bracket can map to ""
23978 Case Folding (both full and simple)
23979 Shouldn't /i be good enough for Full
23980 Decomposition Mapping
23981 Equivalent Unified Ideograph can map to ""
23982 Lowercase Mapping (both full and simple)
23983 NFKC Case Fold can map to ""
23984 Titlecase Mapping (both full and simple)
23985 Uppercase Mapping (both full and simple)
23986 * Handle these the same way Name is done, using say, _wild.pm, but
23987 * having both loose and full, like in charclass_invlists.h.
23988 * Perhaps move block and script to that as they are somewhat large
23989 * in charclass_invlists.h.
23990 * For properties where the default is the code point itself, such
23991 * as any of the case changing mappings, the string would otherwise
23992 * consist of all Unicode code points in UTF-8 strung together.
23993 * This would be impractical. So instead, examine their compiled
23994 * pattern, looking at the ssc. If none, reject the pattern as an
23995 * error. Otherwise run the pattern against every code point in
23996 * the ssc. The ssc is kind of like tr18's 3.9 Possible Match Sets
23997 * And it might be good to create an API to return the ssc.
23998 * Or handle them like the algorithmic names are done
24000 } /* End of is a wildcard subppattern */
24002 /* \p{name=...} is handled specially. Instead of using the normal
24003 * mechanism involving charclass_invlists.h, it uses _charnames.pm
24004 * which has the necessary (huge) data accessible to it, and which
24005 * doesn't get loaded unless necessary. The legal syntax for names is
24006 * somewhat different than other properties due both to the vagaries of
24007 * a few outlier official names, and the fact that only a few ASCII
24008 * characters are permitted in them */
24009 if ( memEQs(lookup_name, j - 1, "name")
24010 || memEQs(lookup_name, j - 1, "na"))
24015 const char * error_msg;
24017 SV * character_name;
24018 STRLEN character_len;
24023 /* Since the RHS (after skipping initial space) is passed unchanged
24024 * to charnames, and there are different criteria for what are
24025 * legal characters in the name, just parse it here. A character
24026 * name must begin with an ASCII alphabetic */
24027 if (! isALPHA(name[i])) {
24030 lookup_name[j++] = name[i];
24032 for (++i; i < name_len; i++) {
24033 /* Official names can only be in the ASCII range, and only
24034 * certain characters */
24035 if (! isASCII(name[i]) || ! isCHARNAME_CONT(name[i])) {
24038 lookup_name[j++] = name[i];
24041 /* Finished parsing, save the name into an SV */
24042 character_name = newSVpvn(lookup_name + equals_pos, j - equals_pos);
24044 /* Make sure _charnames is loaded. (The parameters give context
24045 * for any errors generated */
24046 table = load_charnames(character_name, name, name_len, &error_msg);
24047 if (table == NULL) {
24048 sv_catpv(msg, error_msg);
24049 goto append_name_to_msg;
24052 lookup_loose = get_cvs("_charnames::_loose_regcomp_lookup", 0);
24053 if (! lookup_loose) {
24055 "panic: Can't find '_charnames::_loose_regcomp_lookup");
24058 PUSHSTACKi(PERLSI_REGCOMP);
24064 XPUSHs(character_name);
24066 call_sv(MUTABLE_SV(lookup_loose), G_SCALAR);
24071 SvREFCNT_inc_simple_void_NN(character);
24078 if (! SvOK(character)) {
24082 cp = valid_utf8_to_uvchr((U8 *) SvPVX(character), &character_len);
24083 if (character_len == SvCUR(character)) {
24084 prop_definition = add_cp_to_invlist(NULL, cp);
24089 /* First of the remaining characters in the string. */
24090 char * remaining = SvPVX(character) + character_len;
24092 if (strings == NULL) {
24093 goto failed; /* XXX Perhaps a specific msg instead, like
24094 'not available here' */
24097 if (*strings == NULL) {
24098 *strings = newAV();
24101 this_string = newAV();
24102 av_push(this_string, newSVuv(cp));
24105 cp = valid_utf8_to_uvchr((U8 *) remaining, &character_len);
24106 av_push(this_string, newSVuv(cp));
24107 remaining += character_len;
24108 } while (remaining < SvEND(character));
24110 av_push(*strings, (SV *) this_string);
24113 return prop_definition;
24116 /* Certain properties whose values are numeric need special handling.
24117 * They may optionally be prefixed by 'is'. Ignore that prefix for the
24118 * purposes of checking if this is one of those properties */
24119 if (memBEGINPs(lookup_name, j, "is")) {
24123 /* Then check if it is one of these specially-handled properties. The
24124 * possibilities are hard-coded because easier this way, and the list
24125 * is unlikely to change.
24127 * All numeric value type properties are of this ilk, and are also
24128 * special in a different way later on. So find those first. There
24129 * are several numeric value type properties in the Unihan DB (which is
24130 * unlikely to be compiled with perl, but we handle it here in case it
24131 * does get compiled). They all end with 'numeric'. The interiors
24132 * aren't checked for the precise property. This would stop working if
24133 * a cjk property were to be created that ended with 'numeric' and
24134 * wasn't a numeric type */
24135 is_nv_type = memEQs(lookup_name + lookup_offset,
24136 j - 1 - lookup_offset, "numericvalue")
24137 || memEQs(lookup_name + lookup_offset,
24138 j - 1 - lookup_offset, "nv")
24139 || ( memENDPs(lookup_name + lookup_offset,
24140 j - 1 - lookup_offset, "numeric")
24141 && ( memBEGINPs(lookup_name + lookup_offset,
24142 j - 1 - lookup_offset, "cjk")
24143 || memBEGINPs(lookup_name + lookup_offset,
24144 j - 1 - lookup_offset, "k")));
24146 || memEQs(lookup_name + lookup_offset,
24147 j - 1 - lookup_offset, "canonicalcombiningclass")
24148 || memEQs(lookup_name + lookup_offset,
24149 j - 1 - lookup_offset, "ccc")
24150 || memEQs(lookup_name + lookup_offset,
24151 j - 1 - lookup_offset, "age")
24152 || memEQs(lookup_name + lookup_offset,
24153 j - 1 - lookup_offset, "in")
24154 || memEQs(lookup_name + lookup_offset,
24155 j - 1 - lookup_offset, "presentin"))
24159 /* Since the stuff after the '=' is a number, we can't throw away
24160 * '-' willy-nilly, as those could be a minus sign. Other stricter
24161 * rules also apply. However, these properties all can have the
24162 * rhs not be a number, in which case they contain at least one
24163 * alphabetic. In those cases, the stricter rules don't apply.
24164 * But the numeric type properties can have the alphas [Ee] to
24165 * signify an exponent, and it is still a number with stricter
24166 * rules. So look for an alpha that signifies not-strict */
24168 for (k = i; k < name_len; k++) {
24169 if ( isALPHA_A(name[k])
24170 && (! is_nv_type || ! isALPHA_FOLD_EQ(name[k], 'E')))
24172 stricter = Not_Strict;
24180 /* A number may have a leading '+' or '-'. The latter is retained
24182 if (name[i] == '+') {
24185 else if (name[i] == '-') {
24186 lookup_name[j++] = '-';
24190 /* Skip leading zeros including single underscores separating the
24191 * zeros, or between the final leading zero and the first other
24193 for (; i < name_len - 1; i++) {
24194 if ( name[i] != '0'
24195 && (name[i] != '_' || ! isDIGIT_A(name[i+1])))
24202 else { /* No '=' */
24204 /* Only a few properties without an '=' should be parsed with stricter
24205 * rules. The list is unlikely to change. */
24206 if ( memBEGINPs(lookup_name, j, "perl")
24207 && memNEs(lookup_name + 4, j - 4, "space")
24208 && memNEs(lookup_name + 4, j - 4, "word"))
24212 /* We set the inputs back to 0 and the code below will reparse,
24218 /* Here, we have either finished the property, or are positioned to parse
24219 * the remainder, and we know if stricter rules apply. Finish out, if not
24221 for (; i < name_len; i++) {
24222 char cur = name[i];
24224 /* In all instances, case differences are ignored, and we normalize to
24226 if (isUPPER_A(cur)) {
24227 lookup_name[j++] = toLOWER(cur);
24231 /* An underscore is skipped, but not under strict rules unless it
24232 * separates two digits */
24235 && ( i == 0 || (int) i == equals_pos || i == name_len- 1
24236 || ! isDIGIT_A(name[i-1]) || ! isDIGIT_A(name[i+1])))
24238 lookup_name[j++] = '_';
24243 /* Hyphens are skipped except under strict */
24244 if (cur == '-' && ! stricter) {
24248 /* XXX Bug in documentation. It says white space skipped adjacent to
24249 * non-word char. Maybe we should, but shouldn't skip it next to a dot
24251 if (isSPACE_A(cur) && ! stricter) {
24255 lookup_name[j++] = cur;
24257 /* Unless this is a non-trailing slash, we are done with it */
24258 if (i >= name_len - 1 || cur != '/') {
24264 /* A slash in the 'numeric value' property indicates that what follows
24265 * is a denominator. It can have a leading '+' and '0's that should be
24266 * skipped. But we have never allowed a negative denominator, so treat
24267 * a minus like every other character. (No need to rule out a second
24268 * '/', as that won't match anything anyway */
24271 if (i < name_len && name[i] == '+') {
24275 /* Skip leading zeros including underscores separating digits */
24276 for (; i < name_len - 1; i++) {
24277 if ( name[i] != '0'
24278 && (name[i] != '_' || ! isDIGIT_A(name[i+1])))
24284 /* Store the first real character in the denominator */
24285 if (i < name_len) {
24286 lookup_name[j++] = name[i];
24291 /* Here are completely done parsing the input 'name', and 'lookup_name'
24292 * contains a copy, normalized.
24294 * This special case is grandfathered in: 'L_' and 'GC=L_' are accepted and
24295 * different from without the underscores. */
24296 if ( ( UNLIKELY(memEQs(lookup_name, j, "l"))
24297 || UNLIKELY(memEQs(lookup_name, j, "gc=l")))
24298 && UNLIKELY(name[name_len-1] == '_'))
24300 lookup_name[j++] = '&';
24303 /* If the original input began with 'In' or 'Is', it could be a subroutine
24304 * call to a user-defined property instead of a Unicode property name. */
24305 if ( name_len - non_pkg_begin > 2
24306 && name[non_pkg_begin+0] == 'I'
24307 && (name[non_pkg_begin+1] == 'n' || name[non_pkg_begin+1] == 's'))
24309 /* Names that start with In have different characterstics than those
24310 * that start with Is */
24311 if (name[non_pkg_begin+1] == 's') {
24312 starts_with_Is = TRUE;
24316 could_be_user_defined = FALSE;
24319 if (could_be_user_defined) {
24322 /* If the user defined property returns the empty string, it could
24323 * easily be because the pattern is being compiled before the data it
24324 * actually needs to compile is available. This could be argued to be
24325 * a bug in the perl code, but this is a change of behavior for Perl,
24326 * so we handle it. This means that intentionally returning nothing
24327 * will not be resolved until runtime */
24328 bool empty_return = FALSE;
24330 /* Here, the name could be for a user defined property, which are
24331 * implemented as subs. */
24332 user_sub = get_cvn_flags(name, name_len, 0);
24335 /* Here, the property name could be a user-defined one, but there
24336 * is no subroutine to handle it (as of now). Defer handling it
24337 * until runtime. Otherwise, a block defined by Unicode in a later
24338 * release would get the synonym InFoo added for it, and existing
24339 * code that used that name would suddenly break if it referred to
24340 * the property before the sub was declared. See [perl #134146] */
24342 goto definition_deferred;
24345 /* Here, we are at runtime, and didn't find the user property. It
24346 * could be an official property, but only if no package was
24347 * specified, or just the utf8:: package. */
24348 if (could_be_deferred_official) {
24349 lookup_name += lun_non_pkg_begin;
24350 j -= lun_non_pkg_begin;
24352 else if (! stripped_utf8_pkg) {
24353 goto unknown_user_defined;
24356 /* Drop down to look up in the official properties */
24359 const char insecure[] = "Insecure user-defined property";
24361 /* Here, there is a sub by the correct name. Normally we call it
24362 * to get the property definition */
24364 SV * user_sub_sv = MUTABLE_SV(user_sub);
24365 SV * error; /* Any error returned by calling 'user_sub' */
24366 SV * key; /* The key into the hash of user defined sub names
24369 SV ** saved_user_prop_ptr; /* Hash entry for this property */
24371 /* How many times to retry when another thread is in the middle of
24372 * expanding the same definition we want */
24373 PERL_INT_FAST8_T retry_countdown = 10;
24375 DECLARATION_FOR_GLOBAL_CONTEXT;
24377 /* If we get here, we know this property is user-defined */
24378 *user_defined_ptr = TRUE;
24380 /* We refuse to call a potentially tainted subroutine; returning an
24383 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24384 sv_catpvn(msg, insecure, sizeof(insecure) - 1);
24385 goto append_name_to_msg;
24388 /* In principal, we only call each subroutine property definition
24389 * once during the life of the program. This guarantees that the
24390 * property definition never changes. The results of the single
24391 * sub call are stored in a hash, which is used instead for future
24392 * references to this property. The property definition is thus
24393 * immutable. But, to allow the user to have a /i-dependent
24394 * definition, we call the sub once for non-/i, and once for /i,
24395 * should the need arise, passing the /i status as a parameter.
24397 * We start by constructing the hash key name, consisting of the
24398 * fully qualified subroutine name, preceded by the /i status, so
24399 * that there is a key for /i and a different key for non-/i */
24400 key = newSVpvn(((to_fold) ? "1" : "0"), 1);
24401 fq_name = S_get_fq_name(aTHX_ name, name_len, is_utf8,
24402 non_pkg_begin != 0);
24403 sv_catsv(key, fq_name);
24406 /* We only call the sub once throughout the life of the program
24407 * (with the /i, non-/i exception noted above). That means the
24408 * hash must be global and accessible to all threads. It is
24409 * created at program start-up, before any threads are created, so
24410 * is accessible to all children. But this creates some
24413 * 1) The keys can't be shared, or else problems arise; sharing is
24414 * turned off at hash creation time
24415 * 2) All SVs in it are there for the remainder of the life of the
24416 * program, and must be created in the same interpreter context
24417 * as the hash, or else they will be freed from the wrong pool
24418 * at global destruction time. This is handled by switching to
24419 * the hash's context to create each SV going into it, and then
24420 * immediately switching back
24421 * 3) All accesses to the hash must be controlled by a mutex, to
24422 * prevent two threads from getting an unstable state should
24423 * they simultaneously be accessing it. The code below is
24424 * crafted so that the mutex is locked whenever there is an
24425 * access and unlocked only when the next stable state is
24428 * The hash stores either the definition of the property if it was
24429 * valid, or, if invalid, the error message that was raised. We
24430 * use the type of SV to distinguish.
24432 * There's also the need to guard against the definition expansion
24433 * from infinitely recursing. This is handled by storing the aTHX
24434 * of the expanding thread during the expansion. Again the SV type
24435 * is used to distinguish this from the other two cases. If we
24436 * come to here and the hash entry for this property is our aTHX,
24437 * it means we have recursed, and the code assumes that we would
24438 * infinitely recurse, so instead stops and raises an error.
24439 * (Any recursion has always been treated as infinite recursion in
24442 * If instead, the entry is for a different aTHX, it means that
24443 * that thread has gotten here first, and hasn't finished expanding
24444 * the definition yet. We just have to wait until it is done. We
24445 * sleep and retry a few times, returning an error if the other
24446 * thread doesn't complete. */
24449 USER_PROP_MUTEX_LOCK;
24451 /* If we have an entry for this key, the subroutine has already
24452 * been called once with this /i status. */
24453 saved_user_prop_ptr = hv_fetch(PL_user_def_props,
24454 SvPVX(key), SvCUR(key), 0);
24455 if (saved_user_prop_ptr) {
24457 /* If the saved result is an inversion list, it is the valid
24458 * definition of this property */
24459 if (is_invlist(*saved_user_prop_ptr)) {
24460 prop_definition = *saved_user_prop_ptr;
24462 /* The SV in the hash won't be removed until global
24463 * destruction, so it is stable and we can unlock */
24464 USER_PROP_MUTEX_UNLOCK;
24466 /* The caller shouldn't try to free this SV */
24467 return prop_definition;
24470 /* Otherwise, if it is a string, it is the error message
24471 * that was returned when we first tried to evaluate this
24472 * property. Fail, and append the message */
24473 if (SvPOK(*saved_user_prop_ptr)) {
24474 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24475 sv_catsv(msg, *saved_user_prop_ptr);
24477 /* The SV in the hash won't be removed until global
24478 * destruction, so it is stable and we can unlock */
24479 USER_PROP_MUTEX_UNLOCK;
24484 assert(SvIOK(*saved_user_prop_ptr));
24486 /* Here, we have an unstable entry in the hash. Either another
24487 * thread is in the middle of expanding the property's
24488 * definition, or we are ourselves recursing. We use the aTHX
24489 * in it to distinguish */
24490 if (SvIV(*saved_user_prop_ptr) != PTR2IV(CUR_CONTEXT)) {
24492 /* Here, it's another thread doing the expanding. We've
24493 * looked as much as we are going to at the contents of the
24494 * hash entry. It's safe to unlock. */
24495 USER_PROP_MUTEX_UNLOCK;
24497 /* Retry a few times */
24498 if (retry_countdown-- > 0) {
24503 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24504 sv_catpvs(msg, "Timeout waiting for another thread to "
24506 goto append_name_to_msg;
24509 /* Here, we are recursing; don't dig any deeper */
24510 USER_PROP_MUTEX_UNLOCK;
24512 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24514 "Infinite recursion in user-defined property");
24515 goto append_name_to_msg;
24518 /* Here, this thread has exclusive control, and there is no entry
24519 * for this property in the hash. So we have the go ahead to
24520 * expand the definition ourselves. */
24522 PUSHSTACKi(PERLSI_REGCOMP);
24525 /* Create a temporary placeholder in the hash to detect recursion
24527 SWITCH_TO_GLOBAL_CONTEXT;
24528 placeholder= newSVuv(PTR2IV(ORIGINAL_CONTEXT));
24529 (void) hv_store_ent(PL_user_def_props, key, placeholder, 0);
24532 /* Now that we have a placeholder, we can let other threads
24534 USER_PROP_MUTEX_UNLOCK;
24536 /* Make sure the placeholder always gets destroyed */
24537 SAVEDESTRUCTOR_X(S_delete_recursion_entry, SvPVX(key));
24542 /* Call the user's function, with the /i status as a parameter.
24543 * Note that we have gone to a lot of trouble to keep this call
24544 * from being within the locked mutex region. */
24545 XPUSHs(boolSV(to_fold));
24548 /* The following block was taken from swash_init(). Presumably
24549 * they apply to here as well, though we no longer use a swash --
24553 /* We might get here via a subroutine signature which uses a utf8
24554 * parameter name, at which point PL_subname will have been set
24555 * but not yet used. */
24556 save_item(PL_subname);
24558 /* G_SCALAR guarantees a single return value */
24559 (void) call_sv(user_sub_sv, G_EVAL|G_SCALAR);
24564 if (TAINT_get || SvTRUE(error)) {
24565 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24566 if (SvTRUE(error)) {
24567 sv_catpvs(msg, "Error \"");
24568 sv_catsv(msg, error);
24569 sv_catpvs(msg, "\"");
24572 if (SvTRUE(error)) sv_catpvs(msg, "; ");
24573 sv_catpvn(msg, insecure, sizeof(insecure) - 1);
24576 if (name_len > 0) {
24577 sv_catpvs(msg, " in expansion of ");
24578 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8,
24584 prop_definition = NULL;
24587 SV * contents = POPs;
24589 /* The contents is supposed to be the expansion of the property
24590 * definition. If the definition is deferrable, and we got an
24591 * empty string back, set a flag to later defer it (after clean
24594 && (! SvPOK(contents) || SvCUR(contents) == 0))
24596 empty_return = TRUE;
24598 else { /* Otherwise, call a function to check for valid syntax,
24601 prop_definition = handle_user_defined_property(
24603 is_utf8, to_fold, runtime,
24605 contents, user_defined_ptr,
24611 /* Here, we have the results of the expansion. Delete the
24612 * placeholder, and if the definition is now known, replace it with
24613 * that definition. We need exclusive access to the hash, and we
24614 * can't let anyone else in, between when we delete the placeholder
24615 * and add the permanent entry */
24616 USER_PROP_MUTEX_LOCK;
24618 S_delete_recursion_entry(aTHX_ SvPVX(key));
24620 if ( ! empty_return
24621 && (! prop_definition || is_invlist(prop_definition)))
24623 /* If we got success we use the inversion list defining the
24624 * property; otherwise use the error message */
24625 SWITCH_TO_GLOBAL_CONTEXT;
24626 (void) hv_store_ent(PL_user_def_props,
24629 ? newSVsv(prop_definition)
24635 /* All done, and the hash now has a permanent entry for this
24636 * property. Give up exclusive control */
24637 USER_PROP_MUTEX_UNLOCK;
24643 if (empty_return) {
24644 goto definition_deferred;
24647 if (prop_definition) {
24649 /* If the definition is for something not known at this time,
24650 * we toss it, and go return the main property name, as that's
24651 * the one the user will be aware of */
24652 if (! is_invlist(prop_definition)) {
24653 SvREFCNT_dec_NN(prop_definition);
24654 goto definition_deferred;
24657 sv_2mortal(prop_definition);
24661 return prop_definition;
24663 } /* End of calling the subroutine for the user-defined property */
24664 } /* End of it could be a user-defined property */
24666 /* Here it wasn't a user-defined property that is known at this time. See
24667 * if it is a Unicode property */
24669 lookup_len = j; /* This is a more mnemonic name than 'j' */
24671 /* Get the index into our pointer table of the inversion list corresponding
24672 * to the property */
24673 table_index = do_uniprop_match(lookup_name, lookup_len);
24675 /* If it didn't find the property ... */
24676 if (table_index == 0) {
24678 /* Try again stripping off any initial 'Is'. This is because we
24679 * promise that an initial Is is optional. The same isn't true of
24680 * names that start with 'In'. Those can match only blocks, and the
24681 * lookup table already has those accounted for. The lookup table also
24682 * has already accounted for Perl extensions (without and = sign)
24683 * starting with 'i's'. */
24684 if (starts_with_Is && equals_pos >= 0) {
24690 table_index = do_uniprop_match(lookup_name, lookup_len);
24693 if (table_index == 0) {
24696 /* Here, we didn't find it. If not a numeric type property, and
24697 * can't be a user-defined one, it isn't a legal property */
24698 if (! is_nv_type) {
24699 if (! could_be_user_defined) {
24703 /* Here, the property name is legal as a user-defined one. At
24704 * compile time, it might just be that the subroutine for that
24705 * property hasn't been encountered yet, but at runtime, it's
24706 * an error to try to use an undefined one */
24707 if (! deferrable) {
24708 goto unknown_user_defined;;
24711 goto definition_deferred;
24712 } /* End of isn't a numeric type property */
24714 /* The numeric type properties need more work to decide. What we
24715 * do is make sure we have the number in canonical form and look
24718 if (slash_pos < 0) { /* No slash */
24720 /* When it isn't a rational, take the input, convert it to a
24721 * NV, then create a canonical string representation of that
24725 SSize_t value_len = lookup_len - equals_pos;
24727 /* Get the value */
24728 if ( value_len <= 0
24729 || my_atof3(lookup_name + equals_pos, &value,
24731 != lookup_name + lookup_len)
24736 /* If the value is an integer, the canonical value is integral
24738 if (Perl_ceil(value) == value) {
24739 canonical = Perl_form(aTHX_ "%.*s%.0" NVff,
24740 equals_pos, lookup_name, value);
24742 else { /* Otherwise, it is %e with a known precision */
24745 canonical = Perl_form(aTHX_ "%.*s%.*" NVef,
24746 equals_pos, lookup_name,
24747 PL_E_FORMAT_PRECISION, value);
24749 /* The exponent generated is expecting two digits, whereas
24750 * %e on some systems will generate three. Remove leading
24751 * zeros in excess of 2 from the exponent. We start
24752 * looking for them after the '=' */
24753 exp_ptr = strchr(canonical + equals_pos, 'e');
24755 char * cur_ptr = exp_ptr + 2; /* past the 'e[+-]' */
24756 SSize_t excess_exponent_len = strlen(cur_ptr) - 2;
24758 assert(*(cur_ptr - 1) == '-' || *(cur_ptr - 1) == '+');
24760 if (excess_exponent_len > 0) {
24761 SSize_t leading_zeros = strspn(cur_ptr, "0");
24762 SSize_t excess_leading_zeros
24763 = MIN(leading_zeros, excess_exponent_len);
24764 if (excess_leading_zeros > 0) {
24765 Move(cur_ptr + excess_leading_zeros,
24767 strlen(cur_ptr) - excess_leading_zeros
24768 + 1, /* Copy the NUL as well */
24775 else { /* Has a slash. Create a rational in canonical form */
24776 UV numerator, denominator, gcd, trial;
24777 const char * end_ptr;
24778 const char * sign = "";
24780 /* We can't just find the numerator, denominator, and do the
24781 * division, then use the method above, because that is
24782 * inexact. And the input could be a rational that is within
24783 * epsilon (given our precision) of a valid rational, and would
24784 * then incorrectly compare valid.
24786 * We're only interested in the part after the '=' */
24787 const char * this_lookup_name = lookup_name + equals_pos;
24788 lookup_len -= equals_pos;
24789 slash_pos -= equals_pos;
24791 /* Handle any leading minus */
24792 if (this_lookup_name[0] == '-') {
24794 this_lookup_name++;
24799 /* Convert the numerator to numeric */
24800 end_ptr = this_lookup_name + slash_pos;
24801 if (! grok_atoUV(this_lookup_name, &numerator, &end_ptr)) {
24805 /* It better have included all characters before the slash */
24806 if (*end_ptr != '/') {
24810 /* Set to look at just the denominator */
24811 this_lookup_name += slash_pos;
24812 lookup_len -= slash_pos;
24813 end_ptr = this_lookup_name + lookup_len;
24815 /* Convert the denominator to numeric */
24816 if (! grok_atoUV(this_lookup_name, &denominator, &end_ptr)) {
24820 /* It better be the rest of the characters, and don't divide by
24822 if ( end_ptr != this_lookup_name + lookup_len
24823 || denominator == 0)
24828 /* Get the greatest common denominator using
24829 http://en.wikipedia.org/wiki/Euclidean_algorithm */
24831 trial = denominator;
24832 while (trial != 0) {
24834 trial = gcd % trial;
24838 /* If already in lowest possible terms, we have already tried
24839 * looking this up */
24844 /* Reduce the rational, which should put it in canonical form
24847 denominator /= gcd;
24849 canonical = Perl_form(aTHX_ "%.*s%s%" UVuf "/%" UVuf,
24850 equals_pos, lookup_name, sign, numerator, denominator);
24853 /* Here, we have the number in canonical form. Try that */
24854 table_index = do_uniprop_match(canonical, strlen(canonical));
24855 if (table_index == 0) {
24858 } /* End of still didn't find the property in our table */
24859 } /* End of didn't find the property in our table */
24861 /* Here, we have a non-zero return, which is an index into a table of ptrs.
24862 * A negative return signifies that the real index is the absolute value,
24863 * but the result needs to be inverted */
24864 if (table_index < 0) {
24865 invert_return = TRUE;
24866 table_index = -table_index;
24869 /* Out-of band indices indicate a deprecated property. The proper index is
24870 * modulo it with the table size. And dividing by the table size yields
24871 * an offset into a table constructed by regen/mk_invlists.pl to contain
24872 * the corresponding warning message */
24873 if (table_index > MAX_UNI_KEYWORD_INDEX) {
24874 Size_t warning_offset = table_index / MAX_UNI_KEYWORD_INDEX;
24875 table_index %= MAX_UNI_KEYWORD_INDEX;
24876 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED),
24877 "Use of '%.*s' in \\p{} or \\P{} is deprecated because: %s",
24878 (int) name_len, name,
24879 get_deprecated_property_msg(warning_offset));
24882 /* In a few properties, a different property is used under /i. These are
24883 * unlikely to change, so are hard-coded here. */
24885 if ( table_index == UNI_XPOSIXUPPER
24886 || table_index == UNI_XPOSIXLOWER
24887 || table_index == UNI_TITLE)
24889 table_index = UNI_CASED;
24891 else if ( table_index == UNI_UPPERCASELETTER
24892 || table_index == UNI_LOWERCASELETTER
24893 # ifdef UNI_TITLECASELETTER /* Missing from early Unicodes */
24894 || table_index == UNI_TITLECASELETTER
24897 table_index = UNI_CASEDLETTER;
24899 else if ( table_index == UNI_POSIXUPPER
24900 || table_index == UNI_POSIXLOWER)
24902 table_index = UNI_POSIXALPHA;
24906 /* Create and return the inversion list */
24907 prop_definition = get_prop_definition(table_index);
24908 sv_2mortal(prop_definition);
24910 /* See if there is a private use override to add to this definition */
24912 COPHH * hinthash = (IN_PERL_COMPILETIME)
24913 ? CopHINTHASH_get(&PL_compiling)
24914 : CopHINTHASH_get(PL_curcop);
24915 SV * pu_overrides = cophh_fetch_pv(hinthash, "private_use", 0, 0);
24917 if (UNLIKELY(pu_overrides && SvPOK(pu_overrides))) {
24919 /* See if there is an element in the hints hash for this table */
24920 SV * pu_lookup = Perl_newSVpvf(aTHX_ "%d=", table_index);
24921 const char * pos = strstr(SvPVX(pu_overrides), SvPVX(pu_lookup));
24925 SV * pu_definition;
24927 SV * expanded_prop_definition =
24928 sv_2mortal(invlist_clone(prop_definition, NULL));
24930 /* If so, it's definition is the string from here to the next
24931 * \a character. And its format is the same as a user-defined
24933 pos += SvCUR(pu_lookup);
24934 pu_definition = newSVpvn(pos, strchr(pos, '\a') - pos);
24935 pu_invlist = handle_user_defined_property(lookup_name,
24938 0, /* Not folded */
24946 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24947 sv_catpvs(msg, "Insecure private-use override");
24948 goto append_name_to_msg;
24951 /* For now, as a safety measure, make sure that it doesn't
24952 * override non-private use code points */
24953 _invlist_intersection(pu_invlist, PL_Private_Use, &pu_invlist);
24955 /* Add it to the list to be returned */
24956 _invlist_union(prop_definition, pu_invlist,
24957 &expanded_prop_definition);
24958 prop_definition = expanded_prop_definition;
24959 Perl_ck_warner_d(aTHX_ packWARN(WARN_EXPERIMENTAL__PRIVATE_USE), "The private_use feature is experimental");
24964 if (invert_return) {
24965 _invlist_invert(prop_definition);
24967 return prop_definition;
24969 unknown_user_defined:
24970 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24971 sv_catpvs(msg, "Unknown user-defined property name");
24972 goto append_name_to_msg;
24975 if (non_pkg_begin != 0) {
24976 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24977 sv_catpvs(msg, "Illegal user-defined property name");
24980 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24981 sv_catpvs(msg, "Can't find Unicode property definition");
24985 append_name_to_msg:
24987 const char * prefix = (runtime && level == 0) ? " \\p{" : " \"";
24988 const char * suffix = (runtime && level == 0) ? "}" : "\"";
24990 sv_catpv(msg, prefix);
24991 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8, name_len, name));
24992 sv_catpv(msg, suffix);
24997 definition_deferred:
25000 bool is_qualified = non_pkg_begin != 0; /* If has "::" */
25002 /* Here it could yet to be defined, so defer evaluation of this until
25003 * its needed at runtime. We need the fully qualified property name to
25004 * avoid ambiguity */
25006 fq_name = S_get_fq_name(aTHX_ name, name_len, is_utf8,
25010 /* If it didn't come with a package, or the package is utf8::, this
25011 * actually could be an official Unicode property whose inclusion we
25012 * are deferring until runtime to make sure that it isn't overridden by
25013 * a user-defined property of the same name (which we haven't
25014 * encountered yet). Add a marker to indicate this possibility, for
25015 * use at such time when we first need the definition during pattern
25016 * matching execution */
25017 if (! is_qualified || memBEGINPs(name, non_pkg_begin, "utf8::")) {
25018 sv_catpvs(fq_name, DEFERRED_COULD_BE_OFFICIAL_MARKERs);
25021 /* We also need a trailing newline */
25022 sv_catpvs(fq_name, "\n");
25024 *user_defined_ptr = TRUE;
25030 S_handle_names_wildcard(pTHX_ const char * wname, /* wildcard name to match */
25031 const STRLEN wname_len, /* Its length */
25032 SV ** prop_definition,
25035 /* Deal with Name property wildcard subpatterns; returns TRUE if there were
25036 * any matches, adding them to prop_definition */
25040 CV * get_names_info; /* entry to charnames.pm to get info we need */
25041 SV * names_string; /* Contains all character names, except algo */
25042 SV * algorithmic_names; /* Contains info about algorithmically
25043 generated character names */
25044 REGEXP * subpattern_re; /* The user's pattern to match with */
25045 struct regexp * prog; /* The compiled pattern */
25046 char * all_names_start; /* lib/unicore/Name.pl string of every
25047 (non-algorithmic) character name */
25048 char * cur_pos; /* We match, effectively using /gc; this is
25049 where we are now */
25050 bool found_matches = FALSE; /* Did any name match so far? */
25051 SV * empty; /* For matching zero length names */
25052 SV * must_sv; /* Contains the substring, if any, that must be
25053 in a name for the subpattern to match */
25054 const char * must; /* The PV of 'must' */
25055 STRLEN must_len; /* And its length */
25056 SV * syllable_name = NULL; /* For Hangul syllables */
25057 const char hangul_prefix[] = "HANGUL SYLLABLE ";
25058 const STRLEN hangul_prefix_len = sizeof(hangul_prefix) - 1;
25060 /* By inspection, there are a maximum of 7 bytes in the suffix of a hangul
25061 * syllable name, and these are immutable and guaranteed by the Unicode
25062 * standard to never be extended */
25063 const STRLEN syl_max_len = hangul_prefix_len + 7;
25067 PERL_ARGS_ASSERT_HANDLE_NAMES_WILDCARD;
25069 /* Make sure _charnames is loaded. (The parameters give context
25070 * for any errors generated */
25071 get_names_info = get_cv("_charnames::_get_names_info", 0);
25072 if (! get_names_info) {
25073 Perl_croak(aTHX_ "panic: Can't find '_charnames::_get_names_info");
25076 /* Get the charnames data */
25077 PUSHSTACKi(PERLSI_REGCOMP);
25085 /* Special _charnames entry point that returns the info this routine
25087 call_sv(MUTABLE_SV(get_names_info), G_ARRAY);
25091 /* Data structure for names which end in their very own code points */
25092 algorithmic_names = POPs;
25093 SvREFCNT_inc_simple_void_NN(algorithmic_names);
25095 /* The lib/unicore/Name.pl string */
25096 names_string = POPs;
25097 SvREFCNT_inc_simple_void_NN(names_string);
25104 if ( ! SvROK(names_string)
25105 || ! SvROK(algorithmic_names))
25106 { /* Perhaps should panic instead XXX */
25107 SvREFCNT_dec(names_string);
25108 SvREFCNT_dec(algorithmic_names);
25112 names_string = sv_2mortal(SvRV(names_string));
25113 all_names_start = SvPVX(names_string);
25114 cur_pos = all_names_start;
25116 algorithmic_names= sv_2mortal(SvRV(algorithmic_names));
25118 /* Compile the subpattern consisting of the name being looked for */
25119 subpattern_re = compile_wildcard(wname, wname_len, FALSE /* /-i */ );
25121 must_sv = re_intuit_string(subpattern_re);
25123 /* regexec.c can free the re_intuit_string() return. GH #17734 */
25124 must_sv = sv_2mortal(newSVsv(must_sv));
25125 must = SvPV(must_sv, must_len);
25132 /* (Note: 'must' could contain a NUL. And yet we use strspn() below on it.
25133 * This works because the NUL causes the function to return early, thus
25134 * showing that there are characters in it other than the acceptable ones,
25135 * which is our desired result.) */
25137 prog = ReANY(subpattern_re);
25139 /* If only nothing is matched, skip to where empty names are looked for */
25140 if (prog->maxlen == 0) {
25144 /* And match against the string of all names /gc. Don't even try if it
25145 * must match a character not found in any name. */
25146 if (strspn(must, "\n -0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ()") == must_len)
25148 while (execute_wildcard(subpattern_re,
25150 SvEND(names_string),
25151 all_names_start, 0,
25154 { /* Here, matched. */
25156 /* Note the string entries look like
25157 * 00001\nSTART OF HEADING\n\n
25158 * so we could match anywhere in that string. We have to rule out
25159 * matching a code point line */
25160 char * this_name_start = all_names_start
25161 + RX_OFFS(subpattern_re)->start;
25162 char * this_name_end = all_names_start
25163 + RX_OFFS(subpattern_re)->end;
25166 UV cp = 0; /* Silences some compilers */
25167 AV * this_string = NULL;
25168 bool is_multi = FALSE;
25170 /* If matched nothing, advance to next possible match */
25171 if (this_name_start == this_name_end) {
25172 cur_pos = (char *) memchr(this_name_end + 1, '\n',
25173 SvEND(names_string) - this_name_end);
25174 if (cur_pos == NULL) {
25179 /* Position the next match to start beyond the current returned
25181 cur_pos = (char *) memchr(this_name_end, '\n',
25182 SvEND(names_string) - this_name_end);
25185 /* Back up to the \n just before the beginning of the character. */
25186 cp_end = (char *) my_memrchr(all_names_start,
25188 this_name_start - all_names_start);
25190 /* If we didn't find a \n, it means it matched somewhere in the
25191 * initial '00000' in the string, so isn't a real match */
25192 if (cp_end == NULL) {
25196 this_name_start = cp_end + 1; /* The name starts just after */
25197 cp_end--; /* the \n, and the code point */
25198 /* ends just before it */
25200 /* All code points are 5 digits long */
25201 cp_start = cp_end - 4;
25203 /* This shouldn't happen, as we found a \n, and the first \n is
25204 * further along than what we subtracted */
25205 assert(cp_start >= all_names_start);
25207 if (cp_start == all_names_start) {
25208 *prop_definition = add_cp_to_invlist(*prop_definition, 0);
25212 /* If the character is a blank, we either have a named sequence, or
25213 * something is wrong */
25214 if (*(cp_start - 1) == ' ') {
25215 cp_start = (char *) my_memrchr(all_names_start,
25217 cp_start - all_names_start);
25221 assert(cp_start != NULL && cp_start >= all_names_start + 2);
25223 /* Except for the first line in the string, the sequence before the
25224 * code point is \n\n. If that isn't the case here, we didn't
25225 * match the name of a character. (We could have matched a named
25226 * sequence, not currently handled */
25227 if (*(cp_start - 1) != '\n' || *(cp_start - 2) != '\n') {
25231 /* We matched! Add this to the list */
25232 found_matches = TRUE;
25234 /* Loop through all the code points in the sequence */
25235 while (cp_start < cp_end) {
25237 /* Calculate this code point from its 5 digits */
25238 cp = (XDIGIT_VALUE(cp_start[0]) << 16)
25239 + (XDIGIT_VALUE(cp_start[1]) << 12)
25240 + (XDIGIT_VALUE(cp_start[2]) << 8)
25241 + (XDIGIT_VALUE(cp_start[3]) << 4)
25242 + XDIGIT_VALUE(cp_start[4]);
25244 cp_start += 6; /* Go past any blank */
25246 if (cp_start < cp_end || is_multi) {
25247 if (this_string == NULL) {
25248 this_string = newAV();
25252 av_push(this_string, newSVuv(cp));
25256 if (is_multi) { /* Was more than one code point */
25257 if (*strings == NULL) {
25258 *strings = newAV();
25261 av_push(*strings, (SV *) this_string);
25263 else { /* Only a single code point */
25264 *prop_definition = add_cp_to_invlist(*prop_definition, cp);
25266 } /* End of loop through the non-algorithmic names string */
25269 /* There are also character names not in 'names_string'. These are
25270 * algorithmically generatable. Try this pattern on each possible one.
25271 * (khw originally planned to leave this out given the large number of
25272 * matches attempted; but the speed turned out to be quite acceptable
25274 * There are plenty of opportunities to optimize to skip many of the tests.
25275 * beyond the rudimentary ones already here */
25277 /* First see if the subpattern matches any of the algorithmic generatable
25278 * Hangul syllable names.
25280 * We know none of these syllable names will match if the input pattern
25281 * requires more bytes than any syllable has, or if the input pattern only
25282 * matches an empty name, or if the pattern has something it must match and
25283 * one of the characters in that isn't in any Hangul syllable. */
25284 if ( prog->minlen <= (SSize_t) syl_max_len
25285 && prog->maxlen > 0
25286 && (strspn(must, "\n ABCDEGHIJKLMNOPRSTUWY") == must_len))
25288 /* These constants, names, values, and algorithm are adapted from the
25289 * Unicode standard, version 5.1, section 3.12, and should never
25291 const char * JamoL[] = {
25292 "G", "GG", "N", "D", "DD", "R", "M", "B", "BB",
25293 "S", "SS", "", "J", "JJ", "C", "K", "T", "P", "H"
25295 const int LCount = C_ARRAY_LENGTH(JamoL);
25297 const char * JamoV[] = {
25298 "A", "AE", "YA", "YAE", "EO", "E", "YEO", "YE", "O", "WA",
25299 "WAE", "OE", "YO", "U", "WEO", "WE", "WI", "YU", "EU", "YI",
25302 const int VCount = C_ARRAY_LENGTH(JamoV);
25304 const char * JamoT[] = {
25305 "", "G", "GG", "GS", "N", "NJ", "NH", "D", "L",
25306 "LG", "LM", "LB", "LS", "LT", "LP", "LH", "M", "B",
25307 "BS", "S", "SS", "NG", "J", "C", "K", "T", "P", "H"
25309 const int TCount = C_ARRAY_LENGTH(JamoT);
25313 /* This is the initial Hangul syllable code point; each time through the
25314 * inner loop, it maps to the next higher code point. For more info,
25315 * see the Hangul syllable section of the Unicode standard. */
25318 syllable_name = sv_2mortal(newSV(syl_max_len));
25319 sv_setpvn(syllable_name, hangul_prefix, hangul_prefix_len);
25321 for (L = 0; L < LCount; L++) {
25322 for (V = 0; V < VCount; V++) {
25323 for (T = 0; T < TCount; T++) {
25325 /* Truncate back to the prefix, which is unvarying */
25326 SvCUR_set(syllable_name, hangul_prefix_len);
25328 sv_catpv(syllable_name, JamoL[L]);
25329 sv_catpv(syllable_name, JamoV[V]);
25330 sv_catpv(syllable_name, JamoT[T]);
25332 if (execute_wildcard(subpattern_re,
25333 SvPVX(syllable_name),
25334 SvEND(syllable_name),
25335 SvPVX(syllable_name), 0,
25339 *prop_definition = add_cp_to_invlist(*prop_definition,
25341 found_matches = TRUE;
25350 /* The rest of the algorithmically generatable names are of the form
25351 * "PREFIX-code_point". The prefixes and the code point limits of each
25352 * were returned to us in the array 'algorithmic_names' from data in
25353 * lib/unicore/Name.pm. 'code_point' in the name is expressed in hex. */
25354 for (i = 0; i <= av_top_index((AV *) algorithmic_names); i++) {
25357 /* Each element of the array is a hash, giving the details for the
25358 * series of names it covers. There is the base name of the characters
25359 * in the series, and the low and high code points in the series. And,
25360 * for optimization purposes a string containing all the legal
25361 * characters that could possibly be in a name in this series. */
25362 HV * this_series = (HV *) SvRV(* av_fetch((AV *) algorithmic_names, i, 0));
25363 SV * prefix = * hv_fetchs(this_series, "name", 0);
25364 IV low = SvIV(* hv_fetchs(this_series, "low", 0));
25365 IV high = SvIV(* hv_fetchs(this_series, "high", 0));
25366 char * legal = SvPVX(* hv_fetchs(this_series, "legal", 0));
25368 /* Pre-allocate an SV with enough space */
25369 SV * algo_name = sv_2mortal(Perl_newSVpvf(aTHX_ "%s-0000",
25371 if (high >= 0x10000) {
25372 sv_catpvs(algo_name, "0");
25375 /* This series can be skipped entirely if the pattern requires
25376 * something longer than any name in the series, or can only match an
25377 * empty name, or contains a character not found in any name in the
25379 if ( prog->minlen <= (SSize_t) SvCUR(algo_name)
25380 && prog->maxlen > 0
25381 && (strspn(must, legal) == must_len))
25383 for (j = low; j <= high; j++) { /* For each code point in the series */
25385 /* Get its name, and see if it matches the subpattern */
25386 Perl_sv_setpvf(aTHX_ algo_name, "%s-%X", SvPVX(prefix),
25389 if (execute_wildcard(subpattern_re,
25392 SvPVX(algo_name), 0,
25396 *prop_definition = add_cp_to_invlist(*prop_definition, j);
25397 found_matches = TRUE;
25404 /* Finally, see if the subpattern matches an empty string */
25405 empty = newSVpvs("");
25406 if (execute_wildcard(subpattern_re,
25413 /* Many code points have empty names. Currently these are the \p{GC=C}
25414 * ones, minus CC and CF */
25416 SV * empty_names_ref = get_prop_definition(UNI_C);
25417 SV * empty_names = invlist_clone(empty_names_ref, NULL);
25419 SV * subtract = get_prop_definition(UNI_CC);
25421 _invlist_subtract(empty_names, subtract, &empty_names);
25422 SvREFCNT_dec_NN(empty_names_ref);
25423 SvREFCNT_dec_NN(subtract);
25425 subtract = get_prop_definition(UNI_CF);
25426 _invlist_subtract(empty_names, subtract, &empty_names);
25427 SvREFCNT_dec_NN(subtract);
25429 _invlist_union(*prop_definition, empty_names, prop_definition);
25430 found_matches = TRUE;
25431 SvREFCNT_dec_NN(empty_names);
25433 SvREFCNT_dec_NN(empty);
25436 /* If we ever were to accept aliases for, say private use names, we would
25437 * need to do something fancier to find empty names. The code below works
25438 * (at the time it was written), and is slower than the above */
25439 const char empties_pat[] = "^.";
25440 if (strNE(name, empties_pat)) {
25441 SV * empty = newSVpvs("");
25442 if (execute_wildcard(subpattern_re,
25449 SV * empties = NULL;
25451 (void) handle_names_wildcard(empties_pat, strlen(empties_pat), &empties);
25453 _invlist_union_complement_2nd(*prop_definition, empties, prop_definition);
25454 SvREFCNT_dec_NN(empties);
25456 found_matches = TRUE;
25458 SvREFCNT_dec_NN(empty);
25462 SvREFCNT_dec_NN(subpattern_re);
25463 return found_matches;
25467 * ex: set ts=8 sts=4 sw=4 et: