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"
146 #define HAS_NONLATIN1_FOLD_CLOSURE(i) \
147 _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
148 #define HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(i) \
149 _HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
150 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
151 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
154 #define STATIC static
157 /* this is a chain of data about sub patterns we are processing that
158 need to be handled separately/specially in study_chunk. Its so
159 we can simulate recursion without losing state. */
161 typedef struct scan_frame {
162 regnode *last_regnode; /* last node to process in this frame */
163 regnode *next_regnode; /* next node to process when last is reached */
164 U32 prev_recursed_depth;
165 I32 stopparen; /* what stopparen do we use */
167 struct scan_frame *this_prev_frame; /* this previous frame */
168 struct scan_frame *prev_frame; /* previous frame */
169 struct scan_frame *next_frame; /* next frame */
172 /* Certain characters are output as a sequence with the first being a
174 #define isBACKSLASHED_PUNCT(c) memCHRs("-[]\\^", c)
177 struct RExC_state_t {
178 U32 flags; /* RXf_* are we folding, multilining? */
179 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
180 char *precomp; /* uncompiled string. */
181 char *precomp_end; /* pointer to end of uncompiled string. */
182 REGEXP *rx_sv; /* The SV that is the regexp. */
183 regexp *rx; /* perl core regexp structure */
184 regexp_internal *rxi; /* internal data for regexp object
186 char *start; /* Start of input for compile */
187 char *end; /* End of input for compile */
188 char *parse; /* Input-scan pointer. */
189 char *copy_start; /* start of copy of input within
190 constructed parse string */
191 char *save_copy_start; /* Provides one level of saving
192 and restoring 'copy_start' */
193 char *copy_start_in_input; /* Position in input string
194 corresponding to copy_start */
195 SSize_t whilem_seen; /* number of WHILEM in this expr */
196 regnode *emit_start; /* Start of emitted-code area */
197 regnode_offset emit; /* Code-emit pointer */
198 I32 naughty; /* How bad is this pattern? */
199 I32 sawback; /* Did we see \1, ...? */
200 SSize_t size; /* Number of regnode equivalents in
202 Size_t sets_depth; /* Counts recursion depth of already-
203 compiled regex set patterns */
206 I32 parens_buf_size; /* #slots malloced open/close_parens */
207 regnode_offset *open_parens; /* offsets to open parens */
208 regnode_offset *close_parens; /* offsets to close parens */
209 HV *paren_names; /* Paren names */
211 /* position beyond 'precomp' of the warning message furthest away from
212 * 'precomp'. During the parse, no warnings are raised for any problems
213 * earlier in the parse than this position. This works if warnings are
214 * raised the first time a given spot is parsed, and if only one
215 * independent warning is raised for any given spot */
216 Size_t latest_warn_offset;
218 I32 npar; /* Capture buffer count so far in the
219 parse, (OPEN) plus one. ("par" 0 is
221 I32 total_par; /* During initial parse, is either 0,
222 or -1; the latter indicating a
223 reparse is needed. After that pass,
224 it is what 'npar' became after the
225 pass. Hence, it being > 0 indicates
226 we are in a reparse situation */
227 I32 nestroot; /* root parens we are in - used by
230 regnode *end_op; /* END node in program */
231 I32 utf8; /* whether the pattern is utf8 or not */
232 I32 orig_utf8; /* whether the pattern was originally in utf8 */
233 /* XXX use this for future optimisation of case
234 * where pattern must be upgraded to utf8. */
235 I32 uni_semantics; /* If a d charset modifier should use unicode
236 rules, even if the pattern is not in
239 I32 recurse_count; /* Number of recurse regops we have generated */
240 regnode **recurse; /* Recurse regops */
241 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
243 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
247 I32 override_recoding;
248 I32 recode_x_to_native;
249 I32 in_multi_char_class;
250 int code_index; /* next code_blocks[] slot */
251 struct reg_code_blocks *code_blocks;/* positions of literal (?{})
253 SSize_t maxlen; /* mininum possible number of chars in string to match */
254 scan_frame *frame_head;
255 scan_frame *frame_last;
259 SV *runtime_code_qr; /* qr with the runtime code blocks */
261 const char *lastparse;
263 U32 study_chunk_recursed_count;
264 AV *paren_name_list; /* idx -> name */
268 #define RExC_lastparse (pRExC_state->lastparse)
269 #define RExC_lastnum (pRExC_state->lastnum)
270 #define RExC_paren_name_list (pRExC_state->paren_name_list)
271 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
272 #define RExC_mysv (pRExC_state->mysv1)
273 #define RExC_mysv1 (pRExC_state->mysv1)
274 #define RExC_mysv2 (pRExC_state->mysv2)
282 bool sWARN_EXPERIMENTAL__VLB;
283 bool sWARN_EXPERIMENTAL__REGEX_SETS;
286 #define RExC_flags (pRExC_state->flags)
287 #define RExC_pm_flags (pRExC_state->pm_flags)
288 #define RExC_precomp (pRExC_state->precomp)
289 #define RExC_copy_start_in_input (pRExC_state->copy_start_in_input)
290 #define RExC_copy_start_in_constructed (pRExC_state->copy_start)
291 #define RExC_save_copy_start_in_constructed (pRExC_state->save_copy_start)
292 #define RExC_precomp_end (pRExC_state->precomp_end)
293 #define RExC_rx_sv (pRExC_state->rx_sv)
294 #define RExC_rx (pRExC_state->rx)
295 #define RExC_rxi (pRExC_state->rxi)
296 #define RExC_start (pRExC_state->start)
297 #define RExC_end (pRExC_state->end)
298 #define RExC_parse (pRExC_state->parse)
299 #define RExC_latest_warn_offset (pRExC_state->latest_warn_offset )
300 #define RExC_whilem_seen (pRExC_state->whilem_seen)
301 #define RExC_seen_d_op (pRExC_state->seen_d_op) /* Seen something that differs
302 under /d from /u ? */
304 #ifdef RE_TRACK_PATTERN_OFFSETS
305 # define RExC_offsets (RExC_rxi->u.offsets) /* I am not like the
308 #define RExC_emit (pRExC_state->emit)
309 #define RExC_emit_start (pRExC_state->emit_start)
310 #define RExC_sawback (pRExC_state->sawback)
311 #define RExC_seen (pRExC_state->seen)
312 #define RExC_size (pRExC_state->size)
313 #define RExC_maxlen (pRExC_state->maxlen)
314 #define RExC_npar (pRExC_state->npar)
315 #define RExC_total_parens (pRExC_state->total_par)
316 #define RExC_parens_buf_size (pRExC_state->parens_buf_size)
317 #define RExC_nestroot (pRExC_state->nestroot)
318 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
319 #define RExC_utf8 (pRExC_state->utf8)
320 #define RExC_uni_semantics (pRExC_state->uni_semantics)
321 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
322 #define RExC_open_parens (pRExC_state->open_parens)
323 #define RExC_close_parens (pRExC_state->close_parens)
324 #define RExC_end_op (pRExC_state->end_op)
325 #define RExC_paren_names (pRExC_state->paren_names)
326 #define RExC_recurse (pRExC_state->recurse)
327 #define RExC_recurse_count (pRExC_state->recurse_count)
328 #define RExC_sets_depth (pRExC_state->sets_depth)
329 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
330 #define RExC_study_chunk_recursed_bytes \
331 (pRExC_state->study_chunk_recursed_bytes)
332 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
333 #define RExC_in_lookahead (pRExC_state->in_lookahead)
334 #define RExC_contains_locale (pRExC_state->contains_locale)
335 #define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
338 # define SET_recode_x_to_native(x) \
339 STMT_START { RExC_recode_x_to_native = (x); } STMT_END
341 # define SET_recode_x_to_native(x) NOOP
344 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
345 #define RExC_frame_head (pRExC_state->frame_head)
346 #define RExC_frame_last (pRExC_state->frame_last)
347 #define RExC_frame_count (pRExC_state->frame_count)
348 #define RExC_strict (pRExC_state->strict)
349 #define RExC_study_started (pRExC_state->study_started)
350 #define RExC_warn_text (pRExC_state->warn_text)
351 #define RExC_in_script_run (pRExC_state->in_script_run)
352 #define RExC_use_BRANCHJ (pRExC_state->use_BRANCHJ)
353 #define RExC_warned_WARN_EXPERIMENTAL__VLB (pRExC_state->sWARN_EXPERIMENTAL__VLB)
354 #define RExC_warned_WARN_EXPERIMENTAL__REGEX_SETS (pRExC_state->sWARN_EXPERIMENTAL__REGEX_SETS)
355 #define RExC_unlexed_names (pRExC_state->unlexed_names)
357 /* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
358 * a flag to disable back-off on the fixed/floating substrings - if it's
359 * a high complexity pattern we assume the benefit of avoiding a full match
360 * is worth the cost of checking for the substrings even if they rarely help.
362 #define RExC_naughty (pRExC_state->naughty)
363 #define TOO_NAUGHTY (10)
364 #define MARK_NAUGHTY(add) \
365 if (RExC_naughty < TOO_NAUGHTY) \
366 RExC_naughty += (add)
367 #define MARK_NAUGHTY_EXP(exp, add) \
368 if (RExC_naughty < TOO_NAUGHTY) \
369 RExC_naughty += RExC_naughty / (exp) + (add)
371 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
372 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
373 ((*s) == '{' && regcurly(s)))
376 * Flags to be passed up and down.
378 #define WORST 0 /* Worst case. */
379 #define HASWIDTH 0x01 /* Known to not match null strings, could match
382 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
383 * character. (There needs to be a case: in the switch statement in regexec.c
384 * for any node marked SIMPLE.) Note that this is not the same thing as
387 #define SPSTART 0x04 /* Starts with * or + */
388 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
389 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
390 #define RESTART_PARSE 0x20 /* Need to redo the parse */
391 #define NEED_UTF8 0x40 /* In conjunction with RESTART_PARSE, need to
392 calcuate sizes as UTF-8 */
394 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
396 /* whether trie related optimizations are enabled */
397 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
398 #define TRIE_STUDY_OPT
399 #define FULL_TRIE_STUDY
405 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
406 #define PBITVAL(paren) (1 << ((paren) & 7))
407 #define PAREN_OFFSET(depth) \
408 (RExC_study_chunk_recursed + (depth) * RExC_study_chunk_recursed_bytes)
409 #define PAREN_TEST(depth, paren) \
410 (PBYTE(PAREN_OFFSET(depth), paren) & PBITVAL(paren))
411 #define PAREN_SET(depth, paren) \
412 (PBYTE(PAREN_OFFSET(depth), paren) |= PBITVAL(paren))
413 #define PAREN_UNSET(depth, paren) \
414 (PBYTE(PAREN_OFFSET(depth), paren) &= ~PBITVAL(paren))
416 #define REQUIRE_UTF8(flagp) STMT_START { \
418 *flagp = RESTART_PARSE|NEED_UTF8; \
423 /* Change from /d into /u rules, and restart the parse. RExC_uni_semantics is
424 * a flag that indicates we need to override /d with /u as a result of
425 * something in the pattern. It should only be used in regards to calling
426 * set_regex_charset() or get_regex_charset() */
427 #define REQUIRE_UNI_RULES(flagp, restart_retval) \
429 if (DEPENDS_SEMANTICS) { \
430 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
431 RExC_uni_semantics = 1; \
432 if (RExC_seen_d_op && LIKELY(! IN_PARENS_PASS)) { \
433 /* No need to restart the parse if we haven't seen \
434 * anything that differs between /u and /d, and no need \
435 * to restart immediately if we're going to reparse \
436 * anyway to count parens */ \
437 *flagp |= RESTART_PARSE; \
438 return restart_retval; \
443 #define REQUIRE_BRANCHJ(flagp, restart_retval) \
445 RExC_use_BRANCHJ = 1; \
446 *flagp |= RESTART_PARSE; \
447 return restart_retval; \
450 /* Until we have completed the parse, we leave RExC_total_parens at 0 or
451 * less. After that, it must always be positive, because the whole re is
452 * considered to be surrounded by virtual parens. Setting it to negative
453 * indicates there is some construct that needs to know the actual number of
454 * parens to be properly handled. And that means an extra pass will be
455 * required after we've counted them all */
456 #define ALL_PARENS_COUNTED (RExC_total_parens > 0)
457 #define REQUIRE_PARENS_PASS \
458 STMT_START { /* No-op if have completed a pass */ \
459 if (! ALL_PARENS_COUNTED) RExC_total_parens = -1; \
461 #define IN_PARENS_PASS (RExC_total_parens < 0)
464 /* This is used to return failure (zero) early from the calling function if
465 * various flags in 'flags' are set. Two flags always cause a return:
466 * 'RESTART_PARSE' and 'NEED_UTF8'. 'extra' can be used to specify any
467 * additional flags that should cause a return; 0 if none. If the return will
468 * be done, '*flagp' is first set to be all of the flags that caused the
470 #define RETURN_FAIL_ON_RESTART_OR_FLAGS(flags,flagp,extra) \
472 if ((flags) & (RESTART_PARSE|NEED_UTF8|(extra))) { \
473 *(flagp) = (flags) & (RESTART_PARSE|NEED_UTF8|(extra)); \
478 #define MUST_RESTART(flags) ((flags) & (RESTART_PARSE))
480 #define RETURN_FAIL_ON_RESTART(flags,flagp) \
481 RETURN_FAIL_ON_RESTART_OR_FLAGS( flags, flagp, 0)
482 #define RETURN_FAIL_ON_RESTART_FLAGP(flagp) \
483 if (MUST_RESTART(*(flagp))) return 0
485 /* This converts the named class defined in regcomp.h to its equivalent class
486 * number defined in handy.h. */
487 #define namedclass_to_classnum(class) ((int) ((class) / 2))
488 #define classnum_to_namedclass(classnum) ((classnum) * 2)
490 #define _invlist_union_complement_2nd(a, b, output) \
491 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
492 #define _invlist_intersection_complement_2nd(a, b, output) \
493 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
495 /* We add a marker if we are deferring expansion of a property that is both
496 * 1) potentiallly user-defined; and
497 * 2) could also be an official Unicode property.
499 * Without this marker, any deferred expansion can only be for a user-defined
500 * one. This marker shouldn't conflict with any that could be in a legal name,
501 * and is appended to its name to indicate this. There is a string and
503 #define DEFERRED_COULD_BE_OFFICIAL_MARKERs "~"
504 #define DEFERRED_COULD_BE_OFFICIAL_MARKERc '~'
506 /* What is infinity for optimization purposes */
507 #define OPTIMIZE_INFTY SSize_t_MAX
509 /* About scan_data_t.
511 During optimisation we recurse through the regexp program performing
512 various inplace (keyhole style) optimisations. In addition study_chunk
513 and scan_commit populate this data structure with information about
514 what strings MUST appear in the pattern. We look for the longest
515 string that must appear at a fixed location, and we look for the
516 longest string that may appear at a floating location. So for instance
521 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
522 strings (because they follow a .* construct). study_chunk will identify
523 both FOO and BAR as being the longest fixed and floating strings respectively.
525 The strings can be composites, for instance
529 will result in a composite fixed substring 'foo'.
531 For each string some basic information is maintained:
534 This is the position the string must appear at, or not before.
535 It also implicitly (when combined with minlenp) tells us how many
536 characters must match before the string we are searching for.
537 Likewise when combined with minlenp and the length of the string it
538 tells us how many characters must appear after the string we have
542 Only used for floating strings. This is the rightmost point that
543 the string can appear at. If set to OPTIMIZE_INFTY it indicates that the
544 string can occur infinitely far to the right.
545 For fixed strings, it is equal to min_offset.
548 A pointer to the minimum number of characters of the pattern that the
549 string was found inside. This is important as in the case of positive
550 lookahead or positive lookbehind we can have multiple patterns
555 The minimum length of the pattern overall is 3, the minimum length
556 of the lookahead part is 3, but the minimum length of the part that
557 will actually match is 1. So 'FOO's minimum length is 3, but the
558 minimum length for the F is 1. This is important as the minimum length
559 is used to determine offsets in front of and behind the string being
560 looked for. Since strings can be composites this is the length of the
561 pattern at the time it was committed with a scan_commit. Note that
562 the length is calculated by study_chunk, so that the minimum lengths
563 are not known until the full pattern has been compiled, thus the
564 pointer to the value.
568 In the case of lookbehind the string being searched for can be
569 offset past the start point of the final matching string.
570 If this value was just blithely removed from the min_offset it would
571 invalidate some of the calculations for how many chars must match
572 before or after (as they are derived from min_offset and minlen and
573 the length of the string being searched for).
574 When the final pattern is compiled and the data is moved from the
575 scan_data_t structure into the regexp structure the information
576 about lookbehind is factored in, with the information that would
577 have been lost precalculated in the end_shift field for the
580 The fields pos_min and pos_delta are used to store the minimum offset
581 and the delta to the maximum offset at the current point in the pattern.
585 struct scan_data_substrs {
586 SV *str; /* longest substring found in pattern */
587 SSize_t min_offset; /* earliest point in string it can appear */
588 SSize_t max_offset; /* latest point in string it can appear */
589 SSize_t *minlenp; /* pointer to the minlen relevant to the string */
590 SSize_t lookbehind; /* is the pos of the string modified by LB */
591 I32 flags; /* per substring SF_* and SCF_* flags */
594 typedef struct scan_data_t {
595 /*I32 len_min; unused */
596 /*I32 len_delta; unused */
600 SSize_t last_end; /* min value, <0 unless valid. */
601 SSize_t last_start_min;
602 SSize_t last_start_max;
603 U8 cur_is_floating; /* whether the last_* values should be set as
604 * the next fixed (0) or floating (1)
607 /* [0] is longest fixed substring so far, [1] is longest float so far */
608 struct scan_data_substrs substrs[2];
610 I32 flags; /* common SF_* and SCF_* flags */
612 SSize_t *last_closep;
613 regnode_ssc *start_class;
617 * Forward declarations for pregcomp()'s friends.
620 static const scan_data_t zero_scan_data = {
621 0, 0, NULL, 0, 0, 0, 0,
623 { NULL, 0, 0, 0, 0, 0 },
624 { NULL, 0, 0, 0, 0, 0 },
631 #define SF_BEFORE_SEOL 0x0001
632 #define SF_BEFORE_MEOL 0x0002
633 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
635 #define SF_IS_INF 0x0040
636 #define SF_HAS_PAR 0x0080
637 #define SF_IN_PAR 0x0100
638 #define SF_HAS_EVAL 0x0200
641 /* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
642 * longest substring in the pattern. When it is not set the optimiser keeps
643 * track of position, but does not keep track of the actual strings seen,
645 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
648 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
649 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
650 * turned off because of the alternation (BRANCH). */
651 #define SCF_DO_SUBSTR 0x0400
653 #define SCF_DO_STCLASS_AND 0x0800
654 #define SCF_DO_STCLASS_OR 0x1000
655 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
656 #define SCF_WHILEM_VISITED_POS 0x2000
658 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
659 #define SCF_SEEN_ACCEPT 0x8000
660 #define SCF_TRIE_DOING_RESTUDY 0x10000
661 #define SCF_IN_DEFINE 0x20000
666 #define UTF cBOOL(RExC_utf8)
668 /* The enums for all these are ordered so things work out correctly */
669 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
670 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
671 == REGEX_DEPENDS_CHARSET)
672 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
673 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
674 >= REGEX_UNICODE_CHARSET)
675 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
676 == REGEX_ASCII_RESTRICTED_CHARSET)
677 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
678 >= REGEX_ASCII_RESTRICTED_CHARSET)
679 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
680 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
682 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
684 /* For programs that want to be strictly Unicode compatible by dying if any
685 * attempt is made to match a non-Unicode code point against a Unicode
687 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
689 #define OOB_NAMEDCLASS -1
691 /* There is no code point that is out-of-bounds, so this is problematic. But
692 * its only current use is to initialize a variable that is always set before
694 #define OOB_UNICODE 0xDEADBEEF
696 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
699 /* length of regex to show in messages that don't mark a position within */
700 #define RegexLengthToShowInErrorMessages 127
703 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
704 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
705 * op/pragma/warn/regcomp.
707 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
708 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
710 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
711 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
713 /* The code in this file in places uses one level of recursion with parsing
714 * rebased to an alternate string constructed by us in memory. This can take
715 * the form of something that is completely different from the input, or
716 * something that uses the input as part of the alternate. In the first case,
717 * there should be no possibility of an error, as we are in complete control of
718 * the alternate string. But in the second case we don't completely control
719 * the input portion, so there may be errors in that. Here's an example:
721 * is handled specially because \x{df} folds to a sequence of more than one
722 * character: 'ss'. What is done is to create and parse an alternate string,
723 * which looks like this:
724 * /(?:\x{DF}|[abc\x{DF}def])/ui
725 * where it uses the input unchanged in the middle of something it constructs,
726 * which is a branch for the DF outside the character class, and clustering
727 * parens around the whole thing. (It knows enough to skip the DF inside the
728 * class while in this substitute parse.) 'abc' and 'def' may have errors that
729 * need to be reported. The general situation looks like this:
731 * |<------- identical ------>|
733 * Input: ---------------------------------------------------------------
734 * Constructed: ---------------------------------------------------
736 * |<------- identical ------>|
738 * sI..eI is the portion of the input pattern we are concerned with here.
739 * sC..EC is the constructed substitute parse string.
740 * sC..tC is constructed by us
741 * tC..eC is an exact duplicate of the portion of the input pattern tI..eI.
742 * In the diagram, these are vertically aligned.
743 * eC..EC is also constructed by us.
744 * xC is the position in the substitute parse string where we found a
746 * xI is the position in the original pattern corresponding to xC.
748 * We want to display a message showing the real input string. Thus we need to
749 * translate from xC to xI. We know that xC >= tC, since the portion of the
750 * string sC..tC has been constructed by us, and so shouldn't have errors. We
752 * xI = tI + (xC - tC)
754 * When the substitute parse is constructed, the code needs to set:
757 * RExC_copy_start_in_input (tI)
758 * RExC_copy_start_in_constructed (tC)
759 * and restore them when done.
761 * During normal processing of the input pattern, both
762 * 'RExC_copy_start_in_input' and 'RExC_copy_start_in_constructed' are set to
763 * sI, so that xC equals xI.
766 #define sI RExC_precomp
767 #define eI RExC_precomp_end
768 #define sC RExC_start
770 #define tI RExC_copy_start_in_input
771 #define tC RExC_copy_start_in_constructed
772 #define xI(xC) (tI + (xC - tC))
773 #define xI_offset(xC) (xI(xC) - sI)
775 #define REPORT_LOCATION_ARGS(xC) \
777 (xI(xC) > eI) /* Don't run off end */ \
778 ? eI - sI /* Length before the <--HERE */ \
779 : ((xI_offset(xC) >= 0) \
781 : (Perl_croak(aTHX_ "panic: %s: %d: negative offset: %" \
782 IVdf " trying to output message for " \
784 __FILE__, __LINE__, (IV) xI_offset(xC), \
785 ((int) (eC - sC)), sC), 0)), \
786 sI), /* The input pattern printed up to the <--HERE */ \
788 (xI(xC) > eI) ? 0 : eI - xI(xC), /* Length after <--HERE */ \
789 (xI(xC) > eI) ? eI : xI(xC)) /* pattern after <--HERE */
791 /* Used to point after bad bytes for an error message, but avoid skipping
792 * past a nul byte. */
793 #define SKIP_IF_CHAR(s, e) (!*(s) ? 0 : UTF ? UTF8_SAFE_SKIP(s, e) : 1)
795 /* Set up to clean up after our imminent demise */
796 #define PREPARE_TO_DIE \
799 SAVEFREESV(RExC_rx_sv); \
800 if (RExC_open_parens) \
801 SAVEFREEPV(RExC_open_parens); \
802 if (RExC_close_parens) \
803 SAVEFREEPV(RExC_close_parens); \
807 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
808 * arg. Show regex, up to a maximum length. If it's too long, chop and add
811 #define _FAIL(code) STMT_START { \
812 const char *ellipses = ""; \
813 IV len = RExC_precomp_end - RExC_precomp; \
816 if (len > RegexLengthToShowInErrorMessages) { \
817 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
818 len = RegexLengthToShowInErrorMessages - 10; \
824 #define FAIL(msg) _FAIL( \
825 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
826 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
828 #define FAIL2(msg,arg) _FAIL( \
829 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
830 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
832 #define FAIL3(msg,arg1,arg2) _FAIL( \
833 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
834 arg1, arg2, UTF8fARG(UTF, len, RExC_precomp), ellipses))
837 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
839 #define Simple_vFAIL(m) STMT_START { \
840 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
841 m, REPORT_LOCATION_ARGS(RExC_parse)); \
845 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
847 #define vFAIL(m) STMT_START { \
853 * Like Simple_vFAIL(), but accepts two arguments.
855 #define Simple_vFAIL2(m,a1) STMT_START { \
856 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, \
857 REPORT_LOCATION_ARGS(RExC_parse)); \
861 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
863 #define vFAIL2(m,a1) STMT_START { \
865 Simple_vFAIL2(m, a1); \
870 * Like Simple_vFAIL(), but accepts three arguments.
872 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
873 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, a2, \
874 REPORT_LOCATION_ARGS(RExC_parse)); \
878 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
880 #define vFAIL3(m,a1,a2) STMT_START { \
882 Simple_vFAIL3(m, a1, a2); \
886 * Like Simple_vFAIL(), but accepts four arguments.
888 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
889 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, a2, a3, \
890 REPORT_LOCATION_ARGS(RExC_parse)); \
893 #define vFAIL4(m,a1,a2,a3) STMT_START { \
895 Simple_vFAIL4(m, a1, a2, a3); \
898 /* A specialized version of vFAIL2 that works with UTF8f */
899 #define vFAIL2utf8f(m, a1) STMT_START { \
901 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, \
902 REPORT_LOCATION_ARGS(RExC_parse)); \
905 #define vFAIL3utf8f(m, a1, a2) STMT_START { \
907 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, a2, \
908 REPORT_LOCATION_ARGS(RExC_parse)); \
911 /* Setting this to NULL is a signal to not output warnings */
912 #define TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE \
914 RExC_save_copy_start_in_constructed = RExC_copy_start_in_constructed;\
915 RExC_copy_start_in_constructed = NULL; \
917 #define RESTORE_WARNINGS \
918 RExC_copy_start_in_constructed = RExC_save_copy_start_in_constructed
920 /* Since a warning can be generated multiple times as the input is reparsed, we
921 * output it the first time we come to that point in the parse, but suppress it
922 * otherwise. 'RExC_copy_start_in_constructed' being NULL is a flag to not
923 * generate any warnings */
924 #define TO_OUTPUT_WARNINGS(loc) \
925 ( RExC_copy_start_in_constructed \
926 && ((xI(loc)) - RExC_precomp) > (Ptrdiff_t) RExC_latest_warn_offset)
928 /* After we've emitted a warning, we save the position in the input so we don't
930 #define UPDATE_WARNINGS_LOC(loc) \
932 if (TO_OUTPUT_WARNINGS(loc)) { \
933 RExC_latest_warn_offset = MAX(sI, MIN(eI, xI(loc))) \
938 /* 'warns' is the output of the packWARNx macro used in 'code' */
939 #define _WARN_HELPER(loc, warns, code) \
941 if (! RExC_copy_start_in_constructed) { \
942 Perl_croak( aTHX_ "panic! %s: %d: Tried to warn when none" \
943 " expected at '%s'", \
944 __FILE__, __LINE__, loc); \
946 if (TO_OUTPUT_WARNINGS(loc)) { \
950 UPDATE_WARNINGS_LOC(loc); \
954 /* m is not necessarily a "literal string", in this macro */
955 #define warn_non_literal_string(loc, packed_warn, m) \
956 _WARN_HELPER(loc, packed_warn, \
957 Perl_warner(aTHX_ packed_warn, \
958 "%s" REPORT_LOCATION, \
959 m, REPORT_LOCATION_ARGS(loc)))
960 #define reg_warn_non_literal_string(loc, m) \
961 warn_non_literal_string(loc, packWARN(WARN_REGEXP), m)
963 #define ckWARN2_non_literal_string(loc, packwarn, m, a1) \
966 Size_t format_size = strlen(m) + strlen(REPORT_LOCATION)+ 1;\
967 Newx(format, format_size, char); \
968 my_strlcpy(format, m, format_size); \
969 my_strlcat(format, REPORT_LOCATION, format_size); \
970 SAVEFREEPV(format); \
971 _WARN_HELPER(loc, packwarn, \
972 Perl_ck_warner(aTHX_ packwarn, \
974 a1, REPORT_LOCATION_ARGS(loc))); \
977 #define ckWARNreg(loc,m) \
978 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
979 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
981 REPORT_LOCATION_ARGS(loc)))
983 #define vWARN(loc, m) \
984 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
985 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
987 REPORT_LOCATION_ARGS(loc))) \
989 #define vWARN_dep(loc, m) \
990 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
991 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
993 REPORT_LOCATION_ARGS(loc)))
995 #define ckWARNdep(loc,m) \
996 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
997 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
999 REPORT_LOCATION_ARGS(loc)))
1001 #define ckWARNregdep(loc,m) \
1002 _WARN_HELPER(loc, packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
1003 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
1005 m REPORT_LOCATION, \
1006 REPORT_LOCATION_ARGS(loc)))
1008 #define ckWARN2reg_d(loc,m, a1) \
1009 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1010 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
1011 m REPORT_LOCATION, \
1012 a1, REPORT_LOCATION_ARGS(loc)))
1014 #define ckWARN2reg(loc, m, a1) \
1015 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1016 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
1017 m REPORT_LOCATION, \
1018 a1, REPORT_LOCATION_ARGS(loc)))
1020 #define vWARN3(loc, m, a1, a2) \
1021 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1022 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
1023 m REPORT_LOCATION, \
1024 a1, a2, REPORT_LOCATION_ARGS(loc)))
1026 #define ckWARN3reg(loc, m, a1, a2) \
1027 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1028 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
1029 m REPORT_LOCATION, \
1031 REPORT_LOCATION_ARGS(loc)))
1033 #define vWARN4(loc, m, a1, a2, a3) \
1034 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1035 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
1036 m REPORT_LOCATION, \
1038 REPORT_LOCATION_ARGS(loc)))
1040 #define ckWARN4reg(loc, m, a1, a2, a3) \
1041 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1042 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
1043 m REPORT_LOCATION, \
1045 REPORT_LOCATION_ARGS(loc)))
1047 #define vWARN5(loc, m, a1, a2, a3, a4) \
1048 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1049 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
1050 m REPORT_LOCATION, \
1052 REPORT_LOCATION_ARGS(loc)))
1054 #define ckWARNexperimental(loc, class, m) \
1056 if (! RExC_warned_ ## class) { /* warn once per compilation */ \
1057 RExC_warned_ ## class = 1; \
1058 _WARN_HELPER(loc, packWARN(class), \
1059 Perl_ck_warner_d(aTHX_ packWARN(class), \
1060 m REPORT_LOCATION, \
1061 REPORT_LOCATION_ARGS(loc)));\
1065 /* Convert between a pointer to a node and its offset from the beginning of the
1067 #define REGNODE_p(offset) (RExC_emit_start + (offset))
1068 #define REGNODE_OFFSET(node) ((node) - RExC_emit_start)
1070 /* Macros for recording node offsets. 20001227 mjd@plover.com
1071 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
1072 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
1073 * Element 0 holds the number n.
1074 * Position is 1 indexed.
1076 #ifndef RE_TRACK_PATTERN_OFFSETS
1077 #define Set_Node_Offset_To_R(offset,byte)
1078 #define Set_Node_Offset(node,byte)
1079 #define Set_Cur_Node_Offset
1080 #define Set_Node_Length_To_R(node,len)
1081 #define Set_Node_Length(node,len)
1082 #define Set_Node_Cur_Length(node,start)
1083 #define Node_Offset(n)
1084 #define Node_Length(n)
1085 #define Set_Node_Offset_Length(node,offset,len)
1086 #define ProgLen(ri) ri->u.proglen
1087 #define SetProgLen(ri,x) ri->u.proglen = x
1088 #define Track_Code(code)
1090 #define ProgLen(ri) ri->u.offsets[0]
1091 #define SetProgLen(ri,x) ri->u.offsets[0] = x
1092 #define Set_Node_Offset_To_R(offset,byte) STMT_START { \
1093 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
1094 __LINE__, (int)(offset), (int)(byte))); \
1095 if((offset) < 0) { \
1096 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
1099 RExC_offsets[2*(offset)-1] = (byte); \
1103 #define Set_Node_Offset(node,byte) \
1104 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (byte)-RExC_start)
1105 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
1107 #define Set_Node_Length_To_R(node,len) STMT_START { \
1108 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
1109 __LINE__, (int)(node), (int)(len))); \
1111 Perl_croak(aTHX_ "value of node is %d in Length macro", \
1114 RExC_offsets[2*(node)] = (len); \
1118 #define Set_Node_Length(node,len) \
1119 Set_Node_Length_To_R(REGNODE_OFFSET(node), len)
1120 #define Set_Node_Cur_Length(node, start) \
1121 Set_Node_Length(node, RExC_parse - start)
1123 /* Get offsets and lengths */
1124 #define Node_Offset(n) (RExC_offsets[2*(REGNODE_OFFSET(n))-1])
1125 #define Node_Length(n) (RExC_offsets[2*(REGNODE_OFFSET(n))])
1127 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
1128 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (offset)); \
1129 Set_Node_Length_To_R(REGNODE_OFFSET(node), (len)); \
1132 #define Track_Code(code) STMT_START { code } STMT_END
1135 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
1136 #define EXPERIMENTAL_INPLACESCAN
1137 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
1141 Perl_re_printf(pTHX_ const char *fmt, ...)
1145 PerlIO *f= Perl_debug_log;
1146 PERL_ARGS_ASSERT_RE_PRINTF;
1148 result = PerlIO_vprintf(f, fmt, ap);
1154 Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
1158 PerlIO *f= Perl_debug_log;
1159 PERL_ARGS_ASSERT_RE_INDENTF;
1160 va_start(ap, depth);
1161 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
1162 result = PerlIO_vprintf(f, fmt, ap);
1166 #endif /* DEBUGGING */
1168 #define DEBUG_RExC_seen() \
1169 DEBUG_OPTIMISE_MORE_r({ \
1170 Perl_re_printf( aTHX_ "RExC_seen: "); \
1172 if (RExC_seen & REG_ZERO_LEN_SEEN) \
1173 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
1175 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
1176 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
1178 if (RExC_seen & REG_GPOS_SEEN) \
1179 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
1181 if (RExC_seen & REG_RECURSE_SEEN) \
1182 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
1184 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
1185 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
1187 if (RExC_seen & REG_VERBARG_SEEN) \
1188 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
1190 if (RExC_seen & REG_CUTGROUP_SEEN) \
1191 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
1193 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
1194 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
1196 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
1197 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
1199 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
1200 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
1202 Perl_re_printf( aTHX_ "\n"); \
1205 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
1206 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
1211 S_debug_show_study_flags(pTHX_ U32 flags, const char *open_str,
1212 const char *close_str)
1217 Perl_re_printf( aTHX_ "%s", open_str);
1218 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_SEOL);
1219 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_MEOL);
1220 DEBUG_SHOW_STUDY_FLAG(flags, SF_IS_INF);
1221 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_PAR);
1222 DEBUG_SHOW_STUDY_FLAG(flags, SF_IN_PAR);
1223 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_EVAL);
1224 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_SUBSTR);
1225 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_AND);
1226 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_OR);
1227 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS);
1228 DEBUG_SHOW_STUDY_FLAG(flags, SCF_WHILEM_VISITED_POS);
1229 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_RESTUDY);
1230 DEBUG_SHOW_STUDY_FLAG(flags, SCF_SEEN_ACCEPT);
1231 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_DOING_RESTUDY);
1232 DEBUG_SHOW_STUDY_FLAG(flags, SCF_IN_DEFINE);
1233 Perl_re_printf( aTHX_ "%s", close_str);
1238 S_debug_studydata(pTHX_ const char *where, scan_data_t *data,
1239 U32 depth, int is_inf)
1241 DECLARE_AND_GET_RE_DEBUG_FLAGS;
1243 DEBUG_OPTIMISE_MORE_r({
1246 Perl_re_indentf(aTHX_ "%s: Pos:%" IVdf "/%" IVdf " Flags: 0x%" UVXf,
1250 (IV)data->pos_delta,
1254 S_debug_show_study_flags(aTHX_ data->flags," [","]");
1256 Perl_re_printf( aTHX_
1257 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s",
1259 (IV)(data->last_closep ? *((data)->last_closep) : -1),
1260 is_inf ? "INF " : ""
1263 if (data->last_found) {
1265 Perl_re_printf(aTHX_
1266 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf,
1267 SvPVX_const(data->last_found),
1269 (IV)data->last_start_min,
1270 (IV)data->last_start_max
1273 for (i = 0; i < 2; i++) {
1274 Perl_re_printf(aTHX_
1275 " %s%s: '%s' @ %" IVdf "/%" IVdf,
1276 data->cur_is_floating == i ? "*" : "",
1277 i ? "Float" : "Fixed",
1278 SvPVX_const(data->substrs[i].str),
1279 (IV)data->substrs[i].min_offset,
1280 (IV)data->substrs[i].max_offset
1282 S_debug_show_study_flags(aTHX_ data->substrs[i].flags," [","]");
1286 Perl_re_printf( aTHX_ "\n");
1292 S_debug_peep(pTHX_ const char *str, const RExC_state_t *pRExC_state,
1293 regnode *scan, U32 depth, U32 flags)
1295 DECLARE_AND_GET_RE_DEBUG_FLAGS;
1302 Next = regnext(scan);
1303 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
1304 Perl_re_indentf( aTHX_ "%s>%3d: %s (%d)",
1307 REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),
1308 Next ? (REG_NODE_NUM(Next)) : 0 );
1309 S_debug_show_study_flags(aTHX_ flags," [ ","]");
1310 Perl_re_printf( aTHX_ "\n");
1315 # define DEBUG_STUDYDATA(where, data, depth, is_inf) \
1316 S_debug_studydata(aTHX_ where, data, depth, is_inf)
1318 # define DEBUG_PEEP(str, scan, depth, flags) \
1319 S_debug_peep(aTHX_ str, pRExC_state, scan, depth, flags)
1322 # define DEBUG_STUDYDATA(where, data, depth, is_inf) NOOP
1323 # define DEBUG_PEEP(str, scan, depth, flags) NOOP
1327 /* =========================================================
1328 * BEGIN edit_distance stuff.
1330 * This calculates how many single character changes of any type are needed to
1331 * transform a string into another one. It is taken from version 3.1 of
1333 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1336 /* Our unsorted dictionary linked list. */
1337 /* Note we use UVs, not chars. */
1342 struct dictionary* next;
1344 typedef struct dictionary item;
1347 PERL_STATIC_INLINE item*
1348 push(UV key, item* curr)
1351 Newx(head, 1, item);
1359 PERL_STATIC_INLINE item*
1360 find(item* head, UV key)
1362 item* iterator = head;
1364 if (iterator->key == key){
1367 iterator = iterator->next;
1373 PERL_STATIC_INLINE item*
1374 uniquePush(item* head, UV key)
1376 item* iterator = head;
1379 if (iterator->key == key) {
1382 iterator = iterator->next;
1385 return push(key, head);
1388 PERL_STATIC_INLINE void
1389 dict_free(item* head)
1391 item* iterator = head;
1394 item* temp = iterator;
1395 iterator = iterator->next;
1402 /* End of Dictionary Stuff */
1404 /* All calculations/work are done here */
1406 S_edit_distance(const UV* src,
1408 const STRLEN x, /* length of src[] */
1409 const STRLEN y, /* length of tgt[] */
1410 const SSize_t maxDistance
1414 UV swapCount, swapScore, targetCharCount, i, j;
1416 UV score_ceil = x + y;
1418 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1420 /* intialize matrix start values */
1421 Newx(scores, ( (x + 2) * (y + 2)), UV);
1422 scores[0] = score_ceil;
1423 scores[1 * (y + 2) + 0] = score_ceil;
1424 scores[0 * (y + 2) + 1] = score_ceil;
1425 scores[1 * (y + 2) + 1] = 0;
1426 head = uniquePush(uniquePush(head, src[0]), tgt[0]);
1431 for (i=1;i<=x;i++) {
1433 head = uniquePush(head, src[i]);
1434 scores[(i+1) * (y + 2) + 1] = i;
1435 scores[(i+1) * (y + 2) + 0] = score_ceil;
1438 for (j=1;j<=y;j++) {
1441 head = uniquePush(head, tgt[j]);
1442 scores[1 * (y + 2) + (j + 1)] = j;
1443 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1446 targetCharCount = find(head, tgt[j-1])->value;
1447 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1449 if (src[i-1] != tgt[j-1]){
1450 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));
1454 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1458 find(head, src[i-1])->value = i;
1462 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1465 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1469 /* END of edit_distance() stuff
1470 * ========================================================= */
1472 /* Mark that we cannot extend a found fixed substring at this point.
1473 Update the longest found anchored substring or the longest found
1474 floating substrings if needed. */
1477 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1478 SSize_t *minlenp, int is_inf)
1480 const STRLEN l = CHR_SVLEN(data->last_found);
1481 SV * const longest_sv = data->substrs[data->cur_is_floating].str;
1482 const STRLEN old_l = CHR_SVLEN(longest_sv);
1483 DECLARE_AND_GET_RE_DEBUG_FLAGS;
1485 PERL_ARGS_ASSERT_SCAN_COMMIT;
1487 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
1488 const U8 i = data->cur_is_floating;
1489 SvSetMagicSV(longest_sv, data->last_found);
1490 data->substrs[i].min_offset = l ? data->last_start_min : data->pos_min;
1493 data->substrs[0].max_offset = data->substrs[0].min_offset;
1495 data->substrs[1].max_offset =
1499 ? data->last_start_max
1500 : (data->pos_delta > OPTIMIZE_INFTY - data->pos_min
1502 : data->pos_min + data->pos_delta));
1505 if (data->flags & SF_BEFORE_EOL)
1506 data->substrs[i].flags |= (data->flags & SF_BEFORE_EOL);
1508 data->substrs[i].flags &= ~SF_BEFORE_EOL;
1509 data->substrs[i].minlenp = minlenp;
1510 data->substrs[i].lookbehind = 0;
1513 SvCUR_set(data->last_found, 0);
1515 SV * const sv = data->last_found;
1516 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1517 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1522 data->last_end = -1;
1523 data->flags &= ~SF_BEFORE_EOL;
1524 DEBUG_STUDYDATA("commit", data, 0, is_inf);
1527 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1528 * list that describes which code points it matches */
1531 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1533 /* Set the SSC 'ssc' to match an empty string or any code point */
1535 PERL_ARGS_ASSERT_SSC_ANYTHING;
1537 assert(is_ANYOF_SYNTHETIC(ssc));
1539 /* mortalize so won't leak */
1540 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
1541 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1545 S_ssc_is_anything(const regnode_ssc *ssc)
1547 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1548 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1549 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1550 * in any way, so there's no point in using it */
1555 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1557 assert(is_ANYOF_SYNTHETIC(ssc));
1559 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1563 /* See if the list consists solely of the range 0 - Infinity */
1564 invlist_iterinit(ssc->invlist);
1565 ret = invlist_iternext(ssc->invlist, &start, &end)
1569 invlist_iterfinish(ssc->invlist);
1575 /* If e.g., both \w and \W are set, matches everything */
1576 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1578 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1579 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1589 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1591 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1592 * string, any code point, or any posix class under locale */
1594 PERL_ARGS_ASSERT_SSC_INIT;
1596 Zero(ssc, 1, regnode_ssc);
1597 set_ANYOF_SYNTHETIC(ssc);
1598 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1601 /* If any portion of the regex is to operate under locale rules that aren't
1602 * fully known at compile time, initialization includes it. The reason
1603 * this isn't done for all regexes is that the optimizer was written under
1604 * the assumption that locale was all-or-nothing. Given the complexity and
1605 * lack of documentation in the optimizer, and that there are inadequate
1606 * test cases for locale, many parts of it may not work properly, it is
1607 * safest to avoid locale unless necessary. */
1608 if (RExC_contains_locale) {
1609 ANYOF_POSIXL_SETALL(ssc);
1612 ANYOF_POSIXL_ZERO(ssc);
1617 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1618 const regnode_ssc *ssc)
1620 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1621 * to the list of code points matched, and locale posix classes; hence does
1622 * not check its flags) */
1627 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1629 assert(is_ANYOF_SYNTHETIC(ssc));
1631 invlist_iterinit(ssc->invlist);
1632 ret = invlist_iternext(ssc->invlist, &start, &end)
1636 invlist_iterfinish(ssc->invlist);
1642 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1649 #define INVLIST_INDEX 0
1650 #define ONLY_LOCALE_MATCHES_INDEX 1
1651 #define DEFERRED_USER_DEFINED_INDEX 2
1654 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1655 const regnode_charclass* const node)
1657 /* Returns a mortal inversion list defining which code points are matched
1658 * by 'node', which is of type ANYOF. Handles complementing the result if
1659 * appropriate. If some code points aren't knowable at this time, the
1660 * returned list must, and will, contain every code point that is a
1665 SV* only_utf8_locale_invlist = NULL;
1667 const U32 n = ARG(node);
1668 bool new_node_has_latin1 = FALSE;
1669 const U8 flags = (inRANGE(OP(node), ANYOFH, ANYOFRb))
1671 : ANYOF_FLAGS(node);
1673 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1675 /* Look at the data structure created by S_set_ANYOF_arg() */
1676 if (n != ANYOF_ONLY_HAS_BITMAP) {
1677 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1678 AV * const av = MUTABLE_AV(SvRV(rv));
1679 SV **const ary = AvARRAY(av);
1680 assert(RExC_rxi->data->what[n] == 's');
1682 if (av_tindex_skip_len_mg(av) >= DEFERRED_USER_DEFINED_INDEX) {
1684 /* Here there are things that won't be known until runtime -- we
1685 * have to assume it could be anything */
1686 invlist = sv_2mortal(_new_invlist(1));
1687 return _add_range_to_invlist(invlist, 0, UV_MAX);
1689 else if (ary[INVLIST_INDEX]) {
1691 /* Use the node's inversion list */
1692 invlist = sv_2mortal(invlist_clone(ary[INVLIST_INDEX], NULL));
1695 /* Get the code points valid only under UTF-8 locales */
1696 if ( (flags & ANYOFL_FOLD)
1697 && av_tindex_skip_len_mg(av) >= ONLY_LOCALE_MATCHES_INDEX)
1699 only_utf8_locale_invlist = ary[ONLY_LOCALE_MATCHES_INDEX];
1704 invlist = sv_2mortal(_new_invlist(0));
1707 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1708 * code points, and an inversion list for the others, but if there are code
1709 * points that should match only conditionally on the target string being
1710 * UTF-8, those are placed in the inversion list, and not the bitmap.
1711 * Since there are circumstances under which they could match, they are
1712 * included in the SSC. But if the ANYOF node is to be inverted, we have
1713 * to exclude them here, so that when we invert below, the end result
1714 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1715 * have to do this here before we add the unconditionally matched code
1717 if (flags & ANYOF_INVERT) {
1718 _invlist_intersection_complement_2nd(invlist,
1723 /* Add in the points from the bit map */
1724 if (! inRANGE(OP(node), ANYOFH, ANYOFRb)) {
1725 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1726 if (ANYOF_BITMAP_TEST(node, i)) {
1727 unsigned int start = i++;
1729 for (; i < NUM_ANYOF_CODE_POINTS
1730 && ANYOF_BITMAP_TEST(node, i); ++i)
1734 invlist = _add_range_to_invlist(invlist, start, i-1);
1735 new_node_has_latin1 = TRUE;
1740 /* If this can match all upper Latin1 code points, have to add them
1741 * as well. But don't add them if inverting, as when that gets done below,
1742 * it would exclude all these characters, including the ones it shouldn't
1743 * that were added just above */
1744 if (! (flags & ANYOF_INVERT) && OP(node) == ANYOFD
1745 && (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1747 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1750 /* Similarly for these */
1751 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1752 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1755 if (flags & ANYOF_INVERT) {
1756 _invlist_invert(invlist);
1758 else if (flags & ANYOFL_FOLD) {
1759 if (new_node_has_latin1) {
1761 /* Under /li, any 0-255 could fold to any other 0-255, depending on
1762 * the locale. We can skip this if there are no 0-255 at all. */
1763 _invlist_union(invlist, PL_Latin1, &invlist);
1765 invlist = add_cp_to_invlist(invlist, LATIN_SMALL_LETTER_DOTLESS_I);
1766 invlist = add_cp_to_invlist(invlist, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
1769 if (_invlist_contains_cp(invlist, LATIN_SMALL_LETTER_DOTLESS_I)) {
1770 invlist = add_cp_to_invlist(invlist, 'I');
1772 if (_invlist_contains_cp(invlist,
1773 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE))
1775 invlist = add_cp_to_invlist(invlist, 'i');
1780 /* Similarly add the UTF-8 locale possible matches. These have to be
1781 * deferred until after the non-UTF-8 locale ones are taken care of just
1782 * above, or it leads to wrong results under ANYOF_INVERT */
1783 if (only_utf8_locale_invlist) {
1784 _invlist_union_maybe_complement_2nd(invlist,
1785 only_utf8_locale_invlist,
1786 flags & ANYOF_INVERT,
1793 /* These two functions currently do the exact same thing */
1794 #define ssc_init_zero ssc_init
1796 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1797 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1799 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1800 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1801 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1804 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1805 const regnode_charclass *and_with)
1807 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1808 * another SSC or a regular ANYOF class. Can create false positives. */
1811 U8 and_with_flags = inRANGE(OP(and_with), ANYOFH, ANYOFRb)
1813 : ANYOF_FLAGS(and_with);
1816 PERL_ARGS_ASSERT_SSC_AND;
1818 assert(is_ANYOF_SYNTHETIC(ssc));
1820 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1821 * the code point inversion list and just the relevant flags */
1822 if (is_ANYOF_SYNTHETIC(and_with)) {
1823 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1824 anded_flags = and_with_flags;
1826 /* XXX This is a kludge around what appears to be deficiencies in the
1827 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1828 * there are paths through the optimizer where it doesn't get weeded
1829 * out when it should. And if we don't make some extra provision for
1830 * it like the code just below, it doesn't get added when it should.
1831 * This solution is to add it only when AND'ing, which is here, and
1832 * only when what is being AND'ed is the pristine, original node
1833 * matching anything. Thus it is like adding it to ssc_anything() but
1834 * only when the result is to be AND'ed. Probably the same solution
1835 * could be adopted for the same problem we have with /l matching,
1836 * which is solved differently in S_ssc_init(), and that would lead to
1837 * fewer false positives than that solution has. But if this solution
1838 * creates bugs, the consequences are only that a warning isn't raised
1839 * that should be; while the consequences for having /l bugs is
1840 * incorrect matches */
1841 if (ssc_is_anything((regnode_ssc *)and_with)) {
1842 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1846 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1847 if (OP(and_with) == ANYOFD) {
1848 anded_flags = and_with_flags & ANYOF_COMMON_FLAGS;
1851 anded_flags = and_with_flags
1852 &( ANYOF_COMMON_FLAGS
1853 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1854 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1855 if (ANYOFL_UTF8_LOCALE_REQD(and_with_flags)) {
1857 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1862 ANYOF_FLAGS(ssc) &= anded_flags;
1864 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1865 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1866 * 'and_with' may be inverted. When not inverted, we have the situation of
1868 * (C1 | P1) & (C2 | P2)
1869 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1870 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1871 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1872 * <= ((C1 & C2) | P1 | P2)
1873 * Alternatively, the last few steps could be:
1874 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1875 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1876 * <= (C1 | C2 | (P1 & P2))
1877 * We favor the second approach if either P1 or P2 is non-empty. This is
1878 * because these components are a barrier to doing optimizations, as what
1879 * they match cannot be known until the moment of matching as they are
1880 * dependent on the current locale, 'AND"ing them likely will reduce or
1882 * But we can do better if we know that C1,P1 are in their initial state (a
1883 * frequent occurrence), each matching everything:
1884 * (<everything>) & (C2 | P2) = C2 | P2
1885 * Similarly, if C2,P2 are in their initial state (again a frequent
1886 * occurrence), the result is a no-op
1887 * (C1 | P1) & (<everything>) = C1 | P1
1890 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1891 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1892 * <= (C1 & ~C2) | (P1 & ~P2)
1895 if ((and_with_flags & ANYOF_INVERT)
1896 && ! is_ANYOF_SYNTHETIC(and_with))
1900 ssc_intersection(ssc,
1902 FALSE /* Has already been inverted */
1905 /* If either P1 or P2 is empty, the intersection will be also; can skip
1907 if (! (and_with_flags & ANYOF_MATCHES_POSIXL)) {
1908 ANYOF_POSIXL_ZERO(ssc);
1910 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1912 /* Note that the Posix class component P from 'and_with' actually
1914 * P = Pa | Pb | ... | Pn
1915 * where each component is one posix class, such as in [\w\s].
1917 * ~P = ~(Pa | Pb | ... | Pn)
1918 * = ~Pa & ~Pb & ... & ~Pn
1919 * <= ~Pa | ~Pb | ... | ~Pn
1920 * The last is something we can easily calculate, but unfortunately
1921 * is likely to have many false positives. We could do better
1922 * in some (but certainly not all) instances if two classes in
1923 * P have known relationships. For example
1924 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1926 * :lower: & :print: = :lower:
1927 * And similarly for classes that must be disjoint. For example,
1928 * since \s and \w can have no elements in common based on rules in
1929 * the POSIX standard,
1930 * \w & ^\S = nothing
1931 * Unfortunately, some vendor locales do not meet the Posix
1932 * standard, in particular almost everything by Microsoft.
1933 * The loop below just changes e.g., \w into \W and vice versa */
1935 regnode_charclass_posixl temp;
1936 int add = 1; /* To calculate the index of the complement */
1938 Zero(&temp, 1, regnode_charclass_posixl);
1939 ANYOF_POSIXL_ZERO(&temp);
1940 for (i = 0; i < ANYOF_MAX; i++) {
1942 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1943 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1945 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1946 ANYOF_POSIXL_SET(&temp, i + add);
1948 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1950 ANYOF_POSIXL_AND(&temp, ssc);
1952 } /* else ssc already has no posixes */
1953 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1954 in its initial state */
1955 else if (! is_ANYOF_SYNTHETIC(and_with)
1956 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1958 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1959 * copy it over 'ssc' */
1960 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1961 if (is_ANYOF_SYNTHETIC(and_with)) {
1962 StructCopy(and_with, ssc, regnode_ssc);
1965 ssc->invlist = anded_cp_list;
1966 ANYOF_POSIXL_ZERO(ssc);
1967 if (and_with_flags & ANYOF_MATCHES_POSIXL) {
1968 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1972 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1973 || (and_with_flags & ANYOF_MATCHES_POSIXL))
1975 /* One or the other of P1, P2 is non-empty. */
1976 if (and_with_flags & ANYOF_MATCHES_POSIXL) {
1977 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1979 ssc_union(ssc, anded_cp_list, FALSE);
1981 else { /* P1 = P2 = empty */
1982 ssc_intersection(ssc, anded_cp_list, FALSE);
1988 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1989 const regnode_charclass *or_with)
1991 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1992 * another SSC or a regular ANYOF class. Can create false positives if
1993 * 'or_with' is to be inverted. */
1997 U8 or_with_flags = inRANGE(OP(or_with), ANYOFH, ANYOFRb)
1999 : ANYOF_FLAGS(or_with);
2001 PERL_ARGS_ASSERT_SSC_OR;
2003 assert(is_ANYOF_SYNTHETIC(ssc));
2005 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
2006 * the code point inversion list and just the relevant flags */
2007 if (is_ANYOF_SYNTHETIC(or_with)) {
2008 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
2009 ored_flags = or_with_flags;
2012 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
2013 ored_flags = or_with_flags & ANYOF_COMMON_FLAGS;
2014 if (OP(or_with) != ANYOFD) {
2017 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
2018 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
2019 if (ANYOFL_UTF8_LOCALE_REQD(or_with_flags)) {
2021 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
2026 ANYOF_FLAGS(ssc) |= ored_flags;
2028 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
2029 * C2 is the list of code points in 'or-with'; P2, its posix classes.
2030 * 'or_with' may be inverted. When not inverted, we have the simple
2031 * situation of computing:
2032 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
2033 * If P1|P2 yields a situation with both a class and its complement are
2034 * set, like having both \w and \W, this matches all code points, and we
2035 * can delete these from the P component of the ssc going forward. XXX We
2036 * might be able to delete all the P components, but I (khw) am not certain
2037 * about this, and it is better to be safe.
2040 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
2041 * <= (C1 | P1) | ~C2
2042 * <= (C1 | ~C2) | P1
2043 * (which results in actually simpler code than the non-inverted case)
2046 if ((or_with_flags & ANYOF_INVERT)
2047 && ! is_ANYOF_SYNTHETIC(or_with))
2049 /* We ignore P2, leaving P1 going forward */
2050 } /* else Not inverted */
2051 else if (or_with_flags & ANYOF_MATCHES_POSIXL) {
2052 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
2053 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
2055 for (i = 0; i < ANYOF_MAX; i += 2) {
2056 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
2058 ssc_match_all_cp(ssc);
2059 ANYOF_POSIXL_CLEAR(ssc, i);
2060 ANYOF_POSIXL_CLEAR(ssc, i+1);
2068 FALSE /* Already has been inverted */
2072 PERL_STATIC_INLINE void
2073 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
2075 PERL_ARGS_ASSERT_SSC_UNION;
2077 assert(is_ANYOF_SYNTHETIC(ssc));
2079 _invlist_union_maybe_complement_2nd(ssc->invlist,
2085 PERL_STATIC_INLINE void
2086 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
2088 const bool invert2nd)
2090 PERL_ARGS_ASSERT_SSC_INTERSECTION;
2092 assert(is_ANYOF_SYNTHETIC(ssc));
2094 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
2100 PERL_STATIC_INLINE void
2101 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
2103 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
2105 assert(is_ANYOF_SYNTHETIC(ssc));
2107 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
2110 PERL_STATIC_INLINE void
2111 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
2113 /* AND just the single code point 'cp' into the SSC 'ssc' */
2115 SV* cp_list = _new_invlist(2);
2117 PERL_ARGS_ASSERT_SSC_CP_AND;
2119 assert(is_ANYOF_SYNTHETIC(ssc));
2121 cp_list = add_cp_to_invlist(cp_list, cp);
2122 ssc_intersection(ssc, cp_list,
2123 FALSE /* Not inverted */
2125 SvREFCNT_dec_NN(cp_list);
2128 PERL_STATIC_INLINE void
2129 S_ssc_clear_locale(regnode_ssc *ssc)
2131 /* Set the SSC 'ssc' to not match any locale things */
2132 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
2134 assert(is_ANYOF_SYNTHETIC(ssc));
2136 ANYOF_POSIXL_ZERO(ssc);
2137 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
2140 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
2143 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
2145 /* The synthetic start class is used to hopefully quickly winnow down
2146 * places where a pattern could start a match in the target string. If it
2147 * doesn't really narrow things down that much, there isn't much point to
2148 * having the overhead of using it. This function uses some very crude
2149 * heuristics to decide if to use the ssc or not.
2151 * It returns TRUE if 'ssc' rules out more than half what it considers to
2152 * be the "likely" possible matches, but of course it doesn't know what the
2153 * actual things being matched are going to be; these are only guesses
2155 * For /l matches, it assumes that the only likely matches are going to be
2156 * in the 0-255 range, uniformly distributed, so half of that is 127
2157 * For /a and /d matches, it assumes that the likely matches will be just
2158 * the ASCII range, so half of that is 63
2159 * For /u and there isn't anything matching above the Latin1 range, it
2160 * assumes that that is the only range likely to be matched, and uses
2161 * half that as the cut-off: 127. If anything matches above Latin1,
2162 * it assumes that all of Unicode could match (uniformly), except for
2163 * non-Unicode code points and things in the General Category "Other"
2164 * (unassigned, private use, surrogates, controls and formats). This
2165 * is a much large number. */
2167 U32 count = 0; /* Running total of number of code points matched by
2169 UV start, end; /* Start and end points of current range in inversion
2170 XXX outdated. UTF-8 locales are common, what about invert? list */
2171 const U32 max_code_points = (LOC)
2173 : (( ! UNI_SEMANTICS
2174 || invlist_highest(ssc->invlist) < 256)
2177 const U32 max_match = max_code_points / 2;
2179 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
2181 invlist_iterinit(ssc->invlist);
2182 while (invlist_iternext(ssc->invlist, &start, &end)) {
2183 if (start >= max_code_points) {
2186 end = MIN(end, max_code_points - 1);
2187 count += end - start + 1;
2188 if (count >= max_match) {
2189 invlist_iterfinish(ssc->invlist);
2199 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
2201 /* The inversion list in the SSC is marked mortal; now we need a more
2202 * permanent copy, which is stored the same way that is done in a regular
2203 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
2206 SV* invlist = invlist_clone(ssc->invlist, NULL);
2208 PERL_ARGS_ASSERT_SSC_FINALIZE;
2210 assert(is_ANYOF_SYNTHETIC(ssc));
2212 /* The code in this file assumes that all but these flags aren't relevant
2213 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
2214 * by the time we reach here */
2215 assert(! (ANYOF_FLAGS(ssc)
2216 & ~( ANYOF_COMMON_FLAGS
2217 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
2218 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
2220 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
2222 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist, NULL, NULL);
2223 SvREFCNT_dec(invlist);
2225 /* Make sure is clone-safe */
2226 ssc->invlist = NULL;
2228 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
2229 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
2230 OP(ssc) = ANYOFPOSIXL;
2232 else if (RExC_contains_locale) {
2236 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
2239 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
2240 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
2241 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
2242 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
2243 ? (TRIE_LIST_CUR( idx ) - 1) \
2249 dump_trie(trie,widecharmap,revcharmap)
2250 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
2251 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
2253 These routines dump out a trie in a somewhat readable format.
2254 The _interim_ variants are used for debugging the interim
2255 tables that are used to generate the final compressed
2256 representation which is what dump_trie expects.
2258 Part of the reason for their existence is to provide a form
2259 of documentation as to how the different representations function.
2264 Dumps the final compressed table form of the trie to Perl_debug_log.
2265 Used for debugging make_trie().
2269 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
2270 AV *revcharmap, U32 depth)
2273 SV *sv=sv_newmortal();
2274 int colwidth= widecharmap ? 6 : 4;
2276 DECLARE_AND_GET_RE_DEBUG_FLAGS;
2278 PERL_ARGS_ASSERT_DUMP_TRIE;
2280 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
2281 depth+1, "Match","Base","Ofs" );
2283 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2284 SV ** const tmp = av_fetch( revcharmap, state, 0);
2286 Perl_re_printf( aTHX_ "%*s",
2288 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2289 PL_colors[0], PL_colors[1],
2290 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2291 PERL_PV_ESCAPE_FIRSTCHAR
2296 Perl_re_printf( aTHX_ "\n");
2297 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
2299 for( state = 0 ; state < trie->uniquecharcount ; state++ )
2300 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2301 Perl_re_printf( aTHX_ "\n");
2303 for( state = 1 ; state < trie->statecount ; state++ ) {
2304 const U32 base = trie->states[ state ].trans.base;
2306 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
2308 if ( trie->states[ state ].wordnum ) {
2309 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
2311 Perl_re_printf( aTHX_ "%6s", "" );
2314 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
2319 while( ( base + ofs < trie->uniquecharcount ) ||
2320 ( base + ofs - trie->uniquecharcount < trie->lasttrans
2321 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2325 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
2327 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2328 if ( ( base + ofs >= trie->uniquecharcount )
2329 && ( base + ofs - trie->uniquecharcount
2331 && trie->trans[ base + ofs
2332 - trie->uniquecharcount ].check == state )
2334 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
2335 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2338 Perl_re_printf( aTHX_ "%*s", colwidth," ." );
2342 Perl_re_printf( aTHX_ "]");
2345 Perl_re_printf( aTHX_ "\n" );
2347 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
2349 for (word=1; word <= trie->wordcount; word++) {
2350 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2351 (int)word, (int)(trie->wordinfo[word].prev),
2352 (int)(trie->wordinfo[word].len));
2354 Perl_re_printf( aTHX_ "\n" );
2357 Dumps a fully constructed but uncompressed trie in list form.
2358 List tries normally only are used for construction when the number of
2359 possible chars (trie->uniquecharcount) is very high.
2360 Used for debugging make_trie().
2363 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2364 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2368 SV *sv=sv_newmortal();
2369 int colwidth= widecharmap ? 6 : 4;
2370 DECLARE_AND_GET_RE_DEBUG_FLAGS;
2372 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2374 /* print out the table precompression. */
2375 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
2377 Perl_re_indentf( aTHX_ "%s",
2378 depth+1, "------:-----+-----------------\n" );
2380 for( state=1 ; state < next_alloc ; state ++ ) {
2383 Perl_re_indentf( aTHX_ " %4" UVXf " :",
2384 depth+1, (UV)state );
2385 if ( ! trie->states[ state ].wordnum ) {
2386 Perl_re_printf( aTHX_ "%5s| ","");
2388 Perl_re_printf( aTHX_ "W%4x| ",
2389 trie->states[ state ].wordnum
2392 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2393 SV ** const tmp = av_fetch( revcharmap,
2394 TRIE_LIST_ITEM(state, charid).forid, 0);
2396 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
2398 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2400 PL_colors[0], PL_colors[1],
2401 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2402 | PERL_PV_ESCAPE_FIRSTCHAR
2404 TRIE_LIST_ITEM(state, charid).forid,
2405 (UV)TRIE_LIST_ITEM(state, charid).newstate
2408 Perl_re_printf( aTHX_ "\n%*s| ",
2409 (int)((depth * 2) + 14), "");
2412 Perl_re_printf( aTHX_ "\n");
2417 Dumps a fully constructed but uncompressed trie in table form.
2418 This is the normal DFA style state transition table, with a few
2419 twists to facilitate compression later.
2420 Used for debugging make_trie().
2423 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2424 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2429 SV *sv=sv_newmortal();
2430 int colwidth= widecharmap ? 6 : 4;
2431 DECLARE_AND_GET_RE_DEBUG_FLAGS;
2433 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2436 print out the table precompression so that we can do a visual check
2437 that they are identical.
2440 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
2442 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2443 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2445 Perl_re_printf( aTHX_ "%*s",
2447 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2448 PL_colors[0], PL_colors[1],
2449 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2450 PERL_PV_ESCAPE_FIRSTCHAR
2456 Perl_re_printf( aTHX_ "\n");
2457 Perl_re_indentf( aTHX_ "State+-", depth+1 );
2459 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2460 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
2463 Perl_re_printf( aTHX_ "\n" );
2465 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2467 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
2469 (UV)TRIE_NODENUM( state ) );
2471 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2472 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2474 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
2476 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
2478 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2479 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
2480 (UV)trie->trans[ state ].check );
2482 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
2483 (UV)trie->trans[ state ].check,
2484 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2492 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2493 startbranch: the first branch in the whole branch sequence
2494 first : start branch of sequence of branch-exact nodes.
2495 May be the same as startbranch
2496 last : Thing following the last branch.
2497 May be the same as tail.
2498 tail : item following the branch sequence
2499 count : words in the sequence
2500 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2501 depth : indent depth
2503 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2505 A trie is an N'ary tree where the branches are determined by digital
2506 decomposition of the key. IE, at the root node you look up the 1st character and
2507 follow that branch repeat until you find the end of the branches. Nodes can be
2508 marked as "accepting" meaning they represent a complete word. Eg:
2512 would convert into the following structure. Numbers represent states, letters
2513 following numbers represent valid transitions on the letter from that state, if
2514 the number is in square brackets it represents an accepting state, otherwise it
2515 will be in parenthesis.
2517 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2521 (1) +-i->(6)-+-s->[7]
2523 +-s->(3)-+-h->(4)-+-e->[5]
2525 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2527 This shows that when matching against the string 'hers' we will begin at state 1
2528 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2529 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2530 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2531 single traverse. We store a mapping from accepting to state to which word was
2532 matched, and then when we have multiple possibilities we try to complete the
2533 rest of the regex in the order in which they occurred in the alternation.
2535 The only prior NFA like behaviour that would be changed by the TRIE support is
2536 the silent ignoring of duplicate alternations which are of the form:
2538 / (DUPE|DUPE) X? (?{ ... }) Y /x
2540 Thus EVAL blocks following a trie may be called a different number of times with
2541 and without the optimisation. With the optimisations dupes will be silently
2542 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2543 the following demonstrates:
2545 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2547 which prints out 'word' three times, but
2549 'words'=~/(word|word|word)(?{ print $1 })S/
2551 which doesnt print it out at all. This is due to other optimisations kicking in.
2553 Example of what happens on a structural level:
2555 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2557 1: CURLYM[1] {1,32767}(18)
2568 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2569 and should turn into:
2571 1: CURLYM[1] {1,32767}(18)
2573 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2581 Cases where tail != last would be like /(?foo|bar)baz/:
2591 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2592 and would end up looking like:
2595 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2602 d = uvchr_to_utf8_flags(d, uv, 0);
2604 is the recommended Unicode-aware way of saying
2609 #define TRIE_STORE_REVCHAR(val) \
2612 SV *zlopp = newSV(UTF8_MAXBYTES); \
2613 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2614 unsigned char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2616 SvCUR_set(zlopp, kapow - flrbbbbb); \
2619 av_push(revcharmap, zlopp); \
2621 char ooooff = (char)val; \
2622 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2626 /* This gets the next character from the input, folding it if not already
2628 #define TRIE_READ_CHAR STMT_START { \
2631 /* if it is UTF then it is either already folded, or does not need \
2633 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2635 else if (folder == PL_fold_latin1) { \
2636 /* This folder implies Unicode rules, which in the range expressible \
2637 * by not UTF is the lower case, with the two exceptions, one of \
2638 * which should have been taken care of before calling this */ \
2639 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2640 uvc = toLOWER_L1(*uc); \
2641 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2644 /* raw data, will be folded later if needed */ \
2652 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2653 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2654 U32 ging = TRIE_LIST_LEN( state ) * 2; \
2655 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2656 TRIE_LIST_LEN( state ) = ging; \
2658 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2659 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2660 TRIE_LIST_CUR( state )++; \
2663 #define TRIE_LIST_NEW(state) STMT_START { \
2664 Newx( trie->states[ state ].trans.list, \
2665 4, reg_trie_trans_le ); \
2666 TRIE_LIST_CUR( state ) = 1; \
2667 TRIE_LIST_LEN( state ) = 4; \
2670 #define TRIE_HANDLE_WORD(state) STMT_START { \
2671 U16 dupe= trie->states[ state ].wordnum; \
2672 regnode * const noper_next = regnext( noper ); \
2675 /* store the word for dumping */ \
2677 if (OP(noper) != NOTHING) \
2678 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2680 tmp = newSVpvn_utf8( "", 0, UTF ); \
2681 av_push( trie_words, tmp ); \
2685 trie->wordinfo[curword].prev = 0; \
2686 trie->wordinfo[curword].len = wordlen; \
2687 trie->wordinfo[curword].accept = state; \
2689 if ( noper_next < tail ) { \
2691 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2693 trie->jump[curword] = (U16)(noper_next - convert); \
2695 jumper = noper_next; \
2697 nextbranch= regnext(cur); \
2701 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2702 /* chain, so that when the bits of chain are later */\
2703 /* linked together, the dups appear in the chain */\
2704 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2705 trie->wordinfo[dupe].prev = curword; \
2707 /* we haven't inserted this word yet. */ \
2708 trie->states[ state ].wordnum = curword; \
2713 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2714 ( ( base + charid >= ucharcount \
2715 && base + charid < ubound \
2716 && state == trie->trans[ base - ucharcount + charid ].check \
2717 && trie->trans[ base - ucharcount + charid ].next ) \
2718 ? trie->trans[ base - ucharcount + charid ].next \
2719 : ( state==1 ? special : 0 ) \
2722 #define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2724 TRIE_BITMAP_SET(trie, uvc); \
2725 /* store the folded codepoint */ \
2727 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2730 /* store first byte of utf8 representation of */ \
2731 /* variant codepoints */ \
2732 if (! UVCHR_IS_INVARIANT(uvc)) { \
2733 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2738 #define MADE_JUMP_TRIE 2
2739 #define MADE_EXACT_TRIE 4
2742 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2743 regnode *first, regnode *last, regnode *tail,
2744 U32 word_count, U32 flags, U32 depth)
2746 /* first pass, loop through and scan words */
2747 reg_trie_data *trie;
2748 HV *widecharmap = NULL;
2749 AV *revcharmap = newAV();
2755 regnode *jumper = NULL;
2756 regnode *nextbranch = NULL;
2757 regnode *convert = NULL;
2758 U32 *prev_states; /* temp array mapping each state to previous one */
2759 /* we just use folder as a flag in utf8 */
2760 const U8 * folder = NULL;
2762 /* in the below add_data call we are storing either 'tu' or 'tuaa'
2763 * which stands for one trie structure, one hash, optionally followed
2766 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuaa"));
2767 AV *trie_words = NULL;
2768 /* along with revcharmap, this only used during construction but both are
2769 * useful during debugging so we store them in the struct when debugging.
2772 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2773 STRLEN trie_charcount=0;
2775 SV *re_trie_maxbuff;
2776 DECLARE_AND_GET_RE_DEBUG_FLAGS;
2778 PERL_ARGS_ASSERT_MAKE_TRIE;
2780 PERL_UNUSED_ARG(depth);
2784 case EXACT: case EXACT_REQ8: case EXACTL: break;
2788 case EXACTFLU8: folder = PL_fold_latin1; break;
2789 case EXACTF: folder = PL_fold; break;
2790 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2793 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2795 trie->startstate = 1;
2796 trie->wordcount = word_count;
2797 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2798 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2799 if (flags == EXACT || flags == EXACT_REQ8 || flags == EXACTL)
2800 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2801 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2802 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2805 trie_words = newAV();
2808 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, GV_ADD);
2809 assert(re_trie_maxbuff);
2810 if (!SvIOK(re_trie_maxbuff)) {
2811 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2813 DEBUG_TRIE_COMPILE_r({
2814 Perl_re_indentf( aTHX_
2815 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2817 REG_NODE_NUM(startbranch), REG_NODE_NUM(first),
2818 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2821 /* Find the node we are going to overwrite */
2822 if ( first == startbranch && OP( last ) != BRANCH ) {
2823 /* whole branch chain */
2826 /* branch sub-chain */
2827 convert = NEXTOPER( first );
2830 /* -- First loop and Setup --
2832 We first traverse the branches and scan each word to determine if it
2833 contains widechars, and how many unique chars there are, this is
2834 important as we have to build a table with at least as many columns as we
2837 We use an array of integers to represent the character codes 0..255
2838 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2839 the native representation of the character value as the key and IV's for
2842 *TODO* If we keep track of how many times each character is used we can
2843 remap the columns so that the table compression later on is more
2844 efficient in terms of memory by ensuring the most common value is in the
2845 middle and the least common are on the outside. IMO this would be better
2846 than a most to least common mapping as theres a decent chance the most
2847 common letter will share a node with the least common, meaning the node
2848 will not be compressible. With a middle is most common approach the worst
2849 case is when we have the least common nodes twice.
2853 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2854 regnode *noper = NEXTOPER( cur );
2858 U32 wordlen = 0; /* required init */
2859 STRLEN minchars = 0;
2860 STRLEN maxchars = 0;
2861 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2864 if (OP(noper) == NOTHING) {
2865 /* skip past a NOTHING at the start of an alternation
2866 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2868 * If the next node is not something we are supposed to process
2869 * we will just ignore it due to the condition guarding the
2873 regnode *noper_next= regnext(noper);
2874 if (noper_next < tail)
2879 && ( OP(noper) == flags
2880 || (flags == EXACT && OP(noper) == EXACT_REQ8)
2881 || (flags == EXACTFU && ( OP(noper) == EXACTFU_REQ8
2882 || OP(noper) == EXACTFUP))))
2884 uc= (U8*)STRING(noper);
2885 e= uc + STR_LEN(noper);
2892 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2893 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2894 regardless of encoding */
2895 if (OP( noper ) == EXACTFUP) {
2896 /* false positives are ok, so just set this */
2897 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2901 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2903 TRIE_CHARCOUNT(trie)++;
2906 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2907 * is in effect. Under /i, this character can match itself, or
2908 * anything that folds to it. If not under /i, it can match just
2909 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2910 * all fold to k, and all are single characters. But some folds
2911 * expand to more than one character, so for example LATIN SMALL
2912 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2913 * the string beginning at 'uc' is 'ffi', it could be matched by
2914 * three characters, or just by the one ligature character. (It
2915 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2916 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2917 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2918 * match.) The trie needs to know the minimum and maximum number
2919 * of characters that could match so that it can use size alone to
2920 * quickly reject many match attempts. The max is simple: it is
2921 * the number of folded characters in this branch (since a fold is
2922 * never shorter than what folds to it. */
2926 /* And the min is equal to the max if not under /i (indicated by
2927 * 'folder' being NULL), or there are no multi-character folds. If
2928 * there is a multi-character fold, the min is incremented just
2929 * once, for the character that folds to the sequence. Each
2930 * character in the sequence needs to be added to the list below of
2931 * characters in the trie, but we count only the first towards the
2932 * min number of characters needed. This is done through the
2933 * variable 'foldlen', which is returned by the macros that look
2934 * for these sequences as the number of bytes the sequence
2935 * occupies. Each time through the loop, we decrement 'foldlen' by
2936 * how many bytes the current char occupies. Only when it reaches
2937 * 0 do we increment 'minchars' or look for another multi-character
2939 if (folder == NULL) {
2942 else if (foldlen > 0) {
2943 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2948 /* See if *uc is the beginning of a multi-character fold. If
2949 * so, we decrement the length remaining to look at, to account
2950 * for the current character this iteration. (We can use 'uc'
2951 * instead of the fold returned by TRIE_READ_CHAR because for
2952 * non-UTF, the latin1_safe macro is smart enough to account
2953 * for all the unfolded characters, and because for UTF, the
2954 * string will already have been folded earlier in the
2955 * compilation process */
2957 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2958 foldlen -= UTF8SKIP(uc);
2961 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2966 /* The current character (and any potential folds) should be added
2967 * to the possible matching characters for this position in this
2971 U8 folded= folder[ (U8) uvc ];
2972 if ( !trie->charmap[ folded ] ) {
2973 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2974 TRIE_STORE_REVCHAR( folded );
2977 if ( !trie->charmap[ uvc ] ) {
2978 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2979 TRIE_STORE_REVCHAR( uvc );
2982 /* store the codepoint in the bitmap, and its folded
2984 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
2985 set_bit = 0; /* We've done our bit :-) */
2989 /* XXX We could come up with the list of code points that fold
2990 * to this using PL_utf8_foldclosures, except not for
2991 * multi-char folds, as there may be multiple combinations
2992 * there that could work, which needs to wait until runtime to
2993 * resolve (The comment about LIGATURE FFI above is such an
2998 widecharmap = newHV();
3000 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
3003 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
3005 if ( !SvTRUE( *svpp ) ) {
3006 sv_setiv( *svpp, ++trie->uniquecharcount );
3007 TRIE_STORE_REVCHAR(uvc);
3010 } /* end loop through characters in this branch of the trie */
3012 /* We take the min and max for this branch and combine to find the min
3013 * and max for all branches processed so far */
3014 if( cur == first ) {
3015 trie->minlen = minchars;
3016 trie->maxlen = maxchars;
3017 } else if (minchars < trie->minlen) {
3018 trie->minlen = minchars;
3019 } else if (maxchars > trie->maxlen) {
3020 trie->maxlen = maxchars;
3022 } /* end first pass */
3023 DEBUG_TRIE_COMPILE_r(
3024 Perl_re_indentf( aTHX_
3025 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
3027 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
3028 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
3029 (int)trie->minlen, (int)trie->maxlen )
3033 We now know what we are dealing with in terms of unique chars and
3034 string sizes so we can calculate how much memory a naive
3035 representation using a flat table will take. If it's over a reasonable
3036 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
3037 conservative but potentially much slower representation using an array
3040 At the end we convert both representations into the same compressed
3041 form that will be used in regexec.c for matching with. The latter
3042 is a form that cannot be used to construct with but has memory
3043 properties similar to the list form and access properties similar
3044 to the table form making it both suitable for fast searches and
3045 small enough that its feasable to store for the duration of a program.
3047 See the comment in the code where the compressed table is produced
3048 inplace from the flat tabe representation for an explanation of how
3049 the compression works.
3054 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
3057 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
3058 > SvIV(re_trie_maxbuff) )
3061 Second Pass -- Array Of Lists Representation
3063 Each state will be represented by a list of charid:state records
3064 (reg_trie_trans_le) the first such element holds the CUR and LEN
3065 points of the allocated array. (See defines above).
3067 We build the initial structure using the lists, and then convert
3068 it into the compressed table form which allows faster lookups
3069 (but cant be modified once converted).
3072 STRLEN transcount = 1;
3074 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
3077 trie->states = (reg_trie_state *)
3078 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3079 sizeof(reg_trie_state) );
3083 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3085 regnode *noper = NEXTOPER( cur );
3086 U32 state = 1; /* required init */
3087 U16 charid = 0; /* sanity init */
3088 U32 wordlen = 0; /* required init */
3090 if (OP(noper) == NOTHING) {
3091 regnode *noper_next= regnext(noper);
3092 if (noper_next < tail)
3094 /* we will undo this assignment if noper does not
3095 * point at a trieable type in the else clause of
3096 * the following statement. */
3100 && ( OP(noper) == flags
3101 || (flags == EXACT && OP(noper) == EXACT_REQ8)
3102 || (flags == EXACTFU && ( OP(noper) == EXACTFU_REQ8
3103 || OP(noper) == EXACTFUP))))
3105 const U8 *uc= (U8*)STRING(noper);
3106 const U8 *e= uc + STR_LEN(noper);
3108 for ( ; uc < e ; uc += len ) {
3113 charid = trie->charmap[ uvc ];
3115 SV** const svpp = hv_fetch( widecharmap,
3122 charid=(U16)SvIV( *svpp );
3125 /* charid is now 0 if we dont know the char read, or
3126 * nonzero if we do */
3133 if ( !trie->states[ state ].trans.list ) {
3134 TRIE_LIST_NEW( state );
3137 check <= TRIE_LIST_USED( state );
3140 if ( TRIE_LIST_ITEM( state, check ).forid
3143 newstate = TRIE_LIST_ITEM( state, check ).newstate;
3148 newstate = next_alloc++;
3149 prev_states[newstate] = state;
3150 TRIE_LIST_PUSH( state, charid, newstate );
3155 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3159 /* If we end up here it is because we skipped past a NOTHING, but did not end up
3160 * on a trieable type. So we need to reset noper back to point at the first regop
3161 * in the branch before we call TRIE_HANDLE_WORD()
3163 noper= NEXTOPER(cur);
3165 TRIE_HANDLE_WORD(state);
3167 } /* end second pass */
3169 /* next alloc is the NEXT state to be allocated */
3170 trie->statecount = next_alloc;
3171 trie->states = (reg_trie_state *)
3172 PerlMemShared_realloc( trie->states,
3174 * sizeof(reg_trie_state) );
3176 /* and now dump it out before we compress it */
3177 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
3178 revcharmap, next_alloc,
3182 trie->trans = (reg_trie_trans *)
3183 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
3190 for( state=1 ; state < next_alloc ; state ++ ) {
3194 DEBUG_TRIE_COMPILE_MORE_r(
3195 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
3199 if (trie->states[state].trans.list) {
3200 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
3204 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3205 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
3206 if ( forid < minid ) {
3208 } else if ( forid > maxid ) {
3212 if ( transcount < tp + maxid - minid + 1) {
3214 trie->trans = (reg_trie_trans *)
3215 PerlMemShared_realloc( trie->trans,
3217 * sizeof(reg_trie_trans) );
3218 Zero( trie->trans + (transcount / 2),
3222 base = trie->uniquecharcount + tp - minid;
3223 if ( maxid == minid ) {
3225 for ( ; zp < tp ; zp++ ) {
3226 if ( ! trie->trans[ zp ].next ) {
3227 base = trie->uniquecharcount + zp - minid;
3228 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
3230 trie->trans[ zp ].check = state;
3236 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
3238 trie->trans[ tp ].check = state;
3243 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3244 const U32 tid = base
3245 - trie->uniquecharcount
3246 + TRIE_LIST_ITEM( state, idx ).forid;
3247 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
3249 trie->trans[ tid ].check = state;
3251 tp += ( maxid - minid + 1 );
3253 Safefree(trie->states[ state ].trans.list);
3256 DEBUG_TRIE_COMPILE_MORE_r(
3257 Perl_re_printf( aTHX_ " base: %d\n",base);
3260 trie->states[ state ].trans.base=base;
3262 trie->lasttrans = tp + 1;
3266 Second Pass -- Flat Table Representation.
3268 we dont use the 0 slot of either trans[] or states[] so we add 1 to
3269 each. We know that we will need Charcount+1 trans at most to store
3270 the data (one row per char at worst case) So we preallocate both
3271 structures assuming worst case.
3273 We then construct the trie using only the .next slots of the entry
3276 We use the .check field of the first entry of the node temporarily
3277 to make compression both faster and easier by keeping track of how
3278 many non zero fields are in the node.
3280 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
3283 There are two terms at use here: state as a TRIE_NODEIDX() which is
3284 a number representing the first entry of the node, and state as a
3285 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
3286 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
3287 if there are 2 entrys per node. eg:
3295 The table is internally in the right hand, idx form. However as we
3296 also have to deal with the states array which is indexed by nodenum
3297 we have to use TRIE_NODENUM() to convert.
3300 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
3303 trie->trans = (reg_trie_trans *)
3304 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
3305 * trie->uniquecharcount + 1,
3306 sizeof(reg_trie_trans) );
3307 trie->states = (reg_trie_state *)
3308 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3309 sizeof(reg_trie_state) );
3310 next_alloc = trie->uniquecharcount + 1;
3313 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3315 regnode *noper = NEXTOPER( cur );
3317 U32 state = 1; /* required init */
3319 U16 charid = 0; /* sanity init */
3320 U32 accept_state = 0; /* sanity init */
3322 U32 wordlen = 0; /* required init */
3324 if (OP(noper) == NOTHING) {
3325 regnode *noper_next= regnext(noper);
3326 if (noper_next < tail)
3328 /* we will undo this assignment if noper does not
3329 * point at a trieable type in the else clause of
3330 * the following statement. */
3334 && ( OP(noper) == flags
3335 || (flags == EXACT && OP(noper) == EXACT_REQ8)
3336 || (flags == EXACTFU && ( OP(noper) == EXACTFU_REQ8
3337 || OP(noper) == EXACTFUP))))
3339 const U8 *uc= (U8*)STRING(noper);
3340 const U8 *e= uc + STR_LEN(noper);
3342 for ( ; uc < e ; uc += len ) {
3347 charid = trie->charmap[ uvc ];
3349 SV* const * const svpp = hv_fetch( widecharmap,
3353 charid = svpp ? (U16)SvIV(*svpp) : 0;
3357 if ( !trie->trans[ state + charid ].next ) {
3358 trie->trans[ state + charid ].next = next_alloc;
3359 trie->trans[ state ].check++;
3360 prev_states[TRIE_NODENUM(next_alloc)]
3361 = TRIE_NODENUM(state);
3362 next_alloc += trie->uniquecharcount;
3364 state = trie->trans[ state + charid ].next;
3366 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3368 /* charid is now 0 if we dont know the char read, or
3369 * nonzero if we do */
3372 /* If we end up here it is because we skipped past a NOTHING, but did not end up
3373 * on a trieable type. So we need to reset noper back to point at the first regop
3374 * in the branch before we call TRIE_HANDLE_WORD().
3376 noper= NEXTOPER(cur);
3378 accept_state = TRIE_NODENUM( state );
3379 TRIE_HANDLE_WORD(accept_state);
3381 } /* end second pass */
3383 /* and now dump it out before we compress it */
3384 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3386 next_alloc, depth+1));
3390 * Inplace compress the table.*
3392 For sparse data sets the table constructed by the trie algorithm will
3393 be mostly 0/FAIL transitions or to put it another way mostly empty.
3394 (Note that leaf nodes will not contain any transitions.)
3396 This algorithm compresses the tables by eliminating most such
3397 transitions, at the cost of a modest bit of extra work during lookup:
3399 - Each states[] entry contains a .base field which indicates the
3400 index in the state[] array wheres its transition data is stored.
3402 - If .base is 0 there are no valid transitions from that node.
3404 - If .base is nonzero then charid is added to it to find an entry in
3407 -If trans[states[state].base+charid].check!=state then the
3408 transition is taken to be a 0/Fail transition. Thus if there are fail
3409 transitions at the front of the node then the .base offset will point
3410 somewhere inside the previous nodes data (or maybe even into a node
3411 even earlier), but the .check field determines if the transition is
3415 The following process inplace converts the table to the compressed
3416 table: We first do not compress the root node 1,and mark all its
3417 .check pointers as 1 and set its .base pointer as 1 as well. This
3418 allows us to do a DFA construction from the compressed table later,
3419 and ensures that any .base pointers we calculate later are greater
3422 - We set 'pos' to indicate the first entry of the second node.
3424 - We then iterate over the columns of the node, finding the first and
3425 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3426 and set the .check pointers accordingly, and advance pos
3427 appropriately and repreat for the next node. Note that when we copy
3428 the next pointers we have to convert them from the original
3429 NODEIDX form to NODENUM form as the former is not valid post
3432 - If a node has no transitions used we mark its base as 0 and do not
3433 advance the pos pointer.
3435 - If a node only has one transition we use a second pointer into the
3436 structure to fill in allocated fail transitions from other states.
3437 This pointer is independent of the main pointer and scans forward
3438 looking for null transitions that are allocated to a state. When it
3439 finds one it writes the single transition into the "hole". If the
3440 pointer doesnt find one the single transition is appended as normal.
3442 - Once compressed we can Renew/realloc the structures to release the
3445 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3446 specifically Fig 3.47 and the associated pseudocode.
3450 const U32 laststate = TRIE_NODENUM( next_alloc );
3453 trie->statecount = laststate;
3455 for ( state = 1 ; state < laststate ; state++ ) {
3457 const U32 stateidx = TRIE_NODEIDX( state );
3458 const U32 o_used = trie->trans[ stateidx ].check;
3459 U32 used = trie->trans[ stateidx ].check;
3460 trie->trans[ stateidx ].check = 0;
3463 used && charid < trie->uniquecharcount;
3466 if ( flag || trie->trans[ stateidx + charid ].next ) {
3467 if ( trie->trans[ stateidx + charid ].next ) {
3469 for ( ; zp < pos ; zp++ ) {
3470 if ( ! trie->trans[ zp ].next ) {
3474 trie->states[ state ].trans.base
3476 + trie->uniquecharcount
3478 trie->trans[ zp ].next
3479 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3481 trie->trans[ zp ].check = state;
3482 if ( ++zp > pos ) pos = zp;
3489 trie->states[ state ].trans.base
3490 = pos + trie->uniquecharcount - charid ;
3492 trie->trans[ pos ].next
3493 = SAFE_TRIE_NODENUM(
3494 trie->trans[ stateidx + charid ].next );
3495 trie->trans[ pos ].check = state;
3500 trie->lasttrans = pos + 1;
3501 trie->states = (reg_trie_state *)
3502 PerlMemShared_realloc( trie->states, laststate
3503 * sizeof(reg_trie_state) );
3504 DEBUG_TRIE_COMPILE_MORE_r(
3505 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
3507 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3511 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3514 } /* end table compress */
3516 DEBUG_TRIE_COMPILE_MORE_r(
3517 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
3519 (UV)trie->statecount,
3520 (UV)trie->lasttrans)
3522 /* resize the trans array to remove unused space */
3523 trie->trans = (reg_trie_trans *)
3524 PerlMemShared_realloc( trie->trans, trie->lasttrans
3525 * sizeof(reg_trie_trans) );
3527 { /* Modify the program and insert the new TRIE node */
3528 U8 nodetype =(U8)(flags & 0xFF);
3532 regnode *optimize = NULL;
3533 #ifdef RE_TRACK_PATTERN_OFFSETS
3536 U32 mjd_nodelen = 0;
3537 #endif /* RE_TRACK_PATTERN_OFFSETS */
3538 #endif /* DEBUGGING */
3540 This means we convert either the first branch or the first Exact,
3541 depending on whether the thing following (in 'last') is a branch
3542 or not and whther first is the startbranch (ie is it a sub part of
3543 the alternation or is it the whole thing.)
3544 Assuming its a sub part we convert the EXACT otherwise we convert
3545 the whole branch sequence, including the first.
3547 /* Find the node we are going to overwrite */
3548 if ( first != startbranch || OP( last ) == BRANCH ) {
3549 /* branch sub-chain */
3550 NEXT_OFF( first ) = (U16)(last - first);
3551 #ifdef RE_TRACK_PATTERN_OFFSETS
3553 mjd_offset= Node_Offset((convert));
3554 mjd_nodelen= Node_Length((convert));
3557 /* whole branch chain */
3559 #ifdef RE_TRACK_PATTERN_OFFSETS
3562 const regnode *nop = NEXTOPER( convert );
3563 mjd_offset= Node_Offset((nop));
3564 mjd_nodelen= Node_Length((nop));
3568 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
3570 (UV)mjd_offset, (UV)mjd_nodelen)
3573 /* But first we check to see if there is a common prefix we can
3574 split out as an EXACT and put in front of the TRIE node. */
3575 trie->startstate= 1;
3576 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3577 /* we want to find the first state that has more than
3578 * one transition, if that state is not the first state
3579 * then we have a common prefix which we can remove.
3582 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3584 I32 first_ofs = -1; /* keeps track of the ofs of the first
3585 transition, -1 means none */
3587 const U32 base = trie->states[ state ].trans.base;
3589 /* does this state terminate an alternation? */
3590 if ( trie->states[state].wordnum )
3593 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3594 if ( ( base + ofs >= trie->uniquecharcount ) &&
3595 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3596 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3598 if ( ++count > 1 ) {
3599 /* we have more than one transition */
3602 /* if this is the first state there is no common prefix
3603 * to extract, so we can exit */
3604 if ( state == 1 ) break;
3605 tmp = av_fetch( revcharmap, ofs, 0);
3606 ch = (U8*)SvPV_nolen_const( *tmp );
3608 /* if we are on count 2 then we need to initialize the
3609 * bitmap, and store the previous char if there was one
3612 /* clear the bitmap */
3613 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3615 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
3618 if (first_ofs >= 0) {
3619 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
3620 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3622 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3624 Perl_re_printf( aTHX_ "%s", (char*)ch)
3628 /* store the current firstchar in the bitmap */
3629 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3630 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
3636 /* This state has only one transition, its transition is part
3637 * of a common prefix - we need to concatenate the char it
3638 * represents to what we have so far. */
3639 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
3641 char *ch = SvPV( *tmp, len );
3643 SV *sv=sv_newmortal();
3644 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
3646 (UV)state, (UV)first_ofs,
3647 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3648 PL_colors[0], PL_colors[1],
3649 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3650 PERL_PV_ESCAPE_FIRSTCHAR
3655 OP( convert ) = nodetype;
3656 str=STRING(convert);
3657 setSTR_LEN(convert, 0);
3659 assert( ( STR_LEN(convert) + len ) < 256 );
3660 setSTR_LEN(convert, (U8)(STR_LEN(convert) + len));
3666 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
3671 trie->prefixlen = (state-1);
3673 regnode *n = convert+NODE_SZ_STR(convert);
3674 assert( NODE_SZ_STR(convert) <= U16_MAX );
3675 NEXT_OFF(convert) = (U16)(NODE_SZ_STR(convert));
3676 trie->startstate = state;
3677 trie->minlen -= (state - 1);
3678 trie->maxlen -= (state - 1);
3680 /* At least the UNICOS C compiler choked on this
3681 * being argument to DEBUG_r(), so let's just have
3684 #ifdef PERL_EXT_RE_BUILD
3690 regnode *fix = convert;
3691 U32 word = trie->wordcount;
3692 #ifdef RE_TRACK_PATTERN_OFFSETS
3695 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3696 while( ++fix < n ) {
3697 Set_Node_Offset_Length(fix, 0, 0);
3700 SV ** const tmp = av_fetch( trie_words, word, 0 );
3702 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3703 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3705 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3713 NEXT_OFF(convert) = (U16)(tail - convert);
3714 DEBUG_r(optimize= n);
3720 if ( trie->maxlen ) {
3721 NEXT_OFF( convert ) = (U16)(tail - convert);
3722 ARG_SET( convert, data_slot );
3723 /* Store the offset to the first unabsorbed branch in
3724 jump[0], which is otherwise unused by the jump logic.
3725 We use this when dumping a trie and during optimisation. */
3727 trie->jump[0] = (U16)(nextbranch - convert);
3729 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3730 * and there is a bitmap
3731 * and the first "jump target" node we found leaves enough room
3732 * then convert the TRIE node into a TRIEC node, with the bitmap
3733 * embedded inline in the opcode - this is hypothetically faster.
3735 if ( !trie->states[trie->startstate].wordnum
3737 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3739 OP( convert ) = TRIEC;
3740 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3741 PerlMemShared_free(trie->bitmap);
3744 OP( convert ) = TRIE;
3746 /* store the type in the flags */
3747 convert->flags = nodetype;
3751 + regarglen[ OP( convert ) ];
3753 /* XXX We really should free up the resource in trie now,
3754 as we won't use them - (which resources?) dmq */
3756 /* needed for dumping*/
3757 DEBUG_r(if (optimize) {
3758 regnode *opt = convert;
3760 while ( ++opt < optimize) {
3761 Set_Node_Offset_Length(opt, 0, 0);
3764 Try to clean up some of the debris left after the
3767 while( optimize < jumper ) {
3768 Track_Code( mjd_nodelen += Node_Length((optimize)); );
3769 OP( optimize ) = OPTIMIZED;
3770 Set_Node_Offset_Length(optimize, 0, 0);
3773 Set_Node_Offset_Length(convert, mjd_offset, mjd_nodelen);
3775 } /* end node insert */
3777 /* Finish populating the prev field of the wordinfo array. Walk back
3778 * from each accept state until we find another accept state, and if
3779 * so, point the first word's .prev field at the second word. If the
3780 * second already has a .prev field set, stop now. This will be the
3781 * case either if we've already processed that word's accept state,
3782 * or that state had multiple words, and the overspill words were
3783 * already linked up earlier.
3790 for (word=1; word <= trie->wordcount; word++) {
3792 if (trie->wordinfo[word].prev)
3794 state = trie->wordinfo[word].accept;
3796 state = prev_states[state];
3799 prev = trie->states[state].wordnum;
3803 trie->wordinfo[word].prev = prev;
3805 Safefree(prev_states);
3809 /* and now dump out the compressed format */
3810 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3812 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3814 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3815 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3817 SvREFCNT_dec_NN(revcharmap);
3821 : trie->startstate>1
3827 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3829 /* The Trie is constructed and compressed now so we can build a fail array if
3832 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3834 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3838 We find the fail state for each state in the trie, this state is the longest
3839 proper suffix of the current state's 'word' that is also a proper prefix of
3840 another word in our trie. State 1 represents the word '' and is thus the
3841 default fail state. This allows the DFA not to have to restart after its
3842 tried and failed a word at a given point, it simply continues as though it
3843 had been matching the other word in the first place.
3845 'abcdgu'=~/abcdefg|cdgu/
3846 When we get to 'd' we are still matching the first word, we would encounter
3847 'g' which would fail, which would bring us to the state representing 'd' in
3848 the second word where we would try 'g' and succeed, proceeding to match
3851 /* add a fail transition */
3852 const U32 trie_offset = ARG(source);
3853 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3855 const U32 ucharcount = trie->uniquecharcount;
3856 const U32 numstates = trie->statecount;
3857 const U32 ubound = trie->lasttrans + ucharcount;
3861 U32 base = trie->states[ 1 ].trans.base;
3864 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3866 DECLARE_AND_GET_RE_DEBUG_FLAGS;
3868 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3869 PERL_UNUSED_CONTEXT;
3871 PERL_UNUSED_ARG(depth);
3874 if ( OP(source) == TRIE ) {
3875 struct regnode_1 *op = (struct regnode_1 *)
3876 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3877 StructCopy(source, op, struct regnode_1);
3878 stclass = (regnode *)op;
3880 struct regnode_charclass *op = (struct regnode_charclass *)
3881 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3882 StructCopy(source, op, struct regnode_charclass);
3883 stclass = (regnode *)op;
3885 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3887 ARG_SET( stclass, data_slot );
3888 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3889 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3890 aho->trie=trie_offset;
3891 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3892 Copy( trie->states, aho->states, numstates, reg_trie_state );
3893 Newx( q, numstates, U32);
3894 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3897 /* initialize fail[0..1] to be 1 so that we always have
3898 a valid final fail state */
3899 fail[ 0 ] = fail[ 1 ] = 1;
3901 for ( charid = 0; charid < ucharcount ; charid++ ) {
3902 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3904 q[ q_write ] = newstate;
3905 /* set to point at the root */
3906 fail[ q[ q_write++ ] ]=1;
3909 while ( q_read < q_write) {
3910 const U32 cur = q[ q_read++ % numstates ];
3911 base = trie->states[ cur ].trans.base;
3913 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3914 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3916 U32 fail_state = cur;
3919 fail_state = fail[ fail_state ];
3920 fail_base = aho->states[ fail_state ].trans.base;
3921 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3923 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3924 fail[ ch_state ] = fail_state;
3925 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3927 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3929 q[ q_write++ % numstates] = ch_state;
3933 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3934 when we fail in state 1, this allows us to use the
3935 charclass scan to find a valid start char. This is based on the principle
3936 that theres a good chance the string being searched contains lots of stuff
3937 that cant be a start char.
3939 fail[ 0 ] = fail[ 1 ] = 0;
3940 DEBUG_TRIE_COMPILE_r({
3941 Perl_re_indentf( aTHX_ "Stclass Failtable (%" UVuf " states): 0",
3942 depth, (UV)numstates
3944 for( q_read=1; q_read<numstates; q_read++ ) {
3945 Perl_re_printf( aTHX_ ", %" UVuf, (UV)fail[q_read]);
3947 Perl_re_printf( aTHX_ "\n");
3950 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3955 /* The below joins as many adjacent EXACTish nodes as possible into a single
3956 * one. The regop may be changed if the node(s) contain certain sequences that
3957 * require special handling. The joining is only done if:
3958 * 1) there is room in the current conglomerated node to entirely contain the
3960 * 2) they are compatible node types
3962 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3963 * these get optimized out
3965 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3966 * as possible, even if that means splitting an existing node so that its first
3967 * part is moved to the preceeding node. This would maximise the efficiency of
3968 * memEQ during matching.
3970 * If a node is to match under /i (folded), the number of characters it matches
3971 * can be different than its character length if it contains a multi-character
3972 * fold. *min_subtract is set to the total delta number of characters of the
3975 * And *unfolded_multi_char is set to indicate whether or not the node contains
3976 * an unfolded multi-char fold. This happens when it won't be known until
3977 * runtime whether the fold is valid or not; namely
3978 * 1) for EXACTF nodes that contain LATIN SMALL LETTER SHARP S, as only if the
3979 * target string being matched against turns out to be UTF-8 is that fold
3981 * 2) for EXACTFL nodes whose folding rules depend on the locale in force at
3983 * (Multi-char folds whose components are all above the Latin1 range are not
3984 * run-time locale dependent, and have already been folded by the time this
3985 * function is called.)
3987 * This is as good a place as any to discuss the design of handling these
3988 * multi-character fold sequences. It's been wrong in Perl for a very long
3989 * time. There are three code points in Unicode whose multi-character folds
3990 * were long ago discovered to mess things up. The previous designs for
3991 * dealing with these involved assigning a special node for them. This
3992 * approach doesn't always work, as evidenced by this example:
3993 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3994 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3995 * would match just the \xDF, it won't be able to handle the case where a
3996 * successful match would have to cross the node's boundary. The new approach
3997 * that hopefully generally solves the problem generates an EXACTFUP node
3998 * that is "sss" in this case.
4000 * It turns out that there are problems with all multi-character folds, and not
4001 * just these three. Now the code is general, for all such cases. The
4002 * approach taken is:
4003 * 1) This routine examines each EXACTFish node that could contain multi-
4004 * character folded sequences. Since a single character can fold into
4005 * such a sequence, the minimum match length for this node is less than
4006 * the number of characters in the node. This routine returns in
4007 * *min_subtract how many characters to subtract from the the actual
4008 * length of the string to get a real minimum match length; it is 0 if
4009 * there are no multi-char foldeds. This delta is used by the caller to
4010 * adjust the min length of the match, and the delta between min and max,
4011 * so that the optimizer doesn't reject these possibilities based on size
4014 * 2) For the sequence involving the LATIN SMALL LETTER SHARP S (U+00DF)
4015 * under /u, we fold it to 'ss' in regatom(), and in this routine, after
4016 * joining, we scan for occurrences of the sequence 'ss' in non-UTF-8
4017 * EXACTFU nodes. The node type of such nodes is then changed to
4018 * EXACTFUP, indicating it is problematic, and needs careful handling.
4019 * (The procedures in step 1) above are sufficient to handle this case in
4020 * UTF-8 encoded nodes.) The reason this is problematic is that this is
4021 * the only case where there is a possible fold length change in non-UTF-8
4022 * patterns. By reserving a special node type for problematic cases, the
4023 * far more common regular EXACTFU nodes can be processed faster.
4024 * regexec.c takes advantage of this.
4026 * EXACTFUP has been created as a grab-bag for (hopefully uncommon)
4027 * problematic cases. These all only occur when the pattern is not
4028 * UTF-8. In addition to the 'ss' sequence where there is a possible fold
4029 * length change, it handles the situation where the string cannot be
4030 * entirely folded. The strings in an EXACTFish node are folded as much
4031 * as possible during compilation in regcomp.c. This saves effort in
4032 * regex matching. By using an EXACTFUP node when it is not possible to
4033 * fully fold at compile time, regexec.c can know that everything in an
4034 * EXACTFU node is folded, so folding can be skipped at runtime. The only
4035 * case where folding in EXACTFU nodes can't be done at compile time is
4036 * the presumably uncommon MICRO SIGN, when the pattern isn't UTF-8. This
4037 * is because its fold requires UTF-8 to represent. Thus EXACTFUP nodes
4038 * handle two very different cases. Alternatively, there could have been
4039 * a node type where there are length changes, one for unfolded, and one
4040 * for both. If yet another special case needed to be created, the number
4041 * of required node types would have to go to 7. khw figures that even
4042 * though there are plenty of node types to spare, that the maintenance
4043 * cost wasn't worth the small speedup of doing it that way, especially
4044 * since he thinks the MICRO SIGN is rarely encountered in practice.
4046 * There are other cases where folding isn't done at compile time, but
4047 * none of them are under /u, and hence not for EXACTFU nodes. The folds
4048 * in EXACTFL nodes aren't known until runtime, and vary as the locale
4049 * changes. Some folds in EXACTF depend on if the runtime target string
4050 * is UTF-8 or not. (regatom() will create an EXACTFU node even under /di
4051 * when no fold in it depends on the UTF-8ness of the target string.)
4053 * 3) A problem remains for unfolded multi-char folds. (These occur when the
4054 * validity of the fold won't be known until runtime, and so must remain
4055 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFAA
4056 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
4057 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
4058 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
4059 * The reason this is a problem is that the optimizer part of regexec.c
4060 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
4061 * that a character in the pattern corresponds to at most a single
4062 * character in the target string. (And I do mean character, and not byte
4063 * here, unlike other parts of the documentation that have never been
4064 * updated to account for multibyte Unicode.) Sharp s in EXACTF and
4065 * EXACTFL nodes can match the two character string 'ss'; in EXACTFAA
4066 * nodes it can match "\x{17F}\x{17F}". These, along with other ones in
4067 * EXACTFL nodes, violate the assumption, and they are the only instances
4068 * where it is violated. I'm reluctant to try to change the assumption,
4069 * as the code involved is impenetrable to me (khw), so instead the code
4070 * here punts. This routine examines EXACTFL nodes, and (when the pattern
4071 * isn't UTF-8) EXACTF and EXACTFAA for such unfolded folds, and returns a
4072 * boolean indicating whether or not the node contains such a fold. When
4073 * it is true, the caller sets a flag that later causes the optimizer in
4074 * this file to not set values for the floating and fixed string lengths,
4075 * and thus avoids the optimizer code in regexec.c that makes the invalid
4076 * assumption. Thus, there is no optimization based on string lengths for
4077 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
4078 * EXACTF and EXACTFAA nodes that contain the sharp s. (The reason the
4079 * assumption is wrong only in these cases is that all other non-UTF-8
4080 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
4081 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
4082 * EXACTF nodes because we don't know at compile time if it actually
4083 * matches 'ss' or not. For EXACTF nodes it will match iff the target
4084 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
4085 * always matches; and EXACTFAA where it never does. In an EXACTFAA node
4086 * in a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
4087 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
4088 * string would require the pattern to be forced into UTF-8, the overhead
4089 * of which we want to avoid. Similarly the unfolded multi-char folds in
4090 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
4093 * Similarly, the code that generates tries doesn't currently handle
4094 * not-already-folded multi-char folds, and it looks like a pain to change
4095 * that. Therefore, trie generation of EXACTFAA nodes with the sharp s
4096 * doesn't work. Instead, such an EXACTFAA is turned into a new regnode,
4097 * EXACTFAA_NO_TRIE, which the trie code knows not to handle. Most people
4098 * using /iaa matching will be doing so almost entirely with ASCII
4099 * strings, so this should rarely be encountered in practice */
4102 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
4103 UV *min_subtract, bool *unfolded_multi_char,
4104 U32 flags, regnode *val, U32 depth)
4106 /* Merge several consecutive EXACTish nodes into one. */
4108 regnode *n = regnext(scan);
4110 regnode *next = scan + NODE_SZ_STR(scan);
4114 regnode *stop = scan;
4115 DECLARE_AND_GET_RE_DEBUG_FLAGS;
4117 PERL_UNUSED_ARG(depth);
4120 PERL_ARGS_ASSERT_JOIN_EXACT;
4121 #ifndef EXPERIMENTAL_INPLACESCAN
4122 PERL_UNUSED_ARG(flags);
4123 PERL_UNUSED_ARG(val);
4125 DEBUG_PEEP("join", scan, depth, 0);
4127 assert(PL_regkind[OP(scan)] == EXACT);
4129 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
4130 * EXACT ones that are mergeable to the current one. */
4132 && ( PL_regkind[OP(n)] == NOTHING
4133 || (stringok && PL_regkind[OP(n)] == EXACT))
4135 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
4138 if (OP(n) == TAIL || n > next)
4140 if (PL_regkind[OP(n)] == NOTHING) {
4141 DEBUG_PEEP("skip:", n, depth, 0);
4142 NEXT_OFF(scan) += NEXT_OFF(n);
4143 next = n + NODE_STEP_REGNODE;
4150 else if (stringok) {
4151 const unsigned int oldl = STR_LEN(scan);
4152 regnode * const nnext = regnext(n);
4154 /* XXX I (khw) kind of doubt that this works on platforms (should
4155 * Perl ever run on one) where U8_MAX is above 255 because of lots
4156 * of other assumptions */
4157 /* Don't join if the sum can't fit into a single node */
4158 if (oldl + STR_LEN(n) > U8_MAX)
4161 /* Joining something that requires UTF-8 with something that
4162 * doesn't, means the result requires UTF-8. */
4163 if (OP(scan) == EXACT && (OP(n) == EXACT_REQ8)) {
4164 OP(scan) = EXACT_REQ8;
4166 else if (OP(scan) == EXACT_REQ8 && (OP(n) == EXACT)) {
4167 ; /* join is compatible, no need to change OP */
4169 else if ((OP(scan) == EXACTFU) && (OP(n) == EXACTFU_REQ8)) {
4170 OP(scan) = EXACTFU_REQ8;
4172 else if ((OP(scan) == EXACTFU_REQ8) && (OP(n) == EXACTFU)) {
4173 ; /* join is compatible, no need to change OP */
4175 else if (OP(scan) == EXACTFU && OP(n) == EXACTFU) {
4176 ; /* join is compatible, no need to change OP */
4178 else if (OP(scan) == EXACTFU && OP(n) == EXACTFU_S_EDGE) {
4180 /* Under /di, temporary EXACTFU_S_EDGE nodes are generated,
4181 * which can join with EXACTFU ones. We check for this case
4182 * here. These need to be resolved to either EXACTFU or
4183 * EXACTF at joining time. They have nothing in them that
4184 * would forbid them from being the more desirable EXACTFU
4185 * nodes except that they begin and/or end with a single [Ss].
4186 * The reason this is problematic is because they could be
4187 * joined in this loop with an adjacent node that ends and/or
4188 * begins with [Ss] which would then form the sequence 'ss',
4189 * which matches differently under /di than /ui, in which case
4190 * EXACTFU can't be used. If the 'ss' sequence doesn't get
4191 * formed, the nodes get absorbed into any adjacent EXACTFU
4192 * node. And if the only adjacent node is EXACTF, they get
4193 * absorbed into that, under the theory that a longer node is
4194 * better than two shorter ones, even if one is EXACTFU. Note
4195 * that EXACTFU_REQ8 is generated only for UTF-8 patterns,
4196 * and the EXACTFU_S_EDGE ones only for non-UTF-8. */
4198 if (STRING(n)[STR_LEN(n)-1] == 's') {
4200 /* Here the joined node would end with 's'. If the node
4201 * following the combination is an EXACTF one, it's better to
4202 * join this trailing edge 's' node with that one, leaving the
4203 * current one in 'scan' be the more desirable EXACTFU */
4204 if (OP(nnext) == EXACTF) {
4208 OP(scan) = EXACTFU_S_EDGE;
4210 } /* Otherwise, the beginning 's' of the 2nd node just
4211 becomes an interior 's' in 'scan' */
4213 else if (OP(scan) == EXACTF && OP(n) == EXACTF) {
4214 ; /* join is compatible, no need to change OP */
4216 else if (OP(scan) == EXACTF && OP(n) == EXACTFU_S_EDGE) {
4218 /* EXACTF nodes are compatible for joining with EXACTFU_S_EDGE
4219 * nodes. But the latter nodes can be also joined with EXACTFU
4220 * ones, and that is a better outcome, so if the node following
4221 * 'n' is EXACTFU, quit now so that those two can be joined
4223 if (OP(nnext) == EXACTFU) {
4227 /* The join is compatible, and the combined node will be
4228 * EXACTF. (These don't care if they begin or end with 's' */
4230 else if (OP(scan) == EXACTFU_S_EDGE && OP(n) == EXACTFU_S_EDGE) {
4231 if ( STRING(scan)[STR_LEN(scan)-1] == 's'
4232 && STRING(n)[0] == 's')
4234 /* When combined, we have the sequence 'ss', which means we
4235 * have to remain /di */
4239 else if (OP(scan) == EXACTFU_S_EDGE && OP(n) == EXACTFU) {
4240 if (STRING(n)[0] == 's') {
4241 ; /* Here the join is compatible and the combined node
4242 starts with 's', no need to change OP */
4244 else { /* Now the trailing 's' is in the interior */
4248 else if (OP(scan) == EXACTFU_S_EDGE && OP(n) == EXACTF) {
4250 /* The join is compatible, and the combined node will be
4251 * EXACTF. (These don't care if they begin or end with 's' */
4254 else if (OP(scan) != OP(n)) {
4256 /* The only other compatible joinings are the same node type */
4260 DEBUG_PEEP("merg", n, depth, 0);
4263 NEXT_OFF(scan) += NEXT_OFF(n);
4264 assert( ( STR_LEN(scan) + STR_LEN(n) ) < 256 );
4265 setSTR_LEN(scan, (U8)(STR_LEN(scan) + STR_LEN(n)));
4266 next = n + NODE_SZ_STR(n);
4267 /* Now we can overwrite *n : */
4268 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
4276 #ifdef EXPERIMENTAL_INPLACESCAN
4277 if (flags && !NEXT_OFF(n)) {
4278 DEBUG_PEEP("atch", val, depth, 0);
4279 if (reg_off_by_arg[OP(n)]) {
4280 ARG_SET(n, val - n);
4283 NEXT_OFF(n) = val - n;
4290 /* This temporary node can now be turned into EXACTFU, and must, as
4291 * regexec.c doesn't handle it */
4292 if (OP(scan) == EXACTFU_S_EDGE) {
4297 *unfolded_multi_char = FALSE;
4299 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
4300 * can now analyze for sequences of problematic code points. (Prior to
4301 * this final joining, sequences could have been split over boundaries, and
4302 * hence missed). The sequences only happen in folding, hence for any
4303 * non-EXACT EXACTish node */
4304 if (OP(scan) != EXACT && OP(scan) != EXACT_REQ8 && OP(scan) != EXACTL) {
4305 U8* s0 = (U8*) STRING(scan);
4307 U8* s_end = s0 + STR_LEN(scan);
4309 int total_count_delta = 0; /* Total delta number of characters that
4310 multi-char folds expand to */
4312 /* One pass is made over the node's string looking for all the
4313 * possibilities. To avoid some tests in the loop, there are two main
4314 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
4319 if (OP(scan) == EXACTFL) {
4322 /* An EXACTFL node would already have been changed to another
4323 * node type unless there is at least one character in it that
4324 * is problematic; likely a character whose fold definition
4325 * won't be known until runtime, and so has yet to be folded.
4326 * For all but the UTF-8 locale, folds are 1-1 in length, but
4327 * to handle the UTF-8 case, we need to create a temporary
4328 * folded copy using UTF-8 locale rules in order to analyze it.
4329 * This is because our macros that look to see if a sequence is
4330 * a multi-char fold assume everything is folded (otherwise the
4331 * tests in those macros would be too complicated and slow).
4332 * Note that here, the non-problematic folds will have already
4333 * been done, so we can just copy such characters. We actually
4334 * don't completely fold the EXACTFL string. We skip the
4335 * unfolded multi-char folds, as that would just create work
4336 * below to figure out the size they already are */
4338 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
4341 STRLEN s_len = UTF8SKIP(s);
4342 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
4343 Copy(s, d, s_len, U8);
4346 else if (is_FOLDS_TO_MULTI_utf8(s)) {
4347 *unfolded_multi_char = TRUE;
4348 Copy(s, d, s_len, U8);
4351 else if (isASCII(*s)) {
4352 *(d++) = toFOLD(*s);
4356 _toFOLD_utf8_flags(s, s_end, d, &len, FOLD_FLAGS_FULL);
4362 /* Point the remainder of the routine to look at our temporary
4366 } /* End of creating folded copy of EXACTFL string */
4368 /* Examine the string for a multi-character fold sequence. UTF-8
4369 * patterns have all characters pre-folded by the time this code is
4371 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
4372 length sequence we are looking for is 2 */
4374 int count = 0; /* How many characters in a multi-char fold */
4375 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
4376 if (! len) { /* Not a multi-char fold: get next char */
4381 { /* Here is a generic multi-char fold. */
4382 U8* multi_end = s + len;
4384 /* Count how many characters are in it. In the case of
4385 * /aa, no folds which contain ASCII code points are
4386 * allowed, so check for those, and skip if found. */
4387 if (OP(scan) != EXACTFAA && OP(scan) != EXACTFAA_NO_TRIE) {
4388 count = utf8_length(s, multi_end);
4392 while (s < multi_end) {
4395 goto next_iteration;
4405 /* The delta is how long the sequence is minus 1 (1 is how long
4406 * the character that folds to the sequence is) */
4407 total_count_delta += count - 1;
4411 /* We created a temporary folded copy of the string in EXACTFL
4412 * nodes. Therefore we need to be sure it doesn't go below zero,
4413 * as the real string could be shorter */
4414 if (OP(scan) == EXACTFL) {
4415 int total_chars = utf8_length((U8*) STRING(scan),
4416 (U8*) STRING(scan) + STR_LEN(scan));
4417 if (total_count_delta > total_chars) {
4418 total_count_delta = total_chars;
4422 *min_subtract += total_count_delta;
4425 else if (OP(scan) == EXACTFAA) {
4427 /* Non-UTF-8 pattern, EXACTFAA node. There can't be a multi-char
4428 * fold to the ASCII range (and there are no existing ones in the
4429 * upper latin1 range). But, as outlined in the comments preceding
4430 * this function, we need to flag any occurrences of the sharp s.
4431 * This character forbids trie formation (because of added
4433 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
4434 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
4435 || UNICODE_DOT_DOT_VERSION > 0)
4437 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
4438 OP(scan) = EXACTFAA_NO_TRIE;
4439 *unfolded_multi_char = TRUE;
4447 /* Non-UTF-8 pattern, not EXACTFAA node. Look for the multi-char
4448 * folds that are all Latin1. As explained in the comments
4449 * preceding this function, we look also for the sharp s in EXACTF
4450 * and EXACTFL nodes; it can be in the final position. Otherwise
4451 * we can stop looking 1 byte earlier because have to find at least
4452 * two characters for a multi-fold */
4453 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
4458 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
4459 if (! len) { /* Not a multi-char fold. */
4460 if (*s == LATIN_SMALL_LETTER_SHARP_S
4461 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
4463 *unfolded_multi_char = TRUE;
4470 && isALPHA_FOLD_EQ(*s, 's')
4471 && isALPHA_FOLD_EQ(*(s+1), 's'))
4474 /* EXACTF nodes need to know that the minimum length
4475 * changed so that a sharp s in the string can match this
4476 * ss in the pattern, but they remain EXACTF nodes, as they
4477 * won't match this unless the target string is is UTF-8,
4478 * which we don't know until runtime. EXACTFL nodes can't
4479 * transform into EXACTFU nodes */
4480 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
4481 OP(scan) = EXACTFUP;
4485 *min_subtract += len - 1;
4493 /* Allow dumping but overwriting the collection of skipped
4494 * ops and/or strings with fake optimized ops */
4495 n = scan + NODE_SZ_STR(scan);
4503 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl", scan, depth, 0);});
4507 /* REx optimizer. Converts nodes into quicker variants "in place".
4508 Finds fixed substrings. */
4510 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
4511 to the position after last scanned or to NULL. */
4513 #define INIT_AND_WITHP \
4514 assert(!and_withp); \
4515 Newx(and_withp, 1, regnode_ssc); \
4516 SAVEFREEPV(and_withp)
4520 S_unwind_scan_frames(pTHX_ const void *p)
4522 scan_frame *f= (scan_frame *)p;
4524 scan_frame *n= f->next_frame;
4530 /* the return from this sub is the minimum length that could possibly match */
4532 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
4533 SSize_t *minlenp, SSize_t *deltap,
4538 regnode_ssc *and_withp,
4539 U32 flags, U32 depth)
4540 /* scanp: Start here (read-write). */
4541 /* deltap: Write maxlen-minlen here. */
4542 /* last: Stop before this one. */
4543 /* data: string data about the pattern */
4544 /* stopparen: treat close N as END */
4545 /* recursed: which subroutines have we recursed into */
4546 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
4549 SSize_t final_minlen;
4550 /* There must be at least this number of characters to match */
4553 regnode *scan = *scanp, *next;
4555 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
4556 int is_inf_internal = 0; /* The studied chunk is infinite */
4557 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
4558 scan_data_t data_fake;
4559 SV *re_trie_maxbuff = NULL;
4560 regnode *first_non_open = scan;
4561 SSize_t stopmin = OPTIMIZE_INFTY;
4562 scan_frame *frame = NULL;
4563 DECLARE_AND_GET_RE_DEBUG_FLAGS;
4565 PERL_ARGS_ASSERT_STUDY_CHUNK;
4566 RExC_study_started= 1;
4568 Zero(&data_fake, 1, scan_data_t);
4571 while (first_non_open && OP(first_non_open) == OPEN)
4572 first_non_open=regnext(first_non_open);
4578 RExC_study_chunk_recursed_count++;
4580 DEBUG_OPTIMISE_MORE_r(
4582 Perl_re_indentf( aTHX_ "study_chunk stopparen=%ld recursed_count=%lu depth=%lu recursed_depth=%lu scan=%p last=%p",
4583 depth, (long)stopparen,
4584 (unsigned long)RExC_study_chunk_recursed_count,
4585 (unsigned long)depth, (unsigned long)recursed_depth,
4588 if (recursed_depth) {
4591 for ( j = 0 ; j < recursed_depth ; j++ ) {
4592 for ( i = 0 ; i < (U32)RExC_total_parens ; i++ ) {
4593 if (PAREN_TEST(j, i) && (!j || !PAREN_TEST(j - 1, i))) {
4594 Perl_re_printf( aTHX_ " %d",(int)i);
4598 if ( j + 1 < recursed_depth ) {
4599 Perl_re_printf( aTHX_ ",");
4603 Perl_re_printf( aTHX_ "\n");
4606 while ( scan && OP(scan) != END && scan < last ){
4607 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
4608 node length to get a real minimum (because
4609 the folded version may be shorter) */
4610 bool unfolded_multi_char = FALSE;
4611 /* Peephole optimizer: */
4612 DEBUG_STUDYDATA("Peep", data, depth, is_inf);
4613 DEBUG_PEEP("Peep", scan, depth, flags);
4616 /* The reason we do this here is that we need to deal with things like
4617 * /(?:f)(?:o)(?:o)/ which cant be dealt with by the normal EXACT
4618 * parsing code, as each (?:..) is handled by a different invocation of
4621 if (PL_regkind[OP(scan)] == EXACT && OP(scan) != LEXACT
4622 && OP(scan) != LEXACT_REQ8)
4623 join_exact(pRExC_state, scan, &min_subtract, &unfolded_multi_char,
4624 0, NULL, depth + 1);
4626 /* Follow the next-chain of the current node and optimize
4627 away all the NOTHINGs from it. */
4628 if (OP(scan) != CURLYX) {
4629 const int max = (reg_off_by_arg[OP(scan)]
4631 /* I32 may be smaller than U16 on CRAYs! */
4632 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
4633 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
4637 /* Skip NOTHING and LONGJMP. */
4638 while ( (n = regnext(n))
4639 && ( (PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
4640 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
4641 && off + noff < max)
4643 if (reg_off_by_arg[OP(scan)])
4646 NEXT_OFF(scan) = off;
4649 /* The principal pseudo-switch. Cannot be a switch, since we look into
4650 * several different things. */
4651 if ( OP(scan) == DEFINEP ) {
4653 SSize_t deltanext = 0;
4654 SSize_t fake_last_close = 0;
4655 I32 f = SCF_IN_DEFINE;
4657 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4658 scan = regnext(scan);
4659 assert( OP(scan) == IFTHEN );
4660 DEBUG_PEEP("expect IFTHEN", scan, depth, flags);
4662 data_fake.last_closep= &fake_last_close;
4664 next = regnext(scan);
4665 scan = NEXTOPER(NEXTOPER(scan));
4666 DEBUG_PEEP("scan", scan, depth, flags);
4667 DEBUG_PEEP("next", next, depth, flags);
4669 /* we suppose the run is continuous, last=next...
4670 * NOTE we dont use the return here! */
4671 /* DEFINEP study_chunk() recursion */
4672 (void)study_chunk(pRExC_state, &scan, &minlen,
4673 &deltanext, next, &data_fake, stopparen,
4674 recursed_depth, NULL, f, depth+1);
4679 OP(scan) == BRANCH ||
4680 OP(scan) == BRANCHJ ||
4683 next = regnext(scan);
4686 /* The op(next)==code check below is to see if we
4687 * have "BRANCH-BRANCH", "BRANCHJ-BRANCHJ", "IFTHEN-IFTHEN"
4688 * IFTHEN is special as it might not appear in pairs.
4689 * Not sure whether BRANCH-BRANCHJ is possible, regardless
4690 * we dont handle it cleanly. */
4691 if (OP(next) == code || code == IFTHEN) {
4692 /* NOTE - There is similar code to this block below for
4693 * handling TRIE nodes on a re-study. If you change stuff here
4694 * check there too. */
4695 SSize_t max1 = 0, min1 = OPTIMIZE_INFTY, num = 0;
4697 regnode * const startbranch=scan;
4699 if (flags & SCF_DO_SUBSTR) {
4700 /* Cannot merge strings after this. */
4701 scan_commit(pRExC_state, data, minlenp, is_inf);
4704 if (flags & SCF_DO_STCLASS)
4705 ssc_init_zero(pRExC_state, &accum);
4707 while (OP(scan) == code) {
4708 SSize_t deltanext, minnext, fake;
4710 regnode_ssc this_class;
4712 DEBUG_PEEP("Branch", scan, depth, flags);
4715 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4717 data_fake.whilem_c = data->whilem_c;
4718 data_fake.last_closep = data->last_closep;
4721 data_fake.last_closep = &fake;
4723 data_fake.pos_delta = delta;
4724 next = regnext(scan);
4726 scan = NEXTOPER(scan); /* everything */
4727 if (code != BRANCH) /* everything but BRANCH */
4728 scan = NEXTOPER(scan);
4730 if (flags & SCF_DO_STCLASS) {
4731 ssc_init(pRExC_state, &this_class);
4732 data_fake.start_class = &this_class;
4733 f = SCF_DO_STCLASS_AND;
4735 if (flags & SCF_WHILEM_VISITED_POS)
4736 f |= SCF_WHILEM_VISITED_POS;
4738 /* we suppose the run is continuous, last=next...*/
4739 /* recurse study_chunk() for each BRANCH in an alternation */
4740 minnext = study_chunk(pRExC_state, &scan, minlenp,
4741 &deltanext, next, &data_fake, stopparen,
4742 recursed_depth, NULL, f, depth+1);
4746 if (deltanext == OPTIMIZE_INFTY) {
4747 is_inf = is_inf_internal = 1;
4748 max1 = OPTIMIZE_INFTY;
4749 } else if (max1 < minnext + deltanext)
4750 max1 = minnext + deltanext;
4752 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4754 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4755 if ( stopmin > minnext)
4756 stopmin = min + min1;
4757 flags &= ~SCF_DO_SUBSTR;
4759 data->flags |= SCF_SEEN_ACCEPT;
4762 if (data_fake.flags & SF_HAS_EVAL)
4763 data->flags |= SF_HAS_EVAL;
4764 data->whilem_c = data_fake.whilem_c;
4766 if (flags & SCF_DO_STCLASS)
4767 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
4769 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
4771 if (flags & SCF_DO_SUBSTR) {
4772 data->pos_min += min1;
4773 if (data->pos_delta >= OPTIMIZE_INFTY - (max1 - min1))
4774 data->pos_delta = OPTIMIZE_INFTY;
4776 data->pos_delta += max1 - min1;
4777 if (max1 != min1 || is_inf)
4778 data->cur_is_floating = 1;
4781 if (delta == OPTIMIZE_INFTY
4782 || OPTIMIZE_INFTY - delta - (max1 - min1) < 0)
4783 delta = OPTIMIZE_INFTY;
4785 delta += max1 - min1;
4786 if (flags & SCF_DO_STCLASS_OR) {
4787 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
4789 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4790 flags &= ~SCF_DO_STCLASS;
4793 else if (flags & SCF_DO_STCLASS_AND) {
4795 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
4796 flags &= ~SCF_DO_STCLASS;
4799 /* Switch to OR mode: cache the old value of
4800 * data->start_class */
4802 StructCopy(data->start_class, and_withp, regnode_ssc);
4803 flags &= ~SCF_DO_STCLASS_AND;
4804 StructCopy(&accum, data->start_class, regnode_ssc);
4805 flags |= SCF_DO_STCLASS_OR;
4809 if (PERL_ENABLE_TRIE_OPTIMISATION &&
4810 OP( startbranch ) == BRANCH )
4814 Assuming this was/is a branch we are dealing with: 'scan'
4815 now points at the item that follows the branch sequence,
4816 whatever it is. We now start at the beginning of the
4817 sequence and look for subsequences of
4823 which would be constructed from a pattern like
4826 If we can find such a subsequence we need to turn the first
4827 element into a trie and then add the subsequent branch exact
4828 strings to the trie.
4832 1. patterns where the whole set of branches can be
4835 2. patterns where only a subset can be converted.
4837 In case 1 we can replace the whole set with a single regop
4838 for the trie. In case 2 we need to keep the start and end
4841 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
4842 becomes BRANCH TRIE; BRANCH X;
4844 There is an additional case, that being where there is a
4845 common prefix, which gets split out into an EXACT like node
4846 preceding the TRIE node.
4848 If x(1..n)==tail then we can do a simple trie, if not we make
4849 a "jump" trie, such that when we match the appropriate word
4850 we "jump" to the appropriate tail node. Essentially we turn
4851 a nested if into a case structure of sorts.
4856 if (!re_trie_maxbuff) {
4857 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
4858 if (!SvIOK(re_trie_maxbuff))
4859 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
4861 if ( SvIV(re_trie_maxbuff)>=0 ) {
4863 regnode *first = (regnode *)NULL;
4864 regnode *prev = (regnode *)NULL;
4865 regnode *tail = scan;
4869 /* var tail is used because there may be a TAIL
4870 regop in the way. Ie, the exacts will point to the
4871 thing following the TAIL, but the last branch will
4872 point at the TAIL. So we advance tail. If we
4873 have nested (?:) we may have to move through several
4877 while ( OP( tail ) == TAIL ) {
4878 /* this is the TAIL generated by (?:) */
4879 tail = regnext( tail );
4883 DEBUG_TRIE_COMPILE_r({
4884 regprop(RExC_rx, RExC_mysv, tail, NULL, pRExC_state);
4885 Perl_re_indentf( aTHX_ "%s %" UVuf ":%s\n",
4887 "Looking for TRIE'able sequences. Tail node is ",
4888 (UV) REGNODE_OFFSET(tail),
4889 SvPV_nolen_const( RExC_mysv )
4895 Step through the branches
4896 cur represents each branch,
4897 noper is the first thing to be matched as part
4899 noper_next is the regnext() of that node.
4901 We normally handle a case like this
4902 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
4903 support building with NOJUMPTRIE, which restricts
4904 the trie logic to structures like /FOO|BAR/.
4906 If noper is a trieable nodetype then the branch is
4907 a possible optimization target. If we are building
4908 under NOJUMPTRIE then we require that noper_next is
4909 the same as scan (our current position in the regex
4912 Once we have two or more consecutive such branches
4913 we can create a trie of the EXACT's contents and
4914 stitch it in place into the program.
4916 If the sequence represents all of the branches in
4917 the alternation we replace the entire thing with a
4920 Otherwise when it is a subsequence we need to
4921 stitch it in place and replace only the relevant
4922 branches. This means the first branch has to remain
4923 as it is used by the alternation logic, and its
4924 next pointer, and needs to be repointed at the item
4925 on the branch chain following the last branch we
4926 have optimized away.
4928 This could be either a BRANCH, in which case the
4929 subsequence is internal, or it could be the item
4930 following the branch sequence in which case the
4931 subsequence is at the end (which does not
4932 necessarily mean the first node is the start of the
4935 TRIE_TYPE(X) is a define which maps the optype to a
4939 ----------------+-----------
4944 EXACTFU_REQ8 | EXACTFU
4948 EXACTFLU8 | EXACTFLU8
4952 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) \
4954 : ( EXACT == (X) || EXACT_REQ8 == (X) ) \
4956 : ( EXACTFU == (X) \
4957 || EXACTFU_REQ8 == (X) \
4958 || EXACTFUP == (X) ) \
4960 : ( EXACTFAA == (X) ) \
4962 : ( EXACTL == (X) ) \
4964 : ( EXACTFLU8 == (X) ) \
4968 /* dont use tail as the end marker for this traverse */
4969 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
4970 regnode * const noper = NEXTOPER( cur );
4971 U8 noper_type = OP( noper );
4972 U8 noper_trietype = TRIE_TYPE( noper_type );
4973 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
4974 regnode * const noper_next = regnext( noper );
4975 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4976 U8 noper_next_trietype = (noper_next && noper_next < tail) ? TRIE_TYPE( noper_next_type ) :0;
4979 DEBUG_TRIE_COMPILE_r({
4980 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4981 Perl_re_indentf( aTHX_ "- %d:%s (%d)",
4983 REG_NODE_NUM(cur), SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur) );
4985 regprop(RExC_rx, RExC_mysv, noper, NULL, pRExC_state);
4986 Perl_re_printf( aTHX_ " -> %d:%s",
4987 REG_NODE_NUM(noper), SvPV_nolen_const(RExC_mysv));
4990 regprop(RExC_rx, RExC_mysv, noper_next, NULL, pRExC_state);
4991 Perl_re_printf( aTHX_ "\t=> %d:%s\t",
4992 REG_NODE_NUM(noper_next), SvPV_nolen_const(RExC_mysv));
4994 Perl_re_printf( aTHX_ "(First==%d,Last==%d,Cur==%d,tt==%s,ntt==%s,nntt==%s)\n",
4995 REG_NODE_NUM(first), REG_NODE_NUM(prev), REG_NODE_NUM(cur),
4996 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
5000 /* Is noper a trieable nodetype that can be merged
5001 * with the current trie (if there is one)? */
5005 ( noper_trietype == NOTHING )
5006 || ( trietype == NOTHING )
5007 || ( trietype == noper_trietype )
5010 && noper_next >= tail
5014 /* Handle mergable triable node Either we are
5015 * the first node in a new trieable sequence,
5016 * in which case we do some bookkeeping,
5017 * otherwise we update the end pointer. */
5020 if ( noper_trietype == NOTHING ) {
5021 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
5022 regnode * const noper_next = regnext( noper );
5023 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
5024 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
5027 if ( noper_next_trietype ) {
5028 trietype = noper_next_trietype;
5029 } else if (noper_next_type) {
5030 /* a NOTHING regop is 1 regop wide.
5031 * We need at least two for a trie
5032 * so we can't merge this in */
5036 trietype = noper_trietype;
5039 if ( trietype == NOTHING )
5040 trietype = noper_trietype;
5045 } /* end handle mergable triable node */
5047 /* handle unmergable node -
5048 * noper may either be a triable node which can
5049 * not be tried together with the current trie,
5050 * or a non triable node */
5052 /* If last is set and trietype is not
5053 * NOTHING then we have found at least two
5054 * triable branch sequences in a row of a
5055 * similar trietype so we can turn them
5056 * into a trie. If/when we allow NOTHING to
5057 * start a trie sequence this condition
5058 * will be required, and it isn't expensive
5059 * so we leave it in for now. */
5060 if ( trietype && trietype != NOTHING )
5061 make_trie( pRExC_state,
5062 startbranch, first, cur, tail,
5063 count, trietype, depth+1 );
5064 prev = NULL; /* note: we clear/update
5065 first, trietype etc below,
5066 so we dont do it here */
5070 && noper_next >= tail
5073 /* noper is triable, so we can start a new
5077 trietype = noper_trietype;
5079 /* if we already saw a first but the
5080 * current node is not triable then we have
5081 * to reset the first information. */
5086 } /* end handle unmergable node */
5087 } /* loop over branches */
5088 DEBUG_TRIE_COMPILE_r({
5089 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
5090 Perl_re_indentf( aTHX_ "- %s (%d) <SCAN FINISHED> ",
5091 depth+1, SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur));
5092 Perl_re_printf( aTHX_ "(First==%d, Last==%d, Cur==%d, tt==%s)\n",
5093 REG_NODE_NUM(first), REG_NODE_NUM(prev), REG_NODE_NUM(cur),
5094 PL_reg_name[trietype]
5098 if ( prev && trietype ) {
5099 if ( trietype != NOTHING ) {
5100 /* the last branch of the sequence was part of
5101 * a trie, so we have to construct it here
5102 * outside of the loop */
5103 made= make_trie( pRExC_state, startbranch,
5104 first, scan, tail, count,
5105 trietype, depth+1 );
5106 #ifdef TRIE_STUDY_OPT
5107 if ( ((made == MADE_EXACT_TRIE &&
5108 startbranch == first)
5109 || ( first_non_open == first )) &&
5111 flags |= SCF_TRIE_RESTUDY;
5112 if ( startbranch == first
5115 RExC_seen &=~REG_TOP_LEVEL_BRANCHES_SEEN;
5120 /* at this point we know whatever we have is a
5121 * NOTHING sequence/branch AND if 'startbranch'
5122 * is 'first' then we can turn the whole thing
5125 if ( startbranch == first ) {
5127 /* the entire thing is a NOTHING sequence,
5128 * something like this: (?:|) So we can
5129 * turn it into a plain NOTHING op. */
5130 DEBUG_TRIE_COMPILE_r({
5131 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
5132 Perl_re_indentf( aTHX_ "- %s (%d) <NOTHING BRANCH SEQUENCE>\n",
5134 SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur));
5137 OP(startbranch)= NOTHING;
5138 NEXT_OFF(startbranch)= tail - startbranch;
5139 for ( opt= startbranch + 1; opt < tail ; opt++ )
5143 } /* end if ( prev) */
5144 } /* TRIE_MAXBUF is non zero */
5148 else if ( code == BRANCHJ ) { /* single branch is optimized. */
5149 scan = NEXTOPER(NEXTOPER(scan));
5150 } else /* single branch is optimized. */
5151 scan = NEXTOPER(scan);
5153 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB) {
5155 regnode *start = NULL;
5156 regnode *end = NULL;
5157 U32 my_recursed_depth= recursed_depth;
5159 if (OP(scan) != SUSPEND) { /* GOSUB */
5160 /* Do setup, note this code has side effects beyond
5161 * the rest of this block. Specifically setting
5162 * RExC_recurse[] must happen at least once during
5165 RExC_recurse[ARG2L(scan)] = scan;
5166 start = REGNODE_p(RExC_open_parens[paren]);
5167 end = REGNODE_p(RExC_close_parens[paren]);
5169 /* NOTE we MUST always execute the above code, even
5170 * if we do nothing with a GOSUB */
5172 ( flags & SCF_IN_DEFINE )
5175 (is_inf_internal || is_inf || (data && data->flags & SF_IS_INF))
5177 ( (flags & (SCF_DO_STCLASS | SCF_DO_SUBSTR)) == 0 )
5180 /* no need to do anything here if we are in a define. */
5181 /* or we are after some kind of infinite construct
5182 * so we can skip recursing into this item.
5183 * Since it is infinite we will not change the maxlen
5184 * or delta, and if we miss something that might raise
5185 * the minlen it will merely pessimise a little.
5187 * Iow /(?(DEFINE)(?<foo>foo|food))a+(?&foo)/
5188 * might result in a minlen of 1 and not of 4,
5189 * but this doesn't make us mismatch, just try a bit
5190 * harder than we should.
5192 scan= regnext(scan);
5198 || !PAREN_TEST(recursed_depth - 1, paren)
5200 /* it is quite possible that there are more efficient ways
5201 * to do this. We maintain a bitmap per level of recursion
5202 * of which patterns we have entered so we can detect if a
5203 * pattern creates a possible infinite loop. When we
5204 * recurse down a level we copy the previous levels bitmap
5205 * down. When we are at recursion level 0 we zero the top
5206 * level bitmap. It would be nice to implement a different
5207 * more efficient way of doing this. In particular the top
5208 * level bitmap may be unnecessary.
5210 if (!recursed_depth) {
5211 Zero(RExC_study_chunk_recursed, RExC_study_chunk_recursed_bytes, U8);
5213 Copy(PAREN_OFFSET(recursed_depth - 1),
5214 PAREN_OFFSET(recursed_depth),
5215 RExC_study_chunk_recursed_bytes, U8);
5217 /* we havent recursed into this paren yet, so recurse into it */
5218 DEBUG_STUDYDATA("gosub-set", data, depth, is_inf);
5219 PAREN_SET(recursed_depth, paren);
5220 my_recursed_depth= recursed_depth + 1;
5222 DEBUG_STUDYDATA("gosub-inf", data, depth, is_inf);
5223 /* some form of infinite recursion, assume infinite length
5225 if (flags & SCF_DO_SUBSTR) {
5226 scan_commit(pRExC_state, data, minlenp, is_inf);
5227 data->cur_is_floating = 1;
5229 is_inf = is_inf_internal = 1;
5230 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5231 ssc_anything(data->start_class);
5232 flags &= ~SCF_DO_STCLASS;
5234 start= NULL; /* reset start so we dont recurse later on. */
5239 end = regnext(scan);
5242 scan_frame *newframe;
5244 if (!RExC_frame_last) {
5245 Newxz(newframe, 1, scan_frame);
5246 SAVEDESTRUCTOR_X(S_unwind_scan_frames, newframe);
5247 RExC_frame_head= newframe;
5249 } else if (!RExC_frame_last->next_frame) {
5250 Newxz(newframe, 1, scan_frame);
5251 RExC_frame_last->next_frame= newframe;
5252 newframe->prev_frame= RExC_frame_last;
5255 newframe= RExC_frame_last->next_frame;
5257 RExC_frame_last= newframe;
5259 newframe->next_regnode = regnext(scan);
5260 newframe->last_regnode = last;
5261 newframe->stopparen = stopparen;
5262 newframe->prev_recursed_depth = recursed_depth;
5263 newframe->this_prev_frame= frame;
5265 DEBUG_STUDYDATA("frame-new", data, depth, is_inf);
5266 DEBUG_PEEP("fnew", scan, depth, flags);
5273 recursed_depth= my_recursed_depth;
5278 else if ( OP(scan) == EXACT
5279 || OP(scan) == LEXACT
5280 || OP(scan) == EXACT_REQ8
5281 || OP(scan) == LEXACT_REQ8
5282 || OP(scan) == EXACTL)
5284 SSize_t bytelen = STR_LEN(scan), charlen;
5288 const U8 * const s = (U8*)STRING(scan);
5289 uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
5290 charlen = utf8_length(s, s + bytelen);
5292 uc = *((U8*)STRING(scan));
5296 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
5297 /* The code below prefers earlier match for fixed
5298 offset, later match for variable offset. */
5299 if (data->last_end == -1) { /* Update the start info. */
5300 data->last_start_min = data->pos_min;
5301 data->last_start_max = is_inf
5302 ? OPTIMIZE_INFTY : data->pos_min + data->pos_delta;
5304 sv_catpvn(data->last_found, STRING(scan), bytelen);
5306 SvUTF8_on(data->last_found);
5308 SV * const sv = data->last_found;
5309 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
5310 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5311 if (mg && mg->mg_len >= 0)
5312 mg->mg_len += charlen;
5314 data->last_end = data->pos_min + charlen;
5315 data->pos_min += charlen; /* As in the first entry. */
5316 data->flags &= ~SF_BEFORE_EOL;
5319 /* ANDing the code point leaves at most it, and not in locale, and
5320 * can't match null string */
5321 if (flags & SCF_DO_STCLASS_AND) {
5322 ssc_cp_and(data->start_class, uc);
5323 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5324 ssc_clear_locale(data->start_class);
5326 else if (flags & SCF_DO_STCLASS_OR) {
5327 ssc_add_cp(data->start_class, uc);
5328 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5330 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5331 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5333 flags &= ~SCF_DO_STCLASS;
5335 else if (PL_regkind[OP(scan)] == EXACT) {
5336 /* But OP != EXACT!, so is EXACTFish */
5337 SSize_t bytelen = STR_LEN(scan), charlen;
5338 const U8 * s = (U8*)STRING(scan);
5340 /* Replace a length 1 ASCII fold pair node with an ANYOFM node,
5341 * with the mask set to the complement of the bit that differs
5342 * between upper and lower case, and the lowest code point of the
5343 * pair (which the '&' forces) */
5346 && ( OP(scan) == EXACTFAA
5347 || ( OP(scan) == EXACTFU
5348 && ! HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(*s))))
5350 U8 mask = ~ ('A' ^ 'a'); /* These differ in just one bit */
5353 ARG_SET(scan, *s & mask);
5355 /* we're not EXACTFish any more, so restudy */
5359 /* Search for fixed substrings supports EXACT only. */
5360 if (flags & SCF_DO_SUBSTR) {
5362 scan_commit(pRExC_state, data, minlenp, is_inf);
5364 charlen = UTF ? (SSize_t) utf8_length(s, s + bytelen) : bytelen;
5365 if (unfolded_multi_char) {
5366 RExC_seen |= REG_UNFOLDED_MULTI_SEEN;
5368 min += charlen - min_subtract;
5370 delta += min_subtract;
5371 if (flags & SCF_DO_SUBSTR) {
5372 data->pos_min += charlen - min_subtract;
5373 if (data->pos_min < 0) {
5376 data->pos_delta += min_subtract;
5378 data->cur_is_floating = 1; /* float */
5382 if (flags & SCF_DO_STCLASS) {
5383 SV* EXACTF_invlist = make_exactf_invlist(pRExC_state, scan);
5385 assert(EXACTF_invlist);
5386 if (flags & SCF_DO_STCLASS_AND) {
5387 if (OP(scan) != EXACTFL)
5388 ssc_clear_locale(data->start_class);
5389 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5390 ANYOF_POSIXL_ZERO(data->start_class);
5391 ssc_intersection(data->start_class, EXACTF_invlist, FALSE);
5393 else { /* SCF_DO_STCLASS_OR */
5394 ssc_union(data->start_class, EXACTF_invlist, FALSE);
5395 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5397 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5398 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5400 flags &= ~SCF_DO_STCLASS;
5401 SvREFCNT_dec(EXACTF_invlist);
5404 else if (REGNODE_VARIES(OP(scan))) {
5405 SSize_t mincount, maxcount, minnext, deltanext, pos_before = 0;
5406 I32 fl = 0, f = flags;
5407 regnode * const oscan = scan;
5408 regnode_ssc this_class;
5409 regnode_ssc *oclass = NULL;
5410 I32 next_is_eval = 0;
5412 switch (PL_regkind[OP(scan)]) {
5413 case WHILEM: /* End of (?:...)* . */
5414 scan = NEXTOPER(scan);
5417 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
5418 next = NEXTOPER(scan);
5419 if ( OP(next) == EXACT
5420 || OP(next) == LEXACT
5421 || OP(next) == EXACT_REQ8
5422 || OP(next) == LEXACT_REQ8
5423 || OP(next) == EXACTL
5424 || (flags & SCF_DO_STCLASS))
5427 maxcount = REG_INFTY;
5428 next = regnext(scan);
5429 scan = NEXTOPER(scan);
5433 if (flags & SCF_DO_SUBSTR)
5438 next = NEXTOPER(scan);
5440 /* This temporary node can now be turned into EXACTFU, and
5441 * must, as regexec.c doesn't handle it */
5442 if (OP(next) == EXACTFU_S_EDGE) {
5446 if ( STR_LEN(next) == 1
5447 && isALPHA_A(* STRING(next))
5448 && ( OP(next) == EXACTFAA
5449 || ( OP(next) == EXACTFU
5450 && ! HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(* STRING(next)))))
5452 /* These differ in just one bit */
5453 U8 mask = ~ ('A' ^ 'a');
5455 assert(isALPHA_A(* STRING(next)));
5457 /* Then replace it by an ANYOFM node, with
5458 * the mask set to the complement of the
5459 * bit that differs between upper and lower
5460 * case, and the lowest code point of the
5461 * pair (which the '&' forces) */
5463 ARG_SET(next, *STRING(next) & mask);
5467 if (flags & SCF_DO_STCLASS) {
5469 maxcount = REG_INFTY;
5470 next = regnext(scan);
5471 scan = NEXTOPER(scan);
5474 if (flags & SCF_DO_SUBSTR) {
5475 scan_commit(pRExC_state, data, minlenp, is_inf);
5476 /* Cannot extend fixed substrings */
5477 data->cur_is_floating = 1; /* float */
5479 is_inf = is_inf_internal = 1;
5480 scan = regnext(scan);
5481 goto optimize_curly_tail;
5483 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
5484 && (scan->flags == stopparen))
5489 mincount = ARG1(scan);
5490 maxcount = ARG2(scan);
5492 next = regnext(scan);
5493 if (OP(scan) == CURLYX) {
5494 I32 lp = (data ? *(data->last_closep) : 0);
5495 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
5497 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
5498 next_is_eval = (OP(scan) == EVAL);
5500 if (flags & SCF_DO_SUBSTR) {
5502 scan_commit(pRExC_state, data, minlenp, is_inf);
5503 /* Cannot extend fixed substrings */
5504 pos_before = data->pos_min;
5508 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
5510 data->flags |= SF_IS_INF;
5512 if (flags & SCF_DO_STCLASS) {
5513 ssc_init(pRExC_state, &this_class);
5514 oclass = data->start_class;
5515 data->start_class = &this_class;
5516 f |= SCF_DO_STCLASS_AND;
5517 f &= ~SCF_DO_STCLASS_OR;
5519 /* Exclude from super-linear cache processing any {n,m}
5520 regops for which the combination of input pos and regex
5521 pos is not enough information to determine if a match
5524 For example, in the regex /foo(bar\s*){4,8}baz/ with the
5525 regex pos at the \s*, the prospects for a match depend not
5526 only on the input position but also on how many (bar\s*)
5527 repeats into the {4,8} we are. */
5528 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
5529 f &= ~SCF_WHILEM_VISITED_POS;
5531 /* This will finish on WHILEM, setting scan, or on NULL: */
5532 /* recurse study_chunk() on loop bodies */
5533 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
5534 last, data, stopparen, recursed_depth, NULL,
5536 ? (f & ~SCF_DO_SUBSTR)
5540 if (flags & SCF_DO_STCLASS)
5541 data->start_class = oclass;
5542 if (mincount == 0 || minnext == 0) {
5543 if (flags & SCF_DO_STCLASS_OR) {
5544 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5546 else if (flags & SCF_DO_STCLASS_AND) {
5547 /* Switch to OR mode: cache the old value of
5548 * data->start_class */
5550 StructCopy(data->start_class, and_withp, regnode_ssc);
5551 flags &= ~SCF_DO_STCLASS_AND;
5552 StructCopy(&this_class, data->start_class, regnode_ssc);
5553 flags |= SCF_DO_STCLASS_OR;
5554 ANYOF_FLAGS(data->start_class)
5555 |= SSC_MATCHES_EMPTY_STRING;
5557 } else { /* Non-zero len */
5558 if (flags & SCF_DO_STCLASS_OR) {
5559 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5560 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5562 else if (flags & SCF_DO_STCLASS_AND)
5563 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5564 flags &= ~SCF_DO_STCLASS;
5566 if (!scan) /* It was not CURLYX, but CURLY. */
5568 if (((flags & (SCF_TRIE_DOING_RESTUDY|SCF_DO_SUBSTR))==SCF_DO_SUBSTR)
5569 /* ? quantifier ok, except for (?{ ... }) */
5570 && (next_is_eval || !(mincount == 0 && maxcount == 1))
5571 && (minnext == 0) && (deltanext == 0)
5572 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
5573 && maxcount <= REG_INFTY/3) /* Complement check for big
5576 _WARN_HELPER(RExC_precomp_end, packWARN(WARN_REGEXP),
5577 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP),
5578 "Quantifier unexpected on zero-length expression "
5579 "in regex m/%" UTF8f "/",
5580 UTF8fARG(UTF, RExC_precomp_end - RExC_precomp,
5584 min += minnext * mincount;
5585 is_inf_internal |= deltanext == OPTIMIZE_INFTY
5586 || (maxcount == REG_INFTY && minnext + deltanext > 0);
5587 is_inf |= is_inf_internal;
5589 delta = OPTIMIZE_INFTY;
5591 delta += (minnext + deltanext) * maxcount
5592 - minnext * mincount;
5594 /* Try powerful optimization CURLYX => CURLYN. */
5595 if ( OP(oscan) == CURLYX && data
5596 && data->flags & SF_IN_PAR
5597 && !(data->flags & SF_HAS_EVAL)
5598 && !deltanext && minnext == 1 ) {
5599 /* Try to optimize to CURLYN. */
5600 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
5601 regnode * const nxt1 = nxt;
5608 if (!REGNODE_SIMPLE(OP(nxt))
5609 && !(PL_regkind[OP(nxt)] == EXACT
5610 && STR_LEN(nxt) == 1))
5616 if (OP(nxt) != CLOSE)
5618 if (RExC_open_parens) {
5621 RExC_open_parens[ARG(nxt1)] = REGNODE_OFFSET(oscan);
5624 RExC_close_parens[ARG(nxt1)] = REGNODE_OFFSET(nxt) + 2;
5626 /* Now we know that nxt2 is the only contents: */
5627 oscan->flags = (U8)ARG(nxt);
5629 OP(nxt1) = NOTHING; /* was OPEN. */
5632 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5633 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
5634 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
5635 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5636 OP(nxt + 1) = OPTIMIZED; /* was count. */
5637 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
5642 /* Try optimization CURLYX => CURLYM. */
5643 if ( OP(oscan) == CURLYX && data
5644 && !(data->flags & SF_HAS_PAR)
5645 && !(data->flags & SF_HAS_EVAL)
5646 && !deltanext /* atom is fixed width */
5647 && minnext != 0 /* CURLYM can't handle zero width */
5649 /* Nor characters whose fold at run-time may be
5650 * multi-character */
5651 && ! (RExC_seen & REG_UNFOLDED_MULTI_SEEN)
5653 /* XXXX How to optimize if data == 0? */
5654 /* Optimize to a simpler form. */
5655 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
5659 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
5660 && (OP(nxt2) != WHILEM))
5662 OP(nxt2) = SUCCEED; /* Whas WHILEM */
5663 /* Need to optimize away parenths. */
5664 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
5665 /* Set the parenth number. */
5666 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
5668 oscan->flags = (U8)ARG(nxt);
5669 if (RExC_open_parens) {
5671 RExC_open_parens[ARG(nxt1)] = REGNODE_OFFSET(oscan);
5674 RExC_close_parens[ARG(nxt1)] = REGNODE_OFFSET(nxt2)
5677 OP(nxt1) = OPTIMIZED; /* was OPEN. */
5678 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5681 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5682 OP(nxt + 1) = OPTIMIZED; /* was count. */
5683 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
5684 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
5687 while ( nxt1 && (OP(nxt1) != WHILEM)) {
5688 regnode *nnxt = regnext(nxt1);
5690 if (reg_off_by_arg[OP(nxt1)])
5691 ARG_SET(nxt1, nxt2 - nxt1);
5692 else if (nxt2 - nxt1 < U16_MAX)
5693 NEXT_OFF(nxt1) = nxt2 - nxt1;
5695 OP(nxt) = NOTHING; /* Cannot beautify */
5700 /* Optimize again: */
5701 /* recurse study_chunk() on optimised CURLYX => CURLYM */
5702 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
5703 NULL, stopparen, recursed_depth, NULL, 0,
5709 else if ((OP(oscan) == CURLYX)
5710 && (flags & SCF_WHILEM_VISITED_POS)
5711 /* See the comment on a similar expression above.
5712 However, this time it's not a subexpression
5713 we care about, but the expression itself. */
5714 && (maxcount == REG_INFTY)
5716 /* This stays as CURLYX, we can put the count/of pair. */
5717 /* Find WHILEM (as in regexec.c) */
5718 regnode *nxt = oscan + NEXT_OFF(oscan);
5720 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
5722 nxt = PREVOPER(nxt);
5723 if (nxt->flags & 0xf) {
5724 /* we've already set whilem count on this node */
5725 } else if (++data->whilem_c < 16) {
5726 assert(data->whilem_c <= RExC_whilem_seen);
5727 nxt->flags = (U8)(data->whilem_c
5728 | (RExC_whilem_seen << 4)); /* On WHILEM */
5731 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
5733 if (flags & SCF_DO_SUBSTR) {
5734 SV *last_str = NULL;
5735 STRLEN last_chrs = 0;
5736 int counted = mincount != 0;
5738 if (data->last_end > 0 && mincount != 0) { /* Ends with a
5740 SSize_t b = pos_before >= data->last_start_min
5741 ? pos_before : data->last_start_min;
5743 const char * const s = SvPV_const(data->last_found, l);
5744 SSize_t old = b - data->last_start_min;
5748 old = utf8_hop_forward((U8*)s, old,
5749 (U8 *) SvEND(data->last_found))
5752 /* Get the added string: */
5753 last_str = newSVpvn_utf8(s + old, l, UTF);
5754 last_chrs = UTF ? utf8_length((U8*)(s + old),
5755 (U8*)(s + old + l)) : l;
5756 if (deltanext == 0 && pos_before == b) {
5757 /* What was added is a constant string */
5760 SvGROW(last_str, (mincount * l) + 1);
5761 repeatcpy(SvPVX(last_str) + l,
5762 SvPVX_const(last_str), l,
5764 SvCUR_set(last_str, SvCUR(last_str) * mincount);
5765 /* Add additional parts. */
5766 SvCUR_set(data->last_found,
5767 SvCUR(data->last_found) - l);
5768 sv_catsv(data->last_found, last_str);
5770 SV * sv = data->last_found;
5772 SvUTF8(sv) && SvMAGICAL(sv) ?
5773 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5774 if (mg && mg->mg_len >= 0)
5775 mg->mg_len += last_chrs * (mincount-1);
5777 last_chrs *= mincount;
5778 data->last_end += l * (mincount - 1);
5781 /* start offset must point into the last copy */
5782 data->last_start_min += minnext * (mincount - 1);
5783 data->last_start_max =
5786 : data->last_start_max +
5787 (maxcount - 1) * (minnext + data->pos_delta);
5790 /* It is counted once already... */
5791 data->pos_min += minnext * (mincount - counted);
5793 Perl_re_printf( aTHX_ "counted=%" UVuf " deltanext=%" UVuf
5794 " OPTIMIZE_INFTY=%" UVuf " minnext=%" UVuf
5795 " maxcount=%" UVuf " mincount=%" UVuf "\n",
5796 (UV)counted, (UV)deltanext, (UV)OPTIMIZE_INFTY, (UV)minnext, (UV)maxcount,
5798 if (deltanext != OPTIMIZE_INFTY)
5799 Perl_re_printf( aTHX_ "LHS=%" UVuf " RHS=%" UVuf "\n",
5800 (UV)(-counted * deltanext + (minnext + deltanext) * maxcount
5801 - minnext * mincount), (UV)(OPTIMIZE_INFTY - data->pos_delta));
5803 if (deltanext == OPTIMIZE_INFTY
5804 || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= OPTIMIZE_INFTY - data->pos_delta)
5805 data->pos_delta = OPTIMIZE_INFTY;
5807 data->pos_delta += - counted * deltanext +
5808 (minnext + deltanext) * maxcount - minnext * mincount;
5809 if (mincount != maxcount) {
5810 /* Cannot extend fixed substrings found inside
5812 scan_commit(pRExC_state, data, minlenp, is_inf);
5813 if (mincount && last_str) {
5814 SV * const sv = data->last_found;
5815 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
5816 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5820 sv_setsv(sv, last_str);
5821 data->last_end = data->pos_min;
5822 data->last_start_min = data->pos_min - last_chrs;
5823 data->last_start_max = is_inf
5825 : data->pos_min + data->pos_delta - last_chrs;
5827 data->cur_is_floating = 1; /* float */
5829 SvREFCNT_dec(last_str);
5831 if (data && (fl & SF_HAS_EVAL))
5832 data->flags |= SF_HAS_EVAL;
5833 optimize_curly_tail:
5834 if (OP(oscan) != CURLYX) {
5835 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
5837 NEXT_OFF(oscan) += NEXT_OFF(next);
5842 Perl_croak(aTHX_ "panic: unexpected varying REx opcode %d",
5846 if (flags & SCF_DO_SUBSTR) {
5847 /* Cannot expect anything... */
5848 scan_commit(pRExC_state, data, minlenp, is_inf);
5849 data->cur_is_floating = 1; /* float */
5851 is_inf = is_inf_internal = 1;
5852 if (flags & SCF_DO_STCLASS_OR) {
5853 if (OP(scan) == CLUMP) {
5854 /* Actually is any start char, but very few code points
5855 * aren't start characters */
5856 ssc_match_all_cp(data->start_class);
5859 ssc_anything(data->start_class);
5862 flags &= ~SCF_DO_STCLASS;
5866 else if (OP(scan) == LNBREAK) {
5867 if (flags & SCF_DO_STCLASS) {
5868 if (flags & SCF_DO_STCLASS_AND) {
5869 ssc_intersection(data->start_class,
5870 PL_XPosix_ptrs[_CC_VERTSPACE], FALSE);
5871 ssc_clear_locale(data->start_class);
5872 ANYOF_FLAGS(data->start_class)
5873 &= ~SSC_MATCHES_EMPTY_STRING;
5875 else if (flags & SCF_DO_STCLASS_OR) {
5876 ssc_union(data->start_class,
5877 PL_XPosix_ptrs[_CC_VERTSPACE],
5879 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5881 /* See commit msg for
5882 * 749e076fceedeb708a624933726e7989f2302f6a */
5883 ANYOF_FLAGS(data->start_class)
5884 &= ~SSC_MATCHES_EMPTY_STRING;
5886 flags &= ~SCF_DO_STCLASS;
5889 if (delta != OPTIMIZE_INFTY)
5890 delta++; /* Because of the 2 char string cr-lf */
5891 if (flags & SCF_DO_SUBSTR) {
5892 /* Cannot expect anything... */
5893 scan_commit(pRExC_state, data, minlenp, is_inf);
5895 if (data->pos_delta != OPTIMIZE_INFTY) {
5896 data->pos_delta += 1;
5898 data->cur_is_floating = 1; /* float */
5901 else if (REGNODE_SIMPLE(OP(scan))) {
5903 if (flags & SCF_DO_SUBSTR) {
5904 scan_commit(pRExC_state, data, minlenp, is_inf);
5908 if (flags & SCF_DO_STCLASS) {
5910 SV* my_invlist = NULL;
5913 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5914 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5916 /* Some of the logic below assumes that switching
5917 locale on will only add false positives. */
5922 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d",
5926 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5927 ssc_match_all_cp(data->start_class);
5932 SV* REG_ANY_invlist = _new_invlist(2);
5933 REG_ANY_invlist = add_cp_to_invlist(REG_ANY_invlist,
5935 if (flags & SCF_DO_STCLASS_OR) {
5936 ssc_union(data->start_class,
5938 TRUE /* TRUE => invert, hence all but \n
5942 else if (flags & SCF_DO_STCLASS_AND) {
5943 ssc_intersection(data->start_class,
5945 TRUE /* TRUE => invert */
5947 ssc_clear_locale(data->start_class);
5949 SvREFCNT_dec_NN(REG_ANY_invlist);
5961 if (flags & SCF_DO_STCLASS_AND)
5962 ssc_and(pRExC_state, data->start_class,
5963 (regnode_charclass *) scan);
5965 ssc_or(pRExC_state, data->start_class,
5966 (regnode_charclass *) scan);
5972 SV* cp_list = get_ANYOFM_contents(scan);
5974 if (flags & SCF_DO_STCLASS_OR) {
5975 ssc_union(data->start_class, cp_list, invert);
5977 else if (flags & SCF_DO_STCLASS_AND) {
5978 ssc_intersection(data->start_class, cp_list, invert);
5981 SvREFCNT_dec_NN(cp_list);
5990 cp_list = _add_range_to_invlist(cp_list,
5992 ANYOFRbase(scan) + ANYOFRdelta(scan));
5994 if (flags & SCF_DO_STCLASS_OR) {
5995 ssc_union(data->start_class, cp_list, invert);
5997 else if (flags & SCF_DO_STCLASS_AND) {
5998 ssc_intersection(data->start_class, cp_list, invert);
6001 SvREFCNT_dec_NN(cp_list);
6010 namedclass = classnum_to_namedclass(FLAGS(scan)) + invert;
6011 if (flags & SCF_DO_STCLASS_AND) {
6012 bool was_there = cBOOL(
6013 ANYOF_POSIXL_TEST(data->start_class,
6015 ANYOF_POSIXL_ZERO(data->start_class);
6016 if (was_there) { /* Do an AND */
6017 ANYOF_POSIXL_SET(data->start_class, namedclass);
6019 /* No individual code points can now match */
6020 data->start_class->invlist
6021 = sv_2mortal(_new_invlist(0));
6024 int complement = namedclass + ((invert) ? -1 : 1);
6026 assert(flags & SCF_DO_STCLASS_OR);
6028 /* If the complement of this class was already there,
6029 * the result is that they match all code points,
6030 * (\d + \D == everything). Remove the classes from
6031 * future consideration. Locale is not relevant in
6033 if (ANYOF_POSIXL_TEST(data->start_class, complement)) {
6034 ssc_match_all_cp(data->start_class);
6035 ANYOF_POSIXL_CLEAR(data->start_class, namedclass);
6036 ANYOF_POSIXL_CLEAR(data->start_class, complement);
6038 else { /* The usual case; just add this class to the
6040 ANYOF_POSIXL_SET(data->start_class, namedclass);
6045 case NPOSIXA: /* For these, we always know the exact set of
6050 my_invlist = invlist_clone(PL_Posix_ptrs[FLAGS(scan)], NULL);
6051 goto join_posix_and_ascii;
6059 my_invlist = invlist_clone(PL_XPosix_ptrs[FLAGS(scan)], NULL);
6061 /* NPOSIXD matches all upper Latin1 code points unless the
6062 * target string being matched is UTF-8, which is
6063 * unknowable until match time. Since we are going to
6064 * invert, we want to get rid of all of them so that the
6065 * inversion will match all */
6066 if (OP(scan) == NPOSIXD) {
6067 _invlist_subtract(my_invlist, PL_UpperLatin1,
6071 join_posix_and_ascii:
6073 if (flags & SCF_DO_STCLASS_AND) {
6074 ssc_intersection(data->start_class, my_invlist, invert);
6075 ssc_clear_locale(data->start_class);
6078 assert(flags & SCF_DO_STCLASS_OR);
6079 ssc_union(data->start_class, my_invlist, invert);
6081 SvREFCNT_dec(my_invlist);
6083 if (flags & SCF_DO_STCLASS_OR)
6084 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6085 flags &= ~SCF_DO_STCLASS;
6088 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
6089 data->flags |= (OP(scan) == MEOL
6092 scan_commit(pRExC_state, data, minlenp, is_inf);
6095 else if ( PL_regkind[OP(scan)] == BRANCHJ
6096 /* Lookbehind, or need to calculate parens/evals/stclass: */
6097 && (scan->flags || data || (flags & SCF_DO_STCLASS))
6098 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM))
6100 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
6101 || OP(scan) == UNLESSM )
6103 /* Negative Lookahead/lookbehind
6104 In this case we can't do fixed string optimisation.
6107 SSize_t deltanext, minnext, fake = 0;
6112 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
6114 data_fake.whilem_c = data->whilem_c;
6115 data_fake.last_closep = data->last_closep;
6118 data_fake.last_closep = &fake;
6119 data_fake.pos_delta = delta;
6120 if ( flags & SCF_DO_STCLASS && !scan->flags
6121 && OP(scan) == IFMATCH ) { /* Lookahead */
6122 ssc_init(pRExC_state, &intrnl);
6123 data_fake.start_class = &intrnl;
6124 f |= SCF_DO_STCLASS_AND;
6126 if (flags & SCF_WHILEM_VISITED_POS)
6127 f |= SCF_WHILEM_VISITED_POS;
6128 next = regnext(scan);
6129 nscan = NEXTOPER(NEXTOPER(scan));
6131 /* recurse study_chunk() for lookahead body */
6132 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
6133 last, &data_fake, stopparen,
6134 recursed_depth, NULL, f, depth+1);
6137 || deltanext > (I32) U8_MAX
6138 || minnext > (I32)U8_MAX
6139 || minnext + deltanext > (I32)U8_MAX)
6141 FAIL2("Lookbehind longer than %" UVuf " not implemented",
6145 /* The 'next_off' field has been repurposed to count the
6146 * additional starting positions to try beyond the initial
6147 * one. (This leaves it at 0 for non-variable length
6148 * matches to avoid breakage for those not using this
6151 scan->next_off = deltanext;
6152 ckWARNexperimental(RExC_parse,
6153 WARN_EXPERIMENTAL__VLB,
6154 "Variable length lookbehind is experimental");
6156 scan->flags = (U8)minnext + deltanext;
6159 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
6161 if (data_fake.flags & SF_HAS_EVAL)
6162 data->flags |= SF_HAS_EVAL;
6163 data->whilem_c = data_fake.whilem_c;
6165 if (f & SCF_DO_STCLASS_AND) {
6166 if (flags & SCF_DO_STCLASS_OR) {
6167 /* OR before, AND after: ideally we would recurse with
6168 * data_fake to get the AND applied by study of the
6169 * remainder of the pattern, and then derecurse;
6170 * *** HACK *** for now just treat as "no information".
6171 * See [perl #56690].
6173 ssc_init(pRExC_state, data->start_class);
6175 /* AND before and after: combine and continue. These
6176 * assertions are zero-length, so can match an EMPTY
6178 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
6179 ANYOF_FLAGS(data->start_class)
6180 |= SSC_MATCHES_EMPTY_STRING;
6184 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
6186 /* Positive Lookahead/lookbehind
6187 In this case we can do fixed string optimisation,
6188 but we must be careful about it. Note in the case of
6189 lookbehind the positions will be offset by the minimum
6190 length of the pattern, something we won't know about
6191 until after the recurse.
6193 SSize_t deltanext, fake = 0;
6197 /* We use SAVEFREEPV so that when the full compile
6198 is finished perl will clean up the allocated
6199 minlens when it's all done. This way we don't
6200 have to worry about freeing them when we know
6201 they wont be used, which would be a pain.
6204 Newx( minnextp, 1, SSize_t );
6205 SAVEFREEPV(minnextp);
6208 StructCopy(data, &data_fake, scan_data_t);
6209 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
6212 scan_commit(pRExC_state, &data_fake, minlenp, is_inf);
6213 data_fake.last_found=newSVsv(data->last_found);
6217 data_fake.last_closep = &fake;
6218 data_fake.flags = 0;
6219 data_fake.substrs[0].flags = 0;
6220 data_fake.substrs[1].flags = 0;
6221 data_fake.pos_delta = delta;
6223 data_fake.flags |= SF_IS_INF;
6224 if ( flags & SCF_DO_STCLASS && !scan->flags
6225 && OP(scan) == IFMATCH ) { /* Lookahead */
6226 ssc_init(pRExC_state, &intrnl);
6227 data_fake.start_class = &intrnl;
6228 f |= SCF_DO_STCLASS_AND;
6230 if (flags & SCF_WHILEM_VISITED_POS)
6231 f |= SCF_WHILEM_VISITED_POS;
6232 next = regnext(scan);
6233 nscan = NEXTOPER(NEXTOPER(scan));
6235 /* positive lookahead study_chunk() recursion */
6236 *minnextp = study_chunk(pRExC_state, &nscan, minnextp,
6237 &deltanext, last, &data_fake,
6238 stopparen, recursed_depth, NULL,
6241 assert(0); /* This code has never been tested since this
6242 is normally not compiled */
6244 || deltanext > (I32) U8_MAX
6245 || *minnextp > (I32)U8_MAX
6246 || *minnextp + deltanext > (I32)U8_MAX)
6248 FAIL2("Lookbehind longer than %" UVuf " not implemented",
6253 scan->next_off = deltanext;
6255 scan->flags = (U8)*minnextp + deltanext;
6260 if (f & SCF_DO_STCLASS_AND) {
6261 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
6262 ANYOF_FLAGS(data->start_class) |= SSC_MATCHES_EMPTY_STRING;
6265 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
6267 if (data_fake.flags & SF_HAS_EVAL)
6268 data->flags |= SF_HAS_EVAL;
6269 data->whilem_c = data_fake.whilem_c;
6270 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
6272 if (RExC_rx->minlen<*minnextp)
6273 RExC_rx->minlen=*minnextp;
6274 scan_commit(pRExC_state, &data_fake, minnextp, is_inf);
6275 SvREFCNT_dec_NN(data_fake.last_found);
6277 for (i = 0; i < 2; i++) {
6278 if (data_fake.substrs[i].minlenp != minlenp) {
6279 data->substrs[i].min_offset =
6280 data_fake.substrs[i].min_offset;
6281 data->substrs[i].max_offset =
6282 data_fake.substrs[i].max_offset;
6283 data->substrs[i].minlenp =
6284 data_fake.substrs[i].minlenp;
6285 data->substrs[i].lookbehind += scan->flags;
6293 else if (OP(scan) == OPEN) {
6294 if (stopparen != (I32)ARG(scan))
6297 else if (OP(scan) == CLOSE) {
6298 if (stopparen == (I32)ARG(scan)) {
6301 if ((I32)ARG(scan) == is_par) {
6302 next = regnext(scan);
6304 if ( next && (OP(next) != WHILEM) && next < last)
6305 is_par = 0; /* Disable optimization */
6308 *(data->last_closep) = ARG(scan);
6310 else if (OP(scan) == EVAL) {
6312 data->flags |= SF_HAS_EVAL;
6314 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
6315 if (flags & SCF_DO_SUBSTR) {
6316 scan_commit(pRExC_state, data, minlenp, is_inf);
6317 flags &= ~SCF_DO_SUBSTR;
6319 if (data && OP(scan)==ACCEPT) {
6320 data->flags |= SCF_SEEN_ACCEPT;
6325 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
6327 if (flags & SCF_DO_SUBSTR) {
6328 scan_commit(pRExC_state, data, minlenp, is_inf);
6329 data->cur_is_floating = 1; /* float */
6331 is_inf = is_inf_internal = 1;
6332 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
6333 ssc_anything(data->start_class);
6334 flags &= ~SCF_DO_STCLASS;
6336 else if (OP(scan) == GPOS) {
6337 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
6338 !(delta || is_inf || (data && data->pos_delta)))
6340 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
6341 RExC_rx->intflags |= PREGf_ANCH_GPOS;
6342 if (RExC_rx->gofs < (STRLEN)min)
6343 RExC_rx->gofs = min;
6345 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
6349 #ifdef TRIE_STUDY_OPT
6350 #ifdef FULL_TRIE_STUDY
6351 else if (PL_regkind[OP(scan)] == TRIE) {
6352 /* NOTE - There is similar code to this block above for handling
6353 BRANCH nodes on the initial study. If you change stuff here
6355 regnode *trie_node= scan;
6356 regnode *tail= regnext(scan);
6357 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
6358 SSize_t max1 = 0, min1 = OPTIMIZE_INFTY;
6361 if (flags & SCF_DO_SUBSTR) { /* XXXX Add !SUSPEND? */
6362 /* Cannot merge strings after this. */
6363 scan_commit(pRExC_state, data, minlenp, is_inf);
6365 if (flags & SCF_DO_STCLASS)
6366 ssc_init_zero(pRExC_state, &accum);
6372 const regnode *nextbranch= NULL;
6375 for ( word=1 ; word <= trie->wordcount ; word++)
6377 SSize_t deltanext=0, minnext=0, f = 0, fake;
6378 regnode_ssc this_class;
6380 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
6382 data_fake.whilem_c = data->whilem_c;
6383 data_fake.last_closep = data->last_closep;
6386 data_fake.last_closep = &fake;
6387 data_fake.pos_delta = delta;
6388 if (flags & SCF_DO_STCLASS) {
6389 ssc_init(pRExC_state, &this_class);
6390 data_fake.start_class = &this_class;
6391 f = SCF_DO_STCLASS_AND;
6393 if (flags & SCF_WHILEM_VISITED_POS)
6394 f |= SCF_WHILEM_VISITED_POS;
6396 if (trie->jump[word]) {
6398 nextbranch = trie_node + trie->jump[0];
6399 scan= trie_node + trie->jump[word];
6400 /* We go from the jump point to the branch that follows
6401 it. Note this means we need the vestigal unused
6402 branches even though they arent otherwise used. */
6403 /* optimise study_chunk() for TRIE */
6404 minnext = study_chunk(pRExC_state, &scan, minlenp,
6405 &deltanext, (regnode *)nextbranch, &data_fake,
6406 stopparen, recursed_depth, NULL, f, depth+1);
6408 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
6409 nextbranch= regnext((regnode*)nextbranch);
6411 if (min1 > (SSize_t)(minnext + trie->minlen))
6412 min1 = minnext + trie->minlen;
6413 if (deltanext == OPTIMIZE_INFTY) {
6414 is_inf = is_inf_internal = 1;
6415 max1 = OPTIMIZE_INFTY;
6416 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
6417 max1 = minnext + deltanext + trie->maxlen;
6419 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
6421 if (data_fake.flags & SCF_SEEN_ACCEPT) {
6422 if ( stopmin > min + min1)
6423 stopmin = min + min1;
6424 flags &= ~SCF_DO_SUBSTR;
6426 data->flags |= SCF_SEEN_ACCEPT;
6429 if (data_fake.flags & SF_HAS_EVAL)
6430 data->flags |= SF_HAS_EVAL;
6431 data->whilem_c = data_fake.whilem_c;
6433 if (flags & SCF_DO_STCLASS)
6434 ssc_or(pRExC_state, &accum, (regnode_charclass *) &this_class);
6437 if (flags & SCF_DO_SUBSTR) {
6438 data->pos_min += min1;
6439 data->pos_delta += max1 - min1;
6440 if (max1 != min1 || is_inf)
6441 data->cur_is_floating = 1; /* float */
6444 if (delta != OPTIMIZE_INFTY) {
6445 if (OPTIMIZE_INFTY - (max1 - min1) >= delta)
6446 delta += max1 - min1;
6448 delta = OPTIMIZE_INFTY;
6450 if (flags & SCF_DO_STCLASS_OR) {
6451 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &accum);
6453 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6454 flags &= ~SCF_DO_STCLASS;
6457 else if (flags & SCF_DO_STCLASS_AND) {
6459 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
6460 flags &= ~SCF_DO_STCLASS;
6463 /* Switch to OR mode: cache the old value of
6464 * data->start_class */
6466 StructCopy(data->start_class, and_withp, regnode_ssc);
6467 flags &= ~SCF_DO_STCLASS_AND;
6468 StructCopy(&accum, data->start_class, regnode_ssc);
6469 flags |= SCF_DO_STCLASS_OR;
6476 else if (PL_regkind[OP(scan)] == TRIE) {
6477 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
6480 min += trie->minlen;
6481 delta += (trie->maxlen - trie->minlen);
6482 flags &= ~SCF_DO_STCLASS; /* xxx */
6483 if (flags & SCF_DO_SUBSTR) {
6484 /* Cannot expect anything... */
6485 scan_commit(pRExC_state, data, minlenp, is_inf);
6486 data->pos_min += trie->minlen;
6487 data->pos_delta += (trie->maxlen - trie->minlen);
6488 if (trie->maxlen != trie->minlen)
6489 data->cur_is_floating = 1; /* float */
6491 if (trie->jump) /* no more substrings -- for now /grr*/
6492 flags &= ~SCF_DO_SUBSTR;
6494 else if (OP(scan) == REGEX_SET) {
6495 Perl_croak(aTHX_ "panic: %s regnode should be resolved"
6496 " before optimization", reg_name[REGEX_SET]);
6499 #endif /* old or new */
6500 #endif /* TRIE_STUDY_OPT */
6502 /* Else: zero-length, ignore. */
6503 scan = regnext(scan);
6508 /* we need to unwind recursion. */
6511 DEBUG_STUDYDATA("frame-end", data, depth, is_inf);
6512 DEBUG_PEEP("fend", scan, depth, flags);
6514 /* restore previous context */
6515 last = frame->last_regnode;
6516 scan = frame->next_regnode;
6517 stopparen = frame->stopparen;
6518 recursed_depth = frame->prev_recursed_depth;
6520 RExC_frame_last = frame->prev_frame;
6521 frame = frame->this_prev_frame;
6522 goto fake_study_recurse;
6526 DEBUG_STUDYDATA("pre-fin", data, depth, is_inf);
6529 *deltap = is_inf_internal ? OPTIMIZE_INFTY : delta;
6531 if (flags & SCF_DO_SUBSTR && is_inf)
6532 data->pos_delta = OPTIMIZE_INFTY - data->pos_min;
6533 if (is_par > (I32)U8_MAX)
6535 if (is_par && pars==1 && data) {
6536 data->flags |= SF_IN_PAR;
6537 data->flags &= ~SF_HAS_PAR;
6539 else if (pars && data) {
6540 data->flags |= SF_HAS_PAR;
6541 data->flags &= ~SF_IN_PAR;
6543 if (flags & SCF_DO_STCLASS_OR)
6544 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6545 if (flags & SCF_TRIE_RESTUDY)
6546 data->flags |= SCF_TRIE_RESTUDY;
6548 DEBUG_STUDYDATA("post-fin", data, depth, is_inf);
6550 final_minlen = min < stopmin
6553 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN)) {
6554 if (final_minlen > OPTIMIZE_INFTY - delta)
6555 RExC_maxlen = OPTIMIZE_INFTY;
6556 else if (RExC_maxlen < final_minlen + delta)
6557 RExC_maxlen = final_minlen + delta;
6559 return final_minlen;
6563 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
6565 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
6567 PERL_ARGS_ASSERT_ADD_DATA;
6569 Renewc(RExC_rxi->data,
6570 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
6571 char, struct reg_data);
6573 Renew(RExC_rxi->data->what, count + n, U8);
6575 Newx(RExC_rxi->data->what, n, U8);
6576 RExC_rxi->data->count = count + n;
6577 Copy(s, RExC_rxi->data->what + count, n, U8);
6581 /*XXX: todo make this not included in a non debugging perl, but appears to be
6582 * used anyway there, in 'use re' */
6583 #ifndef PERL_IN_XSUB_RE
6585 Perl_reginitcolors(pTHX)
6587 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
6589 char *t = savepv(s);
6593 t = strchr(t, '\t');
6599 PL_colors[i] = t = (char *)"";
6604 PL_colors[i++] = (char *)"";
6611 #ifdef TRIE_STUDY_OPT
6612 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
6615 (data.flags & SCF_TRIE_RESTUDY) \
6623 #define CHECK_RESTUDY_GOTO_butfirst
6627 * pregcomp - compile a regular expression into internal code
6629 * Decides which engine's compiler to call based on the hint currently in
6633 #ifndef PERL_IN_XSUB_RE
6635 /* return the currently in-scope regex engine (or the default if none) */
6637 regexp_engine const *
6638 Perl_current_re_engine(pTHX)
6640 if (IN_PERL_COMPILETIME) {
6641 HV * const table = GvHV(PL_hintgv);
6644 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
6645 return &PL_core_reg_engine;
6646 ptr = hv_fetchs(table, "regcomp", FALSE);
6647 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
6648 return &PL_core_reg_engine;
6649 return INT2PTR(regexp_engine*, SvIV(*ptr));
6653 if (!PL_curcop->cop_hints_hash)
6654 return &PL_core_reg_engine;
6655 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
6656 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
6657 return &PL_core_reg_engine;
6658 return INT2PTR(regexp_engine*, SvIV(ptr));
6664 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
6666 regexp_engine const *eng = current_re_engine();
6667 DECLARE_AND_GET_RE_DEBUG_FLAGS;
6669 PERL_ARGS_ASSERT_PREGCOMP;
6671 /* Dispatch a request to compile a regexp to correct regexp engine. */
6673 Perl_re_printf( aTHX_ "Using engine %" UVxf "\n",
6676 return CALLREGCOMP_ENG(eng, pattern, flags);
6680 /* public(ish) entry point for the perl core's own regex compiling code.
6681 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
6682 * pattern rather than a list of OPs, and uses the internal engine rather
6683 * than the current one */
6686 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
6688 PERL_ARGS_ASSERT_RE_COMPILE;
6689 return re_op_compile_wrapper(pattern, rx_flags, 0);
6693 S_re_op_compile_wrapper(pTHX_ SV * const pattern, U32 rx_flags, const U32 pm_flags)
6695 SV *pat = pattern; /* defeat constness! */
6697 PERL_ARGS_ASSERT_RE_OP_COMPILE_WRAPPER;
6699 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
6700 #ifdef PERL_IN_XSUB_RE
6703 &PL_core_reg_engine,
6705 NULL, NULL, rx_flags, pm_flags);
6710 S_free_codeblocks(pTHX_ struct reg_code_blocks *cbs)
6714 if (--cbs->refcnt > 0)
6716 for (n = 0; n < cbs->count; n++) {
6717 REGEXP *rx = cbs->cb[n].src_regex;
6719 cbs->cb[n].src_regex = NULL;
6720 SvREFCNT_dec_NN(rx);
6728 static struct reg_code_blocks *
6729 S_alloc_code_blocks(pTHX_ int ncode)
6731 struct reg_code_blocks *cbs;
6732 Newx(cbs, 1, struct reg_code_blocks);
6735 SAVEDESTRUCTOR_X(S_free_codeblocks, cbs);
6737 Newx(cbs->cb, ncode, struct reg_code_block);
6744 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
6745 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
6746 * point to the realloced string and length.
6748 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
6752 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
6753 char **pat_p, STRLEN *plen_p, int num_code_blocks)
6755 U8 *const src = (U8*)*pat_p;
6760 DECLARE_AND_GET_RE_DEBUG_FLAGS;
6762 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6763 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
6765 /* 1 for each byte + 1 for each byte that expands to two, + trailing NUL */
6766 Newx(dst, *plen_p + variant_under_utf8_count(src, src + *plen_p) + 1, U8);
6769 while (s < *plen_p) {
6770 append_utf8_from_native_byte(src[s], &d);
6772 if (n < num_code_blocks) {
6773 assert(pRExC_state->code_blocks);
6774 if (!do_end && pRExC_state->code_blocks->cb[n].start == s) {
6775 pRExC_state->code_blocks->cb[n].start = d - dst - 1;
6776 assert(*(d - 1) == '(');
6779 else if (do_end && pRExC_state->code_blocks->cb[n].end == s) {
6780 pRExC_state->code_blocks->cb[n].end = d - dst - 1;
6781 assert(*(d - 1) == ')');
6790 *pat_p = (char*) dst;
6792 RExC_orig_utf8 = RExC_utf8 = 1;
6797 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
6798 * while recording any code block indices, and handling overloading,
6799 * nested qr// objects etc. If pat is null, it will allocate a new
6800 * string, or just return the first arg, if there's only one.
6802 * Returns the malloced/updated pat.
6803 * patternp and pat_count is the array of SVs to be concatted;
6804 * oplist is the optional list of ops that generated the SVs;
6805 * recompile_p is a pointer to a boolean that will be set if
6806 * the regex will need to be recompiled.
6807 * delim, if non-null is an SV that will be inserted between each element
6811 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
6812 SV *pat, SV ** const patternp, int pat_count,
6813 OP *oplist, bool *recompile_p, SV *delim)
6817 bool use_delim = FALSE;
6818 bool alloced = FALSE;
6820 /* if we know we have at least two args, create an empty string,
6821 * then concatenate args to that. For no args, return an empty string */
6822 if (!pat && pat_count != 1) {
6828 for (svp = patternp; svp < patternp + pat_count; svp++) {
6831 STRLEN orig_patlen = 0;
6833 SV *msv = use_delim ? delim : *svp;
6834 if (!msv) msv = &PL_sv_undef;
6836 /* if we've got a delimiter, we go round the loop twice for each
6837 * svp slot (except the last), using the delimiter the second
6846 if (SvTYPE(msv) == SVt_PVAV) {
6847 /* we've encountered an interpolated array within
6848 * the pattern, e.g. /...@a..../. Expand the list of elements,
6849 * then recursively append elements.
6850 * The code in this block is based on S_pushav() */
6852 AV *const av = (AV*)msv;
6853 const SSize_t maxarg = AvFILL(av) + 1;
6857 assert(oplist->op_type == OP_PADAV
6858 || oplist->op_type == OP_RV2AV);
6859 oplist = OpSIBLING(oplist);
6862 if (SvRMAGICAL(av)) {
6865 Newx(array, maxarg, SV*);
6867 for (i=0; i < maxarg; i++) {
6868 SV ** const svp = av_fetch(av, i, FALSE);
6869 array[i] = svp ? *svp : &PL_sv_undef;
6873 array = AvARRAY(av);
6875 pat = S_concat_pat(aTHX_ pRExC_state, pat,
6876 array, maxarg, NULL, recompile_p,
6878 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
6884 /* we make the assumption here that each op in the list of
6885 * op_siblings maps to one SV pushed onto the stack,
6886 * except for code blocks, with have both an OP_NULL and
6888 * This allows us to match up the list of SVs against the
6889 * list of OPs to find the next code block.
6891 * Note that PUSHMARK PADSV PADSV ..
6893 * PADRANGE PADSV PADSV ..
6894 * so the alignment still works. */
6897 if (oplist->op_type == OP_NULL
6898 && (oplist->op_flags & OPf_SPECIAL))
6900 assert(n < pRExC_state->code_blocks->count);
6901 pRExC_state->code_blocks->cb[n].start = pat ? SvCUR(pat) : 0;
6902 pRExC_state->code_blocks->cb[n].block = oplist;
6903 pRExC_state->code_blocks->cb[n].src_regex = NULL;
6906 oplist = OpSIBLING(oplist); /* skip CONST */
6909 oplist = OpSIBLING(oplist);;
6912 /* apply magic and QR overloading to arg */
6915 if (SvROK(msv) && SvAMAGIC(msv)) {
6916 SV *sv = AMG_CALLunary(msv, regexp_amg);
6920 if (SvTYPE(sv) != SVt_REGEXP)
6921 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
6926 /* try concatenation overload ... */
6927 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
6928 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
6931 /* overloading involved: all bets are off over literal
6932 * code. Pretend we haven't seen it */
6934 pRExC_state->code_blocks->count -= n;
6938 /* ... or failing that, try "" overload */
6939 while (SvAMAGIC(msv)
6940 && (sv = AMG_CALLunary(msv, string_amg))
6944 && SvRV(msv) == SvRV(sv))
6949 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
6953 /* this is a partially unrolled
6954 * sv_catsv_nomg(pat, msv);
6955 * that allows us to adjust code block indices if
6958 char *dst = SvPV_force_nomg(pat, dlen);
6960 if (SvUTF8(msv) && !SvUTF8(pat)) {
6961 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
6962 sv_setpvn(pat, dst, dlen);
6965 sv_catsv_nomg(pat, msv);
6969 /* We have only one SV to process, but we need to verify
6970 * it is properly null terminated or we will fail asserts
6971 * later. In theory we probably shouldn't get such SV's,
6972 * but if we do we should handle it gracefully. */
6973 if ( SvTYPE(msv) != SVt_PV || (SvLEN(msv) > SvCUR(msv) && *(SvEND(msv)) == 0) || SvIsCOW_shared_hash(msv) ) {
6974 /* not a string, or a string with a trailing null */
6977 /* a string with no trailing null, we need to copy it
6978 * so it has a trailing null */
6979 pat = sv_2mortal(newSVsv(msv));
6984 pRExC_state->code_blocks->cb[n-1].end = SvCUR(pat)-1;
6987 /* extract any code blocks within any embedded qr//'s */
6988 if (rx && SvTYPE(rx) == SVt_REGEXP
6989 && RX_ENGINE((REGEXP*)rx)->op_comp)
6992 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
6993 if (ri->code_blocks && ri->code_blocks->count) {
6995 /* the presence of an embedded qr// with code means
6996 * we should always recompile: the text of the
6997 * qr// may not have changed, but it may be a
6998 * different closure than last time */
7000 if (pRExC_state->code_blocks) {
7001 int new_count = pRExC_state->code_blocks->count
7002 + ri->code_blocks->count;
7003 Renew(pRExC_state->code_blocks->cb,
7004 new_count, struct reg_code_block);
7005 pRExC_state->code_blocks->count = new_count;
7008 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_
7009 ri->code_blocks->count);
7011 for (i=0; i < ri->code_blocks->count; i++) {
7012 struct reg_code_block *src, *dst;
7013 STRLEN offset = orig_patlen
7014 + ReANY((REGEXP *)rx)->pre_prefix;
7015 assert(n < pRExC_state->code_blocks->count);
7016 src = &ri->code_blocks->cb[i];
7017 dst = &pRExC_state->code_blocks->cb[n];
7018 dst->start = src->start + offset;
7019 dst->end = src->end + offset;
7020 dst->block = src->block;
7021 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
7030 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
7039 /* see if there are any run-time code blocks in the pattern.
7040 * False positives are allowed */
7043 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
7044 char *pat, STRLEN plen)
7049 PERL_UNUSED_CONTEXT;
7051 for (s = 0; s < plen; s++) {
7052 if ( pRExC_state->code_blocks
7053 && n < pRExC_state->code_blocks->count
7054 && s == pRExC_state->code_blocks->cb[n].start)
7056 s = pRExC_state->code_blocks->cb[n].end;
7060 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
7062 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
7064 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
7071 /* Handle run-time code blocks. We will already have compiled any direct
7072 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
7073 * copy of it, but with any literal code blocks blanked out and
7074 * appropriate chars escaped; then feed it into
7076 * eval "qr'modified_pattern'"
7080 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
7084 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
7086 * After eval_sv()-ing that, grab any new code blocks from the returned qr
7087 * and merge them with any code blocks of the original regexp.
7089 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
7090 * instead, just save the qr and return FALSE; this tells our caller that
7091 * the original pattern needs upgrading to utf8.
7095 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
7096 char *pat, STRLEN plen)
7100 DECLARE_AND_GET_RE_DEBUG_FLAGS;
7102 if (pRExC_state->runtime_code_qr) {
7103 /* this is the second time we've been called; this should
7104 * only happen if the main pattern got upgraded to utf8
7105 * during compilation; re-use the qr we compiled first time
7106 * round (which should be utf8 too)
7108 qr = pRExC_state->runtime_code_qr;
7109 pRExC_state->runtime_code_qr = NULL;
7110 assert(RExC_utf8 && SvUTF8(qr));
7116 int newlen = plen + 7; /* allow for "qr''xx\0" extra chars */
7120 /* determine how many extra chars we need for ' and \ escaping */
7121 for (s = 0; s < plen; s++) {
7122 if (pat[s] == '\'' || pat[s] == '\\')
7126 Newx(newpat, newlen, char);
7128 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
7130 for (s = 0; s < plen; s++) {
7131 if ( pRExC_state->code_blocks
7132 && n < pRExC_state->code_blocks->count
7133 && s == pRExC_state->code_blocks->cb[n].start)
7135 /* blank out literal code block so that they aren't
7136 * recompiled: eg change from/to:
7146 assert(pat[s] == '(');
7147 assert(pat[s+1] == '?');
7151 while (s < pRExC_state->code_blocks->cb[n].end) {
7159 if (pat[s] == '\'' || pat[s] == '\\')
7164 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED) {
7166 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED_MORE) {
7172 Perl_re_printf( aTHX_
7173 "%sre-parsing pattern for runtime code:%s %s\n",
7174 PL_colors[4], PL_colors[5], newpat);
7177 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
7183 PUSHSTACKi(PERLSI_REQUIRE);
7184 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
7185 * parsing qr''; normally only q'' does this. It also alters
7187 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
7188 SvREFCNT_dec_NN(sv);
7193 SV * const errsv = ERRSV;
7194 if (SvTRUE_NN(errsv))
7195 /* use croak_sv ? */
7196 Perl_croak_nocontext("%" SVf, SVfARG(errsv));
7198 assert(SvROK(qr_ref));
7200 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
7201 /* the leaving below frees the tmp qr_ref.
7202 * Give qr a life of its own */
7210 if (!RExC_utf8 && SvUTF8(qr)) {
7211 /* first time through; the pattern got upgraded; save the
7212 * qr for the next time through */
7213 assert(!pRExC_state->runtime_code_qr);
7214 pRExC_state->runtime_code_qr = qr;
7219 /* extract any code blocks within the returned qr// */
7222 /* merge the main (r1) and run-time (r2) code blocks into one */
7224 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
7225 struct reg_code_block *new_block, *dst;
7226 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
7230 if (!r2->code_blocks || !r2->code_blocks->count) /* we guessed wrong */
7232 SvREFCNT_dec_NN(qr);
7236 if (!r1->code_blocks)
7237 r1->code_blocks = S_alloc_code_blocks(aTHX_ 0);
7239 r1c = r1->code_blocks->count;
7240 r2c = r2->code_blocks->count;
7242 Newx(new_block, r1c + r2c, struct reg_code_block);
7246 while (i1 < r1c || i2 < r2c) {
7247 struct reg_code_block *src;
7251 src = &r2->code_blocks->cb[i2++];
7255 src = &r1->code_blocks->cb[i1++];
7256 else if ( r1->code_blocks->cb[i1].start
7257 < r2->code_blocks->cb[i2].start)
7259 src = &r1->code_blocks->cb[i1++];
7260 assert(src->end < r2->code_blocks->cb[i2].start);
7263 assert( r1->code_blocks->cb[i1].start
7264 > r2->code_blocks->cb[i2].start);
7265 src = &r2->code_blocks->cb[i2++];
7267 assert(src->end < r1->code_blocks->cb[i1].start);
7270 assert(pat[src->start] == '(');
7271 assert(pat[src->end] == ')');
7272 dst->start = src->start;
7273 dst->end = src->end;
7274 dst->block = src->block;
7275 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
7279 r1->code_blocks->count += r2c;
7280 Safefree(r1->code_blocks->cb);
7281 r1->code_blocks->cb = new_block;
7284 SvREFCNT_dec_NN(qr);
7290 S_setup_longest(pTHX_ RExC_state_t *pRExC_state,
7291 struct reg_substr_datum *rsd,
7292 struct scan_data_substrs *sub,
7293 STRLEN longest_length)
7295 /* This is the common code for setting up the floating and fixed length
7296 * string data extracted from Perl_re_op_compile() below. Returns a boolean
7297 * as to whether succeeded or not */
7301 bool eol = cBOOL(sub->flags & SF_BEFORE_EOL);
7302 bool meol = cBOOL(sub->flags & SF_BEFORE_MEOL);
7304 if (! (longest_length
7305 || (eol /* Can't have SEOL and MULTI */
7306 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
7308 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
7309 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
7314 /* copy the information about the longest from the reg_scan_data
7315 over to the program. */
7316 if (SvUTF8(sub->str)) {
7318 rsd->utf8_substr = sub->str;
7320 rsd->substr = sub->str;
7321 rsd->utf8_substr = NULL;
7323 /* end_shift is how many chars that must be matched that
7324 follow this item. We calculate it ahead of time as once the
7325 lookbehind offset is added in we lose the ability to correctly
7327 ml = sub->minlenp ? *(sub->minlenp) : (SSize_t)longest_length;
7328 rsd->end_shift = ml - sub->min_offset
7330 /* XXX SvTAIL is always false here - did you mean FBMcf_TAIL
7332 + (SvTAIL(sub->str) != 0)
7336 t = (eol/* Can't have SEOL and MULTI */
7337 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
7338 fbm_compile(sub->str, t ? FBMcf_TAIL : 0);
7344 S_set_regex_pv(pTHX_ RExC_state_t *pRExC_state, REGEXP *Rx)
7346 /* Calculates and sets in the compiled pattern 'Rx' the string to compile,
7347 * properly wrapped with the right modifiers */
7349 bool has_p = ((RExC_rx->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
7350 bool has_charset = RExC_utf8 || (get_regex_charset(RExC_rx->extflags)
7351 != REGEX_DEPENDS_CHARSET);
7353 /* The caret is output if there are any defaults: if not all the STD
7354 * flags are set, or if no character set specifier is needed */
7356 (((RExC_rx->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
7358 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
7359 == REG_RUN_ON_COMMENT_SEEN);
7360 U8 reganch = (U8)((RExC_rx->extflags & RXf_PMf_STD_PMMOD)
7361 >> RXf_PMf_STD_PMMOD_SHIFT);
7362 const char *fptr = STD_PAT_MODS; /*"msixxn"*/
7364 STRLEN pat_len = RExC_precomp_end - RExC_precomp;
7366 /* We output all the necessary flags; we never output a minus, as all
7367 * those are defaults, so are
7368 * covered by the caret */
7369 const STRLEN wraplen = pat_len + has_p + has_runon
7370 + has_default /* If needs a caret */
7371 + PL_bitcount[reganch] /* 1 char for each set standard flag */
7373 /* If needs a character set specifier */
7374 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
7375 + (sizeof("(?:)") - 1);
7377 PERL_ARGS_ASSERT_SET_REGEX_PV;
7379 /* make sure PL_bitcount bounds not exceeded */
7380 assert(sizeof(STD_PAT_MODS) <= 8);
7382 p = sv_grow(MUTABLE_SV(Rx), wraplen + 1); /* +1 for the ending NUL */
7385 SvFLAGS(Rx) |= SVf_UTF8;
7388 /* If a default, cover it using the caret */
7390 *p++= DEFAULT_PAT_MOD;
7396 name = get_regex_charset_name(RExC_rx->extflags, &len);
7397 if (strEQ(name, DEPENDS_PAT_MODS)) { /* /d under UTF-8 => /u */
7399 name = UNICODE_PAT_MODS;
7400 len = sizeof(UNICODE_PAT_MODS) - 1;
7402 Copy(name, p, len, char);
7406 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
7409 while((ch = *fptr++)) {
7417 Copy(RExC_precomp, p, pat_len, char);
7418 assert ((RX_WRAPPED(Rx) - p) < 16);
7419 RExC_rx->pre_prefix = p - RX_WRAPPED(Rx);
7422 /* Adding a trailing \n causes this to compile properly:
7423 my $R = qr / A B C # D E/x; /($R)/
7424 Otherwise the parens are considered part of the comment */
7429 SvCUR_set(Rx, p - RX_WRAPPED(Rx));
7433 * Perl_re_op_compile - the perl internal RE engine's function to compile a
7434 * regular expression into internal code.
7435 * The pattern may be passed either as:
7436 * a list of SVs (patternp plus pat_count)
7437 * a list of OPs (expr)
7438 * If both are passed, the SV list is used, but the OP list indicates
7439 * which SVs are actually pre-compiled code blocks
7441 * The SVs in the list have magic and qr overloading applied to them (and
7442 * the list may be modified in-place with replacement SVs in the latter
7445 * If the pattern hasn't changed from old_re, then old_re will be
7448 * eng is the current engine. If that engine has an op_comp method, then
7449 * handle directly (i.e. we assume that op_comp was us); otherwise, just
7450 * do the initial concatenation of arguments and pass on to the external
7453 * If is_bare_re is not null, set it to a boolean indicating whether the
7454 * arg list reduced (after overloading) to a single bare regex which has
7455 * been returned (i.e. /$qr/).
7457 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
7459 * pm_flags contains the PMf_* flags, typically based on those from the
7460 * pm_flags field of the related PMOP. Currently we're only interested in
7461 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL, PMf_WILDCARD.
7463 * For many years this code had an initial sizing pass that calculated
7464 * (sometimes incorrectly, leading to security holes) the size needed for the
7465 * compiled pattern. That was changed by commit
7466 * 7c932d07cab18751bfc7515b4320436273a459e2 in 5.29, which reallocs the size, a
7467 * node at a time, as parsing goes along. Patches welcome to fix any obsolete
7468 * references to this sizing pass.
7470 * Now, an initial crude guess as to the size needed is made, based on the
7471 * length of the pattern. Patches welcome to improve that guess. That amount
7472 * of space is malloc'd and then immediately freed, and then clawed back node
7473 * by node. This design is to minimze, to the extent possible, memory churn
7474 * when doing the the reallocs.
7476 * A separate parentheses counting pass may be needed in some cases.
7477 * (Previously the sizing pass did this.) Patches welcome to reduce the number
7480 * The existence of a sizing pass necessitated design decisions that are no
7481 * longer needed. There are potential areas of simplification.
7483 * Beware that the optimization-preparation code in here knows about some
7484 * of the structure of the compiled regexp. [I'll say.]
7488 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
7489 OP *expr, const regexp_engine* eng, REGEXP *old_re,
7490 bool *is_bare_re, const U32 orig_rx_flags, const U32 pm_flags)
7493 REGEXP *Rx; /* Capital 'R' means points to a REGEXP */
7501 SV** new_patternp = patternp;
7503 /* these are all flags - maybe they should be turned
7504 * into a single int with different bit masks */
7505 I32 sawlookahead = 0;
7510 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
7512 bool runtime_code = 0;
7514 RExC_state_t RExC_state;
7515 RExC_state_t * const pRExC_state = &RExC_state;
7516 #ifdef TRIE_STUDY_OPT
7518 RExC_state_t copyRExC_state;
7520 DECLARE_AND_GET_RE_DEBUG_FLAGS;
7522 PERL_ARGS_ASSERT_RE_OP_COMPILE;
7524 DEBUG_r(if (!PL_colorset) reginitcolors());
7527 pRExC_state->warn_text = NULL;
7528 pRExC_state->unlexed_names = NULL;
7529 pRExC_state->code_blocks = NULL;
7532 *is_bare_re = FALSE;
7534 if (expr && (expr->op_type == OP_LIST ||
7535 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
7536 /* allocate code_blocks if needed */
7540 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o))
7541 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
7542 ncode++; /* count of DO blocks */
7545 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_ ncode);
7549 /* compile-time pattern with just OP_CONSTs and DO blocks */
7554 /* find how many CONSTs there are */
7557 if (expr->op_type == OP_CONST)
7560 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
7561 if (o->op_type == OP_CONST)
7565 /* fake up an SV array */
7567 assert(!new_patternp);
7568 Newx(new_patternp, n, SV*);
7569 SAVEFREEPV(new_patternp);
7573 if (expr->op_type == OP_CONST)
7574 new_patternp[n] = cSVOPx_sv(expr);
7576 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
7577 if (o->op_type == OP_CONST)
7578 new_patternp[n++] = cSVOPo_sv;
7583 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7584 "Assembling pattern from %d elements%s\n", pat_count,
7585 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
7587 /* set expr to the first arg op */
7589 if (pRExC_state->code_blocks && pRExC_state->code_blocks->count
7590 && expr->op_type != OP_CONST)
7592 expr = cLISTOPx(expr)->op_first;
7593 assert( expr->op_type == OP_PUSHMARK
7594 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
7595 || expr->op_type == OP_PADRANGE);
7596 expr = OpSIBLING(expr);
7599 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
7600 expr, &recompile, NULL);
7602 /* handle bare (possibly after overloading) regex: foo =~ $re */
7607 if (SvTYPE(re) == SVt_REGEXP) {
7611 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7612 "Precompiled pattern%s\n",
7613 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
7619 exp = SvPV_nomg(pat, plen);
7621 if (!eng->op_comp) {
7622 if ((SvUTF8(pat) && IN_BYTES)
7623 || SvGMAGICAL(pat) || SvAMAGIC(pat))
7625 /* make a temporary copy; either to convert to bytes,
7626 * or to avoid repeating get-magic / overloaded stringify */
7627 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
7628 (IN_BYTES ? 0 : SvUTF8(pat)));
7630 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
7633 /* ignore the utf8ness if the pattern is 0 length */
7634 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
7635 RExC_uni_semantics = 0;
7636 RExC_contains_locale = 0;
7637 RExC_strict = cBOOL(pm_flags & RXf_PMf_STRICT);
7638 RExC_in_script_run = 0;
7639 RExC_study_started = 0;
7640 pRExC_state->runtime_code_qr = NULL;
7641 RExC_frame_head= NULL;
7642 RExC_frame_last= NULL;
7643 RExC_frame_count= 0;
7644 RExC_latest_warn_offset = 0;
7645 RExC_use_BRANCHJ = 0;
7646 RExC_warned_WARN_EXPERIMENTAL__VLB = 0;
7647 RExC_warned_WARN_EXPERIMENTAL__REGEX_SETS = 0;
7648 RExC_total_parens = 0;
7649 RExC_open_parens = NULL;
7650 RExC_close_parens = NULL;
7651 RExC_paren_names = NULL;
7653 RExC_seen_d_op = FALSE;
7655 RExC_paren_name_list = NULL;
7659 RExC_mysv1= sv_newmortal();
7660 RExC_mysv2= sv_newmortal();
7664 SV *dsv= sv_newmortal();
7665 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, PL_dump_re_max_len);
7666 Perl_re_printf( aTHX_ "%sCompiling REx%s %s\n",
7667 PL_colors[4], PL_colors[5], s);
7670 /* we jump here if we have to recompile, e.g., from upgrading the pattern
7673 if ((pm_flags & PMf_USE_RE_EVAL)
7674 /* this second condition covers the non-regex literal case,
7675 * i.e. $foo =~ '(?{})'. */
7676 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
7678 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
7681 /* return old regex if pattern hasn't changed */
7682 /* XXX: note in the below we have to check the flags as well as the
7685 * Things get a touch tricky as we have to compare the utf8 flag
7686 * independently from the compile flags. */
7690 && !!RX_UTF8(old_re) == !!RExC_utf8
7691 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
7692 && RX_PRECOMP(old_re)
7693 && RX_PRELEN(old_re) == plen
7694 && memEQ(RX_PRECOMP(old_re), exp, plen)
7695 && !runtime_code /* with runtime code, always recompile */ )
7698 SV *dsv= sv_newmortal();
7699 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, PL_dump_re_max_len);
7700 Perl_re_printf( aTHX_ "%sSkipping recompilation of unchanged REx%s %s\n",
7701 PL_colors[4], PL_colors[5], s);
7706 /* Allocate the pattern's SV */
7707 RExC_rx_sv = Rx = (REGEXP*) newSV_type(SVt_REGEXP);
7708 RExC_rx = ReANY(Rx);
7709 if ( RExC_rx == NULL )
7710 FAIL("Regexp out of space");
7712 rx_flags = orig_rx_flags;
7714 if ( (UTF || RExC_uni_semantics)
7715 && initial_charset == REGEX_DEPENDS_CHARSET)
7718 /* Set to use unicode semantics if the pattern is in utf8 and has the
7719 * 'depends' charset specified, as it means unicode when utf8 */
7720 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
7721 RExC_uni_semantics = 1;
7724 RExC_pm_flags = pm_flags;
7727 assert(TAINTING_get || !TAINT_get);
7729 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
7731 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
7732 /* whoops, we have a non-utf8 pattern, whilst run-time code
7733 * got compiled as utf8. Try again with a utf8 pattern */
7734 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7735 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7739 assert(!pRExC_state->runtime_code_qr);
7745 RExC_in_lookbehind = 0;
7746 RExC_in_lookahead = 0;
7747 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
7748 RExC_recode_x_to_native = 0;
7749 RExC_in_multi_char_class = 0;
7751 RExC_start = RExC_copy_start_in_constructed = RExC_copy_start_in_input = RExC_precomp = exp;
7752 RExC_precomp_end = RExC_end = exp + plen;
7754 RExC_whilem_seen = 0;
7756 RExC_recurse = NULL;
7757 RExC_study_chunk_recursed = NULL;
7758 RExC_study_chunk_recursed_bytes= 0;
7759 RExC_recurse_count = 0;
7760 RExC_sets_depth = 0;
7761 pRExC_state->code_index = 0;
7763 /* Initialize the string in the compiled pattern. This is so that there is
7764 * something to output if necessary */
7765 set_regex_pv(pRExC_state, Rx);
7768 Perl_re_printf( aTHX_
7769 "Starting parse and generation\n");
7771 RExC_lastparse=NULL;
7774 /* Allocate space and zero-initialize. Note, the two step process
7775 of zeroing when in debug mode, thus anything assigned has to
7776 happen after that */
7779 /* On the first pass of the parse, we guess how big this will be. Then
7780 * we grow in one operation to that amount and then give it back. As
7781 * we go along, we re-allocate what we need.
7783 * XXX Currently the guess is essentially that the pattern will be an
7784 * EXACT node with one byte input, one byte output. This is crude, and
7785 * better heuristics are welcome.
7787 * On any subsequent passes, we guess what we actually computed in the
7788 * latest earlier pass. Such a pass probably didn't complete so is
7789 * missing stuff. We could improve those guesses by knowing where the
7790 * parse stopped, and use the length so far plus apply the above
7791 * assumption to what's left. */
7792 RExC_size = STR_SZ(RExC_end - RExC_start);
7795 Newxc(RExC_rxi, sizeof(regexp_internal) + RExC_size, char, regexp_internal);
7796 if ( RExC_rxi == NULL )
7797 FAIL("Regexp out of space");
7799 Zero(RExC_rxi, sizeof(regexp_internal) + RExC_size, char);
7800 RXi_SET( RExC_rx, RExC_rxi );
7802 /* We start from 0 (over from 0 in the case this is a reparse. The first
7803 * node parsed will give back any excess memory we have allocated so far).
7807 /* non-zero initialization begins here */
7808 RExC_rx->engine= eng;
7809 RExC_rx->extflags = rx_flags;
7810 RXp_COMPFLAGS(RExC_rx) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
7812 if (pm_flags & PMf_IS_QR) {
7813 RExC_rxi->code_blocks = pRExC_state->code_blocks;
7814 if (RExC_rxi->code_blocks) {
7815 RExC_rxi->code_blocks->refcnt++;
7819 RExC_rx->intflags = 0;
7821 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
7824 /* This NUL is guaranteed because the pattern comes from an SV*, and the sv
7825 * code makes sure the final byte is an uncounted NUL. But should this
7826 * ever not be the case, lots of things could read beyond the end of the
7827 * buffer: loops like
7828 * while(isFOO(*RExC_parse)) RExC_parse++;
7829 * strchr(RExC_parse, "foo");
7830 * etc. So it is worth noting. */
7831 assert(*RExC_end == '\0');
7835 RExC_parens_buf_size = 0;
7836 RExC_emit_start = RExC_rxi->program;
7837 pRExC_state->code_index = 0;
7839 *((char*) RExC_emit_start) = (char) REG_MAGIC;
7843 if (reg(pRExC_state, 0, &flags, 1)) {
7845 /* Success!, But we may need to redo the parse knowing how many parens
7846 * there actually are */
7847 if (IN_PARENS_PASS) {
7848 flags |= RESTART_PARSE;
7851 /* We have that number in RExC_npar */
7852 RExC_total_parens = RExC_npar;
7854 else if (! MUST_RESTART(flags)) {
7856 Perl_croak(aTHX_ "panic: reg returned failure to re_op_compile, flags=%#" UVxf, (UV) flags);
7859 /* Here, we either have success, or we have to redo the parse for some reason */
7860 if (MUST_RESTART(flags)) {
7862 /* It's possible to write a regexp in ascii that represents Unicode
7863 codepoints outside of the byte range, such as via \x{100}. If we
7864 detect such a sequence we have to convert the entire pattern to utf8
7865 and then recompile, as our sizing calculation will have been based
7866 on 1 byte == 1 character, but we will need to use utf8 to encode
7867 at least some part of the pattern, and therefore must convert the whole
7870 if (flags & NEED_UTF8) {
7872 /* We have stored the offset of the final warning output so far.
7873 * That must be adjusted. Any variant characters between the start
7874 * of the pattern and this warning count for 2 bytes in the final,
7875 * so just add them again */
7876 if (UNLIKELY(RExC_latest_warn_offset > 0)) {
7877 RExC_latest_warn_offset +=
7878 variant_under_utf8_count((U8 *) exp, (U8 *) exp
7879 + RExC_latest_warn_offset);
7881 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7882 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7883 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "Need to redo parse after upgrade\n"));
7886 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "Need to redo parse\n"));
7889 if (ALL_PARENS_COUNTED) {
7890 /* Make enough room for all the known parens, and zero it */
7891 Renew(RExC_open_parens, RExC_total_parens, regnode_offset);
7892 Zero(RExC_open_parens, RExC_total_parens, regnode_offset);
7893 RExC_open_parens[0] = 1; /* +1 for REG_MAGIC */
7895 Renew(RExC_close_parens, RExC_total_parens, regnode_offset);
7896 Zero(RExC_close_parens, RExC_total_parens, regnode_offset);
7898 else { /* Parse did not complete. Reinitialize the parentheses
7900 RExC_total_parens = 0;
7901 if (RExC_open_parens) {
7902 Safefree(RExC_open_parens);
7903 RExC_open_parens = NULL;
7905 if (RExC_close_parens) {
7906 Safefree(RExC_close_parens);
7907 RExC_close_parens = NULL;
7911 /* Clean up what we did in this parse */
7912 SvREFCNT_dec_NN(RExC_rx_sv);
7917 /* Here, we have successfully parsed and generated the pattern's program
7918 * for the regex engine. We are ready to finish things up and look for
7921 /* Update the string to compile, with correct modifiers, etc */
7922 set_regex_pv(pRExC_state, Rx);
7924 RExC_rx->nparens = RExC_total_parens - 1;
7926 /* Uses the upper 4 bits of the FLAGS field, so keep within that size */
7927 if (RExC_whilem_seen > 15)
7928 RExC_whilem_seen = 15;
7931 Perl_re_printf( aTHX_
7932 "Required size %" IVdf " nodes\n", (IV)RExC_size);
7934 RExC_lastparse=NULL;
7937 #ifdef RE_TRACK_PATTERN_OFFSETS
7938 DEBUG_OFFSETS_r(Perl_re_printf( aTHX_
7939 "%s %" UVuf " bytes for offset annotations.\n",
7940 RExC_offsets ? "Got" : "Couldn't get",
7941 (UV)((RExC_offsets[0] * 2 + 1))));
7942 DEBUG_OFFSETS_r(if (RExC_offsets) {
7943 const STRLEN len = RExC_offsets[0];
7945 DECLARE_AND_GET_RE_DEBUG_FLAGS;
7946 Perl_re_printf( aTHX_
7947 "Offsets: [%" UVuf "]\n\t", (UV)RExC_offsets[0]);
7948 for (i = 1; i <= len; i++) {
7949 if (RExC_offsets[i*2-1] || RExC_offsets[i*2])
7950 Perl_re_printf( aTHX_ "%" UVuf ":%" UVuf "[%" UVuf "] ",
7951 (UV)i, (UV)RExC_offsets[i*2-1], (UV)RExC_offsets[i*2]);
7953 Perl_re_printf( aTHX_ "\n");
7957 SetProgLen(RExC_rxi,RExC_size);
7960 DEBUG_DUMP_PRE_OPTIMIZE_r({
7961 SV * const sv = sv_newmortal();
7962 RXi_GET_DECL(RExC_rx, ri);
7964 Perl_re_printf( aTHX_ "Program before optimization:\n");
7966 (void)dumpuntil(RExC_rx, ri->program, ri->program + 1, NULL, NULL,
7971 Perl_re_printf( aTHX_ "Starting post parse optimization\n");
7974 /* XXXX To minimize changes to RE engine we always allocate
7975 3-units-long substrs field. */
7976 Newx(RExC_rx->substrs, 1, struct reg_substr_data);
7977 if (RExC_recurse_count) {
7978 Newx(RExC_recurse, RExC_recurse_count, regnode *);
7979 SAVEFREEPV(RExC_recurse);
7982 if (RExC_seen & REG_RECURSE_SEEN) {
7983 /* Note, RExC_total_parens is 1 + the number of parens in a pattern.
7984 * So its 1 if there are no parens. */
7985 RExC_study_chunk_recursed_bytes= (RExC_total_parens >> 3) +
7986 ((RExC_total_parens & 0x07) != 0);
7987 Newx(RExC_study_chunk_recursed,
7988 RExC_study_chunk_recursed_bytes * RExC_total_parens, U8);
7989 SAVEFREEPV(RExC_study_chunk_recursed);
7993 RExC_rx->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
7995 RExC_study_chunk_recursed_count= 0;
7997 Zero(RExC_rx->substrs, 1, struct reg_substr_data);
7998 if (RExC_study_chunk_recursed) {
7999 Zero(RExC_study_chunk_recursed,
8000 RExC_study_chunk_recursed_bytes * RExC_total_parens, U8);
8004 #ifdef TRIE_STUDY_OPT
8006 StructCopy(&zero_scan_data, &data, scan_data_t);
8007 copyRExC_state = RExC_state;
8010 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "Restudying\n"));
8012 RExC_state = copyRExC_state;
8013 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
8014 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
8016 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
8017 StructCopy(&zero_scan_data, &data, scan_data_t);
8020 StructCopy(&zero_scan_data, &data, scan_data_t);
8023 /* Dig out information for optimizations. */
8024 RExC_rx->extflags = RExC_flags; /* was pm_op */
8025 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
8028 SvUTF8_on(Rx); /* Unicode in it? */
8029 RExC_rxi->regstclass = NULL;
8030 if (RExC_naughty >= TOO_NAUGHTY) /* Probably an expensive pattern. */
8031 RExC_rx->intflags |= PREGf_NAUGHTY;
8032 scan = RExC_rxi->program + 1; /* First BRANCH. */
8034 /* testing for BRANCH here tells us whether there is "must appear"
8035 data in the pattern. If there is then we can use it for optimisations */
8036 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
8039 STRLEN longest_length[2];
8040 regnode_ssc ch_class; /* pointed to by data */
8042 SSize_t last_close = 0; /* pointed to by data */
8043 regnode *first= scan;
8044 regnode *first_next= regnext(first);
8048 * Skip introductions and multiplicators >= 1
8049 * so that we can extract the 'meat' of the pattern that must
8050 * match in the large if() sequence following.
8051 * NOTE that EXACT is NOT covered here, as it is normally
8052 * picked up by the optimiser separately.
8054 * This is unfortunate as the optimiser isnt handling lookahead
8055 * properly currently.
8058 while ((OP(first) == OPEN && (sawopen = 1)) ||
8059 /* An OR of *one* alternative - should not happen now. */
8060 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
8061 /* for now we can't handle lookbehind IFMATCH*/
8062 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
8063 (OP(first) == PLUS) ||
8064 (OP(first) == MINMOD) ||
8065 /* An {n,m} with n>0 */
8066 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
8067 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
8070 * the only op that could be a regnode is PLUS, all the rest
8071 * will be regnode_1 or regnode_2.
8073 * (yves doesn't think this is true)
8075 if (OP(first) == PLUS)
8078 if (OP(first) == MINMOD)
8080 first += regarglen[OP(first)];
8082 first = NEXTOPER(first);
8083 first_next= regnext(first);
8086 /* Starting-point info. */
8088 DEBUG_PEEP("first:", first, 0, 0);
8089 /* Ignore EXACT as we deal with it later. */
8090 if (PL_regkind[OP(first)] == EXACT) {
8091 if ( OP(first) == EXACT
8092 || OP(first) == LEXACT
8093 || OP(first) == EXACT_REQ8
8094 || OP(first) == LEXACT_REQ8
8095 || OP(first) == EXACTL)
8097 NOOP; /* Empty, get anchored substr later. */
8100 RExC_rxi->regstclass = first;
8103 else if (PL_regkind[OP(first)] == TRIE &&
8104 ((reg_trie_data *)RExC_rxi->data->data[ ARG(first) ])->minlen>0)
8106 /* this can happen only on restudy */
8107 RExC_rxi->regstclass = construct_ahocorasick_from_trie(pRExC_state, (regnode *)first, 0);
8110 else if (REGNODE_SIMPLE(OP(first)))
8111 RExC_rxi->regstclass = first;
8112 else if (PL_regkind[OP(first)] == BOUND ||
8113 PL_regkind[OP(first)] == NBOUND)
8114 RExC_rxi->regstclass = first;
8115 else if (PL_regkind[OP(first)] == BOL) {
8116 RExC_rx->intflags |= (OP(first) == MBOL
8119 first = NEXTOPER(first);
8122 else if (OP(first) == GPOS) {
8123 RExC_rx->intflags |= PREGf_ANCH_GPOS;
8124 first = NEXTOPER(first);
8127 else if ((!sawopen || !RExC_sawback) &&
8129 (OP(first) == STAR &&
8130 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
8131 !(RExC_rx->intflags & PREGf_ANCH) && !pRExC_state->code_blocks)
8133 /* turn .* into ^.* with an implied $*=1 */
8135 (OP(NEXTOPER(first)) == REG_ANY)
8138 RExC_rx->intflags |= (type | PREGf_IMPLICIT);
8139 first = NEXTOPER(first);
8142 if (sawplus && !sawminmod && !sawlookahead
8143 && (!sawopen || !RExC_sawback)
8144 && !pRExC_state->code_blocks) /* May examine pos and $& */
8145 /* x+ must match at the 1st pos of run of x's */
8146 RExC_rx->intflags |= PREGf_SKIP;
8148 /* Scan is after the zeroth branch, first is atomic matcher. */
8149 #ifdef TRIE_STUDY_OPT
8152 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
8153 (IV)(first - scan + 1))
8157 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
8158 (IV)(first - scan + 1))
8164 * If there's something expensive in the r.e., find the
8165 * longest literal string that must appear and make it the
8166 * regmust. Resolve ties in favor of later strings, since
8167 * the regstart check works with the beginning of the r.e.
8168 * and avoiding duplication strengthens checking. Not a
8169 * strong reason, but sufficient in the absence of others.
8170 * [Now we resolve ties in favor of the earlier string if
8171 * it happens that c_offset_min has been invalidated, since the
8172 * earlier string may buy us something the later one won't.]
8175 data.substrs[0].str = newSVpvs("");
8176 data.substrs[1].str = newSVpvs("");
8177 data.last_found = newSVpvs("");
8178 data.cur_is_floating = 0; /* initially any found substring is fixed */
8179 ENTER_with_name("study_chunk");
8180 SAVEFREESV(data.substrs[0].str);
8181 SAVEFREESV(data.substrs[1].str);
8182 SAVEFREESV(data.last_found);
8184 if (!RExC_rxi->regstclass) {
8185 ssc_init(pRExC_state, &ch_class);
8186 data.start_class = &ch_class;
8187 stclass_flag = SCF_DO_STCLASS_AND;
8188 } else /* XXXX Check for BOUND? */
8190 data.last_closep = &last_close;
8194 * MAIN ENTRY FOR study_chunk() FOR m/PATTERN/
8195 * (NO top level branches)
8197 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
8198 scan + RExC_size, /* Up to end */
8200 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
8201 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
8205 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
8208 if ( RExC_total_parens == 1 && !data.cur_is_floating
8209 && data.last_start_min == 0 && data.last_end > 0
8210 && !RExC_seen_zerolen
8211 && !(RExC_seen & REG_VERBARG_SEEN)
8212 && !(RExC_seen & REG_GPOS_SEEN)
8214 RExC_rx->extflags |= RXf_CHECK_ALL;
8216 scan_commit(pRExC_state, &data,&minlen, 0);
8219 /* XXX this is done in reverse order because that's the way the
8220 * code was before it was parameterised. Don't know whether it
8221 * actually needs doing in reverse order. DAPM */
8222 for (i = 1; i >= 0; i--) {
8223 longest_length[i] = CHR_SVLEN(data.substrs[i].str);
8226 && SvCUR(data.substrs[0].str) /* ok to leave SvCUR */
8227 && data.substrs[0].min_offset
8228 == data.substrs[1].min_offset
8229 && SvCUR(data.substrs[0].str)
8230 == SvCUR(data.substrs[1].str)
8232 && S_setup_longest (aTHX_ pRExC_state,
8233 &(RExC_rx->substrs->data[i]),
8237 RExC_rx->substrs->data[i].min_offset =
8238 data.substrs[i].min_offset - data.substrs[i].lookbehind;
8240 RExC_rx->substrs->data[i].max_offset = data.substrs[i].max_offset;
8241 /* Don't offset infinity */
8242 if (data.substrs[i].max_offset < OPTIMIZE_INFTY)
8243 RExC_rx->substrs->data[i].max_offset -= data.substrs[i].lookbehind;
8244 SvREFCNT_inc_simple_void_NN(data.substrs[i].str);
8247 RExC_rx->substrs->data[i].substr = NULL;
8248 RExC_rx->substrs->data[i].utf8_substr = NULL;
8249 longest_length[i] = 0;
8253 LEAVE_with_name("study_chunk");
8255 if (RExC_rxi->regstclass
8256 && (OP(RExC_rxi->regstclass) == REG_ANY || OP(RExC_rxi->regstclass) == SANY))
8257 RExC_rxi->regstclass = NULL;
8259 if ((!(RExC_rx->substrs->data[0].substr || RExC_rx->substrs->data[0].utf8_substr)
8260 || RExC_rx->substrs->data[0].min_offset)
8262 && ! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
8263 && is_ssc_worth_it(pRExC_state, data.start_class))
8265 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
8267 ssc_finalize(pRExC_state, data.start_class);
8269 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
8270 StructCopy(data.start_class,
8271 (regnode_ssc*)RExC_rxi->data->data[n],
8273 RExC_rxi->regstclass = (regnode*)RExC_rxi->data->data[n];
8274 RExC_rx->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
8275 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
8276 regprop(RExC_rx, sv, (regnode*)data.start_class, NULL, pRExC_state);
8277 Perl_re_printf( aTHX_
8278 "synthetic stclass \"%s\".\n",
8279 SvPVX_const(sv));});
8280 data.start_class = NULL;
8283 /* A temporary algorithm prefers floated substr to fixed one of
8284 * same length to dig more info. */
8285 i = (longest_length[0] <= longest_length[1]);
8286 RExC_rx->substrs->check_ix = i;
8287 RExC_rx->check_end_shift = RExC_rx->substrs->data[i].end_shift;
8288 RExC_rx->check_substr = RExC_rx->substrs->data[i].substr;
8289 RExC_rx->check_utf8 = RExC_rx->substrs->data[i].utf8_substr;
8290 RExC_rx->check_offset_min = RExC_rx->substrs->data[i].min_offset;
8291 RExC_rx->check_offset_max = RExC_rx->substrs->data[i].max_offset;
8292 if (!i && (RExC_rx->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS)))
8293 RExC_rx->intflags |= PREGf_NOSCAN;
8295 if ((RExC_rx->check_substr || RExC_rx->check_utf8) ) {
8296 RExC_rx->extflags |= RXf_USE_INTUIT;
8297 if (SvTAIL(RExC_rx->check_substr ? RExC_rx->check_substr : RExC_rx->check_utf8))
8298 RExC_rx->extflags |= RXf_INTUIT_TAIL;
8301 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
8302 if ( (STRLEN)minlen < longest_length[1] )
8303 minlen= longest_length[1];
8304 if ( (STRLEN)minlen < longest_length[0] )
8305 minlen= longest_length[0];
8309 /* Several toplevels. Best we can is to set minlen. */
8311 regnode_ssc ch_class;
8312 SSize_t last_close = 0;
8314 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "\nMulti Top Level\n"));
8316 scan = RExC_rxi->program + 1;
8317 ssc_init(pRExC_state, &ch_class);
8318 data.start_class = &ch_class;
8319 data.last_closep = &last_close;
8323 * MAIN ENTRY FOR study_chunk() FOR m/P1|P2|.../
8324 * (patterns WITH top level branches)
8326 minlen = study_chunk(pRExC_state,
8327 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
8328 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
8329 ? SCF_TRIE_DOING_RESTUDY
8333 CHECK_RESTUDY_GOTO_butfirst(NOOP);
8335 RExC_rx->check_substr = NULL;
8336 RExC_rx->check_utf8 = NULL;
8337 RExC_rx->substrs->data[0].substr = NULL;
8338 RExC_rx->substrs->data[0].utf8_substr = NULL;
8339 RExC_rx->substrs->data[1].substr = NULL;
8340 RExC_rx->substrs->data[1].utf8_substr = NULL;
8342 if (! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
8343 && is_ssc_worth_it(pRExC_state, data.start_class))
8345 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
8347 ssc_finalize(pRExC_state, data.start_class);
8349 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
8350 StructCopy(data.start_class,
8351 (regnode_ssc*)RExC_rxi->data->data[n],
8353 RExC_rxi->regstclass = (regnode*)RExC_rxi->data->data[n];
8354 RExC_rx->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
8355 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
8356 regprop(RExC_rx, sv, (regnode*)data.start_class, NULL, pRExC_state);
8357 Perl_re_printf( aTHX_
8358 "synthetic stclass \"%s\".\n",
8359 SvPVX_const(sv));});
8360 data.start_class = NULL;
8364 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
8365 RExC_rx->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
8366 RExC_rx->maxlen = REG_INFTY;
8369 RExC_rx->maxlen = RExC_maxlen;
8372 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
8373 the "real" pattern. */
8375 Perl_re_printf( aTHX_ "minlen: %" IVdf " RExC_rx->minlen:%" IVdf " maxlen:%" IVdf "\n",
8376 (IV)minlen, (IV)RExC_rx->minlen, (IV)RExC_maxlen);
8378 RExC_rx->minlenret = minlen;
8379 if (RExC_rx->minlen < minlen)
8380 RExC_rx->minlen = minlen;
8382 if (RExC_seen & REG_RECURSE_SEEN ) {
8383 RExC_rx->intflags |= PREGf_RECURSE_SEEN;
8384 Newx(RExC_rx->recurse_locinput, RExC_rx->nparens + 1, char *);
8386 if (RExC_seen & REG_GPOS_SEEN)
8387 RExC_rx->intflags |= PREGf_GPOS_SEEN;
8388 if (RExC_seen & REG_LOOKBEHIND_SEEN)
8389 RExC_rx->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
8391 if (pRExC_state->code_blocks)
8392 RExC_rx->extflags |= RXf_EVAL_SEEN;
8393 if (RExC_seen & REG_VERBARG_SEEN)
8395 RExC_rx->intflags |= PREGf_VERBARG_SEEN;
8396 RExC_rx->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
8398 if (RExC_seen & REG_CUTGROUP_SEEN)
8399 RExC_rx->intflags |= PREGf_CUTGROUP_SEEN;
8400 if (pm_flags & PMf_USE_RE_EVAL)
8401 RExC_rx->intflags |= PREGf_USE_RE_EVAL;
8402 if (RExC_paren_names)
8403 RXp_PAREN_NAMES(RExC_rx) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
8405 RXp_PAREN_NAMES(RExC_rx) = NULL;
8407 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
8408 * so it can be used in pp.c */
8409 if (RExC_rx->intflags & PREGf_ANCH)
8410 RExC_rx->extflags |= RXf_IS_ANCHORED;
8414 /* this is used to identify "special" patterns that might result
8415 * in Perl NOT calling the regex engine and instead doing the match "itself",
8416 * particularly special cases in split//. By having the regex compiler
8417 * do this pattern matching at a regop level (instead of by inspecting the pattern)
8418 * we avoid weird issues with equivalent patterns resulting in different behavior,
8419 * AND we allow non Perl engines to get the same optimizations by the setting the
8420 * flags appropriately - Yves */
8421 regnode *first = RExC_rxi->program + 1;
8423 regnode *next = regnext(first);
8426 if (PL_regkind[fop] == NOTHING && nop == END)
8427 RExC_rx->extflags |= RXf_NULL;
8428 else if ((fop == MBOL || (fop == SBOL && !first->flags)) && nop == END)
8429 /* when fop is SBOL first->flags will be true only when it was
8430 * produced by parsing /\A/, and not when parsing /^/. This is
8431 * very important for the split code as there we want to
8432 * treat /^/ as /^/m, but we do not want to treat /\A/ as /^/m.
8433 * See rt #122761 for more details. -- Yves */
8434 RExC_rx->extflags |= RXf_START_ONLY;
8435 else if (fop == PLUS
8436 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
8438 RExC_rx->extflags |= RXf_WHITE;
8439 else if ( RExC_rx->extflags & RXf_SPLIT
8440 && ( fop == EXACT || fop == LEXACT
8441 || fop == EXACT_REQ8 || fop == LEXACT_REQ8
8443 && STR_LEN(first) == 1
8444 && *(STRING(first)) == ' '
8446 RExC_rx->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
8450 if (RExC_contains_locale) {
8451 RXp_EXTFLAGS(RExC_rx) |= RXf_TAINTED;
8455 if (RExC_paren_names) {
8456 RExC_rxi->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
8457 RExC_rxi->data->data[RExC_rxi->name_list_idx]
8458 = (void*)SvREFCNT_inc(RExC_paren_name_list);
8461 RExC_rxi->name_list_idx = 0;
8463 while ( RExC_recurse_count > 0 ) {
8464 const regnode *scan = RExC_recurse[ --RExC_recurse_count ];
8466 * This data structure is set up in study_chunk() and is used
8467 * to calculate the distance between a GOSUB regopcode and
8468 * the OPEN/CURLYM (CURLYM's are special and can act like OPEN's)
8471 * If for some reason someone writes code that optimises
8472 * away a GOSUB opcode then the assert should be changed to
8473 * an if(scan) to guard the ARG2L_SET() - Yves
8476 assert(scan && OP(scan) == GOSUB);
8477 ARG2L_SET( scan, RExC_open_parens[ARG(scan)] - REGNODE_OFFSET(scan));
8480 Newxz(RExC_rx->offs, RExC_total_parens, regexp_paren_pair);
8481 /* assume we don't need to swap parens around before we match */
8483 Perl_re_printf( aTHX_ "study_chunk_recursed_count: %lu\n",
8484 (unsigned long)RExC_study_chunk_recursed_count);
8488 Perl_re_printf( aTHX_ "Final program:\n");
8492 if (RExC_open_parens) {
8493 Safefree(RExC_open_parens);
8494 RExC_open_parens = NULL;
8496 if (RExC_close_parens) {
8497 Safefree(RExC_close_parens);
8498 RExC_close_parens = NULL;
8502 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
8503 * by setting the regexp SV to readonly-only instead. If the
8504 * pattern's been recompiled, the USEDness should remain. */
8505 if (old_re && SvREADONLY(old_re))
8513 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
8516 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
8518 PERL_UNUSED_ARG(value);
8520 if (flags & RXapif_FETCH) {
8521 return reg_named_buff_fetch(rx, key, flags);
8522 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
8523 Perl_croak_no_modify();
8525 } else if (flags & RXapif_EXISTS) {
8526 return reg_named_buff_exists(rx, key, flags)
8529 } else if (flags & RXapif_REGNAMES) {
8530 return reg_named_buff_all(rx, flags);
8531 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
8532 return reg_named_buff_scalar(rx, flags);
8534 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
8540 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
8543 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
8544 PERL_UNUSED_ARG(lastkey);
8546 if (flags & RXapif_FIRSTKEY)
8547 return reg_named_buff_firstkey(rx, flags);
8548 else if (flags & RXapif_NEXTKEY)
8549 return reg_named_buff_nextkey(rx, flags);
8551 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
8558 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
8562 struct regexp *const rx = ReANY(r);
8564 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
8566 if (rx && RXp_PAREN_NAMES(rx)) {
8567 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
8570 SV* sv_dat=HeVAL(he_str);
8571 I32 *nums=(I32*)SvPVX(sv_dat);
8572 AV * const retarray = (flags & RXapif_ALL) ? newAV() : NULL;
8573 for ( i=0; i<SvIVX(sv_dat); i++ ) {
8574 if ((I32)(rx->nparens) >= nums[i]
8575 && rx->offs[nums[i]].start != -1
8576 && rx->offs[nums[i]].end != -1)
8579 CALLREG_NUMBUF_FETCH(r, nums[i], ret);
8584 ret = newSVsv(&PL_sv_undef);
8587 av_push(retarray, ret);
8590 return newRV_noinc(MUTABLE_SV(retarray));
8597 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
8600 struct regexp *const rx = ReANY(r);
8602 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
8604 if (rx && RXp_PAREN_NAMES(rx)) {
8605 if (flags & RXapif_ALL) {
8606 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
8608 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
8610 SvREFCNT_dec_NN(sv);
8622 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
8624 struct regexp *const rx = ReANY(r);
8626 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
8628 if ( rx && RXp_PAREN_NAMES(rx) ) {
8629 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
8631 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
8638 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
8640 struct regexp *const rx = ReANY(r);
8641 DECLARE_AND_GET_RE_DEBUG_FLAGS;
8643 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
8645 if (rx && RXp_PAREN_NAMES(rx)) {
8646 HV *hv = RXp_PAREN_NAMES(rx);
8648 while ( (temphe = hv_iternext_flags(hv, 0)) ) {
8651 SV* sv_dat = HeVAL(temphe);
8652 I32 *nums = (I32*)SvPVX(sv_dat);
8653 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8654 if ((I32)(rx->lastparen) >= nums[i] &&
8655 rx->offs[nums[i]].start != -1 &&
8656 rx->offs[nums[i]].end != -1)
8662 if (parno || flags & RXapif_ALL) {
8663 return newSVhek(HeKEY_hek(temphe));
8671 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
8676 struct regexp *const rx = ReANY(r);
8678 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
8680 if (rx && RXp_PAREN_NAMES(rx)) {
8681 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
8682 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
8683 } else if (flags & RXapif_ONE) {
8684 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
8685 av = MUTABLE_AV(SvRV(ret));
8686 length = av_tindex(av);
8687 SvREFCNT_dec_NN(ret);
8688 return newSViv(length + 1);
8690 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
8695 return &PL_sv_undef;
8699 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
8701 struct regexp *const rx = ReANY(r);
8704 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
8706 if (rx && RXp_PAREN_NAMES(rx)) {
8707 HV *hv= RXp_PAREN_NAMES(rx);
8709 (void)hv_iterinit(hv);
8710 while ( (temphe = hv_iternext_flags(hv, 0)) ) {
8713 SV* sv_dat = HeVAL(temphe);
8714 I32 *nums = (I32*)SvPVX(sv_dat);
8715 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8716 if ((I32)(rx->lastparen) >= nums[i] &&
8717 rx->offs[nums[i]].start != -1 &&
8718 rx->offs[nums[i]].end != -1)
8724 if (parno || flags & RXapif_ALL) {
8725 av_push(av, newSVhek(HeKEY_hek(temphe)));
8730 return newRV_noinc(MUTABLE_SV(av));
8734 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
8737 struct regexp *const rx = ReANY(r);
8743 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
8745 if ( n == RX_BUFF_IDX_CARET_PREMATCH
8746 || n == RX_BUFF_IDX_CARET_FULLMATCH
8747 || n == RX_BUFF_IDX_CARET_POSTMATCH
8750 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8752 /* on something like
8755 * the KEEPCOPY is set on the PMOP rather than the regex */
8756 if (PL_curpm && r == PM_GETRE(PL_curpm))
8757 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8766 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
8767 /* no need to distinguish between them any more */
8768 n = RX_BUFF_IDX_FULLMATCH;
8770 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
8771 && rx->offs[0].start != -1)
8773 /* $`, ${^PREMATCH} */
8774 i = rx->offs[0].start;
8778 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
8779 && rx->offs[0].end != -1)
8781 /* $', ${^POSTMATCH} */
8782 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
8783 i = rx->sublen + rx->suboffset - rx->offs[0].end;
8786 if (inRANGE(n, 0, (I32)rx->nparens) &&
8787 (s1 = rx->offs[n].start) != -1 &&
8788 (t1 = rx->offs[n].end) != -1)
8790 /* $&, ${^MATCH}, $1 ... */
8792 s = rx->subbeg + s1 - rx->suboffset;
8797 assert(s >= rx->subbeg);
8798 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
8800 #ifdef NO_TAINT_SUPPORT
8801 sv_setpvn(sv, s, i);
8803 const int oldtainted = TAINT_get;
8805 sv_setpvn(sv, s, i);
8806 TAINT_set(oldtainted);
8808 if (RXp_MATCH_UTF8(rx))
8813 if (RXp_MATCH_TAINTED(rx)) {
8814 if (SvTYPE(sv) >= SVt_PVMG) {
8815 MAGIC* const mg = SvMAGIC(sv);
8818 SvMAGIC_set(sv, mg->mg_moremagic);
8820 if ((mgt = SvMAGIC(sv))) {
8821 mg->mg_moremagic = mgt;
8822 SvMAGIC_set(sv, mg);
8839 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
8840 SV const * const value)
8842 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
8844 PERL_UNUSED_ARG(rx);
8845 PERL_UNUSED_ARG(paren);
8846 PERL_UNUSED_ARG(value);
8849 Perl_croak_no_modify();
8853 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
8856 struct regexp *const rx = ReANY(r);
8860 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
8862 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
8863 || paren == RX_BUFF_IDX_CARET_FULLMATCH
8864 || paren == RX_BUFF_IDX_CARET_POSTMATCH
8867 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8869 /* on something like
8872 * the KEEPCOPY is set on the PMOP rather than the regex */
8873 if (PL_curpm && r == PM_GETRE(PL_curpm))
8874 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8880 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
8882 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
8883 case RX_BUFF_IDX_PREMATCH: /* $` */
8884 if (rx->offs[0].start != -1) {
8885 i = rx->offs[0].start;
8894 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
8895 case RX_BUFF_IDX_POSTMATCH: /* $' */
8896 if (rx->offs[0].end != -1) {
8897 i = rx->sublen - rx->offs[0].end;
8899 s1 = rx->offs[0].end;
8906 default: /* $& / ${^MATCH}, $1, $2, ... */
8907 if (paren <= (I32)rx->nparens &&
8908 (s1 = rx->offs[paren].start) != -1 &&
8909 (t1 = rx->offs[paren].end) != -1)
8915 if (ckWARN(WARN_UNINITIALIZED))
8916 report_uninit((const SV *)sv);
8921 if (i > 0 && RXp_MATCH_UTF8(rx)) {
8922 const char * const s = rx->subbeg - rx->suboffset + s1;
8927 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
8934 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
8936 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
8937 PERL_UNUSED_ARG(rx);
8941 return newSVpvs("Regexp");
8944 /* Scans the name of a named buffer from the pattern.
8945 * If flags is REG_RSN_RETURN_NULL returns null.
8946 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
8947 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
8948 * to the parsed name as looked up in the RExC_paren_names hash.
8949 * If there is an error throws a vFAIL().. type exception.
8952 #define REG_RSN_RETURN_NULL 0
8953 #define REG_RSN_RETURN_NAME 1
8954 #define REG_RSN_RETURN_DATA 2
8957 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
8959 char *name_start = RExC_parse;
8962 PERL_ARGS_ASSERT_REG_SCAN_NAME;
8964 assert (RExC_parse <= RExC_end);
8965 if (RExC_parse == RExC_end) NOOP;
8966 else if (isIDFIRST_lazy_if_safe(RExC_parse, RExC_end, UTF)) {
8967 /* Note that the code here assumes well-formed UTF-8. Skip IDFIRST by
8968 * using do...while */
8971 RExC_parse += UTF8SKIP(RExC_parse);
8972 } while ( RExC_parse < RExC_end
8973 && isWORDCHAR_utf8_safe((U8*)RExC_parse, (U8*) RExC_end));
8977 } while (RExC_parse < RExC_end && isWORDCHAR(*RExC_parse));
8979 RExC_parse++; /* so the <- from the vFAIL is after the offending
8981 vFAIL("Group name must start with a non-digit word character");
8983 sv_name = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
8984 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
8985 if ( flags == REG_RSN_RETURN_NAME)
8987 else if (flags==REG_RSN_RETURN_DATA) {
8990 if ( ! sv_name ) /* should not happen*/
8991 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
8992 if (RExC_paren_names)
8993 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
8995 sv_dat = HeVAL(he_str);
8996 if ( ! sv_dat ) { /* Didn't find group */
8998 /* It might be a forward reference; we can't fail until we
8999 * know, by completing the parse to get all the groups, and
9001 if (ALL_PARENS_COUNTED) {
9002 vFAIL("Reference to nonexistent named group");
9005 REQUIRE_PARENS_PASS;
9011 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
9012 (unsigned long) flags);
9015 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
9016 if (RExC_lastparse!=RExC_parse) { \
9017 Perl_re_printf( aTHX_ "%s", \
9018 Perl_pv_pretty(aTHX_ RExC_mysv1, RExC_parse, \
9019 RExC_end - RExC_parse, 16, \
9021 PERL_PV_ESCAPE_UNI_DETECT | \
9022 PERL_PV_PRETTY_ELLIPSES | \
9023 PERL_PV_PRETTY_LTGT | \
9024 PERL_PV_ESCAPE_RE | \
9025 PERL_PV_PRETTY_EXACTSIZE \
9029 Perl_re_printf( aTHX_ "%16s",""); \
9031 if (RExC_lastnum!=RExC_emit) \
9032 Perl_re_printf( aTHX_ "|%4zu", RExC_emit); \
9034 Perl_re_printf( aTHX_ "|%4s",""); \
9035 Perl_re_printf( aTHX_ "|%*s%-4s", \
9036 (int)((depth*2)), "", \
9039 RExC_lastnum=RExC_emit; \
9040 RExC_lastparse=RExC_parse; \
9045 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
9046 DEBUG_PARSE_MSG((funcname)); \
9047 Perl_re_printf( aTHX_ "%4s","\n"); \
9049 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({\
9050 DEBUG_PARSE_MSG((funcname)); \
9051 Perl_re_printf( aTHX_ fmt "\n",args); \
9054 /* This section of code defines the inversion list object and its methods. The
9055 * interfaces are highly subject to change, so as much as possible is static to
9056 * this file. An inversion list is here implemented as a malloc'd C UV array
9057 * as an SVt_INVLIST scalar.
9059 * An inversion list for Unicode is an array of code points, sorted by ordinal
9060 * number. Each element gives the code point that begins a range that extends
9061 * up-to but not including the code point given by the next element. The final
9062 * element gives the first code point of a range that extends to the platform's
9063 * infinity. The even-numbered elements (invlist[0], invlist[2], invlist[4],
9064 * ...) give ranges whose code points are all in the inversion list. We say
9065 * that those ranges are in the set. The odd-numbered elements give ranges
9066 * whose code points are not in the inversion list, and hence not in the set.
9067 * Thus, element [0] is the first code point in the list. Element [1]
9068 * is the first code point beyond that not in the list; and element [2] is the
9069 * first code point beyond that that is in the list. In other words, the first
9070 * range is invlist[0]..(invlist[1]-1), and all code points in that range are
9071 * in the inversion list. The second range is invlist[1]..(invlist[2]-1), and
9072 * all code points in that range are not in the inversion list. The third
9073 * range invlist[2]..(invlist[3]-1) gives code points that are in the inversion
9074 * list, and so forth. Thus every element whose index is divisible by two
9075 * gives the beginning of a range that is in the list, and every element whose
9076 * index is not divisible by two gives the beginning of a range not in the
9077 * list. If the final element's index is divisible by two, the inversion list
9078 * extends to the platform's infinity; otherwise the highest code point in the
9079 * inversion list is the contents of that element minus 1.
9081 * A range that contains just a single code point N will look like
9083 * invlist[i+1] == N+1
9085 * If N is UV_MAX (the highest representable code point on the machine), N+1 is
9086 * impossible to represent, so element [i+1] is omitted. The single element
9088 * invlist[0] == UV_MAX
9089 * contains just UV_MAX, but is interpreted as matching to infinity.
9091 * Taking the complement (inverting) an inversion list is quite simple, if the
9092 * first element is 0, remove it; otherwise add a 0 element at the beginning.
9093 * This implementation reserves an element at the beginning of each inversion
9094 * list to always contain 0; there is an additional flag in the header which
9095 * indicates if the list begins at the 0, or is offset to begin at the next
9096 * element. This means that the inversion list can be inverted without any
9097 * copying; just flip the flag.
9099 * More about inversion lists can be found in "Unicode Demystified"
9100 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
9102 * The inversion list data structure is currently implemented as an SV pointing
9103 * to an array of UVs that the SV thinks are bytes. This allows us to have an
9104 * array of UV whose memory management is automatically handled by the existing
9105 * facilities for SV's.
9107 * Some of the methods should always be private to the implementation, and some
9108 * should eventually be made public */
9110 /* The header definitions are in F<invlist_inline.h> */
9112 #ifndef PERL_IN_XSUB_RE
9114 PERL_STATIC_INLINE UV*
9115 S__invlist_array_init(SV* const invlist, const bool will_have_0)
9117 /* Returns a pointer to the first element in the inversion list's array.
9118 * This is called upon initialization of an inversion list. Where the
9119 * array begins depends on whether the list has the code point U+0000 in it
9120 * or not. The other parameter tells it whether the code that follows this
9121 * call is about to put a 0 in the inversion list or not. The first
9122 * element is either the element reserved for 0, if TRUE, or the element
9123 * after it, if FALSE */
9125 bool* offset = get_invlist_offset_addr(invlist);
9126 UV* zero_addr = (UV *) SvPVX(invlist);
9128 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
9131 assert(! _invlist_len(invlist));
9135 /* 1^1 = 0; 1^0 = 1 */
9136 *offset = 1 ^ will_have_0;
9137 return zero_addr + *offset;
9141 S_invlist_replace_list_destroys_src(pTHX_ SV * dest, SV * src)
9143 /* Replaces the inversion list in 'dest' with the one from 'src'. It
9144 * steals the list from 'src', so 'src' is made to have a NULL list. This
9145 * is similar to what SvSetMagicSV() would do, if it were implemented on
9146 * inversion lists, though this routine avoids a copy */
9148 const UV src_len = _invlist_len(src);
9149 const bool src_offset = *get_invlist_offset_addr(src);
9150 const STRLEN src_byte_len = SvLEN(src);
9151 char * array = SvPVX(src);
9153 const int oldtainted = TAINT_get;
9155 PERL_ARGS_ASSERT_INVLIST_REPLACE_LIST_DESTROYS_SRC;
9157 assert(is_invlist(src));
9158 assert(is_invlist(dest));
9159 assert(! invlist_is_iterating(src));
9160 assert(SvCUR(src) == 0 || SvCUR(src) < SvLEN(src));
9162 /* Make sure it ends in the right place with a NUL, as our inversion list
9163 * manipulations aren't careful to keep this true, but sv_usepvn_flags()
9165 array[src_byte_len - 1] = '\0';
9167 TAINT_NOT; /* Otherwise it breaks */
9168 sv_usepvn_flags(dest,
9172 /* This flag is documented to cause a copy to be avoided */
9173 SV_HAS_TRAILING_NUL);
9174 TAINT_set(oldtainted);
9179 /* Finish up copying over the other fields in an inversion list */
9180 *get_invlist_offset_addr(dest) = src_offset;
9181 invlist_set_len(dest, src_len, src_offset);
9182 *get_invlist_previous_index_addr(dest) = 0;
9183 invlist_iterfinish(dest);
9186 PERL_STATIC_INLINE IV*
9187 S_get_invlist_previous_index_addr(SV* invlist)
9189 /* Return the address of the IV that is reserved to hold the cached index
9191 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
9193 assert(is_invlist(invlist));
9195 return &(((XINVLIST*) SvANY(invlist))->prev_index);
9198 PERL_STATIC_INLINE IV
9199 S_invlist_previous_index(SV* const invlist)
9201 /* Returns cached index of previous search */
9203 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
9205 return *get_invlist_previous_index_addr(invlist);
9208 PERL_STATIC_INLINE void
9209 S_invlist_set_previous_index(SV* const invlist, const IV index)
9211 /* Caches <index> for later retrieval */
9213 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
9215 assert(index == 0 || index < (int) _invlist_len(invlist));
9217 *get_invlist_previous_index_addr(invlist) = index;
9220 PERL_STATIC_INLINE void
9221 S_invlist_trim(SV* invlist)
9223 /* Free the not currently-being-used space in an inversion list */
9225 /* But don't free up the space needed for the 0 UV that is always at the
9226 * beginning of the list, nor the trailing NUL */
9227 const UV min_size = TO_INTERNAL_SIZE(1) + 1;
9229 PERL_ARGS_ASSERT_INVLIST_TRIM;
9231 assert(is_invlist(invlist));
9233 SvPV_renew(invlist, MAX(min_size, SvCUR(invlist) + 1));
9236 PERL_STATIC_INLINE void
9237 S_invlist_clear(pTHX_ SV* invlist) /* Empty the inversion list */
9239 PERL_ARGS_ASSERT_INVLIST_CLEAR;
9241 assert(is_invlist(invlist));
9243 invlist_set_len(invlist, 0, 0);
9244 invlist_trim(invlist);
9247 #endif /* ifndef PERL_IN_XSUB_RE */
9249 PERL_STATIC_INLINE bool
9250 S_invlist_is_iterating(SV* const invlist)
9252 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
9254 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
9257 #ifndef PERL_IN_XSUB_RE
9259 PERL_STATIC_INLINE UV
9260 S_invlist_max(SV* const invlist)
9262 /* Returns the maximum number of elements storable in the inversion list's
9263 * array, without having to realloc() */
9265 PERL_ARGS_ASSERT_INVLIST_MAX;
9267 assert(is_invlist(invlist));
9269 /* Assumes worst case, in which the 0 element is not counted in the
9270 * inversion list, so subtracts 1 for that */
9271 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
9272 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
9273 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
9277 S_initialize_invlist_guts(pTHX_ SV* invlist, const Size_t initial_size)
9279 PERL_ARGS_ASSERT_INITIALIZE_INVLIST_GUTS;
9281 /* First 1 is in case the zero element isn't in the list; second 1 is for
9283 SvGROW(invlist, TO_INTERNAL_SIZE(initial_size + 1) + 1);
9284 invlist_set_len(invlist, 0, 0);
9286 /* Force iterinit() to be used to get iteration to work */
9287 invlist_iterfinish(invlist);
9289 *get_invlist_previous_index_addr(invlist) = 0;
9290 SvPOK_on(invlist); /* This allows B to extract the PV */
9294 Perl__new_invlist(pTHX_ IV initial_size)
9297 /* Return a pointer to a newly constructed inversion list, with enough
9298 * space to store 'initial_size' elements. If that number is negative, a
9299 * system default is used instead */
9303 if (initial_size < 0) {
9307 new_list = newSV_type(SVt_INVLIST);
9308 initialize_invlist_guts(new_list, initial_size);
9314 Perl__new_invlist_C_array(pTHX_ const UV* const list)
9316 /* Return a pointer to a newly constructed inversion list, initialized to
9317 * point to <list>, which has to be in the exact correct inversion list
9318 * form, including internal fields. Thus this is a dangerous routine that
9319 * should not be used in the wrong hands. The passed in 'list' contains
9320 * several header fields at the beginning that are not part of the
9321 * inversion list body proper */
9323 const STRLEN length = (STRLEN) list[0];
9324 const UV version_id = list[1];
9325 const bool offset = cBOOL(list[2]);
9326 #define HEADER_LENGTH 3
9327 /* If any of the above changes in any way, you must change HEADER_LENGTH
9328 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
9329 * perl -E 'say int(rand 2**31-1)'
9331 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
9332 data structure type, so that one being
9333 passed in can be validated to be an
9334 inversion list of the correct vintage.
9337 SV* invlist = newSV_type(SVt_INVLIST);
9339 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
9341 if (version_id != INVLIST_VERSION_ID) {
9342 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
9345 /* The generated array passed in includes header elements that aren't part
9346 * of the list proper, so start it just after them */
9347 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
9349 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
9350 shouldn't touch it */
9352 *(get_invlist_offset_addr(invlist)) = offset;
9354 /* The 'length' passed to us is the physical number of elements in the
9355 * inversion list. But if there is an offset the logical number is one
9357 invlist_set_len(invlist, length - offset, offset);
9359 invlist_set_previous_index(invlist, 0);
9361 /* Initialize the iteration pointer. */
9362 invlist_iterfinish(invlist);
9364 SvREADONLY_on(invlist);
9371 S__append_range_to_invlist(pTHX_ SV* const invlist,
9372 const UV start, const UV end)
9374 /* Subject to change or removal. Append the range from 'start' to 'end' at
9375 * the end of the inversion list. The range must be above any existing
9379 UV max = invlist_max(invlist);
9380 UV len = _invlist_len(invlist);
9383 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
9385 if (len == 0) { /* Empty lists must be initialized */
9386 offset = start != 0;
9387 array = _invlist_array_init(invlist, ! offset);
9390 /* Here, the existing list is non-empty. The current max entry in the
9391 * list is generally the first value not in the set, except when the
9392 * set extends to the end of permissible values, in which case it is
9393 * the first entry in that final set, and so this call is an attempt to
9394 * append out-of-order */
9396 UV final_element = len - 1;
9397 array = invlist_array(invlist);
9398 if ( array[final_element] > start
9399 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
9401 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",
9402 array[final_element], start,
9403 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
9406 /* Here, it is a legal append. If the new range begins 1 above the end
9407 * of the range below it, it is extending the range below it, so the
9408 * new first value not in the set is one greater than the newly
9409 * extended range. */
9410 offset = *get_invlist_offset_addr(invlist);
9411 if (array[final_element] == start) {
9412 if (end != UV_MAX) {
9413 array[final_element] = end + 1;
9416 /* But if the end is the maximum representable on the machine,
9417 * assume that infinity was actually what was meant. Just let
9418 * the range that this would extend to have no end */
9419 invlist_set_len(invlist, len - 1, offset);
9425 /* Here the new range doesn't extend any existing set. Add it */
9427 len += 2; /* Includes an element each for the start and end of range */
9429 /* If wll overflow the existing space, extend, which may cause the array to
9432 invlist_extend(invlist, len);
9434 /* Have to set len here to avoid assert failure in invlist_array() */
9435 invlist_set_len(invlist, len, offset);
9437 array = invlist_array(invlist);
9440 invlist_set_len(invlist, len, offset);
9443 /* The next item on the list starts the range, the one after that is
9444 * one past the new range. */
9445 array[len - 2] = start;
9446 if (end != UV_MAX) {
9447 array[len - 1] = end + 1;
9450 /* But if the end is the maximum representable on the machine, just let
9451 * the range have no end */
9452 invlist_set_len(invlist, len - 1, offset);
9457 Perl__invlist_search(SV* const invlist, const UV cp)
9459 /* Searches the inversion list for the entry that contains the input code
9460 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
9461 * return value is the index into the list's array of the range that
9462 * contains <cp>, that is, 'i' such that
9463 * array[i] <= cp < array[i+1]
9468 IV high = _invlist_len(invlist);
9469 const IV highest_element = high - 1;
9472 PERL_ARGS_ASSERT__INVLIST_SEARCH;
9474 /* If list is empty, return failure. */
9479 /* (We can't get the array unless we know the list is non-empty) */
9480 array = invlist_array(invlist);
9482 mid = invlist_previous_index(invlist);
9484 if (mid > highest_element) {
9485 mid = highest_element;
9488 /* <mid> contains the cache of the result of the previous call to this
9489 * function (0 the first time). See if this call is for the same result,
9490 * or if it is for mid-1. This is under the theory that calls to this
9491 * function will often be for related code points that are near each other.
9492 * And benchmarks show that caching gives better results. We also test
9493 * here if the code point is within the bounds of the list. These tests
9494 * replace others that would have had to be made anyway to make sure that
9495 * the array bounds were not exceeded, and these give us extra information
9496 * at the same time */
9497 if (cp >= array[mid]) {
9498 if (cp >= array[highest_element]) {
9499 return highest_element;
9502 /* Here, array[mid] <= cp < array[highest_element]. This means that
9503 * the final element is not the answer, so can exclude it; it also
9504 * means that <mid> is not the final element, so can refer to 'mid + 1'
9506 if (cp < array[mid + 1]) {
9512 else { /* cp < aray[mid] */
9513 if (cp < array[0]) { /* Fail if outside the array */
9517 if (cp >= array[mid - 1]) {
9522 /* Binary search. What we are looking for is <i> such that
9523 * array[i] <= cp < array[i+1]
9524 * The loop below converges on the i+1. Note that there may not be an
9525 * (i+1)th element in the array, and things work nonetheless */
9526 while (low < high) {
9527 mid = (low + high) / 2;
9528 assert(mid <= highest_element);
9529 if (array[mid] <= cp) { /* cp >= array[mid] */
9532 /* We could do this extra test to exit the loop early.
9533 if (cp < array[low]) {
9538 else { /* cp < array[mid] */
9545 invlist_set_previous_index(invlist, high);
9550 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9551 const bool complement_b, SV** output)
9553 /* Take the union of two inversion lists and point '*output' to it. On
9554 * input, '*output' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9555 * even 'a' or 'b'). If to an inversion list, the contents of the original
9556 * list will be replaced by the union. The first list, 'a', may be
9557 * NULL, in which case a copy of the second list is placed in '*output'.
9558 * If 'complement_b' is TRUE, the union is taken of the complement
9559 * (inversion) of 'b' instead of b itself.
9561 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9562 * Richard Gillam, published by Addison-Wesley, and explained at some
9563 * length there. The preface says to incorporate its examples into your
9564 * code at your own risk.
9566 * The algorithm is like a merge sort. */
9568 const UV* array_a; /* a's array */
9570 UV len_a; /* length of a's array */
9573 SV* u; /* the resulting union */
9577 UV i_a = 0; /* current index into a's array */
9581 /* running count, as explained in the algorithm source book; items are
9582 * stopped accumulating and are output when the count changes to/from 0.
9583 * The count is incremented when we start a range that's in an input's set,
9584 * and decremented when we start a range that's not in a set. So this
9585 * variable can be 0, 1, or 2. When it is 0 neither input is in their set,
9586 * and hence nothing goes into the union; 1, just one of the inputs is in
9587 * its set (and its current range gets added to the union); and 2 when both
9588 * inputs are in their sets. */
9591 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
9593 assert(*output == NULL || is_invlist(*output));
9595 len_b = _invlist_len(b);
9598 /* Here, 'b' is empty, hence it's complement is all possible code
9599 * points. So if the union includes the complement of 'b', it includes
9600 * everything, and we need not even look at 'a'. It's easiest to
9601 * create a new inversion list that matches everything. */
9603 SV* everything = _add_range_to_invlist(NULL, 0, UV_MAX);
9605 if (*output == NULL) { /* If the output didn't exist, just point it
9607 *output = everything;
9609 else { /* Otherwise, replace its contents with the new list */
9610 invlist_replace_list_destroys_src(*output, everything);
9611 SvREFCNT_dec_NN(everything);
9617 /* Here, we don't want the complement of 'b', and since 'b' is empty,
9618 * the union will come entirely from 'a'. If 'a' is NULL or empty, the
9619 * output will be empty */
9621 if (a == NULL || _invlist_len(a) == 0) {
9622 if (*output == NULL) {
9623 *output = _new_invlist(0);
9626 invlist_clear(*output);
9631 /* Here, 'a' is not empty, but 'b' is, so 'a' entirely determines the
9632 * union. We can just return a copy of 'a' if '*output' doesn't point
9633 * to an existing list */
9634 if (*output == NULL) {
9635 *output = invlist_clone(a, NULL);
9639 /* If the output is to overwrite 'a', we have a no-op, as it's
9645 /* Here, '*output' is to be overwritten by 'a' */
9646 u = invlist_clone(a, NULL);
9647 invlist_replace_list_destroys_src(*output, u);
9653 /* Here 'b' is not empty. See about 'a' */
9655 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
9657 /* Here, 'a' is empty (and b is not). That means the union will come
9658 * entirely from 'b'. If '*output' is NULL, we can directly return a
9659 * clone of 'b'. Otherwise, we replace the contents of '*output' with
9662 SV ** dest = (*output == NULL) ? output : &u;
9663 *dest = invlist_clone(b, NULL);
9665 _invlist_invert(*dest);
9669 invlist_replace_list_destroys_src(*output, u);
9676 /* Here both lists exist and are non-empty */
9677 array_a = invlist_array(a);
9678 array_b = invlist_array(b);
9680 /* If are to take the union of 'a' with the complement of b, set it
9681 * up so are looking at b's complement. */
9684 /* To complement, we invert: if the first element is 0, remove it. To
9685 * do this, we just pretend the array starts one later */
9686 if (array_b[0] == 0) {
9692 /* But if the first element is not zero, we pretend the list starts
9693 * at the 0 that is always stored immediately before the array. */
9699 /* Size the union for the worst case: that the sets are completely
9701 u = _new_invlist(len_a + len_b);
9703 /* Will contain U+0000 if either component does */
9704 array_u = _invlist_array_init(u, ( len_a > 0 && array_a[0] == 0)
9705 || (len_b > 0 && array_b[0] == 0));
9707 /* Go through each input list item by item, stopping when have exhausted
9709 while (i_a < len_a && i_b < len_b) {
9710 UV cp; /* The element to potentially add to the union's array */
9711 bool cp_in_set; /* is it in the the input list's set or not */
9713 /* We need to take one or the other of the two inputs for the union.
9714 * Since we are merging two sorted lists, we take the smaller of the
9715 * next items. In case of a tie, we take first the one that is in its
9716 * set. If we first took the one not in its set, it would decrement
9717 * the count, possibly to 0 which would cause it to be output as ending
9718 * the range, and the next time through we would take the same number,
9719 * and output it again as beginning the next range. By doing it the
9720 * opposite way, there is no possibility that the count will be
9721 * momentarily decremented to 0, and thus the two adjoining ranges will
9722 * be seamlessly merged. (In a tie and both are in the set or both not
9723 * in the set, it doesn't matter which we take first.) */
9724 if ( array_a[i_a] < array_b[i_b]
9725 || ( array_a[i_a] == array_b[i_b]
9726 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9728 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9729 cp = array_a[i_a++];
9732 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9733 cp = array_b[i_b++];
9736 /* Here, have chosen which of the two inputs to look at. Only output
9737 * if the running count changes to/from 0, which marks the
9738 * beginning/end of a range that's in the set */
9741 array_u[i_u++] = cp;
9748 array_u[i_u++] = cp;
9754 /* The loop above increments the index into exactly one of the input lists
9755 * each iteration, and ends when either index gets to its list end. That
9756 * means the other index is lower than its end, and so something is
9757 * remaining in that one. We decrement 'count', as explained below, if
9758 * that list is in its set. (i_a and i_b each currently index the element
9759 * beyond the one we care about.) */
9760 if ( (i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9761 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9766 /* Above we decremented 'count' if the list that had unexamined elements in
9767 * it was in its set. This has made it so that 'count' being non-zero
9768 * means there isn't anything left to output; and 'count' equal to 0 means
9769 * that what is left to output is precisely that which is left in the
9770 * non-exhausted input list.
9772 * To see why, note first that the exhausted input obviously has nothing
9773 * left to add to the union. If it was in its set at its end, that means
9774 * the set extends from here to the platform's infinity, and hence so does
9775 * the union and the non-exhausted set is irrelevant. The exhausted set
9776 * also contributed 1 to 'count'. If 'count' was 2, it got decremented to
9777 * 1, but if it was 1, the non-exhausted set wasn't in its set, and so
9778 * 'count' remains at 1. This is consistent with the decremented 'count'
9779 * != 0 meaning there's nothing left to add to the union.
9781 * But if the exhausted input wasn't in its set, it contributed 0 to
9782 * 'count', and the rest of the union will be whatever the other input is.
9783 * If 'count' was 0, neither list was in its set, and 'count' remains 0;
9784 * otherwise it gets decremented to 0. This is consistent with 'count'
9785 * == 0 meaning the remainder of the union is whatever is left in the
9786 * non-exhausted list. */
9791 IV copy_count = len_a - i_a;
9792 if (copy_count > 0) { /* The non-exhausted input is 'a' */
9793 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
9795 else { /* The non-exhausted input is b */
9796 copy_count = len_b - i_b;
9797 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
9799 len_u = i_u + copy_count;
9802 /* Set the result to the final length, which can change the pointer to
9803 * array_u, so re-find it. (Note that it is unlikely that this will
9804 * change, as we are shrinking the space, not enlarging it) */
9805 if (len_u != _invlist_len(u)) {
9806 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
9808 array_u = invlist_array(u);
9811 if (*output == NULL) { /* Simply return the new inversion list */
9815 /* Otherwise, overwrite the inversion list that was in '*output'. We
9816 * could instead free '*output', and then set it to 'u', but experience
9817 * has shown [perl #127392] that if the input is a mortal, we can get a
9818 * huge build-up of these during regex compilation before they get
9820 invlist_replace_list_destroys_src(*output, u);
9828 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9829 const bool complement_b, SV** i)
9831 /* Take the intersection of two inversion lists and point '*i' to it. On
9832 * input, '*i' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9833 * even 'a' or 'b'). If to an inversion list, the contents of the original
9834 * list will be replaced by the intersection. The first list, 'a', may be
9835 * NULL, in which case '*i' will be an empty list. If 'complement_b' is
9836 * TRUE, the result will be the intersection of 'a' and the complement (or
9837 * inversion) of 'b' instead of 'b' directly.
9839 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9840 * Richard Gillam, published by Addison-Wesley, and explained at some
9841 * length there. The preface says to incorporate its examples into your
9842 * code at your own risk. In fact, it had bugs
9844 * The algorithm is like a merge sort, and is essentially the same as the
9848 const UV* array_a; /* a's array */
9850 UV len_a; /* length of a's array */
9853 SV* r; /* the resulting intersection */
9857 UV i_a = 0; /* current index into a's array */
9861 /* running count of how many of the two inputs are postitioned at ranges
9862 * that are in their sets. As explained in the algorithm source book,
9863 * items are stopped accumulating and are output when the count changes
9864 * to/from 2. The count is incremented when we start a range that's in an
9865 * input's set, and decremented when we start a range that's not in a set.
9866 * Only when it is 2 are we in the intersection. */
9869 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
9871 assert(*i == NULL || is_invlist(*i));
9873 /* Special case if either one is empty */
9874 len_a = (a == NULL) ? 0 : _invlist_len(a);
9875 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
9876 if (len_a != 0 && complement_b) {
9878 /* Here, 'a' is not empty, therefore from the enclosing 'if', 'b'
9879 * must be empty. Here, also we are using 'b's complement, which
9880 * hence must be every possible code point. Thus the intersection
9883 if (*i == a) { /* No-op */
9888 *i = invlist_clone(a, NULL);
9892 r = invlist_clone(a, NULL);
9893 invlist_replace_list_destroys_src(*i, r);
9898 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
9899 * intersection must be empty */
9901 *i = _new_invlist(0);
9909 /* Here both lists exist and are non-empty */
9910 array_a = invlist_array(a);
9911 array_b = invlist_array(b);
9913 /* If are to take the intersection of 'a' with the complement of b, set it
9914 * up so are looking at b's complement. */
9917 /* To complement, we invert: if the first element is 0, remove it. To
9918 * do this, we just pretend the array starts one later */
9919 if (array_b[0] == 0) {
9925 /* But if the first element is not zero, we pretend the list starts
9926 * at the 0 that is always stored immediately before the array. */
9932 /* Size the intersection for the worst case: that the intersection ends up
9933 * fragmenting everything to be completely disjoint */
9934 r= _new_invlist(len_a + len_b);
9936 /* Will contain U+0000 iff both components do */
9937 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
9938 && len_b > 0 && array_b[0] == 0);
9940 /* Go through each list item by item, stopping when have exhausted one of
9942 while (i_a < len_a && i_b < len_b) {
9943 UV cp; /* The element to potentially add to the intersection's
9945 bool cp_in_set; /* Is it in the input list's set or not */
9947 /* We need to take one or the other of the two inputs for the
9948 * intersection. Since we are merging two sorted lists, we take the
9949 * smaller of the next items. In case of a tie, we take first the one
9950 * that is not in its set (a difference from the union algorithm). If
9951 * we first took the one in its set, it would increment the count,
9952 * possibly to 2 which would cause it to be output as starting a range
9953 * in the intersection, and the next time through we would take that
9954 * same number, and output it again as ending the set. By doing the
9955 * opposite of this, there is no possibility that the count will be
9956 * momentarily incremented to 2. (In a tie and both are in the set or
9957 * both not in the set, it doesn't matter which we take first.) */
9958 if ( array_a[i_a] < array_b[i_b]
9959 || ( array_a[i_a] == array_b[i_b]
9960 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9962 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9963 cp = array_a[i_a++];
9966 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9970 /* Here, have chosen which of the two inputs to look at. Only output
9971 * if the running count changes to/from 2, which marks the
9972 * beginning/end of a range that's in the intersection */
9976 array_r[i_r++] = cp;
9981 array_r[i_r++] = cp;
9988 /* The loop above increments the index into exactly one of the input lists
9989 * each iteration, and ends when either index gets to its list end. That
9990 * means the other index is lower than its end, and so something is
9991 * remaining in that one. We increment 'count', as explained below, if the
9992 * exhausted list was in its set. (i_a and i_b each currently index the
9993 * element beyond the one we care about.) */
9994 if ( (i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9995 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
10000 /* Above we incremented 'count' if the exhausted list was in its set. This
10001 * has made it so that 'count' being below 2 means there is nothing left to
10002 * output; otheriwse what's left to add to the intersection is precisely
10003 * that which is left in the non-exhausted input list.
10005 * To see why, note first that the exhausted input obviously has nothing
10006 * left to affect the intersection. If it was in its set at its end, that
10007 * means the set extends from here to the platform's infinity, and hence
10008 * anything in the non-exhausted's list will be in the intersection, and
10009 * anything not in it won't be. Hence, the rest of the intersection is
10010 * precisely what's in the non-exhausted list The exhausted set also
10011 * contributed 1 to 'count', meaning 'count' was at least 1. Incrementing
10012 * it means 'count' is now at least 2. This is consistent with the
10013 * incremented 'count' being >= 2 means to add the non-exhausted list to
10014 * the intersection.
10016 * But if the exhausted input wasn't in its set, it contributed 0 to
10017 * 'count', and the intersection can't include anything further; the
10018 * non-exhausted set is irrelevant. 'count' was at most 1, and doesn't get
10019 * incremented. This is consistent with 'count' being < 2 meaning nothing
10020 * further to add to the intersection. */
10021 if (count < 2) { /* Nothing left to put in the intersection. */
10024 else { /* copy the non-exhausted list, unchanged. */
10025 IV copy_count = len_a - i_a;
10026 if (copy_count > 0) { /* a is the one with stuff left */
10027 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
10029 else { /* b is the one with stuff left */
10030 copy_count = len_b - i_b;
10031 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
10033 len_r = i_r + copy_count;
10036 /* Set the result to the final length, which can change the pointer to
10037 * array_r, so re-find it. (Note that it is unlikely that this will
10038 * change, as we are shrinking the space, not enlarging it) */
10039 if (len_r != _invlist_len(r)) {
10040 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
10042 array_r = invlist_array(r);
10045 if (*i == NULL) { /* Simply return the calculated intersection */
10048 else { /* Otherwise, replace the existing inversion list in '*i'. We could
10049 instead free '*i', and then set it to 'r', but experience has
10050 shown [perl #127392] that if the input is a mortal, we can get a
10051 huge build-up of these during regex compilation before they get
10054 invlist_replace_list_destroys_src(*i, r);
10059 SvREFCNT_dec_NN(r);
10066 Perl__add_range_to_invlist(pTHX_ SV* invlist, UV start, UV end)
10068 /* Add the range from 'start' to 'end' inclusive to the inversion list's
10069 * set. A pointer to the inversion list is returned. This may actually be
10070 * a new list, in which case the passed in one has been destroyed. The
10071 * passed-in inversion list can be NULL, in which case a new one is created
10072 * with just the one range in it. The new list is not necessarily
10073 * NUL-terminated. Space is not freed if the inversion list shrinks as a
10074 * result of this function. The gain would not be large, and in many
10075 * cases, this is called multiple times on a single inversion list, so
10076 * anything freed may almost immediately be needed again.
10078 * This used to mostly call the 'union' routine, but that is much more
10079 * heavyweight than really needed for a single range addition */
10081 UV* array; /* The array implementing the inversion list */
10082 UV len; /* How many elements in 'array' */
10083 SSize_t i_s; /* index into the invlist array where 'start'
10085 SSize_t i_e = 0; /* And the index where 'end' should go */
10086 UV cur_highest; /* The highest code point in the inversion list
10087 upon entry to this function */
10089 /* This range becomes the whole inversion list if none already existed */
10090 if (invlist == NULL) {
10091 invlist = _new_invlist(2);
10092 _append_range_to_invlist(invlist, start, end);
10096 /* Likewise, if the inversion list is currently empty */
10097 len = _invlist_len(invlist);
10099 _append_range_to_invlist(invlist, start, end);
10103 /* Starting here, we have to know the internals of the list */
10104 array = invlist_array(invlist);
10106 /* If the new range ends higher than the current highest ... */
10107 cur_highest = invlist_highest(invlist);
10108 if (end > cur_highest) {
10110 /* If the whole range is higher, we can just append it */
10111 if (start > cur_highest) {
10112 _append_range_to_invlist(invlist, start, end);
10116 /* Otherwise, add the portion that is higher ... */
10117 _append_range_to_invlist(invlist, cur_highest + 1, end);
10119 /* ... and continue on below to handle the rest. As a result of the
10120 * above append, we know that the index of the end of the range is the
10121 * final even numbered one of the array. Recall that the final element
10122 * always starts a range that extends to infinity. If that range is in
10123 * the set (meaning the set goes from here to infinity), it will be an
10124 * even index, but if it isn't in the set, it's odd, and the final
10125 * range in the set is one less, which is even. */
10126 if (end == UV_MAX) {
10134 /* We have dealt with appending, now see about prepending. If the new
10135 * range starts lower than the current lowest ... */
10136 if (start < array[0]) {
10138 /* Adding something which has 0 in it is somewhat tricky, and uncommon.
10139 * Let the union code handle it, rather than having to know the
10140 * trickiness in two code places. */
10141 if (UNLIKELY(start == 0)) {
10144 range_invlist = _new_invlist(2);
10145 _append_range_to_invlist(range_invlist, start, end);
10147 _invlist_union(invlist, range_invlist, &invlist);
10149 SvREFCNT_dec_NN(range_invlist);
10154 /* If the whole new range comes before the first entry, and doesn't
10155 * extend it, we have to insert it as an additional range */
10156 if (end < array[0] - 1) {
10158 goto splice_in_new_range;
10161 /* Here the new range adjoins the existing first range, extending it
10165 /* And continue on below to handle the rest. We know that the index of
10166 * the beginning of the range is the first one of the array */
10169 else { /* Not prepending any part of the new range to the existing list.
10170 * Find where in the list it should go. This finds i_s, such that:
10171 * invlist[i_s] <= start < array[i_s+1]
10173 i_s = _invlist_search(invlist, start);
10176 /* At this point, any extending before the beginning of the inversion list
10177 * and/or after the end has been done. This has made it so that, in the
10178 * code below, each endpoint of the new range is either in a range that is
10179 * in the set, or is in a gap between two ranges that are. This means we
10180 * don't have to worry about exceeding the array bounds.
10182 * Find where in the list the new range ends (but we can skip this if we
10183 * have already determined what it is, or if it will be the same as i_s,
10184 * which we already have computed) */
10186 i_e = (start == end)
10188 : _invlist_search(invlist, end);
10191 /* Here generally invlist[i_e] <= end < array[i_e+1]. But if invlist[i_e]
10192 * is a range that goes to infinity there is no element at invlist[i_e+1],
10193 * so only the first relation holds. */
10195 if ( ! ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
10197 /* Here, the ranges on either side of the beginning of the new range
10198 * are in the set, and this range starts in the gap between them.
10200 * The new range extends the range above it downwards if the new range
10201 * ends at or above that range's start */
10202 const bool extends_the_range_above = ( end == UV_MAX
10203 || end + 1 >= array[i_s+1]);
10205 /* The new range extends the range below it upwards if it begins just
10206 * after where that range ends */
10207 if (start == array[i_s]) {
10209 /* If the new range fills the entire gap between the other ranges,
10210 * they will get merged together. Other ranges may also get
10211 * merged, depending on how many of them the new range spans. In
10212 * the general case, we do the merge later, just once, after we
10213 * figure out how many to merge. But in the case where the new
10214 * range exactly spans just this one gap (possibly extending into
10215 * the one above), we do the merge here, and an early exit. This
10216 * is done here to avoid having to special case later. */
10217 if (i_e - i_s <= 1) {
10219 /* If i_e - i_s == 1, it means that the new range terminates
10220 * within the range above, and hence 'extends_the_range_above'
10221 * must be true. (If the range above it extends to infinity,
10222 * 'i_s+2' will be above the array's limit, but 'len-i_s-2'
10223 * will be 0, so no harm done.) */
10224 if (extends_the_range_above) {
10225 Move(array + i_s + 2, array + i_s, len - i_s - 2, UV);
10226 invlist_set_len(invlist,
10228 *(get_invlist_offset_addr(invlist)));
10232 /* Here, i_e must == i_s. We keep them in sync, as they apply
10233 * to the same range, and below we are about to decrement i_s
10238 /* Here, the new range is adjacent to the one below. (It may also
10239 * span beyond the range above, but that will get resolved later.)
10240 * Extend the range below to include this one. */
10241 array[i_s] = (end == UV_MAX) ? UV_MAX : end + 1;
10243 start = array[i_s];
10245 else if (extends_the_range_above) {
10247 /* Here the new range only extends the range above it, but not the
10248 * one below. It merges with the one above. Again, we keep i_e
10249 * and i_s in sync if they point to the same range */
10254 array[i_s] = start;
10258 /* Here, we've dealt with the new range start extending any adjoining
10261 * If the new range extends to infinity, it is now the final one,
10262 * regardless of what was there before */
10263 if (UNLIKELY(end == UV_MAX)) {
10264 invlist_set_len(invlist, i_s + 1, *(get_invlist_offset_addr(invlist)));
10268 /* If i_e started as == i_s, it has also been dealt with,
10269 * and been updated to the new i_s, which will fail the following if */
10270 if (! ELEMENT_RANGE_MATCHES_INVLIST(i_e)) {
10272 /* Here, the ranges on either side of the end of the new range are in
10273 * the set, and this range ends in the gap between them.
10275 * If this range is adjacent to (hence extends) the range above it, it
10276 * becomes part of that range; likewise if it extends the range below,
10277 * it becomes part of that range */
10278 if (end + 1 == array[i_e+1]) {
10280 array[i_e] = start;
10282 else if (start <= array[i_e]) {
10283 array[i_e] = end + 1;
10290 /* If the range fits entirely in an existing range (as possibly already
10291 * extended above), it doesn't add anything new */
10292 if (ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
10296 /* Here, no part of the range is in the list. Must add it. It will
10297 * occupy 2 more slots */
10298 splice_in_new_range:
10300 invlist_extend(invlist, len + 2);
10301 array = invlist_array(invlist);
10302 /* Move the rest of the array down two slots. Don't include any
10304 Move(array + i_e + 1, array + i_e + 3, len - i_e - 1, UV);
10306 /* Do the actual splice */
10307 array[i_e+1] = start;
10308 array[i_e+2] = end + 1;
10309 invlist_set_len(invlist, len + 2, *(get_invlist_offset_addr(invlist)));
10313 /* Here the new range crossed the boundaries of a pre-existing range. The
10314 * code above has adjusted things so that both ends are in ranges that are
10315 * in the set. This means everything in between must also be in the set.
10316 * Just squash things together */
10317 Move(array + i_e + 1, array + i_s + 1, len - i_e - 1, UV);
10318 invlist_set_len(invlist,
10320 *(get_invlist_offset_addr(invlist)));
10326 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
10327 UV** other_elements_ptr)
10329 /* Create and return an inversion list whose contents are to be populated
10330 * by the caller. The caller gives the number of elements (in 'size') and
10331 * the very first element ('element0'). This function will set
10332 * '*other_elements_ptr' to an array of UVs, where the remaining elements
10333 * are to be placed.
10335 * Obviously there is some trust involved that the caller will properly
10336 * fill in the other elements of the array.
10338 * (The first element needs to be passed in, as the underlying code does
10339 * things differently depending on whether it is zero or non-zero) */
10341 SV* invlist = _new_invlist(size);
10344 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
10346 invlist = add_cp_to_invlist(invlist, element0);
10347 offset = *get_invlist_offset_addr(invlist);
10349 invlist_set_len(invlist, size, offset);
10350 *other_elements_ptr = invlist_array(invlist) + 1;
10356 #ifndef PERL_IN_XSUB_RE
10358 Perl__invlist_invert(pTHX_ SV* const invlist)
10360 /* Complement the input inversion list. This adds a 0 if the list didn't
10361 * have a zero; removes it otherwise. As described above, the data
10362 * structure is set up so that this is very efficient */
10364 PERL_ARGS_ASSERT__INVLIST_INVERT;
10366 assert(! invlist_is_iterating(invlist));
10368 /* The inverse of matching nothing is matching everything */
10369 if (_invlist_len(invlist) == 0) {
10370 _append_range_to_invlist(invlist, 0, UV_MAX);
10374 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
10378 Perl_invlist_clone(pTHX_ SV* const invlist, SV* new_invlist)
10380 /* Return a new inversion list that is a copy of the input one, which is
10381 * unchanged. The new list will not be mortal even if the old one was. */
10383 const STRLEN nominal_length = _invlist_len(invlist);
10384 const STRLEN physical_length = SvCUR(invlist);
10385 const bool offset = *(get_invlist_offset_addr(invlist));
10387 PERL_ARGS_ASSERT_INVLIST_CLONE;
10389 if (new_invlist == NULL) {
10390 new_invlist = _new_invlist(nominal_length);
10393 sv_upgrade(new_invlist, SVt_INVLIST);
10394 initialize_invlist_guts(new_invlist, nominal_length);
10397 *(get_invlist_offset_addr(new_invlist)) = offset;
10398 invlist_set_len(new_invlist, nominal_length, offset);
10399 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
10401 return new_invlist;
10406 PERL_STATIC_INLINE UV
10407 S_invlist_lowest(SV* const invlist)
10409 /* Returns the lowest code point that matches an inversion list. This API
10410 * has an ambiguity, as it returns 0 under either the lowest is actually
10411 * 0, or if the list is empty. If this distinction matters to you, check
10412 * for emptiness before calling this function */
10414 UV len = _invlist_len(invlist);
10417 PERL_ARGS_ASSERT_INVLIST_LOWEST;
10423 array = invlist_array(invlist);
10429 S_invlist_contents(pTHX_ SV* const invlist, const bool traditional_style)
10431 /* Get the contents of an inversion list into a string SV so that they can
10432 * be printed out. If 'traditional_style' is TRUE, it uses the format
10433 * traditionally done for debug tracing; otherwise it uses a format
10434 * suitable for just copying to the output, with blanks between ranges and
10435 * a dash between range components */
10439 const char intra_range_delimiter = (traditional_style ? '\t' : '-');
10440 const char inter_range_delimiter = (traditional_style ? '\n' : ' ');
10442 if (traditional_style) {
10443 output = newSVpvs("\n");
10446 output = newSVpvs("");
10449 PERL_ARGS_ASSERT_INVLIST_CONTENTS;
10451 assert(! invlist_is_iterating(invlist));
10453 invlist_iterinit(invlist);
10454 while (invlist_iternext(invlist, &start, &end)) {
10455 if (end == UV_MAX) {
10456 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%cINFTY%c",
10457 start, intra_range_delimiter,
10458 inter_range_delimiter);
10460 else if (end != start) {
10461 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c%04" UVXf "%c",
10463 intra_range_delimiter,
10464 end, inter_range_delimiter);
10467 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c",
10468 start, inter_range_delimiter);
10472 if (SvCUR(output) && ! traditional_style) {/* Get rid of trailing blank */
10473 SvCUR_set(output, SvCUR(output) - 1);
10479 #ifndef PERL_IN_XSUB_RE
10481 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
10482 const char * const indent, SV* const invlist)
10484 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
10485 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
10486 * the string 'indent'. The output looks like this:
10487 [0] 0x000A .. 0x000D
10489 [4] 0x2028 .. 0x2029
10490 [6] 0x3104 .. INFTY
10491 * This means that the first range of code points matched by the list are
10492 * 0xA through 0xD; the second range contains only the single code point
10493 * 0x85, etc. An inversion list is an array of UVs. Two array elements
10494 * are used to define each range (except if the final range extends to
10495 * infinity, only a single element is needed). The array index of the
10496 * first element for the corresponding range is given in brackets. */
10501 PERL_ARGS_ASSERT__INVLIST_DUMP;
10503 if (invlist_is_iterating(invlist)) {
10504 Perl_dump_indent(aTHX_ level, file,
10505 "%sCan't dump inversion list because is in middle of iterating\n",
10510 invlist_iterinit(invlist);
10511 while (invlist_iternext(invlist, &start, &end)) {
10512 if (end == UV_MAX) {
10513 Perl_dump_indent(aTHX_ level, file,
10514 "%s[%" UVuf "] 0x%04" UVXf " .. INFTY\n",
10515 indent, (UV)count, start);
10517 else if (end != start) {
10518 Perl_dump_indent(aTHX_ level, file,
10519 "%s[%" UVuf "] 0x%04" UVXf " .. 0x%04" UVXf "\n",
10520 indent, (UV)count, start, end);
10523 Perl_dump_indent(aTHX_ level, file, "%s[%" UVuf "] 0x%04" UVXf "\n",
10524 indent, (UV)count, start);
10532 #if defined(PERL_ARGS_ASSERT__INVLISTEQ) && !defined(PERL_IN_XSUB_RE)
10534 Perl__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
10536 /* Return a boolean as to if the two passed in inversion lists are
10537 * identical. The final argument, if TRUE, says to take the complement of
10538 * the second inversion list before doing the comparison */
10540 const UV len_a = _invlist_len(a);
10541 UV len_b = _invlist_len(b);
10543 const UV* array_a = NULL;
10544 const UV* array_b = NULL;
10546 PERL_ARGS_ASSERT__INVLISTEQ;
10548 /* This code avoids accessing the arrays unless it knows the length is
10553 return ! complement_b;
10557 array_a = invlist_array(a);
10561 array_b = invlist_array(b);
10564 /* If are to compare 'a' with the complement of b, set it
10565 * up so are looking at b's complement. */
10566 if (complement_b) {
10568 /* The complement of nothing is everything, so <a> would have to have
10569 * just one element, starting at zero (ending at infinity) */
10571 return (len_a == 1 && array_a[0] == 0);
10573 if (array_b[0] == 0) {
10575 /* Otherwise, to complement, we invert. Here, the first element is
10576 * 0, just remove it. To do this, we just pretend the array starts
10584 /* But if the first element is not zero, we pretend the list starts
10585 * at the 0 that is always stored immediately before the array. */
10591 return len_a == len_b
10592 && memEQ(array_a, array_b, len_a * sizeof(array_a[0]));
10598 * As best we can, determine the characters that can match the start of
10599 * the given EXACTF-ish node. This is for use in creating ssc nodes, so there
10600 * can be false positive matches
10602 * Returns the invlist as a new SV*; it is the caller's responsibility to
10603 * call SvREFCNT_dec() when done with it.
10606 S_make_exactf_invlist(pTHX_ RExC_state_t *pRExC_state, regnode *node)
10609 const U8 * s = (U8*)STRING(node);
10610 SSize_t bytelen = STR_LEN(node);
10612 /* Start out big enough for 2 separate code points */
10613 SV* invlist = _new_invlist(4);
10615 PERL_ARGS_ASSERT_MAKE_EXACTF_INVLIST;
10620 /* We punt and assume can match anything if the node begins
10621 * with a multi-character fold. Things are complicated. For
10622 * example, /ffi/i could match any of:
10623 * "\N{LATIN SMALL LIGATURE FFI}"
10624 * "\N{LATIN SMALL LIGATURE FF}I"
10625 * "F\N{LATIN SMALL LIGATURE FI}"
10626 * plus several other things; and making sure we have all the
10627 * possibilities is hard. */
10628 if (is_MULTI_CHAR_FOLD_latin1_safe(s, s + bytelen)) {
10629 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10632 /* Any Latin1 range character can potentially match any
10633 * other depending on the locale, and in Turkic locales, U+130 and
10635 if (OP(node) == EXACTFL) {
10636 _invlist_union(invlist, PL_Latin1, &invlist);
10637 invlist = add_cp_to_invlist(invlist,
10638 LATIN_SMALL_LETTER_DOTLESS_I);
10639 invlist = add_cp_to_invlist(invlist,
10640 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
10643 /* But otherwise, it matches at least itself. We can
10644 * quickly tell if it has a distinct fold, and if so,
10645 * it matches that as well */
10646 invlist = add_cp_to_invlist(invlist, uc);
10647 if (IS_IN_SOME_FOLD_L1(uc))
10648 invlist = add_cp_to_invlist(invlist, PL_fold_latin1[uc]);
10651 /* Some characters match above-Latin1 ones under /i. This
10652 * is true of EXACTFL ones when the locale is UTF-8 */
10653 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(uc)
10654 && (! isASCII(uc) || (OP(node) != EXACTFAA
10655 && OP(node) != EXACTFAA_NO_TRIE)))
10657 add_above_Latin1_folds(pRExC_state, (U8) uc, &invlist);
10661 else { /* Pattern is UTF-8 */
10662 U8 folded[UTF8_MAX_FOLD_CHAR_EXPAND * UTF8_MAXBYTES_CASE + 1] = { '\0' };
10663 const U8* e = s + bytelen;
10666 fc = uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
10668 /* The only code points that aren't folded in a UTF EXACTFish
10669 * node are are the problematic ones in EXACTFL nodes */
10670 if (OP(node) == EXACTFL && is_PROBLEMATIC_LOCALE_FOLDEDS_START_cp(uc)) {
10671 /* We need to check for the possibility that this EXACTFL
10672 * node begins with a multi-char fold. Therefore we fold
10673 * the first few characters of it so that we can make that
10679 for (i = 0; i < UTF8_MAX_FOLD_CHAR_EXPAND && s < e; i++) {
10681 *(d++) = (U8) toFOLD(*s);
10682 if (fc < 0) { /* Save the first fold */
10689 UV fold = toFOLD_utf8_safe(s, e, d, &len);
10690 if (fc < 0) { /* Save the first fold */
10698 /* And set up so the code below that looks in this folded
10699 * buffer instead of the node's string */
10704 /* When we reach here 's' points to the fold of the first
10705 * character(s) of the node; and 'e' points to far enough along
10706 * the folded string to be just past any possible multi-char
10709 * Unlike the non-UTF-8 case, the macro for determining if a
10710 * string is a multi-char fold requires all the characters to
10711 * already be folded. This is because of all the complications
10712 * if not. Note that they are folded anyway, except in EXACTFL
10713 * nodes. Like the non-UTF case above, we punt if the node
10714 * begins with a multi-char fold */
10716 if (is_MULTI_CHAR_FOLD_utf8_safe(s, e)) {
10717 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10719 else { /* Single char fold */
10722 const U32 * remaining_folds;
10723 Size_t folds_count;
10725 /* It matches itself */
10726 invlist = add_cp_to_invlist(invlist, fc);
10728 /* ... plus all the things that fold to it, which are found in
10729 * PL_utf8_foldclosures */
10730 folds_count = _inverse_folds(fc, &first_fold,
10732 for (k = 0; k < folds_count; k++) {
10733 UV c = (k == 0) ? first_fold : remaining_folds[k-1];
10735 /* /aa doesn't allow folds between ASCII and non- */
10736 if ( (OP(node) == EXACTFAA || OP(node) == EXACTFAA_NO_TRIE)
10737 && isASCII(c) != isASCII(fc))
10742 invlist = add_cp_to_invlist(invlist, c);
10745 if (OP(node) == EXACTFL) {
10747 /* If either [iI] are present in an EXACTFL node the above code
10748 * should have added its normal case pair, but under a Turkish
10749 * locale they could match instead the case pairs from it. Add
10750 * those as potential matches as well */
10751 if (isALPHA_FOLD_EQ(fc, 'I')) {
10752 invlist = add_cp_to_invlist(invlist,
10753 LATIN_SMALL_LETTER_DOTLESS_I);
10754 invlist = add_cp_to_invlist(invlist,
10755 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
10757 else if (fc == LATIN_SMALL_LETTER_DOTLESS_I) {
10758 invlist = add_cp_to_invlist(invlist, 'I');
10760 else if (fc == LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE) {
10761 invlist = add_cp_to_invlist(invlist, 'i');
10770 #undef HEADER_LENGTH
10771 #undef TO_INTERNAL_SIZE
10772 #undef FROM_INTERNAL_SIZE
10773 #undef INVLIST_VERSION_ID
10775 /* End of inversion list object */
10778 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
10780 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
10781 * constructs, and updates RExC_flags with them. On input, RExC_parse
10782 * should point to the first flag; it is updated on output to point to the
10783 * final ')' or ':'. There needs to be at least one flag, or this will
10786 /* for (?g), (?gc), and (?o) warnings; warning
10787 about (?c) will warn about (?g) -- japhy */
10789 #define WASTED_O 0x01
10790 #define WASTED_G 0x02
10791 #define WASTED_C 0x04
10792 #define WASTED_GC (WASTED_G|WASTED_C)
10793 I32 wastedflags = 0x00;
10794 U32 posflags = 0, negflags = 0;
10795 U32 *flagsp = &posflags;
10796 char has_charset_modifier = '\0';
10798 bool has_use_defaults = FALSE;
10799 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
10800 int x_mod_count = 0;
10802 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
10804 /* '^' as an initial flag sets certain defaults */
10805 if (UCHARAT(RExC_parse) == '^') {
10807 has_use_defaults = TRUE;
10808 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
10809 cs = (RExC_uni_semantics)
10810 ? REGEX_UNICODE_CHARSET
10811 : REGEX_DEPENDS_CHARSET;
10812 set_regex_charset(&RExC_flags, cs);
10815 cs = get_regex_charset(RExC_flags);
10816 if ( cs == REGEX_DEPENDS_CHARSET
10817 && RExC_uni_semantics)
10819 cs = REGEX_UNICODE_CHARSET;
10823 while (RExC_parse < RExC_end) {
10824 /* && memCHRs("iogcmsx", *RExC_parse) */
10825 /* (?g), (?gc) and (?o) are useless here
10826 and must be globally applied -- japhy */
10827 if ((RExC_pm_flags & PMf_WILDCARD)) {
10828 if (flagsp == & negflags) {
10829 if (*RExC_parse == 'm') {
10831 /* diag_listed_as: Use of %s is not allowed in Unicode
10832 property wildcard subpatterns in regex; marked by <--
10834 vFAIL("Use of modifier '-m' is not allowed in Unicode"
10835 " property wildcard subpatterns");
10839 if (*RExC_parse == 's') {
10840 goto modifier_illegal_in_wildcard;
10845 switch (*RExC_parse) {
10847 /* Code for the imsxn flags */
10848 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp, x_mod_count);
10850 case LOCALE_PAT_MOD:
10851 if (has_charset_modifier) {
10852 goto excess_modifier;
10854 else if (flagsp == &negflags) {
10857 cs = REGEX_LOCALE_CHARSET;
10858 has_charset_modifier = LOCALE_PAT_MOD;
10860 case UNICODE_PAT_MOD:
10861 if (has_charset_modifier) {
10862 goto excess_modifier;
10864 else if (flagsp == &negflags) {
10867 cs = REGEX_UNICODE_CHARSET;
10868 has_charset_modifier = UNICODE_PAT_MOD;
10870 case ASCII_RESTRICT_PAT_MOD:
10871 if (flagsp == &negflags) {
10874 if (has_charset_modifier) {
10875 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
10876 goto excess_modifier;
10878 /* Doubled modifier implies more restricted */
10879 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
10882 cs = REGEX_ASCII_RESTRICTED_CHARSET;
10884 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
10886 case DEPENDS_PAT_MOD:
10887 if (has_use_defaults) {
10888 goto fail_modifiers;
10890 else if (flagsp == &negflags) {
10893 else if (has_charset_modifier) {
10894 goto excess_modifier;
10897 /* The dual charset means unicode semantics if the
10898 * pattern (or target, not known until runtime) are
10899 * utf8, or something in the pattern indicates unicode
10901 cs = (RExC_uni_semantics)
10902 ? REGEX_UNICODE_CHARSET
10903 : REGEX_DEPENDS_CHARSET;
10904 has_charset_modifier = DEPENDS_PAT_MOD;
10908 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
10909 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
10911 else if (has_charset_modifier == *(RExC_parse - 1)) {
10912 vFAIL2("Regexp modifier \"%c\" may not appear twice",
10913 *(RExC_parse - 1));
10916 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
10918 NOT_REACHED; /*NOTREACHED*/
10921 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
10922 *(RExC_parse - 1));
10923 NOT_REACHED; /*NOTREACHED*/
10924 case GLOBAL_PAT_MOD: /* 'g' */
10925 if (RExC_pm_flags & PMf_WILDCARD) {
10926 goto modifier_illegal_in_wildcard;
10929 case ONCE_PAT_MOD: /* 'o' */
10930 if (ckWARN(WARN_REGEXP)) {
10931 const I32 wflagbit = *RExC_parse == 'o'
10934 if (! (wastedflags & wflagbit) ) {
10935 wastedflags |= wflagbit;
10936 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10939 "Useless (%s%c) - %suse /%c modifier",
10940 flagsp == &negflags ? "?-" : "?",
10942 flagsp == &negflags ? "don't " : "",
10949 case CONTINUE_PAT_MOD: /* 'c' */
10950 if (RExC_pm_flags & PMf_WILDCARD) {
10951 goto modifier_illegal_in_wildcard;
10953 if (ckWARN(WARN_REGEXP)) {
10954 if (! (wastedflags & WASTED_C) ) {
10955 wastedflags |= WASTED_GC;
10956 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10959 "Useless (%sc) - %suse /gc modifier",
10960 flagsp == &negflags ? "?-" : "?",
10961 flagsp == &negflags ? "don't " : ""
10966 case KEEPCOPY_PAT_MOD: /* 'p' */
10967 if (RExC_pm_flags & PMf_WILDCARD) {
10968 goto modifier_illegal_in_wildcard;
10970 if (flagsp == &negflags) {
10971 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
10973 *flagsp |= RXf_PMf_KEEPCOPY;
10977 /* A flag is a default iff it is following a minus, so
10978 * if there is a minus, it means will be trying to
10979 * re-specify a default which is an error */
10980 if (has_use_defaults || flagsp == &negflags) {
10981 goto fail_modifiers;
10983 flagsp = &negflags;
10984 wastedflags = 0; /* reset so (?g-c) warns twice */
10990 if ( (RExC_pm_flags & PMf_WILDCARD)
10991 && cs != REGEX_ASCII_MORE_RESTRICTED_CHARSET)
10994 /* diag_listed_as: Use of %s is not allowed in Unicode
10995 property wildcard subpatterns in regex; marked by <--
10997 vFAIL2("Use of modifier '%c' is not allowed in Unicode"
10998 " property wildcard subpatterns",
10999 has_charset_modifier);
11002 if ((posflags & (RXf_PMf_EXTENDED|RXf_PMf_EXTENDED_MORE)) == RXf_PMf_EXTENDED) {
11003 negflags |= RXf_PMf_EXTENDED_MORE;
11005 RExC_flags |= posflags;
11007 if (negflags & RXf_PMf_EXTENDED) {
11008 negflags |= RXf_PMf_EXTENDED_MORE;
11010 RExC_flags &= ~negflags;
11011 set_regex_charset(&RExC_flags, cs);
11016 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
11017 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11018 vFAIL2utf8f("Sequence (%" UTF8f "...) not recognized",
11019 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
11020 NOT_REACHED; /*NOTREACHED*/
11023 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11026 vFAIL("Sequence (?... not terminated");
11028 modifier_illegal_in_wildcard:
11030 /* diag_listed_as: Use of %s is not allowed in Unicode property wildcard
11031 subpatterns in regex; marked by <-- HERE in m/%s/ */
11032 vFAIL2("Use of modifier '%c' is not allowed in Unicode property wildcard"
11033 " subpatterns", *(RExC_parse - 1));
11037 - reg - regular expression, i.e. main body or parenthesized thing
11039 * Caller must absorb opening parenthesis.
11041 * Combining parenthesis handling with the base level of regular expression
11042 * is a trifle forced, but the need to tie the tails of the branches to what
11043 * follows makes it hard to avoid.
11045 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
11047 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
11049 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
11052 PERL_STATIC_INLINE regnode_offset
11053 S_handle_named_backref(pTHX_ RExC_state_t *pRExC_state,
11055 char * parse_start,
11059 regnode_offset ret;
11060 char* name_start = RExC_parse;
11062 SV *sv_dat = reg_scan_name(pRExC_state, REG_RSN_RETURN_DATA);
11063 DECLARE_AND_GET_RE_DEBUG_FLAGS;
11065 PERL_ARGS_ASSERT_HANDLE_NAMED_BACKREF;
11067 if (RExC_parse == name_start || *RExC_parse != ch) {
11068 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
11069 vFAIL2("Sequence %.3s... not terminated", parse_start);
11073 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11074 RExC_rxi->data->data[num]=(void*)sv_dat;
11075 SvREFCNT_inc_simple_void_NN(sv_dat);
11078 ret = reganode(pRExC_state,
11081 : (ASCII_FOLD_RESTRICTED)
11083 : (AT_LEAST_UNI_SEMANTICS)
11089 *flagp |= HASWIDTH;
11091 Set_Node_Offset(REGNODE_p(ret), parse_start+1);
11092 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
11094 nextchar(pRExC_state);
11098 /* On success, returns the offset at which any next node should be placed into
11099 * the regex engine program being compiled.
11101 * Returns 0 otherwise, with *flagp set to indicate why:
11102 * TRYAGAIN at the end of (?) that only sets flags.
11103 * RESTART_PARSE if the parse needs to be restarted, or'd with
11104 * NEED_UTF8 if the pattern needs to be upgraded to UTF-8.
11105 * Otherwise would only return 0 if regbranch() returns 0, which cannot
11107 STATIC regnode_offset
11108 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp, U32 depth)
11109 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
11110 * 2 is like 1, but indicates that nextchar() has been called to advance
11111 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
11112 * this flag alerts us to the need to check for that */
11114 regnode_offset ret = 0; /* Will be the head of the group. */
11116 regnode_offset lastbr;
11117 regnode_offset ender = 0;
11120 U32 oregflags = RExC_flags;
11121 bool have_branch = 0;
11123 I32 freeze_paren = 0;
11124 I32 after_freeze = 0;
11125 I32 num; /* numeric backreferences */
11126 SV * max_open; /* Max number of unclosed parens */
11128 char * parse_start = RExC_parse; /* MJD */
11129 char * const oregcomp_parse = RExC_parse;
11131 DECLARE_AND_GET_RE_DEBUG_FLAGS;
11133 PERL_ARGS_ASSERT_REG;
11134 DEBUG_PARSE("reg ");
11136 max_open = get_sv(RE_COMPILE_RECURSION_LIMIT, GV_ADD);
11138 if (!SvIOK(max_open)) {
11139 sv_setiv(max_open, RE_COMPILE_RECURSION_INIT);
11141 if (depth > 4 * (UV) SvIV(max_open)) { /* We increase depth by 4 for each
11143 vFAIL("Too many nested open parens");
11146 *flagp = 0; /* Tentatively. */
11148 if (RExC_in_lookbehind) {
11149 RExC_in_lookbehind++;
11151 if (RExC_in_lookahead) {
11152 RExC_in_lookahead++;
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;
11408 RExC_in_lookbehind++;
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 if (has_intervening_patws) {
11485 vFAIL("In '(?...)', the '(' and '?' must be adjacent");
11488 RExC_parse++; /* past the '?' */
11489 paren = *RExC_parse; /* might be a trailing NUL, if not
11491 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11492 if (RExC_parse > RExC_end) {
11495 ret = 0; /* For look-ahead/behind. */
11498 case 'P': /* (?P...) variants for those used to PCRE/Python */
11499 paren = *RExC_parse;
11500 if ( paren == '<') { /* (?P<...>) named capture */
11502 if (RExC_parse >= RExC_end) {
11503 vFAIL("Sequence (?P<... not terminated");
11505 goto named_capture;
11507 else if (paren == '>') { /* (?P>name) named recursion */
11509 if (RExC_parse >= RExC_end) {
11510 vFAIL("Sequence (?P>... not terminated");
11512 goto named_recursion;
11514 else if (paren == '=') { /* (?P=...) named backref */
11516 return handle_named_backref(pRExC_state, flagp,
11519 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
11520 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11521 vFAIL3("Sequence (%.*s...) not recognized",
11522 (int) (RExC_parse - seqstart), seqstart);
11523 NOT_REACHED; /*NOTREACHED*/
11524 case '<': /* (?<...) */
11525 /* If you want to support (?<*...), first reconcile with GH #17363 */
11526 if (*RExC_parse == '!')
11528 else if (*RExC_parse != '=')
11535 case '\'': /* (?'...') */
11536 name_start = RExC_parse;
11537 svname = reg_scan_name(pRExC_state, REG_RSN_RETURN_NAME);
11538 if ( RExC_parse == name_start
11539 || RExC_parse >= RExC_end
11540 || *RExC_parse != paren)
11542 vFAIL2("Sequence (?%c... not terminated",
11543 paren=='>' ? '<' : (char) paren);
11548 if (!svname) /* shouldn't happen */
11550 "panic: reg_scan_name returned NULL");
11551 if (!RExC_paren_names) {
11552 RExC_paren_names= newHV();
11553 sv_2mortal(MUTABLE_SV(RExC_paren_names));
11555 RExC_paren_name_list= newAV();
11556 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
11559 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
11561 sv_dat = HeVAL(he_str);
11563 /* croak baby croak */
11565 "panic: paren_name hash element allocation failed");
11566 } else if ( SvPOK(sv_dat) ) {
11567 /* (?|...) can mean we have dupes so scan to check
11568 its already been stored. Maybe a flag indicating
11569 we are inside such a construct would be useful,
11570 but the arrays are likely to be quite small, so
11571 for now we punt -- dmq */
11572 IV count = SvIV(sv_dat);
11573 I32 *pv = (I32*)SvPVX(sv_dat);
11575 for ( i = 0 ; i < count ; i++ ) {
11576 if ( pv[i] == RExC_npar ) {
11582 pv = (I32*)SvGROW(sv_dat,
11583 SvCUR(sv_dat) + sizeof(I32)+1);
11584 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
11585 pv[count] = RExC_npar;
11586 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
11589 (void)SvUPGRADE(sv_dat, SVt_PVNV);
11590 sv_setpvn(sv_dat, (char *)&(RExC_npar),
11593 SvIV_set(sv_dat, 1);
11596 /* Yes this does cause a memory leak in debugging Perls
11598 if (!av_store(RExC_paren_name_list,
11599 RExC_npar, SvREFCNT_inc_NN(svname)))
11600 SvREFCNT_dec_NN(svname);
11603 /*sv_dump(sv_dat);*/
11605 nextchar(pRExC_state);
11607 goto capturing_parens;
11610 RExC_seen |= REG_LOOKBEHIND_SEEN;
11611 RExC_in_lookbehind++;
11613 if (RExC_parse >= RExC_end) {
11614 vFAIL("Sequence (?... not terminated");
11616 RExC_seen_zerolen++;
11618 case '=': /* (?=...) */
11619 RExC_seen_zerolen++;
11620 RExC_in_lookahead++;
11622 case '!': /* (?!...) */
11623 RExC_seen_zerolen++;
11624 /* check if we're really just a "FAIL" assertion */
11625 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
11626 FALSE /* Don't force to /x */ );
11627 if (*RExC_parse == ')') {
11628 ret=reganode(pRExC_state, OPFAIL, 0);
11629 nextchar(pRExC_state);
11633 case '|': /* (?|...) */
11634 /* branch reset, behave like a (?:...) except that
11635 buffers in alternations share the same numbers */
11637 after_freeze = freeze_paren = RExC_npar;
11639 /* XXX This construct currently requires an extra pass.
11640 * Investigation would be required to see if that could be
11642 REQUIRE_PARENS_PASS;
11644 case ':': /* (?:...) */
11645 case '>': /* (?>...) */
11647 case '$': /* (?$...) */
11648 case '@': /* (?@...) */
11649 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
11651 case '0' : /* (?0) */
11652 case 'R' : /* (?R) */
11653 if (RExC_parse == RExC_end || *RExC_parse != ')')
11654 FAIL("Sequence (?R) not terminated");
11656 RExC_seen |= REG_RECURSE_SEEN;
11658 /* XXX These constructs currently require an extra pass.
11659 * It probably could be changed */
11660 REQUIRE_PARENS_PASS;
11662 *flagp |= POSTPONED;
11663 goto gen_recurse_regop;
11665 /* named and numeric backreferences */
11666 case '&': /* (?&NAME) */
11667 parse_start = RExC_parse - 1;
11670 SV *sv_dat = reg_scan_name(pRExC_state,
11671 REG_RSN_RETURN_DATA);
11672 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
11674 if (RExC_parse >= RExC_end || *RExC_parse != ')')
11675 vFAIL("Sequence (?&... not terminated");
11676 goto gen_recurse_regop;
11679 if (! inRANGE(RExC_parse[0], '1', '9')) {
11681 vFAIL("Illegal pattern");
11683 goto parse_recursion;
11685 case '-': /* (?-1) */
11686 if (! inRANGE(RExC_parse[0], '1', '9')) {
11687 RExC_parse--; /* rewind to let it be handled later */
11691 case '1': case '2': case '3': case '4': /* (?1) */
11692 case '5': case '6': case '7': case '8': case '9':
11693 RExC_parse = (char *) seqstart + 1; /* Point to the digit */
11696 bool is_neg = FALSE;
11698 parse_start = RExC_parse - 1; /* MJD */
11699 if (*RExC_parse == '-') {
11704 if (grok_atoUV(RExC_parse, &unum, &endptr)
11708 RExC_parse = (char*)endptr;
11712 /* Some limit for num? */
11716 if (*RExC_parse!=')')
11717 vFAIL("Expecting close bracket");
11720 if ( paren == '-' ) {
11722 Diagram of capture buffer numbering.
11723 Top line is the normal capture buffer numbers
11724 Bottom line is the negative indexing as from
11728 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
11732 num = RExC_npar + num;
11735 /* It might be a forward reference; we can't fail until
11736 * we know, by completing the parse to get all the
11737 * groups, and then reparsing */
11738 if (ALL_PARENS_COUNTED) {
11740 vFAIL("Reference to nonexistent group");
11743 REQUIRE_PARENS_PASS;
11746 } else if ( paren == '+' ) {
11747 num = RExC_npar + num - 1;
11749 /* We keep track how many GOSUB items we have produced.
11750 To start off the ARG2L() of the GOSUB holds its "id",
11751 which is used later in conjunction with RExC_recurse
11752 to calculate the offset we need to jump for the GOSUB,
11753 which it will store in the final representation.
11754 We have to defer the actual calculation until much later
11755 as the regop may move.
11758 ret = reg2Lanode(pRExC_state, GOSUB, num, RExC_recurse_count);
11759 if (num >= RExC_npar) {
11761 /* It might be a forward reference; we can't fail until we
11762 * know, by completing the parse to get all the groups, and
11763 * then reparsing */
11764 if (ALL_PARENS_COUNTED) {
11765 if (num >= RExC_total_parens) {
11767 vFAIL("Reference to nonexistent group");
11771 REQUIRE_PARENS_PASS;
11774 RExC_recurse_count++;
11775 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11776 "%*s%*s Recurse #%" UVuf " to %" IVdf "\n",
11777 22, "| |", (int)(depth * 2 + 1), "",
11778 (UV)ARG(REGNODE_p(ret)),
11779 (IV)ARG2L(REGNODE_p(ret))));
11780 RExC_seen |= REG_RECURSE_SEEN;
11782 Set_Node_Length(REGNODE_p(ret),
11783 1 + regarglen[OP(REGNODE_p(ret))]); /* MJD */
11784 Set_Node_Offset(REGNODE_p(ret), parse_start); /* MJD */
11786 *flagp |= POSTPONED;
11787 assert(*RExC_parse == ')');
11788 nextchar(pRExC_state);
11793 case '?': /* (??...) */
11795 if (*RExC_parse != '{') {
11796 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
11797 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11799 "Sequence (%" UTF8f "...) not recognized",
11800 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
11801 NOT_REACHED; /*NOTREACHED*/
11803 *flagp |= POSTPONED;
11807 case '{': /* (?{...}) */
11810 struct reg_code_block *cb;
11813 RExC_seen_zerolen++;
11815 if ( !pRExC_state->code_blocks
11816 || pRExC_state->code_index
11817 >= pRExC_state->code_blocks->count
11818 || pRExC_state->code_blocks->cb[pRExC_state->code_index].start
11819 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
11822 if (RExC_pm_flags & PMf_USE_RE_EVAL)
11823 FAIL("panic: Sequence (?{...}): no code block found\n");
11824 FAIL("Eval-group not allowed at runtime, use re 'eval'");
11826 /* this is a pre-compiled code block (?{...}) */
11827 cb = &pRExC_state->code_blocks->cb[pRExC_state->code_index];
11828 RExC_parse = RExC_start + cb->end;
11830 if (cb->src_regex) {
11831 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
11832 RExC_rxi->data->data[n] =
11833 (void*)SvREFCNT_inc((SV*)cb->src_regex);
11834 RExC_rxi->data->data[n+1] = (void*)o;
11837 n = add_data(pRExC_state,
11838 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
11839 RExC_rxi->data->data[n] = (void*)o;
11841 pRExC_state->code_index++;
11842 nextchar(pRExC_state);
11845 regnode_offset eval;
11846 ret = reg_node(pRExC_state, LOGICAL);
11848 eval = reg2Lanode(pRExC_state, EVAL,
11851 /* for later propagation into (??{})
11853 RExC_flags & RXf_PMf_COMPILETIME
11855 FLAGS(REGNODE_p(ret)) = 2;
11856 if (! REGTAIL(pRExC_state, ret, eval)) {
11857 REQUIRE_BRANCHJ(flagp, 0);
11859 /* deal with the length of this later - MJD */
11862 ret = reg2Lanode(pRExC_state, EVAL, n, 0);
11863 Set_Node_Length(REGNODE_p(ret), RExC_parse - parse_start + 1);
11864 Set_Node_Offset(REGNODE_p(ret), parse_start);
11867 case '(': /* (?(?{...})...) and (?(?=...)...) */
11870 const int DEFINE_len = sizeof("DEFINE") - 1;
11871 if ( RExC_parse < RExC_end - 1
11872 && ( ( RExC_parse[0] == '?' /* (?(?...)) */
11873 && ( RExC_parse[1] == '='
11874 || RExC_parse[1] == '!'
11875 || RExC_parse[1] == '<'
11876 || RExC_parse[1] == '{'))
11877 || ( RExC_parse[0] == '*' /* (?(*...)) */
11878 && ( memBEGINs(RExC_parse + 1,
11879 (Size_t) (RExC_end - (RExC_parse + 1)),
11881 || memBEGINs(RExC_parse + 1,
11882 (Size_t) (RExC_end - (RExC_parse + 1)),
11884 || memBEGINs(RExC_parse + 1,
11885 (Size_t) (RExC_end - (RExC_parse + 1)),
11887 || memBEGINs(RExC_parse + 1,
11888 (Size_t) (RExC_end - (RExC_parse + 1)),
11890 || memBEGINs(RExC_parse + 1,
11891 (Size_t) (RExC_end - (RExC_parse + 1)),
11892 "positive_lookahead:")
11893 || memBEGINs(RExC_parse + 1,
11894 (Size_t) (RExC_end - (RExC_parse + 1)),
11895 "positive_lookbehind:")
11896 || memBEGINs(RExC_parse + 1,
11897 (Size_t) (RExC_end - (RExC_parse + 1)),
11898 "negative_lookahead:")
11899 || memBEGINs(RExC_parse + 1,
11900 (Size_t) (RExC_end - (RExC_parse + 1)),
11901 "negative_lookbehind:"))))
11902 ) { /* Lookahead or eval. */
11904 regnode_offset tail;
11906 ret = reg_node(pRExC_state, LOGICAL);
11907 FLAGS(REGNODE_p(ret)) = 1;
11909 tail = reg(pRExC_state, 1, &flag, depth+1);
11910 RETURN_FAIL_ON_RESTART(flag, flagp);
11911 if (! REGTAIL(pRExC_state, ret, tail)) {
11912 REQUIRE_BRANCHJ(flagp, 0);
11916 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
11917 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
11919 char ch = RExC_parse[0] == '<' ? '>' : '\'';
11920 char *name_start= RExC_parse++;
11922 SV *sv_dat=reg_scan_name(pRExC_state, REG_RSN_RETURN_DATA);
11923 if ( RExC_parse == name_start
11924 || RExC_parse >= RExC_end
11925 || *RExC_parse != ch)
11927 vFAIL2("Sequence (?(%c... not terminated",
11928 (ch == '>' ? '<' : ch));
11932 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11933 RExC_rxi->data->data[num]=(void*)sv_dat;
11934 SvREFCNT_inc_simple_void_NN(sv_dat);
11936 ret = reganode(pRExC_state, GROUPPN, num);
11937 goto insert_if_check_paren;
11939 else if (memBEGINs(RExC_parse,
11940 (STRLEN) (RExC_end - RExC_parse),
11943 ret = reganode(pRExC_state, DEFINEP, 0);
11944 RExC_parse += DEFINE_len;
11946 goto insert_if_check_paren;
11948 else if (RExC_parse[0] == 'R') {
11950 /* parno == 0 => /(?(R)YES|NO)/ "in any form of recursion OR eval"
11951 * parno == 1 => /(?(R0)YES|NO)/ "in GOSUB (?0) / (?R)"
11952 * parno == 2 => /(?(R1)YES|NO)/ "in GOSUB (?1) (parno-1)"
11955 if (RExC_parse[0] == '0') {
11959 else if (inRANGE(RExC_parse[0], '1', '9')) {
11962 if (grok_atoUV(RExC_parse, &uv, &endptr)
11965 parno = (I32)uv + 1;
11966 RExC_parse = (char*)endptr;
11968 /* else "Switch condition not recognized" below */
11969 } else if (RExC_parse[0] == '&') {
11972 sv_dat = reg_scan_name(pRExC_state,
11973 REG_RSN_RETURN_DATA);
11975 parno = 1 + *((I32 *)SvPVX(sv_dat));
11977 ret = reganode(pRExC_state, INSUBP, parno);
11978 goto insert_if_check_paren;
11980 else if (inRANGE(RExC_parse[0], '1', '9')) {
11985 if (grok_atoUV(RExC_parse, &uv, &endptr)
11989 RExC_parse = (char*)endptr;
11992 vFAIL("panic: grok_atoUV returned FALSE");
11994 ret = reganode(pRExC_state, GROUPP, parno);
11996 insert_if_check_paren:
11997 if (UCHARAT(RExC_parse) != ')') {
11999 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
12001 vFAIL("Switch condition not recognized");
12003 nextchar(pRExC_state);
12005 if (! REGTAIL(pRExC_state, ret, reganode(pRExC_state,
12008 REQUIRE_BRANCHJ(flagp, 0);
12010 br = regbranch(pRExC_state, &flags, 1, depth+1);
12012 RETURN_FAIL_ON_RESTART(flags,flagp);
12013 FAIL2("panic: regbranch returned failure, flags=%#" UVxf,
12016 if (! REGTAIL(pRExC_state, br, reganode(pRExC_state,
12019 REQUIRE_BRANCHJ(flagp, 0);
12021 c = UCHARAT(RExC_parse);
12022 nextchar(pRExC_state);
12023 if (flags&HASWIDTH)
12024 *flagp |= HASWIDTH;
12027 vFAIL("(?(DEFINE)....) does not allow branches");
12029 /* Fake one for optimizer. */
12030 lastbr = reganode(pRExC_state, IFTHEN, 0);
12032 if (!regbranch(pRExC_state, &flags, 1, depth+1)) {
12033 RETURN_FAIL_ON_RESTART(flags, flagp);
12034 FAIL2("panic: regbranch returned failure, flags=%#" UVxf,
12037 if (! REGTAIL(pRExC_state, ret, lastbr)) {
12038 REQUIRE_BRANCHJ(flagp, 0);
12040 if (flags&HASWIDTH)
12041 *flagp |= HASWIDTH;
12042 c = UCHARAT(RExC_parse);
12043 nextchar(pRExC_state);
12048 if (RExC_parse >= RExC_end)
12049 vFAIL("Switch (?(condition)... not terminated");
12051 vFAIL("Switch (?(condition)... contains too many branches");
12053 ender = reg_node(pRExC_state, TAIL);
12054 if (! REGTAIL(pRExC_state, br, ender)) {
12055 REQUIRE_BRANCHJ(flagp, 0);
12058 if (! REGTAIL(pRExC_state, lastbr, ender)) {
12059 REQUIRE_BRANCHJ(flagp, 0);
12061 if (! REGTAIL(pRExC_state,
12064 NEXTOPER(REGNODE_p(lastbr)))),
12067 REQUIRE_BRANCHJ(flagp, 0);
12071 if (! REGTAIL(pRExC_state, ret, ender)) {
12072 REQUIRE_BRANCHJ(flagp, 0);
12074 #if 0 /* Removing this doesn't cause failures in the test suite -- khw */
12075 RExC_size++; /* XXX WHY do we need this?!!
12076 For large programs it seems to be required
12077 but I can't figure out why. -- dmq*/
12082 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
12084 vFAIL("Unknown switch condition (?(...))");
12086 case '[': /* (?[ ... ]) */
12087 return handle_regex_sets(pRExC_state, NULL, flagp, depth+1,
12089 case 0: /* A NUL */
12090 RExC_parse--; /* for vFAIL to print correctly */
12091 vFAIL("Sequence (? incomplete");
12095 if (RExC_strict) { /* [perl #132851] */
12096 ckWARNreg(RExC_parse, "Empty (?) without any modifiers");
12099 case '*': /* If you want to support (?*...), first reconcile with GH #17363 */
12101 default: /* e.g., (?i) */
12102 RExC_parse = (char *) seqstart + 1;
12104 parse_lparen_question_flags(pRExC_state);
12105 if (UCHARAT(RExC_parse) != ':') {
12106 if (RExC_parse < RExC_end)
12107 nextchar(pRExC_state);
12112 nextchar(pRExC_state);
12117 else if (!(RExC_flags & RXf_PMf_NOCAPTURE)) { /* (...) */
12121 if (! ALL_PARENS_COUNTED) {
12122 /* If we are in our first pass through (and maybe only pass),
12123 * we need to allocate memory for the capturing parentheses
12127 if (!RExC_parens_buf_size) {
12128 /* first guess at number of parens we might encounter */
12129 RExC_parens_buf_size = 10;
12131 /* setup RExC_open_parens, which holds the address of each
12132 * OPEN tag, and to make things simpler for the 0 index the
12133 * start of the program - this is used later for offsets */
12134 Newxz(RExC_open_parens, RExC_parens_buf_size,
12136 RExC_open_parens[0] = 1; /* +1 for REG_MAGIC */
12138 /* setup RExC_close_parens, which holds the address of each
12139 * CLOSE tag, and to make things simpler for the 0 index
12140 * the end of the program - this is used later for offsets
12142 Newxz(RExC_close_parens, RExC_parens_buf_size,
12144 /* we dont know where end op starts yet, so we dont need to
12145 * set RExC_close_parens[0] like we do RExC_open_parens[0]
12148 else if (RExC_npar > RExC_parens_buf_size) {
12149 I32 old_size = RExC_parens_buf_size;
12151 RExC_parens_buf_size *= 2;
12153 Renew(RExC_open_parens, RExC_parens_buf_size,
12155 Zero(RExC_open_parens + old_size,
12156 RExC_parens_buf_size - old_size, regnode_offset);
12158 Renew(RExC_close_parens, RExC_parens_buf_size,
12160 Zero(RExC_close_parens + old_size,
12161 RExC_parens_buf_size - old_size, regnode_offset);
12165 ret = reganode(pRExC_state, OPEN, parno);
12166 if (!RExC_nestroot)
12167 RExC_nestroot = parno;
12168 if (RExC_open_parens && !RExC_open_parens[parno])
12170 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12171 "%*s%*s Setting open paren #%" IVdf " to %zu\n",
12172 22, "| |", (int)(depth * 2 + 1), "",
12174 RExC_open_parens[parno]= ret;
12177 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
12178 Set_Node_Offset(REGNODE_p(ret), RExC_parse); /* MJD */
12181 /* with RXf_PMf_NOCAPTURE treat (...) as (?:...) */
12190 /* Pick up the branches, linking them together. */
12191 parse_start = RExC_parse; /* MJD */
12192 br = regbranch(pRExC_state, &flags, 1, depth+1);
12194 /* branch_len = (paren != 0); */
12197 RETURN_FAIL_ON_RESTART(flags, flagp);
12198 FAIL2("panic: regbranch returned failure, flags=%#" UVxf, (UV) flags);
12200 if (*RExC_parse == '|') {
12201 if (RExC_use_BRANCHJ) {
12202 reginsert(pRExC_state, BRANCHJ, br, depth+1);
12205 reginsert(pRExC_state, BRANCH, br, depth+1);
12206 Set_Node_Length(REGNODE_p(br), paren != 0);
12207 Set_Node_Offset_To_R(br, parse_start-RExC_start);
12211 else if (paren == ':') {
12212 *flagp |= flags&SIMPLE;
12214 if (is_open) { /* Starts with OPEN. */
12215 if (! REGTAIL(pRExC_state, ret, br)) { /* OPEN -> first. */
12216 REQUIRE_BRANCHJ(flagp, 0);
12219 else if (paren != '?') /* Not Conditional */
12221 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
12223 while (*RExC_parse == '|') {
12224 if (RExC_use_BRANCHJ) {
12227 ender = reganode(pRExC_state, LONGJMP, 0);
12229 /* Append to the previous. */
12230 shut_gcc_up = REGTAIL(pRExC_state,
12231 REGNODE_OFFSET(NEXTOPER(NEXTOPER(REGNODE_p(lastbr)))),
12233 PERL_UNUSED_VAR(shut_gcc_up);
12235 nextchar(pRExC_state);
12236 if (freeze_paren) {
12237 if (RExC_npar > after_freeze)
12238 after_freeze = RExC_npar;
12239 RExC_npar = freeze_paren;
12241 br = regbranch(pRExC_state, &flags, 0, depth+1);
12244 RETURN_FAIL_ON_RESTART(flags, flagp);
12245 FAIL2("panic: regbranch returned failure, flags=%#" UVxf, (UV) flags);
12247 if (! REGTAIL(pRExC_state, lastbr, br)) { /* BRANCH -> BRANCH. */
12248 REQUIRE_BRANCHJ(flagp, 0);
12251 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
12254 if (have_branch || paren != ':') {
12257 /* Make a closing node, and hook it on the end. */
12260 ender = reg_node(pRExC_state, TAIL);
12263 ender = reganode(pRExC_state, CLOSE, parno);
12264 if ( RExC_close_parens ) {
12265 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12266 "%*s%*s Setting close paren #%" IVdf " to %zu\n",
12267 22, "| |", (int)(depth * 2 + 1), "",
12268 (IV)parno, ender));
12269 RExC_close_parens[parno]= ender;
12270 if (RExC_nestroot == parno)
12273 Set_Node_Offset(REGNODE_p(ender), RExC_parse+1); /* MJD */
12274 Set_Node_Length(REGNODE_p(ender), 1); /* MJD */
12277 ender = reg_node(pRExC_state, SRCLOSE);
12278 RExC_in_script_run = 0;
12288 *flagp &= ~HASWIDTH;
12290 case 't': /* aTomic */
12292 ender = reg_node(pRExC_state, SUCCEED);
12295 ender = reg_node(pRExC_state, END);
12296 assert(!RExC_end_op); /* there can only be one! */
12297 RExC_end_op = REGNODE_p(ender);
12298 if (RExC_close_parens) {
12299 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12300 "%*s%*s Setting close paren #0 (END) to %zu\n",
12301 22, "| |", (int)(depth * 2 + 1), "",
12304 RExC_close_parens[0]= ender;
12309 DEBUG_PARSE_MSG("lsbr");
12310 regprop(RExC_rx, RExC_mysv1, REGNODE_p(lastbr), NULL, pRExC_state);
12311 regprop(RExC_rx, RExC_mysv2, REGNODE_p(ender), NULL, pRExC_state);
12312 Perl_re_printf( aTHX_ "~ tying lastbr %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
12313 SvPV_nolen_const(RExC_mysv1),
12315 SvPV_nolen_const(RExC_mysv2),
12317 (IV)(ender - lastbr)
12320 if (! REGTAIL(pRExC_state, lastbr, ender)) {
12321 REQUIRE_BRANCHJ(flagp, 0);
12325 char is_nothing= 1;
12327 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
12329 /* Hook the tails of the branches to the closing node. */
12330 for (br = REGNODE_p(ret); br; br = regnext(br)) {
12331 const U8 op = PL_regkind[OP(br)];
12332 if (op == BRANCH) {
12333 if (! REGTAIL_STUDY(pRExC_state,
12334 REGNODE_OFFSET(NEXTOPER(br)),
12337 REQUIRE_BRANCHJ(flagp, 0);
12339 if ( OP(NEXTOPER(br)) != NOTHING
12340 || regnext(NEXTOPER(br)) != REGNODE_p(ender))
12343 else if (op == BRANCHJ) {
12344 bool shut_gcc_up = REGTAIL_STUDY(pRExC_state,
12345 REGNODE_OFFSET(NEXTOPER(NEXTOPER(br))),
12347 PERL_UNUSED_VAR(shut_gcc_up);
12348 /* for now we always disable this optimisation * /
12349 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
12350 || regnext(NEXTOPER(NEXTOPER(br))) != REGNODE_p(ender))
12356 regnode * ret_as_regnode = REGNODE_p(ret);
12357 br= PL_regkind[OP(ret_as_regnode)] != BRANCH
12358 ? regnext(ret_as_regnode)
12361 DEBUG_PARSE_MSG("NADA");
12362 regprop(RExC_rx, RExC_mysv1, ret_as_regnode,
12363 NULL, pRExC_state);
12364 regprop(RExC_rx, RExC_mysv2, REGNODE_p(ender),
12365 NULL, pRExC_state);
12366 Perl_re_printf( aTHX_ "~ converting ret %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
12367 SvPV_nolen_const(RExC_mysv1),
12368 (IV)REG_NODE_NUM(ret_as_regnode),
12369 SvPV_nolen_const(RExC_mysv2),
12375 if (OP(REGNODE_p(ender)) == TAIL) {
12377 RExC_emit= REGNODE_OFFSET(br) + 1;
12380 for ( opt= br + 1; opt < REGNODE_p(ender) ; opt++ )
12381 OP(opt)= OPTIMIZED;
12382 NEXT_OFF(br)= REGNODE_p(ender) - br;
12390 /* Even/odd or x=don't care: 010101x10x */
12391 static const char parens[] = "=!aA<,>Bbt";
12392 /* flag below is set to 0 up through 'A'; 1 for larger */
12394 if (paren && (p = strchr(parens, paren))) {
12395 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
12396 int flag = (p - parens) > 3;
12398 if (paren == '>' || paren == 't') {
12399 node = SUSPEND, flag = 0;
12402 reginsert(pRExC_state, node, ret, depth+1);
12403 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
12404 Set_Node_Offset(REGNODE_p(ret), parse_start + 1);
12405 FLAGS(REGNODE_p(ret)) = flag;
12406 if (! REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL)))
12408 REQUIRE_BRANCHJ(flagp, 0);
12413 /* Check for proper termination. */
12415 /* restore original flags, but keep (?p) and, if we've encountered
12416 * something in the parse that changes /d rules into /u, keep the /u */
12417 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
12418 if (DEPENDS_SEMANTICS && RExC_uni_semantics) {
12419 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
12421 if (RExC_parse >= RExC_end || UCHARAT(RExC_parse) != ')') {
12422 RExC_parse = oregcomp_parse;
12423 vFAIL("Unmatched (");
12425 nextchar(pRExC_state);
12427 else if (!paren && RExC_parse < RExC_end) {
12428 if (*RExC_parse == ')') {
12430 vFAIL("Unmatched )");
12433 FAIL("Junk on end of regexp"); /* "Can't happen". */
12434 NOT_REACHED; /* NOTREACHED */
12437 if (RExC_in_lookbehind) {
12438 RExC_in_lookbehind--;
12440 if (RExC_in_lookahead) {
12441 RExC_in_lookahead--;
12443 if (after_freeze > RExC_npar)
12444 RExC_npar = after_freeze;
12449 - regbranch - one alternative of an | operator
12451 * Implements the concatenation operator.
12453 * On success, returns the offset at which any next node should be placed into
12454 * the regex engine program being compiled.
12456 * Returns 0 otherwise, setting flagp to RESTART_PARSE if the parse needs
12457 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to
12460 STATIC regnode_offset
12461 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
12463 regnode_offset ret;
12464 regnode_offset chain = 0;
12465 regnode_offset latest;
12466 I32 flags = 0, c = 0;
12467 DECLARE_AND_GET_RE_DEBUG_FLAGS;
12469 PERL_ARGS_ASSERT_REGBRANCH;
12471 DEBUG_PARSE("brnc");
12476 if (RExC_use_BRANCHJ)
12477 ret = reganode(pRExC_state, BRANCHJ, 0);
12479 ret = reg_node(pRExC_state, BRANCH);
12480 Set_Node_Length(REGNODE_p(ret), 1);
12484 *flagp = WORST; /* Tentatively. */
12486 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
12487 FALSE /* Don't force to /x */ );
12488 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
12489 flags &= ~TRYAGAIN;
12490 latest = regpiece(pRExC_state, &flags, depth+1);
12492 if (flags & TRYAGAIN)
12494 RETURN_FAIL_ON_RESTART(flags, flagp);
12495 FAIL2("panic: regpiece returned failure, flags=%#" UVxf, (UV) flags);
12499 *flagp |= flags&(HASWIDTH|POSTPONED);
12500 if (chain == 0) /* First piece. */
12501 *flagp |= flags&SPSTART;
12503 /* FIXME adding one for every branch after the first is probably
12504 * excessive now we have TRIE support. (hv) */
12506 if (! REGTAIL(pRExC_state, chain, latest)) {
12507 /* XXX We could just redo this branch, but figuring out what
12508 * bookkeeping needs to be reset is a pain, and it's likely
12509 * that other branches that goto END will also be too large */
12510 REQUIRE_BRANCHJ(flagp, 0);
12516 if (chain == 0) { /* Loop ran zero times. */
12517 chain = reg_node(pRExC_state, NOTHING);
12522 *flagp |= flags&SIMPLE;
12529 - regpiece - something followed by possible quantifier * + ? {n,m}
12531 * Note that the branching code sequences used for ? and the general cases
12532 * of * and + are somewhat optimized: they use the same NOTHING node as
12533 * both the endmarker for their branch list and the body of the last branch.
12534 * It might seem that this node could be dispensed with entirely, but the
12535 * endmarker role is not redundant.
12537 * On success, returns the offset at which any next node should be placed into
12538 * the regex engine program being compiled.
12540 * Returns 0 otherwise, with *flagp set to indicate why:
12541 * TRYAGAIN if regatom() returns 0 with TRYAGAIN.
12542 * RESTART_PARSE if the parse needs to be restarted, or'd with
12543 * NEED_UTF8 if the pattern needs to be upgraded to UTF-8.
12545 STATIC regnode_offset
12546 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
12548 regnode_offset ret;
12552 const char * const origparse = RExC_parse;
12554 I32 max = REG_INFTY;
12555 #ifdef RE_TRACK_PATTERN_OFFSETS
12558 const char *maxpos = NULL;
12561 /* Save the original in case we change the emitted regop to a FAIL. */
12562 const regnode_offset orig_emit = RExC_emit;
12564 DECLARE_AND_GET_RE_DEBUG_FLAGS;
12566 PERL_ARGS_ASSERT_REGPIECE;
12568 DEBUG_PARSE("piec");
12570 ret = regatom(pRExC_state, &flags, depth+1);
12572 RETURN_FAIL_ON_RESTART_OR_FLAGS(flags, flagp, TRYAGAIN);
12573 FAIL2("panic: regatom returned failure, flags=%#" UVxf, (UV) flags);
12578 if (op == '{' && regcurly(RExC_parse)) {
12580 #ifdef RE_TRACK_PATTERN_OFFSETS
12581 parse_start = RExC_parse; /* MJD */
12583 next = RExC_parse + 1;
12584 while (isDIGIT(*next) || *next == ',') {
12585 if (*next == ',') {
12593 if (*next == '}') { /* got one */
12594 const char* endptr;
12598 if (isDIGIT(*RExC_parse)) {
12600 if (!grok_atoUV(RExC_parse, &uv, &endptr))
12601 vFAIL("Invalid quantifier in {,}");
12602 if (uv >= REG_INFTY)
12603 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
12608 if (*maxpos == ',')
12611 maxpos = RExC_parse;
12612 if (isDIGIT(*maxpos)) {
12614 if (!grok_atoUV(maxpos, &uv, &endptr))
12615 vFAIL("Invalid quantifier in {,}");
12616 if (uv >= REG_INFTY)
12617 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
12620 max = REG_INFTY; /* meaning "infinity" */
12623 nextchar(pRExC_state);
12624 if (max < min) { /* If can't match, warn and optimize to fail
12626 reginsert(pRExC_state, OPFAIL, orig_emit, depth+1);
12627 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
12628 NEXT_OFF(REGNODE_p(orig_emit)) =
12629 regarglen[OPFAIL] + NODE_STEP_REGNODE;
12632 else if (min == max && *RExC_parse == '?')
12634 ckWARN2reg(RExC_parse + 1,
12635 "Useless use of greediness modifier '%c'",
12640 if ((flags&SIMPLE)) {
12641 if (min == 0 && max == REG_INFTY) {
12643 /* Going from 0..inf is currently forbidden in wildcard
12644 * subpatterns. The only reason is to make it harder to
12645 * write patterns that take a long long time to halt, and
12646 * because the use of this construct isn't necessary in
12647 * matching Unicode property values */
12648 if (RExC_pm_flags & PMf_WILDCARD) {
12650 /* diag_listed_as: Use of %s is not allowed in Unicode
12651 property wildcard subpatterns in regex; marked by
12652 <-- HERE in m/%s/ */
12653 vFAIL("Use of quantifier '*' is not allowed in"
12654 " Unicode property wildcard subpatterns");
12655 /* Note, don't need to worry about {0,}, as a '}' isn't
12656 * legal at all in wildcards, so wouldn't get this far
12659 reginsert(pRExC_state, STAR, ret, depth+1);
12661 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
12664 if (min == 1 && max == REG_INFTY) {
12665 reginsert(pRExC_state, PLUS, ret, depth+1);
12667 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
12670 MARK_NAUGHTY_EXP(2, 2);
12671 reginsert(pRExC_state, CURLY, ret, depth+1);
12672 Set_Node_Offset(REGNODE_p(ret), parse_start+1); /* MJD */
12673 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
12676 const regnode_offset w = reg_node(pRExC_state, WHILEM);
12678 FLAGS(REGNODE_p(w)) = 0;
12679 if (! REGTAIL(pRExC_state, ret, w)) {
12680 REQUIRE_BRANCHJ(flagp, 0);
12682 if (RExC_use_BRANCHJ) {
12683 reginsert(pRExC_state, LONGJMP, ret, depth+1);
12684 reginsert(pRExC_state, NOTHING, ret, depth+1);
12685 NEXT_OFF(REGNODE_p(ret)) = 3; /* Go over LONGJMP. */
12687 reginsert(pRExC_state, CURLYX, ret, depth+1);
12689 Set_Node_Offset(REGNODE_p(ret), parse_start+1);
12690 Set_Node_Length(REGNODE_p(ret),
12691 op == '{' ? (RExC_parse - parse_start) : 1);
12693 if (RExC_use_BRANCHJ)
12694 NEXT_OFF(REGNODE_p(ret)) = 3; /* Go over NOTHING to
12696 if (! REGTAIL(pRExC_state, ret, reg_node(pRExC_state,
12699 REQUIRE_BRANCHJ(flagp, 0);
12701 RExC_whilem_seen++;
12702 MARK_NAUGHTY_EXP(1, 4); /* compound interest */
12704 FLAGS(REGNODE_p(ret)) = 0;
12709 *flagp |= HASWIDTH;
12710 ARG1_SET(REGNODE_p(ret), (U16)min);
12711 ARG2_SET(REGNODE_p(ret), (U16)max);
12712 if (max == REG_INFTY)
12713 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
12719 if (!ISMULT1(op)) {
12724 #if 0 /* Now runtime fix should be reliable. */
12726 /* if this is reinstated, don't forget to put this back into perldiag:
12728 =item Regexp *+ operand could be empty at {#} in regex m/%s/
12730 (F) The part of the regexp subject to either the * or + quantifier
12731 could match an empty string. The {#} shows in the regular
12732 expression about where the problem was discovered.
12736 if (!(flags&HASWIDTH) && op != '?')
12737 vFAIL("Regexp *+ operand could be empty");
12740 #ifdef RE_TRACK_PATTERN_OFFSETS
12741 parse_start = RExC_parse;
12743 nextchar(pRExC_state);
12745 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
12751 else if (op == '+') {
12755 else if (op == '?') {
12760 if (!(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
12761 if (origparse[0] == '\\' && origparse[1] == 'K') {
12763 "%" UTF8f " is forbidden - matches null string many times",
12764 UTF8fARG(UTF, (RExC_parse >= origparse
12765 ? RExC_parse - origparse
12770 ckWARN2reg(RExC_parse,
12771 "%" UTF8f " matches null string many times",
12772 UTF8fARG(UTF, (RExC_parse >= origparse
12773 ? RExC_parse - origparse
12779 if (*RExC_parse == '?') {
12780 nextchar(pRExC_state);
12781 reginsert(pRExC_state, MINMOD, ret, depth+1);
12782 if (! REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE)) {
12783 REQUIRE_BRANCHJ(flagp, 0);
12786 else if (*RExC_parse == '+') {
12787 regnode_offset ender;
12788 nextchar(pRExC_state);
12789 ender = reg_node(pRExC_state, SUCCEED);
12790 if (! REGTAIL(pRExC_state, ret, ender)) {
12791 REQUIRE_BRANCHJ(flagp, 0);
12793 reginsert(pRExC_state, SUSPEND, ret, depth+1);
12794 ender = reg_node(pRExC_state, TAIL);
12795 if (! REGTAIL(pRExC_state, ret, ender)) {
12796 REQUIRE_BRANCHJ(flagp, 0);
12800 if (ISMULT2(RExC_parse)) {
12802 vFAIL("Nested quantifiers");
12809 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state,
12810 regnode_offset * node_p,
12818 /* This routine teases apart the various meanings of \N and returns
12819 * accordingly. The input parameters constrain which meaning(s) is/are valid
12820 * in the current context.
12822 * Exactly one of <node_p> and <code_point_p> must be non-NULL.
12824 * If <code_point_p> is not NULL, the context is expecting the result to be a
12825 * single code point. If this \N instance turns out to a single code point,
12826 * the function returns TRUE and sets *code_point_p to that code point.
12828 * If <node_p> is not NULL, the context is expecting the result to be one of
12829 * the things representable by a regnode. If this \N instance turns out to be
12830 * one such, the function generates the regnode, returns TRUE and sets *node_p
12831 * to point to the offset of that regnode into the regex engine program being
12834 * If this instance of \N isn't legal in any context, this function will
12835 * generate a fatal error and not return.
12837 * On input, RExC_parse should point to the first char following the \N at the
12838 * time of the call. On successful return, RExC_parse will have been updated
12839 * to point to just after the sequence identified by this routine. Also
12840 * *flagp has been updated as needed.
12842 * When there is some problem with the current context and this \N instance,
12843 * the function returns FALSE, without advancing RExC_parse, nor setting
12844 * *node_p, nor *code_point_p, nor *flagp.
12846 * If <cp_count> is not NULL, the caller wants to know the length (in code
12847 * points) that this \N sequence matches. This is set, and the input is
12848 * parsed for errors, even if the function returns FALSE, as detailed below.
12850 * There are 6 possibilities here, as detailed in the next 6 paragraphs.
12852 * Probably the most common case is for the \N to specify a single code point.
12853 * *cp_count will be set to 1, and *code_point_p will be set to that code
12856 * Another possibility is for the input to be an empty \N{}. This is no
12857 * longer accepted, and will generate a fatal error.
12859 * Another possibility is for a custom charnames handler to be in effect which
12860 * translates the input name to an empty string. *cp_count will be set to 0.
12861 * *node_p will be set to a generated NOTHING node.
12863 * Still another possibility is for the \N to mean [^\n]. *cp_count will be
12864 * set to 0. *node_p will be set to a generated REG_ANY node.
12866 * The fifth possibility is that \N resolves to a sequence of more than one
12867 * code points. *cp_count will be set to the number of code points in the
12868 * sequence. *node_p will be set to a generated node returned by this
12869 * function calling S_reg().
12871 * The final possibility is that it is premature to be calling this function;
12872 * the parse needs to be restarted. This can happen when this changes from
12873 * /d to /u rules, or when the pattern needs to be upgraded to UTF-8. The
12874 * latter occurs only when the fifth possibility would otherwise be in
12875 * effect, and is because one of those code points requires the pattern to be
12876 * recompiled as UTF-8. The function returns FALSE, and sets the
12877 * RESTART_PARSE and NEED_UTF8 flags in *flagp, as appropriate. When this
12878 * happens, the caller needs to desist from continuing parsing, and return
12879 * this information to its caller. This is not set for when there is only one
12880 * code point, as this can be called as part of an ANYOF node, and they can
12881 * store above-Latin1 code points without the pattern having to be in UTF-8.
12883 * For non-single-quoted regexes, the tokenizer has resolved character and
12884 * sequence names inside \N{...} into their Unicode values, normalizing the
12885 * result into what we should see here: '\N{U+c1.c2...}', where c1... are the
12886 * hex-represented code points in the sequence. This is done there because
12887 * the names can vary based on what charnames pragma is in scope at the time,
12888 * so we need a way to take a snapshot of what they resolve to at the time of
12889 * the original parse. [perl #56444].
12891 * That parsing is skipped for single-quoted regexes, so here we may get
12892 * '\N{NAME}', which is parsed now. If the single-quoted regex is something
12893 * like '\N{U+41}', that code point is Unicode, and has to be translated into
12894 * the native character set for non-ASCII platforms. The other possibilities
12895 * are already native, so no translation is done. */
12897 char * endbrace; /* points to '}' following the name */
12898 char* p = RExC_parse; /* Temporary */
12900 SV * substitute_parse = NULL;
12905 DECLARE_AND_GET_RE_DEBUG_FLAGS;
12907 PERL_ARGS_ASSERT_GROK_BSLASH_N;
12909 assert(cBOOL(node_p) ^ cBOOL(code_point_p)); /* Exactly one should be set */
12910 assert(! (node_p && cp_count)); /* At most 1 should be set */
12912 if (cp_count) { /* Initialize return for the most common case */
12916 /* The [^\n] meaning of \N ignores spaces and comments under the /x
12917 * modifier. The other meanings do not, so use a temporary until we find
12918 * out which we are being called with */
12919 skip_to_be_ignored_text(pRExC_state, &p,
12920 FALSE /* Don't force to /x */ );
12922 /* Disambiguate between \N meaning a named character versus \N meaning
12923 * [^\n]. The latter is assumed when the {...} following the \N is a legal
12924 * quantifier, or if there is no '{' at all */
12925 if (*p != '{' || regcurly(p)) {
12935 *node_p = reg_node(pRExC_state, REG_ANY);
12936 *flagp |= HASWIDTH|SIMPLE;
12938 Set_Node_Length(REGNODE_p(*(node_p)), 1); /* MJD */
12942 /* The test above made sure that the next real character is a '{', but
12943 * under the /x modifier, it could be separated by space (or a comment and
12944 * \n) and this is not allowed (for consistency with \x{...} and the
12945 * tokenizer handling of \N{NAME}). */
12946 if (*RExC_parse != '{') {
12947 vFAIL("Missing braces on \\N{}");
12950 RExC_parse++; /* Skip past the '{' */
12952 endbrace = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
12953 if (! endbrace) { /* no trailing brace */
12954 vFAIL2("Missing right brace on \\%c{}", 'N');
12957 /* Here, we have decided it should be a named character or sequence. These
12958 * imply Unicode semantics */
12959 REQUIRE_UNI_RULES(flagp, FALSE);
12961 /* \N{_} is what toke.c returns to us to indicate a name that evaluates to
12962 * nothing at all (not allowed under strict) */
12963 if (endbrace - RExC_parse == 1 && *RExC_parse == '_') {
12964 RExC_parse = endbrace;
12966 RExC_parse++; /* Position after the "}" */
12967 vFAIL("Zero length \\N{}");
12973 nextchar(pRExC_state);
12978 *node_p = reg_node(pRExC_state, NOTHING);
12982 if (endbrace - RExC_parse < 2 || ! strBEGINs(RExC_parse, "U+")) {
12984 /* Here, the name isn't of the form U+.... This can happen if the
12985 * pattern is single-quoted, so didn't get evaluated in toke.c. Now
12986 * is the time to find out what the name means */
12988 const STRLEN name_len = endbrace - RExC_parse;
12989 SV * value_sv; /* What does this name evaluate to */
12991 const U8 * value; /* string of name's value */
12992 STRLEN value_len; /* and its length */
12994 /* RExC_unlexed_names is a hash of names that weren't evaluated by
12995 * toke.c, and their values. Make sure is initialized */
12996 if (! RExC_unlexed_names) {
12997 RExC_unlexed_names = newHV();
13000 /* If we have already seen this name in this pattern, use that. This
13001 * allows us to only call the charnames handler once per name per
13002 * pattern. A broken or malicious handler could return something
13003 * different each time, which could cause the results to vary depending
13004 * on if something gets added or subtracted from the pattern that
13005 * causes the number of passes to change, for example */
13006 if ((value_svp = hv_fetch(RExC_unlexed_names, RExC_parse,
13009 value_sv = *value_svp;
13011 else { /* Otherwise we have to go out and get the name */
13012 const char * error_msg = NULL;
13013 value_sv = get_and_check_backslash_N_name(RExC_parse, endbrace,
13017 RExC_parse = endbrace;
13021 /* If no error message, should have gotten a valid return */
13024 /* Save the name's meaning for later use */
13025 if (! hv_store(RExC_unlexed_names, RExC_parse, name_len,
13028 Perl_croak(aTHX_ "panic: hv_store() unexpectedly failed");
13032 /* Here, we have the value the name evaluates to in 'value_sv' */
13033 value = (U8 *) SvPV(value_sv, value_len);
13035 /* See if the result is one code point vs 0 or multiple */
13036 if (inRANGE(value_len, 1, ((UV) SvUTF8(value_sv)
13040 /* Here, exactly one code point. If that isn't what is wanted,
13042 if (! code_point_p) {
13047 /* Convert from string to numeric code point */
13048 *code_point_p = (SvUTF8(value_sv))
13049 ? valid_utf8_to_uvchr(value, NULL)
13052 /* Have parsed this entire single code point \N{...}. *cp_count
13053 * has already been set to 1, so don't do it again. */
13054 RExC_parse = endbrace;
13055 nextchar(pRExC_state);
13057 } /* End of is a single code point */
13059 /* Count the code points, if caller desires. The API says to do this
13060 * even if we will later return FALSE */
13064 *cp_count = (SvUTF8(value_sv))
13065 ? utf8_length(value, value + value_len)
13069 /* Fail if caller doesn't want to handle a multi-code-point sequence.
13070 * But don't back the pointer up if the caller wants to know how many
13071 * code points there are (they need to handle it themselves in this
13080 /* Convert this to a sub-pattern of the form "(?: ... )", and then call
13081 * reg recursively to parse it. That way, it retains its atomicness,
13082 * while not having to worry about any special handling that some code
13083 * points may have. */
13085 substitute_parse = newSVpvs("?:");
13086 sv_catsv(substitute_parse, value_sv);
13087 sv_catpv(substitute_parse, ")");
13089 /* The value should already be native, so no need to convert on EBCDIC
13091 assert(! RExC_recode_x_to_native);
13094 else { /* \N{U+...} */
13095 Size_t count = 0; /* code point count kept internally */
13097 /* We can get to here when the input is \N{U+...} or when toke.c has
13098 * converted a name to the \N{U+...} form. This include changing a
13099 * name that evaluates to multiple code points to \N{U+c1.c2.c3 ...} */
13101 RExC_parse += 2; /* Skip past the 'U+' */
13103 /* Code points are separated by dots. The '}' terminates the whole
13106 do { /* Loop until the ending brace */
13107 I32 flags = PERL_SCAN_SILENT_OVERFLOW
13108 | PERL_SCAN_SILENT_ILLDIGIT
13109 | PERL_SCAN_NOTIFY_ILLDIGIT
13110 | PERL_SCAN_ALLOW_MEDIAL_UNDERSCORES
13111 | PERL_SCAN_DISALLOW_PREFIX;
13112 STRLEN len = endbrace - RExC_parse;
13114 char * start_digit = RExC_parse;
13115 UV cp = grok_hex(RExC_parse, &len, &flags, &overflow_value);
13120 vFAIL("Invalid hexadecimal number in \\N{U+...}");
13125 if (cp > MAX_LEGAL_CP) {
13126 vFAIL(form_cp_too_large_msg(16, start_digit, len, 0));
13129 if (RExC_parse >= endbrace) { /* Got to the closing '}' */
13134 /* Here, is a single code point; fail if doesn't want that */
13135 if (! code_point_p) {
13140 /* A single code point is easy to handle; just return it */
13141 *code_point_p = UNI_TO_NATIVE(cp);
13142 RExC_parse = endbrace;
13143 nextchar(pRExC_state);
13147 /* Here, the parse stopped bfore the ending brace. This is legal
13148 * only if that character is a dot separating code points, like a
13149 * multiple character sequence (of the form "\N{U+c1.c2. ... }".
13150 * So the next character must be a dot (and the one after that
13151 * can't be the endbrace, or we'd have something like \N{U+100.} )
13153 if (*RExC_parse != '.' || RExC_parse + 1 >= endbrace) {
13154 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
13155 ? UTF8SKIP(RExC_parse)
13157 RExC_parse = MIN(endbrace, RExC_parse);/* Guard against
13162 /* Here, looks like its really a multiple character sequence. Fail
13163 * if that's not what the caller wants. But continue with counting
13164 * and error checking if they still want a count */
13165 if (! node_p && ! cp_count) {
13169 /* What is done here is to convert this to a sub-pattern of the
13170 * form \x{char1}\x{char2}... and then call reg recursively to
13171 * parse it (enclosing in "(?: ... )" ). That way, it retains its
13172 * atomicness, while not having to worry about special handling
13173 * that some code points may have. We don't create a subpattern,
13174 * but go through the motions of code point counting and error
13175 * checking, if the caller doesn't want a node returned. */
13177 if (node_p && ! substitute_parse) {
13178 substitute_parse = newSVpvs("?:");
13184 /* Convert to notation the rest of the code understands */
13185 sv_catpvs(substitute_parse, "\\x{");
13186 sv_catpvn(substitute_parse, start_digit,
13187 RExC_parse - start_digit);
13188 sv_catpvs(substitute_parse, "}");
13191 /* Move to after the dot (or ending brace the final time through.)
13196 } while (RExC_parse < endbrace);
13198 if (! node_p) { /* Doesn't want the node */
13205 sv_catpvs(substitute_parse, ")");
13207 /* The values are Unicode, and therefore have to be converted to native
13208 * on a non-Unicode (meaning non-ASCII) platform. */
13209 SET_recode_x_to_native(1);
13212 /* Here, we have the string the name evaluates to, ready to be parsed,
13213 * stored in 'substitute_parse' as a series of valid "\x{...}\x{...}"
13214 * constructs. This can be called from within a substitute parse already.
13215 * The error reporting mechanism doesn't work for 2 levels of this, but the
13216 * code above has validated this new construct, so there should be no
13217 * errors generated by the below. And this isn' an exact copy, so the
13218 * mechanism to seamlessly deal with this won't work, so turn off warnings
13220 save_start = RExC_start;
13221 orig_end = RExC_end;
13223 RExC_parse = RExC_start = SvPVX(substitute_parse);
13224 RExC_end = RExC_parse + SvCUR(substitute_parse);
13225 TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE;
13227 *node_p = reg(pRExC_state, 1, &flags, depth+1);
13229 /* Restore the saved values */
13231 RExC_start = save_start;
13232 RExC_parse = endbrace;
13233 RExC_end = orig_end;
13234 SET_recode_x_to_native(0);
13236 SvREFCNT_dec_NN(substitute_parse);
13239 RETURN_FAIL_ON_RESTART(flags, flagp);
13240 FAIL2("panic: reg returned failure to grok_bslash_N, flags=%#" UVxf,
13243 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
13245 nextchar(pRExC_state);
13251 PERL_STATIC_INLINE U8
13252 S_compute_EXACTish(RExC_state_t *pRExC_state)
13256 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
13264 op = get_regex_charset(RExC_flags);
13265 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
13266 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
13267 been, so there is no hole */
13270 return op + EXACTF;
13274 S_new_regcurly(const char *s, const char *e)
13276 /* This is a temporary function designed to match the most lenient form of
13277 * a {m,n} quantifier we ever envision, with either number omitted, and
13278 * spaces anywhere between/before/after them.
13280 * If this function fails, then the string it matches is very unlikely to
13281 * ever be considered a valid quantifier, so we can allow the '{' that
13282 * begins it to be considered as a literal */
13284 bool has_min = FALSE;
13285 bool has_max = FALSE;
13287 PERL_ARGS_ASSERT_NEW_REGCURLY;
13289 if (s >= e || *s++ != '{')
13292 while (s < e && isSPACE(*s)) {
13295 while (s < e && isDIGIT(*s)) {
13299 while (s < e && isSPACE(*s)) {
13305 while (s < e && isSPACE(*s)) {
13308 while (s < e && isDIGIT(*s)) {
13312 while (s < e && isSPACE(*s)) {
13317 return s < e && *s == '}' && (has_min || has_max);
13320 /* Parse backref decimal value, unless it's too big to sensibly be a backref,
13321 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
13324 S_backref_value(char *p, char *e)
13326 const char* endptr = e;
13328 if (grok_atoUV(p, &val, &endptr) && val <= I32_MAX)
13335 - regatom - the lowest level
13337 Try to identify anything special at the start of the current parse position.
13338 If there is, then handle it as required. This may involve generating a
13339 single regop, such as for an assertion; or it may involve recursing, such as
13340 to handle a () structure.
13342 If the string doesn't start with something special then we gobble up
13343 as much literal text as we can. If we encounter a quantifier, we have to
13344 back off the final literal character, as that quantifier applies to just it
13345 and not to the whole string of literals.
13347 Once we have been able to handle whatever type of thing started the
13348 sequence, we return the offset into the regex engine program being compiled
13349 at which any next regnode should be placed.
13351 Returns 0, setting *flagp to TRYAGAIN if reg() returns 0 with TRYAGAIN.
13352 Returns 0, setting *flagp to RESTART_PARSE if the parse needs to be
13353 restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
13354 Otherwise does not return 0.
13356 Note: we have to be careful with escapes, as they can be both literal
13357 and special, and in the case of \10 and friends, context determines which.
13359 A summary of the code structure is:
13361 switch (first_byte) {
13362 cases for each special:
13363 handle this special;
13366 switch (2nd byte) {
13367 cases for each unambiguous special:
13368 handle this special;
13370 cases for each ambigous special/literal:
13372 if (special) handle here
13374 default: // unambiguously literal:
13377 default: // is a literal char
13380 create EXACTish node for literal;
13381 while (more input and node isn't full) {
13382 switch (input_byte) {
13383 cases for each special;
13384 make sure parse pointer is set so that the next call to
13385 regatom will see this special first
13386 goto loopdone; // EXACTish node terminated by prev. char
13388 append char to EXACTISH node;
13390 get next input byte;
13394 return the generated node;
13396 Specifically there are two separate switches for handling
13397 escape sequences, with the one for handling literal escapes requiring
13398 a dummy entry for all of the special escapes that are actually handled
13403 STATIC regnode_offset
13404 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
13407 regnode_offset ret = 0;
13413 DECLARE_AND_GET_RE_DEBUG_FLAGS;
13415 *flagp = WORST; /* Tentatively. */
13417 DEBUG_PARSE("atom");
13419 PERL_ARGS_ASSERT_REGATOM;
13422 parse_start = RExC_parse;
13423 assert(RExC_parse < RExC_end);
13424 switch ((U8)*RExC_parse) {
13426 RExC_seen_zerolen++;
13427 nextchar(pRExC_state);
13428 if (RExC_flags & RXf_PMf_MULTILINE)
13429 ret = reg_node(pRExC_state, MBOL);
13431 ret = reg_node(pRExC_state, SBOL);
13432 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13435 nextchar(pRExC_state);
13437 RExC_seen_zerolen++;
13438 if (RExC_flags & RXf_PMf_MULTILINE)
13439 ret = reg_node(pRExC_state, MEOL);
13441 ret = reg_node(pRExC_state, SEOL);
13442 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13445 nextchar(pRExC_state);
13446 if (RExC_flags & RXf_PMf_SINGLELINE)
13447 ret = reg_node(pRExC_state, SANY);
13449 ret = reg_node(pRExC_state, REG_ANY);
13450 *flagp |= HASWIDTH|SIMPLE;
13452 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13456 char * const oregcomp_parse = ++RExC_parse;
13457 ret = regclass(pRExC_state, flagp, depth+1,
13458 FALSE, /* means parse the whole char class */
13459 TRUE, /* allow multi-char folds */
13460 FALSE, /* don't silence non-portable warnings. */
13461 (bool) RExC_strict,
13462 TRUE, /* Allow an optimized regnode result */
13465 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13466 FAIL2("panic: regclass returned failure to regatom, flags=%#" UVxf,
13469 if (*RExC_parse != ']') {
13470 RExC_parse = oregcomp_parse;
13471 vFAIL("Unmatched [");
13473 nextchar(pRExC_state);
13474 Set_Node_Length(REGNODE_p(ret), RExC_parse - oregcomp_parse + 1); /* MJD */
13478 nextchar(pRExC_state);
13479 ret = reg(pRExC_state, 2, &flags, depth+1);
13481 if (flags & TRYAGAIN) {
13482 if (RExC_parse >= RExC_end) {
13483 /* Make parent create an empty node if needed. */
13484 *flagp |= TRYAGAIN;
13489 RETURN_FAIL_ON_RESTART(flags, flagp);
13490 FAIL2("panic: reg returned failure to regatom, flags=%#" UVxf,
13493 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
13497 if (flags & TRYAGAIN) {
13498 *flagp |= TRYAGAIN;
13501 vFAIL("Internal urp");
13502 /* Supposed to be caught earlier. */
13508 vFAIL("Quantifier follows nothing");
13513 This switch handles escape sequences that resolve to some kind
13514 of special regop and not to literal text. Escape sequences that
13515 resolve to literal text are handled below in the switch marked
13518 Every entry in this switch *must* have a corresponding entry
13519 in the literal escape switch. However, the opposite is not
13520 required, as the default for this switch is to jump to the
13521 literal text handling code.
13524 switch ((U8)*RExC_parse) {
13525 /* Special Escapes */
13527 RExC_seen_zerolen++;
13528 /* Under wildcards, this is changed to match \n; should be
13529 * invisible to the user, as they have to compile under /m */
13530 if (RExC_pm_flags & PMf_WILDCARD) {
13531 ret = reg_node(pRExC_state, MBOL);
13534 ret = reg_node(pRExC_state, SBOL);
13535 /* SBOL is shared with /^/ so we set the flags so we can tell
13536 * /\A/ from /^/ in split. */
13537 FLAGS(REGNODE_p(ret)) = 1;
13540 goto finish_meta_pat;
13542 if (RExC_pm_flags & PMf_WILDCARD) {
13544 /* diag_listed_as: Use of %s is not allowed in Unicode property
13545 wildcard subpatterns in regex; marked by <-- HERE in m/%s/
13547 vFAIL("Use of '\\G' is not allowed in Unicode property"
13548 " wildcard subpatterns");
13550 ret = reg_node(pRExC_state, GPOS);
13551 RExC_seen |= REG_GPOS_SEEN;
13553 goto finish_meta_pat;
13555 if (!RExC_in_lookbehind && !RExC_in_lookahead) {
13556 RExC_seen_zerolen++;
13557 ret = reg_node(pRExC_state, KEEPS);
13559 /* XXX:dmq : disabling in-place substitution seems to
13560 * be necessary here to avoid cases of memory corruption, as
13561 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
13563 RExC_seen |= REG_LOOKBEHIND_SEEN;
13564 goto finish_meta_pat;
13567 ++RExC_parse; /* advance past the 'K' */
13568 vFAIL("\\K not permitted in lookahead/lookbehind");
13571 if (RExC_pm_flags & PMf_WILDCARD) {
13572 /* See comment under \A above */
13573 ret = reg_node(pRExC_state, MEOL);
13576 ret = reg_node(pRExC_state, SEOL);
13579 RExC_seen_zerolen++; /* Do not optimize RE away */
13580 goto finish_meta_pat;
13582 if (RExC_pm_flags & PMf_WILDCARD) {
13583 /* See comment under \A above */
13584 ret = reg_node(pRExC_state, MEOL);
13587 ret = reg_node(pRExC_state, EOS);
13590 RExC_seen_zerolen++; /* Do not optimize RE away */
13591 goto finish_meta_pat;
13593 vFAIL("\\C no longer supported");
13595 ret = reg_node(pRExC_state, CLUMP);
13596 *flagp |= HASWIDTH;
13597 goto finish_meta_pat;
13605 regex_charset charset = get_regex_charset(RExC_flags);
13607 RExC_seen_zerolen++;
13608 RExC_seen |= REG_LOOKBEHIND_SEEN;
13609 op = BOUND + charset;
13611 if (RExC_parse >= RExC_end || *(RExC_parse + 1) != '{') {
13612 flags = TRADITIONAL_BOUND;
13613 if (op > BOUNDA) { /* /aa is same as /a */
13619 char name = *RExC_parse;
13620 char * endbrace = NULL;
13622 endbrace = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
13625 vFAIL2("Missing right brace on \\%c{}", name);
13627 /* XXX Need to decide whether to take spaces or not. Should be
13628 * consistent with \p{}, but that currently is SPACE, which
13629 * means vertical too, which seems wrong
13630 * while (isBLANK(*RExC_parse)) {
13633 if (endbrace == RExC_parse) {
13634 RExC_parse++; /* After the '}' */
13635 vFAIL2("Empty \\%c{}", name);
13637 length = endbrace - RExC_parse;
13638 /*while (isBLANK(*(RExC_parse + length - 1))) {
13641 switch (*RExC_parse) {
13644 && (memNEs(RExC_parse + 1, length - 1, "cb")))
13646 goto bad_bound_type;
13651 if (length != 2 || *(RExC_parse + 1) != 'b') {
13652 goto bad_bound_type;
13657 if (length != 2 || *(RExC_parse + 1) != 'b') {
13658 goto bad_bound_type;
13663 if (length != 2 || *(RExC_parse + 1) != 'b') {
13664 goto bad_bound_type;
13670 RExC_parse = endbrace;
13672 "'%" UTF8f "' is an unknown bound type",
13673 UTF8fARG(UTF, length, endbrace - length));
13674 NOT_REACHED; /*NOTREACHED*/
13676 RExC_parse = endbrace;
13677 REQUIRE_UNI_RULES(flagp, 0);
13682 else if (op >= BOUNDA) { /* /aa is same as /a */
13686 /* Don't have to worry about UTF-8, in this message because
13687 * to get here the contents of the \b must be ASCII */
13688 ckWARN4reg(RExC_parse + 1, /* Include the '}' in msg */
13689 "Using /u for '%.*s' instead of /%s",
13691 endbrace - length + 1,
13692 (charset == REGEX_ASCII_RESTRICTED_CHARSET)
13693 ? ASCII_RESTRICT_PAT_MODS
13694 : ASCII_MORE_RESTRICT_PAT_MODS);
13699 RExC_seen_d_op = TRUE;
13701 else if (op == BOUNDL) {
13702 RExC_contains_locale = 1;
13706 op += NBOUND - BOUND;
13709 ret = reg_node(pRExC_state, op);
13710 FLAGS(REGNODE_p(ret)) = flags;
13714 goto finish_meta_pat;
13718 ret = reg_node(pRExC_state, LNBREAK);
13719 *flagp |= HASWIDTH|SIMPLE;
13720 goto finish_meta_pat;
13734 /* These all have the same meaning inside [brackets], and it knows
13735 * how to do the best optimizations for them. So, pretend we found
13736 * these within brackets, and let it do the work */
13739 ret = regclass(pRExC_state, flagp, depth+1,
13740 TRUE, /* means just parse this element */
13741 FALSE, /* don't allow multi-char folds */
13742 FALSE, /* don't silence non-portable warnings. It
13743 would be a bug if these returned
13745 (bool) RExC_strict,
13746 TRUE, /* Allow an optimized regnode result */
13748 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13749 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
13750 * multi-char folds are allowed. */
13752 FAIL2("panic: regclass returned failure to regatom, flags=%#" UVxf,
13755 RExC_parse--; /* regclass() leaves this one too far ahead */
13758 /* The escapes above that don't take a parameter can't be
13759 * followed by a '{'. But 'pX', 'p{foo}' and
13760 * correspondingly 'P' can be */
13761 if ( RExC_parse - parse_start == 1
13762 && UCHARAT(RExC_parse + 1) == '{'
13763 && UNLIKELY(! new_regcurly(RExC_parse + 1, RExC_end)))
13766 vFAIL("Unescaped left brace in regex is illegal here");
13768 Set_Node_Offset(REGNODE_p(ret), parse_start);
13769 Set_Node_Length(REGNODE_p(ret), RExC_parse - parse_start + 1); /* MJD */
13770 nextchar(pRExC_state);
13773 /* Handle \N, \N{} and \N{NAMED SEQUENCE} (the latter meaning the
13774 * \N{...} evaluates to a sequence of more than one code points).
13775 * The function call below returns a regnode, which is our result.
13776 * The parameters cause it to fail if the \N{} evaluates to a
13777 * single code point; we handle those like any other literal. The
13778 * reason that the multicharacter case is handled here and not as
13779 * part of the EXACtish code is because of quantifiers. In
13780 * /\N{BLAH}+/, the '+' applies to the whole thing, and doing it
13781 * this way makes that Just Happen. dmq.
13782 * join_exact() will join this up with adjacent EXACTish nodes
13783 * later on, if appropriate. */
13785 if (grok_bslash_N(pRExC_state,
13786 &ret, /* Want a regnode returned */
13787 NULL, /* Fail if evaluates to a single code
13789 NULL, /* Don't need a count of how many code
13798 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13800 /* Here, evaluates to a single code point. Go get that */
13801 RExC_parse = parse_start;
13804 case 'k': /* Handle \k<NAME> and \k'NAME' */
13808 if ( RExC_parse >= RExC_end - 1
13809 || (( ch = RExC_parse[1]) != '<'
13814 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
13815 vFAIL2("Sequence %.2s... not terminated", parse_start);
13818 ret = handle_named_backref(pRExC_state,
13830 case '1': case '2': case '3': case '4':
13831 case '5': case '6': case '7': case '8': case '9':
13836 if (*RExC_parse == 'g') {
13840 if (*RExC_parse == '{') {
13844 if (*RExC_parse == '-') {
13848 if (hasbrace && !isDIGIT(*RExC_parse)) {
13849 if (isrel) RExC_parse--;
13851 goto parse_named_seq;
13854 if (RExC_parse >= RExC_end) {
13855 goto unterminated_g;
13857 num = S_backref_value(RExC_parse, RExC_end);
13859 vFAIL("Reference to invalid group 0");
13860 else if (num == I32_MAX) {
13861 if (isDIGIT(*RExC_parse))
13862 vFAIL("Reference to nonexistent group");
13865 vFAIL("Unterminated \\g... pattern");
13869 num = RExC_npar - num;
13871 vFAIL("Reference to nonexistent or unclosed group");
13875 num = S_backref_value(RExC_parse, RExC_end);
13876 /* bare \NNN might be backref or octal - if it is larger
13877 * than or equal RExC_npar then it is assumed to be an
13878 * octal escape. Note RExC_npar is +1 from the actual
13879 * number of parens. */
13880 /* Note we do NOT check if num == I32_MAX here, as that is
13881 * handled by the RExC_npar check */
13884 /* any numeric escape < 10 is always a backref */
13886 /* any numeric escape < RExC_npar is a backref */
13887 && num >= RExC_npar
13888 /* cannot be an octal escape if it starts with 8 */
13889 && *RExC_parse != '8'
13890 /* cannot be an octal escape if it starts with 9 */
13891 && *RExC_parse != '9'
13893 /* Probably not meant to be a backref, instead likely
13894 * to be an octal character escape, e.g. \35 or \777.
13895 * The above logic should make it obvious why using
13896 * octal escapes in patterns is problematic. - Yves */
13897 RExC_parse = parse_start;
13902 /* At this point RExC_parse points at a numeric escape like
13903 * \12 or \88 or something similar, which we should NOT treat
13904 * as an octal escape. It may or may not be a valid backref
13905 * escape. For instance \88888888 is unlikely to be a valid
13907 while (isDIGIT(*RExC_parse))
13910 if (*RExC_parse != '}')
13911 vFAIL("Unterminated \\g{...} pattern");
13914 if (num >= (I32)RExC_npar) {
13916 /* It might be a forward reference; we can't fail until we
13917 * know, by completing the parse to get all the groups, and
13918 * then reparsing */
13919 if (ALL_PARENS_COUNTED) {
13920 if (num >= RExC_total_parens) {
13921 vFAIL("Reference to nonexistent group");
13925 REQUIRE_PARENS_PASS;
13929 ret = reganode(pRExC_state,
13932 : (ASCII_FOLD_RESTRICTED)
13934 : (AT_LEAST_UNI_SEMANTICS)
13940 if (OP(REGNODE_p(ret)) == REFF) {
13941 RExC_seen_d_op = TRUE;
13943 *flagp |= HASWIDTH;
13945 /* override incorrect value set in reganode MJD */
13946 Set_Node_Offset(REGNODE_p(ret), parse_start);
13947 Set_Node_Cur_Length(REGNODE_p(ret), parse_start-1);
13948 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13949 FALSE /* Don't force to /x */ );
13953 if (RExC_parse >= RExC_end)
13954 FAIL("Trailing \\");
13957 /* Do not generate "unrecognized" warnings here, we fall
13958 back into the quick-grab loop below */
13959 RExC_parse = parse_start;
13961 } /* end of switch on a \foo sequence */
13966 /* '#' comments should have been spaced over before this function was
13968 assert((RExC_flags & RXf_PMf_EXTENDED) == 0);
13970 if (RExC_flags & RXf_PMf_EXTENDED) {
13971 RExC_parse = reg_skipcomment( pRExC_state, RExC_parse );
13972 if (RExC_parse < RExC_end)
13982 /* Here, we have determined that the next thing is probably a
13983 * literal character. RExC_parse points to the first byte of its
13984 * definition. (It still may be an escape sequence that evaluates
13985 * to a single character) */
13990 char *s, *old_s = NULL, *old_old_s = NULL;
13992 U32 max_string_len = 255;
13994 /* We may have to reparse the node, artificially stopping filling
13995 * it early, based on info gleaned in the first parse. This
13996 * variable gives where we stop. Make it above the normal stopping
13997 * place first time through; otherwise it would stop too early */
13998 U32 upper_fill = max_string_len + 1;
14000 /* We start out as an EXACT node, even if under /i, until we find a
14001 * character which is in a fold. The algorithm now segregates into
14002 * separate nodes, characters that fold from those that don't under
14003 * /i. (This hopefully will create nodes that are fixed strings
14004 * even under /i, giving the optimizer something to grab on to.)
14005 * So, if a node has something in it and the next character is in
14006 * the opposite category, that node is closed up, and the function
14007 * returns. Then regatom is called again, and a new node is
14008 * created for the new category. */
14009 U8 node_type = EXACT;
14011 /* Assume the node will be fully used; the excess is given back at
14012 * the end. Under /i, we may need to temporarily add the fold of
14013 * an extra character or two at the end to check for splitting
14014 * multi-char folds, so allocate extra space for that. We can't
14015 * make any other length assumptions, as a byte input sequence
14016 * could shrink down. */
14017 Ptrdiff_t current_string_nodes = STR_SZ(max_string_len
14021 ? UTF8_MAXBYTES_CASE
14022 /* Max non-UTF-8 expansion is 2 */ : 2)));
14024 bool next_is_quantifier;
14025 char * oldp = NULL;
14027 /* We can convert EXACTF nodes to EXACTFU if they contain only
14028 * characters that match identically regardless of the target
14029 * string's UTF8ness. The reason to do this is that EXACTF is not
14030 * trie-able, EXACTFU is, and EXACTFU requires fewer operations at
14033 * Similarly, we can convert EXACTFL nodes to EXACTFLU8 if they
14034 * contain only above-Latin1 characters (hence must be in UTF8),
14035 * which don't participate in folds with Latin1-range characters,
14036 * as the latter's folds aren't known until runtime. */
14037 bool maybe_exactfu = FOLD && (DEPENDS_SEMANTICS || LOC);
14039 /* Single-character EXACTish nodes are almost always SIMPLE. This
14040 * allows us to override this as encountered */
14041 U8 maybe_SIMPLE = SIMPLE;
14043 /* Does this node contain something that can't match unless the
14044 * target string is (also) in UTF-8 */
14045 bool requires_utf8_target = FALSE;
14047 /* The sequence 'ss' is problematic in non-UTF-8 patterns. */
14048 bool has_ss = FALSE;
14050 /* So is the MICRO SIGN */
14051 bool has_micro_sign = FALSE;
14053 /* Set when we fill up the current node and there is still more
14054 * text to process */
14057 /* Allocate an EXACT node. The node_type may change below to
14058 * another EXACTish node, but since the size of the node doesn't
14059 * change, it works */
14060 ret = regnode_guts(pRExC_state, node_type, current_string_nodes,
14062 FILL_NODE(ret, node_type);
14065 s = STRING(REGNODE_p(ret));
14076 maybe_exactfu = FOLD && (DEPENDS_SEMANTICS || LOC);
14077 maybe_SIMPLE = SIMPLE;
14078 requires_utf8_target = FALSE;
14080 has_micro_sign = FALSE;
14084 /* This breaks under rare circumstances. If folding, we do not
14085 * want to split a node at a character that is a non-final in a
14086 * multi-char fold, as an input string could just happen to want to
14087 * match across the node boundary. The code at the end of the loop
14088 * looks for this, and backs off until it finds not such a
14089 * character, but it is possible (though extremely, extremely
14090 * unlikely) for all characters in the node to be non-final fold
14091 * ones, in which case we just leave the node fully filled, and
14092 * hope that it doesn't match the string in just the wrong place */
14094 assert( ! UTF /* Is at the beginning of a character */
14095 || UTF8_IS_INVARIANT(UCHARAT(RExC_parse))
14096 || UTF8_IS_START(UCHARAT(RExC_parse)));
14098 overflowed = FALSE;
14100 /* Here, we have a literal character. Find the maximal string of
14101 * them in the input that we can fit into a single EXACTish node.
14102 * We quit at the first non-literal or when the node gets full, or
14103 * under /i the categorization of folding/non-folding character
14105 while (p < RExC_end && len < upper_fill) {
14107 /* In most cases each iteration adds one byte to the output.
14108 * The exceptions override this */
14109 Size_t added_len = 1;
14115 /* White space has already been ignored */
14116 assert( (RExC_flags & RXf_PMf_EXTENDED) == 0
14117 || ! is_PATWS_safe((p), RExC_end, UTF));
14120 const char* message;
14133 /* Literal Escapes Switch
14135 This switch is meant to handle escape sequences that
14136 resolve to a literal character.
14138 Every escape sequence that represents something
14139 else, like an assertion or a char class, is handled
14140 in the switch marked 'Special Escapes' above in this
14141 routine, but also has an entry here as anything that
14142 isn't explicitly mentioned here will be treated as
14143 an unescaped equivalent literal.
14146 switch ((U8)*++p) {
14148 /* These are all the special escapes. */
14149 case 'A': /* Start assertion */
14150 case 'b': case 'B': /* Word-boundary assertion*/
14151 case 'C': /* Single char !DANGEROUS! */
14152 case 'd': case 'D': /* digit class */
14153 case 'g': case 'G': /* generic-backref, pos assertion */
14154 case 'h': case 'H': /* HORIZWS */
14155 case 'k': case 'K': /* named backref, keep marker */
14156 case 'p': case 'P': /* Unicode property */
14157 case 'R': /* LNBREAK */
14158 case 's': case 'S': /* space class */
14159 case 'v': case 'V': /* VERTWS */
14160 case 'w': case 'W': /* word class */
14161 case 'X': /* eXtended Unicode "combining
14162 character sequence" */
14163 case 'z': case 'Z': /* End of line/string assertion */
14167 /* Anything after here is an escape that resolves to a
14168 literal. (Except digits, which may or may not)
14174 case 'N': /* Handle a single-code point named character. */
14175 RExC_parse = p + 1;
14176 if (! grok_bslash_N(pRExC_state,
14177 NULL, /* Fail if evaluates to
14178 anything other than a
14179 single code point */
14180 &ender, /* The returned single code
14182 NULL, /* Don't need a count of
14183 how many code points */
14188 if (*flagp & NEED_UTF8)
14189 FAIL("panic: grok_bslash_N set NEED_UTF8");
14190 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
14192 /* Here, it wasn't a single code point. Go close
14193 * up this EXACTish node. The switch() prior to
14194 * this switch handles the other cases */
14195 RExC_parse = p = oldp;
14199 RExC_parse = parse_start;
14201 /* The \N{} means the pattern, if previously /d,
14202 * becomes /u. That means it can't be an EXACTF node,
14203 * but an EXACTFU */
14204 if (node_type == EXACTF) {
14205 node_type = EXACTFU;
14207 /* If the node already contains something that
14208 * differs between EXACTF and EXACTFU, reparse it
14210 if (! maybe_exactfu) {
14231 ender = ESC_NATIVE;
14239 if (! grok_bslash_o(&p,
14244 (bool) RExC_strict,
14245 FALSE, /* No illegal cp's */
14248 RExC_parse = p; /* going to die anyway; point to
14249 exact spot of failure */
14253 if (message && TO_OUTPUT_WARNINGS(p)) {
14254 warn_non_literal_string(p, packed_warn, message);
14258 if (! grok_bslash_x(&p,
14263 (bool) RExC_strict,
14264 FALSE, /* No illegal cp's */
14267 RExC_parse = p; /* going to die anyway; point
14268 to exact spot of failure */
14272 if (message && TO_OUTPUT_WARNINGS(p)) {
14273 warn_non_literal_string(p, packed_warn, message);
14277 if (ender < 0x100) {
14278 if (RExC_recode_x_to_native) {
14279 ender = LATIN1_TO_NATIVE(ender);
14286 if (! grok_bslash_c(*p, &grok_c_char,
14287 &message, &packed_warn))
14289 /* going to die anyway; point to exact spot of
14291 RExC_parse = p + ((UTF)
14292 ? UTF8_SAFE_SKIP(p, RExC_end)
14297 ender = grok_c_char;
14299 if (message && TO_OUTPUT_WARNINGS(p)) {
14300 warn_non_literal_string(p, packed_warn, message);
14304 case '8': case '9': /* must be a backreference */
14306 /* we have an escape like \8 which cannot be an octal escape
14307 * so we exit the loop, and let the outer loop handle this
14308 * escape which may or may not be a legitimate backref. */
14310 case '1': case '2': case '3':case '4':
14311 case '5': case '6': case '7':
14312 /* When we parse backslash escapes there is ambiguity
14313 * between backreferences and octal escapes. Any escape
14314 * from \1 - \9 is a backreference, any multi-digit
14315 * escape which does not start with 0 and which when
14316 * evaluated as decimal could refer to an already
14317 * parsed capture buffer is a back reference. Anything
14320 * Note this implies that \118 could be interpreted as
14321 * 118 OR as "\11" . "8" depending on whether there
14322 * were 118 capture buffers defined already in the
14325 /* NOTE, RExC_npar is 1 more than the actual number of
14326 * parens we have seen so far, hence the "<" as opposed
14328 if ( !isDIGIT(p[1]) || S_backref_value(p, RExC_end) < RExC_npar)
14329 { /* Not to be treated as an octal constant, go
14337 I32 flags = PERL_SCAN_SILENT_ILLDIGIT
14338 | PERL_SCAN_NOTIFY_ILLDIGIT;
14340 ender = grok_oct(p, &numlen, &flags, NULL);
14342 if ( (flags & PERL_SCAN_NOTIFY_ILLDIGIT)
14343 && isDIGIT(*p) /* like \08, \178 */
14344 && ckWARN(WARN_REGEXP))
14346 reg_warn_non_literal_string(
14348 form_alien_digit_msg(8, numlen, p,
14349 RExC_end, UTF, FALSE));
14355 FAIL("Trailing \\");
14358 if (isALPHANUMERIC(*p)) {
14359 /* An alpha followed by '{' is going to fail next
14360 * iteration, so don't output this warning in that
14362 if (! isALPHA(*p) || *(p + 1) != '{') {
14363 ckWARN2reg(p + 1, "Unrecognized escape \\%.1s"
14364 " passed through", p);
14367 goto normal_default;
14368 } /* End of switch on '\' */
14371 /* Trying to gain new uses for '{' without breaking too
14372 * much existing code is hard. The solution currently
14374 * 1) If there is no ambiguity that a '{' should always
14375 * be taken literally, at the start of a construct, we
14377 * 2) If the literal '{' conflicts with our desired use
14378 * of it as a metacharacter, we die. The deprecation
14379 * cycles for this have come and gone.
14380 * 3) If there is ambiguity, we raise a simple warning.
14381 * This could happen, for example, if the user
14382 * intended it to introduce a quantifier, but slightly
14383 * misspelled the quantifier. Without this warning,
14384 * the quantifier would silently be taken as a literal
14385 * string of characters instead of a meta construct */
14386 if (len || (p > RExC_start && isALPHA_A(*(p - 1)))) {
14388 || ( p > parse_start + 1
14389 && isALPHA_A(*(p - 1))
14390 && *(p - 2) == '\\')
14391 || new_regcurly(p, RExC_end))
14393 RExC_parse = p + 1;
14394 vFAIL("Unescaped left brace in regex is "
14397 ckWARNreg(p + 1, "Unescaped left brace in regex is"
14398 " passed through");
14400 goto normal_default;
14403 if (p > RExC_parse && RExC_strict) {
14404 ckWARN2reg(p + 1, "Unescaped literal '%c'", *p);
14407 default: /* A literal character */
14409 if (! UTF8_IS_INVARIANT(*p) && UTF) {
14411 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
14412 &numlen, UTF8_ALLOW_DEFAULT);
14418 } /* End of switch on the literal */
14420 /* Here, have looked at the literal character, and <ender>
14421 * contains its ordinal; <p> points to the character after it.
14425 REQUIRE_UTF8(flagp);
14426 if ( UNICODE_IS_PERL_EXTENDED(ender)
14427 && TO_OUTPUT_WARNINGS(p))
14429 ckWARN2_non_literal_string(p,
14430 packWARN(WARN_PORTABLE),
14431 PL_extended_cp_format,
14436 /* We need to check if the next non-ignored thing is a
14437 * quantifier. Move <p> to after anything that should be
14438 * ignored, which, as a side effect, positions <p> for the next
14439 * loop iteration */
14440 skip_to_be_ignored_text(pRExC_state, &p,
14441 FALSE /* Don't force to /x */ );
14443 /* If the next thing is a quantifier, it applies to this
14444 * character only, which means that this character has to be in
14445 * its own node and can't just be appended to the string in an
14446 * existing node, so if there are already other characters in
14447 * the node, close the node with just them, and set up to do
14448 * this character again next time through, when it will be the
14449 * only thing in its new node */
14451 next_is_quantifier = LIKELY(p < RExC_end)
14452 && UNLIKELY(ISMULT2(p));
14454 if (next_is_quantifier && LIKELY(len)) {
14459 /* Ready to add 'ender' to the node */
14461 if (! FOLD) { /* The simple case, just append the literal */
14464 /* Don't output if it would overflow */
14465 if (UNLIKELY(len > max_string_len - ((UTF)
14466 ? UVCHR_SKIP(ender)
14473 if (UVCHR_IS_INVARIANT(ender) || ! UTF) {
14474 *(s++) = (char) ender;
14477 U8 * new_s = uvchr_to_utf8((U8*)s, ender);
14478 added_len = (char *) new_s - s;
14479 s = (char *) new_s;
14482 requires_utf8_target = TRUE;
14486 else if (LOC && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)) {
14488 /* Here are folding under /l, and the code point is
14489 * problematic. If this is the first character in the
14490 * node, change the node type to folding. Otherwise, if
14491 * this is the first problematic character, close up the
14492 * existing node, so can start a new node with this one */
14494 node_type = EXACTFL;
14495 RExC_contains_locale = 1;
14497 else if (node_type == EXACT) {
14502 /* This problematic code point means we can't simplify
14504 maybe_exactfu = FALSE;
14506 /* Here, we are adding a problematic fold character.
14507 * "Problematic" in this context means that its fold isn't
14508 * known until runtime. (The non-problematic code points
14509 * are the above-Latin1 ones that fold to also all
14510 * above-Latin1. Their folds don't vary no matter what the
14511 * locale is.) But here we have characters whose fold
14512 * depends on the locale. We just add in the unfolded
14513 * character, and wait until runtime to fold it */
14514 goto not_fold_common;
14516 else /* regular fold; see if actually is in a fold */
14517 if ( (ender < 256 && ! IS_IN_SOME_FOLD_L1(ender))
14519 && ! _invlist_contains_cp(PL_in_some_fold, ender)))
14521 /* Here, folding, but the character isn't in a fold.
14523 * Start a new node if previous characters in the node were
14525 if (len && node_type != EXACT) {
14530 /* Here, continuing a node with non-folded characters. Add
14532 goto not_fold_common;
14534 else { /* Here, does participate in some fold */
14536 /* If this is the first character in the node, change its
14537 * type to folding. Otherwise, if this is the first
14538 * folding character in the node, close up the existing
14539 * node, so can start a new node with this one. */
14541 node_type = compute_EXACTish(pRExC_state);
14543 else if (node_type == EXACT) {
14548 if (UTF) { /* Alway use the folded value for UTF-8
14550 if (UVCHR_IS_INVARIANT(ender)) {
14551 if (UNLIKELY(len + 1 > max_string_len)) {
14556 *(s)++ = (U8) toFOLD(ender);
14559 UV folded = _to_uni_fold_flags(
14561 (U8 *) s, /* We have allocated extra space
14562 in 's' so can't run off the
14565 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
14566 ? FOLD_FLAGS_NOMIX_ASCII
14568 if (UNLIKELY(len + added_len > max_string_len)) {
14576 && LIKELY(folded != GREEK_SMALL_LETTER_MU))
14578 /* U+B5 folds to the MU, so its possible for a
14579 * non-UTF-8 target to match it */
14580 requires_utf8_target = TRUE;
14584 else { /* Here is non-UTF8. */
14586 /* The fold will be one or (rarely) two characters.
14587 * Check that there's room for at least a single one
14588 * before setting any flags, etc. Because otherwise an
14589 * overflowing character could cause a flag to be set
14590 * even though it doesn't end up in this node. (For
14591 * the two character fold, we check again, before
14592 * setting any flags) */
14593 if (UNLIKELY(len + 1 > max_string_len)) {
14598 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
14599 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
14600 || UNICODE_DOT_DOT_VERSION > 0)
14602 /* On non-ancient Unicodes, check for the only possible
14603 * multi-char fold */
14604 if (UNLIKELY(ender == LATIN_SMALL_LETTER_SHARP_S)) {
14606 /* This potential multi-char fold means the node
14607 * can't be simple (because it could match more
14608 * than a single char). And in some cases it will
14609 * match 'ss', so set that flag */
14613 /* It can't change to be an EXACTFU (unless already
14614 * is one). We fold it iff under /u rules. */
14615 if (node_type != EXACTFU) {
14616 maybe_exactfu = FALSE;
14619 if (UNLIKELY(len + 2 > max_string_len)) {
14628 goto done_with_this_char;
14631 else if ( UNLIKELY(isALPHA_FOLD_EQ(ender, 's'))
14633 && UNLIKELY(isALPHA_FOLD_EQ(*(s-1), 's')))
14635 /* Also, the sequence 'ss' is special when not
14636 * under /u. If the target string is UTF-8, it
14637 * should match SHARP S; otherwise it won't. So,
14638 * here we have to exclude the possibility of this
14639 * node moving to /u.*/
14641 maybe_exactfu = FALSE;
14644 /* Here, the fold will be a single character */
14646 if (UNLIKELY(ender == MICRO_SIGN)) {
14647 has_micro_sign = TRUE;
14649 else if (PL_fold[ender] != PL_fold_latin1[ender]) {
14651 /* If the character's fold differs between /d and
14652 * /u, this can't change to be an EXACTFU node */
14653 maybe_exactfu = FALSE;
14656 *(s++) = (DEPENDS_SEMANTICS)
14657 ? (char) toFOLD(ender)
14659 /* Under /u, the fold of any character in
14660 * the 0-255 range happens to be its
14661 * lowercase equivalent, except for LATIN
14662 * SMALL LETTER SHARP S, which was handled
14663 * above, and the MICRO SIGN, whose fold
14664 * requires UTF-8 to represent. */
14665 : (char) toLOWER_L1(ender);
14667 } /* End of adding current character to the node */
14669 done_with_this_char:
14673 if (next_is_quantifier) {
14675 /* Here, the next input is a quantifier, and to get here,
14676 * the current character is the only one in the node. */
14680 } /* End of loop through literal characters */
14682 /* Here we have either exhausted the input or run out of room in
14683 * the node. If the former, we are done. (If we encountered a
14684 * character that can't be in the node, transfer is made directly
14685 * to <loopdone>, and so we wouldn't have fallen off the end of the
14687 if (LIKELY(! overflowed)) {
14691 /* Here we have run out of room. We can grow plain EXACT and
14692 * LEXACT nodes. If the pattern is gigantic enough, though,
14693 * eventually we'll have to artificially chunk the pattern into
14694 * multiple nodes. */
14695 if (! LOC && (node_type == EXACT || node_type == LEXACT)) {
14696 Size_t overhead = 1 + regarglen[OP(REGNODE_p(ret))];
14697 Size_t overhead_expansion = 0;
14699 Size_t max_nodes_for_string;
14703 /* Here we couldn't fit the final character in the current
14704 * node, so it will have to be reparsed, no matter what else we
14708 /* If would have overflowed a regular EXACT node, switch
14709 * instead to an LEXACT. The code below is structured so that
14710 * the actual growing code is common to changing from an EXACT
14711 * or just increasing the LEXACT size. This means that we have
14712 * to save the string in the EXACT case before growing, and
14713 * then copy it afterwards to its new location */
14714 if (node_type == EXACT) {
14715 overhead_expansion = regarglen[LEXACT] - regarglen[EXACT];
14716 RExC_emit += overhead_expansion;
14717 Copy(s0, temp, len, char);
14720 /* Ready to grow. If it was a plain EXACT, the string was
14721 * saved, and the first few bytes of it overwritten by adding
14722 * an argument field. We assume, as we do elsewhere in this
14723 * file, that one byte of remaining input will translate into
14724 * one byte of output, and if that's too small, we grow again,
14725 * if too large the excess memory is freed at the end */
14727 max_nodes_for_string = U16_MAX - overhead - overhead_expansion;
14728 achievable = MIN(max_nodes_for_string,
14729 current_string_nodes + STR_SZ(RExC_end - p));
14730 delta = achievable - current_string_nodes;
14732 /* If there is just no more room, go finish up this chunk of
14738 change_engine_size(pRExC_state, delta + overhead_expansion);
14739 current_string_nodes += delta;
14741 = sizeof(struct regnode) * current_string_nodes;
14742 upper_fill = max_string_len + 1;
14744 /* If the length was small, we know this was originally an
14745 * EXACT node now converted to LEXACT, and the string has to be
14746 * restored. Otherwise the string was untouched. 260 is just
14747 * a number safely above 255 so don't have to worry about
14748 * getting it precise */
14750 node_type = LEXACT;
14751 FILL_NODE(ret, node_type);
14752 s0 = STRING(REGNODE_p(ret));
14753 Copy(temp, s0, len, char);
14757 goto continue_parse;
14760 bool splittable = FALSE;
14761 bool backed_up = FALSE;
14762 char * e; /* should this be U8? */
14763 char * s_start; /* should this be U8? */
14765 /* Here is /i. Running out of room creates a problem if we are
14766 * folding, and the split happens in the middle of a
14767 * multi-character fold, as a match that should have occurred,
14768 * won't, due to the way nodes are matched, and our artificial
14769 * boundary. So back off until we aren't splitting such a
14770 * fold. If there is no such place to back off to, we end up
14771 * taking the entire node as-is. This can happen if the node
14772 * consists entirely of 'f' or entirely of 's' characters (or
14773 * things that fold to them) as 'ff' and 'ss' are
14774 * multi-character folds.
14776 * The Unicode standard says that multi character folds consist
14777 * of either two or three characters. That means we would be
14778 * splitting one if the final character in the node is at the
14779 * beginning of either type, or is the second of a three
14783 * ender is the code point of the character that won't fit
14785 * s points to just beyond the final byte in the node.
14786 * It's where we would place ender if there were
14787 * room, and where in fact we do place ender's fold
14788 * in the code below, as we've over-allocated space
14789 * for s0 (hence s) to allow for this
14790 * e starts at 's' and advances as we append things.
14791 * old_s is the same as 's'. (If ender had fit, 's' would
14792 * have been advanced to beyond it).
14793 * old_old_s points to the beginning byte of the final
14794 * character in the node
14795 * p points to the beginning byte in the input of the
14796 * character beyond 'ender'.
14797 * oldp points to the beginning byte in the input of
14800 * In the case of /il, we haven't folded anything that could be
14801 * affected by the locale. That means only above-Latin1
14802 * characters that fold to other above-latin1 characters get
14803 * folded at compile time. To check where a good place to
14804 * split nodes is, everything in it will have to be folded.
14805 * The boolean 'maybe_exactfu' keeps track in /il if there are
14806 * any unfolded characters in the node. */
14807 bool need_to_fold_loc = LOC && ! maybe_exactfu;
14809 /* If we do need to fold the node, we need a place to store the
14810 * folded copy, and a way to map back to the unfolded original
14812 char * locfold_buf = NULL;
14813 Size_t * loc_correspondence = NULL;
14815 if (! need_to_fold_loc) { /* The normal case. Just
14816 initialize to the actual node */
14819 s = old_old_s; /* Point to the beginning of the final char
14820 that fits in the node */
14824 /* Here, we have filled a /il node, and there are unfolded
14825 * characters in it. If the runtime locale turns out to be
14826 * UTF-8, there are possible multi-character folds, just
14827 * like when not under /l. The node hence can't terminate
14828 * in the middle of such a fold. To determine this, we
14829 * have to create a folded copy of this node. That means
14830 * reparsing the node, folding everything assuming a UTF-8
14831 * locale. (If at runtime it isn't such a locale, the
14832 * actions here wouldn't have been necessary, but we have
14833 * to assume the worst case.) If we find we need to back
14834 * off the folded string, we do so, and then map that
14835 * position back to the original unfolded node, which then
14836 * gets output, truncated at that spot */
14838 char * redo_p = RExC_parse;
14842 /* Allow enough space assuming a single byte input folds to
14843 * a single byte output, plus assume that the two unparsed
14844 * characters (that we may need) fold to the largest number
14845 * of bytes possible, plus extra for one more worst case
14846 * scenario. In the loop below, if we start eating into
14847 * that final spare space, we enlarge this initial space */
14848 Size_t size = max_string_len + (3 * UTF8_MAXBYTES_CASE) + 1;
14850 Newxz(locfold_buf, size, char);
14851 Newxz(loc_correspondence, size, Size_t);
14853 /* Redo this node's parse, folding into 'locfold_buf' */
14854 redo_p = RExC_parse;
14855 old_redo_e = redo_e = locfold_buf;
14856 while (redo_p <= oldp) {
14858 old_redo_e = redo_e;
14859 loc_correspondence[redo_e - locfold_buf]
14860 = redo_p - RExC_parse;
14865 (void) _to_utf8_fold_flags((U8 *) redo_p,
14870 redo_e += added_len;
14871 redo_p += UTF8SKIP(redo_p);
14875 /* Note that if this code is run on some ancient
14876 * Unicode versions, SHARP S doesn't fold to 'ss',
14877 * but rather than clutter the code with #ifdef's,
14878 * as is done above, we ignore that possibility.
14879 * This is ok because this code doesn't affect what
14880 * gets matched, but merely where the node gets
14882 if (UCHARAT(redo_p) != LATIN_SMALL_LETTER_SHARP_S) {
14883 *redo_e++ = toLOWER_L1(UCHARAT(redo_p));
14893 /* If we're getting so close to the end that a
14894 * worst-case fold in the next character would cause us
14895 * to overflow, increase, assuming one byte output byte
14896 * per one byte input one, plus room for another worst
14898 if ( redo_p <= oldp
14899 && redo_e > locfold_buf + size
14900 - (UTF8_MAXBYTES_CASE + 1))
14902 Size_t new_size = size
14904 + UTF8_MAXBYTES_CASE + 1;
14905 Ptrdiff_t e_offset = redo_e - locfold_buf;
14907 Renew(locfold_buf, new_size, char);
14908 Renew(loc_correspondence, new_size, Size_t);
14911 redo_e = locfold_buf + e_offset;
14915 /* Set so that things are in terms of the folded, temporary
14918 s_start = locfold_buf;
14923 /* Here, we have 's', 's_start' and 'e' set up to point to the
14924 * input that goes into the node, folded.
14926 * If the final character of the node and the fold of ender
14927 * form the first two characters of a three character fold, we
14928 * need to peek ahead at the next (unparsed) character in the
14929 * input to determine if the three actually do form such a
14930 * fold. Just looking at that character is not generally
14931 * sufficient, as it could be, for example, an escape sequence
14932 * that evaluates to something else, and it needs to be folded.
14934 * khw originally thought to just go through the parse loop one
14935 * extra time, but that doesn't work easily as that iteration
14936 * could cause things to think that the parse is over and to
14937 * goto loopdone. The character could be a '$' for example, or
14938 * the character beyond could be a quantifier, and other
14939 * glitches as well.
14941 * The solution used here for peeking ahead is to look at that
14942 * next character. If it isn't ASCII punctuation, then it will
14943 * be something that continues in an EXACTish node if there
14944 * were space. We append the fold of it to s, having reserved
14945 * enough room in s0 for the purpose. If we can't reasonably
14946 * peek ahead, we instead assume the worst case: that it is
14947 * something that would form the completion of a multi-char
14950 * If we can't split between s and ender, we work backwards
14951 * character-by-character down to s0. At each current point
14952 * see if we are at the beginning of a multi-char fold. If so,
14953 * that means we would be splitting the fold across nodes, and
14954 * so we back up one and try again.
14956 * If we're not at the beginning, we still could be at the
14957 * final two characters of a (rare) three character fold. We
14958 * check if the sequence starting at the character before the
14959 * current position (and including the current and next
14960 * characters) is a three character fold. If not, the node can
14961 * be split here. If it is, we have to backup two characters
14964 * Otherwise, the node can be split at the current position.
14966 * The same logic is used for UTF-8 patterns and not */
14970 /* Append the fold of ender */
14971 (void) _to_uni_fold_flags(
14975 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
14976 ? FOLD_FLAGS_NOMIX_ASCII
14980 /* 's' and the character folded to by ender may be the
14981 * first two of a three-character fold, in which case the
14982 * node should not be split here. That may mean examining
14983 * the so-far unparsed character starting at 'p'. But if
14984 * ender folded to more than one character, we already have
14985 * three characters to look at. Also, we first check if
14986 * the sequence consisting of s and the next character form
14987 * the first two of some three character fold. If not,
14988 * there's no need to peek ahead. */
14989 if ( added_len <= UTF8SKIP(e - added_len)
14990 && UNLIKELY(is_THREE_CHAR_FOLD_HEAD_utf8_safe(s, e)))
14992 /* Here, the two do form the beginning of a potential
14993 * three character fold. The unexamined character may
14994 * or may not complete it. Peek at it. It might be
14995 * something that ends the node or an escape sequence,
14996 * in which case we don't know without a lot of work
14997 * what it evaluates to, so we have to assume the worst
14998 * case: that it does complete the fold, and so we
14999 * can't split here. All such instances will have
15000 * that character be an ASCII punctuation character,
15001 * like a backslash. So, for that case, backup one and
15002 * drop down to try at that position */
15004 s = (char *) utf8_hop_back((U8 *) s, -1,
15009 /* Here, since it's not punctuation, it must be a
15010 * real character, and we can append its fold to
15011 * 'e' (having deliberately reserved enough space
15012 * for this eventuality) and drop down to check if
15013 * the three actually do form a folded sequence */
15014 (void) _to_utf8_fold_flags(
15015 (U8 *) p, (U8 *) RExC_end,
15018 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
15019 ? FOLD_FLAGS_NOMIX_ASCII
15025 /* Here, we either have three characters available in
15026 * sequence starting at 's', or we have two characters and
15027 * know that the following one can't possibly be part of a
15028 * three character fold. We go through the node backwards
15029 * until we find a place where we can split it without
15030 * breaking apart a multi-character fold. At any given
15031 * point we have to worry about if such a fold begins at
15032 * the current 's', and also if a three-character fold
15033 * begins at s-1, (containing s and s+1). Splitting in
15034 * either case would break apart a fold */
15036 char *prev_s = (char *) utf8_hop_back((U8 *) s, -1,
15039 /* If is a multi-char fold, can't split here. Backup
15040 * one char and try again */
15041 if (UNLIKELY(is_MULTI_CHAR_FOLD_utf8_safe(s, e))) {
15047 /* If the two characters beginning at 's' are part of a
15048 * three character fold starting at the character
15049 * before s, we can't split either before or after s.
15050 * Backup two chars and try again */
15051 if ( LIKELY(s > s_start)
15052 && UNLIKELY(is_THREE_CHAR_FOLD_utf8_safe(prev_s, e)))
15055 s = (char *) utf8_hop_back((U8 *) s, -1, (U8 *) s_start);
15060 /* Here there's no multi-char fold between s and the
15061 * next character following it. We can split */
15065 } while (s > s_start); /* End of loops backing up through the node */
15067 /* Here we either couldn't find a place to split the node,
15068 * or else we broke out of the loop setting 'splittable' to
15069 * true. In the latter case, the place to split is between
15070 * the first and second characters in the sequence starting
15076 else { /* Pattern not UTF-8 */
15077 if ( ender != LATIN_SMALL_LETTER_SHARP_S
15078 || ASCII_FOLD_RESTRICTED)
15080 assert( toLOWER_L1(ender) < 256 );
15081 *e++ = (char)(toLOWER_L1(ender)); /* should e and the cast be U8? */
15089 && UNLIKELY(is_THREE_CHAR_FOLD_HEAD_latin1_safe(s, e)))
15096 if ( UCHARAT(p) != LATIN_SMALL_LETTER_SHARP_S
15097 || ASCII_FOLD_RESTRICTED)
15099 assert( toLOWER_L1(ender) < 256 );
15100 *e++ = (char)(toLOWER_L1(ender)); /* should e and the cast be U8? */
15110 if (UNLIKELY(is_MULTI_CHAR_FOLD_latin1_safe(s, e))) {
15116 if ( LIKELY(s > s_start)
15117 && UNLIKELY(is_THREE_CHAR_FOLD_latin1_safe(s - 1, e)))
15127 } while (s > s_start);
15134 /* Here, we are done backing up. If we didn't backup at all
15135 * (the likely case), just proceed */
15138 /* If we did find a place to split, reparse the entire node
15139 * stopping where we have calculated. */
15142 /* If we created a temporary folded string under /l, we
15143 * have to map that back to the original */
15144 if (need_to_fold_loc) {
15145 upper_fill = loc_correspondence[s - s_start];
15146 Safefree(locfold_buf);
15147 Safefree(loc_correspondence);
15149 if (upper_fill == 0) {
15150 FAIL2("panic: loc_correspondence[%d] is 0",
15151 (int) (s - s_start));
15155 upper_fill = s - s0;
15159 else if (need_to_fold_loc) {
15160 Safefree(locfold_buf);
15161 Safefree(loc_correspondence);
15164 /* Here the node consists entirely of non-final multi-char
15165 * folds. (Likely it is all 'f's or all 's's.) There's no
15166 * decent place to split it, so give up and just take the
15170 } /* End of verifying node ends with an appropriate char */
15172 /* We need to start the next node at the character that didn't fit
15176 loopdone: /* Jumped to when encounters something that shouldn't be
15179 /* Free up any over-allocated space; cast is to silence bogus
15180 * warning in MS VC */
15181 change_engine_size(pRExC_state,
15182 - (Ptrdiff_t) (current_string_nodes - STR_SZ(len)));
15184 /* I (khw) don't know if you can get here with zero length, but the
15185 * old code handled this situation by creating a zero-length EXACT
15186 * node. Might as well be NOTHING instead */
15188 OP(REGNODE_p(ret)) = NOTHING;
15192 /* If the node type is EXACT here, check to see if it
15193 * should be EXACTL, or EXACT_REQ8. */
15194 if (node_type == EXACT) {
15196 node_type = EXACTL;
15198 else if (requires_utf8_target) {
15199 node_type = EXACT_REQ8;
15202 else if (node_type == LEXACT) {
15203 if (requires_utf8_target) {
15204 node_type = LEXACT_REQ8;
15208 if ( UNLIKELY(has_micro_sign || has_ss)
15209 && (node_type == EXACTFU || ( node_type == EXACTF
15210 && maybe_exactfu)))
15211 { /* These two conditions are problematic in non-UTF-8
15214 node_type = EXACTFUP;
15216 else if (node_type == EXACTFL) {
15218 /* 'maybe_exactfu' is deliberately set above to
15219 * indicate this node type, where all code points in it
15221 if (maybe_exactfu) {
15222 node_type = EXACTFLU8;
15225 _invlist_contains_cp(PL_HasMultiCharFold, ender)))
15227 /* A character that folds to more than one will
15228 * match multiple characters, so can't be SIMPLE.
15229 * We don't have to worry about this with EXACTFLU8
15230 * nodes just above, as they have already been
15231 * folded (since the fold doesn't vary at run
15232 * time). Here, if the final character in the node
15233 * folds to multiple, it can't be simple. (This
15234 * only has an effect if the node has only a single
15235 * character, hence the final one, as elsewhere we
15236 * turn off simple for nodes whose length > 1 */
15240 else if (node_type == EXACTF) { /* Means is /di */
15242 /* This intermediate variable is needed solely because
15243 * the asserts in the macro where used exceed Win32's
15244 * literal string capacity */
15245 char first_char = * STRING(REGNODE_p(ret));
15247 /* If 'maybe_exactfu' is clear, then we need to stay
15248 * /di. If it is set, it means there are no code
15249 * points that match differently depending on UTF8ness
15250 * of the target string, so it can become an EXACTFU
15252 if (! maybe_exactfu) {
15253 RExC_seen_d_op = TRUE;
15255 else if ( isALPHA_FOLD_EQ(first_char, 's')
15256 || isALPHA_FOLD_EQ(ender, 's'))
15258 /* But, if the node begins or ends in an 's' we
15259 * have to defer changing it into an EXACTFU, as
15260 * the node could later get joined with another one
15261 * that ends or begins with 's' creating an 'ss'
15262 * sequence which would then wrongly match the
15263 * sharp s without the target being UTF-8. We
15264 * create a special node that we resolve later when
15265 * we join nodes together */
15267 node_type = EXACTFU_S_EDGE;
15270 node_type = EXACTFU;
15274 if (requires_utf8_target && node_type == EXACTFU) {
15275 node_type = EXACTFU_REQ8;
15279 OP(REGNODE_p(ret)) = node_type;
15280 setSTR_LEN(REGNODE_p(ret), len);
15281 RExC_emit += STR_SZ(len);
15283 /* If the node isn't a single character, it can't be SIMPLE */
15284 if (len > (Size_t) ((UTF) ? UTF8SKIP(STRING(REGNODE_p(ret))) : 1)) {
15288 *flagp |= HASWIDTH | maybe_SIMPLE;
15291 Set_Node_Length(REGNODE_p(ret), p - parse_start - 1);
15295 /* len is STRLEN which is unsigned, need to copy to signed */
15298 vFAIL("Internal disaster");
15301 } /* End of label 'defchar:' */
15303 } /* End of giant switch on input character */
15305 /* Position parse to next real character */
15306 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
15307 FALSE /* Don't force to /x */ );
15308 if ( *RExC_parse == '{'
15309 && OP(REGNODE_p(ret)) != SBOL && ! regcurly(RExC_parse))
15311 if (RExC_strict || new_regcurly(RExC_parse, RExC_end)) {
15313 vFAIL("Unescaped left brace in regex is illegal here");
15315 ckWARNreg(RExC_parse + 1, "Unescaped left brace in regex is"
15316 " passed through");
15324 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
15326 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
15327 * sets up the bitmap and any flags, removing those code points from the
15328 * inversion list, setting it to NULL should it become completely empty */
15332 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
15333 assert(PL_regkind[OP(node)] == ANYOF);
15335 /* There is no bitmap for this node type */
15336 if (inRANGE(OP(node), ANYOFH, ANYOFRb)) {
15340 ANYOF_BITMAP_ZERO(node);
15341 if (*invlist_ptr) {
15343 /* This gets set if we actually need to modify things */
15344 bool change_invlist = FALSE;
15348 /* Start looking through *invlist_ptr */
15349 invlist_iterinit(*invlist_ptr);
15350 while (invlist_iternext(*invlist_ptr, &start, &end)) {
15354 if (end == UV_MAX && start <= NUM_ANYOF_CODE_POINTS) {
15355 ANYOF_FLAGS(node) |= ANYOF_MATCHES_ALL_ABOVE_BITMAP;
15358 /* Quit if are above what we should change */
15359 if (start >= NUM_ANYOF_CODE_POINTS) {
15363 change_invlist = TRUE;
15365 /* Set all the bits in the range, up to the max that we are doing */
15366 high = (end < NUM_ANYOF_CODE_POINTS - 1)
15368 : NUM_ANYOF_CODE_POINTS - 1;
15369 for (i = start; i <= (int) high; i++) {
15370 if (! ANYOF_BITMAP_TEST(node, i)) {
15371 ANYOF_BITMAP_SET(node, i);
15375 invlist_iterfinish(*invlist_ptr);
15377 /* Done with loop; remove any code points that are in the bitmap from
15378 * *invlist_ptr; similarly for code points above the bitmap if we have
15379 * a flag to match all of them anyways */
15380 if (change_invlist) {
15381 _invlist_subtract(*invlist_ptr, PL_InBitmap, invlist_ptr);
15383 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
15384 _invlist_intersection(*invlist_ptr, PL_InBitmap, invlist_ptr);
15387 /* If have completely emptied it, remove it completely */
15388 if (_invlist_len(*invlist_ptr) == 0) {
15389 SvREFCNT_dec_NN(*invlist_ptr);
15390 *invlist_ptr = NULL;
15395 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
15396 Character classes ([:foo:]) can also be negated ([:^foo:]).
15397 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
15398 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
15399 but trigger failures because they are currently unimplemented. */
15401 #define POSIXCC_DONE(c) ((c) == ':')
15402 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
15403 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
15404 #define MAYBE_POSIXCC(c) (POSIXCC(c) || (c) == '^' || (c) == ';')
15406 #define WARNING_PREFIX "Assuming NOT a POSIX class since "
15407 #define NO_BLANKS_POSIX_WARNING "no blanks are allowed in one"
15408 #define SEMI_COLON_POSIX_WARNING "a semi-colon was found instead of a colon"
15410 #define NOT_MEANT_TO_BE_A_POSIX_CLASS (OOB_NAMEDCLASS - 1)
15412 /* 'posix_warnings' and 'warn_text' are names of variables in the following
15414 #define ADD_POSIX_WARNING(p, text) STMT_START { \
15415 if (posix_warnings) { \
15416 if (! RExC_warn_text ) RExC_warn_text = \
15417 (AV *) sv_2mortal((SV *) newAV()); \
15418 av_push(RExC_warn_text, Perl_newSVpvf(aTHX_ \
15422 REPORT_LOCATION_ARGS(p))); \
15425 #define CLEAR_POSIX_WARNINGS() \
15427 if (posix_warnings && RExC_warn_text) \
15428 av_clear(RExC_warn_text); \
15431 #define CLEAR_POSIX_WARNINGS_AND_RETURN(ret) \
15433 CLEAR_POSIX_WARNINGS(); \
15438 S_handle_possible_posix(pTHX_ RExC_state_t *pRExC_state,
15440 const char * const s, /* Where the putative posix class begins.
15441 Normally, this is one past the '['. This
15442 parameter exists so it can be somewhere
15443 besides RExC_parse. */
15444 char ** updated_parse_ptr, /* Where to set the updated parse pointer, or
15446 AV ** posix_warnings, /* Where to place any generated warnings, or
15448 const bool check_only /* Don't die if error */
15451 /* This parses what the caller thinks may be one of the three POSIX
15453 * 1) a character class, like [:blank:]
15454 * 2) a collating symbol, like [. .]
15455 * 3) an equivalence class, like [= =]
15456 * In the latter two cases, it croaks if it finds a syntactically legal
15457 * one, as these are not handled by Perl.
15459 * The main purpose is to look for a POSIX character class. It returns:
15460 * a) the class number
15461 * if it is a completely syntactically and semantically legal class.
15462 * 'updated_parse_ptr', if not NULL, is set to point to just after the
15463 * closing ']' of the class
15464 * b) OOB_NAMEDCLASS
15465 * if it appears that one of the three POSIX constructs was meant, but
15466 * its specification was somehow defective. 'updated_parse_ptr', if
15467 * not NULL, is set to point to the character just after the end
15468 * character of the class. See below for handling of warnings.
15469 * c) NOT_MEANT_TO_BE_A_POSIX_CLASS
15470 * if it doesn't appear that a POSIX construct was intended.
15471 * 'updated_parse_ptr' is not changed. No warnings nor errors are
15474 * In b) there may be errors or warnings generated. If 'check_only' is
15475 * TRUE, then any errors are discarded. Warnings are returned to the
15476 * caller via an AV* created into '*posix_warnings' if it is not NULL. If
15477 * instead it is NULL, warnings are suppressed.
15479 * The reason for this function, and its complexity is that a bracketed
15480 * character class can contain just about anything. But it's easy to
15481 * mistype the very specific posix class syntax but yielding a valid
15482 * regular bracketed class, so it silently gets compiled into something
15483 * quite unintended.
15485 * The solution adopted here maintains backward compatibility except that
15486 * it adds a warning if it looks like a posix class was intended but
15487 * improperly specified. The warning is not raised unless what is input
15488 * very closely resembles one of the 14 legal posix classes. To do this,
15489 * it uses fuzzy parsing. It calculates how many single-character edits it
15490 * would take to transform what was input into a legal posix class. Only
15491 * if that number is quite small does it think that the intention was a
15492 * posix class. Obviously these are heuristics, and there will be cases
15493 * where it errs on one side or another, and they can be tweaked as
15494 * experience informs.
15496 * The syntax for a legal posix class is:
15498 * qr/(?xa: \[ : \^? [[:lower:]]{4,6} : \] )/
15500 * What this routine considers syntactically to be an intended posix class
15501 * is this (the comments indicate some restrictions that the pattern
15504 * qr/(?x: \[? # The left bracket, possibly
15506 * \h* # possibly followed by blanks
15507 * (?: \^ \h* )? # possibly a misplaced caret
15508 * [:;]? # The opening class character,
15509 * # possibly omitted. A typo
15510 * # semi-colon can also be used.
15512 * \^? # possibly a correctly placed
15513 * # caret, but not if there was also
15514 * # a misplaced one
15516 * .{3,15} # The class name. If there are
15517 * # deviations from the legal syntax,
15518 * # its edit distance must be close
15519 * # to a real class name in order
15520 * # for it to be considered to be
15521 * # an intended posix class.
15523 * [[:punct:]]? # The closing class character,
15524 * # possibly omitted. If not a colon
15525 * # nor semi colon, the class name
15526 * # must be even closer to a valid
15529 * \]? # The right bracket, possibly
15533 * In the above, \h must be ASCII-only.
15535 * These are heuristics, and can be tweaked as field experience dictates.
15536 * There will be cases when someone didn't intend to specify a posix class
15537 * that this warns as being so. The goal is to minimize these, while
15538 * maximizing the catching of things intended to be a posix class that
15539 * aren't parsed as such.
15543 const char * const e = RExC_end;
15544 unsigned complement = 0; /* If to complement the class */
15545 bool found_problem = FALSE; /* Assume OK until proven otherwise */
15546 bool has_opening_bracket = FALSE;
15547 bool has_opening_colon = FALSE;
15548 int class_number = OOB_NAMEDCLASS; /* Out-of-bounds until find
15550 const char * possible_end = NULL; /* used for a 2nd parse pass */
15551 const char* name_start; /* ptr to class name first char */
15553 /* If the number of single-character typos the input name is away from a
15554 * legal name is no more than this number, it is considered to have meant
15555 * the legal name */
15556 int max_distance = 2;
15558 /* to store the name. The size determines the maximum length before we
15559 * decide that no posix class was intended. Should be at least
15560 * sizeof("alphanumeric") */
15562 STATIC_ASSERT_DECL(C_ARRAY_LENGTH(input_text) >= sizeof "alphanumeric");
15564 PERL_ARGS_ASSERT_HANDLE_POSSIBLE_POSIX;
15566 CLEAR_POSIX_WARNINGS();
15569 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
15572 if (*(p - 1) != '[') {
15573 ADD_POSIX_WARNING(p, "it doesn't start with a '['");
15574 found_problem = TRUE;
15577 has_opening_bracket = TRUE;
15580 /* They could be confused and think you can put spaces between the
15583 found_problem = TRUE;
15587 } while (p < e && isBLANK(*p));
15589 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15592 /* For [. .] and [= =]. These are quite different internally from [: :],
15593 * so they are handled separately. */
15594 if (POSIXCC_NOTYET(*p) && p < e - 3) /* 1 for the close, and 1 for the ']'
15595 and 1 for at least one char in it
15598 const char open_char = *p;
15599 const char * temp_ptr = p + 1;
15601 /* These two constructs are not handled by perl, and if we find a
15602 * syntactically valid one, we croak. khw, who wrote this code, finds
15603 * this explanation of them very unclear:
15604 * http://pubs.opengroup.org/onlinepubs/009696899/basedefs/xbd_chap09.html
15605 * And searching the rest of the internet wasn't very helpful either.
15606 * It looks like just about any byte can be in these constructs,
15607 * depending on the locale. But unless the pattern is being compiled
15608 * under /l, which is very rare, Perl runs under the C or POSIX locale.
15609 * In that case, it looks like [= =] isn't allowed at all, and that
15610 * [. .] could be any single code point, but for longer strings the
15611 * constituent characters would have to be the ASCII alphabetics plus
15612 * the minus-hyphen. Any sensible locale definition would limit itself
15613 * to these. And any portable one definitely should. Trying to parse
15614 * the general case is a nightmare (see [perl #127604]). So, this code
15615 * looks only for interiors of these constructs that match:
15617 * Using \w relaxes the apparent rules a little, without adding much
15618 * danger of mistaking something else for one of these constructs.
15620 * [. .] in some implementations described on the internet is usable to
15621 * escape a character that otherwise is special in bracketed character
15622 * classes. For example [.].] means a literal right bracket instead of
15623 * the ending of the class
15625 * [= =] can legitimately contain a [. .] construct, but we don't
15626 * handle this case, as that [. .] construct will later get parsed
15627 * itself and croak then. And [= =] is checked for even when not under
15628 * /l, as Perl has long done so.
15630 * The code below relies on there being a trailing NUL, so it doesn't
15631 * have to keep checking if the parse ptr < e.
15633 if (temp_ptr[1] == open_char) {
15636 else while ( temp_ptr < e
15637 && (isWORDCHAR(*temp_ptr) || *temp_ptr == '-'))
15642 if (*temp_ptr == open_char) {
15644 if (*temp_ptr == ']') {
15646 if (! found_problem && ! check_only) {
15647 RExC_parse = (char *) temp_ptr;
15648 vFAIL3("POSIX syntax [%c %c] is reserved for future "
15649 "extensions", open_char, open_char);
15652 /* Here, the syntax wasn't completely valid, or else the call
15653 * is to check-only */
15654 if (updated_parse_ptr) {
15655 *updated_parse_ptr = (char *) temp_ptr;
15658 CLEAR_POSIX_WARNINGS_AND_RETURN(OOB_NAMEDCLASS);
15662 /* If we find something that started out to look like one of these
15663 * constructs, but isn't, we continue below so that it can be checked
15664 * for being a class name with a typo of '.' or '=' instead of a colon.
15668 /* Here, we think there is a possibility that a [: :] class was meant, and
15669 * we have the first real character. It could be they think the '^' comes
15672 found_problem = TRUE;
15673 ADD_POSIX_WARNING(p + 1, "the '^' must come after the colon");
15678 found_problem = TRUE;
15682 } while (p < e && isBLANK(*p));
15684 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15688 /* But the first character should be a colon, which they could have easily
15689 * mistyped on a qwerty keyboard as a semi-colon (and which may be hard to
15690 * distinguish from a colon, so treat that as a colon). */
15693 has_opening_colon = TRUE;
15695 else if (*p == ';') {
15696 found_problem = TRUE;
15698 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
15699 has_opening_colon = TRUE;
15702 found_problem = TRUE;
15703 ADD_POSIX_WARNING(p, "there must be a starting ':'");
15705 /* Consider an initial punctuation (not one of the recognized ones) to
15706 * be a left terminator */
15707 if (*p != '^' && *p != ']' && isPUNCT(*p)) {
15712 /* They may think that you can put spaces between the components */
15714 found_problem = TRUE;
15718 } while (p < e && isBLANK(*p));
15720 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15725 /* We consider something like [^:^alnum:]] to not have been intended to
15726 * be a posix class, but XXX maybe we should */
15728 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15735 /* Again, they may think that you can put spaces between the components */
15737 found_problem = TRUE;
15741 } while (p < e && isBLANK(*p));
15743 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15748 /* XXX This ']' may be a typo, and something else was meant. But
15749 * treating it as such creates enough complications, that that
15750 * possibility isn't currently considered here. So we assume that the
15751 * ']' is what is intended, and if we've already found an initial '[',
15752 * this leaves this construct looking like [:] or [:^], which almost
15753 * certainly weren't intended to be posix classes */
15754 if (has_opening_bracket) {
15755 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15758 /* But this function can be called when we parse the colon for
15759 * something like qr/[alpha:]]/, so we back up to look for the
15764 found_problem = TRUE;
15765 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
15767 else if (*p != ':') {
15769 /* XXX We are currently very restrictive here, so this code doesn't
15770 * consider the possibility that, say, /[alpha.]]/ was intended to
15771 * be a posix class. */
15772 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15775 /* Here we have something like 'foo:]'. There was no initial colon,
15776 * and we back up over 'foo. XXX Unlike the going forward case, we
15777 * don't handle typos of non-word chars in the middle */
15778 has_opening_colon = FALSE;
15781 while (p > RExC_start && isWORDCHAR(*p)) {
15786 /* Here, we have positioned ourselves to where we think the first
15787 * character in the potential class is */
15790 /* Now the interior really starts. There are certain key characters that
15791 * can end the interior, or these could just be typos. To catch both
15792 * cases, we may have to do two passes. In the first pass, we keep on
15793 * going unless we come to a sequence that matches
15794 * qr/ [[:punct:]] [[:blank:]]* \] /xa
15795 * This means it takes a sequence to end the pass, so two typos in a row if
15796 * that wasn't what was intended. If the class is perfectly formed, just
15797 * this one pass is needed. We also stop if there are too many characters
15798 * being accumulated, but this number is deliberately set higher than any
15799 * real class. It is set high enough so that someone who thinks that
15800 * 'alphanumeric' is a correct name would get warned that it wasn't.
15801 * While doing the pass, we keep track of where the key characters were in
15802 * it. If we don't find an end to the class, and one of the key characters
15803 * was found, we redo the pass, but stop when we get to that character.
15804 * Thus the key character was considered a typo in the first pass, but a
15805 * terminator in the second. If two key characters are found, we stop at
15806 * the second one in the first pass. Again this can miss two typos, but
15807 * catches a single one
15809 * In the first pass, 'possible_end' starts as NULL, and then gets set to
15810 * point to the first key character. For the second pass, it starts as -1.
15816 bool has_blank = FALSE;
15817 bool has_upper = FALSE;
15818 bool has_terminating_colon = FALSE;
15819 bool has_terminating_bracket = FALSE;
15820 bool has_semi_colon = FALSE;
15821 unsigned int name_len = 0;
15822 int punct_count = 0;
15826 /* Squeeze out blanks when looking up the class name below */
15827 if (isBLANK(*p) ) {
15829 found_problem = TRUE;
15834 /* The name will end with a punctuation */
15836 const char * peek = p + 1;
15838 /* Treat any non-']' punctuation followed by a ']' (possibly
15839 * with intervening blanks) as trying to terminate the class.
15840 * ']]' is very likely to mean a class was intended (but
15841 * missing the colon), but the warning message that gets
15842 * generated shows the error position better if we exit the
15843 * loop at the bottom (eventually), so skip it here. */
15845 if (peek < e && isBLANK(*peek)) {
15847 found_problem = TRUE;
15850 } while (peek < e && isBLANK(*peek));
15853 if (peek < e && *peek == ']') {
15854 has_terminating_bracket = TRUE;
15856 has_terminating_colon = TRUE;
15858 else if (*p == ';') {
15859 has_semi_colon = TRUE;
15860 has_terminating_colon = TRUE;
15863 found_problem = TRUE;
15870 /* Here we have punctuation we thought didn't end the class.
15871 * Keep track of the position of the key characters that are
15872 * more likely to have been class-enders */
15873 if (*p == ']' || *p == '[' || *p == ':' || *p == ';') {
15875 /* Allow just one such possible class-ender not actually
15876 * ending the class. */
15877 if (possible_end) {
15883 /* If we have too many punctuation characters, no use in
15885 if (++punct_count > max_distance) {
15889 /* Treat the punctuation as a typo. */
15890 input_text[name_len++] = *p;
15893 else if (isUPPER(*p)) { /* Use lowercase for lookup */
15894 input_text[name_len++] = toLOWER(*p);
15896 found_problem = TRUE;
15898 } else if (! UTF || UTF8_IS_INVARIANT(*p)) {
15899 input_text[name_len++] = *p;
15903 input_text[name_len++] = utf8_to_uvchr_buf((U8 *) p, e, NULL);
15907 /* The declaration of 'input_text' is how long we allow a potential
15908 * class name to be, before saying they didn't mean a class name at
15910 if (name_len >= C_ARRAY_LENGTH(input_text)) {
15915 /* We get to here when the possible class name hasn't been properly
15916 * terminated before:
15917 * 1) we ran off the end of the pattern; or
15918 * 2) found two characters, each of which might have been intended to
15919 * be the name's terminator
15920 * 3) found so many punctuation characters in the purported name,
15921 * that the edit distance to a valid one is exceeded
15922 * 4) we decided it was more characters than anyone could have
15923 * intended to be one. */
15925 found_problem = TRUE;
15927 /* In the final two cases, we know that looking up what we've
15928 * accumulated won't lead to a match, even a fuzzy one. */
15929 if ( name_len >= C_ARRAY_LENGTH(input_text)
15930 || punct_count > max_distance)
15932 /* If there was an intermediate key character that could have been
15933 * an intended end, redo the parse, but stop there */
15934 if (possible_end && possible_end != (char *) -1) {
15935 possible_end = (char *) -1; /* Special signal value to say
15936 we've done a first pass */
15941 /* Otherwise, it can't have meant to have been a class */
15942 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15945 /* If we ran off the end, and the final character was a punctuation
15946 * one, back up one, to look at that final one just below. Later, we
15947 * will restore the parse pointer if appropriate */
15948 if (name_len && p == e && isPUNCT(*(p-1))) {
15953 if (p < e && isPUNCT(*p)) {
15955 has_terminating_bracket = TRUE;
15957 /* If this is a 2nd ']', and the first one is just below this
15958 * one, consider that to be the real terminator. This gives a
15959 * uniform and better positioning for the warning message */
15961 && possible_end != (char *) -1
15962 && *possible_end == ']'
15963 && name_len && input_text[name_len - 1] == ']')
15968 /* And this is actually equivalent to having done the 2nd
15969 * pass now, so set it to not try again */
15970 possible_end = (char *) -1;
15975 has_terminating_colon = TRUE;
15977 else if (*p == ';') {
15978 has_semi_colon = TRUE;
15979 has_terminating_colon = TRUE;
15987 /* Here, we have a class name to look up. We can short circuit the
15988 * stuff below for short names that can't possibly be meant to be a
15989 * class name. (We can do this on the first pass, as any second pass
15990 * will yield an even shorter name) */
15991 if (name_len < 3) {
15992 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15995 /* Find which class it is. Initially switch on the length of the name.
15997 switch (name_len) {
15999 if (memEQs(name_start, 4, "word")) {
16000 /* this is not POSIX, this is the Perl \w */
16001 class_number = ANYOF_WORDCHAR;
16005 /* Names all of length 5: alnum alpha ascii blank cntrl digit
16006 * graph lower print punct space upper
16007 * Offset 4 gives the best switch position. */
16008 switch (name_start[4]) {
16010 if (memBEGINs(name_start, 5, "alph")) /* alpha */
16011 class_number = ANYOF_ALPHA;
16014 if (memBEGINs(name_start, 5, "spac")) /* space */
16015 class_number = ANYOF_SPACE;
16018 if (memBEGINs(name_start, 5, "grap")) /* graph */
16019 class_number = ANYOF_GRAPH;
16022 if (memBEGINs(name_start, 5, "asci")) /* ascii */
16023 class_number = ANYOF_ASCII;
16026 if (memBEGINs(name_start, 5, "blan")) /* blank */
16027 class_number = ANYOF_BLANK;
16030 if (memBEGINs(name_start, 5, "cntr")) /* cntrl */
16031 class_number = ANYOF_CNTRL;
16034 if (memBEGINs(name_start, 5, "alnu")) /* alnum */
16035 class_number = ANYOF_ALPHANUMERIC;
16038 if (memBEGINs(name_start, 5, "lowe")) /* lower */
16039 class_number = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
16040 else if (memBEGINs(name_start, 5, "uppe")) /* upper */
16041 class_number = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
16044 if (memBEGINs(name_start, 5, "digi")) /* digit */
16045 class_number = ANYOF_DIGIT;
16046 else if (memBEGINs(name_start, 5, "prin")) /* print */
16047 class_number = ANYOF_PRINT;
16048 else if (memBEGINs(name_start, 5, "punc")) /* punct */
16049 class_number = ANYOF_PUNCT;
16054 if (memEQs(name_start, 6, "xdigit"))
16055 class_number = ANYOF_XDIGIT;
16059 /* If the name exactly matches a posix class name the class number will
16060 * here be set to it, and the input almost certainly was meant to be a
16061 * posix class, so we can skip further checking. If instead the syntax
16062 * is exactly correct, but the name isn't one of the legal ones, we
16063 * will return that as an error below. But if neither of these apply,
16064 * it could be that no posix class was intended at all, or that one
16065 * was, but there was a typo. We tease these apart by doing fuzzy
16066 * matching on the name */
16067 if (class_number == OOB_NAMEDCLASS && found_problem) {
16068 const UV posix_names[][6] = {
16069 { 'a', 'l', 'n', 'u', 'm' },
16070 { 'a', 'l', 'p', 'h', 'a' },
16071 { 'a', 's', 'c', 'i', 'i' },
16072 { 'b', 'l', 'a', 'n', 'k' },
16073 { 'c', 'n', 't', 'r', 'l' },
16074 { 'd', 'i', 'g', 'i', 't' },
16075 { 'g', 'r', 'a', 'p', 'h' },
16076 { 'l', 'o', 'w', 'e', 'r' },
16077 { 'p', 'r', 'i', 'n', 't' },
16078 { 'p', 'u', 'n', 'c', 't' },
16079 { 's', 'p', 'a', 'c', 'e' },
16080 { 'u', 'p', 'p', 'e', 'r' },
16081 { 'w', 'o', 'r', 'd' },
16082 { 'x', 'd', 'i', 'g', 'i', 't' }
16084 /* The names of the above all have added NULs to make them the same
16085 * size, so we need to also have the real lengths */
16086 const UV posix_name_lengths[] = {
16087 sizeof("alnum") - 1,
16088 sizeof("alpha") - 1,
16089 sizeof("ascii") - 1,
16090 sizeof("blank") - 1,
16091 sizeof("cntrl") - 1,
16092 sizeof("digit") - 1,
16093 sizeof("graph") - 1,
16094 sizeof("lower") - 1,
16095 sizeof("print") - 1,
16096 sizeof("punct") - 1,
16097 sizeof("space") - 1,
16098 sizeof("upper") - 1,
16099 sizeof("word") - 1,
16100 sizeof("xdigit")- 1
16103 int temp_max = max_distance; /* Use a temporary, so if we
16104 reparse, we haven't changed the
16107 /* Use a smaller max edit distance if we are missing one of the
16109 if ( has_opening_bracket + has_opening_colon < 2
16110 || has_terminating_bracket + has_terminating_colon < 2)
16115 /* See if the input name is close to a legal one */
16116 for (i = 0; i < C_ARRAY_LENGTH(posix_names); i++) {
16118 /* Short circuit call if the lengths are too far apart to be
16120 if (abs( (int) (name_len - posix_name_lengths[i]))
16126 if (edit_distance(input_text,
16129 posix_name_lengths[i],
16133 { /* If it is close, it probably was intended to be a class */
16134 goto probably_meant_to_be;
16138 /* Here the input name is not close enough to a valid class name
16139 * for us to consider it to be intended to be a posix class. If
16140 * we haven't already done so, and the parse found a character that
16141 * could have been terminators for the name, but which we absorbed
16142 * as typos during the first pass, repeat the parse, signalling it
16143 * to stop at that character */
16144 if (possible_end && possible_end != (char *) -1) {
16145 possible_end = (char *) -1;
16150 /* Here neither pass found a close-enough class name */
16151 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
16154 probably_meant_to_be:
16156 /* Here we think that a posix specification was intended. Update any
16158 if (updated_parse_ptr) {
16159 *updated_parse_ptr = (char *) p;
16162 /* If a posix class name was intended but incorrectly specified, we
16163 * output or return the warnings */
16164 if (found_problem) {
16166 /* We set flags for these issues in the parse loop above instead of
16167 * adding them to the list of warnings, because we can parse it
16168 * twice, and we only want one warning instance */
16170 ADD_POSIX_WARNING(p, "the name must be all lowercase letters");
16173 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
16175 if (has_semi_colon) {
16176 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
16178 else if (! has_terminating_colon) {
16179 ADD_POSIX_WARNING(p, "there is no terminating ':'");
16181 if (! has_terminating_bracket) {
16182 ADD_POSIX_WARNING(p, "there is no terminating ']'");
16185 if ( posix_warnings
16187 && av_top_index(RExC_warn_text) > -1)
16189 *posix_warnings = RExC_warn_text;
16192 else if (class_number != OOB_NAMEDCLASS) {
16193 /* If it is a known class, return the class. The class number
16194 * #defines are structured so each complement is +1 to the normal
16196 CLEAR_POSIX_WARNINGS_AND_RETURN(class_number + complement);
16198 else if (! check_only) {
16200 /* Here, it is an unrecognized class. This is an error (unless the
16201 * call is to check only, which we've already handled above) */
16202 const char * const complement_string = (complement)
16205 RExC_parse = (char *) p;
16206 vFAIL3utf8f("POSIX class [:%s%" UTF8f ":] unknown",
16208 UTF8fARG(UTF, RExC_parse - name_start - 2, name_start));
16212 return OOB_NAMEDCLASS;
16214 #undef ADD_POSIX_WARNING
16216 STATIC unsigned int
16217 S_regex_set_precedence(const U8 my_operator) {
16219 /* Returns the precedence in the (?[...]) construct of the input operator,
16220 * specified by its character representation. The precedence follows
16221 * general Perl rules, but it extends this so that ')' and ']' have (low)
16222 * precedence even though they aren't really operators */
16224 switch (my_operator) {
16240 NOT_REACHED; /* NOTREACHED */
16241 return 0; /* Silence compiler warning */
16244 STATIC regnode_offset
16245 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
16246 I32 *flagp, U32 depth,
16247 char * const oregcomp_parse)
16249 /* Handle the (?[...]) construct to do set operations */
16251 U8 curchar; /* Current character being parsed */
16252 UV start, end; /* End points of code point ranges */
16253 SV* final = NULL; /* The end result inversion list */
16254 SV* result_string; /* 'final' stringified */
16255 AV* stack; /* stack of operators and operands not yet
16257 AV* fence_stack = NULL; /* A stack containing the positions in
16258 'stack' of where the undealt-with left
16259 parens would be if they were actually
16261 /* The 'volatile' is a workaround for an optimiser bug
16262 * in Solaris Studio 12.3. See RT #127455 */
16263 volatile IV fence = 0; /* Position of where most recent undealt-
16264 with left paren in stack is; -1 if none.
16266 STRLEN len; /* Temporary */
16267 regnode_offset node; /* Temporary, and final regnode returned by
16269 const bool save_fold = FOLD; /* Temporary */
16270 char *save_end, *save_parse; /* Temporaries */
16271 const bool in_locale = LOC; /* we turn off /l during processing */
16273 DECLARE_AND_GET_RE_DEBUG_FLAGS;
16275 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
16276 PERL_UNUSED_ARG(oregcomp_parse); /* Only for Set_Node_Length */
16278 DEBUG_PARSE("xcls");
16281 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
16284 /* The use of this operator implies /u. This is required so that the
16285 * compile time values are valid in all runtime cases */
16286 REQUIRE_UNI_RULES(flagp, 0);
16288 ckWARNexperimental(RExC_parse,
16289 WARN_EXPERIMENTAL__REGEX_SETS,
16290 "The regex_sets feature is experimental");
16292 /* Everything in this construct is a metacharacter. Operands begin with
16293 * either a '\' (for an escape sequence), or a '[' for a bracketed
16294 * character class. Any other character should be an operator, or
16295 * parenthesis for grouping. Both types of operands are handled by calling
16296 * regclass() to parse them. It is called with a parameter to indicate to
16297 * return the computed inversion list. The parsing here is implemented via
16298 * a stack. Each entry on the stack is a single character representing one
16299 * of the operators; or else a pointer to an operand inversion list. */
16301 #define IS_OPERATOR(a) SvIOK(a)
16302 #define IS_OPERAND(a) (! IS_OPERATOR(a))
16304 /* The stack is kept in Łukasiewicz order. (That's pronounced similar
16305 * to luke-a-shave-itch (or -itz), but people who didn't want to bother
16306 * with pronouncing it called it Reverse Polish instead, but now that YOU
16307 * know how to pronounce it you can use the correct term, thus giving due
16308 * credit to the person who invented it, and impressing your geek friends.
16309 * Wikipedia says that the pronounciation of "Ł" has been changing so that
16310 * it is now more like an English initial W (as in wonk) than an L.)
16312 * This means that, for example, 'a | b & c' is stored on the stack as
16320 * where the numbers in brackets give the stack [array] element number.
16321 * In this implementation, parentheses are not stored on the stack.
16322 * Instead a '(' creates a "fence" so that the part of the stack below the
16323 * fence is invisible except to the corresponding ')' (this allows us to
16324 * replace testing for parens, by using instead subtraction of the fence
16325 * position). As new operands are processed they are pushed onto the stack
16326 * (except as noted in the next paragraph). New operators of higher
16327 * precedence than the current final one are inserted on the stack before
16328 * the lhs operand (so that when the rhs is pushed next, everything will be
16329 * in the correct positions shown above. When an operator of equal or
16330 * lower precedence is encountered in parsing, all the stacked operations
16331 * of equal or higher precedence are evaluated, leaving the result as the
16332 * top entry on the stack. This makes higher precedence operations
16333 * evaluate before lower precedence ones, and causes operations of equal
16334 * precedence to left associate.
16336 * The only unary operator '!' is immediately pushed onto the stack when
16337 * encountered. When an operand is encountered, if the top of the stack is
16338 * a '!", the complement is immediately performed, and the '!' popped. The
16339 * resulting value is treated as a new operand, and the logic in the
16340 * previous paragraph is executed. Thus in the expression
16342 * the stack looks like
16348 * as 'b' gets parsed, the latter gets evaluated to '!b', and the stack
16355 * A ')' is treated as an operator with lower precedence than all the
16356 * aforementioned ones, which causes all operations on the stack above the
16357 * corresponding '(' to be evaluated down to a single resultant operand.
16358 * Then the fence for the '(' is removed, and the operand goes through the
16359 * algorithm above, without the fence.
16361 * A separate stack is kept of the fence positions, so that the position of
16362 * the latest so-far unbalanced '(' is at the top of it.
16364 * The ']' ending the construct is treated as the lowest operator of all,
16365 * so that everything gets evaluated down to a single operand, which is the
16368 sv_2mortal((SV *)(stack = newAV()));
16369 sv_2mortal((SV *)(fence_stack = newAV()));
16371 while (RExC_parse < RExC_end) {
16372 I32 top_index; /* Index of top-most element in 'stack' */
16373 SV** top_ptr; /* Pointer to top 'stack' element */
16374 SV* current = NULL; /* To contain the current inversion list
16376 SV* only_to_avoid_leaks;
16378 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
16379 TRUE /* Force /x */ );
16380 if (RExC_parse >= RExC_end) { /* Fail */
16384 curchar = UCHARAT(RExC_parse);
16388 #ifdef ENABLE_REGEX_SETS_DEBUGGING
16389 /* Enable with -Accflags=-DENABLE_REGEX_SETS_DEBUGGING */
16390 DEBUG_U(dump_regex_sets_structures(pRExC_state,
16391 stack, fence, fence_stack));
16394 top_index = av_tindex_skip_len_mg(stack);
16397 SV** stacked_ptr; /* Ptr to something already on 'stack' */
16398 char stacked_operator; /* The topmost operator on the 'stack'. */
16399 SV* lhs; /* Operand to the left of the operator */
16400 SV* rhs; /* Operand to the right of the operator */
16401 SV* fence_ptr; /* Pointer to top element of the fence
16405 if ( RExC_parse < RExC_end - 2
16406 && UCHARAT(RExC_parse + 1) == '?'
16407 && UCHARAT(RExC_parse + 2) == '^')
16409 const regnode_offset orig_emit = RExC_emit;
16410 SV * resultant_invlist;
16412 /* If is a '(?^', could be an embedded '(?^flags:(?[...])'.
16413 * This happens when we have some thing like
16415 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
16417 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
16419 * Here we would be handling the interpolated
16420 * '$thai_or_lao'. We handle this by a recursive call to
16421 * reg which returns the inversion list the
16422 * interpolated expression evaluates to. Actually, the
16423 * return is a special regnode containing a pointer to that
16424 * inversion list. If the return isn't that regnode alone,
16425 * we know that this wasn't such an interpolation, which is
16426 * an error: we need to get a single inversion list back
16427 * from the recursion */
16432 node = reg(pRExC_state, 2, flagp, depth+1);
16433 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16435 if ( OP(REGNODE_p(node)) != REGEX_SET
16436 /* If more than a single node returned, the nested
16437 * parens evaluated to more than just a (?[...]),
16438 * which isn't legal */
16440 vFAIL("Expecting interpolated extended charclass");
16442 resultant_invlist = (SV *) ARGp(REGNODE_p(node));
16443 current = invlist_clone(resultant_invlist, NULL);
16444 SvREFCNT_dec(resultant_invlist);
16447 RExC_emit = orig_emit;
16448 goto handle_operand;
16451 /* A regular '('. Look behind for illegal syntax */
16452 if (top_index - fence >= 0) {
16453 /* If the top entry on the stack is an operator, it had
16454 * better be a '!', otherwise the entry below the top
16455 * operand should be an operator */
16456 if ( ! (top_ptr = av_fetch(stack, top_index, FALSE))
16457 || (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) != '!')
16458 || ( IS_OPERAND(*top_ptr)
16459 && ( top_index - fence < 1
16460 || ! (stacked_ptr = av_fetch(stack,
16463 || ! IS_OPERATOR(*stacked_ptr))))
16466 vFAIL("Unexpected '(' with no preceding operator");
16470 /* Stack the position of this undealt-with left paren */
16471 av_push(fence_stack, newSViv(fence));
16472 fence = top_index + 1;
16476 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
16477 * multi-char folds are allowed. */
16478 if (!regclass(pRExC_state, flagp, depth+1,
16479 TRUE, /* means parse just the next thing */
16480 FALSE, /* don't allow multi-char folds */
16481 FALSE, /* don't silence non-portable warnings. */
16483 FALSE, /* Require return to be an ANYOF */
16486 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16487 goto regclass_failed;
16490 /* regclass() will return with parsing just the \ sequence,
16491 * leaving the parse pointer at the next thing to parse */
16493 goto handle_operand;
16495 case '[': /* Is a bracketed character class */
16497 /* See if this is a [:posix:] class. */
16498 bool is_posix_class = (OOB_NAMEDCLASS
16499 < handle_possible_posix(pRExC_state,
16503 TRUE /* checking only */));
16504 /* If it is a posix class, leave the parse pointer at the '['
16505 * to fool regclass() into thinking it is part of a
16506 * '[[:posix:]]'. */
16507 if (! is_posix_class) {
16511 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
16512 * multi-char folds are allowed. */
16513 if (!regclass(pRExC_state, flagp, depth+1,
16514 is_posix_class, /* parse the whole char
16515 class only if not a
16517 FALSE, /* don't allow multi-char folds */
16518 TRUE, /* silence non-portable warnings. */
16520 FALSE, /* Require return to be an ANYOF */
16523 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16524 goto regclass_failed;
16531 /* function call leaves parse pointing to the ']', except if we
16533 if (is_posix_class) {
16537 goto handle_operand;
16541 if (top_index >= 1) {
16542 goto join_operators;
16545 /* Only a single operand on the stack: are done */
16549 if (av_tindex_skip_len_mg(fence_stack) < 0) {
16550 if (UCHARAT(RExC_parse - 1) == ']') {
16554 vFAIL("Unexpected ')'");
16557 /* If nothing after the fence, is missing an operand */
16558 if (top_index - fence < 0) {
16562 /* If at least two things on the stack, treat this as an
16564 if (top_index - fence >= 1) {
16565 goto join_operators;
16568 /* Here only a single thing on the fenced stack, and there is a
16569 * fence. Get rid of it */
16570 fence_ptr = av_pop(fence_stack);
16572 fence = SvIV(fence_ptr);
16573 SvREFCNT_dec_NN(fence_ptr);
16580 /* Having gotten rid of the fence, we pop the operand at the
16581 * stack top and process it as a newly encountered operand */
16582 current = av_pop(stack);
16583 if (IS_OPERAND(current)) {
16584 goto handle_operand;
16596 /* These binary operators should have a left operand already
16598 if ( top_index - fence < 0
16599 || top_index - fence == 1
16600 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
16601 || ! IS_OPERAND(*top_ptr))
16603 goto unexpected_binary;
16606 /* If only the one operand is on the part of the stack visible
16607 * to us, we just place this operator in the proper position */
16608 if (top_index - fence < 2) {
16610 /* Place the operator before the operand */
16612 SV* lhs = av_pop(stack);
16613 av_push(stack, newSVuv(curchar));
16614 av_push(stack, lhs);
16618 /* But if there is something else on the stack, we need to
16619 * process it before this new operator if and only if the
16620 * stacked operation has equal or higher precedence than the
16625 /* The operator on the stack is supposed to be below both its
16627 if ( ! (stacked_ptr = av_fetch(stack, top_index - 2, FALSE))
16628 || IS_OPERAND(*stacked_ptr))
16630 /* But if not, it's legal and indicates we are completely
16631 * done if and only if we're currently processing a ']',
16632 * which should be the final thing in the expression */
16633 if (curchar == ']') {
16639 vFAIL2("Unexpected binary operator '%c' with no "
16640 "preceding operand", curchar);
16642 stacked_operator = (char) SvUV(*stacked_ptr);
16644 if (regex_set_precedence(curchar)
16645 > regex_set_precedence(stacked_operator))
16647 /* Here, the new operator has higher precedence than the
16648 * stacked one. This means we need to add the new one to
16649 * the stack to await its rhs operand (and maybe more
16650 * stuff). We put it before the lhs operand, leaving
16651 * untouched the stacked operator and everything below it
16653 lhs = av_pop(stack);
16654 assert(IS_OPERAND(lhs));
16656 av_push(stack, newSVuv(curchar));
16657 av_push(stack, lhs);
16661 /* Here, the new operator has equal or lower precedence than
16662 * what's already there. This means the operation already
16663 * there should be performed now, before the new one. */
16665 rhs = av_pop(stack);
16666 if (! IS_OPERAND(rhs)) {
16668 /* This can happen when a ! is not followed by an operand,
16669 * like in /(?[\t &!])/ */
16673 lhs = av_pop(stack);
16675 if (! IS_OPERAND(lhs)) {
16677 /* This can happen when there is an empty (), like in
16678 * /(?[[0]+()+])/ */
16682 switch (stacked_operator) {
16684 _invlist_intersection(lhs, rhs, &rhs);
16689 _invlist_union(lhs, rhs, &rhs);
16693 _invlist_subtract(lhs, rhs, &rhs);
16696 case '^': /* The union minus the intersection */
16701 _invlist_union(lhs, rhs, &u);
16702 _invlist_intersection(lhs, rhs, &i);
16703 _invlist_subtract(u, i, &rhs);
16704 SvREFCNT_dec_NN(i);
16705 SvREFCNT_dec_NN(u);
16711 /* Here, the higher precedence operation has been done, and the
16712 * result is in 'rhs'. We overwrite the stacked operator with
16713 * the result. Then we redo this code to either push the new
16714 * operator onto the stack or perform any higher precedence
16715 * stacked operation */
16716 only_to_avoid_leaks = av_pop(stack);
16717 SvREFCNT_dec(only_to_avoid_leaks);
16718 av_push(stack, rhs);
16721 case '!': /* Highest priority, right associative */
16723 /* If what's already at the top of the stack is another '!",
16724 * they just cancel each other out */
16725 if ( (top_ptr = av_fetch(stack, top_index, FALSE))
16726 && (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) == '!'))
16728 only_to_avoid_leaks = av_pop(stack);
16729 SvREFCNT_dec(only_to_avoid_leaks);
16731 else { /* Otherwise, since it's right associative, just push
16733 av_push(stack, newSVuv(curchar));
16738 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16739 if (RExC_parse >= RExC_end) {
16742 vFAIL("Unexpected character");
16746 /* Here 'current' is the operand. If something is already on the
16747 * stack, we have to check if it is a !. But first, the code above
16748 * may have altered the stack in the time since we earlier set
16751 top_index = av_tindex_skip_len_mg(stack);
16752 if (top_index - fence >= 0) {
16753 /* If the top entry on the stack is an operator, it had better
16754 * be a '!', otherwise the entry below the top operand should
16755 * be an operator */
16756 top_ptr = av_fetch(stack, top_index, FALSE);
16758 if (IS_OPERATOR(*top_ptr)) {
16760 /* The only permissible operator at the top of the stack is
16761 * '!', which is applied immediately to this operand. */
16762 curchar = (char) SvUV(*top_ptr);
16763 if (curchar != '!') {
16764 SvREFCNT_dec(current);
16765 vFAIL2("Unexpected binary operator '%c' with no "
16766 "preceding operand", curchar);
16769 _invlist_invert(current);
16771 only_to_avoid_leaks = av_pop(stack);
16772 SvREFCNT_dec(only_to_avoid_leaks);
16774 /* And we redo with the inverted operand. This allows
16775 * handling multiple ! in a row */
16776 goto handle_operand;
16778 /* Single operand is ok only for the non-binary ')'
16780 else if ((top_index - fence == 0 && curchar != ')')
16781 || (top_index - fence > 0
16782 && (! (stacked_ptr = av_fetch(stack,
16785 || IS_OPERAND(*stacked_ptr))))
16787 SvREFCNT_dec(current);
16788 vFAIL("Operand with no preceding operator");
16792 /* Here there was nothing on the stack or the top element was
16793 * another operand. Just add this new one */
16794 av_push(stack, current);
16796 } /* End of switch on next parse token */
16798 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16799 } /* End of loop parsing through the construct */
16801 vFAIL("Syntax error in (?[...])");
16805 if (RExC_parse >= RExC_end || RExC_parse[1] != ')') {
16806 if (RExC_parse < RExC_end) {
16810 vFAIL("Unexpected ']' with no following ')' in (?[...");
16813 if (av_tindex_skip_len_mg(fence_stack) >= 0) {
16814 vFAIL("Unmatched (");
16817 if (av_tindex_skip_len_mg(stack) < 0 /* Was empty */
16818 || ((final = av_pop(stack)) == NULL)
16819 || ! IS_OPERAND(final)
16820 || ! is_invlist(final)
16821 || av_tindex_skip_len_mg(stack) >= 0) /* More left on stack */
16824 SvREFCNT_dec(final);
16825 vFAIL("Incomplete expression within '(?[ ])'");
16828 /* Here, 'final' is the resultant inversion list from evaluating the
16829 * expression. Return it if so requested */
16830 if (return_invlist) {
16831 *return_invlist = final;
16835 if (RExC_sets_depth) { /* If within a recursive call, return in a special
16838 node = regpnode(pRExC_state, REGEX_SET, (void *) final);
16842 /* Otherwise generate a resultant node, based on 'final'. regclass()
16843 * is expecting a string of ranges and individual code points */
16844 invlist_iterinit(final);
16845 result_string = newSVpvs("");
16846 while (invlist_iternext(final, &start, &end)) {
16847 if (start == end) {
16848 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}", start);
16851 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}-\\x{%"
16852 UVXf "}", start, end);
16856 /* About to generate an ANYOF (or similar) node from the inversion list
16857 * we have calculated */
16858 save_parse = RExC_parse;
16859 RExC_parse = SvPV(result_string, len);
16860 save_end = RExC_end;
16861 RExC_end = RExC_parse + len;
16862 TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE;
16864 /* We turn off folding around the call, as the class we have
16865 * constructed already has all folding taken into consideration, and we
16866 * don't want regclass() to add to that */
16867 RExC_flags &= ~RXf_PMf_FOLD;
16868 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if multi-char
16869 * folds are allowed. */
16870 node = regclass(pRExC_state, flagp, depth+1,
16871 FALSE, /* means parse the whole char class */
16872 FALSE, /* don't allow multi-char folds */
16873 TRUE, /* silence non-portable warnings. The above may
16874 very well have generated non-portable code
16875 points, but they're valid on this machine */
16876 FALSE, /* similarly, no need for strict */
16878 /* We can optimize into something besides an ANYOF,
16879 * except under /l, which needs to be ANYOF because of
16880 * runtime checks for locale sanity, etc */
16886 RExC_parse = save_parse + 1;
16887 RExC_end = save_end;
16888 SvREFCNT_dec_NN(final);
16889 SvREFCNT_dec_NN(result_string);
16892 RExC_flags |= RXf_PMf_FOLD;
16896 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16897 goto regclass_failed;
16900 /* Fix up the node type if we are in locale. (We have pretended we are
16901 * under /u for the purposes of regclass(), as this construct will only
16902 * work under UTF-8 locales. But now we change the opcode to be ANYOFL
16903 * (so as to cause any warnings about bad locales to be output in
16904 * regexec.c), and add the flag that indicates to check if not in a
16905 * UTF-8 locale. The reason we above forbid optimization into
16906 * something other than an ANYOF node is simply to minimize the number
16907 * of code changes in regexec.c. Otherwise we would have to create new
16908 * EXACTish node types and deal with them. This decision could be
16909 * revisited should this construct become popular.
16911 * (One might think we could look at the resulting ANYOF node and
16912 * suppress the flag if everything is above 255, as those would be
16913 * UTF-8 only, but this isn't true, as the components that led to that
16914 * result could have been locale-affected, and just happen to cancel
16915 * each other out under UTF-8 locales.) */
16917 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
16919 assert(OP(REGNODE_p(node)) == ANYOF);
16921 OP(REGNODE_p(node)) = ANYOFL;
16922 ANYOF_FLAGS(REGNODE_p(node))
16923 |= ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
16927 nextchar(pRExC_state);
16928 Set_Node_Length(REGNODE_p(node), RExC_parse - oregcomp_parse + 1); /* MJD */
16932 FAIL2("panic: regclass returned failure to handle_sets, " "flags=%#" UVxf,
16936 #ifdef ENABLE_REGEX_SETS_DEBUGGING
16939 S_dump_regex_sets_structures(pTHX_ RExC_state_t *pRExC_state,
16940 AV * stack, const IV fence, AV * fence_stack)
16941 { /* Dumps the stacks in handle_regex_sets() */
16943 const SSize_t stack_top = av_tindex_skip_len_mg(stack);
16944 const SSize_t fence_stack_top = av_tindex_skip_len_mg(fence_stack);
16947 PERL_ARGS_ASSERT_DUMP_REGEX_SETS_STRUCTURES;
16949 PerlIO_printf(Perl_debug_log, "\nParse position is:%s\n", RExC_parse);
16951 if (stack_top < 0) {
16952 PerlIO_printf(Perl_debug_log, "Nothing on stack\n");
16955 PerlIO_printf(Perl_debug_log, "Stack: (fence=%d)\n", (int) fence);
16956 for (i = stack_top; i >= 0; i--) {
16957 SV ** element_ptr = av_fetch(stack, i, FALSE);
16958 if (! element_ptr) {
16961 if (IS_OPERATOR(*element_ptr)) {
16962 PerlIO_printf(Perl_debug_log, "[%d]: %c\n",
16963 (int) i, (int) SvIV(*element_ptr));
16966 PerlIO_printf(Perl_debug_log, "[%d] ", (int) i);
16967 sv_dump(*element_ptr);
16972 if (fence_stack_top < 0) {
16973 PerlIO_printf(Perl_debug_log, "Nothing on fence_stack\n");
16976 PerlIO_printf(Perl_debug_log, "Fence_stack: \n");
16977 for (i = fence_stack_top; i >= 0; i--) {
16978 SV ** element_ptr = av_fetch(fence_stack, i, FALSE);
16979 if (! element_ptr) {
16982 PerlIO_printf(Perl_debug_log, "[%d]: %d\n",
16983 (int) i, (int) SvIV(*element_ptr));
16994 S_add_above_Latin1_folds(pTHX_ RExC_state_t *pRExC_state, const U8 cp, SV** invlist)
16996 /* This adds the Latin1/above-Latin1 folding rules.
16998 * This should be called only for a Latin1-range code points, cp, which is
16999 * known to be involved in a simple fold with other code points above
17000 * Latin1. It would give false results if /aa has been specified.
17001 * Multi-char folds are outside the scope of this, and must be handled
17004 PERL_ARGS_ASSERT_ADD_ABOVE_LATIN1_FOLDS;
17006 assert(HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(cp));
17008 /* The rules that are valid for all Unicode versions are hard-coded in */
17013 add_cp_to_invlist(*invlist, KELVIN_SIGN);
17017 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_LONG_S);
17020 *invlist = add_cp_to_invlist(*invlist, GREEK_CAPITAL_LETTER_MU);
17021 *invlist = add_cp_to_invlist(*invlist, GREEK_SMALL_LETTER_MU);
17023 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
17024 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
17025 *invlist = add_cp_to_invlist(*invlist, ANGSTROM_SIGN);
17027 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
17028 *invlist = add_cp_to_invlist(*invlist,
17029 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
17032 default: /* Other code points are checked against the data for the
17033 current Unicode version */
17035 Size_t folds_count;
17037 const U32 * remaining_folds;
17041 folded_cp = toFOLD(cp);
17044 U8 dummy_fold[UTF8_MAXBYTES_CASE+1];
17046 folded_cp = _to_fold_latin1(cp, dummy_fold, &dummy_len, 0);
17049 if (folded_cp > 255) {
17050 *invlist = add_cp_to_invlist(*invlist, folded_cp);
17053 folds_count = _inverse_folds(folded_cp, &first_fold,
17055 if (folds_count == 0) {
17057 /* Use deprecated warning to increase the chances of this being
17059 ckWARN2reg_d(RExC_parse,
17060 "Perl folding rules are not up-to-date for 0x%02X;"
17061 " please use the perlbug utility to report;", cp);
17066 if (first_fold > 255) {
17067 *invlist = add_cp_to_invlist(*invlist, first_fold);
17069 for (i = 0; i < folds_count - 1; i++) {
17070 if (remaining_folds[i] > 255) {
17071 *invlist = add_cp_to_invlist(*invlist,
17072 remaining_folds[i]);
17082 S_output_posix_warnings(pTHX_ RExC_state_t *pRExC_state, AV* posix_warnings)
17084 /* Output the elements of the array given by '*posix_warnings' as REGEXP
17088 const bool first_is_fatal = ckDEAD(packWARN(WARN_REGEXP));
17090 PERL_ARGS_ASSERT_OUTPUT_POSIX_WARNINGS;
17092 if (! TO_OUTPUT_WARNINGS(RExC_parse)) {
17093 CLEAR_POSIX_WARNINGS();
17097 while ((msg = av_shift(posix_warnings)) != &PL_sv_undef) {
17098 if (first_is_fatal) { /* Avoid leaking this */
17099 av_undef(posix_warnings); /* This isn't necessary if the
17100 array is mortal, but is a
17102 (void) sv_2mortal(msg);
17105 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s", SvPVX(msg));
17106 SvREFCNT_dec_NN(msg);
17109 UPDATE_WARNINGS_LOC(RExC_parse);
17112 PERL_STATIC_INLINE Size_t
17113 S_find_first_differing_byte_pos(const U8 * s1, const U8 * s2, const Size_t max)
17115 const U8 * const start = s1;
17116 const U8 * const send = start + max;
17118 PERL_ARGS_ASSERT_FIND_FIRST_DIFFERING_BYTE_POS;
17120 while (s1 < send && *s1 == *s2) {
17129 S_add_multi_match(pTHX_ AV* multi_char_matches, SV* multi_string, const STRLEN cp_count)
17131 /* This adds the string scalar <multi_string> to the array
17132 * <multi_char_matches>. <multi_string> is known to have exactly
17133 * <cp_count> code points in it. This is used when constructing a
17134 * bracketed character class and we find something that needs to match more
17135 * than a single character.
17137 * <multi_char_matches> is actually an array of arrays. Each top-level
17138 * element is an array that contains all the strings known so far that are
17139 * the same length. And that length (in number of code points) is the same
17140 * as the index of the top-level array. Hence, the [2] element is an
17141 * array, each element thereof is a string containing TWO code points;
17142 * while element [3] is for strings of THREE characters, and so on. Since
17143 * this is for multi-char strings there can never be a [0] nor [1] element.
17145 * When we rewrite the character class below, we will do so such that the
17146 * longest strings are written first, so that it prefers the longest
17147 * matching strings first. This is done even if it turns out that any
17148 * quantifier is non-greedy, out of this programmer's (khw) laziness. Tom
17149 * Christiansen has agreed that this is ok. This makes the test for the
17150 * ligature 'ffi' come before the test for 'ff', for example */
17153 AV** this_array_ptr;
17155 PERL_ARGS_ASSERT_ADD_MULTI_MATCH;
17157 if (! multi_char_matches) {
17158 multi_char_matches = newAV();
17161 if (av_exists(multi_char_matches, cp_count)) {
17162 this_array_ptr = (AV**) av_fetch(multi_char_matches, cp_count, FALSE);
17163 this_array = *this_array_ptr;
17166 this_array = newAV();
17167 av_store(multi_char_matches, cp_count,
17170 av_push(this_array, multi_string);
17172 return multi_char_matches;
17175 /* The names of properties whose definitions are not known at compile time are
17176 * stored in this SV, after a constant heading. So if the length has been
17177 * changed since initialization, then there is a run-time definition. */
17178 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
17179 (SvCUR(listsv) != initial_listsv_len)
17181 /* There is a restricted set of white space characters that are legal when
17182 * ignoring white space in a bracketed character class. This generates the
17183 * code to skip them.
17185 * There is a line below that uses the same white space criteria but is outside
17186 * this macro. Both here and there must use the same definition */
17187 #define SKIP_BRACKETED_WHITE_SPACE(do_skip, p) \
17190 while (isBLANK_A(UCHARAT(p))) \
17197 STATIC regnode_offset
17198 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
17199 const bool stop_at_1, /* Just parse the next thing, don't
17200 look for a full character class */
17201 bool allow_mutiple_chars,
17202 const bool silence_non_portable, /* Don't output warnings
17206 bool optimizable, /* ? Allow a non-ANYOF return
17208 SV** ret_invlist /* Return an inversion list, not a node */
17211 /* parse a bracketed class specification. Most of these will produce an
17212 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
17213 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
17214 * under /i with multi-character folds: it will be rewritten following the
17215 * paradigm of this example, where the <multi-fold>s are characters which
17216 * fold to multiple character sequences:
17217 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
17218 * gets effectively rewritten as:
17219 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
17220 * reg() gets called (recursively) on the rewritten version, and this
17221 * function will return what it constructs. (Actually the <multi-fold>s
17222 * aren't physically removed from the [abcdefghi], it's just that they are
17223 * ignored in the recursion by means of a flag:
17224 * <RExC_in_multi_char_class>.)
17226 * ANYOF nodes contain a bit map for the first NUM_ANYOF_CODE_POINTS
17227 * characters, with the corresponding bit set if that character is in the
17228 * list. For characters above this, an inversion list is used. There
17229 * are extra bits for \w, etc. in locale ANYOFs, as what these match is not
17230 * determinable at compile time
17232 * On success, returns the offset at which any next node should be placed
17233 * into the regex engine program being compiled.
17235 * Returns 0 otherwise, setting flagp to RESTART_PARSE if the parse needs
17236 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to
17241 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
17243 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
17244 regnode_offset ret = -1; /* Initialized to an illegal value */
17246 int namedclass = OOB_NAMEDCLASS;
17247 char *rangebegin = NULL;
17248 SV *listsv = NULL; /* List of \p{user-defined} whose definitions
17249 aren't available at the time this was called */
17250 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
17251 than just initialized. */
17252 SV* properties = NULL; /* Code points that match \p{} \P{} */
17253 SV* posixes = NULL; /* Code points that match classes like [:word:],
17254 extended beyond the Latin1 range. These have to
17255 be kept separate from other code points for much
17256 of this function because their handling is
17257 different under /i, and for most classes under
17259 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
17260 separate for a while from the non-complemented
17261 versions because of complications with /d
17263 SV* simple_posixes = NULL; /* But under some conditions, the classes can be
17264 treated more simply than the general case,
17265 leading to less compilation and execution
17267 UV element_count = 0; /* Number of distinct elements in the class.
17268 Optimizations may be possible if this is tiny */
17269 AV * multi_char_matches = NULL; /* Code points that fold to more than one
17270 character; used under /i */
17272 char * stop_ptr = RExC_end; /* where to stop parsing */
17274 /* ignore unescaped whitespace? */
17275 const bool skip_white = cBOOL( ret_invlist
17276 || (RExC_flags & RXf_PMf_EXTENDED_MORE));
17278 /* inversion list of code points this node matches only when the target
17279 * string is in UTF-8. These are all non-ASCII, < 256. (Because is under
17281 SV* upper_latin1_only_utf8_matches = NULL;
17283 /* Inversion list of code points this node matches regardless of things
17284 * like locale, folding, utf8ness of the target string */
17285 SV* cp_list = NULL;
17287 /* Like cp_list, but code points on this list need to be checked for things
17288 * that fold to/from them under /i */
17289 SV* cp_foldable_list = NULL;
17291 /* Like cp_list, but code points on this list are valid only when the
17292 * runtime locale is UTF-8 */
17293 SV* only_utf8_locale_list = NULL;
17295 /* In a range, if one of the endpoints is non-character-set portable,
17296 * meaning that it hard-codes a code point that may mean a different
17297 * charactger in ASCII vs. EBCDIC, as opposed to, say, a literal 'A' or a
17298 * mnemonic '\t' which each mean the same character no matter which
17299 * character set the platform is on. */
17300 unsigned int non_portable_endpoint = 0;
17302 /* Is the range unicode? which means on a platform that isn't 1-1 native
17303 * to Unicode (i.e. non-ASCII), each code point in it should be considered
17304 * to be a Unicode value. */
17305 bool unicode_range = FALSE;
17306 bool invert = FALSE; /* Is this class to be complemented */
17308 bool warn_super = ALWAYS_WARN_SUPER;
17310 const char * orig_parse = RExC_parse;
17312 /* This variable is used to mark where the end in the input is of something
17313 * that looks like a POSIX construct but isn't. During the parse, when
17314 * something looks like it could be such a construct is encountered, it is
17315 * checked for being one, but not if we've already checked this area of the
17316 * input. Only after this position is reached do we check again */
17317 char *not_posix_region_end = RExC_parse - 1;
17319 AV* posix_warnings = NULL;
17320 const bool do_posix_warnings = ckWARN(WARN_REGEXP);
17321 U8 op = END; /* The returned node-type, initialized to an impossible
17323 U8 anyof_flags = 0; /* flag bits if the node is an ANYOF-type */
17324 U32 posixl = 0; /* bit field of posix classes matched under /l */
17327 /* Flags as to what things aren't knowable until runtime. (Note that these are
17328 * mutually exclusive.) */
17329 #define HAS_USER_DEFINED_PROPERTY 0x01 /* /u any user-defined properties that
17330 haven't been defined as of yet */
17331 #define HAS_D_RUNTIME_DEPENDENCY 0x02 /* /d if the target being matched is
17333 #define HAS_L_RUNTIME_DEPENDENCY 0x04 /* /l what the posix classes match and
17334 what gets folded */
17335 U32 has_runtime_dependency = 0; /* OR of the above flags */
17337 DECLARE_AND_GET_RE_DEBUG_FLAGS;
17339 PERL_ARGS_ASSERT_REGCLASS;
17341 PERL_UNUSED_ARG(depth);
17345 /* If wants an inversion list returned, we can't optimize to something
17348 optimizable = FALSE;
17351 DEBUG_PARSE("clas");
17353 #if UNICODE_MAJOR_VERSION < 3 /* no multifolds in early Unicode */ \
17354 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0 \
17355 && UNICODE_DOT_DOT_VERSION == 0)
17356 allow_mutiple_chars = FALSE;
17359 /* We include the /i status at the beginning of this so that we can
17360 * know it at runtime */
17361 listsv = sv_2mortal(Perl_newSVpvf(aTHX_ "#%d\n", cBOOL(FOLD)));
17362 initial_listsv_len = SvCUR(listsv);
17363 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
17365 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
17367 assert(RExC_parse <= RExC_end);
17369 if (UCHARAT(RExC_parse) == '^') { /* Complement the class */
17372 allow_mutiple_chars = FALSE;
17374 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
17377 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
17378 if (! ret_invlist && MAYBE_POSIXCC(UCHARAT(RExC_parse))) {
17379 int maybe_class = handle_possible_posix(pRExC_state,
17381 ¬_posix_region_end,
17383 TRUE /* checking only */);
17384 if (maybe_class >= OOB_NAMEDCLASS && do_posix_warnings) {
17385 ckWARN4reg(not_posix_region_end,
17386 "POSIX syntax [%c %c] belongs inside character classes%s",
17387 *RExC_parse, *RExC_parse,
17388 (maybe_class == OOB_NAMEDCLASS)
17389 ? ((POSIXCC_NOTYET(*RExC_parse))
17390 ? " (but this one isn't implemented)"
17391 : " (but this one isn't fully valid)")
17397 /* If the caller wants us to just parse a single element, accomplish this
17398 * by faking the loop ending condition */
17399 if (stop_at_1 && RExC_end > RExC_parse) {
17400 stop_ptr = RExC_parse + 1;
17403 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
17404 if (UCHARAT(RExC_parse) == ']')
17405 goto charclassloop;
17409 if ( posix_warnings
17410 && av_tindex_skip_len_mg(posix_warnings) >= 0
17411 && RExC_parse > not_posix_region_end)
17413 /* Warnings about posix class issues are considered tentative until
17414 * we are far enough along in the parse that we can no longer
17415 * change our mind, at which point we output them. This is done
17416 * each time through the loop so that a later class won't zap them
17417 * before they have been dealt with. */
17418 output_posix_warnings(pRExC_state, posix_warnings);
17421 if (RExC_parse >= stop_ptr) {
17425 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
17427 if (UCHARAT(RExC_parse) == ']') {
17433 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
17434 save_value = value;
17435 save_prevvalue = prevvalue;
17438 rangebegin = RExC_parse;
17440 non_portable_endpoint = 0;
17442 if (UTF && ! UTF8_IS_INVARIANT(* RExC_parse)) {
17443 value = utf8n_to_uvchr((U8*)RExC_parse,
17444 RExC_end - RExC_parse,
17445 &numlen, UTF8_ALLOW_DEFAULT);
17446 RExC_parse += numlen;
17449 value = UCHARAT(RExC_parse++);
17451 if (value == '[') {
17452 char * posix_class_end;
17453 namedclass = handle_possible_posix(pRExC_state,
17456 do_posix_warnings ? &posix_warnings : NULL,
17457 FALSE /* die if error */);
17458 if (namedclass > OOB_NAMEDCLASS) {
17460 /* If there was an earlier attempt to parse this particular
17461 * posix class, and it failed, it was a false alarm, as this
17462 * successful one proves */
17463 if ( posix_warnings
17464 && av_tindex_skip_len_mg(posix_warnings) >= 0
17465 && not_posix_region_end >= RExC_parse
17466 && not_posix_region_end <= posix_class_end)
17468 av_undef(posix_warnings);
17471 RExC_parse = posix_class_end;
17473 else if (namedclass == OOB_NAMEDCLASS) {
17474 not_posix_region_end = posix_class_end;
17477 namedclass = OOB_NAMEDCLASS;
17480 else if ( RExC_parse - 1 > not_posix_region_end
17481 && MAYBE_POSIXCC(value))
17483 (void) handle_possible_posix(
17485 RExC_parse - 1, /* -1 because parse has already been
17487 ¬_posix_region_end,
17488 do_posix_warnings ? &posix_warnings : NULL,
17489 TRUE /* checking only */);
17491 else if ( strict && ! skip_white
17492 && ( _generic_isCC(value, _CC_VERTSPACE)
17493 || is_VERTWS_cp_high(value)))
17495 vFAIL("Literal vertical space in [] is illegal except under /x");
17497 else if (value == '\\') {
17498 /* Is a backslash; get the code point of the char after it */
17500 if (RExC_parse >= RExC_end) {
17501 vFAIL("Unmatched [");
17504 if (UTF && ! UTF8_IS_INVARIANT(UCHARAT(RExC_parse))) {
17505 value = utf8n_to_uvchr((U8*)RExC_parse,
17506 RExC_end - RExC_parse,
17507 &numlen, UTF8_ALLOW_DEFAULT);
17508 RExC_parse += numlen;
17511 value = UCHARAT(RExC_parse++);
17513 /* Some compilers cannot handle switching on 64-bit integer
17514 * values, therefore value cannot be an UV. Yes, this will
17515 * be a problem later if we want switch on Unicode.
17516 * A similar issue a little bit later when switching on
17517 * namedclass. --jhi */
17519 /* If the \ is escaping white space when white space is being
17520 * skipped, it means that that white space is wanted literally, and
17521 * is already in 'value'. Otherwise, need to translate the escape
17522 * into what it signifies. */
17523 if (! skip_white || ! isBLANK_A(value)) switch ((I32)value) {
17524 const char * message;
17528 case 'w': namedclass = ANYOF_WORDCHAR; break;
17529 case 'W': namedclass = ANYOF_NWORDCHAR; break;
17530 case 's': namedclass = ANYOF_SPACE; break;
17531 case 'S': namedclass = ANYOF_NSPACE; break;
17532 case 'd': namedclass = ANYOF_DIGIT; break;
17533 case 'D': namedclass = ANYOF_NDIGIT; break;
17534 case 'v': namedclass = ANYOF_VERTWS; break;
17535 case 'V': namedclass = ANYOF_NVERTWS; break;
17536 case 'h': namedclass = ANYOF_HORIZWS; break;
17537 case 'H': namedclass = ANYOF_NHORIZWS; break;
17538 case 'N': /* Handle \N{NAME} in class */
17540 const char * const backslash_N_beg = RExC_parse - 2;
17543 if (! grok_bslash_N(pRExC_state,
17544 NULL, /* No regnode */
17545 &value, /* Yes single value */
17546 &cp_count, /* Multiple code pt count */
17552 if (*flagp & NEED_UTF8)
17553 FAIL("panic: grok_bslash_N set NEED_UTF8");
17555 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
17557 if (cp_count < 0) {
17558 vFAIL("\\N in a character class must be a named character: \\N{...}");
17560 else if (cp_count == 0) {
17561 ckWARNreg(RExC_parse,
17562 "Ignoring zero length \\N{} in character class");
17564 else { /* cp_count > 1 */
17565 assert(cp_count > 1);
17566 if (! RExC_in_multi_char_class) {
17567 if ( ! allow_mutiple_chars
17570 || *RExC_parse == '-')
17574 vFAIL("\\N{} here is restricted to one character");
17576 ckWARNreg(RExC_parse, "Using just the first character returned by \\N{} in character class");
17577 break; /* <value> contains the first code
17578 point. Drop out of the switch to
17582 SV * multi_char_N = newSVpvn(backslash_N_beg,
17583 RExC_parse - backslash_N_beg);
17585 = add_multi_match(multi_char_matches,
17590 } /* End of cp_count != 1 */
17592 /* This element should not be processed further in this
17595 value = save_value;
17596 prevvalue = save_prevvalue;
17597 continue; /* Back to top of loop to get next char */
17600 /* Here, is a single code point, and <value> contains it */
17601 unicode_range = TRUE; /* \N{} are Unicode */
17609 if (RExC_pm_flags & PMf_WILDCARD) {
17611 /* diag_listed_as: Use of %s is not allowed in Unicode
17612 property wildcard subpatterns in regex; marked by <--
17614 vFAIL3("Use of '\\%c%c' is not allowed in Unicode property"
17615 " wildcard subpatterns", (char) value, *(RExC_parse - 1));
17618 /* \p means they want Unicode semantics */
17619 REQUIRE_UNI_RULES(flagp, 0);
17621 if (RExC_parse >= RExC_end)
17622 vFAIL2("Empty \\%c", (U8)value);
17623 if (*RExC_parse == '{') {
17624 const U8 c = (U8)value;
17625 e = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
17628 vFAIL2("Missing right brace on \\%c{}", c);
17633 /* White space is allowed adjacent to the braces and after
17634 * any '^', even when not under /x */
17635 while (isSPACE(*RExC_parse)) {
17639 if (UCHARAT(RExC_parse) == '^') {
17641 /* toggle. (The rhs xor gets the single bit that
17642 * differs between P and p; the other xor inverts just
17644 value ^= 'P' ^ 'p';
17647 while (isSPACE(*RExC_parse)) {
17652 if (e == RExC_parse)
17653 vFAIL2("Empty \\%c{}", c);
17655 n = e - RExC_parse;
17656 while (isSPACE(*(RExC_parse + n - 1)))
17659 } /* The \p isn't immediately followed by a '{' */
17660 else if (! isALPHA(*RExC_parse)) {
17661 RExC_parse += (UTF)
17662 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
17664 vFAIL2("Character following \\%c must be '{' or a "
17665 "single-character Unicode property name",
17673 char* name = RExC_parse;
17675 /* Any message returned about expanding the definition */
17676 SV* msg = newSVpvs_flags("", SVs_TEMP);
17678 /* If set TRUE, the property is user-defined as opposed to
17679 * official Unicode */
17680 bool user_defined = FALSE;
17682 SV * prop_definition = parse_uniprop_string(
17683 name, n, UTF, FOLD,
17684 FALSE, /* This is compile-time */
17686 /* We can't defer this defn when
17687 * the full result is required in
17689 ! cBOOL(ret_invlist),
17695 if (SvCUR(msg)) { /* Assumes any error causes a msg */
17696 assert(prop_definition == NULL);
17697 RExC_parse = e + 1;
17698 if (SvUTF8(msg)) { /* msg being UTF-8 makes the whole
17699 thing so, or else the display is
17703 /* diag_listed_as: Can't find Unicode property definition "%s" in regex; marked by <-- HERE in m/%s/ */
17704 vFAIL2utf8f("%" UTF8f, UTF8fARG(SvUTF8(msg),
17705 SvCUR(msg), SvPVX(msg)));
17708 if (! is_invlist(prop_definition)) {
17710 /* Here, the definition isn't known, so we have gotten
17711 * returned a string that will be evaluated if and when
17712 * encountered at runtime. We add it to the list of
17713 * such properties, along with whether it should be
17714 * complemented or not */
17715 if (value == 'P') {
17716 sv_catpvs(listsv, "!");
17719 sv_catpvs(listsv, "+");
17721 sv_catsv(listsv, prop_definition);
17723 has_runtime_dependency |= HAS_USER_DEFINED_PROPERTY;
17725 /* We don't know yet what this matches, so have to flag
17727 anyof_flags |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
17730 assert (prop_definition && is_invlist(prop_definition));
17732 /* Here we do have the complete property definition
17734 * Temporary workaround for [perl #133136]. For this
17735 * precise input that is in the .t that is failing,
17736 * load utf8.pm, which is what the test wants, so that
17737 * that .t passes */
17738 if ( memEQs(RExC_start, e + 1 - RExC_start,
17740 && ! hv_common(GvHVn(PL_incgv),
17742 "utf8.pm", sizeof("utf8.pm") - 1,
17743 0, HV_FETCH_ISEXISTS, NULL, 0))
17745 require_pv("utf8.pm");
17748 if (! user_defined &&
17749 /* We warn on matching an above-Unicode code point
17750 * if the match would return true, except don't
17751 * warn for \p{All}, which has exactly one element
17753 (_invlist_contains_cp(prop_definition, 0x110000)
17754 && (! (_invlist_len(prop_definition) == 1
17755 && *invlist_array(prop_definition) == 0))))
17760 /* Invert if asking for the complement */
17761 if (value == 'P') {
17762 _invlist_union_complement_2nd(properties,
17767 _invlist_union(properties, prop_definition, &properties);
17772 RExC_parse = e + 1;
17773 namedclass = ANYOF_UNIPROP; /* no official name, but it's
17777 case 'n': value = '\n'; break;
17778 case 'r': value = '\r'; break;
17779 case 't': value = '\t'; break;
17780 case 'f': value = '\f'; break;
17781 case 'b': value = '\b'; break;
17782 case 'e': value = ESC_NATIVE; break;
17783 case 'a': value = '\a'; break;
17785 RExC_parse--; /* function expects to be pointed at the 'o' */
17786 if (! grok_bslash_o(&RExC_parse,
17792 cBOOL(range), /* MAX_UV allowed for range
17798 else if (message && TO_OUTPUT_WARNINGS(RExC_parse)) {
17799 warn_non_literal_string(RExC_parse, packed_warn, message);
17803 non_portable_endpoint++;
17807 RExC_parse--; /* function expects to be pointed at the 'x' */
17808 if (! grok_bslash_x(&RExC_parse,
17814 cBOOL(range), /* MAX_UV allowed for range
17820 else if (message && TO_OUTPUT_WARNINGS(RExC_parse)) {
17821 warn_non_literal_string(RExC_parse, packed_warn, message);
17825 non_portable_endpoint++;
17829 if (! grok_bslash_c(*RExC_parse, &grok_c_char, &message,
17832 /* going to die anyway; point to exact spot of
17834 RExC_parse += (UTF)
17835 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
17840 value = grok_c_char;
17842 if (message && TO_OUTPUT_WARNINGS(RExC_parse)) {
17843 warn_non_literal_string(RExC_parse, packed_warn, message);
17846 non_portable_endpoint++;
17848 case '0': case '1': case '2': case '3': case '4':
17849 case '5': case '6': case '7':
17851 /* Take 1-3 octal digits */
17852 I32 flags = PERL_SCAN_SILENT_ILLDIGIT
17853 | PERL_SCAN_NOTIFY_ILLDIGIT;
17854 numlen = (strict) ? 4 : 3;
17855 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
17856 RExC_parse += numlen;
17859 RExC_parse += (UTF)
17860 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
17862 vFAIL("Need exactly 3 octal digits");
17864 else if ( (flags & PERL_SCAN_NOTIFY_ILLDIGIT)
17865 && RExC_parse < RExC_end
17866 && isDIGIT(*RExC_parse)
17867 && ckWARN(WARN_REGEXP))
17869 reg_warn_non_literal_string(
17871 form_alien_digit_msg(8, numlen, RExC_parse,
17872 RExC_end, UTF, FALSE));
17876 non_portable_endpoint++;
17881 /* Allow \_ to not give an error */
17882 if (isWORDCHAR(value) && value != '_') {
17884 vFAIL2("Unrecognized escape \\%c in character class",
17888 ckWARN2reg(RExC_parse,
17889 "Unrecognized escape \\%c in character class passed through",
17894 } /* End of switch on char following backslash */
17895 } /* end of handling backslash escape sequences */
17897 /* Here, we have the current token in 'value' */
17899 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
17902 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
17903 * literal, as is the character that began the false range, i.e.
17904 * the 'a' in the examples */
17906 const int w = (RExC_parse >= rangebegin)
17907 ? RExC_parse - rangebegin
17911 "False [] range \"%" UTF8f "\"",
17912 UTF8fARG(UTF, w, rangebegin));
17915 ckWARN2reg(RExC_parse,
17916 "False [] range \"%" UTF8f "\"",
17917 UTF8fARG(UTF, w, rangebegin));
17918 cp_list = add_cp_to_invlist(cp_list, '-');
17919 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
17923 range = 0; /* this was not a true range */
17924 element_count += 2; /* So counts for three values */
17927 classnum = namedclass_to_classnum(namedclass);
17929 if (LOC && namedclass < ANYOF_POSIXL_MAX
17930 #ifndef HAS_ISASCII
17931 && classnum != _CC_ASCII
17934 SV* scratch_list = NULL;
17936 /* What the Posix classes (like \w, [:space:]) match isn't
17937 * generally knowable under locale until actual match time. A
17938 * special node is used for these which has extra space for a
17939 * bitmap, with a bit reserved for each named class that is to
17940 * be matched against. (This isn't needed for \p{} and
17941 * pseudo-classes, as they are not affected by locale, and
17942 * hence are dealt with separately.) However, if a named class
17943 * and its complement are both present, then it matches
17944 * everything, and there is no runtime dependency. Odd numbers
17945 * are the complements of the next lower number, so xor works.
17946 * (Note that something like [\w\D] should match everything,
17947 * because \d should be a proper subset of \w. But rather than
17948 * trust that the locale is well behaved, we leave this to
17949 * runtime to sort out) */
17950 if (POSIXL_TEST(posixl, namedclass ^ 1)) {
17951 cp_list = _add_range_to_invlist(cp_list, 0, UV_MAX);
17952 POSIXL_ZERO(posixl);
17953 has_runtime_dependency &= ~HAS_L_RUNTIME_DEPENDENCY;
17954 anyof_flags &= ~ANYOF_MATCHES_POSIXL;
17955 continue; /* We could ignore the rest of the class, but
17956 best to parse it for any errors */
17958 else { /* Here, isn't the complement of any already parsed
17960 POSIXL_SET(posixl, namedclass);
17961 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
17962 anyof_flags |= ANYOF_MATCHES_POSIXL;
17964 /* The above-Latin1 characters are not subject to locale
17965 * rules. Just add them to the unconditionally-matched
17968 /* Get the list of the above-Latin1 code points this
17970 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
17971 PL_XPosix_ptrs[classnum],
17973 /* Odd numbers are complements,
17974 * like NDIGIT, NASCII, ... */
17975 namedclass % 2 != 0,
17977 /* Checking if 'cp_list' is NULL first saves an extra
17978 * clone. Its reference count will be decremented at the
17979 * next union, etc, or if this is the only instance, at the
17980 * end of the routine */
17982 cp_list = scratch_list;
17985 _invlist_union(cp_list, scratch_list, &cp_list);
17986 SvREFCNT_dec_NN(scratch_list);
17988 continue; /* Go get next character */
17993 /* Here, is not /l, or is a POSIX class for which /l doesn't
17994 * matter (or is a Unicode property, which is skipped here). */
17995 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
17996 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
17998 /* Here, should be \h, \H, \v, or \V. None of /d, /i
17999 * nor /l make a difference in what these match,
18000 * therefore we just add what they match to cp_list. */
18001 if (classnum != _CC_VERTSPACE) {
18002 assert( namedclass == ANYOF_HORIZWS
18003 || namedclass == ANYOF_NHORIZWS);
18005 /* It turns out that \h is just a synonym for
18007 classnum = _CC_BLANK;
18010 _invlist_union_maybe_complement_2nd(
18012 PL_XPosix_ptrs[classnum],
18013 namedclass % 2 != 0, /* Complement if odd
18014 (NHORIZWS, NVERTWS)
18019 else if ( AT_LEAST_UNI_SEMANTICS
18020 || classnum == _CC_ASCII
18021 || (DEPENDS_SEMANTICS && ( classnum == _CC_DIGIT
18022 || classnum == _CC_XDIGIT)))
18024 /* We usually have to worry about /d affecting what POSIX
18025 * classes match, with special code needed because we won't
18026 * know until runtime what all matches. But there is no
18027 * extra work needed under /u and /a; and [:ascii:] is
18028 * unaffected by /d; and :digit: and :xdigit: don't have
18029 * runtime differences under /d. So we can special case
18030 * these, and avoid some extra work below, and at runtime.
18032 _invlist_union_maybe_complement_2nd(
18034 ((AT_LEAST_ASCII_RESTRICTED)
18035 ? PL_Posix_ptrs[classnum]
18036 : PL_XPosix_ptrs[classnum]),
18037 namedclass % 2 != 0,
18040 else { /* Garden variety class. If is NUPPER, NALPHA, ...
18041 complement and use nposixes */
18042 SV** posixes_ptr = namedclass % 2 == 0
18045 _invlist_union_maybe_complement_2nd(
18047 PL_XPosix_ptrs[classnum],
18048 namedclass % 2 != 0,
18052 } /* end of namedclass \blah */
18054 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
18056 /* If 'range' is set, 'value' is the ending of a range--check its
18057 * validity. (If value isn't a single code point in the case of a
18058 * range, we should have figured that out above in the code that
18059 * catches false ranges). Later, we will handle each individual code
18060 * point in the range. If 'range' isn't set, this could be the
18061 * beginning of a range, so check for that by looking ahead to see if
18062 * the next real character to be processed is the range indicator--the
18067 /* For unicode ranges, we have to test that the Unicode as opposed
18068 * to the native values are not decreasing. (Above 255, there is
18069 * no difference between native and Unicode) */
18070 if (unicode_range && prevvalue < 255 && value < 255) {
18071 if (NATIVE_TO_LATIN1(prevvalue) > NATIVE_TO_LATIN1(value)) {
18072 goto backwards_range;
18077 if (prevvalue > value) /* b-a */ {
18082 w = RExC_parse - rangebegin;
18084 "Invalid [] range \"%" UTF8f "\"",
18085 UTF8fARG(UTF, w, rangebegin));
18086 NOT_REACHED; /* NOTREACHED */
18090 prevvalue = value; /* save the beginning of the potential range */
18091 if (! stop_at_1 /* Can't be a range if parsing just one thing */
18092 && *RExC_parse == '-')
18094 char* next_char_ptr = RExC_parse + 1;
18096 /* Get the next real char after the '-' */
18097 SKIP_BRACKETED_WHITE_SPACE(skip_white, next_char_ptr);
18099 /* If the '-' is at the end of the class (just before the ']',
18100 * it is a literal minus; otherwise it is a range */
18101 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
18102 RExC_parse = next_char_ptr;
18104 /* a bad range like \w-, [:word:]- ? */
18105 if (namedclass > OOB_NAMEDCLASS) {
18106 if (strict || ckWARN(WARN_REGEXP)) {
18107 const int w = RExC_parse >= rangebegin
18108 ? RExC_parse - rangebegin
18111 vFAIL4("False [] range \"%*.*s\"",
18116 "False [] range \"%*.*s\"",
18120 cp_list = add_cp_to_invlist(cp_list, '-');
18123 range = 1; /* yeah, it's a range! */
18124 continue; /* but do it the next time */
18129 if (namedclass > OOB_NAMEDCLASS) {
18133 /* Here, we have a single value this time through the loop, and
18134 * <prevvalue> is the beginning of the range, if any; or <value> if
18137 /* non-Latin1 code point implies unicode semantics. */
18139 if (value > MAX_LEGAL_CP && ( value != UV_MAX
18140 || prevvalue > MAX_LEGAL_CP))
18142 vFAIL(form_cp_too_large_msg(16, NULL, 0, value));
18144 REQUIRE_UNI_RULES(flagp, 0);
18145 if ( ! silence_non_portable
18146 && UNICODE_IS_PERL_EXTENDED(value)
18147 && TO_OUTPUT_WARNINGS(RExC_parse))
18149 ckWARN2_non_literal_string(RExC_parse,
18150 packWARN(WARN_PORTABLE),
18151 PL_extended_cp_format,
18156 /* Ready to process either the single value, or the completed range.
18157 * For single-valued non-inverted ranges, we consider the possibility
18158 * of multi-char folds. (We made a conscious decision to not do this
18159 * for the other cases because it can often lead to non-intuitive
18160 * results. For example, you have the peculiar case that:
18161 * "s s" =~ /^[^\xDF]+$/i => Y
18162 * "ss" =~ /^[^\xDF]+$/i => N
18164 * See [perl #89750] */
18165 if (FOLD && allow_mutiple_chars && value == prevvalue) {
18166 if ( value == LATIN_SMALL_LETTER_SHARP_S
18167 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
18170 /* Here <value> is indeed a multi-char fold. Get what it is */
18172 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
18175 UV folded = _to_uni_fold_flags(
18179 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
18180 ? FOLD_FLAGS_NOMIX_ASCII
18184 /* Here, <folded> should be the first character of the
18185 * multi-char fold of <value>, with <foldbuf> containing the
18186 * whole thing. But, if this fold is not allowed (because of
18187 * the flags), <fold> will be the same as <value>, and should
18188 * be processed like any other character, so skip the special
18190 if (folded != value) {
18192 /* Skip if we are recursed, currently parsing the class
18193 * again. Otherwise add this character to the list of
18194 * multi-char folds. */
18195 if (! RExC_in_multi_char_class) {
18196 STRLEN cp_count = utf8_length(foldbuf,
18197 foldbuf + foldlen);
18198 SV* multi_fold = sv_2mortal(newSVpvs(""));
18200 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%" UVXf "}", value);
18203 = add_multi_match(multi_char_matches,
18209 /* This element should not be processed further in this
18212 value = save_value;
18213 prevvalue = save_prevvalue;
18219 if (strict && ckWARN(WARN_REGEXP)) {
18222 /* If the range starts above 255, everything is portable and
18223 * likely to be so for any forseeable character set, so don't
18225 if (unicode_range && non_portable_endpoint && prevvalue < 256) {
18226 vWARN(RExC_parse, "Both or neither range ends should be Unicode");
18228 else if (prevvalue != value) {
18230 /* Under strict, ranges that stop and/or end in an ASCII
18231 * printable should have each end point be a portable value
18232 * for it (preferably like 'A', but we don't warn if it is
18233 * a (portable) Unicode name or code point), and the range
18234 * must be be all digits or all letters of the same case.
18235 * Otherwise, the range is non-portable and unclear as to
18236 * what it contains */
18237 if ( (isPRINT_A(prevvalue) || isPRINT_A(value))
18238 && ( non_portable_endpoint
18239 || ! ( (isDIGIT_A(prevvalue) && isDIGIT_A(value))
18240 || (isLOWER_A(prevvalue) && isLOWER_A(value))
18241 || (isUPPER_A(prevvalue) && isUPPER_A(value))
18243 vWARN(RExC_parse, "Ranges of ASCII printables should"
18244 " be some subset of \"0-9\","
18245 " \"A-Z\", or \"a-z\"");
18247 else if (prevvalue >= FIRST_NON_ASCII_DECIMAL_DIGIT) {
18248 SSize_t index_start;
18249 SSize_t index_final;
18251 /* But the nature of Unicode and languages mean we
18252 * can't do the same checks for above-ASCII ranges,
18253 * except in the case of digit ones. These should
18254 * contain only digits from the same group of 10. The
18255 * ASCII case is handled just above. Hence here, the
18256 * range could be a range of digits. First some
18257 * unlikely special cases. Grandfather in that a range
18258 * ending in 19DA (NEW TAI LUE THAM DIGIT ONE) is bad
18259 * if its starting value is one of the 10 digits prior
18260 * to it. This is because it is an alternate way of
18261 * writing 19D1, and some people may expect it to be in
18262 * that group. But it is bad, because it won't give
18263 * the expected results. In Unicode 5.2 it was
18264 * considered to be in that group (of 11, hence), but
18265 * this was fixed in the next version */
18267 if (UNLIKELY(value == 0x19DA && prevvalue >= 0x19D0)) {
18268 goto warn_bad_digit_range;
18270 else if (UNLIKELY( prevvalue >= 0x1D7CE
18271 && value <= 0x1D7FF))
18273 /* This is the only other case currently in Unicode
18274 * where the algorithm below fails. The code
18275 * points just above are the end points of a single
18276 * range containing only decimal digits. It is 5
18277 * different series of 0-9. All other ranges of
18278 * digits currently in Unicode are just a single
18279 * series. (And mktables will notify us if a later
18280 * Unicode version breaks this.)
18282 * If the range being checked is at most 9 long,
18283 * and the digit values represented are in
18284 * numerical order, they are from the same series.
18286 if ( value - prevvalue > 9
18287 || ((( value - 0x1D7CE) % 10)
18288 <= (prevvalue - 0x1D7CE) % 10))
18290 goto warn_bad_digit_range;
18295 /* For all other ranges of digits in Unicode, the
18296 * algorithm is just to check if both end points
18297 * are in the same series, which is the same range.
18299 index_start = _invlist_search(
18300 PL_XPosix_ptrs[_CC_DIGIT],
18303 /* Warn if the range starts and ends with a digit,
18304 * and they are not in the same group of 10. */
18305 if ( index_start >= 0
18306 && ELEMENT_RANGE_MATCHES_INVLIST(index_start)
18308 _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
18309 value)) != index_start
18310 && index_final >= 0
18311 && ELEMENT_RANGE_MATCHES_INVLIST(index_final))
18313 warn_bad_digit_range:
18314 vWARN(RExC_parse, "Ranges of digits should be"
18315 " from the same group of"
18322 if ((! range || prevvalue == value) && non_portable_endpoint) {
18323 if (isPRINT_A(value)) {
18326 if (isBACKSLASHED_PUNCT(value)) {
18327 literal[d++] = '\\';
18329 literal[d++] = (char) value;
18330 literal[d++] = '\0';
18333 "\"%.*s\" is more clearly written simply as \"%s\"",
18334 (int) (RExC_parse - rangebegin),
18339 else if (isMNEMONIC_CNTRL(value)) {
18341 "\"%.*s\" is more clearly written simply as \"%s\"",
18342 (int) (RExC_parse - rangebegin),
18344 cntrl_to_mnemonic((U8) value)
18350 /* Deal with this element of the class */
18353 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
18356 /* On non-ASCII platforms, for ranges that span all of 0..255, and ones
18357 * that don't require special handling, we can just add the range like
18358 * we do for ASCII platforms */
18359 if ((UNLIKELY(prevvalue == 0) && value >= 255)
18360 || ! (prevvalue < 256
18362 || (! non_portable_endpoint
18363 && ((isLOWER_A(prevvalue) && isLOWER_A(value))
18364 || (isUPPER_A(prevvalue)
18365 && isUPPER_A(value)))))))
18367 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
18371 /* Here, requires special handling. This can be because it is a
18372 * range whose code points are considered to be Unicode, and so
18373 * must be individually translated into native, or because its a
18374 * subrange of 'A-Z' or 'a-z' which each aren't contiguous in
18375 * EBCDIC, but we have defined them to include only the "expected"
18376 * upper or lower case ASCII alphabetics. Subranges above 255 are
18377 * the same in native and Unicode, so can be added as a range */
18378 U8 start = NATIVE_TO_LATIN1(prevvalue);
18380 U8 end = (value < 256) ? NATIVE_TO_LATIN1(value) : 255;
18381 for (j = start; j <= end; j++) {
18382 cp_foldable_list = add_cp_to_invlist(cp_foldable_list, LATIN1_TO_NATIVE(j));
18385 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
18391 range = 0; /* this range (if it was one) is done now */
18392 } /* End of loop through all the text within the brackets */
18394 if ( posix_warnings && av_tindex_skip_len_mg(posix_warnings) >= 0) {
18395 output_posix_warnings(pRExC_state, posix_warnings);
18398 /* If anything in the class expands to more than one character, we have to
18399 * deal with them by building up a substitute parse string, and recursively
18400 * calling reg() on it, instead of proceeding */
18401 if (multi_char_matches) {
18402 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
18405 char *save_end = RExC_end;
18406 char *save_parse = RExC_parse;
18407 char *save_start = RExC_start;
18408 Size_t constructed_prefix_len = 0; /* This gives the length of the
18409 constructed portion of the
18410 substitute parse. */
18411 bool first_time = TRUE; /* First multi-char occurrence doesn't get
18416 /* Only one level of recursion allowed */
18417 assert(RExC_copy_start_in_constructed == RExC_precomp);
18419 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
18420 because too confusing */
18422 sv_catpvs(substitute_parse, "(?:");
18426 /* Look at the longest folds first */
18427 for (cp_count = av_tindex_skip_len_mg(multi_char_matches);
18432 if (av_exists(multi_char_matches, cp_count)) {
18433 AV** this_array_ptr;
18436 this_array_ptr = (AV**) av_fetch(multi_char_matches,
18438 while ((this_sequence = av_pop(*this_array_ptr)) !=
18441 if (! first_time) {
18442 sv_catpvs(substitute_parse, "|");
18444 first_time = FALSE;
18446 sv_catpv(substitute_parse, SvPVX(this_sequence));
18451 /* If the character class contains anything else besides these
18452 * multi-character folds, have to include it in recursive parsing */
18453 if (element_count) {
18454 sv_catpvs(substitute_parse, "|[");
18455 constructed_prefix_len = SvCUR(substitute_parse);
18456 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
18458 /* Put in a closing ']' only if not going off the end, as otherwise
18459 * we are adding something that really isn't there */
18460 if (RExC_parse < RExC_end) {
18461 sv_catpvs(substitute_parse, "]");
18465 sv_catpvs(substitute_parse, ")");
18468 /* This is a way to get the parse to skip forward a whole named
18469 * sequence instead of matching the 2nd character when it fails the
18471 sv_catpvs(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
18475 /* Set up the data structure so that any errors will be properly
18476 * reported. See the comments at the definition of
18477 * REPORT_LOCATION_ARGS for details */
18478 RExC_copy_start_in_input = (char *) orig_parse;
18479 RExC_start = RExC_parse = SvPV(substitute_parse, len);
18480 RExC_copy_start_in_constructed = RExC_start + constructed_prefix_len;
18481 RExC_end = RExC_parse + len;
18482 RExC_in_multi_char_class = 1;
18484 ret = reg(pRExC_state, 1, ®_flags, depth+1);
18486 *flagp |= reg_flags & (HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_PARSE|NEED_UTF8);
18488 /* And restore so can parse the rest of the pattern */
18489 RExC_parse = save_parse;
18490 RExC_start = RExC_copy_start_in_constructed = RExC_copy_start_in_input = save_start;
18491 RExC_end = save_end;
18492 RExC_in_multi_char_class = 0;
18493 SvREFCNT_dec_NN(multi_char_matches);
18497 /* If folding, we calculate all characters that could fold to or from the
18498 * ones already on the list */
18499 if (cp_foldable_list) {
18501 UV start, end; /* End points of code point ranges */
18503 SV* fold_intersection = NULL;
18506 /* Our calculated list will be for Unicode rules. For locale
18507 * matching, we have to keep a separate list that is consulted at
18508 * runtime only when the locale indicates Unicode rules (and we
18509 * don't include potential matches in the ASCII/Latin1 range, as
18510 * any code point could fold to any other, based on the run-time
18511 * locale). For non-locale, we just use the general list */
18513 use_list = &only_utf8_locale_list;
18516 use_list = &cp_list;
18519 /* Only the characters in this class that participate in folds need
18520 * be checked. Get the intersection of this class and all the
18521 * possible characters that are foldable. This can quickly narrow
18522 * down a large class */
18523 _invlist_intersection(PL_in_some_fold, cp_foldable_list,
18524 &fold_intersection);
18526 /* Now look at the foldable characters in this class individually */
18527 invlist_iterinit(fold_intersection);
18528 while (invlist_iternext(fold_intersection, &start, &end)) {
18532 /* Look at every character in the range */
18533 for (j = start; j <= end; j++) {
18534 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
18537 Size_t folds_count;
18539 const U32 * remaining_folds;
18543 /* Under /l, we don't know what code points below 256
18544 * fold to, except we do know the MICRO SIGN folds to
18545 * an above-255 character if the locale is UTF-8, so we
18546 * add it to the special list (in *use_list) Otherwise
18547 * we know now what things can match, though some folds
18548 * are valid under /d only if the target is UTF-8.
18549 * Those go in a separate list */
18550 if ( IS_IN_SOME_FOLD_L1(j)
18551 && ! (LOC && j != MICRO_SIGN))
18554 /* ASCII is always matched; non-ASCII is matched
18555 * only under Unicode rules (which could happen
18556 * under /l if the locale is a UTF-8 one */
18557 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
18558 *use_list = add_cp_to_invlist(*use_list,
18559 PL_fold_latin1[j]);
18561 else if (j != PL_fold_latin1[j]) {
18562 upper_latin1_only_utf8_matches
18563 = add_cp_to_invlist(
18564 upper_latin1_only_utf8_matches,
18565 PL_fold_latin1[j]);
18569 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(j)
18570 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
18572 add_above_Latin1_folds(pRExC_state,
18579 /* Here is an above Latin1 character. We don't have the
18580 * rules hard-coded for it. First, get its fold. This is
18581 * the simple fold, as the multi-character folds have been
18582 * handled earlier and separated out */
18583 folded = _to_uni_fold_flags(j, foldbuf, &foldlen,
18584 (ASCII_FOLD_RESTRICTED)
18585 ? FOLD_FLAGS_NOMIX_ASCII
18588 /* Single character fold of above Latin1. Add everything
18589 * in its fold closure to the list that this node should
18591 folds_count = _inverse_folds(folded, &first_fold,
18593 for (k = 0; k <= folds_count; k++) {
18594 UV c = (k == 0) /* First time through use itself */
18596 : (k == 1) /* 2nd time use, the first fold */
18599 /* Then the remaining ones */
18600 : remaining_folds[k-2];
18602 /* /aa doesn't allow folds between ASCII and non- */
18603 if (( ASCII_FOLD_RESTRICTED
18604 && (isASCII(c) != isASCII(j))))
18609 /* Folds under /l which cross the 255/256 boundary are
18610 * added to a separate list. (These are valid only
18611 * when the locale is UTF-8.) */
18612 if (c < 256 && LOC) {
18613 *use_list = add_cp_to_invlist(*use_list, c);
18617 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
18619 cp_list = add_cp_to_invlist(cp_list, c);
18622 /* Similarly folds involving non-ascii Latin1
18623 * characters under /d are added to their list */
18624 upper_latin1_only_utf8_matches
18625 = add_cp_to_invlist(
18626 upper_latin1_only_utf8_matches,
18632 SvREFCNT_dec_NN(fold_intersection);
18635 /* Now that we have finished adding all the folds, there is no reason
18636 * to keep the foldable list separate */
18637 _invlist_union(cp_list, cp_foldable_list, &cp_list);
18638 SvREFCNT_dec_NN(cp_foldable_list);
18641 /* And combine the result (if any) with any inversion lists from posix
18642 * classes. The lists are kept separate up to now because we don't want to
18643 * fold the classes */
18644 if (simple_posixes) { /* These are the classes known to be unaffected by
18647 _invlist_union(cp_list, simple_posixes, &cp_list);
18648 SvREFCNT_dec_NN(simple_posixes);
18651 cp_list = simple_posixes;
18654 if (posixes || nposixes) {
18655 if (! DEPENDS_SEMANTICS) {
18657 /* For everything but /d, we can just add the current 'posixes' and
18658 * 'nposixes' to the main list */
18661 _invlist_union(cp_list, posixes, &cp_list);
18662 SvREFCNT_dec_NN(posixes);
18670 _invlist_union(cp_list, nposixes, &cp_list);
18671 SvREFCNT_dec_NN(nposixes);
18674 cp_list = nposixes;
18679 /* Under /d, things like \w match upper Latin1 characters only if
18680 * the target string is in UTF-8. But things like \W match all the
18681 * upper Latin1 characters if the target string is not in UTF-8.
18683 * Handle the case with something like \W separately */
18685 SV* only_non_utf8_list = invlist_clone(PL_UpperLatin1, NULL);
18687 /* A complemented posix class matches all upper Latin1
18688 * characters if not in UTF-8. And it matches just certain
18689 * ones when in UTF-8. That means those certain ones are
18690 * matched regardless, so can just be added to the
18691 * unconditional list */
18693 _invlist_union(cp_list, nposixes, &cp_list);
18694 SvREFCNT_dec_NN(nposixes);
18698 cp_list = nposixes;
18701 /* Likewise for 'posixes' */
18702 _invlist_union(posixes, cp_list, &cp_list);
18703 SvREFCNT_dec(posixes);
18705 /* Likewise for anything else in the range that matched only
18707 if (upper_latin1_only_utf8_matches) {
18708 _invlist_union(cp_list,
18709 upper_latin1_only_utf8_matches,
18711 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
18712 upper_latin1_only_utf8_matches = NULL;
18715 /* If we don't match all the upper Latin1 characters regardless
18716 * of UTF-8ness, we have to set a flag to match the rest when
18718 _invlist_subtract(only_non_utf8_list, cp_list,
18719 &only_non_utf8_list);
18720 if (_invlist_len(only_non_utf8_list) != 0) {
18721 anyof_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
18723 SvREFCNT_dec_NN(only_non_utf8_list);
18726 /* Here there were no complemented posix classes. That means
18727 * the upper Latin1 characters in 'posixes' match only when the
18728 * target string is in UTF-8. So we have to add them to the
18729 * list of those types of code points, while adding the
18730 * remainder to the unconditional list.
18732 * First calculate what they are */
18733 SV* nonascii_but_latin1_properties = NULL;
18734 _invlist_intersection(posixes, PL_UpperLatin1,
18735 &nonascii_but_latin1_properties);
18737 /* And add them to the final list of such characters. */
18738 _invlist_union(upper_latin1_only_utf8_matches,
18739 nonascii_but_latin1_properties,
18740 &upper_latin1_only_utf8_matches);
18742 /* Remove them from what now becomes the unconditional list */
18743 _invlist_subtract(posixes, nonascii_but_latin1_properties,
18746 /* And add those unconditional ones to the final list */
18748 _invlist_union(cp_list, posixes, &cp_list);
18749 SvREFCNT_dec_NN(posixes);
18756 SvREFCNT_dec(nonascii_but_latin1_properties);
18758 /* Get rid of any characters from the conditional list that we
18759 * now know are matched unconditionally, which may make that
18761 _invlist_subtract(upper_latin1_only_utf8_matches,
18763 &upper_latin1_only_utf8_matches);
18764 if (_invlist_len(upper_latin1_only_utf8_matches) == 0) {
18765 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
18766 upper_latin1_only_utf8_matches = NULL;
18772 /* And combine the result (if any) with any inversion list from properties.
18773 * The lists are kept separate up to now so that we can distinguish the two
18774 * in regards to matching above-Unicode. A run-time warning is generated
18775 * if a Unicode property is matched against a non-Unicode code point. But,
18776 * we allow user-defined properties to match anything, without any warning,
18777 * and we also suppress the warning if there is a portion of the character
18778 * class that isn't a Unicode property, and which matches above Unicode, \W
18779 * or [\x{110000}] for example.
18780 * (Note that in this case, unlike the Posix one above, there is no
18781 * <upper_latin1_only_utf8_matches>, because having a Unicode property
18782 * forces Unicode semantics */
18786 /* If it matters to the final outcome, see if a non-property
18787 * component of the class matches above Unicode. If so, the
18788 * warning gets suppressed. This is true even if just a single
18789 * such code point is specified, as, though not strictly correct if
18790 * another such code point is matched against, the fact that they
18791 * are using above-Unicode code points indicates they should know
18792 * the issues involved */
18794 warn_super = ! (invert
18795 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
18798 _invlist_union(properties, cp_list, &cp_list);
18799 SvREFCNT_dec_NN(properties);
18802 cp_list = properties;
18807 |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
18809 /* Because an ANYOF node is the only one that warns, this node
18810 * can't be optimized into something else */
18811 optimizable = FALSE;
18815 /* Here, we have calculated what code points should be in the character
18818 * Now we can see about various optimizations. Fold calculation (which we
18819 * did above) needs to take place before inversion. Otherwise /[^k]/i
18820 * would invert to include K, which under /i would match k, which it
18821 * shouldn't. Therefore we can't invert folded locale now, as it won't be
18822 * folded until runtime */
18824 /* If we didn't do folding, it's because some information isn't available
18825 * until runtime; set the run-time fold flag for these We know to set the
18826 * flag if we have a non-NULL list for UTF-8 locales, or the class matches
18827 * at least one 0-255 range code point */
18830 /* Some things on the list might be unconditionally included because of
18831 * other components. Remove them, and clean up the list if it goes to
18833 if (only_utf8_locale_list && cp_list) {
18834 _invlist_subtract(only_utf8_locale_list, cp_list,
18835 &only_utf8_locale_list);
18837 if (_invlist_len(only_utf8_locale_list) == 0) {
18838 SvREFCNT_dec_NN(only_utf8_locale_list);
18839 only_utf8_locale_list = NULL;
18842 if ( only_utf8_locale_list
18843 || (cp_list && ( _invlist_contains_cp(cp_list, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE)
18844 || _invlist_contains_cp(cp_list, LATIN_SMALL_LETTER_DOTLESS_I))))
18846 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
18849 | ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
18851 else if (cp_list && invlist_lowest(cp_list) < 256) {
18852 /* If nothing is below 256, has no locale dependency; otherwise it
18854 anyof_flags |= ANYOFL_FOLD;
18855 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
18858 else if ( DEPENDS_SEMANTICS
18859 && ( upper_latin1_only_utf8_matches
18860 || (anyof_flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)))
18862 RExC_seen_d_op = TRUE;
18863 has_runtime_dependency |= HAS_D_RUNTIME_DEPENDENCY;
18866 /* Optimize inverted patterns (e.g. [^a-z]) when everything is known at
18870 && ! has_runtime_dependency)
18872 _invlist_invert(cp_list);
18874 /* Clear the invert flag since have just done it here */
18878 /* All possible optimizations below still have these characteristics.
18879 * (Multi-char folds aren't SIMPLE, but they don't get this far in this
18881 *flagp |= HASWIDTH|SIMPLE;
18884 *ret_invlist = cp_list;
18889 if (anyof_flags & ANYOF_LOCALE_FLAGS) {
18890 RExC_contains_locale = 1;
18893 /* Some character classes are equivalent to other nodes. Such nodes take
18894 * up less room, and some nodes require fewer operations to execute, than
18895 * ANYOF nodes. EXACTish nodes may be joinable with adjacent nodes to
18896 * improve efficiency. */
18899 PERL_UINT_FAST8_T i;
18900 UV partial_cp_count = 0;
18901 UV start[MAX_FOLD_FROMS+1] = { 0 }; /* +1 for the folded-to char */
18902 UV end[MAX_FOLD_FROMS+1] = { 0 };
18903 bool single_range = FALSE;
18905 if (cp_list) { /* Count the code points in enough ranges that we would
18906 see all the ones possible in any fold in this version
18909 invlist_iterinit(cp_list);
18910 for (i = 0; i <= MAX_FOLD_FROMS; i++) {
18911 if (! invlist_iternext(cp_list, &start[i], &end[i])) {
18914 partial_cp_count += end[i] - start[i] + 1;
18918 single_range = TRUE;
18920 invlist_iterfinish(cp_list);
18923 /* If we know at compile time that this matches every possible code
18924 * point, any run-time dependencies don't matter */
18925 if (start[0] == 0 && end[0] == UV_MAX) {
18927 ret = reganode(pRExC_state, OPFAIL, 0);
18930 ret = reg_node(pRExC_state, SANY);
18936 /* Similarly, for /l posix classes, if both a class and its
18937 * complement match, any run-time dependencies don't matter */
18939 for (namedclass = 0; namedclass < ANYOF_POSIXL_MAX;
18942 if ( POSIXL_TEST(posixl, namedclass) /* class */
18943 && POSIXL_TEST(posixl, namedclass + 1)) /* its complement */
18946 ret = reganode(pRExC_state, OPFAIL, 0);
18949 ret = reg_node(pRExC_state, SANY);
18956 /* For well-behaved locales, some classes are subsets of others,
18957 * so complementing the subset and including the non-complemented
18958 * superset should match everything, like [\D[:alnum:]], and
18959 * [[:^alpha:][:alnum:]], but some implementations of locales are
18960 * buggy, and khw thinks its a bad idea to have optimization change
18961 * behavior, even if it avoids an OS bug in a given case */
18963 #define isSINGLE_BIT_SET(n) isPOWER_OF_2(n)
18965 /* If is a single posix /l class, can optimize to just that op.
18966 * Such a node will not match anything in the Latin1 range, as that
18967 * is not determinable until runtime, but will match whatever the
18968 * class does outside that range. (Note that some classes won't
18969 * match anything outside the range, like [:ascii:]) */
18970 if ( isSINGLE_BIT_SET(posixl)
18971 && (partial_cp_count == 0 || start[0] > 255))
18974 SV * class_above_latin1 = NULL;
18975 bool already_inverted;
18976 bool are_equivalent;
18978 /* Compute which bit is set, which is the same thing as, e.g.,
18979 * ANYOF_CNTRL. From
18980 * https://graphics.stanford.edu/~seander/bithacks.html#IntegerLogDeBruijn
18982 static const int MultiplyDeBruijnBitPosition2[32] =
18984 0, 1, 28, 2, 29, 14, 24, 3, 30, 22, 20, 15, 25, 17, 4, 8,
18985 31, 27, 13, 23, 21, 19, 16, 7, 26, 12, 18, 6, 11, 5, 10, 9
18988 namedclass = MultiplyDeBruijnBitPosition2[(posixl
18989 * 0x077CB531U) >> 27];
18990 classnum = namedclass_to_classnum(namedclass);
18992 /* The named classes are such that the inverted number is one
18993 * larger than the non-inverted one */
18994 already_inverted = namedclass
18995 - classnum_to_namedclass(classnum);
18997 /* Create an inversion list of the official property, inverted
18998 * if the constructed node list is inverted, and restricted to
18999 * only the above latin1 code points, which are the only ones
19000 * known at compile time */
19001 _invlist_intersection_maybe_complement_2nd(
19003 PL_XPosix_ptrs[classnum],
19005 &class_above_latin1);
19006 are_equivalent = _invlistEQ(class_above_latin1, cp_list,
19008 SvREFCNT_dec_NN(class_above_latin1);
19010 if (are_equivalent) {
19012 /* Resolve the run-time inversion flag with this possibly
19013 * inverted class */
19014 invert = invert ^ already_inverted;
19016 ret = reg_node(pRExC_state,
19017 POSIXL + invert * (NPOSIXL - POSIXL));
19018 FLAGS(REGNODE_p(ret)) = classnum;
19024 /* khw can't think of any other possible transformation involving
19026 if (has_runtime_dependency & HAS_USER_DEFINED_PROPERTY) {
19030 if (! has_runtime_dependency) {
19032 /* If the list is empty, nothing matches. This happens, for
19033 * example, when a Unicode property that doesn't match anything is
19034 * the only element in the character class (perluniprops.pod notes
19035 * such properties). */
19036 if (partial_cp_count == 0) {
19038 ret = reg_node(pRExC_state, SANY);
19041 ret = reganode(pRExC_state, OPFAIL, 0);
19047 /* If matches everything but \n */
19048 if ( start[0] == 0 && end[0] == '\n' - 1
19049 && start[1] == '\n' + 1 && end[1] == UV_MAX)
19052 ret = reg_node(pRExC_state, REG_ANY);
19058 /* Next see if can optimize classes that contain just a few code points
19059 * into an EXACTish node. The reason to do this is to let the
19060 * optimizer join this node with adjacent EXACTish ones, and ANYOF
19061 * nodes require conversion to code point from UTF-8.
19063 * An EXACTFish node can be generated even if not under /i, and vice
19064 * versa. But care must be taken. An EXACTFish node has to be such
19065 * that it only matches precisely the code points in the class, but we
19066 * want to generate the least restrictive one that does that, to
19067 * increase the odds of being able to join with an adjacent node. For
19068 * example, if the class contains [kK], we have to make it an EXACTFAA
19069 * node to prevent the KELVIN SIGN from matching. Whether we are under
19070 * /i or not is irrelevant in this case. Less obvious is the pattern
19071 * qr/[\x{02BC}]n/i. U+02BC is MODIFIER LETTER APOSTROPHE. That is
19072 * supposed to match the single character U+0149 LATIN SMALL LETTER N
19073 * PRECEDED BY APOSTROPHE. And so even though there is no simple fold
19074 * that includes \X{02BC}, there is a multi-char fold that does, and so
19075 * the node generated for it must be an EXACTFish one. On the other
19076 * hand qr/:/i should generate a plain EXACT node since the colon
19077 * participates in no fold whatsoever, and having it EXACT tells the
19078 * optimizer the target string cannot match unless it has a colon in
19084 /* Only try if there are no more code points in the class than
19085 * in the max possible fold */
19086 && inRANGE(partial_cp_count, 1, MAX_FOLD_FROMS + 1))
19088 if (partial_cp_count == 1 && ! upper_latin1_only_utf8_matches)
19090 /* We can always make a single code point class into an
19091 * EXACTish node. */
19095 /* Here is /l: Use EXACTL, except if there is a fold not
19096 * known until runtime so shows as only a single code point
19097 * here. For code points above 255, we know which can
19098 * cause problems by having a potential fold to the Latin1
19101 || ( start[0] > 255
19102 && ! is_PROBLEMATIC_LOCALE_FOLD_cp(start[0])))
19110 else if (! FOLD) { /* Not /l and not /i */
19111 op = (start[0] < 256) ? EXACT : EXACT_REQ8;
19113 else if (start[0] < 256) { /* /i, not /l, and the code point is
19116 /* Under /i, it gets a little tricky. A code point that
19117 * doesn't participate in a fold should be an EXACT node.
19118 * We know this one isn't the result of a simple fold, or
19119 * there'd be more than one code point in the list, but it
19120 * could be part of a multi- character fold. In that case
19121 * we better not create an EXACT node, as we would wrongly
19122 * be telling the optimizer that this code point must be in
19123 * the target string, and that is wrong. This is because
19124 * if the sequence around this code point forms a
19125 * multi-char fold, what needs to be in the string could be
19126 * the code point that folds to the sequence.
19128 * This handles the case of below-255 code points, as we
19129 * have an easy look up for those. The next clause handles
19130 * the above-256 one */
19131 op = IS_IN_SOME_FOLD_L1(start[0])
19135 else { /* /i, larger code point. Since we are under /i, and
19136 have just this code point, we know that it can't
19137 fold to something else, so PL_InMultiCharFold
19139 op = _invlist_contains_cp(PL_InMultiCharFold,
19147 else if ( ! (has_runtime_dependency & ~HAS_D_RUNTIME_DEPENDENCY)
19148 && _invlist_contains_cp(PL_in_some_fold, start[0]))
19150 /* Here, the only runtime dependency, if any, is from /d, and
19151 * the class matches more than one code point, and the lowest
19152 * code point participates in some fold. It might be that the
19153 * other code points are /i equivalent to this one, and hence
19154 * they would representable by an EXACTFish node. Above, we
19155 * eliminated classes that contain too many code points to be
19156 * EXACTFish, with the test for MAX_FOLD_FROMS
19158 * First, special case the ASCII fold pairs, like 'B' and 'b'.
19159 * We do this because we have EXACTFAA at our disposal for the
19161 if (partial_cp_count == 2 && isASCII(start[0])) {
19163 /* The only ASCII characters that participate in folds are
19165 assert(isALPHA(start[0]));
19166 if ( end[0] == start[0] /* First range is a single
19167 character, so 2nd exists */
19168 && isALPHA_FOLD_EQ(start[0], start[1]))
19171 /* Here, is part of an ASCII fold pair */
19173 if ( ASCII_FOLD_RESTRICTED
19174 || HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(start[0]))
19176 /* If the second clause just above was true, it
19177 * means we can't be under /i, or else the list
19178 * would have included more than this fold pair.
19179 * Therefore we have to exclude the possibility of
19180 * whatever else it is that folds to these, by
19181 * using EXACTFAA */
19184 else if (HAS_NONLATIN1_FOLD_CLOSURE(start[0])) {
19186 /* Here, there's no simple fold that start[0] is part
19187 * of, but there is a multi-character one. If we
19188 * are not under /i, we want to exclude that
19189 * possibility; if under /i, we want to include it
19191 op = (FOLD) ? EXACTFU : EXACTFAA;
19195 /* Here, the only possible fold start[0] particpates in
19196 * is with start[1]. /i or not isn't relevant */
19200 value = toFOLD(start[0]);
19203 else if ( ! upper_latin1_only_utf8_matches
19204 || ( _invlist_len(upper_latin1_only_utf8_matches)
19207 invlist_highest(upper_latin1_only_utf8_matches)]
19210 /* Here, the smallest character is non-ascii or there are
19211 * more than 2 code points matched by this node. Also, we
19212 * either don't have /d UTF-8 dependent matches, or if we
19213 * do, they look like they could be a single character that
19214 * is the fold of the lowest one in the always-match list.
19215 * This test quickly excludes most of the false positives
19216 * when there are /d UTF-8 depdendent matches. These are
19217 * like LATIN CAPITAL LETTER A WITH GRAVE matching LATIN
19218 * SMALL LETTER A WITH GRAVE iff the target string is
19219 * UTF-8. (We don't have to worry above about exceeding
19220 * the array bounds of PL_fold_latin1[] because any code
19221 * point in 'upper_latin1_only_utf8_matches' is below 256.)
19223 * EXACTFAA would apply only to pairs (hence exactly 2 code
19224 * points) in the ASCII range, so we can't use it here to
19225 * artificially restrict the fold domain, so we check if
19226 * the class does or does not match some EXACTFish node.
19227 * Further, if we aren't under /i, and and the folded-to
19228 * character is part of a multi-character fold, we can't do
19229 * this optimization, as the sequence around it could be
19230 * that multi-character fold, and we don't here know the
19231 * context, so we have to assume it is that multi-char
19232 * fold, to prevent potential bugs.
19234 * To do the general case, we first find the fold of the
19235 * lowest code point (which may be higher than the lowest
19236 * one), then find everything that folds to it. (The data
19237 * structure we have only maps from the folded code points,
19238 * so we have to do the earlier step.) */
19241 U8 foldbuf[UTF8_MAXBYTES_CASE];
19242 UV folded = _to_uni_fold_flags(start[0],
19243 foldbuf, &foldlen, 0);
19245 const U32 * remaining_folds;
19246 Size_t folds_to_this_cp_count = _inverse_folds(
19250 Size_t folds_count = folds_to_this_cp_count + 1;
19251 SV * fold_list = _new_invlist(folds_count);
19254 /* If there are UTF-8 dependent matches, create a temporary
19255 * list of what this node matches, including them. */
19256 SV * all_cp_list = NULL;
19257 SV ** use_this_list = &cp_list;
19259 if (upper_latin1_only_utf8_matches) {
19260 all_cp_list = _new_invlist(0);
19261 use_this_list = &all_cp_list;
19262 _invlist_union(cp_list,
19263 upper_latin1_only_utf8_matches,
19267 /* Having gotten everything that participates in the fold
19268 * containing the lowest code point, we turn that into an
19269 * inversion list, making sure everything is included. */
19270 fold_list = add_cp_to_invlist(fold_list, start[0]);
19271 fold_list = add_cp_to_invlist(fold_list, folded);
19272 if (folds_to_this_cp_count > 0) {
19273 fold_list = add_cp_to_invlist(fold_list, first_fold);
19274 for (i = 0; i + 1 < folds_to_this_cp_count; i++) {
19275 fold_list = add_cp_to_invlist(fold_list,
19276 remaining_folds[i]);
19280 /* If the fold list is identical to what's in this ANYOF
19281 * node, the node can be represented by an EXACTFish one
19283 if (_invlistEQ(*use_this_list, fold_list,
19284 0 /* Don't complement */ )
19287 /* But, we have to be careful, as mentioned above.
19288 * Just the right sequence of characters could match
19289 * this if it is part of a multi-character fold. That
19290 * IS what we want if we are under /i. But it ISN'T
19291 * what we want if not under /i, as it could match when
19292 * it shouldn't. So, when we aren't under /i and this
19293 * character participates in a multi-char fold, we
19294 * don't optimize into an EXACTFish node. So, for each
19295 * case below we have to check if we are folding
19296 * and if not, if it is not part of a multi-char fold.
19298 if (start[0] > 255) { /* Highish code point */
19299 if (FOLD || ! _invlist_contains_cp(
19300 PL_InMultiCharFold, folded))
19304 : (ASCII_FOLD_RESTRICTED)
19309 } /* Below, the lowest code point < 256 */
19312 && DEPENDS_SEMANTICS)
19313 { /* An EXACTF node containing a single character
19314 's', can be an EXACTFU if it doesn't get
19315 joined with an adjacent 's' */
19316 op = EXACTFU_S_EDGE;
19320 || ! HAS_NONLATIN1_FOLD_CLOSURE(start[0]))
19322 if (upper_latin1_only_utf8_matches) {
19325 /* We can't use the fold, as that only matches
19329 else if ( UNLIKELY(start[0] == MICRO_SIGN)
19331 { /* EXACTFUP is a special node for this
19333 op = (ASCII_FOLD_RESTRICTED)
19336 value = MICRO_SIGN;
19338 else if ( ASCII_FOLD_RESTRICTED
19339 && ! isASCII(start[0]))
19340 { /* For ASCII under /iaa, we can use EXACTFU
19352 SvREFCNT_dec_NN(fold_list);
19353 SvREFCNT_dec(all_cp_list);
19360 /* Here, we have calculated what EXACTish node to use. Have to
19361 * convert to UTF-8 if not already there */
19364 SvREFCNT_dec(cp_list);;
19365 REQUIRE_UTF8(flagp);
19368 /* This is a kludge to the special casing issues with this
19369 * ligature under /aa. FB05 should fold to FB06, but the
19370 * call above to _to_uni_fold_flags() didn't find this, as
19371 * it didn't use the /aa restriction in order to not miss
19372 * other folds that would be affected. This is the only
19373 * instance likely to ever be a problem in all of Unicode.
19374 * So special case it. */
19375 if ( value == LATIN_SMALL_LIGATURE_LONG_S_T
19376 && ASCII_FOLD_RESTRICTED)
19378 value = LATIN_SMALL_LIGATURE_ST;
19382 len = (UTF) ? UVCHR_SKIP(value) : 1;
19384 ret = regnode_guts(pRExC_state, op, len, "exact");
19385 FILL_NODE(ret, op);
19386 RExC_emit += 1 + STR_SZ(len);
19387 setSTR_LEN(REGNODE_p(ret), len);
19389 *STRINGs(REGNODE_p(ret)) = (U8) value;
19392 uvchr_to_utf8((U8 *) STRINGs(REGNODE_p(ret)), value);
19398 if (! has_runtime_dependency) {
19400 /* See if this can be turned into an ANYOFM node. Think about the
19401 * bit patterns in two different bytes. In some positions, the
19402 * bits in each will be 1; and in other positions both will be 0;
19403 * and in some positions the bit will be 1 in one byte, and 0 in
19404 * the other. Let 'n' be the number of positions where the bits
19405 * differ. We create a mask which has exactly 'n' 0 bits, each in
19406 * a position where the two bytes differ. Now take the set of all
19407 * bytes that when ANDed with the mask yield the same result. That
19408 * set has 2**n elements, and is representable by just two 8 bit
19409 * numbers: the result and the mask. Importantly, matching the set
19410 * can be vectorized by creating a word full of the result bytes,
19411 * and a word full of the mask bytes, yielding a significant speed
19412 * up. Here, see if this node matches such a set. As a concrete
19413 * example consider [01], and the byte representing '0' which is
19414 * 0x30 on ASCII machines. It has the bits 0011 0000. Take the
19415 * mask 1111 1110. If we AND 0x31 and 0x30 with that mask we get
19416 * 0x30. Any other bytes ANDed yield something else. So [01],
19417 * which is a common usage, is optimizable into ANYOFM, and can
19418 * benefit from the speed up. We can only do this on UTF-8
19419 * invariant bytes, because they have the same bit patterns under
19421 PERL_UINT_FAST8_T inverted = 0;
19423 const PERL_UINT_FAST8_T max_permissible = 0xFF;
19425 const PERL_UINT_FAST8_T max_permissible = 0x7F;
19427 /* If doesn't fit the criteria for ANYOFM, invert and try again.
19428 * If that works we will instead later generate an NANYOFM, and
19429 * invert back when through */
19430 if (invlist_highest(cp_list) > max_permissible) {
19431 _invlist_invert(cp_list);
19435 if (invlist_highest(cp_list) <= max_permissible) {
19436 UV this_start, this_end;
19437 UV lowest_cp = UV_MAX; /* init'ed to suppress compiler warn */
19438 U8 bits_differing = 0;
19439 Size_t full_cp_count = 0;
19440 bool first_time = TRUE;
19442 /* Go through the bytes and find the bit positions that differ
19444 invlist_iterinit(cp_list);
19445 while (invlist_iternext(cp_list, &this_start, &this_end)) {
19446 unsigned int i = this_start;
19449 if (! UVCHR_IS_INVARIANT(i)) {
19453 first_time = FALSE;
19454 lowest_cp = this_start;
19456 /* We have set up the code point to compare with.
19457 * Don't compare it with itself */
19461 /* Find the bit positions that differ from the lowest code
19462 * point in the node. Keep track of all such positions by
19464 for (; i <= this_end; i++) {
19465 if (! UVCHR_IS_INVARIANT(i)) {
19469 bits_differing |= i ^ lowest_cp;
19472 full_cp_count += this_end - this_start + 1;
19475 /* At the end of the loop, we count how many bits differ from
19476 * the bits in lowest code point, call the count 'd'. If the
19477 * set we found contains 2**d elements, it is the closure of
19478 * all code points that differ only in those bit positions. To
19479 * convince yourself of that, first note that the number in the
19480 * closure must be a power of 2, which we test for. The only
19481 * way we could have that count and it be some differing set,
19482 * is if we got some code points that don't differ from the
19483 * lowest code point in any position, but do differ from each
19484 * other in some other position. That means one code point has
19485 * a 1 in that position, and another has a 0. But that would
19486 * mean that one of them differs from the lowest code point in
19487 * that position, which possibility we've already excluded. */
19488 if ( (inverted || full_cp_count > 1)
19489 && full_cp_count == 1U << PL_bitcount[bits_differing])
19493 op = ANYOFM + inverted;;
19495 /* We need to make the bits that differ be 0's */
19496 ANYOFM_mask = ~ bits_differing; /* This goes into FLAGS */
19498 /* The argument is the lowest code point */
19499 ret = reganode(pRExC_state, op, lowest_cp);
19500 FLAGS(REGNODE_p(ret)) = ANYOFM_mask;
19504 invlist_iterfinish(cp_list);
19508 _invlist_invert(cp_list);
19515 /* XXX We could create an ANYOFR_LOW node here if we saved above if
19516 * all were invariants, it wasn't inverted, and there is a single
19517 * range. This would be faster than some of the posix nodes we
19518 * create below like /\d/a, but would be twice the size. Without
19519 * having actually measured the gain, khw doesn't think the
19520 * tradeoff is really worth it */
19523 if (! (anyof_flags & ANYOF_LOCALE_FLAGS)) {
19524 PERL_UINT_FAST8_T type;
19525 SV * intersection = NULL;
19526 SV* d_invlist = NULL;
19528 /* See if this matches any of the POSIX classes. The POSIXA and
19529 * POSIXD ones are about the same speed as ANYOF ops, but take less
19530 * room; the ones that have above-Latin1 code point matches are
19531 * somewhat faster than ANYOF. */
19533 for (type = POSIXA; type >= POSIXD; type--) {
19536 if (type == POSIXL) { /* But not /l posix classes */
19540 for (posix_class = 0;
19541 posix_class <= _HIGHEST_REGCOMP_DOT_H_SYNC;
19544 SV** our_code_points = &cp_list;
19545 SV** official_code_points;
19548 if (type == POSIXA) {
19549 official_code_points = &PL_Posix_ptrs[posix_class];
19552 official_code_points = &PL_XPosix_ptrs[posix_class];
19555 /* Skip non-existent classes of this type. e.g. \v only
19556 * has an entry in PL_XPosix_ptrs */
19557 if (! *official_code_points) {
19561 /* Try both the regular class, and its inversion */
19562 for (try_inverted = 0; try_inverted < 2; try_inverted++) {
19563 bool this_inverted = invert ^ try_inverted;
19565 if (type != POSIXD) {
19567 /* This class that isn't /d can't match if we have
19568 * /d dependencies */
19569 if (has_runtime_dependency
19570 & HAS_D_RUNTIME_DEPENDENCY)
19575 else /* is /d */ if (! this_inverted) {
19577 /* /d classes don't match anything non-ASCII below
19578 * 256 unconditionally (which cp_list contains) */
19579 _invlist_intersection(cp_list, PL_UpperLatin1,
19581 if (_invlist_len(intersection) != 0) {
19585 SvREFCNT_dec(d_invlist);
19586 d_invlist = invlist_clone(cp_list, NULL);
19588 /* But under UTF-8 it turns into using /u rules.
19589 * Add the things it matches under these conditions
19590 * so that we check below that these are identical
19591 * to what the tested class should match */
19592 if (upper_latin1_only_utf8_matches) {
19595 upper_latin1_only_utf8_matches,
19598 our_code_points = &d_invlist;
19600 else { /* POSIXD, inverted. If this doesn't have this
19601 flag set, it isn't /d. */
19602 if (! (anyof_flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
19606 our_code_points = &cp_list;
19609 /* Here, have weeded out some things. We want to see
19610 * if the list of characters this node contains
19611 * ('*our_code_points') precisely matches those of the
19612 * class we are currently checking against
19613 * ('*official_code_points'). */
19614 if (_invlistEQ(*our_code_points,
19615 *official_code_points,
19618 /* Here, they precisely match. Optimize this ANYOF
19619 * node into its equivalent POSIX one of the
19620 * correct type, possibly inverted */
19621 ret = reg_node(pRExC_state, (try_inverted)
19625 FLAGS(REGNODE_p(ret)) = posix_class;
19626 SvREFCNT_dec(d_invlist);
19627 SvREFCNT_dec(intersection);
19633 SvREFCNT_dec(d_invlist);
19634 SvREFCNT_dec(intersection);
19637 /* If it is a single contiguous range, ANYOFR is an efficient regnode,
19638 * both in size and speed. Currently, a 20 bit range base (smallest
19639 * code point in the range), and a 12 bit maximum delta are packed into
19640 * a 32 bit word. This allows for using it on all of the Unicode code
19641 * points except for the highest plane, which is only for private use
19642 * code points. khw doubts that a bigger delta is likely in real world
19645 && ! has_runtime_dependency
19646 && anyof_flags == 0
19647 && start[0] < (1 << ANYOFR_BASE_BITS)
19648 && end[0] - start[0]
19649 < ((1U << (sizeof(((struct regnode_1 *)NULL)->arg1)
19650 * CHARBITS - ANYOFR_BASE_BITS))))
19653 U8 low_utf8[UTF8_MAXBYTES+1];
19654 U8 high_utf8[UTF8_MAXBYTES+1];
19656 ret = reganode(pRExC_state, ANYOFR,
19657 (start[0] | (end[0] - start[0]) << ANYOFR_BASE_BITS));
19659 /* Place the lowest UTF-8 start byte in the flags field, so as to
19660 * allow efficient ruling out at run time of many possible inputs.
19662 (void) uvchr_to_utf8(low_utf8, start[0]);
19663 (void) uvchr_to_utf8(high_utf8, end[0]);
19665 /* If all code points share the same first byte, this can be an
19666 * ANYOFRb. Otherwise store the lowest UTF-8 start byte which can
19667 * quickly rule out many inputs at run-time without having to
19668 * compute the code point from UTF-8. For EBCDIC, we use I8, as
19669 * not doing that transformation would not rule out nearly so many
19671 if (low_utf8[0] == high_utf8[0]) {
19672 OP(REGNODE_p(ret)) = ANYOFRb;
19673 ANYOF_FLAGS(REGNODE_p(ret)) = low_utf8[0];
19676 ANYOF_FLAGS(REGNODE_p(ret))
19677 = NATIVE_UTF8_TO_I8(low_utf8[0]);
19683 /* If didn't find an optimization and there is no need for a bitmap,
19684 * optimize to indicate that */
19685 if ( start[0] >= NUM_ANYOF_CODE_POINTS
19687 && ! upper_latin1_only_utf8_matches
19688 && anyof_flags == 0)
19690 U8 low_utf8[UTF8_MAXBYTES+1];
19691 UV highest_cp = invlist_highest(cp_list);
19693 /* Currently the maximum allowed code point by the system is
19694 * IV_MAX. Higher ones are reserved for future internal use. This
19695 * particular regnode can be used for higher ones, but we can't
19696 * calculate the code point of those. IV_MAX suffices though, as
19697 * it will be a large first byte */
19698 Size_t low_len = uvchr_to_utf8(low_utf8, MIN(start[0], IV_MAX))
19701 /* We store the lowest possible first byte of the UTF-8
19702 * representation, using the flags field. This allows for quick
19703 * ruling out of some inputs without having to convert from UTF-8
19704 * to code point. For EBCDIC, we use I8, as not doing that
19705 * transformation would not rule out nearly so many things */
19706 anyof_flags = NATIVE_UTF8_TO_I8(low_utf8[0]);
19710 /* If the first UTF-8 start byte for the highest code point in the
19711 * range is suitably small, we may be able to get an upper bound as
19713 if (highest_cp <= IV_MAX) {
19714 U8 high_utf8[UTF8_MAXBYTES+1];
19715 Size_t high_len = uvchr_to_utf8(high_utf8, highest_cp)
19718 /* If the lowest and highest are the same, we can get an exact
19719 * first byte instead of a just minimum or even a sequence of
19720 * exact leading bytes. We signal these with different
19722 if (low_utf8[0] == high_utf8[0]) {
19723 Size_t len = find_first_differing_byte_pos(low_utf8,
19725 MIN(low_len, high_len));
19729 /* No need to convert to I8 for EBCDIC as this is an
19731 anyof_flags = low_utf8[0];
19736 ret = regnode_guts(pRExC_state, op,
19737 regarglen[op] + STR_SZ(len),
19739 FILL_NODE(ret, op);
19740 ((struct regnode_anyofhs *) REGNODE_p(ret))->str_len
19742 Copy(low_utf8, /* Add the common bytes */
19743 ((struct regnode_anyofhs *) REGNODE_p(ret))->string,
19745 RExC_emit += NODE_SZ_STR(REGNODE_p(ret));
19746 set_ANYOF_arg(pRExC_state, REGNODE_p(ret), cp_list,
19747 NULL, only_utf8_locale_list);
19751 else if (NATIVE_UTF8_TO_I8(high_utf8[0]) <= MAX_ANYOF_HRx_BYTE)
19754 /* Here, the high byte is not the same as the low, but is
19755 * small enough that its reasonable to have a loose upper
19756 * bound, which is packed in with the strict lower bound.
19757 * See comments at the definition of MAX_ANYOF_HRx_BYTE.
19758 * On EBCDIC platforms, I8 is used. On ASCII platforms I8
19759 * is the same thing as UTF-8 */
19762 U8 max_range_diff = MAX_ANYOF_HRx_BYTE - anyof_flags;
19763 U8 range_diff = NATIVE_UTF8_TO_I8(high_utf8[0])
19766 if (range_diff <= max_range_diff / 8) {
19769 else if (range_diff <= max_range_diff / 4) {
19772 else if (range_diff <= max_range_diff / 2) {
19775 anyof_flags = (anyof_flags - 0xC0) << 2 | bits;
19780 goto done_finding_op;
19782 } /* End of seeing if can optimize it into a different node */
19784 is_anyof: /* It's going to be an ANYOF node. */
19785 op = (has_runtime_dependency & HAS_D_RUNTIME_DEPENDENCY)
19795 ret = regnode_guts(pRExC_state, op, regarglen[op], "anyof");
19796 FILL_NODE(ret, op); /* We set the argument later */
19797 RExC_emit += 1 + regarglen[op];
19798 ANYOF_FLAGS(REGNODE_p(ret)) = anyof_flags;
19800 /* Here, <cp_list> contains all the code points we can determine at
19801 * compile time that match under all conditions. Go through it, and
19802 * for things that belong in the bitmap, put them there, and delete from
19803 * <cp_list>. While we are at it, see if everything above 255 is in the
19804 * list, and if so, set a flag to speed up execution */
19806 populate_ANYOF_from_invlist(REGNODE_p(ret), &cp_list);
19809 ANYOF_POSIXL_SET_TO_BITMAP(REGNODE_p(ret), posixl);
19813 ANYOF_FLAGS(REGNODE_p(ret)) |= ANYOF_INVERT;
19816 /* Here, the bitmap has been populated with all the Latin1 code points that
19817 * always match. Can now add to the overall list those that match only
19818 * when the target string is UTF-8 (<upper_latin1_only_utf8_matches>).
19820 if (upper_latin1_only_utf8_matches) {
19822 _invlist_union(cp_list,
19823 upper_latin1_only_utf8_matches,
19825 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
19828 cp_list = upper_latin1_only_utf8_matches;
19830 ANYOF_FLAGS(REGNODE_p(ret)) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
19833 set_ANYOF_arg(pRExC_state, REGNODE_p(ret), cp_list,
19834 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
19837 only_utf8_locale_list);
19838 SvREFCNT_dec(cp_list);;
19839 SvREFCNT_dec(only_utf8_locale_list);
19844 /* Here, the node is getting optimized into something that's not an ANYOF
19845 * one. Finish up. */
19847 Set_Node_Offset_Length(REGNODE_p(ret), orig_parse - RExC_start,
19848 RExC_parse - orig_parse);;
19849 SvREFCNT_dec(cp_list);;
19850 SvREFCNT_dec(only_utf8_locale_list);
19854 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
19857 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
19858 regnode* const node,
19860 SV* const runtime_defns,
19861 SV* const only_utf8_locale_list)
19863 /* Sets the arg field of an ANYOF-type node 'node', using information about
19864 * the node passed-in. If there is nothing outside the node's bitmap, the
19865 * arg is set to ANYOF_ONLY_HAS_BITMAP. Otherwise, it sets the argument to
19866 * the count returned by add_data(), having allocated and stored an array,
19869 * av[0] stores the inversion list defining this class as far as known at
19870 * this time, or PL_sv_undef if nothing definite is now known.
19871 * av[1] stores the inversion list of code points that match only if the
19872 * current locale is UTF-8, or if none, PL_sv_undef if there is an
19873 * av[2], or no entry otherwise.
19874 * av[2] stores the list of user-defined properties whose subroutine
19875 * definitions aren't known at this time, or no entry if none. */
19879 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
19881 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
19882 assert(! (ANYOF_FLAGS(node)
19883 & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP));
19884 ARG_SET(node, ANYOF_ONLY_HAS_BITMAP);
19887 AV * const av = newAV();
19891 av_store(av, INVLIST_INDEX, SvREFCNT_inc_NN(cp_list));
19894 if (only_utf8_locale_list) {
19895 av_store(av, ONLY_LOCALE_MATCHES_INDEX,
19896 SvREFCNT_inc_NN(only_utf8_locale_list));
19899 if (runtime_defns) {
19900 av_store(av, DEFERRED_USER_DEFINED_INDEX,
19901 SvREFCNT_inc_NN(runtime_defns));
19904 rv = newRV_noinc(MUTABLE_SV(av));
19905 n = add_data(pRExC_state, STR_WITH_LEN("s"));
19906 RExC_rxi->data->data[n] = (void*)rv;
19911 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
19913 Perl__get_regclass_nonbitmap_data(pTHX_ const regexp *prog,
19914 const regnode* node,
19917 SV** only_utf8_locale_ptr,
19918 SV** output_invlist)
19921 /* For internal core use only.
19922 * Returns the inversion list for the input 'node' in the regex 'prog'.
19923 * If <doinit> is 'true', will attempt to create the inversion list if not
19925 * If <listsvp> is non-null, will return the printable contents of the
19926 * property definition. This can be used to get debugging information
19927 * even before the inversion list exists, by calling this function with
19928 * 'doinit' set to false, in which case the components that will be used
19929 * to eventually create the inversion list are returned (in a printable
19931 * If <only_utf8_locale_ptr> is not NULL, it is where this routine is to
19932 * store an inversion list of code points that should match only if the
19933 * execution-time locale is a UTF-8 one.
19934 * If <output_invlist> is not NULL, it is where this routine is to store an
19935 * inversion list of the code points that would be instead returned in
19936 * <listsvp> if this were NULL. Thus, what gets output in <listsvp>
19937 * when this parameter is used, is just the non-code point data that
19938 * will go into creating the inversion list. This currently should be just
19939 * user-defined properties whose definitions were not known at compile
19940 * time. Using this parameter allows for easier manipulation of the
19941 * inversion list's data by the caller. It is illegal to call this
19942 * function with this parameter set, but not <listsvp>
19944 * Tied intimately to how S_set_ANYOF_arg sets up the data structure. Note
19945 * that, in spite of this function's name, the inversion list it returns
19946 * may include the bitmap data as well */
19948 SV *si = NULL; /* Input initialization string */
19949 SV* invlist = NULL;
19951 RXi_GET_DECL(prog, progi);
19952 const struct reg_data * const data = prog ? progi->data : NULL;
19954 PERL_ARGS_ASSERT__GET_REGCLASS_NONBITMAP_DATA;
19955 assert(! output_invlist || listsvp);
19957 if (data && data->count) {
19958 const U32 n = ARG(node);
19960 if (data->what[n] == 's') {
19961 SV * const rv = MUTABLE_SV(data->data[n]);
19962 AV * const av = MUTABLE_AV(SvRV(rv));
19963 SV **const ary = AvARRAY(av);
19965 invlist = ary[INVLIST_INDEX];
19967 if (av_tindex_skip_len_mg(av) >= ONLY_LOCALE_MATCHES_INDEX) {
19968 *only_utf8_locale_ptr = ary[ONLY_LOCALE_MATCHES_INDEX];
19971 if (av_tindex_skip_len_mg(av) >= DEFERRED_USER_DEFINED_INDEX) {
19972 si = ary[DEFERRED_USER_DEFINED_INDEX];
19975 if (doinit && (si || invlist)) {
19978 SV * msg = newSVpvs_flags("", SVs_TEMP);
19980 SV * prop_definition = handle_user_defined_property(
19981 "", 0, FALSE, /* There is no \p{}, \P{} */
19982 SvPVX_const(si)[1] - '0', /* /i or not has been
19983 stored here for just
19985 TRUE, /* run time */
19986 FALSE, /* This call must find the defn */
19987 si, /* The property definition */
19990 0 /* base level call */
19994 assert(prop_definition == NULL);
19996 Perl_croak(aTHX_ "%" UTF8f,
19997 UTF8fARG(SvUTF8(msg), SvCUR(msg), SvPVX(msg)));
20001 _invlist_union(invlist, prop_definition, &invlist);
20002 SvREFCNT_dec_NN(prop_definition);
20005 invlist = prop_definition;
20008 STATIC_ASSERT_STMT(ONLY_LOCALE_MATCHES_INDEX == 1 + INVLIST_INDEX);
20009 STATIC_ASSERT_STMT(DEFERRED_USER_DEFINED_INDEX == 1 + ONLY_LOCALE_MATCHES_INDEX);
20011 ary[INVLIST_INDEX] = invlist;
20012 av_fill(av, (ary[ONLY_LOCALE_MATCHES_INDEX])
20013 ? ONLY_LOCALE_MATCHES_INDEX
20021 /* If requested, return a printable version of what this ANYOF node matches
20024 SV* matches_string = NULL;
20026 /* This function can be called at compile-time, before everything gets
20027 * resolved, in which case we return the currently best available
20028 * information, which is the string that will eventually be used to do
20029 * that resolving, 'si' */
20031 /* Here, we only have 'si' (and possibly some passed-in data in
20032 * 'invlist', which is handled below) If the caller only wants
20033 * 'si', use that. */
20034 if (! output_invlist) {
20035 matches_string = newSVsv(si);
20038 /* But if the caller wants an inversion list of the node, we
20039 * need to parse 'si' and place as much as possible in the
20040 * desired output inversion list, making 'matches_string' only
20041 * contain the currently unresolvable things */
20042 const char *si_string = SvPVX(si);
20043 STRLEN remaining = SvCUR(si);
20047 /* Ignore everything before and including the first new-line */
20048 si_string = (const char *) memchr(si_string, '\n', SvCUR(si));
20049 assert (si_string != NULL);
20051 remaining = SvPVX(si) + SvCUR(si) - si_string;
20053 while (remaining > 0) {
20055 /* The data consists of just strings defining user-defined
20056 * property names, but in prior incarnations, and perhaps
20057 * somehow from pluggable regex engines, it could still
20058 * hold hex code point definitions, all of which should be
20059 * legal (or it wouldn't have gotten this far). Each
20060 * component of a range would be separated by a tab, and
20061 * each range by a new-line. If these are found, instead
20062 * add them to the inversion list */
20063 I32 grok_flags = PERL_SCAN_SILENT_ILLDIGIT
20064 |PERL_SCAN_SILENT_NON_PORTABLE;
20065 STRLEN len = remaining;
20066 UV cp = grok_hex(si_string, &len, &grok_flags, NULL);
20068 /* If the hex decode routine found something, it should go
20069 * up to the next \n */
20070 if ( *(si_string + len) == '\n') {
20071 if (count) { /* 2nd code point on line */
20072 *output_invlist = _add_range_to_invlist(*output_invlist, prev_cp, cp);
20075 *output_invlist = add_cp_to_invlist(*output_invlist, cp);
20078 goto prepare_for_next_iteration;
20081 /* If the hex decode was instead for the lower range limit,
20082 * save it, and go parse the upper range limit */
20083 if (*(si_string + len) == '\t') {
20084 assert(count == 0);
20088 prepare_for_next_iteration:
20089 si_string += len + 1;
20090 remaining -= len + 1;
20094 /* Here, didn't find a legal hex number. Just add the text
20095 * from here up to the next \n, omitting any trailing
20099 len = strcspn(si_string,
20100 DEFERRED_COULD_BE_OFFICIAL_MARKERs "\n");
20102 if (matches_string) {
20103 sv_catpvn(matches_string, si_string, len);
20106 matches_string = newSVpvn(si_string, len);
20108 sv_catpvs(matches_string, " ");
20112 && UCHARAT(si_string)
20113 == DEFERRED_COULD_BE_OFFICIAL_MARKERc)
20118 if (remaining && UCHARAT(si_string) == '\n') {
20122 } /* end of loop through the text */
20124 assert(matches_string);
20125 if (SvCUR(matches_string)) { /* Get rid of trailing blank */
20126 SvCUR_set(matches_string, SvCUR(matches_string) - 1);
20128 } /* end of has an 'si' */
20131 /* Add the stuff that's already known */
20134 /* Again, if the caller doesn't want the output inversion list, put
20135 * everything in 'matches-string' */
20136 if (! output_invlist) {
20137 if ( ! matches_string) {
20138 matches_string = newSVpvs("\n");
20140 sv_catsv(matches_string, invlist_contents(invlist,
20141 TRUE /* traditional style */
20144 else if (! *output_invlist) {
20145 *output_invlist = invlist_clone(invlist, NULL);
20148 _invlist_union(*output_invlist, invlist, output_invlist);
20152 *listsvp = matches_string;
20157 #endif /* !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION) */
20159 /* reg_skipcomment()
20161 Absorbs an /x style # comment from the input stream,
20162 returning a pointer to the first character beyond the comment, or if the
20163 comment terminates the pattern without anything following it, this returns
20164 one past the final character of the pattern (in other words, RExC_end) and
20165 sets the REG_RUN_ON_COMMENT_SEEN flag.
20167 Note it's the callers responsibility to ensure that we are
20168 actually in /x mode
20172 PERL_STATIC_INLINE char*
20173 S_reg_skipcomment(RExC_state_t *pRExC_state, char* p)
20175 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
20179 while (p < RExC_end) {
20180 if (*(++p) == '\n') {
20185 /* we ran off the end of the pattern without ending the comment, so we have
20186 * to add an \n when wrapping */
20187 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
20192 S_skip_to_be_ignored_text(pTHX_ RExC_state_t *pRExC_state,
20194 const bool force_to_xmod
20197 /* If the text at the current parse position '*p' is a '(?#...)' comment,
20198 * or if we are under /x or 'force_to_xmod' is TRUE, and the text at '*p'
20199 * is /x whitespace, advance '*p' so that on exit it points to the first
20200 * byte past all such white space and comments */
20202 const bool use_xmod = force_to_xmod || (RExC_flags & RXf_PMf_EXTENDED);
20204 PERL_ARGS_ASSERT_SKIP_TO_BE_IGNORED_TEXT;
20206 assert( ! UTF || UTF8_IS_INVARIANT(**p) || UTF8_IS_START(**p));
20209 if (RExC_end - (*p) >= 3
20211 && *(*p + 1) == '?'
20212 && *(*p + 2) == '#')
20214 while (*(*p) != ')') {
20215 if ((*p) == RExC_end)
20216 FAIL("Sequence (?#... not terminated");
20224 const char * save_p = *p;
20225 while ((*p) < RExC_end) {
20227 if ((len = is_PATWS_safe((*p), RExC_end, UTF))) {
20230 else if (*(*p) == '#') {
20231 (*p) = reg_skipcomment(pRExC_state, (*p));
20237 if (*p != save_p) {
20250 Advances the parse position by one byte, unless that byte is the beginning
20251 of a '(?#...)' style comment, or is /x whitespace and /x is in effect. In
20252 those two cases, the parse position is advanced beyond all such comments and
20255 This is the UTF, (?#...), and /x friendly way of saying RExC_parse++.
20259 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
20261 PERL_ARGS_ASSERT_NEXTCHAR;
20263 if (RExC_parse < RExC_end) {
20265 || UTF8_IS_INVARIANT(*RExC_parse)
20266 || UTF8_IS_START(*RExC_parse));
20268 RExC_parse += (UTF)
20269 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
20272 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
20273 FALSE /* Don't force /x */ );
20278 S_change_engine_size(pTHX_ RExC_state_t *pRExC_state, const Ptrdiff_t size)
20280 /* 'size' is the delta number of smallest regnode equivalents to add or
20281 * subtract from the current memory allocated to the regex engine being
20284 PERL_ARGS_ASSERT_CHANGE_ENGINE_SIZE;
20289 sizeof(regexp_internal) + (RExC_size + 1) * sizeof(regnode),
20290 /* +1 for REG_MAGIC */
20293 if ( RExC_rxi == NULL )
20294 FAIL("Regexp out of space");
20295 RXi_SET(RExC_rx, RExC_rxi);
20297 RExC_emit_start = RExC_rxi->program;
20299 Zero(REGNODE_p(RExC_emit), size, regnode);
20302 #ifdef RE_TRACK_PATTERN_OFFSETS
20303 Renew(RExC_offsets, 2*RExC_size+1, U32);
20305 Zero(RExC_offsets + 2*(RExC_size - size) + 1, 2 * size, U32);
20307 RExC_offsets[0] = RExC_size;
20311 STATIC regnode_offset
20312 S_regnode_guts(pTHX_ RExC_state_t *pRExC_state, const U8 op, const STRLEN extra_size, const char* const name)
20314 /* Allocate a regnode for 'op', with 'extra_size' extra (smallest) regnode
20315 * equivalents space. It aligns and increments RExC_size
20317 * It returns the regnode's offset into the regex engine program */
20319 const regnode_offset ret = RExC_emit;
20321 DECLARE_AND_GET_RE_DEBUG_FLAGS;
20323 PERL_ARGS_ASSERT_REGNODE_GUTS;
20325 SIZE_ALIGN(RExC_size);
20326 change_engine_size(pRExC_state, (Ptrdiff_t) 1 + extra_size);
20327 NODE_ALIGN_FILL(REGNODE_p(ret));
20328 #ifndef RE_TRACK_PATTERN_OFFSETS
20329 PERL_UNUSED_ARG(name);
20330 PERL_UNUSED_ARG(op);
20332 assert(extra_size >= regarglen[op] || PL_regkind[op] == ANYOF);
20334 if (RExC_offsets) { /* MJD */
20336 ("%s:%d: (op %s) %s %" UVuf " (len %" UVuf ") (max %" UVuf ").\n",
20339 (UV)(RExC_emit) > RExC_offsets[0]
20340 ? "Overwriting end of array!\n" : "OK",
20342 (UV)(RExC_parse - RExC_start),
20343 (UV)RExC_offsets[0]));
20344 Set_Node_Offset(REGNODE_p(RExC_emit), RExC_parse + (op == END));
20351 - reg_node - emit a node
20353 STATIC regnode_offset /* Location. */
20354 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
20356 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reg_node");
20357 regnode_offset ptr = ret;
20359 PERL_ARGS_ASSERT_REG_NODE;
20361 assert(regarglen[op] == 0);
20363 FILL_ADVANCE_NODE(ptr, op);
20369 - reganode - emit a node with an argument
20371 STATIC regnode_offset /* Location. */
20372 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
20374 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reganode");
20375 regnode_offset ptr = ret;
20377 PERL_ARGS_ASSERT_REGANODE;
20379 /* ANYOF are special cased to allow non-length 1 args */
20380 assert(regarglen[op] == 1);
20382 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
20388 - regpnode - emit a temporary node with a void* argument
20390 STATIC regnode_offset /* Location. */
20391 S_regpnode(pTHX_ RExC_state_t *pRExC_state, U8 op, void * arg)
20393 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "regvnode");
20394 regnode_offset ptr = ret;
20396 PERL_ARGS_ASSERT_REGPNODE;
20398 FILL_ADVANCE_NODE_ARGp(ptr, op, arg);
20403 STATIC regnode_offset
20404 S_reg2Lanode(pTHX_ RExC_state_t *pRExC_state, const U8 op, const U32 arg1, const I32 arg2)
20406 /* emit a node with U32 and I32 arguments */
20408 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reg2Lanode");
20409 regnode_offset ptr = ret;
20411 PERL_ARGS_ASSERT_REG2LANODE;
20413 assert(regarglen[op] == 2);
20415 FILL_ADVANCE_NODE_2L_ARG(ptr, op, arg1, arg2);
20421 - reginsert - insert an operator in front of already-emitted operand
20423 * That means that on exit 'operand' is the offset of the newly inserted
20424 * operator, and the original operand has been relocated.
20426 * IMPORTANT NOTE - it is the *callers* responsibility to correctly
20427 * set up NEXT_OFF() of the inserted node if needed. Something like this:
20429 * reginsert(pRExC, OPFAIL, orig_emit, depth+1);
20430 * NEXT_OFF(orig_emit) = regarglen[OPFAIL] + NODE_STEP_REGNODE;
20432 * ALSO NOTE - FLAGS(newly-inserted-operator) will be set to 0 as well.
20435 S_reginsert(pTHX_ RExC_state_t *pRExC_state, const U8 op,
20436 const regnode_offset operand, const U32 depth)
20441 const int offset = regarglen[(U8)op];
20442 const int size = NODE_STEP_REGNODE + offset;
20443 DECLARE_AND_GET_RE_DEBUG_FLAGS;
20445 PERL_ARGS_ASSERT_REGINSERT;
20446 PERL_UNUSED_CONTEXT;
20447 PERL_UNUSED_ARG(depth);
20448 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
20449 DEBUG_PARSE_FMT("inst"," - %s", PL_reg_name[op]);
20450 assert(!RExC_study_started); /* I believe we should never use reginsert once we have started
20451 studying. If this is wrong then we need to adjust RExC_recurse
20452 below like we do with RExC_open_parens/RExC_close_parens. */
20453 change_engine_size(pRExC_state, (Ptrdiff_t) size);
20454 src = REGNODE_p(RExC_emit);
20456 dst = REGNODE_p(RExC_emit);
20458 /* If we are in a "count the parentheses" pass, the numbers are unreliable,
20459 * and [perl #133871] shows this can lead to problems, so skip this
20460 * realignment of parens until a later pass when they are reliable */
20461 if (! IN_PARENS_PASS && RExC_open_parens) {
20463 /*DEBUG_PARSE_FMT("inst"," - %" IVdf, (IV)RExC_npar);*/
20464 /* remember that RExC_npar is rex->nparens + 1,
20465 * iow it is 1 more than the number of parens seen in
20466 * the pattern so far. */
20467 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
20468 /* note, RExC_open_parens[0] is the start of the
20469 * regex, it can't move. RExC_close_parens[0] is the end
20470 * of the regex, it *can* move. */
20471 if ( paren && RExC_open_parens[paren] >= operand ) {
20472 /*DEBUG_PARSE_FMT("open"," - %d", size);*/
20473 RExC_open_parens[paren] += size;
20475 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
20477 if ( RExC_close_parens[paren] >= operand ) {
20478 /*DEBUG_PARSE_FMT("close"," - %d", size);*/
20479 RExC_close_parens[paren] += size;
20481 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
20486 RExC_end_op += size;
20488 while (src > REGNODE_p(operand)) {
20489 StructCopy(--src, --dst, regnode);
20490 #ifdef RE_TRACK_PATTERN_OFFSETS
20491 if (RExC_offsets) { /* MJD 20010112 */
20493 ("%s(%d): (op %s) %s copy %" UVuf " -> %" UVuf " (max %" UVuf ").\n",
20497 (UV)(REGNODE_OFFSET(dst)) > RExC_offsets[0]
20498 ? "Overwriting end of array!\n" : "OK",
20499 (UV)REGNODE_OFFSET(src),
20500 (UV)REGNODE_OFFSET(dst),
20501 (UV)RExC_offsets[0]));
20502 Set_Node_Offset_To_R(REGNODE_OFFSET(dst), Node_Offset(src));
20503 Set_Node_Length_To_R(REGNODE_OFFSET(dst), Node_Length(src));
20508 place = REGNODE_p(operand); /* Op node, where operand used to be. */
20509 #ifdef RE_TRACK_PATTERN_OFFSETS
20510 if (RExC_offsets) { /* MJD */
20512 ("%s(%d): (op %s) %s %" UVuf " <- %" UVuf " (max %" UVuf ").\n",
20516 (UV)REGNODE_OFFSET(place) > RExC_offsets[0]
20517 ? "Overwriting end of array!\n" : "OK",
20518 (UV)REGNODE_OFFSET(place),
20519 (UV)(RExC_parse - RExC_start),
20520 (UV)RExC_offsets[0]));
20521 Set_Node_Offset(place, RExC_parse);
20522 Set_Node_Length(place, 1);
20525 src = NEXTOPER(place);
20527 FILL_NODE(operand, op);
20529 /* Zero out any arguments in the new node */
20530 Zero(src, offset, regnode);
20534 - regtail - set the next-pointer at the end of a node chain of p to val. If
20535 that value won't fit in the space available, instead returns FALSE.
20536 (Except asserts if we can't fit in the largest space the regex
20537 engine is designed for.)
20538 - SEE ALSO: regtail_study
20541 S_regtail(pTHX_ RExC_state_t * pRExC_state,
20542 const regnode_offset p,
20543 const regnode_offset val,
20546 regnode_offset scan;
20547 DECLARE_AND_GET_RE_DEBUG_FLAGS;
20549 PERL_ARGS_ASSERT_REGTAIL;
20551 PERL_UNUSED_ARG(depth);
20554 /* Find last node. */
20555 scan = (regnode_offset) p;
20557 regnode * const temp = regnext(REGNODE_p(scan));
20559 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
20560 regprop(RExC_rx, RExC_mysv, REGNODE_p(scan), NULL, pRExC_state);
20561 Perl_re_printf( aTHX_ "~ %s (%zu) %s %s\n",
20562 SvPV_nolen_const(RExC_mysv), scan,
20563 (temp == NULL ? "->" : ""),
20564 (temp == NULL ? PL_reg_name[OP(REGNODE_p(val))] : "")
20569 scan = REGNODE_OFFSET(temp);
20572 assert(val >= scan);
20573 if (reg_off_by_arg[OP(REGNODE_p(scan))]) {
20574 assert((UV) (val - scan) <= U32_MAX);
20575 ARG_SET(REGNODE_p(scan), val - scan);
20578 if (val - scan > U16_MAX) {
20579 /* Populate this with something that won't loop and will likely
20580 * lead to a crash if the caller ignores the failure return, and
20581 * execution continues */
20582 NEXT_OFF(REGNODE_p(scan)) = U16_MAX;
20585 NEXT_OFF(REGNODE_p(scan)) = val - scan;
20593 - regtail_study - set the next-pointer at the end of a node chain of p to val.
20594 - Look for optimizable sequences at the same time.
20595 - currently only looks for EXACT chains.
20597 This is experimental code. The idea is to use this routine to perform
20598 in place optimizations on branches and groups as they are constructed,
20599 with the long term intention of removing optimization from study_chunk so
20600 that it is purely analytical.
20602 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
20603 to control which is which.
20605 This used to return a value that was ignored. It was a problem that it is
20606 #ifdef'd to be another function that didn't return a value. khw has changed it
20607 so both currently return a pass/fail return.
20610 /* TODO: All four parms should be const */
20613 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode_offset p,
20614 const regnode_offset val, U32 depth)
20616 regnode_offset scan;
20618 #ifdef EXPERIMENTAL_INPLACESCAN
20621 DECLARE_AND_GET_RE_DEBUG_FLAGS;
20623 PERL_ARGS_ASSERT_REGTAIL_STUDY;
20626 /* Find last node. */
20630 regnode * const temp = regnext(REGNODE_p(scan));
20631 #ifdef EXPERIMENTAL_INPLACESCAN
20632 if (PL_regkind[OP(REGNODE_p(scan))] == EXACT) {
20633 bool unfolded_multi_char; /* Unexamined in this routine */
20634 if (join_exact(pRExC_state, scan, &min,
20635 &unfolded_multi_char, 1, REGNODE_p(val), depth+1))
20636 return TRUE; /* Was return EXACT */
20640 switch (OP(REGNODE_p(scan))) {
20647 case EXACTFU_S_EDGE:
20648 case EXACTFAA_NO_TRIE:
20655 if( exact == PSEUDO )
20656 exact= OP(REGNODE_p(scan));
20657 else if ( exact != OP(REGNODE_p(scan)) )
20666 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
20667 regprop(RExC_rx, RExC_mysv, REGNODE_p(scan), NULL, pRExC_state);
20668 Perl_re_printf( aTHX_ "~ %s (%zu) -> %s\n",
20669 SvPV_nolen_const(RExC_mysv),
20671 PL_reg_name[exact]);
20675 scan = REGNODE_OFFSET(temp);
20678 DEBUG_PARSE_MSG("");
20679 regprop(RExC_rx, RExC_mysv, REGNODE_p(val), NULL, pRExC_state);
20680 Perl_re_printf( aTHX_
20681 "~ attach to %s (%" IVdf ") offset to %" IVdf "\n",
20682 SvPV_nolen_const(RExC_mysv),
20687 if (reg_off_by_arg[OP(REGNODE_p(scan))]) {
20688 assert((UV) (val - scan) <= U32_MAX);
20689 ARG_SET(REGNODE_p(scan), val - scan);
20692 if (val - scan > U16_MAX) {
20693 /* Populate this with something that won't loop and will likely
20694 * lead to a crash if the caller ignores the failure return, and
20695 * execution continues */
20696 NEXT_OFF(REGNODE_p(scan)) = U16_MAX;
20699 NEXT_OFF(REGNODE_p(scan)) = val - scan;
20702 return TRUE; /* Was 'return exact' */
20707 S_get_ANYOFM_contents(pTHX_ const regnode * n) {
20709 /* Returns an inversion list of all the code points matched by the
20710 * ANYOFM/NANYOFM node 'n' */
20712 SV * cp_list = _new_invlist(-1);
20713 const U8 lowest = (U8) ARG(n);
20716 U8 needed = 1U << PL_bitcount[ (U8) ~ FLAGS(n)];
20718 PERL_ARGS_ASSERT_GET_ANYOFM_CONTENTS;
20720 /* Starting with the lowest code point, any code point that ANDed with the
20721 * mask yields the lowest code point is in the set */
20722 for (i = lowest; i <= 0xFF; i++) {
20723 if ((i & FLAGS(n)) == ARG(n)) {
20724 cp_list = add_cp_to_invlist(cp_list, i);
20727 /* We know how many code points (a power of two) that are in the
20728 * set. No use looking once we've got that number */
20729 if (count >= needed) break;
20733 if (OP(n) == NANYOFM) {
20734 _invlist_invert(cp_list);
20740 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
20745 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
20750 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
20752 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
20753 if (flags & (1<<bit)) {
20754 if (!set++ && lead)
20755 Perl_re_printf( aTHX_ "%s", lead);
20756 Perl_re_printf( aTHX_ "%s ", PL_reg_intflags_name[bit]);
20761 Perl_re_printf( aTHX_ "\n");
20763 Perl_re_printf( aTHX_ "%s[none-set]\n", lead);
20768 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
20774 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
20776 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
20777 if (flags & (1<<bit)) {
20778 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
20781 if (!set++ && lead)
20782 Perl_re_printf( aTHX_ "%s", lead);
20783 Perl_re_printf( aTHX_ "%s ", PL_reg_extflags_name[bit]);
20786 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
20787 if (!set++ && lead) {
20788 Perl_re_printf( aTHX_ "%s", lead);
20791 case REGEX_UNICODE_CHARSET:
20792 Perl_re_printf( aTHX_ "UNICODE");
20794 case REGEX_LOCALE_CHARSET:
20795 Perl_re_printf( aTHX_ "LOCALE");
20797 case REGEX_ASCII_RESTRICTED_CHARSET:
20798 Perl_re_printf( aTHX_ "ASCII-RESTRICTED");
20800 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
20801 Perl_re_printf( aTHX_ "ASCII-MORE_RESTRICTED");
20804 Perl_re_printf( aTHX_ "UNKNOWN CHARACTER SET");
20810 Perl_re_printf( aTHX_ "\n");
20812 Perl_re_printf( aTHX_ "%s[none-set]\n", lead);
20818 Perl_regdump(pTHX_ const regexp *r)
20822 SV * const sv = sv_newmortal();
20823 SV *dsv= sv_newmortal();
20824 RXi_GET_DECL(r, ri);
20825 DECLARE_AND_GET_RE_DEBUG_FLAGS;
20827 PERL_ARGS_ASSERT_REGDUMP;
20829 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
20831 /* Header fields of interest. */
20832 for (i = 0; i < 2; i++) {
20833 if (r->substrs->data[i].substr) {
20834 RE_PV_QUOTED_DECL(s, 0, dsv,
20835 SvPVX_const(r->substrs->data[i].substr),
20836 RE_SV_DUMPLEN(r->substrs->data[i].substr),
20837 PL_dump_re_max_len);
20838 Perl_re_printf( aTHX_
20839 "%s %s%s at %" IVdf "..%" UVuf " ",
20840 i ? "floating" : "anchored",
20842 RE_SV_TAIL(r->substrs->data[i].substr),
20843 (IV)r->substrs->data[i].min_offset,
20844 (UV)r->substrs->data[i].max_offset);
20846 else if (r->substrs->data[i].utf8_substr) {
20847 RE_PV_QUOTED_DECL(s, 1, dsv,
20848 SvPVX_const(r->substrs->data[i].utf8_substr),
20849 RE_SV_DUMPLEN(r->substrs->data[i].utf8_substr),
20851 Perl_re_printf( aTHX_
20852 "%s utf8 %s%s at %" IVdf "..%" UVuf " ",
20853 i ? "floating" : "anchored",
20855 RE_SV_TAIL(r->substrs->data[i].utf8_substr),
20856 (IV)r->substrs->data[i].min_offset,
20857 (UV)r->substrs->data[i].max_offset);
20861 if (r->check_substr || r->check_utf8)
20862 Perl_re_printf( aTHX_
20864 ( r->check_substr == r->substrs->data[1].substr
20865 && r->check_utf8 == r->substrs->data[1].utf8_substr
20866 ? "(checking floating" : "(checking anchored"));
20867 if (r->intflags & PREGf_NOSCAN)
20868 Perl_re_printf( aTHX_ " noscan");
20869 if (r->extflags & RXf_CHECK_ALL)
20870 Perl_re_printf( aTHX_ " isall");
20871 if (r->check_substr || r->check_utf8)
20872 Perl_re_printf( aTHX_ ") ");
20874 if (ri->regstclass) {
20875 regprop(r, sv, ri->regstclass, NULL, NULL);
20876 Perl_re_printf( aTHX_ "stclass %s ", SvPVX_const(sv));
20878 if (r->intflags & PREGf_ANCH) {
20879 Perl_re_printf( aTHX_ "anchored");
20880 if (r->intflags & PREGf_ANCH_MBOL)
20881 Perl_re_printf( aTHX_ "(MBOL)");
20882 if (r->intflags & PREGf_ANCH_SBOL)
20883 Perl_re_printf( aTHX_ "(SBOL)");
20884 if (r->intflags & PREGf_ANCH_GPOS)
20885 Perl_re_printf( aTHX_ "(GPOS)");
20886 Perl_re_printf( aTHX_ " ");
20888 if (r->intflags & PREGf_GPOS_SEEN)
20889 Perl_re_printf( aTHX_ "GPOS:%" UVuf " ", (UV)r->gofs);
20890 if (r->intflags & PREGf_SKIP)
20891 Perl_re_printf( aTHX_ "plus ");
20892 if (r->intflags & PREGf_IMPLICIT)
20893 Perl_re_printf( aTHX_ "implicit ");
20894 Perl_re_printf( aTHX_ "minlen %" IVdf " ", (IV)r->minlen);
20895 if (r->extflags & RXf_EVAL_SEEN)
20896 Perl_re_printf( aTHX_ "with eval ");
20897 Perl_re_printf( aTHX_ "\n");
20899 regdump_extflags("r->extflags: ", r->extflags);
20900 regdump_intflags("r->intflags: ", r->intflags);
20903 PERL_ARGS_ASSERT_REGDUMP;
20904 PERL_UNUSED_CONTEXT;
20905 PERL_UNUSED_ARG(r);
20906 #endif /* DEBUGGING */
20909 /* Should be synchronized with ANYOF_ #defines in regcomp.h */
20912 # if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 \
20913 || _CC_LOWER != 3 || _CC_UPPER != 4 || _CC_PUNCT != 5 \
20914 || _CC_PRINT != 6 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 \
20915 || _CC_CASED != 9 || _CC_SPACE != 10 || _CC_BLANK != 11 \
20916 || _CC_XDIGIT != 12 || _CC_CNTRL != 13 || _CC_ASCII != 14 \
20917 || _CC_VERTSPACE != 15
20918 # error Need to adjust order of anyofs[]
20920 static const char * const anyofs[] = {
20957 - regprop - printable representation of opcode, with run time support
20961 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo, const RExC_state_t *pRExC_state)
20966 RXi_GET_DECL(prog, progi);
20967 DECLARE_AND_GET_RE_DEBUG_FLAGS;
20969 PERL_ARGS_ASSERT_REGPROP;
20973 if (OP(o) > REGNODE_MAX) { /* regnode.type is unsigned */
20974 if (pRExC_state) { /* This gives more info, if we have it */
20975 FAIL3("panic: corrupted regexp opcode %d > %d",
20976 (int)OP(o), (int)REGNODE_MAX);
20979 Perl_croak(aTHX_ "panic: corrupted regexp opcode %d > %d",
20980 (int)OP(o), (int)REGNODE_MAX);
20983 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
20985 k = PL_regkind[OP(o)];
20988 sv_catpvs(sv, " ");
20989 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
20990 * is a crude hack but it may be the best for now since
20991 * we have no flag "this EXACTish node was UTF-8"
20993 pv_pretty(sv, STRING(o), STR_LEN(o), PL_dump_re_max_len,
20994 PL_colors[0], PL_colors[1],
20995 PERL_PV_ESCAPE_UNI_DETECT |
20996 PERL_PV_ESCAPE_NONASCII |
20997 PERL_PV_PRETTY_ELLIPSES |
20998 PERL_PV_PRETTY_LTGT |
20999 PERL_PV_PRETTY_NOCLEAR
21001 } else if (k == TRIE) {
21002 /* print the details of the trie in dumpuntil instead, as
21003 * progi->data isn't available here */
21004 const char op = OP(o);
21005 const U32 n = ARG(o);
21006 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
21007 (reg_ac_data *)progi->data->data[n] :
21009 const reg_trie_data * const trie
21010 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
21012 Perl_sv_catpvf(aTHX_ sv, "-%s", PL_reg_name[o->flags]);
21013 DEBUG_TRIE_COMPILE_r({
21015 sv_catpvs(sv, "(JUMP)");
21016 Perl_sv_catpvf(aTHX_ sv,
21017 "<S:%" UVuf "/%" IVdf " W:%" UVuf " L:%" UVuf "/%" UVuf " C:%" UVuf "/%" UVuf ">",
21018 (UV)trie->startstate,
21019 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
21020 (UV)trie->wordcount,
21023 (UV)TRIE_CHARCOUNT(trie),
21024 (UV)trie->uniquecharcount
21027 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
21028 sv_catpvs(sv, "[");
21029 (void) put_charclass_bitmap_innards(sv,
21030 ((IS_ANYOF_TRIE(op))
21032 : TRIE_BITMAP(trie)),
21039 sv_catpvs(sv, "]");
21041 } else if (k == CURLY) {
21042 U32 lo = ARG1(o), hi = ARG2(o);
21043 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
21044 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
21045 Perl_sv_catpvf(aTHX_ sv, "{%u,", (unsigned) lo);
21046 if (hi == REG_INFTY)
21047 sv_catpvs(sv, "INFTY");
21049 Perl_sv_catpvf(aTHX_ sv, "%u", (unsigned) hi);
21050 sv_catpvs(sv, "}");
21052 else if (k == WHILEM && o->flags) /* Ordinal/of */
21053 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
21054 else if (k == REF || k == OPEN || k == CLOSE
21055 || k == GROUPP || OP(o)==ACCEPT)
21057 AV *name_list= NULL;
21058 U32 parno= OP(o) == ACCEPT ? (U32)ARG2L(o) : ARG(o);
21059 Perl_sv_catpvf(aTHX_ sv, "%" UVuf, (UV)parno); /* Parenth number */
21060 if ( RXp_PAREN_NAMES(prog) ) {
21061 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
21062 } else if ( pRExC_state ) {
21063 name_list= RExC_paren_name_list;
21066 if ( k != REF || (OP(o) < REFN)) {
21067 SV **name= av_fetch(name_list, parno, 0 );
21069 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
21072 SV *sv_dat= MUTABLE_SV(progi->data->data[ parno ]);
21073 I32 *nums=(I32*)SvPVX(sv_dat);
21074 SV **name= av_fetch(name_list, nums[0], 0 );
21077 for ( n=0; n<SvIVX(sv_dat); n++ ) {
21078 Perl_sv_catpvf(aTHX_ sv, "%s%" IVdf,
21079 (n ? "," : ""), (IV)nums[n]);
21081 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
21085 if ( k == REF && reginfo) {
21086 U32 n = ARG(o); /* which paren pair */
21087 I32 ln = prog->offs[n].start;
21088 if (prog->lastparen < n || ln == -1 || prog->offs[n].end == -1)
21089 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
21090 else if (ln == prog->offs[n].end)
21091 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
21093 const char *s = reginfo->strbeg + ln;
21094 Perl_sv_catpvf(aTHX_ sv, ": ");
21095 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
21096 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
21099 } else if (k == GOSUB) {
21100 AV *name_list= NULL;
21101 if ( RXp_PAREN_NAMES(prog) ) {
21102 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
21103 } else if ( pRExC_state ) {
21104 name_list= RExC_paren_name_list;
21107 /* Paren and offset */
21108 Perl_sv_catpvf(aTHX_ sv, "%d[%+d:%d]", (int)ARG(o),(int)ARG2L(o),
21109 (int)((o + (int)ARG2L(o)) - progi->program) );
21111 SV **name= av_fetch(name_list, ARG(o), 0 );
21113 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
21116 else if (k == LOGICAL)
21117 /* 2: embedded, otherwise 1 */
21118 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
21119 else if (k == ANYOF || k == ANYOFR) {
21123 bool do_sep = FALSE; /* Do we need to separate various components of
21125 /* Set if there is still an unresolved user-defined property */
21126 SV *unresolved = NULL;
21128 /* Things that are ignored except when the runtime locale is UTF-8 */
21129 SV *only_utf8_locale_invlist = NULL;
21131 /* Code points that don't fit in the bitmap */
21132 SV *nonbitmap_invlist = NULL;
21134 /* And things that aren't in the bitmap, but are small enough to be */
21135 SV* bitmap_range_not_in_bitmap = NULL;
21139 if (inRANGE(OP(o), ANYOFH, ANYOFRb)) {
21145 flags = ANYOF_FLAGS(o);
21146 bitmap = ANYOF_BITMAP(o);
21150 if (OP(o) == ANYOFL || OP(o) == ANYOFPOSIXL) {
21151 if (ANYOFL_UTF8_LOCALE_REQD(flags)) {
21152 sv_catpvs(sv, "{utf8-locale-reqd}");
21154 if (flags & ANYOFL_FOLD) {
21155 sv_catpvs(sv, "{i}");
21159 inverted = flags & ANYOF_INVERT;
21161 /* If there is stuff outside the bitmap, get it */
21162 if (arg != ANYOF_ONLY_HAS_BITMAP) {
21163 if (inRANGE(OP(o), ANYOFR, ANYOFRb)) {
21164 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
21166 ANYOFRbase(o) + ANYOFRdelta(o));
21169 (void) _get_regclass_nonbitmap_data(prog, o, FALSE,
21171 &only_utf8_locale_invlist,
21172 &nonbitmap_invlist);
21175 /* The non-bitmap data may contain stuff that could fit in the
21176 * bitmap. This could come from a user-defined property being
21177 * finally resolved when this call was done; or much more likely
21178 * because there are matches that require UTF-8 to be valid, and so
21179 * aren't in the bitmap (or ANYOFR). This is teased apart later */
21180 _invlist_intersection(nonbitmap_invlist,
21182 &bitmap_range_not_in_bitmap);
21183 /* Leave just the things that don't fit into the bitmap */
21184 _invlist_subtract(nonbitmap_invlist,
21186 &nonbitmap_invlist);
21189 /* Obey this flag to add all above-the-bitmap code points */
21190 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
21191 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
21192 NUM_ANYOF_CODE_POINTS,
21196 /* Ready to start outputting. First, the initial left bracket */
21197 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
21199 /* ANYOFH by definition doesn't have anything that will fit inside the
21200 * bitmap; ANYOFR may or may not. */
21201 if ( ! inRANGE(OP(o), ANYOFH, ANYOFHr)
21202 && ( ! inRANGE(OP(o), ANYOFR, ANYOFRb)
21203 || ANYOFRbase(o) < NUM_ANYOF_CODE_POINTS))
21205 /* Then all the things that could fit in the bitmap */
21206 do_sep = put_charclass_bitmap_innards(sv,
21208 bitmap_range_not_in_bitmap,
21209 only_utf8_locale_invlist,
21213 /* Can't try inverting for a
21214 * better display if there
21215 * are things that haven't
21218 || inRANGE(OP(o), ANYOFR, ANYOFRb));
21219 SvREFCNT_dec(bitmap_range_not_in_bitmap);
21221 /* If there are user-defined properties which haven't been defined
21222 * yet, output them. If the result is not to be inverted, it is
21223 * clearest to output them in a separate [] from the bitmap range
21224 * stuff. If the result is to be complemented, we have to show
21225 * everything in one [], as the inversion applies to the whole
21226 * thing. Use {braces} to separate them from anything in the
21227 * bitmap and anything above the bitmap. */
21230 if (! do_sep) { /* If didn't output anything in the bitmap
21232 sv_catpvs(sv, "^");
21234 sv_catpvs(sv, "{");
21237 Perl_sv_catpvf(aTHX_ sv,"%s][%s", PL_colors[1],
21240 sv_catsv(sv, unresolved);
21242 sv_catpvs(sv, "}");
21244 do_sep = ! inverted;
21248 /* And, finally, add the above-the-bitmap stuff */
21249 if (nonbitmap_invlist && _invlist_len(nonbitmap_invlist)) {
21252 /* See if truncation size is overridden */
21253 const STRLEN dump_len = (PL_dump_re_max_len > 256)
21254 ? PL_dump_re_max_len
21257 /* This is output in a separate [] */
21259 Perl_sv_catpvf(aTHX_ sv,"%s][%s", PL_colors[1], PL_colors[0]);
21262 /* And, for easy of understanding, it is shown in the
21263 * uncomplemented form if possible. The one exception being if
21264 * there are unresolved items, where the inversion has to be
21265 * delayed until runtime */
21266 if (inverted && ! unresolved) {
21267 _invlist_invert(nonbitmap_invlist);
21268 _invlist_subtract(nonbitmap_invlist, PL_InBitmap, &nonbitmap_invlist);
21271 contents = invlist_contents(nonbitmap_invlist,
21272 FALSE /* output suitable for catsv */
21275 /* If the output is shorter than the permissible maximum, just do it. */
21276 if (SvCUR(contents) <= dump_len) {
21277 sv_catsv(sv, contents);
21280 const char * contents_string = SvPVX(contents);
21281 STRLEN i = dump_len;
21283 /* Otherwise, start at the permissible max and work back to the
21284 * first break possibility */
21285 while (i > 0 && contents_string[i] != ' ') {
21288 if (i == 0) { /* Fail-safe. Use the max if we couldn't
21289 find a legal break */
21293 sv_catpvn(sv, contents_string, i);
21294 sv_catpvs(sv, "...");
21297 SvREFCNT_dec_NN(contents);
21298 SvREFCNT_dec_NN(nonbitmap_invlist);
21301 /* And finally the matching, closing ']' */
21302 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
21304 if (OP(o) == ANYOFHs) {
21305 Perl_sv_catpvf(aTHX_ sv, " (Leading UTF-8 bytes=%s", _byte_dump_string((U8 *) ((struct regnode_anyofhs *) o)->string, FLAGS(o), 1));
21307 else if (inRANGE(OP(o), ANYOFH, ANYOFRb)) {
21308 U8 lowest = (OP(o) != ANYOFHr)
21310 : LOWEST_ANYOF_HRx_BYTE(FLAGS(o));
21311 U8 highest = (OP(o) == ANYOFHr)
21312 ? HIGHEST_ANYOF_HRx_BYTE(FLAGS(o))
21313 : (OP(o) == ANYOFH || OP(o) == ANYOFR)
21316 Perl_sv_catpvf(aTHX_ sv, " (First UTF-8 byte=%02X", lowest);
21317 if (lowest != highest) {
21318 Perl_sv_catpvf(aTHX_ sv, "-%02X", highest);
21320 Perl_sv_catpvf(aTHX_ sv, ")");
21323 SvREFCNT_dec(unresolved);
21325 else if (k == ANYOFM) {
21326 SV * cp_list = get_ANYOFM_contents(o);
21328 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
21329 if (OP(o) == NANYOFM) {
21330 _invlist_invert(cp_list);
21333 put_charclass_bitmap_innards(sv, NULL, cp_list, NULL, NULL, 0, TRUE);
21334 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
21336 SvREFCNT_dec(cp_list);
21338 else if (k == POSIXD || k == NPOSIXD) {
21339 U8 index = FLAGS(o) * 2;
21340 if (index < C_ARRAY_LENGTH(anyofs)) {
21341 if (*anyofs[index] != '[') {
21342 sv_catpvs(sv, "[");
21344 sv_catpv(sv, anyofs[index]);
21345 if (*anyofs[index] != '[') {
21346 sv_catpvs(sv, "]");
21350 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
21353 else if (k == BOUND || k == NBOUND) {
21354 /* Must be synced with order of 'bound_type' in regcomp.h */
21355 const char * const bounds[] = {
21356 "", /* Traditional */
21362 assert(FLAGS(o) < C_ARRAY_LENGTH(bounds));
21363 sv_catpv(sv, bounds[FLAGS(o)]);
21365 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH)) {
21366 Perl_sv_catpvf(aTHX_ sv, "[%d", -(o->flags));
21368 Perl_sv_catpvf(aTHX_ sv, "..-%d", o->flags - o->next_off);
21370 Perl_sv_catpvf(aTHX_ sv, "]");
21372 else if (OP(o) == SBOL)
21373 Perl_sv_catpvf(aTHX_ sv, " /%s/", o->flags ? "\\A" : "^");
21375 /* add on the verb argument if there is one */
21376 if ( ( k == VERB || OP(o) == ACCEPT || OP(o) == OPFAIL ) && o->flags) {
21378 Perl_sv_catpvf(aTHX_ sv, ":%" SVf,
21379 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
21381 sv_catpvs(sv, ":NULL");
21384 PERL_UNUSED_CONTEXT;
21385 PERL_UNUSED_ARG(sv);
21386 PERL_UNUSED_ARG(o);
21387 PERL_UNUSED_ARG(prog);
21388 PERL_UNUSED_ARG(reginfo);
21389 PERL_UNUSED_ARG(pRExC_state);
21390 #endif /* DEBUGGING */
21396 Perl_re_intuit_string(pTHX_ REGEXP * const r)
21397 { /* Assume that RE_INTUIT is set */
21398 /* Returns an SV containing a string that must appear in the target for it
21401 struct regexp *const prog = ReANY(r);
21402 DECLARE_AND_GET_RE_DEBUG_FLAGS;
21404 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
21405 PERL_UNUSED_CONTEXT;
21409 if (prog->maxlen > 0) {
21410 const char * const s = SvPV_nolen_const(RX_UTF8(r)
21411 ? prog->check_utf8 : prog->check_substr);
21413 if (!PL_colorset) reginitcolors();
21414 Perl_re_printf( aTHX_
21415 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
21417 RX_UTF8(r) ? "utf8 " : "",
21418 PL_colors[5], PL_colors[0],
21421 (strlen(s) > PL_dump_re_max_len ? "..." : ""));
21425 /* use UTF8 check substring if regexp pattern itself is in UTF8 */
21426 return RX_UTF8(r) ? prog->check_utf8 : prog->check_substr;
21432 handles refcounting and freeing the perl core regexp structure. When
21433 it is necessary to actually free the structure the first thing it
21434 does is call the 'free' method of the regexp_engine associated to
21435 the regexp, allowing the handling of the void *pprivate; member
21436 first. (This routine is not overridable by extensions, which is why
21437 the extensions free is called first.)
21439 See regdupe and regdupe_internal if you change anything here.
21441 #ifndef PERL_IN_XSUB_RE
21443 Perl_pregfree(pTHX_ REGEXP *r)
21449 Perl_pregfree2(pTHX_ REGEXP *rx)
21451 struct regexp *const r = ReANY(rx);
21452 DECLARE_AND_GET_RE_DEBUG_FLAGS;
21454 PERL_ARGS_ASSERT_PREGFREE2;
21459 if (r->mother_re) {
21460 ReREFCNT_dec(r->mother_re);
21462 CALLREGFREE_PVT(rx); /* free the private data */
21463 SvREFCNT_dec(RXp_PAREN_NAMES(r));
21467 for (i = 0; i < 2; i++) {
21468 SvREFCNT_dec(r->substrs->data[i].substr);
21469 SvREFCNT_dec(r->substrs->data[i].utf8_substr);
21471 Safefree(r->substrs);
21473 RX_MATCH_COPY_FREE(rx);
21474 #ifdef PERL_ANY_COW
21475 SvREFCNT_dec(r->saved_copy);
21478 SvREFCNT_dec(r->qr_anoncv);
21479 if (r->recurse_locinput)
21480 Safefree(r->recurse_locinput);
21486 Copy ssv to dsv, both of which should of type SVt_REGEXP or SVt_PVLV,
21487 except that dsv will be created if NULL.
21489 This function is used in two main ways. First to implement
21490 $r = qr/....; $s = $$r;
21492 Secondly, it is used as a hacky workaround to the structural issue of
21494 being stored in the regexp structure which is in turn stored in
21495 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
21496 could be PL_curpm in multiple contexts, and could require multiple
21497 result sets being associated with the pattern simultaneously, such
21498 as when doing a recursive match with (??{$qr})
21500 The solution is to make a lightweight copy of the regexp structure
21501 when a qr// is returned from the code executed by (??{$qr}) this
21502 lightweight copy doesn't actually own any of its data except for
21503 the starp/end and the actual regexp structure itself.
21509 Perl_reg_temp_copy(pTHX_ REGEXP *dsv, REGEXP *ssv)
21511 struct regexp *drx;
21512 struct regexp *const srx = ReANY(ssv);
21513 const bool islv = dsv && SvTYPE(dsv) == SVt_PVLV;
21515 PERL_ARGS_ASSERT_REG_TEMP_COPY;
21518 dsv = (REGEXP*) newSV_type(SVt_REGEXP);
21520 assert(SvTYPE(dsv) == SVt_REGEXP || (SvTYPE(dsv) == SVt_PVLV));
21522 /* our only valid caller, sv_setsv_flags(), should have done
21523 * a SV_CHECK_THINKFIRST_COW_DROP() by now */
21524 assert(!SvOOK(dsv));
21525 assert(!SvIsCOW(dsv));
21526 assert(!SvROK(dsv));
21528 if (SvPVX_const(dsv)) {
21530 Safefree(SvPVX(dsv));
21535 SvOK_off((SV *)dsv);
21538 /* For PVLVs, the head (sv_any) points to an XPVLV, while
21539 * the LV's xpvlenu_rx will point to a regexp body, which
21540 * we allocate here */
21541 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
21542 assert(!SvPVX(dsv));
21543 ((XPV*)SvANY(dsv))->xpv_len_u.xpvlenu_rx = temp->sv_any;
21544 temp->sv_any = NULL;
21545 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
21546 SvREFCNT_dec_NN(temp);
21547 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
21548 ing below will not set it. */
21549 SvCUR_set(dsv, SvCUR(ssv));
21552 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
21553 sv_force_normal(sv) is called. */
21557 SvFLAGS(dsv) |= SvFLAGS(ssv) & (SVf_POK|SVp_POK|SVf_UTF8);
21558 SvPV_set(dsv, RX_WRAPPED(ssv));
21559 /* We share the same string buffer as the original regexp, on which we
21560 hold a reference count, incremented when mother_re is set below.
21561 The string pointer is copied here, being part of the regexp struct.
21563 memcpy(&(drx->xpv_cur), &(srx->xpv_cur),
21564 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
21568 const I32 npar = srx->nparens+1;
21569 Newx(drx->offs, npar, regexp_paren_pair);
21570 Copy(srx->offs, drx->offs, npar, regexp_paren_pair);
21572 if (srx->substrs) {
21574 Newx(drx->substrs, 1, struct reg_substr_data);
21575 StructCopy(srx->substrs, drx->substrs, struct reg_substr_data);
21577 for (i = 0; i < 2; i++) {
21578 SvREFCNT_inc_void(drx->substrs->data[i].substr);
21579 SvREFCNT_inc_void(drx->substrs->data[i].utf8_substr);
21582 /* check_substr and check_utf8, if non-NULL, point to either their
21583 anchored or float namesakes, and don't hold a second reference. */
21585 RX_MATCH_COPIED_off(dsv);
21586 #ifdef PERL_ANY_COW
21587 drx->saved_copy = NULL;
21589 drx->mother_re = ReREFCNT_inc(srx->mother_re ? srx->mother_re : ssv);
21590 SvREFCNT_inc_void(drx->qr_anoncv);
21591 if (srx->recurse_locinput)
21592 Newx(drx->recurse_locinput, srx->nparens + 1, char *);
21599 /* regfree_internal()
21601 Free the private data in a regexp. This is overloadable by
21602 extensions. Perl takes care of the regexp structure in pregfree(),
21603 this covers the *pprivate pointer which technically perl doesn't
21604 know about, however of course we have to handle the
21605 regexp_internal structure when no extension is in use.
21607 Note this is called before freeing anything in the regexp
21612 Perl_regfree_internal(pTHX_ REGEXP * const rx)
21614 struct regexp *const r = ReANY(rx);
21615 RXi_GET_DECL(r, ri);
21616 DECLARE_AND_GET_RE_DEBUG_FLAGS;
21618 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
21628 SV *dsv= sv_newmortal();
21629 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
21630 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), PL_dump_re_max_len);
21631 Perl_re_printf( aTHX_ "%sFreeing REx:%s %s\n",
21632 PL_colors[4], PL_colors[5], s);
21636 #ifdef RE_TRACK_PATTERN_OFFSETS
21638 Safefree(ri->u.offsets); /* 20010421 MJD */
21640 if (ri->code_blocks)
21641 S_free_codeblocks(aTHX_ ri->code_blocks);
21644 int n = ri->data->count;
21647 /* If you add a ->what type here, update the comment in regcomp.h */
21648 switch (ri->data->what[n]) {
21654 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
21657 Safefree(ri->data->data[n]);
21663 { /* Aho Corasick add-on structure for a trie node.
21664 Used in stclass optimization only */
21666 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
21667 #ifdef USE_ITHREADS
21671 refcount = --aho->refcount;
21674 PerlMemShared_free(aho->states);
21675 PerlMemShared_free(aho->fail);
21676 /* do this last!!!! */
21677 PerlMemShared_free(ri->data->data[n]);
21678 /* we should only ever get called once, so
21679 * assert as much, and also guard the free
21680 * which /might/ happen twice. At the least
21681 * it will make code anlyzers happy and it
21682 * doesn't cost much. - Yves */
21683 assert(ri->regstclass);
21684 if (ri->regstclass) {
21685 PerlMemShared_free(ri->regstclass);
21686 ri->regstclass = 0;
21693 /* trie structure. */
21695 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
21696 #ifdef USE_ITHREADS
21700 refcount = --trie->refcount;
21703 PerlMemShared_free(trie->charmap);
21704 PerlMemShared_free(trie->states);
21705 PerlMemShared_free(trie->trans);
21707 PerlMemShared_free(trie->bitmap);
21709 PerlMemShared_free(trie->jump);
21710 PerlMemShared_free(trie->wordinfo);
21711 /* do this last!!!! */
21712 PerlMemShared_free(ri->data->data[n]);
21717 Perl_croak(aTHX_ "panic: regfree data code '%c'",
21718 ri->data->what[n]);
21721 Safefree(ri->data->what);
21722 Safefree(ri->data);
21728 #define av_dup_inc(s, t) MUTABLE_AV(sv_dup_inc((const SV *)s, t))
21729 #define hv_dup_inc(s, t) MUTABLE_HV(sv_dup_inc((const SV *)s, t))
21730 #define SAVEPVN(p, n) ((p) ? savepvn(p, n) : NULL)
21733 re_dup_guts - duplicate a regexp.
21735 This routine is expected to clone a given regexp structure. It is only
21736 compiled under USE_ITHREADS.
21738 After all of the core data stored in struct regexp is duplicated
21739 the regexp_engine.dupe method is used to copy any private data
21740 stored in the *pprivate pointer. This allows extensions to handle
21741 any duplication it needs to do.
21743 See pregfree() and regfree_internal() if you change anything here.
21745 #if defined(USE_ITHREADS)
21746 #ifndef PERL_IN_XSUB_RE
21748 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
21752 const struct regexp *r = ReANY(sstr);
21753 struct regexp *ret = ReANY(dstr);
21755 PERL_ARGS_ASSERT_RE_DUP_GUTS;
21757 npar = r->nparens+1;
21758 Newx(ret->offs, npar, regexp_paren_pair);
21759 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
21761 if (ret->substrs) {
21762 /* Do it this way to avoid reading from *r after the StructCopy().
21763 That way, if any of the sv_dup_inc()s dislodge *r from the L1
21764 cache, it doesn't matter. */
21766 const bool anchored = r->check_substr
21767 ? r->check_substr == r->substrs->data[0].substr
21768 : r->check_utf8 == r->substrs->data[0].utf8_substr;
21769 Newx(ret->substrs, 1, struct reg_substr_data);
21770 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
21772 for (i = 0; i < 2; i++) {
21773 ret->substrs->data[i].substr =
21774 sv_dup_inc(ret->substrs->data[i].substr, param);
21775 ret->substrs->data[i].utf8_substr =
21776 sv_dup_inc(ret->substrs->data[i].utf8_substr, param);
21779 /* check_substr and check_utf8, if non-NULL, point to either their
21780 anchored or float namesakes, and don't hold a second reference. */
21782 if (ret->check_substr) {
21784 assert(r->check_utf8 == r->substrs->data[0].utf8_substr);
21786 ret->check_substr = ret->substrs->data[0].substr;
21787 ret->check_utf8 = ret->substrs->data[0].utf8_substr;
21789 assert(r->check_substr == r->substrs->data[1].substr);
21790 assert(r->check_utf8 == r->substrs->data[1].utf8_substr);
21792 ret->check_substr = ret->substrs->data[1].substr;
21793 ret->check_utf8 = ret->substrs->data[1].utf8_substr;
21795 } else if (ret->check_utf8) {
21797 ret->check_utf8 = ret->substrs->data[0].utf8_substr;
21799 ret->check_utf8 = ret->substrs->data[1].utf8_substr;
21804 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
21805 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
21806 if (r->recurse_locinput)
21807 Newx(ret->recurse_locinput, r->nparens + 1, char *);
21810 RXi_SET(ret, CALLREGDUPE_PVT(dstr, param));
21812 if (RX_MATCH_COPIED(dstr))
21813 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
21815 ret->subbeg = NULL;
21816 #ifdef PERL_ANY_COW
21817 ret->saved_copy = NULL;
21820 /* Whether mother_re be set or no, we need to copy the string. We
21821 cannot refrain from copying it when the storage points directly to
21822 our mother regexp, because that's
21823 1: a buffer in a different thread
21824 2: something we no longer hold a reference on
21825 so we need to copy it locally. */
21826 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED_const(sstr), SvCUR(sstr)+1);
21827 /* set malloced length to a non-zero value so it will be freed
21828 * (otherwise in combination with SVf_FAKE it looks like an alien
21829 * buffer). It doesn't have to be the actual malloced size, since it
21830 * should never be grown */
21831 SvLEN_set(dstr, SvCUR(sstr)+1);
21832 ret->mother_re = NULL;
21834 #endif /* PERL_IN_XSUB_RE */
21839 This is the internal complement to regdupe() which is used to copy
21840 the structure pointed to by the *pprivate pointer in the regexp.
21841 This is the core version of the extension overridable cloning hook.
21842 The regexp structure being duplicated will be copied by perl prior
21843 to this and will be provided as the regexp *r argument, however
21844 with the /old/ structures pprivate pointer value. Thus this routine
21845 may override any copying normally done by perl.
21847 It returns a pointer to the new regexp_internal structure.
21851 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
21854 struct regexp *const r = ReANY(rx);
21855 regexp_internal *reti;
21857 RXi_GET_DECL(r, ri);
21859 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
21863 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
21864 char, regexp_internal);
21865 Copy(ri->program, reti->program, len+1, regnode);
21868 if (ri->code_blocks) {
21870 Newx(reti->code_blocks, 1, struct reg_code_blocks);
21871 Newx(reti->code_blocks->cb, ri->code_blocks->count,
21872 struct reg_code_block);
21873 Copy(ri->code_blocks->cb, reti->code_blocks->cb,
21874 ri->code_blocks->count, struct reg_code_block);
21875 for (n = 0; n < ri->code_blocks->count; n++)
21876 reti->code_blocks->cb[n].src_regex = (REGEXP*)
21877 sv_dup_inc((SV*)(ri->code_blocks->cb[n].src_regex), param);
21878 reti->code_blocks->count = ri->code_blocks->count;
21879 reti->code_blocks->refcnt = 1;
21882 reti->code_blocks = NULL;
21884 reti->regstclass = NULL;
21887 struct reg_data *d;
21888 const int count = ri->data->count;
21891 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
21892 char, struct reg_data);
21893 Newx(d->what, count, U8);
21896 for (i = 0; i < count; i++) {
21897 d->what[i] = ri->data->what[i];
21898 switch (d->what[i]) {
21899 /* see also regcomp.h and regfree_internal() */
21900 case 'a': /* actually an AV, but the dup function is identical.
21901 values seem to be "plain sv's" generally. */
21902 case 'r': /* a compiled regex (but still just another SV) */
21903 case 's': /* an RV (currently only used for an RV to an AV by the ANYOF code)
21904 this use case should go away, the code could have used
21905 'a' instead - see S_set_ANYOF_arg() for array contents. */
21906 case 'S': /* actually an SV, but the dup function is identical. */
21907 case 'u': /* actually an HV, but the dup function is identical.
21908 values are "plain sv's" */
21909 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
21912 /* Synthetic Start Class - "Fake" charclass we generate to optimize
21913 * patterns which could start with several different things. Pre-TRIE
21914 * this was more important than it is now, however this still helps
21915 * in some places, for instance /x?a+/ might produce a SSC equivalent
21916 * to [xa]. This is used by Perl_re_intuit_start() and S_find_byclass()
21919 /* This is cheating. */
21920 Newx(d->data[i], 1, regnode_ssc);
21921 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
21922 reti->regstclass = (regnode*)d->data[i];
21925 /* AHO-CORASICK fail table */
21926 /* Trie stclasses are readonly and can thus be shared
21927 * without duplication. We free the stclass in pregfree
21928 * when the corresponding reg_ac_data struct is freed.
21930 reti->regstclass= ri->regstclass;
21933 /* TRIE transition table */
21935 ((reg_trie_data*)ri->data->data[i])->refcount++;
21938 case 'l': /* (?{...}) or (??{ ... }) code (cb->block) */
21939 case 'L': /* same when RExC_pm_flags & PMf_HAS_CV and code
21940 is not from another regexp */
21941 d->data[i] = ri->data->data[i];
21944 Perl_croak(aTHX_ "panic: re_dup_guts unknown data code '%c'",
21945 ri->data->what[i]);
21954 reti->name_list_idx = ri->name_list_idx;
21956 #ifdef RE_TRACK_PATTERN_OFFSETS
21957 if (ri->u.offsets) {
21958 Newx(reti->u.offsets, 2*len+1, U32);
21959 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
21962 SetProgLen(reti, len);
21965 return (void*)reti;
21968 #endif /* USE_ITHREADS */
21970 #ifndef PERL_IN_XSUB_RE
21973 - regnext - dig the "next" pointer out of a node
21976 Perl_regnext(pTHX_ regnode *p)
21983 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
21984 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
21985 (int)OP(p), (int)REGNODE_MAX);
21988 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
21998 S_re_croak(pTHX_ bool utf8, const char* pat,...)
22001 STRLEN len = strlen(pat);
22004 const char *message;
22006 PERL_ARGS_ASSERT_RE_CROAK;
22010 Copy(pat, buf, len , char);
22012 buf[len + 1] = '\0';
22013 va_start(args, pat);
22014 msv = vmess(buf, &args);
22016 message = SvPV_const(msv, len);
22019 Copy(message, buf, len , char);
22020 /* len-1 to avoid \n */
22021 Perl_croak(aTHX_ "%" UTF8f, UTF8fARG(utf8, len-1, buf));
22024 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
22026 #ifndef PERL_IN_XSUB_RE
22028 Perl_save_re_context(pTHX)
22033 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
22036 const REGEXP * const rx = PM_GETRE(PL_curpm);
22038 nparens = RX_NPARENS(rx);
22041 /* RT #124109. This is a complete hack; in the SWASHNEW case we know
22042 * that PL_curpm will be null, but that utf8.pm and the modules it
22043 * loads will only use $1..$3.
22044 * The t/porting/re_context.t test file checks this assumption.
22049 for (i = 1; i <= nparens; i++) {
22050 char digits[TYPE_CHARS(long)];
22051 const STRLEN len = my_snprintf(digits, sizeof(digits),
22053 GV *const *const gvp
22054 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
22057 GV * const gv = *gvp;
22058 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
22068 S_put_code_point(pTHX_ SV *sv, UV c)
22070 PERL_ARGS_ASSERT_PUT_CODE_POINT;
22073 Perl_sv_catpvf(aTHX_ sv, "\\x{%04" UVXf "}", c);
22075 else if (isPRINT(c)) {
22076 const char string = (char) c;
22078 /* We use {phrase} as metanotation in the class, so also escape literal
22080 if (isBACKSLASHED_PUNCT(c) || c == '{' || c == '}')
22081 sv_catpvs(sv, "\\");
22082 sv_catpvn(sv, &string, 1);
22084 else if (isMNEMONIC_CNTRL(c)) {
22085 Perl_sv_catpvf(aTHX_ sv, "%s", cntrl_to_mnemonic((U8) c));
22088 Perl_sv_catpvf(aTHX_ sv, "\\x%02X", (U8) c);
22092 #define MAX_PRINT_A MAX_PRINT_A_FOR_USE_ONLY_BY_REGCOMP_DOT_C
22095 S_put_range(pTHX_ SV *sv, UV start, const UV end, const bool allow_literals)
22097 /* Appends to 'sv' a displayable version of the range of code points from
22098 * 'start' to 'end'. Mnemonics (like '\r') are used for the few controls
22099 * that have them, when they occur at the beginning or end of the range.
22100 * It uses hex to output the remaining code points, unless 'allow_literals'
22101 * is true, in which case the printable ASCII ones are output as-is (though
22102 * some of these will be escaped by put_code_point()).
22104 * NOTE: This is designed only for printing ranges of code points that fit
22105 * inside an ANYOF bitmap. Higher code points are simply suppressed
22108 const unsigned int min_range_count = 3;
22110 assert(start <= end);
22112 PERL_ARGS_ASSERT_PUT_RANGE;
22114 while (start <= end) {
22116 const char * format;
22118 if (end - start < min_range_count) {
22120 /* Output chars individually when they occur in short ranges */
22121 for (; start <= end; start++) {
22122 put_code_point(sv, start);
22127 /* If permitted by the input options, and there is a possibility that
22128 * this range contains a printable literal, look to see if there is
22130 if (allow_literals && start <= MAX_PRINT_A) {
22132 /* If the character at the beginning of the range isn't an ASCII
22133 * printable, effectively split the range into two parts:
22134 * 1) the portion before the first such printable,
22136 * and output them separately. */
22137 if (! isPRINT_A(start)) {
22138 UV temp_end = start + 1;
22140 /* There is no point looking beyond the final possible
22141 * printable, in MAX_PRINT_A */
22142 UV max = MIN(end, MAX_PRINT_A);
22144 while (temp_end <= max && ! isPRINT_A(temp_end)) {
22148 /* Here, temp_end points to one beyond the first printable if
22149 * found, or to one beyond 'max' if not. If none found, make
22150 * sure that we use the entire range */
22151 if (temp_end > MAX_PRINT_A) {
22152 temp_end = end + 1;
22155 /* Output the first part of the split range: the part that
22156 * doesn't have printables, with the parameter set to not look
22157 * for literals (otherwise we would infinitely recurse) */
22158 put_range(sv, start, temp_end - 1, FALSE);
22160 /* The 2nd part of the range (if any) starts here. */
22163 /* We do a continue, instead of dropping down, because even if
22164 * the 2nd part is non-empty, it could be so short that we want
22165 * to output it as individual characters, as tested for at the
22166 * top of this loop. */
22170 /* Here, 'start' is a printable ASCII. If it is an alphanumeric,
22171 * output a sub-range of just the digits or letters, then process
22172 * the remaining portion as usual. */
22173 if (isALPHANUMERIC_A(start)) {
22174 UV mask = (isDIGIT_A(start))
22179 UV temp_end = start + 1;
22181 /* Find the end of the sub-range that includes just the
22182 * characters in the same class as the first character in it */
22183 while (temp_end <= end && _generic_isCC_A(temp_end, mask)) {
22188 /* For short ranges, don't duplicate the code above to output
22189 * them; just call recursively */
22190 if (temp_end - start < min_range_count) {
22191 put_range(sv, start, temp_end, FALSE);
22193 else { /* Output as a range */
22194 put_code_point(sv, start);
22195 sv_catpvs(sv, "-");
22196 put_code_point(sv, temp_end);
22198 start = temp_end + 1;
22202 /* We output any other printables as individual characters */
22203 if (isPUNCT_A(start) || isSPACE_A(start)) {
22204 while (start <= end && (isPUNCT_A(start)
22205 || isSPACE_A(start)))
22207 put_code_point(sv, start);
22212 } /* End of looking for literals */
22214 /* Here is not to output as a literal. Some control characters have
22215 * mnemonic names. Split off any of those at the beginning and end of
22216 * the range to print mnemonically. It isn't possible for many of
22217 * these to be in a row, so this won't overwhelm with output */
22219 && (isMNEMONIC_CNTRL(start) || isMNEMONIC_CNTRL(end)))
22221 while (isMNEMONIC_CNTRL(start) && start <= end) {
22222 put_code_point(sv, start);
22226 /* If this didn't take care of the whole range ... */
22227 if (start <= end) {
22229 /* Look backwards from the end to find the final non-mnemonic
22232 while (isMNEMONIC_CNTRL(temp_end)) {
22236 /* And separately output the interior range that doesn't start
22237 * or end with mnemonics */
22238 put_range(sv, start, temp_end, FALSE);
22240 /* Then output the mnemonic trailing controls */
22241 start = temp_end + 1;
22242 while (start <= end) {
22243 put_code_point(sv, start);
22250 /* As a final resort, output the range or subrange as hex. */
22252 if (start >= NUM_ANYOF_CODE_POINTS) {
22255 else { /* Have to split range at the bitmap boundary */
22256 this_end = (end < NUM_ANYOF_CODE_POINTS)
22258 : NUM_ANYOF_CODE_POINTS - 1;
22260 #if NUM_ANYOF_CODE_POINTS > 256
22261 format = (this_end < 256)
22262 ? "\\x%02" UVXf "-\\x%02" UVXf
22263 : "\\x{%04" UVXf "}-\\x{%04" UVXf "}";
22265 format = "\\x%02" UVXf "-\\x%02" UVXf;
22267 GCC_DIAG_IGNORE_STMT(-Wformat-nonliteral);
22268 Perl_sv_catpvf(aTHX_ sv, format, start, this_end);
22269 GCC_DIAG_RESTORE_STMT;
22275 S_put_charclass_bitmap_innards_invlist(pTHX_ SV *sv, SV* invlist)
22277 /* Concatenate onto the PV in 'sv' a displayable form of the inversion list
22281 bool allow_literals = TRUE;
22283 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_INVLIST;
22285 /* Generally, it is more readable if printable characters are output as
22286 * literals, but if a range (nearly) spans all of them, it's best to output
22287 * it as a single range. This code will use a single range if all but 2
22288 * ASCII printables are in it */
22289 invlist_iterinit(invlist);
22290 while (invlist_iternext(invlist, &start, &end)) {
22292 /* If the range starts beyond the final printable, it doesn't have any
22294 if (start > MAX_PRINT_A) {
22298 /* In both ASCII and EBCDIC, a SPACE is the lowest printable. To span
22299 * all but two, the range must start and end no later than 2 from
22301 if (start < ' ' + 2 && end > MAX_PRINT_A - 2) {
22302 if (end > MAX_PRINT_A) {
22308 if (end - start >= MAX_PRINT_A - ' ' - 2) {
22309 allow_literals = FALSE;
22314 invlist_iterfinish(invlist);
22316 /* Here we have figured things out. Output each range */
22317 invlist_iterinit(invlist);
22318 while (invlist_iternext(invlist, &start, &end)) {
22319 if (start >= NUM_ANYOF_CODE_POINTS) {
22322 put_range(sv, start, end, allow_literals);
22324 invlist_iterfinish(invlist);
22330 S_put_charclass_bitmap_innards_common(pTHX_
22331 SV* invlist, /* The bitmap */
22332 SV* posixes, /* Under /l, things like [:word:], \S */
22333 SV* only_utf8, /* Under /d, matches iff the target is UTF-8 */
22334 SV* not_utf8, /* /d, matches iff the target isn't UTF-8 */
22335 SV* only_utf8_locale, /* Under /l, matches if the locale is UTF-8 */
22336 const bool invert /* Is the result to be inverted? */
22339 /* Create and return an SV containing a displayable version of the bitmap
22340 * and associated information determined by the input parameters. If the
22341 * output would have been only the inversion indicator '^', NULL is instead
22347 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_COMMON;
22350 output = newSVpvs("^");
22353 output = newSVpvs("");
22356 /* First, the code points in the bitmap that are unconditionally there */
22357 put_charclass_bitmap_innards_invlist(output, invlist);
22359 /* Traditionally, these have been placed after the main code points */
22361 sv_catsv(output, posixes);
22364 if (only_utf8 && _invlist_len(only_utf8)) {
22365 Perl_sv_catpvf(aTHX_ output, "%s{utf8}%s", PL_colors[1], PL_colors[0]);
22366 put_charclass_bitmap_innards_invlist(output, only_utf8);
22369 if (not_utf8 && _invlist_len(not_utf8)) {
22370 Perl_sv_catpvf(aTHX_ output, "%s{not utf8}%s", PL_colors[1], PL_colors[0]);
22371 put_charclass_bitmap_innards_invlist(output, not_utf8);
22374 if (only_utf8_locale && _invlist_len(only_utf8_locale)) {
22375 Perl_sv_catpvf(aTHX_ output, "%s{utf8 locale}%s", PL_colors[1], PL_colors[0]);
22376 put_charclass_bitmap_innards_invlist(output, only_utf8_locale);
22378 /* This is the only list in this routine that can legally contain code
22379 * points outside the bitmap range. The call just above to
22380 * 'put_charclass_bitmap_innards_invlist' will simply suppress them, so
22381 * output them here. There's about a half-dozen possible, and none in
22382 * contiguous ranges longer than 2 */
22383 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
22385 SV* above_bitmap = NULL;
22387 _invlist_subtract(only_utf8_locale, PL_InBitmap, &above_bitmap);
22389 invlist_iterinit(above_bitmap);
22390 while (invlist_iternext(above_bitmap, &start, &end)) {
22393 for (i = start; i <= end; i++) {
22394 put_code_point(output, i);
22397 invlist_iterfinish(above_bitmap);
22398 SvREFCNT_dec_NN(above_bitmap);
22402 if (invert && SvCUR(output) == 1) {
22410 S_put_charclass_bitmap_innards(pTHX_ SV *sv,
22412 SV *nonbitmap_invlist,
22413 SV *only_utf8_locale_invlist,
22414 const regnode * const node,
22416 const bool force_as_is_display)
22418 /* Appends to 'sv' a displayable version of the innards of the bracketed
22419 * character class defined by the other arguments:
22420 * 'bitmap' points to the bitmap, or NULL if to ignore that.
22421 * 'nonbitmap_invlist' is an inversion list of the code points that are in
22422 * the bitmap range, but for some reason aren't in the bitmap; NULL if
22423 * none. The reasons for this could be that they require some
22424 * condition such as the target string being or not being in UTF-8
22425 * (under /d), or because they came from a user-defined property that
22426 * was not resolved at the time of the regex compilation (under /u)
22427 * 'only_utf8_locale_invlist' is an inversion list of the code points that
22428 * are valid only if the runtime locale is a UTF-8 one; NULL if none
22429 * 'node' is the regex pattern ANYOF node. It is needed only when the
22430 * above two parameters are not null, and is passed so that this
22431 * routine can tease apart the various reasons for them.
22432 * 'flags' is the flags field of 'node'
22433 * 'force_as_is_display' is TRUE if this routine should definitely NOT try
22434 * to invert things to see if that leads to a cleaner display. If
22435 * FALSE, this routine is free to use its judgment about doing this.
22437 * It returns TRUE if there was actually something output. (It may be that
22438 * the bitmap, etc is empty.)
22440 * When called for outputting the bitmap of a non-ANYOF node, just pass the
22441 * bitmap, with the succeeding parameters set to NULL, and the final one to
22445 /* In general, it tries to display the 'cleanest' representation of the
22446 * innards, choosing whether to display them inverted or not, regardless of
22447 * whether the class itself is to be inverted. However, there are some
22448 * cases where it can't try inverting, as what actually matches isn't known
22449 * until runtime, and hence the inversion isn't either. */
22452 bool inverting_allowed = ! force_as_is_display;
22455 STRLEN orig_sv_cur = SvCUR(sv);
22457 SV* invlist; /* Inversion list we accumulate of code points that
22458 are unconditionally matched */
22459 SV* only_utf8 = NULL; /* Under /d, list of matches iff the target is
22461 SV* not_utf8 = NULL; /* /d, list of matches iff the target isn't UTF-8
22463 SV* posixes = NULL; /* Under /l, string of things like [:word:], \D */
22464 SV* only_utf8_locale = NULL; /* Under /l, list of matches if the locale
22467 SV* as_is_display; /* The output string when we take the inputs
22469 SV* inverted_display; /* The output string when we invert the inputs */
22471 bool invert = cBOOL(flags & ANYOF_INVERT); /* Is the input to be inverted
22473 /* We are biased in favor of displaying things without them being inverted,
22474 * as that is generally easier to understand */
22475 const int bias = 5;
22477 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS;
22479 /* Start off with whatever code points are passed in. (We clone, so we
22480 * don't change the caller's list) */
22481 if (nonbitmap_invlist) {
22482 assert(invlist_highest(nonbitmap_invlist) < NUM_ANYOF_CODE_POINTS);
22483 invlist = invlist_clone(nonbitmap_invlist, NULL);
22485 else { /* Worst case size is every other code point is matched */
22486 invlist = _new_invlist(NUM_ANYOF_CODE_POINTS / 2);
22490 if (OP(node) == ANYOFD) {
22492 /* This flag indicates that the code points below 0x100 in the
22493 * nonbitmap list are precisely the ones that match only when the
22494 * target is UTF-8 (they should all be non-ASCII). */
22495 if (flags & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)
22497 _invlist_intersection(invlist, PL_UpperLatin1, &only_utf8);
22498 _invlist_subtract(invlist, only_utf8, &invlist);
22501 /* And this flag for matching all non-ASCII 0xFF and below */
22502 if (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)
22504 not_utf8 = invlist_clone(PL_UpperLatin1, NULL);
22507 else if (OP(node) == ANYOFL || OP(node) == ANYOFPOSIXL) {
22509 /* If either of these flags are set, what matches isn't
22510 * determinable except during execution, so don't know enough here
22512 if (flags & (ANYOFL_FOLD|ANYOF_MATCHES_POSIXL)) {
22513 inverting_allowed = FALSE;
22516 /* What the posix classes match also varies at runtime, so these
22517 * will be output symbolically. */
22518 if (ANYOF_POSIXL_TEST_ANY_SET(node)) {
22521 posixes = newSVpvs("");
22522 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
22523 if (ANYOF_POSIXL_TEST(node, i)) {
22524 sv_catpv(posixes, anyofs[i]);
22531 /* Accumulate the bit map into the unconditional match list */
22533 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
22534 if (BITMAP_TEST(bitmap, i)) {
22537 i < NUM_ANYOF_CODE_POINTS && BITMAP_TEST(bitmap, i);
22540 invlist = _add_range_to_invlist(invlist, start, i-1);
22545 /* Make sure that the conditional match lists don't have anything in them
22546 * that match unconditionally; otherwise the output is quite confusing.
22547 * This could happen if the code that populates these misses some
22550 _invlist_subtract(only_utf8, invlist, &only_utf8);
22553 _invlist_subtract(not_utf8, invlist, ¬_utf8);
22556 if (only_utf8_locale_invlist) {
22558 /* Since this list is passed in, we have to make a copy before
22560 only_utf8_locale = invlist_clone(only_utf8_locale_invlist, NULL);
22562 _invlist_subtract(only_utf8_locale, invlist, &only_utf8_locale);
22564 /* And, it can get really weird for us to try outputting an inverted
22565 * form of this list when it has things above the bitmap, so don't even
22567 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
22568 inverting_allowed = FALSE;
22572 /* Calculate what the output would be if we take the input as-is */
22573 as_is_display = put_charclass_bitmap_innards_common(invlist,
22580 /* If have to take the output as-is, just do that */
22581 if (! inverting_allowed) {
22582 if (as_is_display) {
22583 sv_catsv(sv, as_is_display);
22584 SvREFCNT_dec_NN(as_is_display);
22587 else { /* But otherwise, create the output again on the inverted input, and
22588 use whichever version is shorter */
22590 int inverted_bias, as_is_bias;
22592 /* We will apply our bias to whichever of the the results doesn't have
22602 inverted_bias = bias;
22605 /* Now invert each of the lists that contribute to the output,
22606 * excluding from the result things outside the possible range */
22608 /* For the unconditional inversion list, we have to add in all the
22609 * conditional code points, so that when inverted, they will be gone
22611 _invlist_union(only_utf8, invlist, &invlist);
22612 _invlist_union(not_utf8, invlist, &invlist);
22613 _invlist_union(only_utf8_locale, invlist, &invlist);
22614 _invlist_invert(invlist);
22615 _invlist_intersection(invlist, PL_InBitmap, &invlist);
22618 _invlist_invert(only_utf8);
22619 _invlist_intersection(only_utf8, PL_UpperLatin1, &only_utf8);
22621 else if (not_utf8) {
22623 /* If a code point matches iff the target string is not in UTF-8,
22624 * then complementing the result has it not match iff not in UTF-8,
22625 * which is the same thing as matching iff it is UTF-8. */
22626 only_utf8 = not_utf8;
22630 if (only_utf8_locale) {
22631 _invlist_invert(only_utf8_locale);
22632 _invlist_intersection(only_utf8_locale,
22634 &only_utf8_locale);
22637 inverted_display = put_charclass_bitmap_innards_common(
22642 only_utf8_locale, invert);
22644 /* Use the shortest representation, taking into account our bias
22645 * against showing it inverted */
22646 if ( inverted_display
22647 && ( ! as_is_display
22648 || ( SvCUR(inverted_display) + inverted_bias
22649 < SvCUR(as_is_display) + as_is_bias)))
22651 sv_catsv(sv, inverted_display);
22653 else if (as_is_display) {
22654 sv_catsv(sv, as_is_display);
22657 SvREFCNT_dec(as_is_display);
22658 SvREFCNT_dec(inverted_display);
22661 SvREFCNT_dec_NN(invlist);
22662 SvREFCNT_dec(only_utf8);
22663 SvREFCNT_dec(not_utf8);
22664 SvREFCNT_dec(posixes);
22665 SvREFCNT_dec(only_utf8_locale);
22667 return SvCUR(sv) > orig_sv_cur;
22670 #define CLEAR_OPTSTART \
22671 if (optstart) STMT_START { \
22672 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ \
22673 " (%" IVdf " nodes)\n", (IV)(node - optstart))); \
22677 #define DUMPUNTIL(b,e) \
22679 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
22681 STATIC const regnode *
22682 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
22683 const regnode *last, const regnode *plast,
22684 SV* sv, I32 indent, U32 depth)
22686 U8 op = PSEUDO; /* Arbitrary non-END op. */
22687 const regnode *next;
22688 const regnode *optstart= NULL;
22690 RXi_GET_DECL(r, ri);
22691 DECLARE_AND_GET_RE_DEBUG_FLAGS;
22693 PERL_ARGS_ASSERT_DUMPUNTIL;
22695 #ifdef DEBUG_DUMPUNTIL
22696 Perl_re_printf( aTHX_ "--- %d : %d - %d - %d\n", indent, node-start,
22697 last ? last-start : 0, plast ? plast-start : 0);
22700 if (plast && plast < last)
22703 while (PL_regkind[op] != END && (!last || node < last)) {
22705 /* While that wasn't END last time... */
22708 if (op == CLOSE || op == SRCLOSE || op == WHILEM)
22710 next = regnext((regnode *)node);
22713 if (OP(node) == OPTIMIZED) {
22714 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
22721 regprop(r, sv, node, NULL, NULL);
22722 Perl_re_printf( aTHX_ "%4" IVdf ":%*s%s", (IV)(node - start),
22723 (int)(2*indent + 1), "", SvPVX_const(sv));
22725 if (OP(node) != OPTIMIZED) {
22726 if (next == NULL) /* Next ptr. */
22727 Perl_re_printf( aTHX_ " (0)");
22728 else if (PL_regkind[(U8)op] == BRANCH
22729 && PL_regkind[OP(next)] != BRANCH )
22730 Perl_re_printf( aTHX_ " (FAIL)");
22732 Perl_re_printf( aTHX_ " (%" IVdf ")", (IV)(next - start));
22733 Perl_re_printf( aTHX_ "\n");
22737 if (PL_regkind[(U8)op] == BRANCHJ) {
22740 const regnode *nnode = (OP(next) == LONGJMP
22741 ? regnext((regnode *)next)
22743 if (last && nnode > last)
22745 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
22748 else if (PL_regkind[(U8)op] == BRANCH) {
22750 DUMPUNTIL(NEXTOPER(node), next);
22752 else if ( PL_regkind[(U8)op] == TRIE ) {
22753 const regnode *this_trie = node;
22754 const char op = OP(node);
22755 const U32 n = ARG(node);
22756 const reg_ac_data * const ac = op>=AHOCORASICK ?
22757 (reg_ac_data *)ri->data->data[n] :
22759 const reg_trie_data * const trie =
22760 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
22762 AV *const trie_words
22763 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
22765 const regnode *nextbranch= NULL;
22768 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
22769 SV ** const elem_ptr = av_fetch(trie_words, word_idx, 0);
22771 Perl_re_indentf( aTHX_ "%s ",
22774 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
22775 SvCUR(*elem_ptr), PL_dump_re_max_len,
22776 PL_colors[0], PL_colors[1],
22778 ? PERL_PV_ESCAPE_UNI
22780 | PERL_PV_PRETTY_ELLIPSES
22781 | PERL_PV_PRETTY_LTGT
22786 U16 dist= trie->jump[word_idx+1];
22787 Perl_re_printf( aTHX_ "(%" UVuf ")\n",
22788 (UV)((dist ? this_trie + dist : next) - start));
22791 nextbranch= this_trie + trie->jump[0];
22792 DUMPUNTIL(this_trie + dist, nextbranch);
22794 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
22795 nextbranch= regnext((regnode *)nextbranch);
22797 Perl_re_printf( aTHX_ "\n");
22800 if (last && next > last)
22805 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
22806 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
22807 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
22809 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
22811 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
22813 else if ( op == PLUS || op == STAR) {
22814 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
22816 else if (PL_regkind[(U8)op] == EXACT || op == ANYOFHs) {
22817 /* Literal string, where present. */
22818 node += NODE_SZ_STR(node) - 1;
22819 node = NEXTOPER(node);
22822 node = NEXTOPER(node);
22823 node += regarglen[(U8)op];
22825 if (op == CURLYX || op == OPEN || op == SROPEN)
22829 #ifdef DEBUG_DUMPUNTIL
22830 Perl_re_printf( aTHX_ "--- %d\n", (int)indent);
22835 #endif /* DEBUGGING */
22837 #ifndef PERL_IN_XSUB_RE
22839 # include "uni_keywords.h"
22842 Perl_init_uniprops(pTHX)
22847 char * dump_len_string;
22849 dump_len_string = PerlEnv_getenv("PERL_DUMP_RE_MAX_LEN");
22850 if ( ! dump_len_string
22851 || ! grok_atoUV(dump_len_string, (UV *)&PL_dump_re_max_len, NULL))
22853 PL_dump_re_max_len = 60; /* A reasonable default */
22857 PL_user_def_props = newHV();
22859 # ifdef USE_ITHREADS
22861 HvSHAREKEYS_off(PL_user_def_props);
22862 PL_user_def_props_aTHX = aTHX;
22866 /* Set up the inversion list interpreter-level variables */
22868 PL_XPosix_ptrs[_CC_ASCII] = _new_invlist_C_array(uni_prop_ptrs[UNI_ASCII]);
22869 PL_XPosix_ptrs[_CC_ALPHANUMERIC] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXALNUM]);
22870 PL_XPosix_ptrs[_CC_ALPHA] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXALPHA]);
22871 PL_XPosix_ptrs[_CC_BLANK] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXBLANK]);
22872 PL_XPosix_ptrs[_CC_CASED] = _new_invlist_C_array(uni_prop_ptrs[UNI_CASED]);
22873 PL_XPosix_ptrs[_CC_CNTRL] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXCNTRL]);
22874 PL_XPosix_ptrs[_CC_DIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXDIGIT]);
22875 PL_XPosix_ptrs[_CC_GRAPH] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXGRAPH]);
22876 PL_XPosix_ptrs[_CC_LOWER] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXLOWER]);
22877 PL_XPosix_ptrs[_CC_PRINT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXPRINT]);
22878 PL_XPosix_ptrs[_CC_PUNCT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXPUNCT]);
22879 PL_XPosix_ptrs[_CC_SPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXSPACE]);
22880 PL_XPosix_ptrs[_CC_UPPER] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXUPPER]);
22881 PL_XPosix_ptrs[_CC_VERTSPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_VERTSPACE]);
22882 PL_XPosix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXWORD]);
22883 PL_XPosix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXXDIGIT]);
22885 PL_Posix_ptrs[_CC_ASCII] = _new_invlist_C_array(uni_prop_ptrs[UNI_ASCII]);
22886 PL_Posix_ptrs[_CC_ALPHANUMERIC] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXALNUM]);
22887 PL_Posix_ptrs[_CC_ALPHA] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXALPHA]);
22888 PL_Posix_ptrs[_CC_BLANK] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXBLANK]);
22889 PL_Posix_ptrs[_CC_CASED] = PL_Posix_ptrs[_CC_ALPHA];
22890 PL_Posix_ptrs[_CC_CNTRL] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXCNTRL]);
22891 PL_Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXDIGIT]);
22892 PL_Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXGRAPH]);
22893 PL_Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXLOWER]);
22894 PL_Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXPRINT]);
22895 PL_Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXPUNCT]);
22896 PL_Posix_ptrs[_CC_SPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXSPACE]);
22897 PL_Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXUPPER]);
22898 PL_Posix_ptrs[_CC_VERTSPACE] = NULL;
22899 PL_Posix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXWORD]);
22900 PL_Posix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXXDIGIT]);
22902 PL_GCB_invlist = _new_invlist_C_array(_Perl_GCB_invlist);
22903 PL_SB_invlist = _new_invlist_C_array(_Perl_SB_invlist);
22904 PL_WB_invlist = _new_invlist_C_array(_Perl_WB_invlist);
22905 PL_LB_invlist = _new_invlist_C_array(_Perl_LB_invlist);
22906 PL_SCX_invlist = _new_invlist_C_array(_Perl_SCX_invlist);
22908 PL_InBitmap = _new_invlist_C_array(InBitmap_invlist);
22909 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
22910 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
22911 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
22913 PL_Assigned_invlist = _new_invlist_C_array(uni_prop_ptrs[UNI_ASSIGNED]);
22915 PL_utf8_perl_idstart = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_IDSTART]);
22916 PL_utf8_perl_idcont = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_IDCONT]);
22918 PL_utf8_charname_begin = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_CHARNAME_BEGIN]);
22919 PL_utf8_charname_continue = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_CHARNAME_CONTINUE]);
22921 PL_in_some_fold = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_ANY_FOLDS]);
22922 PL_HasMultiCharFold = _new_invlist_C_array(uni_prop_ptrs[
22923 UNI__PERL_FOLDS_TO_MULTI_CHAR]);
22924 PL_InMultiCharFold = _new_invlist_C_array(uni_prop_ptrs[
22925 UNI__PERL_IS_IN_MULTI_CHAR_FOLD]);
22926 PL_utf8_toupper = _new_invlist_C_array(Uppercase_Mapping_invlist);
22927 PL_utf8_tolower = _new_invlist_C_array(Lowercase_Mapping_invlist);
22928 PL_utf8_totitle = _new_invlist_C_array(Titlecase_Mapping_invlist);
22929 PL_utf8_tofold = _new_invlist_C_array(Case_Folding_invlist);
22930 PL_utf8_tosimplefold = _new_invlist_C_array(Simple_Case_Folding_invlist);
22931 PL_utf8_foldclosures = _new_invlist_C_array(_Perl_IVCF_invlist);
22932 PL_utf8_mark = _new_invlist_C_array(uni_prop_ptrs[UNI_M]);
22933 PL_CCC_non0_non230 = _new_invlist_C_array(_Perl_CCC_non0_non230_invlist);
22934 PL_Private_Use = _new_invlist_C_array(uni_prop_ptrs[UNI_CO]);
22937 /* The below are used only by deprecated functions. They could be removed */
22938 PL_utf8_xidcont = _new_invlist_C_array(uni_prop_ptrs[UNI_XIDC]);
22939 PL_utf8_idcont = _new_invlist_C_array(uni_prop_ptrs[UNI_IDC]);
22940 PL_utf8_xidstart = _new_invlist_C_array(uni_prop_ptrs[UNI_XIDS]);
22944 /* These four functions are compiled only in regcomp.c, where they have access
22945 * to the data they return. They are a way for re_comp.c to get access to that
22946 * data without having to compile the whole data structures. */
22949 Perl_do_uniprop_match(const char * const key, const U16 key_len)
22951 PERL_ARGS_ASSERT_DO_UNIPROP_MATCH;
22953 return match_uniprop((U8 *) key, key_len);
22957 Perl_get_prop_definition(pTHX_ const int table_index)
22959 PERL_ARGS_ASSERT_GET_PROP_DEFINITION;
22961 /* Create and return the inversion list */
22962 return _new_invlist_C_array(uni_prop_ptrs[table_index]);
22965 const char * const *
22966 Perl_get_prop_values(const int table_index)
22968 PERL_ARGS_ASSERT_GET_PROP_VALUES;
22970 return UNI_prop_value_ptrs[table_index];
22974 Perl_get_deprecated_property_msg(const Size_t warning_offset)
22976 PERL_ARGS_ASSERT_GET_DEPRECATED_PROPERTY_MSG;
22978 return deprecated_property_msgs[warning_offset];
22983 This code was mainly added for backcompat to give a warning for non-portable
22984 code points in user-defined properties. But experiments showed that the
22985 warning in earlier perls were only omitted on overflow, which should be an
22986 error, so there really isnt a backcompat issue, and actually adding the
22987 warning when none was present before might cause breakage, for little gain. So
22988 khw left this code in, but not enabled. Tests were never added.
22991 Ei |const char *|get_extended_utf8_msg|const UV cp
22993 PERL_STATIC_INLINE const char *
22994 S_get_extended_utf8_msg(pTHX_ const UV cp)
22996 U8 dummy[UTF8_MAXBYTES + 1];
23000 uvchr_to_utf8_flags_msgs(dummy, cp, UNICODE_WARN_PERL_EXTENDED,
23003 msg = hv_fetchs(msgs, "text", 0);
23006 (void) sv_2mortal((SV *) msgs);
23008 return SvPVX(*msg);
23012 #endif /* end of ! PERL_IN_XSUB_RE */
23015 S_compile_wildcard(pTHX_ const char * subpattern, const STRLEN len,
23016 const bool ignore_case)
23018 /* Pretends that the input subpattern is qr/subpattern/aam, compiling it
23019 * possibly with /i if the 'ignore_case' parameter is true. Use /aa
23020 * because nothing outside of ASCII will match. Use /m because the input
23021 * string may be a bunch of lines strung together.
23023 * Also sets up the debugging info */
23025 U32 flags = PMf_MULTILINE|PMf_WILDCARD;
23027 SV * subpattern_sv = sv_2mortal(newSVpvn(subpattern, len));
23028 REGEXP * subpattern_re;
23029 DECLARE_AND_GET_RE_DEBUG_FLAGS;
23031 PERL_ARGS_ASSERT_COMPILE_WILDCARD;
23036 set_regex_charset(&flags, REGEX_ASCII_MORE_RESTRICTED_CHARSET);
23038 /* Like in op.c, we copy the compile time pm flags to the rx ones */
23039 rx_flags = flags & RXf_PMf_COMPILETIME;
23041 #ifndef PERL_IN_XSUB_RE
23042 /* Use the core engine if this file is regcomp.c. That means no
23043 * 'use re "Debug ..." is in effect, so the core engine is sufficient */
23044 subpattern_re = Perl_re_op_compile(aTHX_ &subpattern_sv, 1, NULL,
23045 &PL_core_reg_engine,
23049 if (isDEBUG_WILDCARD) {
23050 /* Use the special debugging engine if this file is re_comp.c and wants
23051 * to output the wildcard matching. This uses whatever
23052 * 'use re "Debug ..." is in effect */
23053 subpattern_re = Perl_re_op_compile(aTHX_ &subpattern_sv, 1, NULL,
23059 /* Use the special wildcard engine if this file is re_comp.c and
23060 * doesn't want to output the wildcard matching. This uses whatever
23061 * 'use re "Debug ..." is in effect for compilation, but this engine
23062 * structure has been set up so that it uses the core engine for
23063 * execution, so no execution debugging as a result of re.pm will be
23065 subpattern_re = Perl_re_op_compile(aTHX_ &subpattern_sv, 1, NULL,
23069 /* XXX The above has the effect that any user-supplied regex engine
23070 * won't be called for matching wildcards. That might be good, or bad.
23071 * It could be changed in several ways. The reason it is done the
23072 * current way is to avoid having to save and restore
23073 * ^{^RE_DEBUG_FLAGS} around the execution. save_scalar() perhaps
23074 * could be used. Another suggestion is to keep the authoritative
23075 * value of the debug flags in a thread-local variable and add set/get
23076 * magic to ${^RE_DEBUG_FLAGS} to keep the C level variable up to date.
23077 * Still another is to pass a flag, say in the engine's intflags that
23078 * would be checked each time before doing the debug output */
23082 assert(subpattern_re); /* Should have died if didn't compile successfully */
23083 return subpattern_re;
23087 S_execute_wildcard(pTHX_ REGEXP * const prog, char* stringarg, char *strend,
23088 char *strbeg, SSize_t minend, SV *screamer, U32 nosave)
23091 DECLARE_AND_GET_RE_DEBUG_FLAGS;
23093 PERL_ARGS_ASSERT_EXECUTE_WILDCARD;
23097 /* The compilation has set things up so that if the program doesn't want to
23098 * see the wildcard matching procedure, it will get the core execution
23099 * engine, which is subject only to -Dr. So we have to turn that off
23100 * around this procedure */
23101 if (! isDEBUG_WILDCARD) {
23102 /* Note! Casts away 'volatile' */
23104 PL_debug &= ~ DEBUG_r_FLAG;
23107 result = CALLREGEXEC(prog, stringarg, strend, strbeg, minend, screamer,
23115 S_handle_user_defined_property(pTHX_
23117 /* Parses the contents of a user-defined property definition; returning the
23118 * expanded definition if possible. If so, the return is an inversion
23121 * If there are subroutines that are part of the expansion and which aren't
23122 * known at the time of the call to this function, this returns what
23123 * parse_uniprop_string() returned for the first one encountered.
23125 * If an error was found, NULL is returned, and 'msg' gets a suitable
23126 * message appended to it. (Appending allows the back trace of how we got
23127 * to the faulty definition to be displayed through nested calls of
23128 * user-defined subs.)
23130 * The caller IS responsible for freeing any returned SV.
23132 * The syntax of the contents is pretty much described in perlunicode.pod,
23133 * but we also allow comments on each line */
23135 const char * name, /* Name of property */
23136 const STRLEN name_len, /* The name's length in bytes */
23137 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
23138 const bool to_fold, /* ? Is this under /i */
23139 const bool runtime, /* ? Are we in compile- or run-time */
23140 const bool deferrable, /* Is it ok for this property's full definition
23141 to be deferred until later? */
23142 SV* contents, /* The property's definition */
23143 bool *user_defined_ptr, /* This will be set TRUE as we wouldn't be
23144 getting called unless this is thought to be
23145 a user-defined property */
23146 SV * msg, /* Any error or warning msg(s) are appended to
23148 const STRLEN level) /* Recursion level of this call */
23151 const char * string = SvPV_const(contents, len);
23152 const char * const e = string + len;
23153 const bool is_contents_utf8 = cBOOL(SvUTF8(contents));
23154 const STRLEN msgs_length_on_entry = SvCUR(msg);
23156 const char * s0 = string; /* Points to first byte in the current line
23157 being parsed in 'string' */
23158 const char overflow_msg[] = "Code point too large in \"";
23159 SV* running_definition = NULL;
23161 PERL_ARGS_ASSERT_HANDLE_USER_DEFINED_PROPERTY;
23163 *user_defined_ptr = TRUE;
23165 /* Look at each line */
23167 const char * s; /* Current byte */
23168 char op = '+'; /* Default operation is 'union' */
23169 IV min = 0; /* range begin code point */
23170 IV max = -1; /* and range end */
23171 SV* this_definition;
23173 /* Skip comment lines */
23175 s0 = strchr(s0, '\n');
23183 /* For backcompat, allow an empty first line */
23189 /* First character in the line may optionally be the operation */
23198 /* If the line is one or two hex digits separated by blank space, its
23199 * a range; otherwise it is either another user-defined property or an
23204 if (! isXDIGIT(*s)) {
23205 goto check_if_property;
23208 do { /* Each new hex digit will add 4 bits. */
23209 if (min > ( (IV) MAX_LEGAL_CP >> 4)) {
23210 s = strchr(s, '\n');
23214 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23215 sv_catpv(msg, overflow_msg);
23216 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
23217 UTF8fARG(is_contents_utf8, s - s0, s0));
23218 sv_catpvs(msg, "\"");
23219 goto return_failure;
23222 /* Accumulate this digit into the value */
23223 min = (min << 4) + READ_XDIGIT(s);
23224 } while (isXDIGIT(*s));
23226 while (isBLANK(*s)) { s++; }
23228 /* We allow comments at the end of the line */
23230 s = strchr(s, '\n');
23236 else if (s < e && *s != '\n') {
23237 if (! isXDIGIT(*s)) {
23238 goto check_if_property;
23241 /* Look for the high point of the range */
23244 if (max > ( (IV) MAX_LEGAL_CP >> 4)) {
23245 s = strchr(s, '\n');
23249 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23250 sv_catpv(msg, overflow_msg);
23251 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
23252 UTF8fARG(is_contents_utf8, s - s0, s0));
23253 sv_catpvs(msg, "\"");
23254 goto return_failure;
23257 max = (max << 4) + READ_XDIGIT(s);
23258 } while (isXDIGIT(*s));
23260 while (isBLANK(*s)) { s++; }
23263 s = strchr(s, '\n');
23268 else if (s < e && *s != '\n') {
23269 goto check_if_property;
23273 if (max == -1) { /* The line only had one entry */
23276 else if (max < min) {
23277 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23278 sv_catpvs(msg, "Illegal range in \"");
23279 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
23280 UTF8fARG(is_contents_utf8, s - s0, s0));
23281 sv_catpvs(msg, "\"");
23282 goto return_failure;
23285 # if 0 /* See explanation at definition above of get_extended_utf8_msg() */
23287 if ( UNICODE_IS_PERL_EXTENDED(min)
23288 || UNICODE_IS_PERL_EXTENDED(max))
23290 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23292 /* If both code points are non-portable, warn only on the lower
23294 sv_catpv(msg, get_extended_utf8_msg(
23295 (UNICODE_IS_PERL_EXTENDED(min))
23297 sv_catpvs(msg, " in \"");
23298 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
23299 UTF8fARG(is_contents_utf8, s - s0, s0));
23300 sv_catpvs(msg, "\"");
23305 /* Here, this line contains a legal range */
23306 this_definition = sv_2mortal(_new_invlist(2));
23307 this_definition = _add_range_to_invlist(this_definition, min, max);
23312 /* Here it isn't a legal range line. See if it is a legal property
23313 * line. First find the end of the meat of the line */
23314 s = strpbrk(s, "#\n");
23319 /* Ignore trailing blanks in keeping with the requirements of
23320 * parse_uniprop_string() */
23322 while (s > s0 && isBLANK_A(*s)) {
23327 this_definition = parse_uniprop_string(s0, s - s0,
23328 is_utf8, to_fold, runtime,
23330 user_defined_ptr, msg,
23332 ? level /* Don't increase level
23333 if input is empty */
23336 if (this_definition == NULL) {
23337 goto return_failure; /* 'msg' should have had the reason
23338 appended to it by the above call */
23341 if (! is_invlist(this_definition)) { /* Unknown at this time */
23342 return newSVsv(this_definition);
23346 s = strchr(s, '\n');
23356 _invlist_union(running_definition, this_definition,
23357 &running_definition);
23360 _invlist_subtract(running_definition, this_definition,
23361 &running_definition);
23364 _invlist_intersection(running_definition, this_definition,
23365 &running_definition);
23368 _invlist_union_complement_2nd(running_definition,
23369 this_definition, &running_definition);
23372 Perl_croak(aTHX_ "panic: %s: %d: Unexpected operation %d",
23373 __FILE__, __LINE__, op);
23377 /* Position past the '\n' */
23379 } /* End of loop through the lines of 'contents' */
23381 /* Here, we processed all the lines in 'contents' without error. If we
23382 * didn't add any warnings, simply return success */
23383 if (msgs_length_on_entry == SvCUR(msg)) {
23385 /* If the expansion was empty, the answer isn't nothing: its an empty
23386 * inversion list */
23387 if (running_definition == NULL) {
23388 running_definition = _new_invlist(1);
23391 return running_definition;
23394 /* Otherwise, add some explanatory text, but we will return success */
23398 running_definition = NULL;
23402 if (name_len > 0) {
23403 sv_catpvs(msg, " in expansion of ");
23404 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8, name_len, name));
23407 return running_definition;
23410 /* As explained below, certain operations need to take place in the first
23411 * thread created. These macros switch contexts */
23412 # ifdef USE_ITHREADS
23413 # define DECLARATION_FOR_GLOBAL_CONTEXT \
23414 PerlInterpreter * save_aTHX = aTHX;
23415 # define SWITCH_TO_GLOBAL_CONTEXT \
23416 PERL_SET_CONTEXT((aTHX = PL_user_def_props_aTHX))
23417 # define RESTORE_CONTEXT PERL_SET_CONTEXT((aTHX = save_aTHX));
23418 # define CUR_CONTEXT aTHX
23419 # define ORIGINAL_CONTEXT save_aTHX
23421 # define DECLARATION_FOR_GLOBAL_CONTEXT
23422 # define SWITCH_TO_GLOBAL_CONTEXT NOOP
23423 # define RESTORE_CONTEXT NOOP
23424 # define CUR_CONTEXT NULL
23425 # define ORIGINAL_CONTEXT NULL
23429 S_delete_recursion_entry(pTHX_ void *key)
23431 /* Deletes the entry used to detect recursion when expanding user-defined
23432 * properties. This is a function so it can be set up to be called even if
23433 * the program unexpectedly quits */
23436 SV ** current_entry;
23437 const STRLEN key_len = strlen((const char *) key);
23438 DECLARATION_FOR_GLOBAL_CONTEXT;
23440 SWITCH_TO_GLOBAL_CONTEXT;
23442 /* If the entry is one of these types, it is a permanent entry, and not the
23443 * one used to detect recursions. This function should delete only the
23444 * recursion entry */
23445 current_entry = hv_fetch(PL_user_def_props, (const char *) key, key_len, 0);
23447 && ! is_invlist(*current_entry)
23448 && ! SvPOK(*current_entry))
23450 (void) hv_delete(PL_user_def_props, (const char *) key, key_len,
23458 S_get_fq_name(pTHX_
23459 const char * const name, /* The first non-blank in the \p{}, \P{} */
23460 const Size_t name_len, /* Its length in bytes, not including any trailing space */
23461 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
23462 const bool has_colon_colon
23465 /* Returns a mortal SV containing the fully qualified version of the input
23470 fq_name = newSVpvs_flags("", SVs_TEMP);
23472 /* Use the current package if it wasn't included in our input */
23473 if (! has_colon_colon) {
23474 const HV * pkg = (IN_PERL_COMPILETIME)
23476 : CopSTASH(PL_curcop);
23477 const char* pkgname = HvNAME(pkg);
23479 Perl_sv_catpvf(aTHX_ fq_name, "%" UTF8f,
23480 UTF8fARG(is_utf8, strlen(pkgname), pkgname));
23481 sv_catpvs(fq_name, "::");
23484 Perl_sv_catpvf(aTHX_ fq_name, "%" UTF8f,
23485 UTF8fARG(is_utf8, name_len, name));
23490 S_parse_uniprop_string(pTHX_
23492 /* Parse the interior of a \p{}, \P{}. Returns its definition if knowable
23493 * now. If so, the return is an inversion list.
23495 * If the property is user-defined, it is a subroutine, which in turn
23496 * may call other subroutines. This function will call the whole nest of
23497 * them to get the definition they return; if some aren't known at the time
23498 * of the call to this function, the fully qualified name of the highest
23499 * level sub is returned. It is an error to call this function at runtime
23500 * without every sub defined.
23502 * If an error was found, NULL is returned, and 'msg' gets a suitable
23503 * message appended to it. (Appending allows the back trace of how we got
23504 * to the faulty definition to be displayed through nested calls of
23505 * user-defined subs.)
23507 * The caller should NOT try to free any returned inversion list.
23509 * Other parameters will be set on return as described below */
23511 const char * const name, /* The first non-blank in the \p{}, \P{} */
23512 Size_t name_len, /* Its length in bytes, not including any
23514 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
23515 const bool to_fold, /* ? Is this under /i */
23516 const bool runtime, /* TRUE if this is being called at run time */
23517 const bool deferrable, /* TRUE if it's ok for the definition to not be
23518 known at this call */
23519 bool *user_defined_ptr, /* Upon return from this function it will be
23520 set to TRUE if any component is a
23521 user-defined property */
23522 SV * msg, /* Any error or warning msg(s) are appended to
23524 const STRLEN level) /* Recursion level of this call */
23527 char* lookup_name; /* normalized name for lookup in our tables */
23528 unsigned lookup_len; /* Its length */
23529 enum { Not_Strict = 0, /* Some properties have stricter name */
23530 Strict, /* normalization rules, which we decide */
23531 As_Is /* upon based on parsing */
23532 } stricter = Not_Strict;
23534 /* nv= or numeric_value=, or possibly one of the cjk numeric properties
23535 * (though it requires extra effort to download them from Unicode and
23536 * compile perl to know about them) */
23537 bool is_nv_type = FALSE;
23539 unsigned int i, j = 0;
23540 int equals_pos = -1; /* Where the '=' is found, or negative if none */
23541 int slash_pos = -1; /* Where the '/' is found, or negative if none */
23542 int table_index = 0; /* The entry number for this property in the table
23543 of all Unicode property names */
23544 bool starts_with_Is = FALSE; /* ? Does the name start with 'Is' */
23545 Size_t lookup_offset = 0; /* Used to ignore the first few characters of
23546 the normalized name in certain situations */
23547 Size_t non_pkg_begin = 0; /* Offset of first byte in 'name' that isn't
23548 part of a package name */
23549 Size_t lun_non_pkg_begin = 0; /* Similarly for 'lookup_name' */
23550 bool could_be_user_defined = TRUE; /* ? Could this be a user-defined
23551 property rather than a Unicode
23553 SV * prop_definition = NULL; /* The returned definition of 'name' or NULL
23554 if an error. If it is an inversion list,
23555 it is the definition. Otherwise it is a
23556 string containing the fully qualified sub
23558 SV * fq_name = NULL; /* For user-defined properties, the fully
23560 bool invert_return = FALSE; /* ? Do we need to complement the result before
23562 bool stripped_utf8_pkg = FALSE; /* Set TRUE if the input includes an
23563 explicit utf8:: package that we strip
23565 /* The expansion of properties that could be either user-defined or
23566 * official unicode ones is deferred until runtime, including a marker for
23567 * those that might be in the latter category. This boolean indicates if
23568 * we've seen that marker. If not, what we're parsing can't be such an
23569 * official Unicode property whose expansion was deferred */
23570 bool could_be_deferred_official = FALSE;
23572 PERL_ARGS_ASSERT_PARSE_UNIPROP_STRING;
23574 /* The input will be normalized into 'lookup_name' */
23575 Newx(lookup_name, name_len, char);
23576 SAVEFREEPV(lookup_name);
23578 /* Parse the input. */
23579 for (i = 0; i < name_len; i++) {
23580 char cur = name[i];
23582 /* Most of the characters in the input will be of this ilk, being parts
23584 if (isIDCONT_A(cur)) {
23586 /* Case differences are ignored. Our lookup routine assumes
23587 * everything is lowercase, so normalize to that */
23588 if (isUPPER_A(cur)) {
23589 lookup_name[j++] = toLOWER_A(cur);
23593 if (cur == '_') { /* Don't include these in the normalized name */
23597 lookup_name[j++] = cur;
23599 /* The first character in a user-defined name must be of this type.
23601 if (i - non_pkg_begin == 0 && ! isIDFIRST_A(cur)) {
23602 could_be_user_defined = FALSE;
23608 /* Here, the character is not something typically in a name, But these
23609 * two types of characters (and the '_' above) can be freely ignored in
23610 * most situations. Later it may turn out we shouldn't have ignored
23611 * them, and we have to reparse, but we don't have enough information
23612 * yet to make that decision */
23613 if (cur == '-' || isSPACE_A(cur)) {
23614 could_be_user_defined = FALSE;
23618 /* An equals sign or single colon mark the end of the first part of
23619 * the property name */
23621 || (cur == ':' && (i >= name_len - 1 || name[i+1] != ':')))
23623 lookup_name[j++] = '='; /* Treat the colon as an '=' */
23624 equals_pos = j; /* Note where it occurred in the input */
23625 could_be_user_defined = FALSE;
23629 /* If this looks like it is a marker we inserted at compile time,
23630 * set a flag and otherwise ignore it. If it isn't in the final
23631 * position, keep it as it would have been user input. */
23632 if ( UNLIKELY(cur == DEFERRED_COULD_BE_OFFICIAL_MARKERc)
23634 && could_be_user_defined
23635 && i == name_len - 1)
23638 could_be_deferred_official = TRUE;
23642 /* Otherwise, this character is part of the name. */
23643 lookup_name[j++] = cur;
23645 /* Here it isn't a single colon, so if it is a colon, it must be a
23649 /* A double colon should be a package qualifier. We note its
23650 * position and continue. Note that one could have
23651 * pkg1::pkg2::...::foo
23652 * so that the position at the end of the loop will be just after
23653 * the final qualifier */
23656 non_pkg_begin = i + 1;
23657 lookup_name[j++] = ':';
23658 lun_non_pkg_begin = j;
23660 else { /* Only word chars (and '::') can be in a user-defined name */
23661 could_be_user_defined = FALSE;
23663 } /* End of parsing through the lhs of the property name (or all of it if
23666 # define STRLENs(s) (sizeof("" s "") - 1)
23668 /* If there is a single package name 'utf8::', it is ambiguous. It could
23669 * be for a user-defined property, or it could be a Unicode property, as
23670 * all of them are considered to be for that package. For the purposes of
23671 * parsing the rest of the property, strip it off */
23672 if (non_pkg_begin == STRLENs("utf8::") && memBEGINPs(name, name_len, "utf8::")) {
23673 lookup_name += STRLENs("utf8::");
23674 j -= STRLENs("utf8::");
23675 equals_pos -= STRLENs("utf8::");
23676 stripped_utf8_pkg = TRUE;
23679 /* Here, we are either done with the whole property name, if it was simple;
23680 * or are positioned just after the '=' if it is compound. */
23682 if (equals_pos >= 0) {
23683 assert(stricter == Not_Strict); /* We shouldn't have set this yet */
23685 /* Space immediately after the '=' is ignored */
23687 for (; i < name_len; i++) {
23688 if (! isSPACE_A(name[i])) {
23693 /* Most punctuation after the equals indicates a subpattern, like
23695 if ( isPUNCT_A(name[i])
23700 /* A backslash means the real delimitter is the next character,
23701 * but it must be punctuation */
23702 && (name[i] != '\\' || (i < name_len && isPUNCT_A(name[i+1]))))
23704 /* Find the property. The table includes the equals sign, so we
23706 table_index = do_uniprop_match(lookup_name, j);
23708 const char * const * prop_values
23709 = get_prop_values(table_index);
23710 REGEXP * subpattern_re;
23711 char open = name[i++];
23713 const char * pos_in_brackets;
23716 /* Backslash => delimitter is the character following. We
23717 * already checked that it is punctuation */
23718 if (open == '\\') {
23723 /* This data structure is constructed so that the matching
23724 * closing bracket is 3 past its matching opening. The second
23725 * set of closing is so that if the opening is something like
23726 * ']', the closing will be that as well. Something similar is
23727 * done in toke.c */
23728 pos_in_brackets = memCHRs("([<)]>)]>", open);
23729 close = (pos_in_brackets) ? pos_in_brackets[3] : open;
23732 || name[name_len-1] != close
23733 || (escaped && name[name_len-2] != '\\')
23734 /* Also make sure that there are enough characters.
23735 * e.g., '\\\' would show up incorrectly as legal even
23736 * though it is too short */
23737 || (SSize_t) (name_len - i - 1 - escaped) < 0)
23739 sv_catpvs(msg, "Unicode property wildcard not terminated");
23740 goto append_name_to_msg;
23743 Perl_ck_warner_d(aTHX_
23744 packWARN(WARN_EXPERIMENTAL__UNIPROP_WILDCARDS),
23745 "The Unicode property wildcards feature is experimental");
23747 /* Now create and compile the wildcard subpattern. Use /i
23748 * because the property values are supposed to match with case
23750 subpattern_re = compile_wildcard(name + i,
23751 name_len - i - 1 - escaped,
23755 /* For each legal property value, see if the supplied pattern
23757 while (*prop_values) {
23758 const char * const entry = *prop_values;
23759 const Size_t len = strlen(entry);
23760 SV* entry_sv = newSVpvn_flags(entry, len, SVs_TEMP);
23762 if (execute_wildcard(subpattern_re,
23764 (char *) entry + len,
23768 { /* Here, matched. Add to the returned list */
23769 Size_t total_len = j + len;
23770 SV * sub_invlist = NULL;
23771 char * this_string;
23773 /* We know this is a legal \p{property=value}. Call
23774 * the function to return the list of code points that
23776 Newxz(this_string, total_len + 1, char);
23777 Copy(lookup_name, this_string, j, char);
23778 my_strlcat(this_string, entry, total_len + 1);
23779 SAVEFREEPV(this_string);
23780 sub_invlist = parse_uniprop_string(this_string,
23789 _invlist_union(prop_definition, sub_invlist,
23793 prop_values++; /* Next iteration, look at next propvalue */
23794 } /* End of looking through property values; (the data
23795 structure is terminated by a NULL ptr) */
23797 SvREFCNT_dec_NN(subpattern_re);
23799 if (prop_definition) {
23800 return prop_definition;
23803 sv_catpvs(msg, "No Unicode property value wildcard matches:");
23804 goto append_name_to_msg;
23807 /* Here's how khw thinks we should proceed to handle the properties
23808 * not yet done: Bidi Mirroring Glyph
23809 Bidi Paired Bracket
23810 Case Folding (both full and simple)
23811 Decomposition Mapping
23812 Equivalent Unified Ideograph
23815 Lowercase Mapping (both full and simple)
23817 Titlecase Mapping (both full and simple)
23818 Uppercase Mapping (both full and simple)
23819 * Move the part that looks at the property values into a perl
23820 * script, like utf8_heavy.pl was done. This makes things somewhat
23821 * easier, but most importantly, it avoids always adding all these
23822 * strings to the memory usage when the feature is little-used.
23824 * The property values would all be concatenated into a single
23825 * string per property with each value on a separate line, and the
23826 * code point it's for on alternating lines. Then we match the
23827 * user's input pattern m//mg, without having to worry about their
23828 * uses of '^' and '$'. Only the values that aren't the default
23829 * would be in the strings. Code points would be in UTF-8. The
23830 * search pattern that we would construct would look like
23831 * (?: \n (code-point_re) \n (?aam: user-re ) \n )
23832 * And so $1 would contain the code point that matched the user-re.
23833 * For properties where the default is the code point itself, such
23834 * as any of the case changing mappings, the string would otherwise
23835 * consist of all Unicode code points in UTF-8 strung together.
23836 * This would be impractical. So instead, examine their compiled
23837 * pattern, looking at the ssc. If none, reject the pattern as an
23838 * error. Otherwise run the pattern against every code point in
23839 * the ssc. The ssc is kind of like tr18's 3.9 Possible Match Sets
23840 * And it might be good to create an API to return the ssc.
23842 * For the name properties, a new function could be created in
23843 * charnames which essentially does the same thing as above,
23844 * sharing Name.pl with the other charname functions. Don't know
23845 * about loose name matching, or algorithmically determined names.
23846 * Decomposition.pl similarly.
23848 * It might be that a new pattern modifier would have to be
23849 * created, like /t for resTricTed, which changed the behavior of
23850 * some constructs in their subpattern, like \A. */
23851 } /* End of is a wildcard subppattern */
23853 /* \p{name=...} is handled specially. Instead of using the normal
23854 * mechanism involving charclass_invlists.h, it uses _charnames.pm
23855 * which has the necessary (huge) data accessible to it, and which
23856 * doesn't get loaded unless necessary. The legal syntax for names is
23857 * somewhat different than other properties due both to the vagaries of
23858 * a few outlier official names, and the fact that only a few ASCII
23859 * characters are permitted in them */
23860 if ( memEQs(lookup_name, j - 1, "name")
23861 || memEQs(lookup_name, j - 1, "na"))
23866 const char * error_msg;
23868 SV * character_name;
23869 STRLEN character_len;
23874 /* Since the RHS (after skipping initial space) is passed unchanged
23875 * to charnames, and there are different criteria for what are
23876 * legal characters in the name, just parse it here. A character
23877 * name must begin with an ASCII alphabetic */
23878 if (! isALPHA(name[i])) {
23881 lookup_name[j++] = name[i];
23883 for (++i; i < name_len; i++) {
23884 /* Official names can only be in the ASCII range, and only
23885 * certain characters */
23886 if (! isASCII(name[i]) || ! isCHARNAME_CONT(name[i])) {
23889 lookup_name[j++] = name[i];
23892 /* Finished parsing, save the name into an SV */
23893 character_name = newSVpvn(lookup_name + equals_pos, j - equals_pos);
23895 /* Make sure _charnames is loaded. (The parameters give context
23896 * for any errors generated */
23897 table = load_charnames(character_name, name, name_len, &error_msg);
23898 if (table == NULL) {
23899 sv_catpv(msg, error_msg);
23900 goto append_name_to_msg;
23903 lookup_loose = get_cv("_charnames::_loose_regcomp_lookup", 0);
23904 if (! lookup_loose) {
23906 "panic: Can't find '_charnames::_loose_regcomp_lookup");
23909 PUSHSTACKi(PERLSI_OVERLOAD);
23915 XPUSHs(character_name);
23917 call_sv(MUTABLE_SV(lookup_loose), G_SCALAR);
23922 SvREFCNT_inc_simple_void_NN(character);
23929 if (! SvOK(character)) {
23933 cp = valid_utf8_to_uvchr((U8 *) SvPVX(character), &character_len);
23934 if (character_len < SvCUR(character)) {
23938 prop_definition = add_cp_to_invlist(NULL, cp);
23939 return prop_definition;
23942 /* Certain properties whose values are numeric need special handling.
23943 * They may optionally be prefixed by 'is'. Ignore that prefix for the
23944 * purposes of checking if this is one of those properties */
23945 if (memBEGINPs(lookup_name, j, "is")) {
23949 /* Then check if it is one of these specially-handled properties. The
23950 * possibilities are hard-coded because easier this way, and the list
23951 * is unlikely to change.
23953 * All numeric value type properties are of this ilk, and are also
23954 * special in a different way later on. So find those first. There
23955 * are several numeric value type properties in the Unihan DB (which is
23956 * unlikely to be compiled with perl, but we handle it here in case it
23957 * does get compiled). They all end with 'numeric'. The interiors
23958 * aren't checked for the precise property. This would stop working if
23959 * a cjk property were to be created that ended with 'numeric' and
23960 * wasn't a numeric type */
23961 is_nv_type = memEQs(lookup_name + lookup_offset,
23962 j - 1 - lookup_offset, "numericvalue")
23963 || memEQs(lookup_name + lookup_offset,
23964 j - 1 - lookup_offset, "nv")
23965 || ( memENDPs(lookup_name + lookup_offset,
23966 j - 1 - lookup_offset, "numeric")
23967 && ( memBEGINPs(lookup_name + lookup_offset,
23968 j - 1 - lookup_offset, "cjk")
23969 || memBEGINPs(lookup_name + lookup_offset,
23970 j - 1 - lookup_offset, "k")));
23972 || memEQs(lookup_name + lookup_offset,
23973 j - 1 - lookup_offset, "canonicalcombiningclass")
23974 || memEQs(lookup_name + lookup_offset,
23975 j - 1 - lookup_offset, "ccc")
23976 || memEQs(lookup_name + lookup_offset,
23977 j - 1 - lookup_offset, "age")
23978 || memEQs(lookup_name + lookup_offset,
23979 j - 1 - lookup_offset, "in")
23980 || memEQs(lookup_name + lookup_offset,
23981 j - 1 - lookup_offset, "presentin"))
23985 /* Since the stuff after the '=' is a number, we can't throw away
23986 * '-' willy-nilly, as those could be a minus sign. Other stricter
23987 * rules also apply. However, these properties all can have the
23988 * rhs not be a number, in which case they contain at least one
23989 * alphabetic. In those cases, the stricter rules don't apply.
23990 * But the numeric type properties can have the alphas [Ee] to
23991 * signify an exponent, and it is still a number with stricter
23992 * rules. So look for an alpha that signifies not-strict */
23994 for (k = i; k < name_len; k++) {
23995 if ( isALPHA_A(name[k])
23996 && (! is_nv_type || ! isALPHA_FOLD_EQ(name[k], 'E')))
23998 stricter = Not_Strict;
24006 /* A number may have a leading '+' or '-'. The latter is retained
24008 if (name[i] == '+') {
24011 else if (name[i] == '-') {
24012 lookup_name[j++] = '-';
24016 /* Skip leading zeros including single underscores separating the
24017 * zeros, or between the final leading zero and the first other
24019 for (; i < name_len - 1; i++) {
24020 if ( name[i] != '0'
24021 && (name[i] != '_' || ! isDIGIT_A(name[i+1])))
24028 else { /* No '=' */
24030 /* Only a few properties without an '=' should be parsed with stricter
24031 * rules. The list is unlikely to change. */
24032 if ( memBEGINPs(lookup_name, j, "perl")
24033 && memNEs(lookup_name + 4, j - 4, "space")
24034 && memNEs(lookup_name + 4, j - 4, "word"))
24038 /* We set the inputs back to 0 and the code below will reparse,
24044 /* Here, we have either finished the property, or are positioned to parse
24045 * the remainder, and we know if stricter rules apply. Finish out, if not
24047 for (; i < name_len; i++) {
24048 char cur = name[i];
24050 /* In all instances, case differences are ignored, and we normalize to
24052 if (isUPPER_A(cur)) {
24053 lookup_name[j++] = toLOWER(cur);
24057 /* An underscore is skipped, but not under strict rules unless it
24058 * separates two digits */
24061 && ( i == 0 || (int) i == equals_pos || i == name_len- 1
24062 || ! isDIGIT_A(name[i-1]) || ! isDIGIT_A(name[i+1])))
24064 lookup_name[j++] = '_';
24069 /* Hyphens are skipped except under strict */
24070 if (cur == '-' && ! stricter) {
24074 /* XXX Bug in documentation. It says white space skipped adjacent to
24075 * non-word char. Maybe we should, but shouldn't skip it next to a dot
24077 if (isSPACE_A(cur) && ! stricter) {
24081 lookup_name[j++] = cur;
24083 /* Unless this is a non-trailing slash, we are done with it */
24084 if (i >= name_len - 1 || cur != '/') {
24090 /* A slash in the 'numeric value' property indicates that what follows
24091 * is a denominator. It can have a leading '+' and '0's that should be
24092 * skipped. But we have never allowed a negative denominator, so treat
24093 * a minus like every other character. (No need to rule out a second
24094 * '/', as that won't match anything anyway */
24097 if (i < name_len && name[i] == '+') {
24101 /* Skip leading zeros including underscores separating digits */
24102 for (; i < name_len - 1; i++) {
24103 if ( name[i] != '0'
24104 && (name[i] != '_' || ! isDIGIT_A(name[i+1])))
24110 /* Store the first real character in the denominator */
24111 if (i < name_len) {
24112 lookup_name[j++] = name[i];
24117 /* Here are completely done parsing the input 'name', and 'lookup_name'
24118 * contains a copy, normalized.
24120 * This special case is grandfathered in: 'L_' and 'GC=L_' are accepted and
24121 * different from without the underscores. */
24122 if ( ( UNLIKELY(memEQs(lookup_name, j, "l"))
24123 || UNLIKELY(memEQs(lookup_name, j, "gc=l")))
24124 && UNLIKELY(name[name_len-1] == '_'))
24126 lookup_name[j++] = '&';
24129 /* If the original input began with 'In' or 'Is', it could be a subroutine
24130 * call to a user-defined property instead of a Unicode property name. */
24131 if ( name_len - non_pkg_begin > 2
24132 && name[non_pkg_begin+0] == 'I'
24133 && (name[non_pkg_begin+1] == 'n' || name[non_pkg_begin+1] == 's'))
24135 /* Names that start with In have different characterstics than those
24136 * that start with Is */
24137 if (name[non_pkg_begin+1] == 's') {
24138 starts_with_Is = TRUE;
24142 could_be_user_defined = FALSE;
24145 if (could_be_user_defined) {
24148 /* If the user defined property returns the empty string, it could
24149 * easily be because the pattern is being compiled before the data it
24150 * actually needs to compile is available. This could be argued to be
24151 * a bug in the perl code, but this is a change of behavior for Perl,
24152 * so we handle it. This means that intentionally returning nothing
24153 * will not be resolved until runtime */
24154 bool empty_return = FALSE;
24156 /* Here, the name could be for a user defined property, which are
24157 * implemented as subs. */
24158 user_sub = get_cvn_flags(name, name_len, 0);
24161 /* Here, the property name could be a user-defined one, but there
24162 * is no subroutine to handle it (as of now). Defer handling it
24163 * until runtime. Otherwise, a block defined by Unicode in a later
24164 * release would get the synonym InFoo added for it, and existing
24165 * code that used that name would suddenly break if it referred to
24166 * the property before the sub was declared. See [perl #134146] */
24168 goto definition_deferred;
24171 /* Here, we are at runtime, and didn't find the user property. It
24172 * could be an official property, but only if no package was
24173 * specified, or just the utf8:: package. */
24174 if (could_be_deferred_official) {
24175 lookup_name += lun_non_pkg_begin;
24176 j -= lun_non_pkg_begin;
24178 else if (! stripped_utf8_pkg) {
24179 goto unknown_user_defined;
24182 /* Drop down to look up in the official properties */
24185 const char insecure[] = "Insecure user-defined property";
24187 /* Here, there is a sub by the correct name. Normally we call it
24188 * to get the property definition */
24190 SV * user_sub_sv = MUTABLE_SV(user_sub);
24191 SV * error; /* Any error returned by calling 'user_sub' */
24192 SV * key; /* The key into the hash of user defined sub names
24195 SV ** saved_user_prop_ptr; /* Hash entry for this property */
24197 /* How many times to retry when another thread is in the middle of
24198 * expanding the same definition we want */
24199 PERL_INT_FAST8_T retry_countdown = 10;
24201 DECLARATION_FOR_GLOBAL_CONTEXT;
24203 /* If we get here, we know this property is user-defined */
24204 *user_defined_ptr = TRUE;
24206 /* We refuse to call a potentially tainted subroutine; returning an
24209 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24210 sv_catpvn(msg, insecure, sizeof(insecure) - 1);
24211 goto append_name_to_msg;
24214 /* In principal, we only call each subroutine property definition
24215 * once during the life of the program. This guarantees that the
24216 * property definition never changes. The results of the single
24217 * sub call are stored in a hash, which is used instead for future
24218 * references to this property. The property definition is thus
24219 * immutable. But, to allow the user to have a /i-dependent
24220 * definition, we call the sub once for non-/i, and once for /i,
24221 * should the need arise, passing the /i status as a parameter.
24223 * We start by constructing the hash key name, consisting of the
24224 * fully qualified subroutine name, preceded by the /i status, so
24225 * that there is a key for /i and a different key for non-/i */
24226 key = newSVpvn(((to_fold) ? "1" : "0"), 1);
24227 fq_name = S_get_fq_name(aTHX_ name, name_len, is_utf8,
24228 non_pkg_begin != 0);
24229 sv_catsv(key, fq_name);
24232 /* We only call the sub once throughout the life of the program
24233 * (with the /i, non-/i exception noted above). That means the
24234 * hash must be global and accessible to all threads. It is
24235 * created at program start-up, before any threads are created, so
24236 * is accessible to all children. But this creates some
24239 * 1) The keys can't be shared, or else problems arise; sharing is
24240 * turned off at hash creation time
24241 * 2) All SVs in it are there for the remainder of the life of the
24242 * program, and must be created in the same interpreter context
24243 * as the hash, or else they will be freed from the wrong pool
24244 * at global destruction time. This is handled by switching to
24245 * the hash's context to create each SV going into it, and then
24246 * immediately switching back
24247 * 3) All accesses to the hash must be controlled by a mutex, to
24248 * prevent two threads from getting an unstable state should
24249 * they simultaneously be accessing it. The code below is
24250 * crafted so that the mutex is locked whenever there is an
24251 * access and unlocked only when the next stable state is
24254 * The hash stores either the definition of the property if it was
24255 * valid, or, if invalid, the error message that was raised. We
24256 * use the type of SV to distinguish.
24258 * There's also the need to guard against the definition expansion
24259 * from infinitely recursing. This is handled by storing the aTHX
24260 * of the expanding thread during the expansion. Again the SV type
24261 * is used to distinguish this from the other two cases. If we
24262 * come to here and the hash entry for this property is our aTHX,
24263 * it means we have recursed, and the code assumes that we would
24264 * infinitely recurse, so instead stops and raises an error.
24265 * (Any recursion has always been treated as infinite recursion in
24268 * If instead, the entry is for a different aTHX, it means that
24269 * that thread has gotten here first, and hasn't finished expanding
24270 * the definition yet. We just have to wait until it is done. We
24271 * sleep and retry a few times, returning an error if the other
24272 * thread doesn't complete. */
24275 USER_PROP_MUTEX_LOCK;
24277 /* If we have an entry for this key, the subroutine has already
24278 * been called once with this /i status. */
24279 saved_user_prop_ptr = hv_fetch(PL_user_def_props,
24280 SvPVX(key), SvCUR(key), 0);
24281 if (saved_user_prop_ptr) {
24283 /* If the saved result is an inversion list, it is the valid
24284 * definition of this property */
24285 if (is_invlist(*saved_user_prop_ptr)) {
24286 prop_definition = *saved_user_prop_ptr;
24288 /* The SV in the hash won't be removed until global
24289 * destruction, so it is stable and we can unlock */
24290 USER_PROP_MUTEX_UNLOCK;
24292 /* The caller shouldn't try to free this SV */
24293 return prop_definition;
24296 /* Otherwise, if it is a string, it is the error message
24297 * that was returned when we first tried to evaluate this
24298 * property. Fail, and append the message */
24299 if (SvPOK(*saved_user_prop_ptr)) {
24300 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24301 sv_catsv(msg, *saved_user_prop_ptr);
24303 /* The SV in the hash won't be removed until global
24304 * destruction, so it is stable and we can unlock */
24305 USER_PROP_MUTEX_UNLOCK;
24310 assert(SvIOK(*saved_user_prop_ptr));
24312 /* Here, we have an unstable entry in the hash. Either another
24313 * thread is in the middle of expanding the property's
24314 * definition, or we are ourselves recursing. We use the aTHX
24315 * in it to distinguish */
24316 if (SvIV(*saved_user_prop_ptr) != PTR2IV(CUR_CONTEXT)) {
24318 /* Here, it's another thread doing the expanding. We've
24319 * looked as much as we are going to at the contents of the
24320 * hash entry. It's safe to unlock. */
24321 USER_PROP_MUTEX_UNLOCK;
24323 /* Retry a few times */
24324 if (retry_countdown-- > 0) {
24329 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24330 sv_catpvs(msg, "Timeout waiting for another thread to "
24332 goto append_name_to_msg;
24335 /* Here, we are recursing; don't dig any deeper */
24336 USER_PROP_MUTEX_UNLOCK;
24338 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24340 "Infinite recursion in user-defined property");
24341 goto append_name_to_msg;
24344 /* Here, this thread has exclusive control, and there is no entry
24345 * for this property in the hash. So we have the go ahead to
24346 * expand the definition ourselves. */
24348 PUSHSTACKi(PERLSI_MAGIC);
24351 /* Create a temporary placeholder in the hash to detect recursion
24353 SWITCH_TO_GLOBAL_CONTEXT;
24354 placeholder= newSVuv(PTR2IV(ORIGINAL_CONTEXT));
24355 (void) hv_store_ent(PL_user_def_props, key, placeholder, 0);
24358 /* Now that we have a placeholder, we can let other threads
24360 USER_PROP_MUTEX_UNLOCK;
24362 /* Make sure the placeholder always gets destroyed */
24363 SAVEDESTRUCTOR_X(S_delete_recursion_entry, SvPVX(key));
24368 /* Call the user's function, with the /i status as a parameter.
24369 * Note that we have gone to a lot of trouble to keep this call
24370 * from being within the locked mutex region. */
24371 XPUSHs(boolSV(to_fold));
24374 /* The following block was taken from swash_init(). Presumably
24375 * they apply to here as well, though we no longer use a swash --
24379 /* We might get here via a subroutine signature which uses a utf8
24380 * parameter name, at which point PL_subname will have been set
24381 * but not yet used. */
24382 save_item(PL_subname);
24384 /* G_SCALAR guarantees a single return value */
24385 (void) call_sv(user_sub_sv, G_EVAL|G_SCALAR);
24390 if (TAINT_get || SvTRUE(error)) {
24391 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24392 if (SvTRUE(error)) {
24393 sv_catpvs(msg, "Error \"");
24394 sv_catsv(msg, error);
24395 sv_catpvs(msg, "\"");
24398 if (SvTRUE(error)) sv_catpvs(msg, "; ");
24399 sv_catpvn(msg, insecure, sizeof(insecure) - 1);
24402 if (name_len > 0) {
24403 sv_catpvs(msg, " in expansion of ");
24404 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8,
24410 prop_definition = NULL;
24413 SV * contents = POPs;
24415 /* The contents is supposed to be the expansion of the property
24416 * definition. If the definition is deferrable, and we got an
24417 * empty string back, set a flag to later defer it (after clean
24420 && (! SvPOK(contents) || SvCUR(contents) == 0))
24422 empty_return = TRUE;
24424 else { /* Otherwise, call a function to check for valid syntax,
24427 prop_definition = handle_user_defined_property(
24429 is_utf8, to_fold, runtime,
24431 contents, user_defined_ptr,
24437 /* Here, we have the results of the expansion. Delete the
24438 * placeholder, and if the definition is now known, replace it with
24439 * that definition. We need exclusive access to the hash, and we
24440 * can't let anyone else in, between when we delete the placeholder
24441 * and add the permanent entry */
24442 USER_PROP_MUTEX_LOCK;
24444 S_delete_recursion_entry(aTHX_ SvPVX(key));
24446 if ( ! empty_return
24447 && (! prop_definition || is_invlist(prop_definition)))
24449 /* If we got success we use the inversion list defining the
24450 * property; otherwise use the error message */
24451 SWITCH_TO_GLOBAL_CONTEXT;
24452 (void) hv_store_ent(PL_user_def_props,
24455 ? newSVsv(prop_definition)
24461 /* All done, and the hash now has a permanent entry for this
24462 * property. Give up exclusive control */
24463 USER_PROP_MUTEX_UNLOCK;
24469 if (empty_return) {
24470 goto definition_deferred;
24473 if (prop_definition) {
24475 /* If the definition is for something not known at this time,
24476 * we toss it, and go return the main property name, as that's
24477 * the one the user will be aware of */
24478 if (! is_invlist(prop_definition)) {
24479 SvREFCNT_dec_NN(prop_definition);
24480 goto definition_deferred;
24483 sv_2mortal(prop_definition);
24487 return prop_definition;
24489 } /* End of calling the subroutine for the user-defined property */
24490 } /* End of it could be a user-defined property */
24492 /* Here it wasn't a user-defined property that is known at this time. See
24493 * if it is a Unicode property */
24495 lookup_len = j; /* This is a more mnemonic name than 'j' */
24497 /* Get the index into our pointer table of the inversion list corresponding
24498 * to the property */
24499 table_index = do_uniprop_match(lookup_name, lookup_len);
24501 /* If it didn't find the property ... */
24502 if (table_index == 0) {
24504 /* Try again stripping off any initial 'Is'. This is because we
24505 * promise that an initial Is is optional. The same isn't true of
24506 * names that start with 'In'. Those can match only blocks, and the
24507 * lookup table already has those accounted for. */
24508 if (starts_with_Is) {
24514 table_index = do_uniprop_match(lookup_name, lookup_len);
24517 if (table_index == 0) {
24520 /* Here, we didn't find it. If not a numeric type property, and
24521 * can't be a user-defined one, it isn't a legal property */
24522 if (! is_nv_type) {
24523 if (! could_be_user_defined) {
24527 /* Here, the property name is legal as a user-defined one. At
24528 * compile time, it might just be that the subroutine for that
24529 * property hasn't been encountered yet, but at runtime, it's
24530 * an error to try to use an undefined one */
24531 if (! deferrable) {
24532 goto unknown_user_defined;;
24535 goto definition_deferred;
24536 } /* End of isn't a numeric type property */
24538 /* The numeric type properties need more work to decide. What we
24539 * do is make sure we have the number in canonical form and look
24542 if (slash_pos < 0) { /* No slash */
24544 /* When it isn't a rational, take the input, convert it to a
24545 * NV, then create a canonical string representation of that
24549 SSize_t value_len = lookup_len - equals_pos;
24551 /* Get the value */
24552 if ( value_len <= 0
24553 || my_atof3(lookup_name + equals_pos, &value,
24555 != lookup_name + lookup_len)
24560 /* If the value is an integer, the canonical value is integral
24562 if (Perl_ceil(value) == value) {
24563 canonical = Perl_form(aTHX_ "%.*s%.0" NVff,
24564 equals_pos, lookup_name, value);
24566 else { /* Otherwise, it is %e with a known precision */
24569 canonical = Perl_form(aTHX_ "%.*s%.*" NVef,
24570 equals_pos, lookup_name,
24571 PL_E_FORMAT_PRECISION, value);
24573 /* The exponent generated is expecting two digits, whereas
24574 * %e on some systems will generate three. Remove leading
24575 * zeros in excess of 2 from the exponent. We start
24576 * looking for them after the '=' */
24577 exp_ptr = strchr(canonical + equals_pos, 'e');
24579 char * cur_ptr = exp_ptr + 2; /* past the 'e[+-]' */
24580 SSize_t excess_exponent_len = strlen(cur_ptr) - 2;
24582 assert(*(cur_ptr - 1) == '-' || *(cur_ptr - 1) == '+');
24584 if (excess_exponent_len > 0) {
24585 SSize_t leading_zeros = strspn(cur_ptr, "0");
24586 SSize_t excess_leading_zeros
24587 = MIN(leading_zeros, excess_exponent_len);
24588 if (excess_leading_zeros > 0) {
24589 Move(cur_ptr + excess_leading_zeros,
24591 strlen(cur_ptr) - excess_leading_zeros
24592 + 1, /* Copy the NUL as well */
24599 else { /* Has a slash. Create a rational in canonical form */
24600 UV numerator, denominator, gcd, trial;
24601 const char * end_ptr;
24602 const char * sign = "";
24604 /* We can't just find the numerator, denominator, and do the
24605 * division, then use the method above, because that is
24606 * inexact. And the input could be a rational that is within
24607 * epsilon (given our precision) of a valid rational, and would
24608 * then incorrectly compare valid.
24610 * We're only interested in the part after the '=' */
24611 const char * this_lookup_name = lookup_name + equals_pos;
24612 lookup_len -= equals_pos;
24613 slash_pos -= equals_pos;
24615 /* Handle any leading minus */
24616 if (this_lookup_name[0] == '-') {
24618 this_lookup_name++;
24623 /* Convert the numerator to numeric */
24624 end_ptr = this_lookup_name + slash_pos;
24625 if (! grok_atoUV(this_lookup_name, &numerator, &end_ptr)) {
24629 /* It better have included all characters before the slash */
24630 if (*end_ptr != '/') {
24634 /* Set to look at just the denominator */
24635 this_lookup_name += slash_pos;
24636 lookup_len -= slash_pos;
24637 end_ptr = this_lookup_name + lookup_len;
24639 /* Convert the denominator to numeric */
24640 if (! grok_atoUV(this_lookup_name, &denominator, &end_ptr)) {
24644 /* It better be the rest of the characters, and don't divide by
24646 if ( end_ptr != this_lookup_name + lookup_len
24647 || denominator == 0)
24652 /* Get the greatest common denominator using
24653 http://en.wikipedia.org/wiki/Euclidean_algorithm */
24655 trial = denominator;
24656 while (trial != 0) {
24658 trial = gcd % trial;
24662 /* If already in lowest possible terms, we have already tried
24663 * looking this up */
24668 /* Reduce the rational, which should put it in canonical form
24671 denominator /= gcd;
24673 canonical = Perl_form(aTHX_ "%.*s%s%" UVuf "/%" UVuf,
24674 equals_pos, lookup_name, sign, numerator, denominator);
24677 /* Here, we have the number in canonical form. Try that */
24678 table_index = do_uniprop_match(canonical, strlen(canonical));
24679 if (table_index == 0) {
24682 } /* End of still didn't find the property in our table */
24683 } /* End of didn't find the property in our table */
24685 /* Here, we have a non-zero return, which is an index into a table of ptrs.
24686 * A negative return signifies that the real index is the absolute value,
24687 * but the result needs to be inverted */
24688 if (table_index < 0) {
24689 invert_return = TRUE;
24690 table_index = -table_index;
24693 /* Out-of band indices indicate a deprecated property. The proper index is
24694 * modulo it with the table size. And dividing by the table size yields
24695 * an offset into a table constructed by regen/mk_invlists.pl to contain
24696 * the corresponding warning message */
24697 if (table_index > MAX_UNI_KEYWORD_INDEX) {
24698 Size_t warning_offset = table_index / MAX_UNI_KEYWORD_INDEX;
24699 table_index %= MAX_UNI_KEYWORD_INDEX;
24700 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED),
24701 "Use of '%.*s' in \\p{} or \\P{} is deprecated because: %s",
24702 (int) name_len, name,
24703 get_deprecated_property_msg(warning_offset));
24706 /* In a few properties, a different property is used under /i. These are
24707 * unlikely to change, so are hard-coded here. */
24709 if ( table_index == UNI_XPOSIXUPPER
24710 || table_index == UNI_XPOSIXLOWER
24711 || table_index == UNI_TITLE)
24713 table_index = UNI_CASED;
24715 else if ( table_index == UNI_UPPERCASELETTER
24716 || table_index == UNI_LOWERCASELETTER
24717 # ifdef UNI_TITLECASELETTER /* Missing from early Unicodes */
24718 || table_index == UNI_TITLECASELETTER
24721 table_index = UNI_CASEDLETTER;
24723 else if ( table_index == UNI_POSIXUPPER
24724 || table_index == UNI_POSIXLOWER)
24726 table_index = UNI_POSIXALPHA;
24730 /* Create and return the inversion list */
24731 prop_definition = get_prop_definition(table_index);
24732 sv_2mortal(prop_definition);
24734 /* See if there is a private use override to add to this definition */
24736 COPHH * hinthash = (IN_PERL_COMPILETIME)
24737 ? CopHINTHASH_get(&PL_compiling)
24738 : CopHINTHASH_get(PL_curcop);
24739 SV * pu_overrides = cophh_fetch_pv(hinthash, "private_use", 0, 0);
24741 if (UNLIKELY(pu_overrides && SvPOK(pu_overrides))) {
24743 /* See if there is an element in the hints hash for this table */
24744 SV * pu_lookup = Perl_newSVpvf(aTHX_ "%d=", table_index);
24745 const char * pos = strstr(SvPVX(pu_overrides), SvPVX(pu_lookup));
24749 SV * pu_definition;
24751 SV * expanded_prop_definition =
24752 sv_2mortal(invlist_clone(prop_definition, NULL));
24754 /* If so, it's definition is the string from here to the next
24755 * \a character. And its format is the same as a user-defined
24757 pos += SvCUR(pu_lookup);
24758 pu_definition = newSVpvn(pos, strchr(pos, '\a') - pos);
24759 pu_invlist = handle_user_defined_property(lookup_name,
24762 0, /* Not folded */
24770 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24771 sv_catpvs(msg, "Insecure private-use override");
24772 goto append_name_to_msg;
24775 /* For now, as a safety measure, make sure that it doesn't
24776 * override non-private use code points */
24777 _invlist_intersection(pu_invlist, PL_Private_Use, &pu_invlist);
24779 /* Add it to the list to be returned */
24780 _invlist_union(prop_definition, pu_invlist,
24781 &expanded_prop_definition);
24782 prop_definition = expanded_prop_definition;
24783 Perl_ck_warner_d(aTHX_ packWARN(WARN_EXPERIMENTAL__PRIVATE_USE), "The private_use feature is experimental");
24788 if (invert_return) {
24789 _invlist_invert(prop_definition);
24791 return prop_definition;
24793 unknown_user_defined:
24794 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24795 sv_catpvs(msg, "Unknown user-defined property name");
24796 goto append_name_to_msg;
24799 if (non_pkg_begin != 0) {
24800 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24801 sv_catpvs(msg, "Illegal user-defined property name");
24804 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24805 sv_catpvs(msg, "Can't find Unicode property definition");
24809 append_name_to_msg:
24811 const char * prefix = (runtime && level == 0) ? " \\p{" : " \"";
24812 const char * suffix = (runtime && level == 0) ? "}" : "\"";
24814 sv_catpv(msg, prefix);
24815 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8, name_len, name));
24816 sv_catpv(msg, suffix);
24821 definition_deferred:
24824 bool is_qualified = non_pkg_begin != 0; /* If has "::" */
24826 /* Here it could yet to be defined, so defer evaluation of this until
24827 * its needed at runtime. We need the fully qualified property name to
24828 * avoid ambiguity */
24830 fq_name = S_get_fq_name(aTHX_ name, name_len, is_utf8,
24834 /* If it didn't come with a package, or the package is utf8::, this
24835 * actually could be an official Unicode property whose inclusion we
24836 * are deferring until runtime to make sure that it isn't overridden by
24837 * a user-defined property of the same name (which we haven't
24838 * encountered yet). Add a marker to indicate this possibility, for
24839 * use at such time when we first need the definition during pattern
24840 * matching execution */
24841 if (! is_qualified || memBEGINPs(name, non_pkg_begin, "utf8::")) {
24842 sv_catpvs(fq_name, DEFERRED_COULD_BE_OFFICIAL_MARKERs);
24845 /* We also need a trailing newline */
24846 sv_catpvs(fq_name, "\n");
24848 *user_defined_ptr = TRUE;
24854 * ex: set ts=8 sts=4 sw=4 et: