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 #define PERL_IN_REGCOMP_C
78 #ifdef PERL_IN_XSUB_RE
80 EXTERN_C const struct regexp_engine my_reg_engine;
85 #include "dquote_inline.h"
86 #include "invlist_inline.h"
87 #include "unicode_constants.h"
89 #define HAS_NONLATIN1_FOLD_CLOSURE(i) \
90 _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
91 #define HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(i) \
92 _HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
93 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
94 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
100 /* this is a chain of data about sub patterns we are processing that
101 need to be handled separately/specially in study_chunk. Its so
102 we can simulate recursion without losing state. */
104 typedef struct scan_frame {
105 regnode *last_regnode; /* last node to process in this frame */
106 regnode *next_regnode; /* next node to process when last is reached */
107 U32 prev_recursed_depth;
108 I32 stopparen; /* what stopparen do we use */
110 struct scan_frame *this_prev_frame; /* this previous frame */
111 struct scan_frame *prev_frame; /* previous frame */
112 struct scan_frame *next_frame; /* next frame */
115 /* Certain characters are output as a sequence with the first being a
117 #define isBACKSLASHED_PUNCT(c) strchr("-[]\\^", c)
120 struct RExC_state_t {
121 U32 flags; /* RXf_* are we folding, multilining? */
122 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
123 char *precomp; /* uncompiled string. */
124 char *precomp_end; /* pointer to end of uncompiled string. */
125 REGEXP *rx_sv; /* The SV that is the regexp. */
126 regexp *rx; /* perl core regexp structure */
127 regexp_internal *rxi; /* internal data for regexp object
129 char *start; /* Start of input for compile */
130 char *end; /* End of input for compile */
131 char *parse; /* Input-scan pointer. */
132 char *copy_start; /* start of copy of input within
133 constructed parse string */
134 char *copy_start_in_input; /* Position in input string
135 corresponding to copy_start */
136 SSize_t whilem_seen; /* number of WHILEM in this expr */
137 regnode *emit_start; /* Start of emitted-code area */
138 regnode_offset emit; /* Code-emit pointer */
139 I32 naughty; /* How bad is this pattern? */
140 I32 sawback; /* Did we see \1, ...? */
142 SSize_t size; /* Number of regnode equivalents in
145 /* position beyond 'precomp' of the warning message furthest away from
146 * 'precomp'. During the parse, no warnings are raised for any problems
147 * earlier in the parse than this position. This works if warnings are
148 * raised the first time a given spot is parsed, and if only one
149 * independent warning is raised for any given spot */
150 Size_t latest_warn_offset;
152 I32 npar; /* Capture buffer count so far in the
153 parse, (OPEN) plus one. ("par" 0 is
155 I32 total_par; /* During initial parse, is either 0,
156 or -1; the latter indicating a
157 reparse is needed. After that pass,
158 it is what 'npar' became after the
159 pass. Hence, it being > 0 indicates
160 we are in a reparse situation */
161 I32 nestroot; /* root parens we are in - used by
164 regnode_offset *open_parens; /* offsets to open parens */
165 regnode_offset *close_parens; /* offsets to close parens */
166 I32 parens_buf_size; /* #slots malloced open/close_parens */
167 regnode *end_op; /* END node in program */
168 I32 utf8; /* whether the pattern is utf8 or not */
169 I32 orig_utf8; /* whether the pattern was originally in utf8 */
170 /* XXX use this for future optimisation of case
171 * where pattern must be upgraded to utf8. */
172 I32 uni_semantics; /* If a d charset modifier should use unicode
173 rules, even if the pattern is not in
175 HV *paren_names; /* Paren names */
177 regnode **recurse; /* Recurse regops */
178 I32 recurse_count; /* Number of recurse regops we have generated */
179 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
181 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
184 I32 override_recoding;
186 I32 recode_x_to_native;
188 I32 in_multi_char_class;
189 struct reg_code_blocks *code_blocks;/* positions of literal (?{})
191 int code_index; /* next code_blocks[] slot */
192 SSize_t maxlen; /* mininum possible number of chars in string to match */
193 scan_frame *frame_head;
194 scan_frame *frame_last;
198 #ifdef ADD_TO_REGEXEC
199 char *starttry; /* -Dr: where regtry was called. */
200 #define RExC_starttry (pRExC_state->starttry)
202 SV *runtime_code_qr; /* qr with the runtime code blocks */
204 const char *lastparse;
206 AV *paren_name_list; /* idx -> name */
207 U32 study_chunk_recursed_count;
211 #define RExC_lastparse (pRExC_state->lastparse)
212 #define RExC_lastnum (pRExC_state->lastnum)
213 #define RExC_paren_name_list (pRExC_state->paren_name_list)
214 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
215 #define RExC_mysv (pRExC_state->mysv1)
216 #define RExC_mysv1 (pRExC_state->mysv1)
217 #define RExC_mysv2 (pRExC_state->mysv2)
227 #define RExC_flags (pRExC_state->flags)
228 #define RExC_pm_flags (pRExC_state->pm_flags)
229 #define RExC_precomp (pRExC_state->precomp)
230 #define RExC_copy_start_in_input (pRExC_state->copy_start_in_input)
231 #define RExC_copy_start_in_constructed (pRExC_state->copy_start)
232 #define RExC_precomp_end (pRExC_state->precomp_end)
233 #define RExC_rx_sv (pRExC_state->rx_sv)
234 #define RExC_rx (pRExC_state->rx)
235 #define RExC_rxi (pRExC_state->rxi)
236 #define RExC_start (pRExC_state->start)
237 #define RExC_end (pRExC_state->end)
238 #define RExC_parse (pRExC_state->parse)
239 #define RExC_latest_warn_offset (pRExC_state->latest_warn_offset )
240 #define RExC_whilem_seen (pRExC_state->whilem_seen)
241 #define RExC_seen_d_op (pRExC_state->seen_d_op) /* Seen something that differs
242 under /d from /u ? */
245 #ifdef RE_TRACK_PATTERN_OFFSETS
246 # define RExC_offsets (RExC_rxi->u.offsets) /* I am not like the
249 #define RExC_emit (pRExC_state->emit)
250 #define RExC_emit_start (pRExC_state->emit_start)
251 #define RExC_sawback (pRExC_state->sawback)
252 #define RExC_seen (pRExC_state->seen)
253 #define RExC_size (pRExC_state->size)
254 #define RExC_maxlen (pRExC_state->maxlen)
255 #define RExC_npar (pRExC_state->npar)
256 #define RExC_total_parens (pRExC_state->total_par)
257 #define RExC_parens_buf_size (pRExC_state->parens_buf_size)
258 #define RExC_nestroot (pRExC_state->nestroot)
259 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
260 #define RExC_utf8 (pRExC_state->utf8)
261 #define RExC_uni_semantics (pRExC_state->uni_semantics)
262 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
263 #define RExC_open_parens (pRExC_state->open_parens)
264 #define RExC_close_parens (pRExC_state->close_parens)
265 #define RExC_end_op (pRExC_state->end_op)
266 #define RExC_paren_names (pRExC_state->paren_names)
267 #define RExC_recurse (pRExC_state->recurse)
268 #define RExC_recurse_count (pRExC_state->recurse_count)
269 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
270 #define RExC_study_chunk_recursed_bytes \
271 (pRExC_state->study_chunk_recursed_bytes)
272 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
273 #define RExC_contains_locale (pRExC_state->contains_locale)
275 # define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
277 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
278 #define RExC_frame_head (pRExC_state->frame_head)
279 #define RExC_frame_last (pRExC_state->frame_last)
280 #define RExC_frame_count (pRExC_state->frame_count)
281 #define RExC_strict (pRExC_state->strict)
282 #define RExC_study_started (pRExC_state->study_started)
283 #define RExC_warn_text (pRExC_state->warn_text)
284 #define RExC_in_script_run (pRExC_state->in_script_run)
285 #define RExC_use_BRANCHJ (pRExC_state->use_BRANCHJ)
286 #define RExC_unlexed_names (pRExC_state->unlexed_names)
288 /* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
289 * a flag to disable back-off on the fixed/floating substrings - if it's
290 * a high complexity pattern we assume the benefit of avoiding a full match
291 * is worth the cost of checking for the substrings even if they rarely help.
293 #define RExC_naughty (pRExC_state->naughty)
294 #define TOO_NAUGHTY (10)
295 #define MARK_NAUGHTY(add) \
296 if (RExC_naughty < TOO_NAUGHTY) \
297 RExC_naughty += (add)
298 #define MARK_NAUGHTY_EXP(exp, add) \
299 if (RExC_naughty < TOO_NAUGHTY) \
300 RExC_naughty += RExC_naughty / (exp) + (add)
302 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
303 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
304 ((*s) == '{' && regcurly(s)))
307 * Flags to be passed up and down.
309 #define WORST 0 /* Worst case. */
310 #define HASWIDTH 0x01 /* Known to not match null strings, could match
313 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
314 * character. (There needs to be a case: in the switch statement in regexec.c
315 * for any node marked SIMPLE.) Note that this is not the same thing as
318 #define SPSTART 0x04 /* Starts with * or + */
319 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
320 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
321 #define RESTART_PARSE 0x20 /* Need to redo the parse */
322 #define NEED_UTF8 0x40 /* In conjunction with RESTART_PARSE, need to
323 calcuate sizes as UTF-8 */
325 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
327 /* whether trie related optimizations are enabled */
328 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
329 #define TRIE_STUDY_OPT
330 #define FULL_TRIE_STUDY
336 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
337 #define PBITVAL(paren) (1 << ((paren) & 7))
338 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
339 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
340 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
342 #define REQUIRE_UTF8(flagp) STMT_START { \
344 *flagp = RESTART_PARSE|NEED_UTF8; \
349 /* Change from /d into /u rules, and restart the parse. RExC_uni_semantics is
350 * a flag that indicates we need to override /d with /u as a result of
351 * something in the pattern. It should only be used in regards to calling
352 * set_regex_charset() or get_regex_charse() */
353 #define REQUIRE_UNI_RULES(flagp, restart_retval) \
355 if (DEPENDS_SEMANTICS) { \
356 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
357 RExC_uni_semantics = 1; \
358 if (RExC_seen_d_op && LIKELY(! IN_PARENS_PASS)) { \
359 /* No need to restart the parse if we haven't seen \
360 * anything that differs between /u and /d, and no need \
361 * to restart immediately if we're going to reparse \
362 * anyway to count parens */ \
363 *flagp |= RESTART_PARSE; \
364 return restart_retval; \
369 #define REQUIRE_BRANCHJ(flagp, restart_retval) \
371 RExC_use_BRANCHJ = 1; \
372 if (LIKELY(! IN_PARENS_PASS)) { \
373 /* No need to restart the parse immediately if we're \
374 * going to reparse anyway to count parens */ \
375 *flagp |= RESTART_PARSE; \
376 return restart_retval; \
380 /* Until we have completed the parse, we leave RExC_total_parens at 0 or
381 * less. After that, it must always be positive, because the whole re is
382 * considered to be surrounded by virtual parens. Setting it to negative
383 * indicates there is some construct that needs to know the actual number of
384 * parens to be properly handled. And that means an extra pass will be
385 * required after we've counted them all */
386 #define ALL_PARENS_COUNTED (RExC_total_parens > 0)
387 #define REQUIRE_PARENS_PASS \
388 STMT_START { /* No-op if have completed a pass */ \
389 if (! ALL_PARENS_COUNTED) RExC_total_parens = -1; \
391 #define IN_PARENS_PASS (RExC_total_parens < 0)
394 /* This is used to return failure (zero) early from the calling function if
395 * various flags in 'flags' are set. Two flags always cause a return:
396 * 'RESTART_PARSE' and 'NEED_UTF8'. 'extra' can be used to specify any
397 * additional flags that should cause a return; 0 if none. If the return will
398 * be done, '*flagp' is first set to be all of the flags that caused the
400 #define RETURN_FAIL_ON_RESTART_OR_FLAGS(flags,flagp,extra) \
402 if ((flags) & (RESTART_PARSE|NEED_UTF8|(extra))) { \
403 *(flagp) = (flags) & (RESTART_PARSE|NEED_UTF8|(extra)); \
408 #define MUST_RESTART(flags) ((flags) & (RESTART_PARSE))
410 #define RETURN_FAIL_ON_RESTART(flags,flagp) \
411 RETURN_FAIL_ON_RESTART_OR_FLAGS( flags, flagp, 0)
412 #define RETURN_FAIL_ON_RESTART_FLAGP(flagp) \
413 if (MUST_RESTART(*(flagp))) return 0
415 /* This converts the named class defined in regcomp.h to its equivalent class
416 * number defined in handy.h. */
417 #define namedclass_to_classnum(class) ((int) ((class) / 2))
418 #define classnum_to_namedclass(classnum) ((classnum) * 2)
420 #define _invlist_union_complement_2nd(a, b, output) \
421 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
422 #define _invlist_intersection_complement_2nd(a, b, output) \
423 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
425 /* About scan_data_t.
427 During optimisation we recurse through the regexp program performing
428 various inplace (keyhole style) optimisations. In addition study_chunk
429 and scan_commit populate this data structure with information about
430 what strings MUST appear in the pattern. We look for the longest
431 string that must appear at a fixed location, and we look for the
432 longest string that may appear at a floating location. So for instance
437 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
438 strings (because they follow a .* construct). study_chunk will identify
439 both FOO and BAR as being the longest fixed and floating strings respectively.
441 The strings can be composites, for instance
445 will result in a composite fixed substring 'foo'.
447 For each string some basic information is maintained:
450 This is the position the string must appear at, or not before.
451 It also implicitly (when combined with minlenp) tells us how many
452 characters must match before the string we are searching for.
453 Likewise when combined with minlenp and the length of the string it
454 tells us how many characters must appear after the string we have
458 Only used for floating strings. This is the rightmost point that
459 the string can appear at. If set to SSize_t_MAX it indicates that the
460 string can occur infinitely far to the right.
461 For fixed strings, it is equal to min_offset.
464 A pointer to the minimum number of characters of the pattern that the
465 string was found inside. This is important as in the case of positive
466 lookahead or positive lookbehind we can have multiple patterns
471 The minimum length of the pattern overall is 3, the minimum length
472 of the lookahead part is 3, but the minimum length of the part that
473 will actually match is 1. So 'FOO's minimum length is 3, but the
474 minimum length for the F is 1. This is important as the minimum length
475 is used to determine offsets in front of and behind the string being
476 looked for. Since strings can be composites this is the length of the
477 pattern at the time it was committed with a scan_commit. Note that
478 the length is calculated by study_chunk, so that the minimum lengths
479 are not known until the full pattern has been compiled, thus the
480 pointer to the value.
484 In the case of lookbehind the string being searched for can be
485 offset past the start point of the final matching string.
486 If this value was just blithely removed from the min_offset it would
487 invalidate some of the calculations for how many chars must match
488 before or after (as they are derived from min_offset and minlen and
489 the length of the string being searched for).
490 When the final pattern is compiled and the data is moved from the
491 scan_data_t structure into the regexp structure the information
492 about lookbehind is factored in, with the information that would
493 have been lost precalculated in the end_shift field for the
496 The fields pos_min and pos_delta are used to store the minimum offset
497 and the delta to the maximum offset at the current point in the pattern.
501 struct scan_data_substrs {
502 SV *str; /* longest substring found in pattern */
503 SSize_t min_offset; /* earliest point in string it can appear */
504 SSize_t max_offset; /* latest point in string it can appear */
505 SSize_t *minlenp; /* pointer to the minlen relevant to the string */
506 SSize_t lookbehind; /* is the pos of the string modified by LB */
507 I32 flags; /* per substring SF_* and SCF_* flags */
510 typedef struct scan_data_t {
511 /*I32 len_min; unused */
512 /*I32 len_delta; unused */
516 SSize_t last_end; /* min value, <0 unless valid. */
517 SSize_t last_start_min;
518 SSize_t last_start_max;
519 U8 cur_is_floating; /* whether the last_* values should be set as
520 * the next fixed (0) or floating (1)
523 /* [0] is longest fixed substring so far, [1] is longest float so far */
524 struct scan_data_substrs substrs[2];
526 I32 flags; /* common SF_* and SCF_* flags */
528 SSize_t *last_closep;
529 regnode_ssc *start_class;
533 * Forward declarations for pregcomp()'s friends.
536 static const scan_data_t zero_scan_data = {
537 0, 0, NULL, 0, 0, 0, 0,
539 { NULL, 0, 0, 0, 0, 0 },
540 { NULL, 0, 0, 0, 0, 0 },
547 #define SF_BEFORE_SEOL 0x0001
548 #define SF_BEFORE_MEOL 0x0002
549 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
551 #define SF_IS_INF 0x0040
552 #define SF_HAS_PAR 0x0080
553 #define SF_IN_PAR 0x0100
554 #define SF_HAS_EVAL 0x0200
557 /* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
558 * longest substring in the pattern. When it is not set the optimiser keeps
559 * track of position, but does not keep track of the actual strings seen,
561 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
564 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
565 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
566 * turned off because of the alternation (BRANCH). */
567 #define SCF_DO_SUBSTR 0x0400
569 #define SCF_DO_STCLASS_AND 0x0800
570 #define SCF_DO_STCLASS_OR 0x1000
571 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
572 #define SCF_WHILEM_VISITED_POS 0x2000
574 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
575 #define SCF_SEEN_ACCEPT 0x8000
576 #define SCF_TRIE_DOING_RESTUDY 0x10000
577 #define SCF_IN_DEFINE 0x20000
582 #define UTF cBOOL(RExC_utf8)
584 /* The enums for all these are ordered so things work out correctly */
585 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
586 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
587 == REGEX_DEPENDS_CHARSET)
588 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
589 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
590 >= REGEX_UNICODE_CHARSET)
591 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
592 == REGEX_ASCII_RESTRICTED_CHARSET)
593 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
594 >= REGEX_ASCII_RESTRICTED_CHARSET)
595 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
596 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
598 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
600 /* For programs that want to be strictly Unicode compatible by dying if any
601 * attempt is made to match a non-Unicode code point against a Unicode
603 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
605 #define OOB_NAMEDCLASS -1
607 /* There is no code point that is out-of-bounds, so this is problematic. But
608 * its only current use is to initialize a variable that is always set before
610 #define OOB_UNICODE 0xDEADBEEF
612 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
615 /* length of regex to show in messages that don't mark a position within */
616 #define RegexLengthToShowInErrorMessages 127
619 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
620 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
621 * op/pragma/warn/regcomp.
623 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
624 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
626 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
627 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
629 /* The code in this file in places uses one level of recursion with parsing
630 * rebased to an alternate string constructed by us in memory. This can take
631 * the form of something that is completely different from the input, or
632 * something that uses the input as part of the alternate. In the first case,
633 * there should be no possibility of an error, as we are in complete control of
634 * the alternate string. But in the second case we don't completely control
635 * the input portion, so there may be errors in that. Here's an example:
637 * is handled specially because \x{df} folds to a sequence of more than one
638 * character: 'ss'. What is done is to create and parse an alternate string,
639 * which looks like this:
640 * /(?:\x{DF}|[abc\x{DF}def])/ui
641 * where it uses the input unchanged in the middle of something it constructs,
642 * which is a branch for the DF outside the character class, and clustering
643 * parens around the whole thing. (It knows enough to skip the DF inside the
644 * class while in this substitute parse.) 'abc' and 'def' may have errors that
645 * need to be reported. The general situation looks like this:
647 * |<------- identical ------>|
649 * Input: ---------------------------------------------------------------
650 * Constructed: ---------------------------------------------------
652 * |<------- identical ------>|
654 * sI..eI is the portion of the input pattern we are concerned with here.
655 * sC..EC is the constructed substitute parse string.
656 * sC..tC is constructed by us
657 * tC..eC is an exact duplicate of the portion of the input pattern tI..eI.
658 * In the diagram, these are vertically aligned.
659 * eC..EC is also constructed by us.
660 * xC is the position in the substitute parse string where we found a
662 * xI is the position in the original pattern corresponding to xC.
664 * We want to display a message showing the real input string. Thus we need to
665 * translate from xC to xI. We know that xC >= tC, since the portion of the
666 * string sC..tC has been constructed by us, and so shouldn't have errors. We
668 * xI = tI + (xC - tC)
670 * When the substitute parse is constructed, the code needs to set:
673 * RExC_copy_start_in_input (tI)
674 * RExC_copy_start_in_constructed (tC)
675 * and restore them when done.
677 * During normal processing of the input pattern, both
678 * 'RExC_copy_start_in_input' and 'RExC_copy_start_in_constructed' are set to
679 * sI, so that xC equals xI.
682 #define sI RExC_precomp
683 #define eI RExC_precomp_end
684 #define sC RExC_start
686 #define tI RExC_copy_start_in_input
687 #define tC RExC_copy_start_in_constructed
688 #define xI(xC) (tI + (xC - tC))
689 #define xI_offset(xC) (xI(xC) - sI)
691 #define REPORT_LOCATION_ARGS(xC) \
693 (xI(xC) > eI) /* Don't run off end */ \
694 ? eI - sI /* Length before the <--HERE */ \
695 : ((xI_offset(xC) >= 0) \
697 : (Perl_croak(aTHX_ "panic: %s: %d: negative offset: %" \
698 IVdf " trying to output message for " \
700 __FILE__, __LINE__, (IV) xI_offset(xC), \
701 ((int) (eC - sC)), sC), 0)), \
702 sI), /* The input pattern printed up to the <--HERE */ \
704 (xI(xC) > eI) ? 0 : eI - xI(xC), /* Length after <--HERE */ \
705 (xI(xC) > eI) ? eI : xI(xC)) /* pattern after <--HERE */
707 /* Used to point after bad bytes for an error message, but avoid skipping
708 * past a nul byte. */
709 #define SKIP_IF_CHAR(s, e) (!*(s) ? 0 : UTF ? UTF8_SAFE_SKIP(s, e) : 1)
711 /* Set up to clean up after our imminent demise */
712 #define PREPARE_TO_DIE \
715 SAVEFREESV(RExC_rx_sv); \
716 if (RExC_open_parens) \
717 SAVEFREEPV(RExC_open_parens); \
718 if (RExC_close_parens) \
719 SAVEFREEPV(RExC_close_parens); \
723 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
724 * arg. Show regex, up to a maximum length. If it's too long, chop and add
727 #define _FAIL(code) STMT_START { \
728 const char *ellipses = ""; \
729 IV len = RExC_precomp_end - RExC_precomp; \
732 if (len > RegexLengthToShowInErrorMessages) { \
733 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
734 len = RegexLengthToShowInErrorMessages - 10; \
740 #define FAIL(msg) _FAIL( \
741 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
742 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
744 #define FAIL2(msg,arg) _FAIL( \
745 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
746 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
749 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
751 #define Simple_vFAIL(m) STMT_START { \
752 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
753 m, REPORT_LOCATION_ARGS(RExC_parse)); \
757 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
759 #define vFAIL(m) STMT_START { \
765 * Like Simple_vFAIL(), but accepts two arguments.
767 #define Simple_vFAIL2(m,a1) STMT_START { \
768 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
769 REPORT_LOCATION_ARGS(RExC_parse)); \
773 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
775 #define vFAIL2(m,a1) STMT_START { \
777 Simple_vFAIL2(m, a1); \
782 * Like Simple_vFAIL(), but accepts three arguments.
784 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
785 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
786 REPORT_LOCATION_ARGS(RExC_parse)); \
790 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
792 #define vFAIL3(m,a1,a2) STMT_START { \
794 Simple_vFAIL3(m, a1, a2); \
798 * Like Simple_vFAIL(), but accepts four arguments.
800 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
801 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
802 REPORT_LOCATION_ARGS(RExC_parse)); \
805 #define vFAIL4(m,a1,a2,a3) STMT_START { \
807 Simple_vFAIL4(m, a1, a2, a3); \
810 /* A specialized version of vFAIL2 that works with UTF8f */
811 #define vFAIL2utf8f(m, a1) STMT_START { \
813 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
814 REPORT_LOCATION_ARGS(RExC_parse)); \
817 #define vFAIL3utf8f(m, a1, a2) STMT_START { \
819 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
820 REPORT_LOCATION_ARGS(RExC_parse)); \
823 /* Setting this to NULL is a signal to not output warnings */
824 #define TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE RExC_copy_start_in_constructed = NULL
825 #define RESTORE_WARNINGS RExC_copy_start_in_constructed = RExC_precomp
827 /* Since a warning can be generated multiple times as the input is reparsed, we
828 * output it the first time we come to that point in the parse, but suppress it
829 * otherwise. 'RExC_copy_start_in_constructed' being NULL is a flag to not
830 * generate any warnings */
831 #define TO_OUTPUT_WARNINGS(loc) \
832 ( RExC_copy_start_in_constructed \
833 && ((xI(loc)) - RExC_precomp) > (Ptrdiff_t) RExC_latest_warn_offset)
835 /* After we've emitted a warning, we save the position in the input so we don't
837 #define UPDATE_WARNINGS_LOC(loc) \
839 if (TO_OUTPUT_WARNINGS(loc)) { \
840 RExC_latest_warn_offset = (xI(loc)) - RExC_precomp; \
844 /* 'warns' is the output of the packWARNx macro used in 'code' */
845 #define _WARN_HELPER(loc, warns, code) \
847 if (! RExC_copy_start_in_constructed) { \
848 Perl_croak( aTHX_ "panic! %s: %d: Tried to warn when none" \
849 " expected at '%s'", \
850 __FILE__, __LINE__, loc); \
852 if (TO_OUTPUT_WARNINGS(loc)) { \
856 UPDATE_WARNINGS_LOC(loc); \
860 /* m is not necessarily a "literal string", in this macro */
861 #define reg_warn_non_literal_string(loc, m) \
862 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
863 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
864 "%s" REPORT_LOCATION, \
865 m, REPORT_LOCATION_ARGS(loc)))
867 #define ckWARNreg(loc,m) \
868 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
869 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
871 REPORT_LOCATION_ARGS(loc)))
873 #define vWARN(loc, m) \
874 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
875 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
877 REPORT_LOCATION_ARGS(loc))) \
879 #define vWARN_dep(loc, m) \
880 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
881 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
883 REPORT_LOCATION_ARGS(loc)))
885 #define ckWARNdep(loc,m) \
886 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
887 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
889 REPORT_LOCATION_ARGS(loc)))
891 #define ckWARNregdep(loc,m) \
892 _WARN_HELPER(loc, packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
893 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
896 REPORT_LOCATION_ARGS(loc)))
898 #define ckWARN2reg_d(loc,m, a1) \
899 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
900 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
902 a1, REPORT_LOCATION_ARGS(loc)))
904 #define ckWARN2reg(loc, m, a1) \
905 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
906 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
908 a1, REPORT_LOCATION_ARGS(loc)))
910 #define vWARN3(loc, m, a1, a2) \
911 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
912 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
914 a1, a2, REPORT_LOCATION_ARGS(loc)))
916 #define ckWARN3reg(loc, m, a1, a2) \
917 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
918 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
921 REPORT_LOCATION_ARGS(loc)))
923 #define vWARN4(loc, m, a1, a2, a3) \
924 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
925 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
928 REPORT_LOCATION_ARGS(loc)))
930 #define ckWARN4reg(loc, m, a1, a2, a3) \
931 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
932 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
935 REPORT_LOCATION_ARGS(loc)))
937 #define vWARN5(loc, m, a1, a2, a3, a4) \
938 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
939 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
942 REPORT_LOCATION_ARGS(loc)))
944 #define ckWARNexperimental(loc, class, m) \
945 _WARN_HELPER(loc, packWARN(class), \
946 Perl_ck_warner_d(aTHX_ packWARN(class), \
948 REPORT_LOCATION_ARGS(loc)))
950 /* Convert between a pointer to a node and its offset from the beginning of the
952 #define REGNODE_p(offset) (RExC_emit_start + (offset))
953 #define REGNODE_OFFSET(node) ((node) - RExC_emit_start)
955 /* Macros for recording node offsets. 20001227 mjd@plover.com
956 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
957 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
958 * Element 0 holds the number n.
959 * Position is 1 indexed.
961 #ifndef RE_TRACK_PATTERN_OFFSETS
962 #define Set_Node_Offset_To_R(offset,byte)
963 #define Set_Node_Offset(node,byte)
964 #define Set_Cur_Node_Offset
965 #define Set_Node_Length_To_R(node,len)
966 #define Set_Node_Length(node,len)
967 #define Set_Node_Cur_Length(node,start)
968 #define Node_Offset(n)
969 #define Node_Length(n)
970 #define Set_Node_Offset_Length(node,offset,len)
971 #define ProgLen(ri) ri->u.proglen
972 #define SetProgLen(ri,x) ri->u.proglen = x
973 #define Track_Code(code)
975 #define ProgLen(ri) ri->u.offsets[0]
976 #define SetProgLen(ri,x) ri->u.offsets[0] = x
977 #define Set_Node_Offset_To_R(offset,byte) STMT_START { \
978 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
979 __LINE__, (int)(offset), (int)(byte))); \
981 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
984 RExC_offsets[2*(offset)-1] = (byte); \
988 #define Set_Node_Offset(node,byte) \
989 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (byte)-RExC_start)
990 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
992 #define Set_Node_Length_To_R(node,len) STMT_START { \
993 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
994 __LINE__, (int)(node), (int)(len))); \
996 Perl_croak(aTHX_ "value of node is %d in Length macro", \
999 RExC_offsets[2*(node)] = (len); \
1003 #define Set_Node_Length(node,len) \
1004 Set_Node_Length_To_R(REGNODE_OFFSET(node), len)
1005 #define Set_Node_Cur_Length(node, start) \
1006 Set_Node_Length(node, RExC_parse - start)
1008 /* Get offsets and lengths */
1009 #define Node_Offset(n) (RExC_offsets[2*(REGNODE_OFFSET(n))-1])
1010 #define Node_Length(n) (RExC_offsets[2*(REGNODE_OFFSET(n))])
1012 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
1013 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (offset)); \
1014 Set_Node_Length_To_R(REGNODE_OFFSET(node), (len)); \
1017 #define Track_Code(code) STMT_START { code } STMT_END
1020 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
1021 #define EXPERIMENTAL_INPLACESCAN
1022 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
1026 Perl_re_printf(pTHX_ const char *fmt, ...)
1030 PerlIO *f= Perl_debug_log;
1031 PERL_ARGS_ASSERT_RE_PRINTF;
1033 result = PerlIO_vprintf(f, fmt, ap);
1039 Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
1043 PerlIO *f= Perl_debug_log;
1044 PERL_ARGS_ASSERT_RE_INDENTF;
1045 va_start(ap, depth);
1046 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
1047 result = PerlIO_vprintf(f, fmt, ap);
1051 #endif /* DEBUGGING */
1053 #define DEBUG_RExC_seen() \
1054 DEBUG_OPTIMISE_MORE_r({ \
1055 Perl_re_printf( aTHX_ "RExC_seen: "); \
1057 if (RExC_seen & REG_ZERO_LEN_SEEN) \
1058 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
1060 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
1061 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
1063 if (RExC_seen & REG_GPOS_SEEN) \
1064 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
1066 if (RExC_seen & REG_RECURSE_SEEN) \
1067 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
1069 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
1070 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
1072 if (RExC_seen & REG_VERBARG_SEEN) \
1073 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
1075 if (RExC_seen & REG_CUTGROUP_SEEN) \
1076 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
1078 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
1079 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
1081 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
1082 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
1084 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
1085 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
1087 Perl_re_printf( aTHX_ "\n"); \
1090 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
1091 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
1096 S_debug_show_study_flags(pTHX_ U32 flags, const char *open_str,
1097 const char *close_str)
1102 Perl_re_printf( aTHX_ "%s", open_str);
1103 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_SEOL);
1104 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_MEOL);
1105 DEBUG_SHOW_STUDY_FLAG(flags, SF_IS_INF);
1106 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_PAR);
1107 DEBUG_SHOW_STUDY_FLAG(flags, SF_IN_PAR);
1108 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_EVAL);
1109 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_SUBSTR);
1110 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_AND);
1111 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_OR);
1112 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS);
1113 DEBUG_SHOW_STUDY_FLAG(flags, SCF_WHILEM_VISITED_POS);
1114 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_RESTUDY);
1115 DEBUG_SHOW_STUDY_FLAG(flags, SCF_SEEN_ACCEPT);
1116 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_DOING_RESTUDY);
1117 DEBUG_SHOW_STUDY_FLAG(flags, SCF_IN_DEFINE);
1118 Perl_re_printf( aTHX_ "%s", close_str);
1123 S_debug_studydata(pTHX_ const char *where, scan_data_t *data,
1124 U32 depth, int is_inf)
1126 GET_RE_DEBUG_FLAGS_DECL;
1128 DEBUG_OPTIMISE_MORE_r({
1131 Perl_re_indentf(aTHX_ "%s: Pos:%" IVdf "/%" IVdf " Flags: 0x%" UVXf,
1135 (IV)data->pos_delta,
1139 S_debug_show_study_flags(aTHX_ data->flags," [","]");
1141 Perl_re_printf( aTHX_
1142 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s",
1144 (IV)(data->last_closep ? *((data)->last_closep) : -1),
1145 is_inf ? "INF " : ""
1148 if (data->last_found) {
1150 Perl_re_printf(aTHX_
1151 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf,
1152 SvPVX_const(data->last_found),
1154 (IV)data->last_start_min,
1155 (IV)data->last_start_max
1158 for (i = 0; i < 2; i++) {
1159 Perl_re_printf(aTHX_
1160 " %s%s: '%s' @ %" IVdf "/%" IVdf,
1161 data->cur_is_floating == i ? "*" : "",
1162 i ? "Float" : "Fixed",
1163 SvPVX_const(data->substrs[i].str),
1164 (IV)data->substrs[i].min_offset,
1165 (IV)data->substrs[i].max_offset
1167 S_debug_show_study_flags(aTHX_ data->substrs[i].flags," [","]");
1171 Perl_re_printf( aTHX_ "\n");
1177 S_debug_peep(pTHX_ const char *str, const RExC_state_t *pRExC_state,
1178 regnode *scan, U32 depth, U32 flags)
1180 GET_RE_DEBUG_FLAGS_DECL;
1187 Next = regnext(scan);
1188 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
1189 Perl_re_indentf( aTHX_ "%s>%3d: %s (%d)",
1192 REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),
1193 Next ? (REG_NODE_NUM(Next)) : 0 );
1194 S_debug_show_study_flags(aTHX_ flags," [ ","]");
1195 Perl_re_printf( aTHX_ "\n");
1200 # define DEBUG_STUDYDATA(where, data, depth, is_inf) \
1201 S_debug_studydata(aTHX_ where, data, depth, is_inf)
1203 # define DEBUG_PEEP(str, scan, depth, flags) \
1204 S_debug_peep(aTHX_ str, pRExC_state, scan, depth, flags)
1207 # define DEBUG_STUDYDATA(where, data, depth, is_inf) NOOP
1208 # define DEBUG_PEEP(str, scan, depth, flags) NOOP
1212 /* =========================================================
1213 * BEGIN edit_distance stuff.
1215 * This calculates how many single character changes of any type are needed to
1216 * transform a string into another one. It is taken from version 3.1 of
1218 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1221 /* Our unsorted dictionary linked list. */
1222 /* Note we use UVs, not chars. */
1227 struct dictionary* next;
1229 typedef struct dictionary item;
1232 PERL_STATIC_INLINE item*
1233 push(UV key, item* curr)
1236 Newx(head, 1, item);
1244 PERL_STATIC_INLINE item*
1245 find(item* head, UV key)
1247 item* iterator = head;
1249 if (iterator->key == key){
1252 iterator = iterator->next;
1258 PERL_STATIC_INLINE item*
1259 uniquePush(item* head, UV key)
1261 item* iterator = head;
1264 if (iterator->key == key) {
1267 iterator = iterator->next;
1270 return push(key, head);
1273 PERL_STATIC_INLINE void
1274 dict_free(item* head)
1276 item* iterator = head;
1279 item* temp = iterator;
1280 iterator = iterator->next;
1287 /* End of Dictionary Stuff */
1289 /* All calculations/work are done here */
1291 S_edit_distance(const UV* src,
1293 const STRLEN x, /* length of src[] */
1294 const STRLEN y, /* length of tgt[] */
1295 const SSize_t maxDistance
1299 UV swapCount, swapScore, targetCharCount, i, j;
1301 UV score_ceil = x + y;
1303 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1305 /* intialize matrix start values */
1306 Newx(scores, ( (x + 2) * (y + 2)), UV);
1307 scores[0] = score_ceil;
1308 scores[1 * (y + 2) + 0] = score_ceil;
1309 scores[0 * (y + 2) + 1] = score_ceil;
1310 scores[1 * (y + 2) + 1] = 0;
1311 head = uniquePush(uniquePush(head, src[0]), tgt[0]);
1316 for (i=1;i<=x;i++) {
1318 head = uniquePush(head, src[i]);
1319 scores[(i+1) * (y + 2) + 1] = i;
1320 scores[(i+1) * (y + 2) + 0] = score_ceil;
1323 for (j=1;j<=y;j++) {
1326 head = uniquePush(head, tgt[j]);
1327 scores[1 * (y + 2) + (j + 1)] = j;
1328 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1331 targetCharCount = find(head, tgt[j-1])->value;
1332 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1334 if (src[i-1] != tgt[j-1]){
1335 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));
1339 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1343 find(head, src[i-1])->value = i;
1347 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1350 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1354 /* END of edit_distance() stuff
1355 * ========================================================= */
1357 /* is c a control character for which we have a mnemonic? */
1358 #define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
1361 S_cntrl_to_mnemonic(const U8 c)
1363 /* Returns the mnemonic string that represents character 'c', if one
1364 * exists; NULL otherwise. The only ones that exist for the purposes of
1365 * this routine are a few control characters */
1368 case '\a': return "\\a";
1369 case '\b': return "\\b";
1370 case ESC_NATIVE: return "\\e";
1371 case '\f': return "\\f";
1372 case '\n': return "\\n";
1373 case '\r': return "\\r";
1374 case '\t': return "\\t";
1380 /* Mark that we cannot extend a found fixed substring at this point.
1381 Update the longest found anchored substring or the longest found
1382 floating substrings if needed. */
1385 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1386 SSize_t *minlenp, int is_inf)
1388 const STRLEN l = CHR_SVLEN(data->last_found);
1389 SV * const longest_sv = data->substrs[data->cur_is_floating].str;
1390 const STRLEN old_l = CHR_SVLEN(longest_sv);
1391 GET_RE_DEBUG_FLAGS_DECL;
1393 PERL_ARGS_ASSERT_SCAN_COMMIT;
1395 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
1396 const U8 i = data->cur_is_floating;
1397 SvSetMagicSV(longest_sv, data->last_found);
1398 data->substrs[i].min_offset = l ? data->last_start_min : data->pos_min;
1401 data->substrs[0].max_offset = data->substrs[0].min_offset;
1403 data->substrs[1].max_offset = (l
1404 ? data->last_start_max
1405 : (data->pos_delta > SSize_t_MAX - data->pos_min
1407 : data->pos_min + data->pos_delta));
1409 || (STRLEN)data->substrs[1].max_offset > (STRLEN)SSize_t_MAX)
1410 data->substrs[1].max_offset = SSize_t_MAX;
1413 if (data->flags & SF_BEFORE_EOL)
1414 data->substrs[i].flags |= (data->flags & SF_BEFORE_EOL);
1416 data->substrs[i].flags &= ~SF_BEFORE_EOL;
1417 data->substrs[i].minlenp = minlenp;
1418 data->substrs[i].lookbehind = 0;
1421 SvCUR_set(data->last_found, 0);
1423 SV * const sv = data->last_found;
1424 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1425 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1430 data->last_end = -1;
1431 data->flags &= ~SF_BEFORE_EOL;
1432 DEBUG_STUDYDATA("commit", data, 0, is_inf);
1435 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1436 * list that describes which code points it matches */
1439 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1441 /* Set the SSC 'ssc' to match an empty string or any code point */
1443 PERL_ARGS_ASSERT_SSC_ANYTHING;
1445 assert(is_ANYOF_SYNTHETIC(ssc));
1447 /* mortalize so won't leak */
1448 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
1449 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1453 S_ssc_is_anything(const regnode_ssc *ssc)
1455 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1456 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1457 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1458 * in any way, so there's no point in using it */
1463 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1465 assert(is_ANYOF_SYNTHETIC(ssc));
1467 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1471 /* See if the list consists solely of the range 0 - Infinity */
1472 invlist_iterinit(ssc->invlist);
1473 ret = invlist_iternext(ssc->invlist, &start, &end)
1477 invlist_iterfinish(ssc->invlist);
1483 /* If e.g., both \w and \W are set, matches everything */
1484 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1486 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1487 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1497 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1499 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1500 * string, any code point, or any posix class under locale */
1502 PERL_ARGS_ASSERT_SSC_INIT;
1504 Zero(ssc, 1, regnode_ssc);
1505 set_ANYOF_SYNTHETIC(ssc);
1506 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1509 /* If any portion of the regex is to operate under locale rules that aren't
1510 * fully known at compile time, initialization includes it. The reason
1511 * this isn't done for all regexes is that the optimizer was written under
1512 * the assumption that locale was all-or-nothing. Given the complexity and
1513 * lack of documentation in the optimizer, and that there are inadequate
1514 * test cases for locale, many parts of it may not work properly, it is
1515 * safest to avoid locale unless necessary. */
1516 if (RExC_contains_locale) {
1517 ANYOF_POSIXL_SETALL(ssc);
1520 ANYOF_POSIXL_ZERO(ssc);
1525 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1526 const regnode_ssc *ssc)
1528 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1529 * to the list of code points matched, and locale posix classes; hence does
1530 * not check its flags) */
1535 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1537 assert(is_ANYOF_SYNTHETIC(ssc));
1539 invlist_iterinit(ssc->invlist);
1540 ret = invlist_iternext(ssc->invlist, &start, &end)
1544 invlist_iterfinish(ssc->invlist);
1550 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1557 #define INVLIST_INDEX 0
1558 #define ONLY_LOCALE_MATCHES_INDEX 1
1559 #define DEFERRED_USER_DEFINED_INDEX 2
1562 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1563 const regnode_charclass* const node)
1565 /* Returns a mortal inversion list defining which code points are matched
1566 * by 'node', which is of type ANYOF. Handles complementing the result if
1567 * appropriate. If some code points aren't knowable at this time, the
1568 * returned list must, and will, contain every code point that is a
1573 SV* only_utf8_locale_invlist = NULL;
1575 const U32 n = ARG(node);
1576 bool new_node_has_latin1 = FALSE;
1577 const U8 flags = OP(node) == ANYOFH ? 0 : ANYOF_FLAGS(node);
1579 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1581 /* Look at the data structure created by S_set_ANYOF_arg() */
1582 if (n != ANYOF_ONLY_HAS_BITMAP) {
1583 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1584 AV * const av = MUTABLE_AV(SvRV(rv));
1585 SV **const ary = AvARRAY(av);
1586 assert(RExC_rxi->data->what[n] == 's');
1588 if (av_tindex_skip_len_mg(av) >= DEFERRED_USER_DEFINED_INDEX) {
1590 /* Here there are things that won't be known until runtime -- we
1591 * have to assume it could be anything */
1592 invlist = sv_2mortal(_new_invlist(1));
1593 return _add_range_to_invlist(invlist, 0, UV_MAX);
1595 else if (ary[INVLIST_INDEX]) {
1597 /* Use the node's inversion list */
1598 invlist = sv_2mortal(invlist_clone(ary[INVLIST_INDEX], NULL));
1601 /* Get the code points valid only under UTF-8 locales */
1602 if ( (flags & ANYOFL_FOLD)
1603 && av_tindex_skip_len_mg(av) >= ONLY_LOCALE_MATCHES_INDEX)
1605 only_utf8_locale_invlist = ary[ONLY_LOCALE_MATCHES_INDEX];
1610 invlist = sv_2mortal(_new_invlist(0));
1613 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1614 * code points, and an inversion list for the others, but if there are code
1615 * points that should match only conditionally on the target string being
1616 * UTF-8, those are placed in the inversion list, and not the bitmap.
1617 * Since there are circumstances under which they could match, they are
1618 * included in the SSC. But if the ANYOF node is to be inverted, we have
1619 * to exclude them here, so that when we invert below, the end result
1620 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1621 * have to do this here before we add the unconditionally matched code
1623 if (flags & ANYOF_INVERT) {
1624 _invlist_intersection_complement_2nd(invlist,
1629 /* Add in the points from the bit map */
1630 if (OP(node) != ANYOFH) {
1631 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1632 if (ANYOF_BITMAP_TEST(node, i)) {
1633 unsigned int start = i++;
1635 for (; i < NUM_ANYOF_CODE_POINTS
1636 && ANYOF_BITMAP_TEST(node, i); ++i)
1640 invlist = _add_range_to_invlist(invlist, start, i-1);
1641 new_node_has_latin1 = TRUE;
1646 /* If this can match all upper Latin1 code points, have to add them
1647 * as well. But don't add them if inverting, as when that gets done below,
1648 * it would exclude all these characters, including the ones it shouldn't
1649 * that were added just above */
1650 if (! (flags & ANYOF_INVERT) && OP(node) == ANYOFD
1651 && (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1653 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1656 /* Similarly for these */
1657 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1658 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1661 if (flags & ANYOF_INVERT) {
1662 _invlist_invert(invlist);
1664 else if (flags & ANYOFL_FOLD) {
1665 if (new_node_has_latin1) {
1667 /* Under /li, any 0-255 could fold to any other 0-255, depending on
1668 * the locale. We can skip this if there are no 0-255 at all. */
1669 _invlist_union(invlist, PL_Latin1, &invlist);
1671 invlist = add_cp_to_invlist(invlist, LATIN_SMALL_LETTER_DOTLESS_I);
1672 invlist = add_cp_to_invlist(invlist, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
1675 if (_invlist_contains_cp(invlist, LATIN_SMALL_LETTER_DOTLESS_I)) {
1676 invlist = add_cp_to_invlist(invlist, 'I');
1678 if (_invlist_contains_cp(invlist,
1679 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE))
1681 invlist = add_cp_to_invlist(invlist, 'i');
1686 /* Similarly add the UTF-8 locale possible matches. These have to be
1687 * deferred until after the non-UTF-8 locale ones are taken care of just
1688 * above, or it leads to wrong results under ANYOF_INVERT */
1689 if (only_utf8_locale_invlist) {
1690 _invlist_union_maybe_complement_2nd(invlist,
1691 only_utf8_locale_invlist,
1692 flags & ANYOF_INVERT,
1699 /* These two functions currently do the exact same thing */
1700 #define ssc_init_zero ssc_init
1702 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1703 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1705 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1706 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1707 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1710 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1711 const regnode_charclass *and_with)
1713 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1714 * another SSC or a regular ANYOF class. Can create false positives. */
1717 U8 and_with_flags = (OP(and_with) == ANYOFH) ? 0 : ANYOF_FLAGS(and_with);
1720 PERL_ARGS_ASSERT_SSC_AND;
1722 assert(is_ANYOF_SYNTHETIC(ssc));
1724 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1725 * the code point inversion list and just the relevant flags */
1726 if (is_ANYOF_SYNTHETIC(and_with)) {
1727 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1728 anded_flags = and_with_flags;
1730 /* XXX This is a kludge around what appears to be deficiencies in the
1731 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1732 * there are paths through the optimizer where it doesn't get weeded
1733 * out when it should. And if we don't make some extra provision for
1734 * it like the code just below, it doesn't get added when it should.
1735 * This solution is to add it only when AND'ing, which is here, and
1736 * only when what is being AND'ed is the pristine, original node
1737 * matching anything. Thus it is like adding it to ssc_anything() but
1738 * only when the result is to be AND'ed. Probably the same solution
1739 * could be adopted for the same problem we have with /l matching,
1740 * which is solved differently in S_ssc_init(), and that would lead to
1741 * fewer false positives than that solution has. But if this solution
1742 * creates bugs, the consequences are only that a warning isn't raised
1743 * that should be; while the consequences for having /l bugs is
1744 * incorrect matches */
1745 if (ssc_is_anything((regnode_ssc *)and_with)) {
1746 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1750 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1751 if (OP(and_with) == ANYOFD) {
1752 anded_flags = and_with_flags & ANYOF_COMMON_FLAGS;
1755 anded_flags = and_with_flags
1756 &( ANYOF_COMMON_FLAGS
1757 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1758 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1759 if (ANYOFL_UTF8_LOCALE_REQD(and_with_flags)) {
1761 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1766 ANYOF_FLAGS(ssc) &= anded_flags;
1768 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1769 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1770 * 'and_with' may be inverted. When not inverted, we have the situation of
1772 * (C1 | P1) & (C2 | P2)
1773 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1774 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1775 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1776 * <= ((C1 & C2) | P1 | P2)
1777 * Alternatively, the last few steps could be:
1778 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1779 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1780 * <= (C1 | C2 | (P1 & P2))
1781 * We favor the second approach if either P1 or P2 is non-empty. This is
1782 * because these components are a barrier to doing optimizations, as what
1783 * they match cannot be known until the moment of matching as they are
1784 * dependent on the current locale, 'AND"ing them likely will reduce or
1786 * But we can do better if we know that C1,P1 are in their initial state (a
1787 * frequent occurrence), each matching everything:
1788 * (<everything>) & (C2 | P2) = C2 | P2
1789 * Similarly, if C2,P2 are in their initial state (again a frequent
1790 * occurrence), the result is a no-op
1791 * (C1 | P1) & (<everything>) = C1 | P1
1794 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1795 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1796 * <= (C1 & ~C2) | (P1 & ~P2)
1799 if ((and_with_flags & ANYOF_INVERT)
1800 && ! is_ANYOF_SYNTHETIC(and_with))
1804 ssc_intersection(ssc,
1806 FALSE /* Has already been inverted */
1809 /* If either P1 or P2 is empty, the intersection will be also; can skip
1811 if (! (and_with_flags & ANYOF_MATCHES_POSIXL)) {
1812 ANYOF_POSIXL_ZERO(ssc);
1814 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1816 /* Note that the Posix class component P from 'and_with' actually
1818 * P = Pa | Pb | ... | Pn
1819 * where each component is one posix class, such as in [\w\s].
1821 * ~P = ~(Pa | Pb | ... | Pn)
1822 * = ~Pa & ~Pb & ... & ~Pn
1823 * <= ~Pa | ~Pb | ... | ~Pn
1824 * The last is something we can easily calculate, but unfortunately
1825 * is likely to have many false positives. We could do better
1826 * in some (but certainly not all) instances if two classes in
1827 * P have known relationships. For example
1828 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1830 * :lower: & :print: = :lower:
1831 * And similarly for classes that must be disjoint. For example,
1832 * since \s and \w can have no elements in common based on rules in
1833 * the POSIX standard,
1834 * \w & ^\S = nothing
1835 * Unfortunately, some vendor locales do not meet the Posix
1836 * standard, in particular almost everything by Microsoft.
1837 * The loop below just changes e.g., \w into \W and vice versa */
1839 regnode_charclass_posixl temp;
1840 int add = 1; /* To calculate the index of the complement */
1842 Zero(&temp, 1, regnode_charclass_posixl);
1843 ANYOF_POSIXL_ZERO(&temp);
1844 for (i = 0; i < ANYOF_MAX; i++) {
1846 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1847 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1849 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1850 ANYOF_POSIXL_SET(&temp, i + add);
1852 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1854 ANYOF_POSIXL_AND(&temp, ssc);
1856 } /* else ssc already has no posixes */
1857 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1858 in its initial state */
1859 else if (! is_ANYOF_SYNTHETIC(and_with)
1860 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1862 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1863 * copy it over 'ssc' */
1864 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1865 if (is_ANYOF_SYNTHETIC(and_with)) {
1866 StructCopy(and_with, ssc, regnode_ssc);
1869 ssc->invlist = anded_cp_list;
1870 ANYOF_POSIXL_ZERO(ssc);
1871 if (and_with_flags & ANYOF_MATCHES_POSIXL) {
1872 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1876 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1877 || (and_with_flags & ANYOF_MATCHES_POSIXL))
1879 /* One or the other of P1, P2 is non-empty. */
1880 if (and_with_flags & ANYOF_MATCHES_POSIXL) {
1881 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1883 ssc_union(ssc, anded_cp_list, FALSE);
1885 else { /* P1 = P2 = empty */
1886 ssc_intersection(ssc, anded_cp_list, FALSE);
1892 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1893 const regnode_charclass *or_with)
1895 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1896 * another SSC or a regular ANYOF class. Can create false positives if
1897 * 'or_with' is to be inverted. */
1901 U8 or_with_flags = (OP(or_with) == ANYOFH) ? 0 : ANYOF_FLAGS(or_with);
1903 PERL_ARGS_ASSERT_SSC_OR;
1905 assert(is_ANYOF_SYNTHETIC(ssc));
1907 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1908 * the code point inversion list and just the relevant flags */
1909 if (is_ANYOF_SYNTHETIC(or_with)) {
1910 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1911 ored_flags = or_with_flags;
1914 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1915 ored_flags = or_with_flags & ANYOF_COMMON_FLAGS;
1916 if (OP(or_with) != ANYOFD) {
1919 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1920 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1921 if (ANYOFL_UTF8_LOCALE_REQD(or_with_flags)) {
1923 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1928 ANYOF_FLAGS(ssc) |= ored_flags;
1930 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1931 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1932 * 'or_with' may be inverted. When not inverted, we have the simple
1933 * situation of computing:
1934 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1935 * If P1|P2 yields a situation with both a class and its complement are
1936 * set, like having both \w and \W, this matches all code points, and we
1937 * can delete these from the P component of the ssc going forward. XXX We
1938 * might be able to delete all the P components, but I (khw) am not certain
1939 * about this, and it is better to be safe.
1942 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1943 * <= (C1 | P1) | ~C2
1944 * <= (C1 | ~C2) | P1
1945 * (which results in actually simpler code than the non-inverted case)
1948 if ((or_with_flags & ANYOF_INVERT)
1949 && ! is_ANYOF_SYNTHETIC(or_with))
1951 /* We ignore P2, leaving P1 going forward */
1952 } /* else Not inverted */
1953 else if (or_with_flags & ANYOF_MATCHES_POSIXL) {
1954 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1955 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1957 for (i = 0; i < ANYOF_MAX; i += 2) {
1958 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1960 ssc_match_all_cp(ssc);
1961 ANYOF_POSIXL_CLEAR(ssc, i);
1962 ANYOF_POSIXL_CLEAR(ssc, i+1);
1970 FALSE /* Already has been inverted */
1974 PERL_STATIC_INLINE void
1975 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1977 PERL_ARGS_ASSERT_SSC_UNION;
1979 assert(is_ANYOF_SYNTHETIC(ssc));
1981 _invlist_union_maybe_complement_2nd(ssc->invlist,
1987 PERL_STATIC_INLINE void
1988 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1990 const bool invert2nd)
1992 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1994 assert(is_ANYOF_SYNTHETIC(ssc));
1996 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
2002 PERL_STATIC_INLINE void
2003 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
2005 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
2007 assert(is_ANYOF_SYNTHETIC(ssc));
2009 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
2012 PERL_STATIC_INLINE void
2013 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
2015 /* AND just the single code point 'cp' into the SSC 'ssc' */
2017 SV* cp_list = _new_invlist(2);
2019 PERL_ARGS_ASSERT_SSC_CP_AND;
2021 assert(is_ANYOF_SYNTHETIC(ssc));
2023 cp_list = add_cp_to_invlist(cp_list, cp);
2024 ssc_intersection(ssc, cp_list,
2025 FALSE /* Not inverted */
2027 SvREFCNT_dec_NN(cp_list);
2030 PERL_STATIC_INLINE void
2031 S_ssc_clear_locale(regnode_ssc *ssc)
2033 /* Set the SSC 'ssc' to not match any locale things */
2034 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
2036 assert(is_ANYOF_SYNTHETIC(ssc));
2038 ANYOF_POSIXL_ZERO(ssc);
2039 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
2042 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
2045 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
2047 /* The synthetic start class is used to hopefully quickly winnow down
2048 * places where a pattern could start a match in the target string. If it
2049 * doesn't really narrow things down that much, there isn't much point to
2050 * having the overhead of using it. This function uses some very crude
2051 * heuristics to decide if to use the ssc or not.
2053 * It returns TRUE if 'ssc' rules out more than half what it considers to
2054 * be the "likely" possible matches, but of course it doesn't know what the
2055 * actual things being matched are going to be; these are only guesses
2057 * For /l matches, it assumes that the only likely matches are going to be
2058 * in the 0-255 range, uniformly distributed, so half of that is 127
2059 * For /a and /d matches, it assumes that the likely matches will be just
2060 * the ASCII range, so half of that is 63
2061 * For /u and there isn't anything matching above the Latin1 range, it
2062 * assumes that that is the only range likely to be matched, and uses
2063 * half that as the cut-off: 127. If anything matches above Latin1,
2064 * it assumes that all of Unicode could match (uniformly), except for
2065 * non-Unicode code points and things in the General Category "Other"
2066 * (unassigned, private use, surrogates, controls and formats). This
2067 * is a much large number. */
2069 U32 count = 0; /* Running total of number of code points matched by
2071 UV start, end; /* Start and end points of current range in inversion
2072 XXX outdated. UTF-8 locales are common, what about invert? list */
2073 const U32 max_code_points = (LOC)
2075 : (( ! UNI_SEMANTICS
2076 || invlist_highest(ssc->invlist) < 256)
2079 const U32 max_match = max_code_points / 2;
2081 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
2083 invlist_iterinit(ssc->invlist);
2084 while (invlist_iternext(ssc->invlist, &start, &end)) {
2085 if (start >= max_code_points) {
2088 end = MIN(end, max_code_points - 1);
2089 count += end - start + 1;
2090 if (count >= max_match) {
2091 invlist_iterfinish(ssc->invlist);
2101 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
2103 /* The inversion list in the SSC is marked mortal; now we need a more
2104 * permanent copy, which is stored the same way that is done in a regular
2105 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
2108 SV* invlist = invlist_clone(ssc->invlist, NULL);
2110 PERL_ARGS_ASSERT_SSC_FINALIZE;
2112 assert(is_ANYOF_SYNTHETIC(ssc));
2114 /* The code in this file assumes that all but these flags aren't relevant
2115 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
2116 * by the time we reach here */
2117 assert(! (ANYOF_FLAGS(ssc)
2118 & ~( ANYOF_COMMON_FLAGS
2119 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
2120 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
2122 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
2124 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist, NULL, NULL);
2126 /* Make sure is clone-safe */
2127 ssc->invlist = NULL;
2129 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
2130 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
2131 OP(ssc) = ANYOFPOSIXL;
2133 else if (RExC_contains_locale) {
2137 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
2140 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
2141 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
2142 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
2143 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
2144 ? (TRIE_LIST_CUR( idx ) - 1) \
2150 dump_trie(trie,widecharmap,revcharmap)
2151 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
2152 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
2154 These routines dump out a trie in a somewhat readable format.
2155 The _interim_ variants are used for debugging the interim
2156 tables that are used to generate the final compressed
2157 representation which is what dump_trie expects.
2159 Part of the reason for their existence is to provide a form
2160 of documentation as to how the different representations function.
2165 Dumps the final compressed table form of the trie to Perl_debug_log.
2166 Used for debugging make_trie().
2170 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
2171 AV *revcharmap, U32 depth)
2174 SV *sv=sv_newmortal();
2175 int colwidth= widecharmap ? 6 : 4;
2177 GET_RE_DEBUG_FLAGS_DECL;
2179 PERL_ARGS_ASSERT_DUMP_TRIE;
2181 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
2182 depth+1, "Match","Base","Ofs" );
2184 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2185 SV ** const tmp = av_fetch( revcharmap, state, 0);
2187 Perl_re_printf( aTHX_ "%*s",
2189 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2190 PL_colors[0], PL_colors[1],
2191 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2192 PERL_PV_ESCAPE_FIRSTCHAR
2197 Perl_re_printf( aTHX_ "\n");
2198 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
2200 for( state = 0 ; state < trie->uniquecharcount ; state++ )
2201 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2202 Perl_re_printf( aTHX_ "\n");
2204 for( state = 1 ; state < trie->statecount ; state++ ) {
2205 const U32 base = trie->states[ state ].trans.base;
2207 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
2209 if ( trie->states[ state ].wordnum ) {
2210 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
2212 Perl_re_printf( aTHX_ "%6s", "" );
2215 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
2220 while( ( base + ofs < trie->uniquecharcount ) ||
2221 ( base + ofs - trie->uniquecharcount < trie->lasttrans
2222 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2226 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
2228 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2229 if ( ( base + ofs >= trie->uniquecharcount )
2230 && ( base + ofs - trie->uniquecharcount
2232 && trie->trans[ base + ofs
2233 - trie->uniquecharcount ].check == state )
2235 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
2236 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2239 Perl_re_printf( aTHX_ "%*s", colwidth," ." );
2243 Perl_re_printf( aTHX_ "]");
2246 Perl_re_printf( aTHX_ "\n" );
2248 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
2250 for (word=1; word <= trie->wordcount; word++) {
2251 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2252 (int)word, (int)(trie->wordinfo[word].prev),
2253 (int)(trie->wordinfo[word].len));
2255 Perl_re_printf( aTHX_ "\n" );
2258 Dumps a fully constructed but uncompressed trie in list form.
2259 List tries normally only are used for construction when the number of
2260 possible chars (trie->uniquecharcount) is very high.
2261 Used for debugging make_trie().
2264 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2265 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2269 SV *sv=sv_newmortal();
2270 int colwidth= widecharmap ? 6 : 4;
2271 GET_RE_DEBUG_FLAGS_DECL;
2273 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2275 /* print out the table precompression. */
2276 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
2278 Perl_re_indentf( aTHX_ "%s",
2279 depth+1, "------:-----+-----------------\n" );
2281 for( state=1 ; state < next_alloc ; state ++ ) {
2284 Perl_re_indentf( aTHX_ " %4" UVXf " :",
2285 depth+1, (UV)state );
2286 if ( ! trie->states[ state ].wordnum ) {
2287 Perl_re_printf( aTHX_ "%5s| ","");
2289 Perl_re_printf( aTHX_ "W%4x| ",
2290 trie->states[ state ].wordnum
2293 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2294 SV ** const tmp = av_fetch( revcharmap,
2295 TRIE_LIST_ITEM(state, charid).forid, 0);
2297 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
2299 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2301 PL_colors[0], PL_colors[1],
2302 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2303 | PERL_PV_ESCAPE_FIRSTCHAR
2305 TRIE_LIST_ITEM(state, charid).forid,
2306 (UV)TRIE_LIST_ITEM(state, charid).newstate
2309 Perl_re_printf( aTHX_ "\n%*s| ",
2310 (int)((depth * 2) + 14), "");
2313 Perl_re_printf( aTHX_ "\n");
2318 Dumps a fully constructed but uncompressed trie in table form.
2319 This is the normal DFA style state transition table, with a few
2320 twists to facilitate compression later.
2321 Used for debugging make_trie().
2324 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2325 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2330 SV *sv=sv_newmortal();
2331 int colwidth= widecharmap ? 6 : 4;
2332 GET_RE_DEBUG_FLAGS_DECL;
2334 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2337 print out the table precompression so that we can do a visual check
2338 that they are identical.
2341 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
2343 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2344 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2346 Perl_re_printf( aTHX_ "%*s",
2348 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2349 PL_colors[0], PL_colors[1],
2350 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2351 PERL_PV_ESCAPE_FIRSTCHAR
2357 Perl_re_printf( aTHX_ "\n");
2358 Perl_re_indentf( aTHX_ "State+-", depth+1 );
2360 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2361 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
2364 Perl_re_printf( aTHX_ "\n" );
2366 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2368 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
2370 (UV)TRIE_NODENUM( state ) );
2372 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2373 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2375 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
2377 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
2379 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2380 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
2381 (UV)trie->trans[ state ].check );
2383 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
2384 (UV)trie->trans[ state ].check,
2385 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2393 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2394 startbranch: the first branch in the whole branch sequence
2395 first : start branch of sequence of branch-exact nodes.
2396 May be the same as startbranch
2397 last : Thing following the last branch.
2398 May be the same as tail.
2399 tail : item following the branch sequence
2400 count : words in the sequence
2401 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2402 depth : indent depth
2404 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2406 A trie is an N'ary tree where the branches are determined by digital
2407 decomposition of the key. IE, at the root node you look up the 1st character and
2408 follow that branch repeat until you find the end of the branches. Nodes can be
2409 marked as "accepting" meaning they represent a complete word. Eg:
2413 would convert into the following structure. Numbers represent states, letters
2414 following numbers represent valid transitions on the letter from that state, if
2415 the number is in square brackets it represents an accepting state, otherwise it
2416 will be in parenthesis.
2418 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2422 (1) +-i->(6)-+-s->[7]
2424 +-s->(3)-+-h->(4)-+-e->[5]
2426 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2428 This shows that when matching against the string 'hers' we will begin at state 1
2429 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2430 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2431 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2432 single traverse. We store a mapping from accepting to state to which word was
2433 matched, and then when we have multiple possibilities we try to complete the
2434 rest of the regex in the order in which they occurred in the alternation.
2436 The only prior NFA like behaviour that would be changed by the TRIE support is
2437 the silent ignoring of duplicate alternations which are of the form:
2439 / (DUPE|DUPE) X? (?{ ... }) Y /x
2441 Thus EVAL blocks following a trie may be called a different number of times with
2442 and without the optimisation. With the optimisations dupes will be silently
2443 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2444 the following demonstrates:
2446 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2448 which prints out 'word' three times, but
2450 'words'=~/(word|word|word)(?{ print $1 })S/
2452 which doesnt print it out at all. This is due to other optimisations kicking in.
2454 Example of what happens on a structural level:
2456 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2458 1: CURLYM[1] {1,32767}(18)
2469 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2470 and should turn into:
2472 1: CURLYM[1] {1,32767}(18)
2474 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2482 Cases where tail != last would be like /(?foo|bar)baz/:
2492 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2493 and would end up looking like:
2496 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2503 d = uvchr_to_utf8_flags(d, uv, 0);
2505 is the recommended Unicode-aware way of saying
2510 #define TRIE_STORE_REVCHAR(val) \
2513 SV *zlopp = newSV(UTF8_MAXBYTES); \
2514 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2515 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2516 SvCUR_set(zlopp, kapow - flrbbbbb); \
2519 av_push(revcharmap, zlopp); \
2521 char ooooff = (char)val; \
2522 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2526 /* This gets the next character from the input, folding it if not already
2528 #define TRIE_READ_CHAR STMT_START { \
2531 /* if it is UTF then it is either already folded, or does not need \
2533 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2535 else if (folder == PL_fold_latin1) { \
2536 /* This folder implies Unicode rules, which in the range expressible \
2537 * by not UTF is the lower case, with the two exceptions, one of \
2538 * which should have been taken care of before calling this */ \
2539 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2540 uvc = toLOWER_L1(*uc); \
2541 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2544 /* raw data, will be folded later if needed */ \
2552 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2553 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2554 U32 ging = TRIE_LIST_LEN( state ) * 2; \
2555 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2556 TRIE_LIST_LEN( state ) = ging; \
2558 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2559 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2560 TRIE_LIST_CUR( state )++; \
2563 #define TRIE_LIST_NEW(state) STMT_START { \
2564 Newx( trie->states[ state ].trans.list, \
2565 4, reg_trie_trans_le ); \
2566 TRIE_LIST_CUR( state ) = 1; \
2567 TRIE_LIST_LEN( state ) = 4; \
2570 #define TRIE_HANDLE_WORD(state) STMT_START { \
2571 U16 dupe= trie->states[ state ].wordnum; \
2572 regnode * const noper_next = regnext( noper ); \
2575 /* store the word for dumping */ \
2577 if (OP(noper) != NOTHING) \
2578 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2580 tmp = newSVpvn_utf8( "", 0, UTF ); \
2581 av_push( trie_words, tmp ); \
2585 trie->wordinfo[curword].prev = 0; \
2586 trie->wordinfo[curword].len = wordlen; \
2587 trie->wordinfo[curword].accept = state; \
2589 if ( noper_next < tail ) { \
2591 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2593 trie->jump[curword] = (U16)(noper_next - convert); \
2595 jumper = noper_next; \
2597 nextbranch= regnext(cur); \
2601 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2602 /* chain, so that when the bits of chain are later */\
2603 /* linked together, the dups appear in the chain */\
2604 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2605 trie->wordinfo[dupe].prev = curword; \
2607 /* we haven't inserted this word yet. */ \
2608 trie->states[ state ].wordnum = curword; \
2613 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2614 ( ( base + charid >= ucharcount \
2615 && base + charid < ubound \
2616 && state == trie->trans[ base - ucharcount + charid ].check \
2617 && trie->trans[ base - ucharcount + charid ].next ) \
2618 ? trie->trans[ base - ucharcount + charid ].next \
2619 : ( state==1 ? special : 0 ) \
2622 #define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2624 TRIE_BITMAP_SET(trie, uvc); \
2625 /* store the folded codepoint */ \
2627 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2630 /* store first byte of utf8 representation of */ \
2631 /* variant codepoints */ \
2632 if (! UVCHR_IS_INVARIANT(uvc)) { \
2633 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2638 #define MADE_JUMP_TRIE 2
2639 #define MADE_EXACT_TRIE 4
2642 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2643 regnode *first, regnode *last, regnode *tail,
2644 U32 word_count, U32 flags, U32 depth)
2646 /* first pass, loop through and scan words */
2647 reg_trie_data *trie;
2648 HV *widecharmap = NULL;
2649 AV *revcharmap = newAV();
2655 regnode *jumper = NULL;
2656 regnode *nextbranch = NULL;
2657 regnode *convert = NULL;
2658 U32 *prev_states; /* temp array mapping each state to previous one */
2659 /* we just use folder as a flag in utf8 */
2660 const U8 * folder = NULL;
2662 /* in the below add_data call we are storing either 'tu' or 'tuaa'
2663 * which stands for one trie structure, one hash, optionally followed
2666 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuaa"));
2667 AV *trie_words = NULL;
2668 /* along with revcharmap, this only used during construction but both are
2669 * useful during debugging so we store them in the struct when debugging.
2672 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2673 STRLEN trie_charcount=0;
2675 SV *re_trie_maxbuff;
2676 GET_RE_DEBUG_FLAGS_DECL;
2678 PERL_ARGS_ASSERT_MAKE_TRIE;
2680 PERL_UNUSED_ARG(depth);
2684 case EXACT: case EXACT_ONLY8: case EXACTL: break;
2688 case EXACTFLU8: folder = PL_fold_latin1; break;
2689 case EXACTF: folder = PL_fold; break;
2690 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2693 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2695 trie->startstate = 1;
2696 trie->wordcount = word_count;
2697 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2698 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2699 if (flags == EXACT || flags == EXACT_ONLY8 || flags == EXACTL)
2700 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2701 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2702 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2705 trie_words = newAV();
2708 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, GV_ADD);
2709 assert(re_trie_maxbuff);
2710 if (!SvIOK(re_trie_maxbuff)) {
2711 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2713 DEBUG_TRIE_COMPILE_r({
2714 Perl_re_indentf( aTHX_
2715 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2717 REG_NODE_NUM(startbranch), REG_NODE_NUM(first),
2718 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2721 /* Find the node we are going to overwrite */
2722 if ( first == startbranch && OP( last ) != BRANCH ) {
2723 /* whole branch chain */
2726 /* branch sub-chain */
2727 convert = NEXTOPER( first );
2730 /* -- First loop and Setup --
2732 We first traverse the branches and scan each word to determine if it
2733 contains widechars, and how many unique chars there are, this is
2734 important as we have to build a table with at least as many columns as we
2737 We use an array of integers to represent the character codes 0..255
2738 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2739 the native representation of the character value as the key and IV's for
2742 *TODO* If we keep track of how many times each character is used we can
2743 remap the columns so that the table compression later on is more
2744 efficient in terms of memory by ensuring the most common value is in the
2745 middle and the least common are on the outside. IMO this would be better
2746 than a most to least common mapping as theres a decent chance the most
2747 common letter will share a node with the least common, meaning the node
2748 will not be compressible. With a middle is most common approach the worst
2749 case is when we have the least common nodes twice.
2753 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2754 regnode *noper = NEXTOPER( cur );
2758 U32 wordlen = 0; /* required init */
2759 STRLEN minchars = 0;
2760 STRLEN maxchars = 0;
2761 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2764 if (OP(noper) == NOTHING) {
2765 /* skip past a NOTHING at the start of an alternation
2766 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2768 regnode *noper_next= regnext(noper);
2769 if (noper_next < tail)
2774 && ( OP(noper) == flags
2775 || (flags == EXACT && OP(noper) == EXACT_ONLY8)
2776 || (flags == EXACTFU && ( OP(noper) == EXACTFU_ONLY8
2777 || OP(noper) == EXACTFUP))))
2779 uc= (U8*)STRING(noper);
2780 e= uc + STR_LEN(noper);
2787 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2788 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2789 regardless of encoding */
2790 if (OP( noper ) == EXACTFUP) {
2791 /* false positives are ok, so just set this */
2792 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2796 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2798 TRIE_CHARCOUNT(trie)++;
2801 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2802 * is in effect. Under /i, this character can match itself, or
2803 * anything that folds to it. If not under /i, it can match just
2804 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2805 * all fold to k, and all are single characters. But some folds
2806 * expand to more than one character, so for example LATIN SMALL
2807 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2808 * the string beginning at 'uc' is 'ffi', it could be matched by
2809 * three characters, or just by the one ligature character. (It
2810 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2811 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2812 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2813 * match.) The trie needs to know the minimum and maximum number
2814 * of characters that could match so that it can use size alone to
2815 * quickly reject many match attempts. The max is simple: it is
2816 * the number of folded characters in this branch (since a fold is
2817 * never shorter than what folds to it. */
2821 /* And the min is equal to the max if not under /i (indicated by
2822 * 'folder' being NULL), or there are no multi-character folds. If
2823 * there is a multi-character fold, the min is incremented just
2824 * once, for the character that folds to the sequence. Each
2825 * character in the sequence needs to be added to the list below of
2826 * characters in the trie, but we count only the first towards the
2827 * min number of characters needed. This is done through the
2828 * variable 'foldlen', which is returned by the macros that look
2829 * for these sequences as the number of bytes the sequence
2830 * occupies. Each time through the loop, we decrement 'foldlen' by
2831 * how many bytes the current char occupies. Only when it reaches
2832 * 0 do we increment 'minchars' or look for another multi-character
2834 if (folder == NULL) {
2837 else if (foldlen > 0) {
2838 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2843 /* See if *uc is the beginning of a multi-character fold. If
2844 * so, we decrement the length remaining to look at, to account
2845 * for the current character this iteration. (We can use 'uc'
2846 * instead of the fold returned by TRIE_READ_CHAR because for
2847 * non-UTF, the latin1_safe macro is smart enough to account
2848 * for all the unfolded characters, and because for UTF, the
2849 * string will already have been folded earlier in the
2850 * compilation process */
2852 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2853 foldlen -= UTF8SKIP(uc);
2856 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2861 /* The current character (and any potential folds) should be added
2862 * to the possible matching characters for this position in this
2866 U8 folded= folder[ (U8) uvc ];
2867 if ( !trie->charmap[ folded ] ) {
2868 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2869 TRIE_STORE_REVCHAR( folded );
2872 if ( !trie->charmap[ uvc ] ) {
2873 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2874 TRIE_STORE_REVCHAR( uvc );
2877 /* store the codepoint in the bitmap, and its folded
2879 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
2880 set_bit = 0; /* We've done our bit :-) */
2884 /* XXX We could come up with the list of code points that fold
2885 * to this using PL_utf8_foldclosures, except not for
2886 * multi-char folds, as there may be multiple combinations
2887 * there that could work, which needs to wait until runtime to
2888 * resolve (The comment about LIGATURE FFI above is such an
2893 widecharmap = newHV();
2895 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2898 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
2900 if ( !SvTRUE( *svpp ) ) {
2901 sv_setiv( *svpp, ++trie->uniquecharcount );
2902 TRIE_STORE_REVCHAR(uvc);
2905 } /* end loop through characters in this branch of the trie */
2907 /* We take the min and max for this branch and combine to find the min
2908 * and max for all branches processed so far */
2909 if( cur == first ) {
2910 trie->minlen = minchars;
2911 trie->maxlen = maxchars;
2912 } else if (minchars < trie->minlen) {
2913 trie->minlen = minchars;
2914 } else if (maxchars > trie->maxlen) {
2915 trie->maxlen = maxchars;
2917 } /* end first pass */
2918 DEBUG_TRIE_COMPILE_r(
2919 Perl_re_indentf( aTHX_
2920 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2922 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2923 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2924 (int)trie->minlen, (int)trie->maxlen )
2928 We now know what we are dealing with in terms of unique chars and
2929 string sizes so we can calculate how much memory a naive
2930 representation using a flat table will take. If it's over a reasonable
2931 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2932 conservative but potentially much slower representation using an array
2935 At the end we convert both representations into the same compressed
2936 form that will be used in regexec.c for matching with. The latter
2937 is a form that cannot be used to construct with but has memory
2938 properties similar to the list form and access properties similar
2939 to the table form making it both suitable for fast searches and
2940 small enough that its feasable to store for the duration of a program.
2942 See the comment in the code where the compressed table is produced
2943 inplace from the flat tabe representation for an explanation of how
2944 the compression works.
2949 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2952 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2953 > SvIV(re_trie_maxbuff) )
2956 Second Pass -- Array Of Lists Representation
2958 Each state will be represented by a list of charid:state records
2959 (reg_trie_trans_le) the first such element holds the CUR and LEN
2960 points of the allocated array. (See defines above).
2962 We build the initial structure using the lists, and then convert
2963 it into the compressed table form which allows faster lookups
2964 (but cant be modified once converted).
2967 STRLEN transcount = 1;
2969 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
2972 trie->states = (reg_trie_state *)
2973 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2974 sizeof(reg_trie_state) );
2978 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2980 regnode *noper = NEXTOPER( cur );
2981 U32 state = 1; /* required init */
2982 U16 charid = 0; /* sanity init */
2983 U32 wordlen = 0; /* required init */
2985 if (OP(noper) == NOTHING) {
2986 regnode *noper_next= regnext(noper);
2987 if (noper_next < tail)
2992 && ( OP(noper) == flags
2993 || (flags == EXACT && OP(noper) == EXACT_ONLY8)
2994 || (flags == EXACTFU && ( OP(noper) == EXACTFU_ONLY8
2995 || OP(noper) == EXACTFUP))))
2997 const U8 *uc= (U8*)STRING(noper);
2998 const U8 *e= uc + STR_LEN(noper);
3000 for ( ; uc < e ; uc += len ) {
3005 charid = trie->charmap[ uvc ];
3007 SV** const svpp = hv_fetch( widecharmap,
3014 charid=(U16)SvIV( *svpp );
3017 /* charid is now 0 if we dont know the char read, or
3018 * nonzero if we do */
3025 if ( !trie->states[ state ].trans.list ) {
3026 TRIE_LIST_NEW( state );
3029 check <= TRIE_LIST_USED( state );
3032 if ( TRIE_LIST_ITEM( state, check ).forid
3035 newstate = TRIE_LIST_ITEM( state, check ).newstate;
3040 newstate = next_alloc++;
3041 prev_states[newstate] = state;
3042 TRIE_LIST_PUSH( state, charid, newstate );
3047 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3051 TRIE_HANDLE_WORD(state);
3053 } /* end second pass */
3055 /* next alloc is the NEXT state to be allocated */
3056 trie->statecount = next_alloc;
3057 trie->states = (reg_trie_state *)
3058 PerlMemShared_realloc( trie->states,
3060 * sizeof(reg_trie_state) );
3062 /* and now dump it out before we compress it */
3063 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
3064 revcharmap, next_alloc,
3068 trie->trans = (reg_trie_trans *)
3069 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
3076 for( state=1 ; state < next_alloc ; state ++ ) {
3080 DEBUG_TRIE_COMPILE_MORE_r(
3081 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
3085 if (trie->states[state].trans.list) {
3086 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
3090 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3091 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
3092 if ( forid < minid ) {
3094 } else if ( forid > maxid ) {
3098 if ( transcount < tp + maxid - minid + 1) {
3100 trie->trans = (reg_trie_trans *)
3101 PerlMemShared_realloc( trie->trans,
3103 * sizeof(reg_trie_trans) );
3104 Zero( trie->trans + (transcount / 2),
3108 base = trie->uniquecharcount + tp - minid;
3109 if ( maxid == minid ) {
3111 for ( ; zp < tp ; zp++ ) {
3112 if ( ! trie->trans[ zp ].next ) {
3113 base = trie->uniquecharcount + zp - minid;
3114 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
3116 trie->trans[ zp ].check = state;
3122 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
3124 trie->trans[ tp ].check = state;
3129 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3130 const U32 tid = base
3131 - trie->uniquecharcount
3132 + TRIE_LIST_ITEM( state, idx ).forid;
3133 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
3135 trie->trans[ tid ].check = state;
3137 tp += ( maxid - minid + 1 );
3139 Safefree(trie->states[ state ].trans.list);
3142 DEBUG_TRIE_COMPILE_MORE_r(
3143 Perl_re_printf( aTHX_ " base: %d\n",base);
3146 trie->states[ state ].trans.base=base;
3148 trie->lasttrans = tp + 1;
3152 Second Pass -- Flat Table Representation.
3154 we dont use the 0 slot of either trans[] or states[] so we add 1 to
3155 each. We know that we will need Charcount+1 trans at most to store
3156 the data (one row per char at worst case) So we preallocate both
3157 structures assuming worst case.
3159 We then construct the trie using only the .next slots of the entry
3162 We use the .check field of the first entry of the node temporarily
3163 to make compression both faster and easier by keeping track of how
3164 many non zero fields are in the node.
3166 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
3169 There are two terms at use here: state as a TRIE_NODEIDX() which is
3170 a number representing the first entry of the node, and state as a
3171 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
3172 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
3173 if there are 2 entrys per node. eg:
3181 The table is internally in the right hand, idx form. However as we
3182 also have to deal with the states array which is indexed by nodenum
3183 we have to use TRIE_NODENUM() to convert.
3186 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
3189 trie->trans = (reg_trie_trans *)
3190 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
3191 * trie->uniquecharcount + 1,
3192 sizeof(reg_trie_trans) );
3193 trie->states = (reg_trie_state *)
3194 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3195 sizeof(reg_trie_state) );
3196 next_alloc = trie->uniquecharcount + 1;
3199 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3201 regnode *noper = NEXTOPER( cur );
3203 U32 state = 1; /* required init */
3205 U16 charid = 0; /* sanity init */
3206 U32 accept_state = 0; /* sanity init */
3208 U32 wordlen = 0; /* required init */
3210 if (OP(noper) == NOTHING) {
3211 regnode *noper_next= regnext(noper);
3212 if (noper_next < tail)
3217 && ( OP(noper) == flags
3218 || (flags == EXACT && OP(noper) == EXACT_ONLY8)
3219 || (flags == EXACTFU && ( OP(noper) == EXACTFU_ONLY8
3220 || OP(noper) == EXACTFUP))))
3222 const U8 *uc= (U8*)STRING(noper);
3223 const U8 *e= uc + STR_LEN(noper);
3225 for ( ; uc < e ; uc += len ) {
3230 charid = trie->charmap[ uvc ];
3232 SV* const * const svpp = hv_fetch( widecharmap,
3236 charid = svpp ? (U16)SvIV(*svpp) : 0;
3240 if ( !trie->trans[ state + charid ].next ) {
3241 trie->trans[ state + charid ].next = next_alloc;
3242 trie->trans[ state ].check++;
3243 prev_states[TRIE_NODENUM(next_alloc)]
3244 = TRIE_NODENUM(state);
3245 next_alloc += trie->uniquecharcount;
3247 state = trie->trans[ state + charid ].next;
3249 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3251 /* charid is now 0 if we dont know the char read, or
3252 * nonzero if we do */
3255 accept_state = TRIE_NODENUM( state );
3256 TRIE_HANDLE_WORD(accept_state);
3258 } /* end second pass */
3260 /* and now dump it out before we compress it */
3261 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3263 next_alloc, depth+1));
3267 * Inplace compress the table.*
3269 For sparse data sets the table constructed by the trie algorithm will
3270 be mostly 0/FAIL transitions or to put it another way mostly empty.
3271 (Note that leaf nodes will not contain any transitions.)
3273 This algorithm compresses the tables by eliminating most such
3274 transitions, at the cost of a modest bit of extra work during lookup:
3276 - Each states[] entry contains a .base field which indicates the
3277 index in the state[] array wheres its transition data is stored.
3279 - If .base is 0 there are no valid transitions from that node.
3281 - If .base is nonzero then charid is added to it to find an entry in
3284 -If trans[states[state].base+charid].check!=state then the
3285 transition is taken to be a 0/Fail transition. Thus if there are fail
3286 transitions at the front of the node then the .base offset will point
3287 somewhere inside the previous nodes data (or maybe even into a node
3288 even earlier), but the .check field determines if the transition is
3292 The following process inplace converts the table to the compressed
3293 table: We first do not compress the root node 1,and mark all its
3294 .check pointers as 1 and set its .base pointer as 1 as well. This
3295 allows us to do a DFA construction from the compressed table later,
3296 and ensures that any .base pointers we calculate later are greater
3299 - We set 'pos' to indicate the first entry of the second node.
3301 - We then iterate over the columns of the node, finding the first and
3302 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3303 and set the .check pointers accordingly, and advance pos
3304 appropriately and repreat for the next node. Note that when we copy
3305 the next pointers we have to convert them from the original
3306 NODEIDX form to NODENUM form as the former is not valid post
3309 - If a node has no transitions used we mark its base as 0 and do not
3310 advance the pos pointer.
3312 - If a node only has one transition we use a second pointer into the
3313 structure to fill in allocated fail transitions from other states.
3314 This pointer is independent of the main pointer and scans forward
3315 looking for null transitions that are allocated to a state. When it
3316 finds one it writes the single transition into the "hole". If the
3317 pointer doesnt find one the single transition is appended as normal.
3319 - Once compressed we can Renew/realloc the structures to release the
3322 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3323 specifically Fig 3.47 and the associated pseudocode.
3327 const U32 laststate = TRIE_NODENUM( next_alloc );
3330 trie->statecount = laststate;
3332 for ( state = 1 ; state < laststate ; state++ ) {
3334 const U32 stateidx = TRIE_NODEIDX( state );
3335 const U32 o_used = trie->trans[ stateidx ].check;
3336 U32 used = trie->trans[ stateidx ].check;
3337 trie->trans[ stateidx ].check = 0;
3340 used && charid < trie->uniquecharcount;
3343 if ( flag || trie->trans[ stateidx + charid ].next ) {
3344 if ( trie->trans[ stateidx + charid ].next ) {
3346 for ( ; zp < pos ; zp++ ) {
3347 if ( ! trie->trans[ zp ].next ) {
3351 trie->states[ state ].trans.base
3353 + trie->uniquecharcount
3355 trie->trans[ zp ].next
3356 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3358 trie->trans[ zp ].check = state;
3359 if ( ++zp > pos ) pos = zp;
3366 trie->states[ state ].trans.base
3367 = pos + trie->uniquecharcount - charid ;
3369 trie->trans[ pos ].next
3370 = SAFE_TRIE_NODENUM(
3371 trie->trans[ stateidx + charid ].next );
3372 trie->trans[ pos ].check = state;
3377 trie->lasttrans = pos + 1;
3378 trie->states = (reg_trie_state *)
3379 PerlMemShared_realloc( trie->states, laststate
3380 * sizeof(reg_trie_state) );
3381 DEBUG_TRIE_COMPILE_MORE_r(
3382 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
3384 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3388 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3391 } /* end table compress */
3393 DEBUG_TRIE_COMPILE_MORE_r(
3394 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
3396 (UV)trie->statecount,
3397 (UV)trie->lasttrans)
3399 /* resize the trans array to remove unused space */
3400 trie->trans = (reg_trie_trans *)
3401 PerlMemShared_realloc( trie->trans, trie->lasttrans
3402 * sizeof(reg_trie_trans) );
3404 { /* Modify the program and insert the new TRIE node */
3405 U8 nodetype =(U8)(flags & 0xFF);
3409 regnode *optimize = NULL;
3410 #ifdef RE_TRACK_PATTERN_OFFSETS
3413 U32 mjd_nodelen = 0;
3414 #endif /* RE_TRACK_PATTERN_OFFSETS */
3415 #endif /* DEBUGGING */
3417 This means we convert either the first branch or the first Exact,
3418 depending on whether the thing following (in 'last') is a branch
3419 or not and whther first is the startbranch (ie is it a sub part of
3420 the alternation or is it the whole thing.)
3421 Assuming its a sub part we convert the EXACT otherwise we convert
3422 the whole branch sequence, including the first.
3424 /* Find the node we are going to overwrite */
3425 if ( first != startbranch || OP( last ) == BRANCH ) {
3426 /* branch sub-chain */
3427 NEXT_OFF( first ) = (U16)(last - first);
3428 #ifdef RE_TRACK_PATTERN_OFFSETS
3430 mjd_offset= Node_Offset((convert));
3431 mjd_nodelen= Node_Length((convert));
3434 /* whole branch chain */
3436 #ifdef RE_TRACK_PATTERN_OFFSETS
3439 const regnode *nop = NEXTOPER( convert );
3440 mjd_offset= Node_Offset((nop));
3441 mjd_nodelen= Node_Length((nop));
3445 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
3447 (UV)mjd_offset, (UV)mjd_nodelen)
3450 /* But first we check to see if there is a common prefix we can
3451 split out as an EXACT and put in front of the TRIE node. */
3452 trie->startstate= 1;
3453 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3454 /* we want to find the first state that has more than
3455 * one transition, if that state is not the first state
3456 * then we have a common prefix which we can remove.
3459 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3461 I32 first_ofs = -1; /* keeps track of the ofs of the first
3462 transition, -1 means none */
3464 const U32 base = trie->states[ state ].trans.base;
3466 /* does this state terminate an alternation? */
3467 if ( trie->states[state].wordnum )
3470 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3471 if ( ( base + ofs >= trie->uniquecharcount ) &&
3472 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3473 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3475 if ( ++count > 1 ) {
3476 /* we have more than one transition */
3479 /* if this is the first state there is no common prefix
3480 * to extract, so we can exit */
3481 if ( state == 1 ) break;
3482 tmp = av_fetch( revcharmap, ofs, 0);
3483 ch = (U8*)SvPV_nolen_const( *tmp );
3485 /* if we are on count 2 then we need to initialize the
3486 * bitmap, and store the previous char if there was one
3489 /* clear the bitmap */
3490 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3492 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
3495 if (first_ofs >= 0) {
3496 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
3497 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3499 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3501 Perl_re_printf( aTHX_ "%s", (char*)ch)
3505 /* store the current firstchar in the bitmap */
3506 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3507 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
3513 /* This state has only one transition, its transition is part
3514 * of a common prefix - we need to concatenate the char it
3515 * represents to what we have so far. */
3516 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
3518 char *ch = SvPV( *tmp, len );
3520 SV *sv=sv_newmortal();
3521 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
3523 (UV)state, (UV)first_ofs,
3524 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3525 PL_colors[0], PL_colors[1],
3526 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3527 PERL_PV_ESCAPE_FIRSTCHAR
3532 OP( convert ) = nodetype;
3533 str=STRING(convert);
3536 STR_LEN(convert) += len;
3542 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
3547 trie->prefixlen = (state-1);
3549 regnode *n = convert+NODE_SZ_STR(convert);
3550 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3551 trie->startstate = state;
3552 trie->minlen -= (state - 1);
3553 trie->maxlen -= (state - 1);
3555 /* At least the UNICOS C compiler choked on this
3556 * being argument to DEBUG_r(), so let's just have
3559 #ifdef PERL_EXT_RE_BUILD
3565 regnode *fix = convert;
3566 U32 word = trie->wordcount;
3567 #ifdef RE_TRACK_PATTERN_OFFSETS
3570 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3571 while( ++fix < n ) {
3572 Set_Node_Offset_Length(fix, 0, 0);
3575 SV ** const tmp = av_fetch( trie_words, word, 0 );
3577 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3578 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3580 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3588 NEXT_OFF(convert) = (U16)(tail - convert);
3589 DEBUG_r(optimize= n);
3595 if ( trie->maxlen ) {
3596 NEXT_OFF( convert ) = (U16)(tail - convert);
3597 ARG_SET( convert, data_slot );
3598 /* Store the offset to the first unabsorbed branch in
3599 jump[0], which is otherwise unused by the jump logic.
3600 We use this when dumping a trie and during optimisation. */
3602 trie->jump[0] = (U16)(nextbranch - convert);
3604 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3605 * and there is a bitmap
3606 * and the first "jump target" node we found leaves enough room
3607 * then convert the TRIE node into a TRIEC node, with the bitmap
3608 * embedded inline in the opcode - this is hypothetically faster.
3610 if ( !trie->states[trie->startstate].wordnum
3612 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3614 OP( convert ) = TRIEC;
3615 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3616 PerlMemShared_free(trie->bitmap);
3619 OP( convert ) = TRIE;
3621 /* store the type in the flags */
3622 convert->flags = nodetype;
3626 + regarglen[ OP( convert ) ];
3628 /* XXX We really should free up the resource in trie now,
3629 as we won't use them - (which resources?) dmq */
3631 /* needed for dumping*/
3632 DEBUG_r(if (optimize) {
3633 regnode *opt = convert;
3635 while ( ++opt < optimize) {
3636 Set_Node_Offset_Length(opt, 0, 0);
3639 Try to clean up some of the debris left after the
3642 while( optimize < jumper ) {
3643 Track_Code( mjd_nodelen += Node_Length((optimize)); );
3644 OP( optimize ) = OPTIMIZED;
3645 Set_Node_Offset_Length(optimize, 0, 0);
3648 Set_Node_Offset_Length(convert, mjd_offset, mjd_nodelen);
3650 } /* end node insert */
3652 /* Finish populating the prev field of the wordinfo array. Walk back
3653 * from each accept state until we find another accept state, and if
3654 * so, point the first word's .prev field at the second word. If the
3655 * second already has a .prev field set, stop now. This will be the
3656 * case either if we've already processed that word's accept state,
3657 * or that state had multiple words, and the overspill words were
3658 * already linked up earlier.
3665 for (word=1; word <= trie->wordcount; word++) {
3667 if (trie->wordinfo[word].prev)
3669 state = trie->wordinfo[word].accept;
3671 state = prev_states[state];
3674 prev = trie->states[state].wordnum;
3678 trie->wordinfo[word].prev = prev;
3680 Safefree(prev_states);
3684 /* and now dump out the compressed format */
3685 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3687 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3689 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3690 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3692 SvREFCNT_dec_NN(revcharmap);
3696 : trie->startstate>1
3702 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3704 /* The Trie is constructed and compressed now so we can build a fail array if
3707 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3709 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3713 We find the fail state for each state in the trie, this state is the longest
3714 proper suffix of the current state's 'word' that is also a proper prefix of
3715 another word in our trie. State 1 represents the word '' and is thus the
3716 default fail state. This allows the DFA not to have to restart after its
3717 tried and failed a word at a given point, it simply continues as though it
3718 had been matching the other word in the first place.
3720 'abcdgu'=~/abcdefg|cdgu/
3721 When we get to 'd' we are still matching the first word, we would encounter
3722 'g' which would fail, which would bring us to the state representing 'd' in
3723 the second word where we would try 'g' and succeed, proceeding to match
3726 /* add a fail transition */
3727 const U32 trie_offset = ARG(source);
3728 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3730 const U32 ucharcount = trie->uniquecharcount;
3731 const U32 numstates = trie->statecount;
3732 const U32 ubound = trie->lasttrans + ucharcount;
3736 U32 base = trie->states[ 1 ].trans.base;
3739 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3741 GET_RE_DEBUG_FLAGS_DECL;
3743 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3744 PERL_UNUSED_CONTEXT;
3746 PERL_UNUSED_ARG(depth);
3749 if ( OP(source) == TRIE ) {
3750 struct regnode_1 *op = (struct regnode_1 *)
3751 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3752 StructCopy(source, op, struct regnode_1);
3753 stclass = (regnode *)op;
3755 struct regnode_charclass *op = (struct regnode_charclass *)
3756 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3757 StructCopy(source, op, struct regnode_charclass);
3758 stclass = (regnode *)op;
3760 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3762 ARG_SET( stclass, data_slot );
3763 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3764 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3765 aho->trie=trie_offset;
3766 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3767 Copy( trie->states, aho->states, numstates, reg_trie_state );
3768 Newx( q, numstates, U32);
3769 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3772 /* initialize fail[0..1] to be 1 so that we always have
3773 a valid final fail state */
3774 fail[ 0 ] = fail[ 1 ] = 1;
3776 for ( charid = 0; charid < ucharcount ; charid++ ) {
3777 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3779 q[ q_write ] = newstate;
3780 /* set to point at the root */
3781 fail[ q[ q_write++ ] ]=1;
3784 while ( q_read < q_write) {
3785 const U32 cur = q[ q_read++ % numstates ];
3786 base = trie->states[ cur ].trans.base;
3788 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3789 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3791 U32 fail_state = cur;
3794 fail_state = fail[ fail_state ];
3795 fail_base = aho->states[ fail_state ].trans.base;
3796 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3798 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3799 fail[ ch_state ] = fail_state;
3800 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3802 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3804 q[ q_write++ % numstates] = ch_state;
3808 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3809 when we fail in state 1, this allows us to use the
3810 charclass scan to find a valid start char. This is based on the principle
3811 that theres a good chance the string being searched contains lots of stuff
3812 that cant be a start char.
3814 fail[ 0 ] = fail[ 1 ] = 0;
3815 DEBUG_TRIE_COMPILE_r({
3816 Perl_re_indentf( aTHX_ "Stclass Failtable (%" UVuf " states): 0",
3817 depth, (UV)numstates
3819 for( q_read=1; q_read<numstates; q_read++ ) {
3820 Perl_re_printf( aTHX_ ", %" UVuf, (UV)fail[q_read]);
3822 Perl_re_printf( aTHX_ "\n");
3825 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3830 /* The below joins as many adjacent EXACTish nodes as possible into a single
3831 * one. The regop may be changed if the node(s) contain certain sequences that
3832 * require special handling. The joining is only done if:
3833 * 1) there is room in the current conglomerated node to entirely contain the
3835 * 2) they are compatible node types
3837 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3838 * these get optimized out
3840 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3841 * as possible, even if that means splitting an existing node so that its first
3842 * part is moved to the preceeding node. This would maximise the efficiency of
3843 * memEQ during matching.
3845 * If a node is to match under /i (folded), the number of characters it matches
3846 * can be different than its character length if it contains a multi-character
3847 * fold. *min_subtract is set to the total delta number of characters of the
3850 * And *unfolded_multi_char is set to indicate whether or not the node contains
3851 * an unfolded multi-char fold. This happens when it won't be known until
3852 * runtime whether the fold is valid or not; namely
3853 * 1) for EXACTF nodes that contain LATIN SMALL LETTER SHARP S, as only if the
3854 * target string being matched against turns out to be UTF-8 is that fold
3856 * 2) for EXACTFL nodes whose folding rules depend on the locale in force at
3858 * (Multi-char folds whose components are all above the Latin1 range are not
3859 * run-time locale dependent, and have already been folded by the time this
3860 * function is called.)
3862 * This is as good a place as any to discuss the design of handling these
3863 * multi-character fold sequences. It's been wrong in Perl for a very long
3864 * time. There are three code points in Unicode whose multi-character folds
3865 * were long ago discovered to mess things up. The previous designs for
3866 * dealing with these involved assigning a special node for them. This
3867 * approach doesn't always work, as evidenced by this example:
3868 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3869 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3870 * would match just the \xDF, it won't be able to handle the case where a
3871 * successful match would have to cross the node's boundary. The new approach
3872 * that hopefully generally solves the problem generates an EXACTFUP node
3873 * that is "sss" in this case.
3875 * It turns out that there are problems with all multi-character folds, and not
3876 * just these three. Now the code is general, for all such cases. The
3877 * approach taken is:
3878 * 1) This routine examines each EXACTFish node that could contain multi-
3879 * character folded sequences. Since a single character can fold into
3880 * such a sequence, the minimum match length for this node is less than
3881 * the number of characters in the node. This routine returns in
3882 * *min_subtract how many characters to subtract from the the actual
3883 * length of the string to get a real minimum match length; it is 0 if
3884 * there are no multi-char foldeds. This delta is used by the caller to
3885 * adjust the min length of the match, and the delta between min and max,
3886 * so that the optimizer doesn't reject these possibilities based on size
3889 * 2) For the sequence involving the LATIN SMALL LETTER SHARP S (U+00DF)
3890 * under /u, we fold it to 'ss' in regatom(), and in this routine, after
3891 * joining, we scan for occurrences of the sequence 'ss' in non-UTF-8
3892 * EXACTFU nodes. The node type of such nodes is then changed to
3893 * EXACTFUP, indicating it is problematic, and needs careful handling.
3894 * (The procedures in step 1) above are sufficient to handle this case in
3895 * UTF-8 encoded nodes.) The reason this is problematic is that this is
3896 * the only case where there is a possible fold length change in non-UTF-8
3897 * patterns. By reserving a special node type for problematic cases, the
3898 * far more common regular EXACTFU nodes can be processed faster.
3899 * regexec.c takes advantage of this.
3901 * EXACTFUP has been created as a grab-bag for (hopefully uncommon)
3902 * problematic cases. These all only occur when the pattern is not
3903 * UTF-8. In addition to the 'ss' sequence where there is a possible fold
3904 * length change, it handles the situation where the string cannot be
3905 * entirely folded. The strings in an EXACTFish node are folded as much
3906 * as possible during compilation in regcomp.c. This saves effort in
3907 * regex matching. By using an EXACTFUP node when it is not possible to
3908 * fully fold at compile time, regexec.c can know that everything in an
3909 * EXACTFU node is folded, so folding can be skipped at runtime. The only
3910 * case where folding in EXACTFU nodes can't be done at compile time is
3911 * the presumably uncommon MICRO SIGN, when the pattern isn't UTF-8. This
3912 * is because its fold requires UTF-8 to represent. Thus EXACTFUP nodes
3913 * handle two very different cases. Alternatively, there could have been
3914 * a node type where there are length changes, one for unfolded, and one
3915 * for both. If yet another special case needed to be created, the number
3916 * of required node types would have to go to 7. khw figures that even
3917 * though there are plenty of node types to spare, that the maintenance
3918 * cost wasn't worth the small speedup of doing it that way, especially
3919 * since he thinks the MICRO SIGN is rarely encountered in practice.
3921 * There are other cases where folding isn't done at compile time, but
3922 * none of them are under /u, and hence not for EXACTFU nodes. The folds
3923 * in EXACTFL nodes aren't known until runtime, and vary as the locale
3924 * changes. Some folds in EXACTF depend on if the runtime target string
3925 * is UTF-8 or not. (regatom() will create an EXACTFU node even under /di
3926 * when no fold in it depends on the UTF-8ness of the target string.)
3928 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3929 * validity of the fold won't be known until runtime, and so must remain
3930 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFAA
3931 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3932 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3933 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3934 * The reason this is a problem is that the optimizer part of regexec.c
3935 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3936 * that a character in the pattern corresponds to at most a single
3937 * character in the target string. (And I do mean character, and not byte
3938 * here, unlike other parts of the documentation that have never been
3939 * updated to account for multibyte Unicode.) Sharp s in EXACTF and
3940 * EXACTFL nodes can match the two character string 'ss'; in EXACTFAA
3941 * nodes it can match "\x{17F}\x{17F}". These, along with other ones in
3942 * EXACTFL nodes, violate the assumption, and they are the only instances
3943 * where it is violated. I'm reluctant to try to change the assumption,
3944 * as the code involved is impenetrable to me (khw), so instead the code
3945 * here punts. This routine examines EXACTFL nodes, and (when the pattern
3946 * isn't UTF-8) EXACTF and EXACTFAA for such unfolded folds, and returns a
3947 * boolean indicating whether or not the node contains such a fold. When
3948 * it is true, the caller sets a flag that later causes the optimizer in
3949 * this file to not set values for the floating and fixed string lengths,
3950 * and thus avoids the optimizer code in regexec.c that makes the invalid
3951 * assumption. Thus, there is no optimization based on string lengths for
3952 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3953 * EXACTF and EXACTFAA nodes that contain the sharp s. (The reason the
3954 * assumption is wrong only in these cases is that all other non-UTF-8
3955 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3956 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3957 * EXACTF nodes because we don't know at compile time if it actually
3958 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3959 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3960 * always matches; and EXACTFAA where it never does. In an EXACTFAA node
3961 * in a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3962 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3963 * string would require the pattern to be forced into UTF-8, the overhead
3964 * of which we want to avoid. Similarly the unfolded multi-char folds in
3965 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3968 * Similarly, the code that generates tries doesn't currently handle
3969 * not-already-folded multi-char folds, and it looks like a pain to change
3970 * that. Therefore, trie generation of EXACTFAA nodes with the sharp s
3971 * doesn't work. Instead, such an EXACTFAA is turned into a new regnode,
3972 * EXACTFAA_NO_TRIE, which the trie code knows not to handle. Most people
3973 * using /iaa matching will be doing so almost entirely with ASCII
3974 * strings, so this should rarely be encountered in practice */
3976 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3977 if (PL_regkind[OP(scan)] == EXACT) \
3978 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags), NULL, depth+1)
3981 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3982 UV *min_subtract, bool *unfolded_multi_char,
3983 U32 flags, regnode *val, U32 depth)
3985 /* Merge several consecutive EXACTish nodes into one. */
3987 regnode *n = regnext(scan);
3989 regnode *next = scan + NODE_SZ_STR(scan);
3993 regnode *stop = scan;
3994 GET_RE_DEBUG_FLAGS_DECL;
3996 PERL_UNUSED_ARG(depth);
3999 PERL_ARGS_ASSERT_JOIN_EXACT;
4000 #ifndef EXPERIMENTAL_INPLACESCAN
4001 PERL_UNUSED_ARG(flags);
4002 PERL_UNUSED_ARG(val);
4004 DEBUG_PEEP("join", scan, depth, 0);
4006 assert(PL_regkind[OP(scan)] == EXACT);
4008 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
4009 * EXACT ones that are mergeable to the current one. */
4011 && ( PL_regkind[OP(n)] == NOTHING
4012 || (stringok && PL_regkind[OP(n)] == EXACT))
4014 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
4017 if (OP(n) == TAIL || n > next)
4019 if (PL_regkind[OP(n)] == NOTHING) {
4020 DEBUG_PEEP("skip:", n, depth, 0);
4021 NEXT_OFF(scan) += NEXT_OFF(n);
4022 next = n + NODE_STEP_REGNODE;
4029 else if (stringok) {
4030 const unsigned int oldl = STR_LEN(scan);
4031 regnode * const nnext = regnext(n);
4033 /* XXX I (khw) kind of doubt that this works on platforms (should
4034 * Perl ever run on one) where U8_MAX is above 255 because of lots
4035 * of other assumptions */
4036 /* Don't join if the sum can't fit into a single node */
4037 if (oldl + STR_LEN(n) > U8_MAX)
4040 /* Joining something that requires UTF-8 with something that
4041 * doesn't, means the result requires UTF-8. */
4042 if (OP(scan) == EXACT && (OP(n) == EXACT_ONLY8)) {
4043 OP(scan) = EXACT_ONLY8;
4045 else if (OP(scan) == EXACT_ONLY8 && (OP(n) == EXACT)) {
4046 ; /* join is compatible, no need to change OP */
4048 else if ((OP(scan) == EXACTFU) && (OP(n) == EXACTFU_ONLY8)) {
4049 OP(scan) = EXACTFU_ONLY8;
4051 else if ((OP(scan) == EXACTFU_ONLY8) && (OP(n) == EXACTFU)) {
4052 ; /* join is compatible, no need to change OP */
4054 else if (OP(scan) == EXACTFU && OP(n) == EXACTFU) {
4055 ; /* join is compatible, no need to change OP */
4057 else if (OP(scan) == EXACTFU && OP(n) == EXACTFU_S_EDGE) {
4059 /* Under /di, temporary EXACTFU_S_EDGE nodes are generated,
4060 * which can join with EXACTFU ones. We check for this case
4061 * here. These need to be resolved to either EXACTFU or
4062 * EXACTF at joining time. They have nothing in them that
4063 * would forbid them from being the more desirable EXACTFU
4064 * nodes except that they begin and/or end with a single [Ss].
4065 * The reason this is problematic is because they could be
4066 * joined in this loop with an adjacent node that ends and/or
4067 * begins with [Ss] which would then form the sequence 'ss',
4068 * which matches differently under /di than /ui, in which case
4069 * EXACTFU can't be used. If the 'ss' sequence doesn't get
4070 * formed, the nodes get absorbed into any adjacent EXACTFU
4071 * node. And if the only adjacent node is EXACTF, they get
4072 * absorbed into that, under the theory that a longer node is
4073 * better than two shorter ones, even if one is EXACTFU. Note
4074 * that EXACTFU_ONLY8 is generated only for UTF-8 patterns,
4075 * and the EXACTFU_S_EDGE ones only for non-UTF-8. */
4077 if (STRING(n)[STR_LEN(n)-1] == 's') {
4079 /* Here the joined node would end with 's'. If the node
4080 * following the combination is an EXACTF one, it's better to
4081 * join this trailing edge 's' node with that one, leaving the
4082 * current one in 'scan' be the more desirable EXACTFU */
4083 if (OP(nnext) == EXACTF) {
4087 OP(scan) = EXACTFU_S_EDGE;
4089 } /* Otherwise, the beginning 's' of the 2nd node just
4090 becomes an interior 's' in 'scan' */
4092 else if (OP(scan) == EXACTF && OP(n) == EXACTF) {
4093 ; /* join is compatible, no need to change OP */
4095 else if (OP(scan) == EXACTF && OP(n) == EXACTFU_S_EDGE) {
4097 /* EXACTF nodes are compatible for joining with EXACTFU_S_EDGE
4098 * nodes. But the latter nodes can be also joined with EXACTFU
4099 * ones, and that is a better outcome, so if the node following
4100 * 'n' is EXACTFU, quit now so that those two can be joined
4102 if (OP(nnext) == EXACTFU) {
4106 /* The join is compatible, and the combined node will be
4107 * EXACTF. (These don't care if they begin or end with 's' */
4109 else if (OP(scan) == EXACTFU_S_EDGE && OP(n) == EXACTFU_S_EDGE) {
4110 if ( STRING(scan)[STR_LEN(scan)-1] == 's'
4111 && STRING(n)[0] == 's')
4113 /* When combined, we have the sequence 'ss', which means we
4114 * have to remain /di */
4118 else if (OP(scan) == EXACTFU_S_EDGE && OP(n) == EXACTFU) {
4119 if (STRING(n)[0] == 's') {
4120 ; /* Here the join is compatible and the combined node
4121 starts with 's', no need to change OP */
4123 else { /* Now the trailing 's' is in the interior */
4127 else if (OP(scan) == EXACTFU_S_EDGE && OP(n) == EXACTF) {
4129 /* The join is compatible, and the combined node will be
4130 * EXACTF. (These don't care if they begin or end with 's' */
4133 else if (OP(scan) != OP(n)) {
4135 /* The only other compatible joinings are the same node type */
4139 DEBUG_PEEP("merg", n, depth, 0);
4142 NEXT_OFF(scan) += NEXT_OFF(n);
4143 STR_LEN(scan) += STR_LEN(n);
4144 next = n + NODE_SZ_STR(n);
4145 /* Now we can overwrite *n : */
4146 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
4154 #ifdef EXPERIMENTAL_INPLACESCAN
4155 if (flags && !NEXT_OFF(n)) {
4156 DEBUG_PEEP("atch", val, depth, 0);
4157 if (reg_off_by_arg[OP(n)]) {
4158 ARG_SET(n, val - n);
4161 NEXT_OFF(n) = val - n;
4168 /* This temporary node can now be turned into EXACTFU, and must, as
4169 * regexec.c doesn't handle it */
4170 if (OP(scan) == EXACTFU_S_EDGE) {
4175 *unfolded_multi_char = FALSE;
4177 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
4178 * can now analyze for sequences of problematic code points. (Prior to
4179 * this final joining, sequences could have been split over boundaries, and
4180 * hence missed). The sequences only happen in folding, hence for any
4181 * non-EXACT EXACTish node */
4182 if (OP(scan) != EXACT && OP(scan) != EXACT_ONLY8 && OP(scan) != EXACTL) {
4183 U8* s0 = (U8*) STRING(scan);
4185 U8* s_end = s0 + STR_LEN(scan);
4187 int total_count_delta = 0; /* Total delta number of characters that
4188 multi-char folds expand to */
4190 /* One pass is made over the node's string looking for all the
4191 * possibilities. To avoid some tests in the loop, there are two main
4192 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
4197 if (OP(scan) == EXACTFL) {
4200 /* An EXACTFL node would already have been changed to another
4201 * node type unless there is at least one character in it that
4202 * is problematic; likely a character whose fold definition
4203 * won't be known until runtime, and so has yet to be folded.
4204 * For all but the UTF-8 locale, folds are 1-1 in length, but
4205 * to handle the UTF-8 case, we need to create a temporary
4206 * folded copy using UTF-8 locale rules in order to analyze it.
4207 * This is because our macros that look to see if a sequence is
4208 * a multi-char fold assume everything is folded (otherwise the
4209 * tests in those macros would be too complicated and slow).
4210 * Note that here, the non-problematic folds will have already
4211 * been done, so we can just copy such characters. We actually
4212 * don't completely fold the EXACTFL string. We skip the
4213 * unfolded multi-char folds, as that would just create work
4214 * below to figure out the size they already are */
4216 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
4219 STRLEN s_len = UTF8SKIP(s);
4220 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
4221 Copy(s, d, s_len, U8);
4224 else if (is_FOLDS_TO_MULTI_utf8(s)) {
4225 *unfolded_multi_char = TRUE;
4226 Copy(s, d, s_len, U8);
4229 else if (isASCII(*s)) {
4230 *(d++) = toFOLD(*s);
4234 _toFOLD_utf8_flags(s, s_end, d, &len, FOLD_FLAGS_FULL);
4240 /* Point the remainder of the routine to look at our temporary
4244 } /* End of creating folded copy of EXACTFL string */
4246 /* Examine the string for a multi-character fold sequence. UTF-8
4247 * patterns have all characters pre-folded by the time this code is
4249 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
4250 length sequence we are looking for is 2 */
4252 int count = 0; /* How many characters in a multi-char fold */
4253 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
4254 if (! len) { /* Not a multi-char fold: get next char */
4259 { /* Here is a generic multi-char fold. */
4260 U8* multi_end = s + len;
4262 /* Count how many characters are in it. In the case of
4263 * /aa, no folds which contain ASCII code points are
4264 * allowed, so check for those, and skip if found. */
4265 if (OP(scan) != EXACTFAA && OP(scan) != EXACTFAA_NO_TRIE) {
4266 count = utf8_length(s, multi_end);
4270 while (s < multi_end) {
4273 goto next_iteration;
4283 /* The delta is how long the sequence is minus 1 (1 is how long
4284 * the character that folds to the sequence is) */
4285 total_count_delta += count - 1;
4289 /* We created a temporary folded copy of the string in EXACTFL
4290 * nodes. Therefore we need to be sure it doesn't go below zero,
4291 * as the real string could be shorter */
4292 if (OP(scan) == EXACTFL) {
4293 int total_chars = utf8_length((U8*) STRING(scan),
4294 (U8*) STRING(scan) + STR_LEN(scan));
4295 if (total_count_delta > total_chars) {
4296 total_count_delta = total_chars;
4300 *min_subtract += total_count_delta;
4303 else if (OP(scan) == EXACTFAA) {
4305 /* Non-UTF-8 pattern, EXACTFAA node. There can't be a multi-char
4306 * fold to the ASCII range (and there are no existing ones in the
4307 * upper latin1 range). But, as outlined in the comments preceding
4308 * this function, we need to flag any occurrences of the sharp s.
4309 * This character forbids trie formation (because of added
4311 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
4312 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
4313 || UNICODE_DOT_DOT_VERSION > 0)
4315 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
4316 OP(scan) = EXACTFAA_NO_TRIE;
4317 *unfolded_multi_char = TRUE;
4325 /* Non-UTF-8 pattern, not EXACTFAA node. Look for the multi-char
4326 * folds that are all Latin1. As explained in the comments
4327 * preceding this function, we look also for the sharp s in EXACTF
4328 * and EXACTFL nodes; it can be in the final position. Otherwise
4329 * we can stop looking 1 byte earlier because have to find at least
4330 * two characters for a multi-fold */
4331 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
4336 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
4337 if (! len) { /* Not a multi-char fold. */
4338 if (*s == LATIN_SMALL_LETTER_SHARP_S
4339 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
4341 *unfolded_multi_char = TRUE;
4348 && isALPHA_FOLD_EQ(*s, 's')
4349 && isALPHA_FOLD_EQ(*(s+1), 's'))
4352 /* EXACTF nodes need to know that the minimum length
4353 * changed so that a sharp s in the string can match this
4354 * ss in the pattern, but they remain EXACTF nodes, as they
4355 * won't match this unless the target string is is UTF-8,
4356 * which we don't know until runtime. EXACTFL nodes can't
4357 * transform into EXACTFU nodes */
4358 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
4359 OP(scan) = EXACTFUP;
4363 *min_subtract += len - 1;
4369 if ( STR_LEN(scan) == 1
4370 && isALPHA_A(* STRING(scan))
4371 && ( OP(scan) == EXACTFAA
4372 || ( OP(scan) == EXACTFU
4373 && ! HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(* STRING(scan)))))
4375 U8 mask = ~ ('A' ^ 'a'); /* These differ in just one bit */
4377 /* Replace a length 1 ASCII fold pair node with an ANYOFM node,
4378 * with the mask set to the complement of the bit that differs
4379 * between upper and lower case, and the lowest code point of the
4380 * pair (which the '&' forces) */
4382 ARG_SET(scan, *STRING(scan) & mask);
4388 /* Allow dumping but overwriting the collection of skipped
4389 * ops and/or strings with fake optimized ops */
4390 n = scan + NODE_SZ_STR(scan);
4398 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl", scan, depth, 0);});
4402 /* REx optimizer. Converts nodes into quicker variants "in place".
4403 Finds fixed substrings. */
4405 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
4406 to the position after last scanned or to NULL. */
4408 #define INIT_AND_WITHP \
4409 assert(!and_withp); \
4410 Newx(and_withp, 1, regnode_ssc); \
4411 SAVEFREEPV(and_withp)
4415 S_unwind_scan_frames(pTHX_ const void *p)
4417 scan_frame *f= (scan_frame *)p;
4419 scan_frame *n= f->next_frame;
4425 /* the return from this sub is the minimum length that could possibly match */
4427 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
4428 SSize_t *minlenp, SSize_t *deltap,
4433 regnode_ssc *and_withp,
4434 U32 flags, U32 depth)
4435 /* scanp: Start here (read-write). */
4436 /* deltap: Write maxlen-minlen here. */
4437 /* last: Stop before this one. */
4438 /* data: string data about the pattern */
4439 /* stopparen: treat close N as END */
4440 /* recursed: which subroutines have we recursed into */
4441 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
4444 /* There must be at least this number of characters to match */
4447 regnode *scan = *scanp, *next;
4449 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
4450 int is_inf_internal = 0; /* The studied chunk is infinite */
4451 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
4452 scan_data_t data_fake;
4453 SV *re_trie_maxbuff = NULL;
4454 regnode *first_non_open = scan;
4455 SSize_t stopmin = SSize_t_MAX;
4456 scan_frame *frame = NULL;
4457 GET_RE_DEBUG_FLAGS_DECL;
4459 PERL_ARGS_ASSERT_STUDY_CHUNK;
4460 RExC_study_started= 1;
4462 Zero(&data_fake, 1, scan_data_t);
4465 while (first_non_open && OP(first_non_open) == OPEN)
4466 first_non_open=regnext(first_non_open);
4472 RExC_study_chunk_recursed_count++;
4474 DEBUG_OPTIMISE_MORE_r(
4476 Perl_re_indentf( aTHX_ "study_chunk stopparen=%ld recursed_count=%lu depth=%lu recursed_depth=%lu scan=%p last=%p",
4477 depth, (long)stopparen,
4478 (unsigned long)RExC_study_chunk_recursed_count,
4479 (unsigned long)depth, (unsigned long)recursed_depth,
4482 if (recursed_depth) {
4485 for ( j = 0 ; j < recursed_depth ; j++ ) {
4486 for ( i = 0 ; i < (U32)RExC_total_parens ; i++ ) {
4488 PAREN_TEST(RExC_study_chunk_recursed +
4489 ( j * RExC_study_chunk_recursed_bytes), i )
4492 !PAREN_TEST(RExC_study_chunk_recursed +
4493 (( j - 1 ) * RExC_study_chunk_recursed_bytes), i)
4496 Perl_re_printf( aTHX_ " %d",(int)i);
4500 if ( j + 1 < recursed_depth ) {
4501 Perl_re_printf( aTHX_ ",");
4505 Perl_re_printf( aTHX_ "\n");
4508 while ( scan && OP(scan) != END && scan < last ){
4509 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
4510 node length to get a real minimum (because
4511 the folded version may be shorter) */
4512 bool unfolded_multi_char = FALSE;
4513 /* Peephole optimizer: */
4514 DEBUG_STUDYDATA("Peep", data, depth, is_inf);
4515 DEBUG_PEEP("Peep", scan, depth, flags);
4518 /* The reason we do this here is that we need to deal with things like
4519 * /(?:f)(?:o)(?:o)/ which cant be dealt with by the normal EXACT
4520 * parsing code, as each (?:..) is handled by a different invocation of
4523 JOIN_EXACT(scan,&min_subtract, &unfolded_multi_char, 0);
4525 /* Follow the next-chain of the current node and optimize
4526 away all the NOTHINGs from it. */
4527 if (OP(scan) != CURLYX) {
4528 const int max = (reg_off_by_arg[OP(scan)]
4530 /* I32 may be smaller than U16 on CRAYs! */
4531 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
4532 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
4536 /* Skip NOTHING and LONGJMP. */
4537 while ((n = regnext(n))
4538 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
4539 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
4540 && off + noff < max)
4542 if (reg_off_by_arg[OP(scan)])
4545 NEXT_OFF(scan) = off;
4548 /* The principal pseudo-switch. Cannot be a switch, since we
4549 look into several different things. */
4550 if ( OP(scan) == DEFINEP ) {
4552 SSize_t deltanext = 0;
4553 SSize_t fake_last_close = 0;
4554 I32 f = SCF_IN_DEFINE;
4556 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4557 scan = regnext(scan);
4558 assert( OP(scan) == IFTHEN );
4559 DEBUG_PEEP("expect IFTHEN", scan, depth, flags);
4561 data_fake.last_closep= &fake_last_close;
4563 next = regnext(scan);
4564 scan = NEXTOPER(NEXTOPER(scan));
4565 DEBUG_PEEP("scan", scan, depth, flags);
4566 DEBUG_PEEP("next", next, depth, flags);
4568 /* we suppose the run is continuous, last=next...
4569 * NOTE we dont use the return here! */
4570 /* DEFINEP study_chunk() recursion */
4571 (void)study_chunk(pRExC_state, &scan, &minlen,
4572 &deltanext, next, &data_fake, stopparen,
4573 recursed_depth, NULL, f, depth+1);
4578 OP(scan) == BRANCH ||
4579 OP(scan) == BRANCHJ ||
4582 next = regnext(scan);
4585 /* The op(next)==code check below is to see if we
4586 * have "BRANCH-BRANCH", "BRANCHJ-BRANCHJ", "IFTHEN-IFTHEN"
4587 * IFTHEN is special as it might not appear in pairs.
4588 * Not sure whether BRANCH-BRANCHJ is possible, regardless
4589 * we dont handle it cleanly. */
4590 if (OP(next) == code || code == IFTHEN) {
4591 /* NOTE - There is similar code to this block below for
4592 * handling TRIE nodes on a re-study. If you change stuff here
4593 * check there too. */
4594 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
4596 regnode * const startbranch=scan;
4598 if (flags & SCF_DO_SUBSTR) {
4599 /* Cannot merge strings after this. */
4600 scan_commit(pRExC_state, data, minlenp, is_inf);
4603 if (flags & SCF_DO_STCLASS)
4604 ssc_init_zero(pRExC_state, &accum);
4606 while (OP(scan) == code) {
4607 SSize_t deltanext, minnext, fake;
4609 regnode_ssc this_class;
4611 DEBUG_PEEP("Branch", scan, depth, flags);
4614 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4616 data_fake.whilem_c = data->whilem_c;
4617 data_fake.last_closep = data->last_closep;
4620 data_fake.last_closep = &fake;
4622 data_fake.pos_delta = delta;
4623 next = regnext(scan);
4625 scan = NEXTOPER(scan); /* everything */
4626 if (code != BRANCH) /* everything but BRANCH */
4627 scan = NEXTOPER(scan);
4629 if (flags & SCF_DO_STCLASS) {
4630 ssc_init(pRExC_state, &this_class);
4631 data_fake.start_class = &this_class;
4632 f = SCF_DO_STCLASS_AND;
4634 if (flags & SCF_WHILEM_VISITED_POS)
4635 f |= SCF_WHILEM_VISITED_POS;
4637 /* we suppose the run is continuous, last=next...*/
4638 /* recurse study_chunk() for each BRANCH in an alternation */
4639 minnext = study_chunk(pRExC_state, &scan, minlenp,
4640 &deltanext, next, &data_fake, stopparen,
4641 recursed_depth, NULL, f, depth+1);
4645 if (deltanext == SSize_t_MAX) {
4646 is_inf = is_inf_internal = 1;
4648 } else if (max1 < minnext + deltanext)
4649 max1 = minnext + deltanext;
4651 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4653 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4654 if ( stopmin > minnext)
4655 stopmin = min + min1;
4656 flags &= ~SCF_DO_SUBSTR;
4658 data->flags |= SCF_SEEN_ACCEPT;
4661 if (data_fake.flags & SF_HAS_EVAL)
4662 data->flags |= SF_HAS_EVAL;
4663 data->whilem_c = data_fake.whilem_c;
4665 if (flags & SCF_DO_STCLASS)
4666 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
4668 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
4670 if (flags & SCF_DO_SUBSTR) {
4671 data->pos_min += min1;
4672 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
4673 data->pos_delta = SSize_t_MAX;
4675 data->pos_delta += max1 - min1;
4676 if (max1 != min1 || is_inf)
4677 data->cur_is_floating = 1;
4680 if (delta == SSize_t_MAX
4681 || SSize_t_MAX - delta - (max1 - min1) < 0)
4682 delta = SSize_t_MAX;
4684 delta += max1 - min1;
4685 if (flags & SCF_DO_STCLASS_OR) {
4686 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
4688 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4689 flags &= ~SCF_DO_STCLASS;
4692 else if (flags & SCF_DO_STCLASS_AND) {
4694 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
4695 flags &= ~SCF_DO_STCLASS;
4698 /* Switch to OR mode: cache the old value of
4699 * data->start_class */
4701 StructCopy(data->start_class, and_withp, regnode_ssc);
4702 flags &= ~SCF_DO_STCLASS_AND;
4703 StructCopy(&accum, data->start_class, regnode_ssc);
4704 flags |= SCF_DO_STCLASS_OR;
4708 if (PERL_ENABLE_TRIE_OPTIMISATION &&
4709 OP( startbranch ) == BRANCH )
4713 Assuming this was/is a branch we are dealing with: 'scan'
4714 now points at the item that follows the branch sequence,
4715 whatever it is. We now start at the beginning of the
4716 sequence and look for subsequences of
4722 which would be constructed from a pattern like
4725 If we can find such a subsequence we need to turn the first
4726 element into a trie and then add the subsequent branch exact
4727 strings to the trie.
4731 1. patterns where the whole set of branches can be
4734 2. patterns where only a subset can be converted.
4736 In case 1 we can replace the whole set with a single regop
4737 for the trie. In case 2 we need to keep the start and end
4740 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
4741 becomes BRANCH TRIE; BRANCH X;
4743 There is an additional case, that being where there is a
4744 common prefix, which gets split out into an EXACT like node
4745 preceding the TRIE node.
4747 If x(1..n)==tail then we can do a simple trie, if not we make
4748 a "jump" trie, such that when we match the appropriate word
4749 we "jump" to the appropriate tail node. Essentially we turn
4750 a nested if into a case structure of sorts.
4755 if (!re_trie_maxbuff) {
4756 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
4757 if (!SvIOK(re_trie_maxbuff))
4758 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
4760 if ( SvIV(re_trie_maxbuff)>=0 ) {
4762 regnode *first = (regnode *)NULL;
4763 regnode *last = (regnode *)NULL;
4764 regnode *tail = scan;
4768 /* var tail is used because there may be a TAIL
4769 regop in the way. Ie, the exacts will point to the
4770 thing following the TAIL, but the last branch will
4771 point at the TAIL. So we advance tail. If we
4772 have nested (?:) we may have to move through several
4776 while ( OP( tail ) == TAIL ) {
4777 /* this is the TAIL generated by (?:) */
4778 tail = regnext( tail );
4782 DEBUG_TRIE_COMPILE_r({
4783 regprop(RExC_rx, RExC_mysv, tail, NULL, pRExC_state);
4784 Perl_re_indentf( aTHX_ "%s %" UVuf ":%s\n",
4786 "Looking for TRIE'able sequences. Tail node is ",
4787 (UV) REGNODE_OFFSET(tail),
4788 SvPV_nolen_const( RExC_mysv )
4794 Step through the branches
4795 cur represents each branch,
4796 noper is the first thing to be matched as part
4798 noper_next is the regnext() of that node.
4800 We normally handle a case like this
4801 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
4802 support building with NOJUMPTRIE, which restricts
4803 the trie logic to structures like /FOO|BAR/.
4805 If noper is a trieable nodetype then the branch is
4806 a possible optimization target. If we are building
4807 under NOJUMPTRIE then we require that noper_next is
4808 the same as scan (our current position in the regex
4811 Once we have two or more consecutive such branches
4812 we can create a trie of the EXACT's contents and
4813 stitch it in place into the program.
4815 If the sequence represents all of the branches in
4816 the alternation we replace the entire thing with a
4819 Otherwise when it is a subsequence we need to
4820 stitch it in place and replace only the relevant
4821 branches. This means the first branch has to remain
4822 as it is used by the alternation logic, and its
4823 next pointer, and needs to be repointed at the item
4824 on the branch chain following the last branch we
4825 have optimized away.
4827 This could be either a BRANCH, in which case the
4828 subsequence is internal, or it could be the item
4829 following the branch sequence in which case the
4830 subsequence is at the end (which does not
4831 necessarily mean the first node is the start of the
4834 TRIE_TYPE(X) is a define which maps the optype to a
4838 ----------------+-----------
4843 EXACTFU_ONLY8 | EXACTFU
4847 EXACTFLU8 | EXACTFLU8
4851 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) \
4853 : ( EXACT == (X) || EXACT_ONLY8 == (X) ) \
4855 : ( EXACTFU == (X) \
4856 || EXACTFU_ONLY8 == (X) \
4857 || EXACTFUP == (X) ) \
4859 : ( EXACTFAA == (X) ) \
4861 : ( EXACTL == (X) ) \
4863 : ( EXACTFLU8 == (X) ) \
4867 /* dont use tail as the end marker for this traverse */
4868 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
4869 regnode * const noper = NEXTOPER( cur );
4870 U8 noper_type = OP( noper );
4871 U8 noper_trietype = TRIE_TYPE( noper_type );
4872 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
4873 regnode * const noper_next = regnext( noper );
4874 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4875 U8 noper_next_trietype = (noper_next && noper_next < tail) ? TRIE_TYPE( noper_next_type ) :0;
4878 DEBUG_TRIE_COMPILE_r({
4879 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4880 Perl_re_indentf( aTHX_ "- %d:%s (%d)",
4882 REG_NODE_NUM(cur), SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur) );
4884 regprop(RExC_rx, RExC_mysv, noper, NULL, pRExC_state);
4885 Perl_re_printf( aTHX_ " -> %d:%s",
4886 REG_NODE_NUM(noper), SvPV_nolen_const(RExC_mysv));
4889 regprop(RExC_rx, RExC_mysv, noper_next, NULL, pRExC_state);
4890 Perl_re_printf( aTHX_ "\t=> %d:%s\t",
4891 REG_NODE_NUM(noper_next), SvPV_nolen_const(RExC_mysv));
4893 Perl_re_printf( aTHX_ "(First==%d,Last==%d,Cur==%d,tt==%s,ntt==%s,nntt==%s)\n",
4894 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
4895 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
4899 /* Is noper a trieable nodetype that can be merged
4900 * with the current trie (if there is one)? */
4904 ( noper_trietype == NOTHING )
4905 || ( trietype == NOTHING )
4906 || ( trietype == noper_trietype )
4909 && noper_next >= tail
4913 /* Handle mergable triable node Either we are
4914 * the first node in a new trieable sequence,
4915 * in which case we do some bookkeeping,
4916 * otherwise we update the end pointer. */
4919 if ( noper_trietype == NOTHING ) {
4920 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
4921 regnode * const noper_next = regnext( noper );
4922 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4923 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
4926 if ( noper_next_trietype ) {
4927 trietype = noper_next_trietype;
4928 } else if (noper_next_type) {
4929 /* a NOTHING regop is 1 regop wide.
4930 * We need at least two for a trie
4931 * so we can't merge this in */
4935 trietype = noper_trietype;
4938 if ( trietype == NOTHING )
4939 trietype = noper_trietype;
4944 } /* end handle mergable triable node */
4946 /* handle unmergable node -
4947 * noper may either be a triable node which can
4948 * not be tried together with the current trie,
4949 * or a non triable node */
4951 /* If last is set and trietype is not
4952 * NOTHING then we have found at least two
4953 * triable branch sequences in a row of a
4954 * similar trietype so we can turn them
4955 * into a trie. If/when we allow NOTHING to
4956 * start a trie sequence this condition
4957 * will be required, and it isn't expensive
4958 * so we leave it in for now. */
4959 if ( trietype && trietype != NOTHING )
4960 make_trie( pRExC_state,
4961 startbranch, first, cur, tail,
4962 count, trietype, depth+1 );
4963 last = NULL; /* note: we clear/update
4964 first, trietype etc below,
4965 so we dont do it here */
4969 && noper_next >= tail
4972 /* noper is triable, so we can start a new
4976 trietype = noper_trietype;
4978 /* if we already saw a first but the
4979 * current node is not triable then we have
4980 * to reset the first information. */
4985 } /* end handle unmergable node */
4986 } /* loop over branches */
4987 DEBUG_TRIE_COMPILE_r({
4988 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4989 Perl_re_indentf( aTHX_ "- %s (%d) <SCAN FINISHED> ",
4990 depth+1, SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur));
4991 Perl_re_printf( aTHX_ "(First==%d, Last==%d, Cur==%d, tt==%s)\n",
4992 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
4993 PL_reg_name[trietype]
4997 if ( last && trietype ) {
4998 if ( trietype != NOTHING ) {
4999 /* the last branch of the sequence was part of
5000 * a trie, so we have to construct it here
5001 * outside of the loop */
5002 made= make_trie( pRExC_state, startbranch,
5003 first, scan, tail, count,
5004 trietype, depth+1 );
5005 #ifdef TRIE_STUDY_OPT
5006 if ( ((made == MADE_EXACT_TRIE &&
5007 startbranch == first)
5008 || ( first_non_open == first )) &&
5010 flags |= SCF_TRIE_RESTUDY;
5011 if ( startbranch == first
5014 RExC_seen &=~REG_TOP_LEVEL_BRANCHES_SEEN;
5019 /* at this point we know whatever we have is a
5020 * NOTHING sequence/branch AND if 'startbranch'
5021 * is 'first' then we can turn the whole thing
5024 if ( startbranch == first ) {
5026 /* the entire thing is a NOTHING sequence,
5027 * something like this: (?:|) So we can
5028 * turn it into a plain NOTHING op. */
5029 DEBUG_TRIE_COMPILE_r({
5030 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
5031 Perl_re_indentf( aTHX_ "- %s (%d) <NOTHING BRANCH SEQUENCE>\n",
5033 SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur));
5036 OP(startbranch)= NOTHING;
5037 NEXT_OFF(startbranch)= tail - startbranch;
5038 for ( opt= startbranch + 1; opt < tail ; opt++ )
5042 } /* end if ( last) */
5043 } /* TRIE_MAXBUF is non zero */
5048 else if ( code == BRANCHJ ) { /* single branch is optimized. */
5049 scan = NEXTOPER(NEXTOPER(scan));
5050 } else /* single branch is optimized. */
5051 scan = NEXTOPER(scan);
5053 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB) {
5055 regnode *start = NULL;
5056 regnode *end = NULL;
5057 U32 my_recursed_depth= recursed_depth;
5059 if (OP(scan) != SUSPEND) { /* GOSUB */
5060 /* Do setup, note this code has side effects beyond
5061 * the rest of this block. Specifically setting
5062 * RExC_recurse[] must happen at least once during
5065 RExC_recurse[ARG2L(scan)] = scan;
5066 start = REGNODE_p(RExC_open_parens[paren]);
5067 end = REGNODE_p(RExC_close_parens[paren]);
5069 /* NOTE we MUST always execute the above code, even
5070 * if we do nothing with a GOSUB */
5072 ( flags & SCF_IN_DEFINE )
5075 (is_inf_internal || is_inf || (data && data->flags & SF_IS_INF))
5077 ( (flags & (SCF_DO_STCLASS | SCF_DO_SUBSTR)) == 0 )
5080 /* no need to do anything here if we are in a define. */
5081 /* or we are after some kind of infinite construct
5082 * so we can skip recursing into this item.
5083 * Since it is infinite we will not change the maxlen
5084 * or delta, and if we miss something that might raise
5085 * the minlen it will merely pessimise a little.
5087 * Iow /(?(DEFINE)(?<foo>foo|food))a+(?&foo)/
5088 * might result in a minlen of 1 and not of 4,
5089 * but this doesn't make us mismatch, just try a bit
5090 * harder than we should.
5092 scan= regnext(scan);
5099 !PAREN_TEST(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes), paren)
5101 /* it is quite possible that there are more efficient ways
5102 * to do this. We maintain a bitmap per level of recursion
5103 * of which patterns we have entered so we can detect if a
5104 * pattern creates a possible infinite loop. When we
5105 * recurse down a level we copy the previous levels bitmap
5106 * down. When we are at recursion level 0 we zero the top
5107 * level bitmap. It would be nice to implement a different
5108 * more efficient way of doing this. In particular the top
5109 * level bitmap may be unnecessary.
5111 if (!recursed_depth) {
5112 Zero(RExC_study_chunk_recursed, RExC_study_chunk_recursed_bytes, U8);
5114 Copy(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes),
5115 RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes),
5116 RExC_study_chunk_recursed_bytes, U8);
5118 /* we havent recursed into this paren yet, so recurse into it */
5119 DEBUG_STUDYDATA("gosub-set", data, depth, is_inf);
5120 PAREN_SET(RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes), paren);
5121 my_recursed_depth= recursed_depth + 1;
5123 DEBUG_STUDYDATA("gosub-inf", data, depth, is_inf);
5124 /* some form of infinite recursion, assume infinite length
5126 if (flags & SCF_DO_SUBSTR) {
5127 scan_commit(pRExC_state, data, minlenp, is_inf);
5128 data->cur_is_floating = 1;
5130 is_inf = is_inf_internal = 1;
5131 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5132 ssc_anything(data->start_class);
5133 flags &= ~SCF_DO_STCLASS;
5135 start= NULL; /* reset start so we dont recurse later on. */
5140 end = regnext(scan);
5143 scan_frame *newframe;
5145 if (!RExC_frame_last) {
5146 Newxz(newframe, 1, scan_frame);
5147 SAVEDESTRUCTOR_X(S_unwind_scan_frames, newframe);
5148 RExC_frame_head= newframe;
5150 } else if (!RExC_frame_last->next_frame) {
5151 Newxz(newframe, 1, scan_frame);
5152 RExC_frame_last->next_frame= newframe;
5153 newframe->prev_frame= RExC_frame_last;
5156 newframe= RExC_frame_last->next_frame;
5158 RExC_frame_last= newframe;
5160 newframe->next_regnode = regnext(scan);
5161 newframe->last_regnode = last;
5162 newframe->stopparen = stopparen;
5163 newframe->prev_recursed_depth = recursed_depth;
5164 newframe->this_prev_frame= frame;
5166 DEBUG_STUDYDATA("frame-new", data, depth, is_inf);
5167 DEBUG_PEEP("fnew", scan, depth, flags);
5174 recursed_depth= my_recursed_depth;
5179 else if ( OP(scan) == EXACT
5180 || OP(scan) == EXACT_ONLY8
5181 || OP(scan) == EXACTL)
5183 SSize_t l = STR_LEN(scan);
5187 const U8 * const s = (U8*)STRING(scan);
5188 uc = utf8_to_uvchr_buf(s, s + l, NULL);
5189 l = utf8_length(s, s + l);
5191 uc = *((U8*)STRING(scan));
5194 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
5195 /* The code below prefers earlier match for fixed
5196 offset, later match for variable offset. */
5197 if (data->last_end == -1) { /* Update the start info. */
5198 data->last_start_min = data->pos_min;
5199 data->last_start_max = is_inf
5200 ? SSize_t_MAX : data->pos_min + data->pos_delta;
5202 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
5204 SvUTF8_on(data->last_found);
5206 SV * const sv = data->last_found;
5207 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
5208 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5209 if (mg && mg->mg_len >= 0)
5210 mg->mg_len += utf8_length((U8*)STRING(scan),
5211 (U8*)STRING(scan)+STR_LEN(scan));
5213 data->last_end = data->pos_min + l;
5214 data->pos_min += l; /* As in the first entry. */
5215 data->flags &= ~SF_BEFORE_EOL;
5218 /* ANDing the code point leaves at most it, and not in locale, and
5219 * can't match null string */
5220 if (flags & SCF_DO_STCLASS_AND) {
5221 ssc_cp_and(data->start_class, uc);
5222 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5223 ssc_clear_locale(data->start_class);
5225 else if (flags & SCF_DO_STCLASS_OR) {
5226 ssc_add_cp(data->start_class, uc);
5227 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5229 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5230 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5232 flags &= ~SCF_DO_STCLASS;
5234 else if (PL_regkind[OP(scan)] == EXACT) {
5235 /* But OP != EXACT!, so is EXACTFish */
5236 SSize_t l = STR_LEN(scan);
5237 const U8 * s = (U8*)STRING(scan);
5239 /* Search for fixed substrings supports EXACT only. */
5240 if (flags & SCF_DO_SUBSTR) {
5242 scan_commit(pRExC_state, data, minlenp, is_inf);
5245 l = utf8_length(s, s + l);
5247 if (unfolded_multi_char) {
5248 RExC_seen |= REG_UNFOLDED_MULTI_SEEN;
5250 min += l - min_subtract;
5252 delta += min_subtract;
5253 if (flags & SCF_DO_SUBSTR) {
5254 data->pos_min += l - min_subtract;
5255 if (data->pos_min < 0) {
5258 data->pos_delta += min_subtract;
5260 data->cur_is_floating = 1; /* float */
5264 if (flags & SCF_DO_STCLASS) {
5265 SV* EXACTF_invlist = _make_exactf_invlist(pRExC_state, scan);
5267 assert(EXACTF_invlist);
5268 if (flags & SCF_DO_STCLASS_AND) {
5269 if (OP(scan) != EXACTFL)
5270 ssc_clear_locale(data->start_class);
5271 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5272 ANYOF_POSIXL_ZERO(data->start_class);
5273 ssc_intersection(data->start_class, EXACTF_invlist, FALSE);
5275 else { /* SCF_DO_STCLASS_OR */
5276 ssc_union(data->start_class, EXACTF_invlist, FALSE);
5277 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5279 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5280 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5282 flags &= ~SCF_DO_STCLASS;
5283 SvREFCNT_dec(EXACTF_invlist);
5286 else if (REGNODE_VARIES(OP(scan))) {
5287 SSize_t mincount, maxcount, minnext, deltanext, pos_before = 0;
5288 I32 fl = 0, f = flags;
5289 regnode * const oscan = scan;
5290 regnode_ssc this_class;
5291 regnode_ssc *oclass = NULL;
5292 I32 next_is_eval = 0;
5294 switch (PL_regkind[OP(scan)]) {
5295 case WHILEM: /* End of (?:...)* . */
5296 scan = NEXTOPER(scan);
5299 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
5300 next = NEXTOPER(scan);
5301 if ( OP(next) == EXACT
5302 || OP(next) == EXACT_ONLY8
5303 || OP(next) == EXACTL
5304 || (flags & SCF_DO_STCLASS))
5307 maxcount = REG_INFTY;
5308 next = regnext(scan);
5309 scan = NEXTOPER(scan);
5313 if (flags & SCF_DO_SUBSTR)
5318 next = NEXTOPER(scan);
5320 /* This temporary node can now be turned into EXACTFU, and
5321 * must, as regexec.c doesn't handle it */
5322 if (OP(next) == EXACTFU_S_EDGE) {
5326 if ( STR_LEN(next) == 1
5327 && isALPHA_A(* STRING(next))
5328 && ( OP(next) == EXACTFAA
5329 || ( OP(next) == EXACTFU
5330 && ! HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(* STRING(next)))))
5332 /* These differ in just one bit */
5333 U8 mask = ~ ('A' ^ 'a');
5335 assert(isALPHA_A(* STRING(next)));
5337 /* Then replace it by an ANYOFM node, with
5338 * the mask set to the complement of the
5339 * bit that differs between upper and lower
5340 * case, and the lowest code point of the
5341 * pair (which the '&' forces) */
5343 ARG_SET(next, *STRING(next) & mask);
5347 if (flags & SCF_DO_STCLASS) {
5349 maxcount = REG_INFTY;
5350 next = regnext(scan);
5351 scan = NEXTOPER(scan);
5354 if (flags & SCF_DO_SUBSTR) {
5355 scan_commit(pRExC_state, data, minlenp, is_inf);
5356 /* Cannot extend fixed substrings */
5357 data->cur_is_floating = 1; /* float */
5359 is_inf = is_inf_internal = 1;
5360 scan = regnext(scan);
5361 goto optimize_curly_tail;
5363 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
5364 && (scan->flags == stopparen))
5369 mincount = ARG1(scan);
5370 maxcount = ARG2(scan);
5372 next = regnext(scan);
5373 if (OP(scan) == CURLYX) {
5374 I32 lp = (data ? *(data->last_closep) : 0);
5375 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
5377 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
5378 next_is_eval = (OP(scan) == EVAL);
5380 if (flags & SCF_DO_SUBSTR) {
5382 scan_commit(pRExC_state, data, minlenp, is_inf);
5383 /* Cannot extend fixed substrings */
5384 pos_before = data->pos_min;
5388 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
5390 data->flags |= SF_IS_INF;
5392 if (flags & SCF_DO_STCLASS) {
5393 ssc_init(pRExC_state, &this_class);
5394 oclass = data->start_class;
5395 data->start_class = &this_class;
5396 f |= SCF_DO_STCLASS_AND;
5397 f &= ~SCF_DO_STCLASS_OR;
5399 /* Exclude from super-linear cache processing any {n,m}
5400 regops for which the combination of input pos and regex
5401 pos is not enough information to determine if a match
5404 For example, in the regex /foo(bar\s*){4,8}baz/ with the
5405 regex pos at the \s*, the prospects for a match depend not
5406 only on the input position but also on how many (bar\s*)
5407 repeats into the {4,8} we are. */
5408 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
5409 f &= ~SCF_WHILEM_VISITED_POS;
5411 /* This will finish on WHILEM, setting scan, or on NULL: */
5412 /* recurse study_chunk() on loop bodies */
5413 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
5414 last, data, stopparen, recursed_depth, NULL,
5416 ? (f & ~SCF_DO_SUBSTR)
5420 if (flags & SCF_DO_STCLASS)
5421 data->start_class = oclass;
5422 if (mincount == 0 || minnext == 0) {
5423 if (flags & SCF_DO_STCLASS_OR) {
5424 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5426 else if (flags & SCF_DO_STCLASS_AND) {
5427 /* Switch to OR mode: cache the old value of
5428 * data->start_class */
5430 StructCopy(data->start_class, and_withp, regnode_ssc);
5431 flags &= ~SCF_DO_STCLASS_AND;
5432 StructCopy(&this_class, data->start_class, regnode_ssc);
5433 flags |= SCF_DO_STCLASS_OR;
5434 ANYOF_FLAGS(data->start_class)
5435 |= SSC_MATCHES_EMPTY_STRING;
5437 } else { /* Non-zero len */
5438 if (flags & SCF_DO_STCLASS_OR) {
5439 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5440 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5442 else if (flags & SCF_DO_STCLASS_AND)
5443 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5444 flags &= ~SCF_DO_STCLASS;
5446 if (!scan) /* It was not CURLYX, but CURLY. */
5448 if (((flags & (SCF_TRIE_DOING_RESTUDY|SCF_DO_SUBSTR))==SCF_DO_SUBSTR)
5449 /* ? quantifier ok, except for (?{ ... }) */
5450 && (next_is_eval || !(mincount == 0 && maxcount == 1))
5451 && (minnext == 0) && (deltanext == 0)
5452 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
5453 && maxcount <= REG_INFTY/3) /* Complement check for big
5456 _WARN_HELPER(RExC_precomp_end, packWARN(WARN_REGEXP),
5457 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP),
5458 "Quantifier unexpected on zero-length expression "
5459 "in regex m/%" UTF8f "/",
5460 UTF8fARG(UTF, RExC_precomp_end - RExC_precomp,
5464 min += minnext * mincount;
5465 is_inf_internal |= deltanext == SSize_t_MAX
5466 || (maxcount == REG_INFTY && minnext + deltanext > 0);
5467 is_inf |= is_inf_internal;
5469 delta = SSize_t_MAX;
5471 delta += (minnext + deltanext) * maxcount
5472 - minnext * mincount;
5474 /* Try powerful optimization CURLYX => CURLYN. */
5475 if ( OP(oscan) == CURLYX && data
5476 && data->flags & SF_IN_PAR
5477 && !(data->flags & SF_HAS_EVAL)
5478 && !deltanext && minnext == 1 ) {
5479 /* Try to optimize to CURLYN. */
5480 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
5481 regnode * const nxt1 = nxt;
5488 if (!REGNODE_SIMPLE(OP(nxt))
5489 && !(PL_regkind[OP(nxt)] == EXACT
5490 && STR_LEN(nxt) == 1))
5496 if (OP(nxt) != CLOSE)
5498 if (RExC_open_parens) {
5501 RExC_open_parens[ARG(nxt1)] = REGNODE_OFFSET(oscan);
5504 RExC_close_parens[ARG(nxt1)] = REGNODE_OFFSET(nxt) + 2;
5506 /* Now we know that nxt2 is the only contents: */
5507 oscan->flags = (U8)ARG(nxt);
5509 OP(nxt1) = NOTHING; /* was OPEN. */
5512 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5513 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
5514 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
5515 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5516 OP(nxt + 1) = OPTIMIZED; /* was count. */
5517 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
5522 /* Try optimization CURLYX => CURLYM. */
5523 if ( OP(oscan) == CURLYX && data
5524 && !(data->flags & SF_HAS_PAR)
5525 && !(data->flags & SF_HAS_EVAL)
5526 && !deltanext /* atom is fixed width */
5527 && minnext != 0 /* CURLYM can't handle zero width */
5529 /* Nor characters whose fold at run-time may be
5530 * multi-character */
5531 && ! (RExC_seen & REG_UNFOLDED_MULTI_SEEN)
5533 /* XXXX How to optimize if data == 0? */
5534 /* Optimize to a simpler form. */
5535 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
5539 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
5540 && (OP(nxt2) != WHILEM))
5542 OP(nxt2) = SUCCEED; /* Whas WHILEM */
5543 /* Need to optimize away parenths. */
5544 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
5545 /* Set the parenth number. */
5546 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
5548 oscan->flags = (U8)ARG(nxt);
5549 if (RExC_open_parens) {
5551 RExC_open_parens[ARG(nxt1)] = REGNODE_OFFSET(oscan);
5554 RExC_close_parens[ARG(nxt1)] = REGNODE_OFFSET(nxt2)
5557 OP(nxt1) = OPTIMIZED; /* was OPEN. */
5558 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5561 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5562 OP(nxt + 1) = OPTIMIZED; /* was count. */
5563 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
5564 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
5567 while ( nxt1 && (OP(nxt1) != WHILEM)) {
5568 regnode *nnxt = regnext(nxt1);
5570 if (reg_off_by_arg[OP(nxt1)])
5571 ARG_SET(nxt1, nxt2 - nxt1);
5572 else if (nxt2 - nxt1 < U16_MAX)
5573 NEXT_OFF(nxt1) = nxt2 - nxt1;
5575 OP(nxt) = NOTHING; /* Cannot beautify */
5580 /* Optimize again: */
5581 /* recurse study_chunk() on optimised CURLYX => CURLYM */
5582 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
5583 NULL, stopparen, recursed_depth, NULL, 0,
5589 else if ((OP(oscan) == CURLYX)
5590 && (flags & SCF_WHILEM_VISITED_POS)
5591 /* See the comment on a similar expression above.
5592 However, this time it's not a subexpression
5593 we care about, but the expression itself. */
5594 && (maxcount == REG_INFTY)
5596 /* This stays as CURLYX, we can put the count/of pair. */
5597 /* Find WHILEM (as in regexec.c) */
5598 regnode *nxt = oscan + NEXT_OFF(oscan);
5600 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
5602 nxt = PREVOPER(nxt);
5603 if (nxt->flags & 0xf) {
5604 /* we've already set whilem count on this node */
5605 } else if (++data->whilem_c < 16) {
5606 assert(data->whilem_c <= RExC_whilem_seen);
5607 nxt->flags = (U8)(data->whilem_c
5608 | (RExC_whilem_seen << 4)); /* On WHILEM */
5611 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
5613 if (flags & SCF_DO_SUBSTR) {
5614 SV *last_str = NULL;
5615 STRLEN last_chrs = 0;
5616 int counted = mincount != 0;
5618 if (data->last_end > 0 && mincount != 0) { /* Ends with a
5620 SSize_t b = pos_before >= data->last_start_min
5621 ? pos_before : data->last_start_min;
5623 const char * const s = SvPV_const(data->last_found, l);
5624 SSize_t old = b - data->last_start_min;
5628 old = utf8_hop_forward((U8*)s, old,
5629 (U8 *) SvEND(data->last_found))
5632 /* Get the added string: */
5633 last_str = newSVpvn_utf8(s + old, l, UTF);
5634 last_chrs = UTF ? utf8_length((U8*)(s + old),
5635 (U8*)(s + old + l)) : l;
5636 if (deltanext == 0 && pos_before == b) {
5637 /* What was added is a constant string */
5640 SvGROW(last_str, (mincount * l) + 1);
5641 repeatcpy(SvPVX(last_str) + l,
5642 SvPVX_const(last_str), l,
5644 SvCUR_set(last_str, SvCUR(last_str) * mincount);
5645 /* Add additional parts. */
5646 SvCUR_set(data->last_found,
5647 SvCUR(data->last_found) - l);
5648 sv_catsv(data->last_found, last_str);
5650 SV * sv = data->last_found;
5652 SvUTF8(sv) && SvMAGICAL(sv) ?
5653 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5654 if (mg && mg->mg_len >= 0)
5655 mg->mg_len += last_chrs * (mincount-1);
5657 last_chrs *= mincount;
5658 data->last_end += l * (mincount - 1);
5661 /* start offset must point into the last copy */
5662 data->last_start_min += minnext * (mincount - 1);
5663 data->last_start_max =
5666 : data->last_start_max +
5667 (maxcount - 1) * (minnext + data->pos_delta);
5670 /* It is counted once already... */
5671 data->pos_min += minnext * (mincount - counted);
5673 Perl_re_printf( aTHX_ "counted=%" UVuf " deltanext=%" UVuf
5674 " SSize_t_MAX=%" UVuf " minnext=%" UVuf
5675 " maxcount=%" UVuf " mincount=%" UVuf "\n",
5676 (UV)counted, (UV)deltanext, (UV)SSize_t_MAX, (UV)minnext, (UV)maxcount,
5678 if (deltanext != SSize_t_MAX)
5679 Perl_re_printf( aTHX_ "LHS=%" UVuf " RHS=%" UVuf "\n",
5680 (UV)(-counted * deltanext + (minnext + deltanext) * maxcount
5681 - minnext * mincount), (UV)(SSize_t_MAX - data->pos_delta));
5683 if (deltanext == SSize_t_MAX
5684 || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= SSize_t_MAX - data->pos_delta)
5685 data->pos_delta = SSize_t_MAX;
5687 data->pos_delta += - counted * deltanext +
5688 (minnext + deltanext) * maxcount - minnext * mincount;
5689 if (mincount != maxcount) {
5690 /* Cannot extend fixed substrings found inside
5692 scan_commit(pRExC_state, data, minlenp, is_inf);
5693 if (mincount && last_str) {
5694 SV * const sv = data->last_found;
5695 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
5696 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5700 sv_setsv(sv, last_str);
5701 data->last_end = data->pos_min;
5702 data->last_start_min = data->pos_min - last_chrs;
5703 data->last_start_max = is_inf
5705 : data->pos_min + data->pos_delta - last_chrs;
5707 data->cur_is_floating = 1; /* float */
5709 SvREFCNT_dec(last_str);
5711 if (data && (fl & SF_HAS_EVAL))
5712 data->flags |= SF_HAS_EVAL;
5713 optimize_curly_tail:
5714 if (OP(oscan) != CURLYX) {
5715 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
5717 NEXT_OFF(oscan) += NEXT_OFF(next);
5723 Perl_croak(aTHX_ "panic: unexpected varying REx opcode %d",
5728 if (flags & SCF_DO_SUBSTR) {
5729 /* Cannot expect anything... */
5730 scan_commit(pRExC_state, data, minlenp, is_inf);
5731 data->cur_is_floating = 1; /* float */
5733 is_inf = is_inf_internal = 1;
5734 if (flags & SCF_DO_STCLASS_OR) {
5735 if (OP(scan) == CLUMP) {
5736 /* Actually is any start char, but very few code points
5737 * aren't start characters */
5738 ssc_match_all_cp(data->start_class);
5741 ssc_anything(data->start_class);
5744 flags &= ~SCF_DO_STCLASS;
5748 else if (OP(scan) == LNBREAK) {
5749 if (flags & SCF_DO_STCLASS) {
5750 if (flags & SCF_DO_STCLASS_AND) {
5751 ssc_intersection(data->start_class,
5752 PL_XPosix_ptrs[_CC_VERTSPACE], FALSE);
5753 ssc_clear_locale(data->start_class);
5754 ANYOF_FLAGS(data->start_class)
5755 &= ~SSC_MATCHES_EMPTY_STRING;
5757 else if (flags & SCF_DO_STCLASS_OR) {
5758 ssc_union(data->start_class,
5759 PL_XPosix_ptrs[_CC_VERTSPACE],
5761 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5763 /* See commit msg for
5764 * 749e076fceedeb708a624933726e7989f2302f6a */
5765 ANYOF_FLAGS(data->start_class)
5766 &= ~SSC_MATCHES_EMPTY_STRING;
5768 flags &= ~SCF_DO_STCLASS;
5771 if (delta != SSize_t_MAX)
5772 delta++; /* Because of the 2 char string cr-lf */
5773 if (flags & SCF_DO_SUBSTR) {
5774 /* Cannot expect anything... */
5775 scan_commit(pRExC_state, data, minlenp, is_inf);
5777 if (data->pos_delta != SSize_t_MAX) {
5778 data->pos_delta += 1;
5780 data->cur_is_floating = 1; /* float */
5783 else if (REGNODE_SIMPLE(OP(scan))) {
5785 if (flags & SCF_DO_SUBSTR) {
5786 scan_commit(pRExC_state, data, minlenp, is_inf);
5790 if (flags & SCF_DO_STCLASS) {
5792 SV* my_invlist = NULL;
5795 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5796 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5798 /* Some of the logic below assumes that switching
5799 locale on will only add false positives. */
5804 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d",
5808 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5809 ssc_match_all_cp(data->start_class);
5814 SV* REG_ANY_invlist = _new_invlist(2);
5815 REG_ANY_invlist = add_cp_to_invlist(REG_ANY_invlist,
5817 if (flags & SCF_DO_STCLASS_OR) {
5818 ssc_union(data->start_class,
5820 TRUE /* TRUE => invert, hence all but \n
5824 else if (flags & SCF_DO_STCLASS_AND) {
5825 ssc_intersection(data->start_class,
5827 TRUE /* TRUE => invert */
5829 ssc_clear_locale(data->start_class);
5831 SvREFCNT_dec_NN(REG_ANY_invlist);
5840 if (flags & SCF_DO_STCLASS_AND)
5841 ssc_and(pRExC_state, data->start_class,
5842 (regnode_charclass *) scan);
5844 ssc_or(pRExC_state, data->start_class,
5845 (regnode_charclass *) scan);
5851 SV* cp_list = get_ANYOFM_contents(scan);
5853 if (flags & SCF_DO_STCLASS_OR) {
5854 ssc_union(data->start_class, cp_list, invert);
5856 else if (flags & SCF_DO_STCLASS_AND) {
5857 ssc_intersection(data->start_class, cp_list, invert);
5860 SvREFCNT_dec_NN(cp_list);
5869 namedclass = classnum_to_namedclass(FLAGS(scan)) + invert;
5870 if (flags & SCF_DO_STCLASS_AND) {
5871 bool was_there = cBOOL(
5872 ANYOF_POSIXL_TEST(data->start_class,
5874 ANYOF_POSIXL_ZERO(data->start_class);
5875 if (was_there) { /* Do an AND */
5876 ANYOF_POSIXL_SET(data->start_class, namedclass);
5878 /* No individual code points can now match */
5879 data->start_class->invlist
5880 = sv_2mortal(_new_invlist(0));
5883 int complement = namedclass + ((invert) ? -1 : 1);
5885 assert(flags & SCF_DO_STCLASS_OR);
5887 /* If the complement of this class was already there,
5888 * the result is that they match all code points,
5889 * (\d + \D == everything). Remove the classes from
5890 * future consideration. Locale is not relevant in
5892 if (ANYOF_POSIXL_TEST(data->start_class, complement)) {
5893 ssc_match_all_cp(data->start_class);
5894 ANYOF_POSIXL_CLEAR(data->start_class, namedclass);
5895 ANYOF_POSIXL_CLEAR(data->start_class, complement);
5897 else { /* The usual case; just add this class to the
5899 ANYOF_POSIXL_SET(data->start_class, namedclass);
5904 case NPOSIXA: /* For these, we always know the exact set of
5909 my_invlist = invlist_clone(PL_Posix_ptrs[FLAGS(scan)], NULL);
5910 goto join_posix_and_ascii;
5918 my_invlist = invlist_clone(PL_XPosix_ptrs[FLAGS(scan)], NULL);
5920 /* NPOSIXD matches all upper Latin1 code points unless the
5921 * target string being matched is UTF-8, which is
5922 * unknowable until match time. Since we are going to
5923 * invert, we want to get rid of all of them so that the
5924 * inversion will match all */
5925 if (OP(scan) == NPOSIXD) {
5926 _invlist_subtract(my_invlist, PL_UpperLatin1,
5930 join_posix_and_ascii:
5932 if (flags & SCF_DO_STCLASS_AND) {
5933 ssc_intersection(data->start_class, my_invlist, invert);
5934 ssc_clear_locale(data->start_class);
5937 assert(flags & SCF_DO_STCLASS_OR);
5938 ssc_union(data->start_class, my_invlist, invert);
5940 SvREFCNT_dec(my_invlist);
5942 if (flags & SCF_DO_STCLASS_OR)
5943 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5944 flags &= ~SCF_DO_STCLASS;
5947 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
5948 data->flags |= (OP(scan) == MEOL
5951 scan_commit(pRExC_state, data, minlenp, is_inf);
5954 else if ( PL_regkind[OP(scan)] == BRANCHJ
5955 /* Lookbehind, or need to calculate parens/evals/stclass: */
5956 && (scan->flags || data || (flags & SCF_DO_STCLASS))
5957 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM))
5959 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5960 || OP(scan) == UNLESSM )
5962 /* Negative Lookahead/lookbehind
5963 In this case we can't do fixed string optimisation.
5966 SSize_t deltanext, minnext, fake = 0;
5971 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5973 data_fake.whilem_c = data->whilem_c;
5974 data_fake.last_closep = data->last_closep;
5977 data_fake.last_closep = &fake;
5978 data_fake.pos_delta = delta;
5979 if ( flags & SCF_DO_STCLASS && !scan->flags
5980 && OP(scan) == IFMATCH ) { /* Lookahead */
5981 ssc_init(pRExC_state, &intrnl);
5982 data_fake.start_class = &intrnl;
5983 f |= SCF_DO_STCLASS_AND;
5985 if (flags & SCF_WHILEM_VISITED_POS)
5986 f |= SCF_WHILEM_VISITED_POS;
5987 next = regnext(scan);
5988 nscan = NEXTOPER(NEXTOPER(scan));
5990 /* recurse study_chunk() for lookahead body */
5991 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
5992 last, &data_fake, stopparen,
5993 recursed_depth, NULL, f, depth+1);
5996 || deltanext > (I32) U8_MAX
5997 || minnext > (I32)U8_MAX
5998 || minnext + deltanext > (I32)U8_MAX)
6000 FAIL2("Lookbehind longer than %" UVuf " not implemented",
6004 /* The 'next_off' field has been repurposed to count the
6005 * additional starting positions to try beyond the initial
6006 * one. (This leaves it at 0 for non-variable length
6007 * matches to avoid breakage for those not using this
6010 scan->next_off = deltanext;
6011 ckWARNexperimental(RExC_parse,
6012 WARN_EXPERIMENTAL__VLB,
6013 "Variable length lookbehind is experimental");
6015 scan->flags = (U8)minnext + deltanext;
6018 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
6020 if (data_fake.flags & SF_HAS_EVAL)
6021 data->flags |= SF_HAS_EVAL;
6022 data->whilem_c = data_fake.whilem_c;
6024 if (f & SCF_DO_STCLASS_AND) {
6025 if (flags & SCF_DO_STCLASS_OR) {
6026 /* OR before, AND after: ideally we would recurse with
6027 * data_fake to get the AND applied by study of the
6028 * remainder of the pattern, and then derecurse;
6029 * *** HACK *** for now just treat as "no information".
6030 * See [perl #56690].
6032 ssc_init(pRExC_state, data->start_class);
6034 /* AND before and after: combine and continue. These
6035 * assertions are zero-length, so can match an EMPTY
6037 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
6038 ANYOF_FLAGS(data->start_class)
6039 |= SSC_MATCHES_EMPTY_STRING;
6043 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
6045 /* Positive Lookahead/lookbehind
6046 In this case we can do fixed string optimisation,
6047 but we must be careful about it. Note in the case of
6048 lookbehind the positions will be offset by the minimum
6049 length of the pattern, something we won't know about
6050 until after the recurse.
6052 SSize_t deltanext, fake = 0;
6056 /* We use SAVEFREEPV so that when the full compile
6057 is finished perl will clean up the allocated
6058 minlens when it's all done. This way we don't
6059 have to worry about freeing them when we know
6060 they wont be used, which would be a pain.
6063 Newx( minnextp, 1, SSize_t );
6064 SAVEFREEPV(minnextp);
6067 StructCopy(data, &data_fake, scan_data_t);
6068 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
6071 scan_commit(pRExC_state, &data_fake, minlenp, is_inf);
6072 data_fake.last_found=newSVsv(data->last_found);
6076 data_fake.last_closep = &fake;
6077 data_fake.flags = 0;
6078 data_fake.substrs[0].flags = 0;
6079 data_fake.substrs[1].flags = 0;
6080 data_fake.pos_delta = delta;
6082 data_fake.flags |= SF_IS_INF;
6083 if ( flags & SCF_DO_STCLASS && !scan->flags
6084 && OP(scan) == IFMATCH ) { /* Lookahead */
6085 ssc_init(pRExC_state, &intrnl);
6086 data_fake.start_class = &intrnl;
6087 f |= SCF_DO_STCLASS_AND;
6089 if (flags & SCF_WHILEM_VISITED_POS)
6090 f |= SCF_WHILEM_VISITED_POS;
6091 next = regnext(scan);
6092 nscan = NEXTOPER(NEXTOPER(scan));
6094 /* positive lookahead study_chunk() recursion */
6095 *minnextp = study_chunk(pRExC_state, &nscan, minnextp,
6096 &deltanext, last, &data_fake,
6097 stopparen, recursed_depth, NULL,
6100 assert(0); /* This code has never been tested since this
6101 is normally not compiled */
6103 || deltanext > (I32) U8_MAX
6104 || *minnextp > (I32)U8_MAX
6105 || *minnextp + deltanext > (I32)U8_MAX)
6107 FAIL2("Lookbehind longer than %" UVuf " not implemented",
6112 scan->next_off = deltanext;
6114 scan->flags = (U8)*minnextp + deltanext;
6119 if (f & SCF_DO_STCLASS_AND) {
6120 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
6121 ANYOF_FLAGS(data->start_class) |= SSC_MATCHES_EMPTY_STRING;
6124 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
6126 if (data_fake.flags & SF_HAS_EVAL)
6127 data->flags |= SF_HAS_EVAL;
6128 data->whilem_c = data_fake.whilem_c;
6129 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
6131 if (RExC_rx->minlen<*minnextp)
6132 RExC_rx->minlen=*minnextp;
6133 scan_commit(pRExC_state, &data_fake, minnextp, is_inf);
6134 SvREFCNT_dec_NN(data_fake.last_found);
6136 for (i = 0; i < 2; i++) {
6137 if (data_fake.substrs[i].minlenp != minlenp) {
6138 data->substrs[i].min_offset =
6139 data_fake.substrs[i].min_offset;
6140 data->substrs[i].max_offset =
6141 data_fake.substrs[i].max_offset;
6142 data->substrs[i].minlenp =
6143 data_fake.substrs[i].minlenp;
6144 data->substrs[i].lookbehind += scan->flags;
6153 else if (OP(scan) == OPEN) {
6154 if (stopparen != (I32)ARG(scan))
6157 else if (OP(scan) == CLOSE) {
6158 if (stopparen == (I32)ARG(scan)) {
6161 if ((I32)ARG(scan) == is_par) {
6162 next = regnext(scan);
6164 if ( next && (OP(next) != WHILEM) && next < last)
6165 is_par = 0; /* Disable optimization */
6168 *(data->last_closep) = ARG(scan);
6170 else if (OP(scan) == EVAL) {
6172 data->flags |= SF_HAS_EVAL;
6174 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
6175 if (flags & SCF_DO_SUBSTR) {
6176 scan_commit(pRExC_state, data, minlenp, is_inf);
6177 flags &= ~SCF_DO_SUBSTR;
6179 if (data && OP(scan)==ACCEPT) {
6180 data->flags |= SCF_SEEN_ACCEPT;
6185 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
6187 if (flags & SCF_DO_SUBSTR) {
6188 scan_commit(pRExC_state, data, minlenp, is_inf);
6189 data->cur_is_floating = 1; /* float */
6191 is_inf = is_inf_internal = 1;
6192 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
6193 ssc_anything(data->start_class);
6194 flags &= ~SCF_DO_STCLASS;
6196 else if (OP(scan) == GPOS) {
6197 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
6198 !(delta || is_inf || (data && data->pos_delta)))
6200 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
6201 RExC_rx->intflags |= PREGf_ANCH_GPOS;
6202 if (RExC_rx->gofs < (STRLEN)min)
6203 RExC_rx->gofs = min;
6205 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
6209 #ifdef TRIE_STUDY_OPT
6210 #ifdef FULL_TRIE_STUDY
6211 else if (PL_regkind[OP(scan)] == TRIE) {
6212 /* NOTE - There is similar code to this block above for handling
6213 BRANCH nodes on the initial study. If you change stuff here
6215 regnode *trie_node= scan;
6216 regnode *tail= regnext(scan);
6217 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
6218 SSize_t max1 = 0, min1 = SSize_t_MAX;
6221 if (flags & SCF_DO_SUBSTR) { /* XXXX Add !SUSPEND? */
6222 /* Cannot merge strings after this. */
6223 scan_commit(pRExC_state, data, minlenp, is_inf);
6225 if (flags & SCF_DO_STCLASS)
6226 ssc_init_zero(pRExC_state, &accum);
6232 const regnode *nextbranch= NULL;
6235 for ( word=1 ; word <= trie->wordcount ; word++)
6237 SSize_t deltanext=0, minnext=0, f = 0, fake;
6238 regnode_ssc this_class;
6240 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
6242 data_fake.whilem_c = data->whilem_c;
6243 data_fake.last_closep = data->last_closep;
6246 data_fake.last_closep = &fake;
6247 data_fake.pos_delta = delta;
6248 if (flags & SCF_DO_STCLASS) {
6249 ssc_init(pRExC_state, &this_class);
6250 data_fake.start_class = &this_class;
6251 f = SCF_DO_STCLASS_AND;
6253 if (flags & SCF_WHILEM_VISITED_POS)
6254 f |= SCF_WHILEM_VISITED_POS;
6256 if (trie->jump[word]) {
6258 nextbranch = trie_node + trie->jump[0];
6259 scan= trie_node + trie->jump[word];
6260 /* We go from the jump point to the branch that follows
6261 it. Note this means we need the vestigal unused
6262 branches even though they arent otherwise used. */
6263 /* optimise study_chunk() for TRIE */
6264 minnext = study_chunk(pRExC_state, &scan, minlenp,
6265 &deltanext, (regnode *)nextbranch, &data_fake,
6266 stopparen, recursed_depth, NULL, f, depth+1);
6268 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
6269 nextbranch= regnext((regnode*)nextbranch);
6271 if (min1 > (SSize_t)(minnext + trie->minlen))
6272 min1 = minnext + trie->minlen;
6273 if (deltanext == SSize_t_MAX) {
6274 is_inf = is_inf_internal = 1;
6276 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
6277 max1 = minnext + deltanext + trie->maxlen;
6279 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
6281 if (data_fake.flags & SCF_SEEN_ACCEPT) {
6282 if ( stopmin > min + min1)
6283 stopmin = min + min1;
6284 flags &= ~SCF_DO_SUBSTR;
6286 data->flags |= SCF_SEEN_ACCEPT;
6289 if (data_fake.flags & SF_HAS_EVAL)
6290 data->flags |= SF_HAS_EVAL;
6291 data->whilem_c = data_fake.whilem_c;
6293 if (flags & SCF_DO_STCLASS)
6294 ssc_or(pRExC_state, &accum, (regnode_charclass *) &this_class);
6297 if (flags & SCF_DO_SUBSTR) {
6298 data->pos_min += min1;
6299 data->pos_delta += max1 - min1;
6300 if (max1 != min1 || is_inf)
6301 data->cur_is_floating = 1; /* float */
6304 if (delta != SSize_t_MAX) {
6305 if (SSize_t_MAX - (max1 - min1) >= delta)
6306 delta += max1 - min1;
6308 delta = SSize_t_MAX;
6310 if (flags & SCF_DO_STCLASS_OR) {
6311 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &accum);
6313 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6314 flags &= ~SCF_DO_STCLASS;
6317 else if (flags & SCF_DO_STCLASS_AND) {
6319 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
6320 flags &= ~SCF_DO_STCLASS;
6323 /* Switch to OR mode: cache the old value of
6324 * data->start_class */
6326 StructCopy(data->start_class, and_withp, regnode_ssc);
6327 flags &= ~SCF_DO_STCLASS_AND;
6328 StructCopy(&accum, data->start_class, regnode_ssc);
6329 flags |= SCF_DO_STCLASS_OR;
6336 else if (PL_regkind[OP(scan)] == TRIE) {
6337 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
6340 min += trie->minlen;
6341 delta += (trie->maxlen - trie->minlen);
6342 flags &= ~SCF_DO_STCLASS; /* xxx */
6343 if (flags & SCF_DO_SUBSTR) {
6344 /* Cannot expect anything... */
6345 scan_commit(pRExC_state, data, minlenp, is_inf);
6346 data->pos_min += trie->minlen;
6347 data->pos_delta += (trie->maxlen - trie->minlen);
6348 if (trie->maxlen != trie->minlen)
6349 data->cur_is_floating = 1; /* float */
6351 if (trie->jump) /* no more substrings -- for now /grr*/
6352 flags &= ~SCF_DO_SUBSTR;
6354 #endif /* old or new */
6355 #endif /* TRIE_STUDY_OPT */
6357 /* Else: zero-length, ignore. */
6358 scan = regnext(scan);
6363 /* we need to unwind recursion. */
6366 DEBUG_STUDYDATA("frame-end", data, depth, is_inf);
6367 DEBUG_PEEP("fend", scan, depth, flags);
6369 /* restore previous context */
6370 last = frame->last_regnode;
6371 scan = frame->next_regnode;
6372 stopparen = frame->stopparen;
6373 recursed_depth = frame->prev_recursed_depth;
6375 RExC_frame_last = frame->prev_frame;
6376 frame = frame->this_prev_frame;
6377 goto fake_study_recurse;
6381 DEBUG_STUDYDATA("pre-fin", data, depth, is_inf);
6384 *deltap = is_inf_internal ? SSize_t_MAX : delta;
6386 if (flags & SCF_DO_SUBSTR && is_inf)
6387 data->pos_delta = SSize_t_MAX - data->pos_min;
6388 if (is_par > (I32)U8_MAX)
6390 if (is_par && pars==1 && data) {
6391 data->flags |= SF_IN_PAR;
6392 data->flags &= ~SF_HAS_PAR;
6394 else if (pars && data) {
6395 data->flags |= SF_HAS_PAR;
6396 data->flags &= ~SF_IN_PAR;
6398 if (flags & SCF_DO_STCLASS_OR)
6399 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6400 if (flags & SCF_TRIE_RESTUDY)
6401 data->flags |= SCF_TRIE_RESTUDY;
6403 DEBUG_STUDYDATA("post-fin", data, depth, is_inf);
6406 SSize_t final_minlen= min < stopmin ? min : stopmin;
6408 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN)) {
6409 if (final_minlen > SSize_t_MAX - delta)
6410 RExC_maxlen = SSize_t_MAX;
6411 else if (RExC_maxlen < final_minlen + delta)
6412 RExC_maxlen = final_minlen + delta;
6414 return final_minlen;
6416 NOT_REACHED; /* NOTREACHED */
6420 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
6422 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
6424 PERL_ARGS_ASSERT_ADD_DATA;
6426 Renewc(RExC_rxi->data,
6427 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
6428 char, struct reg_data);
6430 Renew(RExC_rxi->data->what, count + n, U8);
6432 Newx(RExC_rxi->data->what, n, U8);
6433 RExC_rxi->data->count = count + n;
6434 Copy(s, RExC_rxi->data->what + count, n, U8);
6438 /*XXX: todo make this not included in a non debugging perl, but appears to be
6439 * used anyway there, in 'use re' */
6440 #ifndef PERL_IN_XSUB_RE
6442 Perl_reginitcolors(pTHX)
6444 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
6446 char *t = savepv(s);
6450 t = strchr(t, '\t');
6456 PL_colors[i] = t = (char *)"";
6461 PL_colors[i++] = (char *)"";
6468 #ifdef TRIE_STUDY_OPT
6469 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
6472 (data.flags & SCF_TRIE_RESTUDY) \
6480 #define CHECK_RESTUDY_GOTO_butfirst
6484 * pregcomp - compile a regular expression into internal code
6486 * Decides which engine's compiler to call based on the hint currently in
6490 #ifndef PERL_IN_XSUB_RE
6492 /* return the currently in-scope regex engine (or the default if none) */
6494 regexp_engine const *
6495 Perl_current_re_engine(pTHX)
6497 if (IN_PERL_COMPILETIME) {
6498 HV * const table = GvHV(PL_hintgv);
6501 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
6502 return &PL_core_reg_engine;
6503 ptr = hv_fetchs(table, "regcomp", FALSE);
6504 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
6505 return &PL_core_reg_engine;
6506 return INT2PTR(regexp_engine*, SvIV(*ptr));
6510 if (!PL_curcop->cop_hints_hash)
6511 return &PL_core_reg_engine;
6512 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
6513 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
6514 return &PL_core_reg_engine;
6515 return INT2PTR(regexp_engine*, SvIV(ptr));
6521 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
6523 regexp_engine const *eng = current_re_engine();
6524 GET_RE_DEBUG_FLAGS_DECL;
6526 PERL_ARGS_ASSERT_PREGCOMP;
6528 /* Dispatch a request to compile a regexp to correct regexp engine. */
6530 Perl_re_printf( aTHX_ "Using engine %" UVxf "\n",
6533 return CALLREGCOMP_ENG(eng, pattern, flags);
6537 /* public(ish) entry point for the perl core's own regex compiling code.
6538 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
6539 * pattern rather than a list of OPs, and uses the internal engine rather
6540 * than the current one */
6543 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
6545 SV *pat = pattern; /* defeat constness! */
6546 PERL_ARGS_ASSERT_RE_COMPILE;
6547 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
6548 #ifdef PERL_IN_XSUB_RE
6551 &PL_core_reg_engine,
6553 NULL, NULL, rx_flags, 0);
6558 S_free_codeblocks(pTHX_ struct reg_code_blocks *cbs)
6562 if (--cbs->refcnt > 0)
6564 for (n = 0; n < cbs->count; n++) {
6565 REGEXP *rx = cbs->cb[n].src_regex;
6567 cbs->cb[n].src_regex = NULL;
6568 SvREFCNT_dec_NN(rx);
6576 static struct reg_code_blocks *
6577 S_alloc_code_blocks(pTHX_ int ncode)
6579 struct reg_code_blocks *cbs;
6580 Newx(cbs, 1, struct reg_code_blocks);
6583 SAVEDESTRUCTOR_X(S_free_codeblocks, cbs);
6585 Newx(cbs->cb, ncode, struct reg_code_block);
6592 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
6593 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
6594 * point to the realloced string and length.
6596 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
6600 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
6601 char **pat_p, STRLEN *plen_p, int num_code_blocks)
6603 U8 *const src = (U8*)*pat_p;
6608 GET_RE_DEBUG_FLAGS_DECL;
6610 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6611 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
6613 /* 1 for each byte + 1 for each byte that expands to two, + trailing NUL */
6614 Newx(dst, *plen_p + variant_under_utf8_count(src, src + *plen_p) + 1, U8);
6617 while (s < *plen_p) {
6618 append_utf8_from_native_byte(src[s], &d);
6620 if (n < num_code_blocks) {
6621 assert(pRExC_state->code_blocks);
6622 if (!do_end && pRExC_state->code_blocks->cb[n].start == s) {
6623 pRExC_state->code_blocks->cb[n].start = d - dst - 1;
6624 assert(*(d - 1) == '(');
6627 else if (do_end && pRExC_state->code_blocks->cb[n].end == s) {
6628 pRExC_state->code_blocks->cb[n].end = d - dst - 1;
6629 assert(*(d - 1) == ')');
6638 *pat_p = (char*) dst;
6640 RExC_orig_utf8 = RExC_utf8 = 1;
6645 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
6646 * while recording any code block indices, and handling overloading,
6647 * nested qr// objects etc. If pat is null, it will allocate a new
6648 * string, or just return the first arg, if there's only one.
6650 * Returns the malloced/updated pat.
6651 * patternp and pat_count is the array of SVs to be concatted;
6652 * oplist is the optional list of ops that generated the SVs;
6653 * recompile_p is a pointer to a boolean that will be set if
6654 * the regex will need to be recompiled.
6655 * delim, if non-null is an SV that will be inserted between each element
6659 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
6660 SV *pat, SV ** const patternp, int pat_count,
6661 OP *oplist, bool *recompile_p, SV *delim)
6665 bool use_delim = FALSE;
6666 bool alloced = FALSE;
6668 /* if we know we have at least two args, create an empty string,
6669 * then concatenate args to that. For no args, return an empty string */
6670 if (!pat && pat_count != 1) {
6676 for (svp = patternp; svp < patternp + pat_count; svp++) {
6679 STRLEN orig_patlen = 0;
6681 SV *msv = use_delim ? delim : *svp;
6682 if (!msv) msv = &PL_sv_undef;
6684 /* if we've got a delimiter, we go round the loop twice for each
6685 * svp slot (except the last), using the delimiter the second
6694 if (SvTYPE(msv) == SVt_PVAV) {
6695 /* we've encountered an interpolated array within
6696 * the pattern, e.g. /...@a..../. Expand the list of elements,
6697 * then recursively append elements.
6698 * The code in this block is based on S_pushav() */
6700 AV *const av = (AV*)msv;
6701 const SSize_t maxarg = AvFILL(av) + 1;
6705 assert(oplist->op_type == OP_PADAV
6706 || oplist->op_type == OP_RV2AV);
6707 oplist = OpSIBLING(oplist);
6710 if (SvRMAGICAL(av)) {
6713 Newx(array, maxarg, SV*);
6715 for (i=0; i < maxarg; i++) {
6716 SV ** const svp = av_fetch(av, i, FALSE);
6717 array[i] = svp ? *svp : &PL_sv_undef;
6721 array = AvARRAY(av);
6723 pat = S_concat_pat(aTHX_ pRExC_state, pat,
6724 array, maxarg, NULL, recompile_p,
6726 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
6732 /* we make the assumption here that each op in the list of
6733 * op_siblings maps to one SV pushed onto the stack,
6734 * except for code blocks, with have both an OP_NULL and
6736 * This allows us to match up the list of SVs against the
6737 * list of OPs to find the next code block.
6739 * Note that PUSHMARK PADSV PADSV ..
6741 * PADRANGE PADSV PADSV ..
6742 * so the alignment still works. */
6745 if (oplist->op_type == OP_NULL
6746 && (oplist->op_flags & OPf_SPECIAL))
6748 assert(n < pRExC_state->code_blocks->count);
6749 pRExC_state->code_blocks->cb[n].start = pat ? SvCUR(pat) : 0;
6750 pRExC_state->code_blocks->cb[n].block = oplist;
6751 pRExC_state->code_blocks->cb[n].src_regex = NULL;
6754 oplist = OpSIBLING(oplist); /* skip CONST */
6757 oplist = OpSIBLING(oplist);;
6760 /* apply magic and QR overloading to arg */
6763 if (SvROK(msv) && SvAMAGIC(msv)) {
6764 SV *sv = AMG_CALLunary(msv, regexp_amg);
6768 if (SvTYPE(sv) != SVt_REGEXP)
6769 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
6774 /* try concatenation overload ... */
6775 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
6776 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
6779 /* overloading involved: all bets are off over literal
6780 * code. Pretend we haven't seen it */
6782 pRExC_state->code_blocks->count -= n;
6786 /* ... or failing that, try "" overload */
6787 while (SvAMAGIC(msv)
6788 && (sv = AMG_CALLunary(msv, string_amg))
6792 && SvRV(msv) == SvRV(sv))
6797 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
6801 /* this is a partially unrolled
6802 * sv_catsv_nomg(pat, msv);
6803 * that allows us to adjust code block indices if
6806 char *dst = SvPV_force_nomg(pat, dlen);
6808 if (SvUTF8(msv) && !SvUTF8(pat)) {
6809 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
6810 sv_setpvn(pat, dst, dlen);
6813 sv_catsv_nomg(pat, msv);
6817 /* We have only one SV to process, but we need to verify
6818 * it is properly null terminated or we will fail asserts
6819 * later. In theory we probably shouldn't get such SV's,
6820 * but if we do we should handle it gracefully. */
6821 if ( SvTYPE(msv) != SVt_PV || (SvLEN(msv) > SvCUR(msv) && *(SvEND(msv)) == 0) || SvIsCOW_shared_hash(msv) ) {
6822 /* not a string, or a string with a trailing null */
6825 /* a string with no trailing null, we need to copy it
6826 * so it has a trailing null */
6827 pat = sv_2mortal(newSVsv(msv));
6832 pRExC_state->code_blocks->cb[n-1].end = SvCUR(pat)-1;
6835 /* extract any code blocks within any embedded qr//'s */
6836 if (rx && SvTYPE(rx) == SVt_REGEXP
6837 && RX_ENGINE((REGEXP*)rx)->op_comp)
6840 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
6841 if (ri->code_blocks && ri->code_blocks->count) {
6843 /* the presence of an embedded qr// with code means
6844 * we should always recompile: the text of the
6845 * qr// may not have changed, but it may be a
6846 * different closure than last time */
6848 if (pRExC_state->code_blocks) {
6849 int new_count = pRExC_state->code_blocks->count
6850 + ri->code_blocks->count;
6851 Renew(pRExC_state->code_blocks->cb,
6852 new_count, struct reg_code_block);
6853 pRExC_state->code_blocks->count = new_count;
6856 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_
6857 ri->code_blocks->count);
6859 for (i=0; i < ri->code_blocks->count; i++) {
6860 struct reg_code_block *src, *dst;
6861 STRLEN offset = orig_patlen
6862 + ReANY((REGEXP *)rx)->pre_prefix;
6863 assert(n < pRExC_state->code_blocks->count);
6864 src = &ri->code_blocks->cb[i];
6865 dst = &pRExC_state->code_blocks->cb[n];
6866 dst->start = src->start + offset;
6867 dst->end = src->end + offset;
6868 dst->block = src->block;
6869 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
6878 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
6887 /* see if there are any run-time code blocks in the pattern.
6888 * False positives are allowed */
6891 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6892 char *pat, STRLEN plen)
6897 PERL_UNUSED_CONTEXT;
6899 for (s = 0; s < plen; s++) {
6900 if ( pRExC_state->code_blocks
6901 && n < pRExC_state->code_blocks->count
6902 && s == pRExC_state->code_blocks->cb[n].start)
6904 s = pRExC_state->code_blocks->cb[n].end;
6908 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
6910 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
6912 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
6919 /* Handle run-time code blocks. We will already have compiled any direct
6920 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
6921 * copy of it, but with any literal code blocks blanked out and
6922 * appropriate chars escaped; then feed it into
6924 * eval "qr'modified_pattern'"
6928 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
6932 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
6934 * After eval_sv()-ing that, grab any new code blocks from the returned qr
6935 * and merge them with any code blocks of the original regexp.
6937 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
6938 * instead, just save the qr and return FALSE; this tells our caller that
6939 * the original pattern needs upgrading to utf8.
6943 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6944 char *pat, STRLEN plen)
6948 GET_RE_DEBUG_FLAGS_DECL;
6950 if (pRExC_state->runtime_code_qr) {
6951 /* this is the second time we've been called; this should
6952 * only happen if the main pattern got upgraded to utf8
6953 * during compilation; re-use the qr we compiled first time
6954 * round (which should be utf8 too)
6956 qr = pRExC_state->runtime_code_qr;
6957 pRExC_state->runtime_code_qr = NULL;
6958 assert(RExC_utf8 && SvUTF8(qr));
6964 int newlen = plen + 7; /* allow for "qr''xx\0" extra chars */
6968 /* determine how many extra chars we need for ' and \ escaping */
6969 for (s = 0; s < plen; s++) {
6970 if (pat[s] == '\'' || pat[s] == '\\')
6974 Newx(newpat, newlen, char);
6976 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
6978 for (s = 0; s < plen; s++) {
6979 if ( pRExC_state->code_blocks
6980 && n < pRExC_state->code_blocks->count
6981 && s == pRExC_state->code_blocks->cb[n].start)
6983 /* blank out literal code block so that they aren't
6984 * recompiled: eg change from/to:
6994 assert(pat[s] == '(');
6995 assert(pat[s+1] == '?');
6999 while (s < pRExC_state->code_blocks->cb[n].end) {
7007 if (pat[s] == '\'' || pat[s] == '\\')
7012 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED) {
7014 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED_MORE) {
7020 Perl_re_printf( aTHX_
7021 "%sre-parsing pattern for runtime code:%s %s\n",
7022 PL_colors[4], PL_colors[5], newpat);
7025 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
7031 PUSHSTACKi(PERLSI_REQUIRE);
7032 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
7033 * parsing qr''; normally only q'' does this. It also alters
7035 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
7036 SvREFCNT_dec_NN(sv);
7041 SV * const errsv = ERRSV;
7042 if (SvTRUE_NN(errsv))
7043 /* use croak_sv ? */
7044 Perl_croak_nocontext("%" SVf, SVfARG(errsv));
7046 assert(SvROK(qr_ref));
7048 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
7049 /* the leaving below frees the tmp qr_ref.
7050 * Give qr a life of its own */
7058 if (!RExC_utf8 && SvUTF8(qr)) {
7059 /* first time through; the pattern got upgraded; save the
7060 * qr for the next time through */
7061 assert(!pRExC_state->runtime_code_qr);
7062 pRExC_state->runtime_code_qr = qr;
7067 /* extract any code blocks within the returned qr// */
7070 /* merge the main (r1) and run-time (r2) code blocks into one */
7072 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
7073 struct reg_code_block *new_block, *dst;
7074 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
7078 if (!r2->code_blocks || !r2->code_blocks->count) /* we guessed wrong */
7080 SvREFCNT_dec_NN(qr);
7084 if (!r1->code_blocks)
7085 r1->code_blocks = S_alloc_code_blocks(aTHX_ 0);
7087 r1c = r1->code_blocks->count;
7088 r2c = r2->code_blocks->count;
7090 Newx(new_block, r1c + r2c, struct reg_code_block);
7094 while (i1 < r1c || i2 < r2c) {
7095 struct reg_code_block *src;
7099 src = &r2->code_blocks->cb[i2++];
7103 src = &r1->code_blocks->cb[i1++];
7104 else if ( r1->code_blocks->cb[i1].start
7105 < r2->code_blocks->cb[i2].start)
7107 src = &r1->code_blocks->cb[i1++];
7108 assert(src->end < r2->code_blocks->cb[i2].start);
7111 assert( r1->code_blocks->cb[i1].start
7112 > r2->code_blocks->cb[i2].start);
7113 src = &r2->code_blocks->cb[i2++];
7115 assert(src->end < r1->code_blocks->cb[i1].start);
7118 assert(pat[src->start] == '(');
7119 assert(pat[src->end] == ')');
7120 dst->start = src->start;
7121 dst->end = src->end;
7122 dst->block = src->block;
7123 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
7127 r1->code_blocks->count += r2c;
7128 Safefree(r1->code_blocks->cb);
7129 r1->code_blocks->cb = new_block;
7132 SvREFCNT_dec_NN(qr);
7138 S_setup_longest(pTHX_ RExC_state_t *pRExC_state,
7139 struct reg_substr_datum *rsd,
7140 struct scan_data_substrs *sub,
7141 STRLEN longest_length)
7143 /* This is the common code for setting up the floating and fixed length
7144 * string data extracted from Perl_re_op_compile() below. Returns a boolean
7145 * as to whether succeeded or not */
7149 bool eol = cBOOL(sub->flags & SF_BEFORE_EOL);
7150 bool meol = cBOOL(sub->flags & SF_BEFORE_MEOL);
7152 if (! (longest_length
7153 || (eol /* Can't have SEOL and MULTI */
7154 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
7156 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
7157 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
7162 /* copy the information about the longest from the reg_scan_data
7163 over to the program. */
7164 if (SvUTF8(sub->str)) {
7166 rsd->utf8_substr = sub->str;
7168 rsd->substr = sub->str;
7169 rsd->utf8_substr = NULL;
7171 /* end_shift is how many chars that must be matched that
7172 follow this item. We calculate it ahead of time as once the
7173 lookbehind offset is added in we lose the ability to correctly
7175 ml = sub->minlenp ? *(sub->minlenp) : (SSize_t)longest_length;
7176 rsd->end_shift = ml - sub->min_offset
7178 /* XXX SvTAIL is always false here - did you mean FBMcf_TAIL
7180 + (SvTAIL(sub->str) != 0)
7184 t = (eol/* Can't have SEOL and MULTI */
7185 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
7186 fbm_compile(sub->str, t ? FBMcf_TAIL : 0);
7192 S_set_regex_pv(pTHX_ RExC_state_t *pRExC_state, REGEXP *Rx)
7194 /* Calculates and sets in the compiled pattern 'Rx' the string to compile,
7195 * properly wrapped with the right modifiers */
7197 bool has_p = ((RExC_rx->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
7198 bool has_charset = RExC_utf8 || (get_regex_charset(RExC_rx->extflags)
7199 != REGEX_DEPENDS_CHARSET);
7201 /* The caret is output if there are any defaults: if not all the STD
7202 * flags are set, or if no character set specifier is needed */
7204 (((RExC_rx->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
7206 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
7207 == REG_RUN_ON_COMMENT_SEEN);
7208 U8 reganch = (U8)((RExC_rx->extflags & RXf_PMf_STD_PMMOD)
7209 >> RXf_PMf_STD_PMMOD_SHIFT);
7210 const char *fptr = STD_PAT_MODS; /*"msixxn"*/
7212 STRLEN pat_len = RExC_precomp_end - RExC_precomp;
7214 /* We output all the necessary flags; we never output a minus, as all
7215 * those are defaults, so are
7216 * covered by the caret */
7217 const STRLEN wraplen = pat_len + has_p + has_runon
7218 + has_default /* If needs a caret */
7219 + PL_bitcount[reganch] /* 1 char for each set standard flag */
7221 /* If needs a character set specifier */
7222 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
7223 + (sizeof("(?:)") - 1);
7225 PERL_ARGS_ASSERT_SET_REGEX_PV;
7227 /* make sure PL_bitcount bounds not exceeded */
7228 assert(sizeof(STD_PAT_MODS) <= 8);
7230 p = sv_grow(MUTABLE_SV(Rx), wraplen + 1); /* +1 for the ending NUL */
7233 SvFLAGS(Rx) |= SVf_UTF8;
7236 /* If a default, cover it using the caret */
7238 *p++= DEFAULT_PAT_MOD;
7244 name = get_regex_charset_name(RExC_rx->extflags, &len);
7245 if strEQ(name, DEPENDS_PAT_MODS) { /* /d under UTF-8 => /u */
7247 name = UNICODE_PAT_MODS;
7248 len = sizeof(UNICODE_PAT_MODS) - 1;
7250 Copy(name, p, len, char);
7254 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
7257 while((ch = *fptr++)) {
7265 Copy(RExC_precomp, p, pat_len, char);
7266 assert ((RX_WRAPPED(Rx) - p) < 16);
7267 RExC_rx->pre_prefix = p - RX_WRAPPED(Rx);
7270 /* Adding a trailing \n causes this to compile properly:
7271 my $R = qr / A B C # D E/x; /($R)/
7272 Otherwise the parens are considered part of the comment */
7277 SvCUR_set(Rx, p - RX_WRAPPED(Rx));
7281 * Perl_re_op_compile - the perl internal RE engine's function to compile a
7282 * regular expression into internal code.
7283 * The pattern may be passed either as:
7284 * a list of SVs (patternp plus pat_count)
7285 * a list of OPs (expr)
7286 * If both are passed, the SV list is used, but the OP list indicates
7287 * which SVs are actually pre-compiled code blocks
7289 * The SVs in the list have magic and qr overloading applied to them (and
7290 * the list may be modified in-place with replacement SVs in the latter
7293 * If the pattern hasn't changed from old_re, then old_re will be
7296 * eng is the current engine. If that engine has an op_comp method, then
7297 * handle directly (i.e. we assume that op_comp was us); otherwise, just
7298 * do the initial concatenation of arguments and pass on to the external
7301 * If is_bare_re is not null, set it to a boolean indicating whether the
7302 * arg list reduced (after overloading) to a single bare regex which has
7303 * been returned (i.e. /$qr/).
7305 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
7307 * pm_flags contains the PMf_* flags, typically based on those from the
7308 * pm_flags field of the related PMOP. Currently we're only interested in
7309 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
7311 * For many years this code had an initial sizing pass that calculated
7312 * (sometimes incorrectly, leading to security holes) the size needed for the
7313 * compiled pattern. That was changed by commit
7314 * 7c932d07cab18751bfc7515b4320436273a459e2 in 5.29, which reallocs the size, a
7315 * node at a time, as parsing goes along. Patches welcome to fix any obsolete
7316 * references to this sizing pass.
7318 * Now, an initial crude guess as to the size needed is made, based on the
7319 * length of the pattern. Patches welcome to improve that guess. That amount
7320 * of space is malloc'd and then immediately freed, and then clawed back node
7321 * by node. This design is to minimze, to the extent possible, memory churn
7322 * when doing the the reallocs.
7324 * A separate parentheses counting pass may be needed in some cases.
7325 * (Previously the sizing pass did this.) Patches welcome to reduce the number
7328 * The existence of a sizing pass necessitated design decisions that are no
7329 * longer needed. There are potential areas of simplification.
7331 * Beware that the optimization-preparation code in here knows about some
7332 * of the structure of the compiled regexp. [I'll say.]
7336 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
7337 OP *expr, const regexp_engine* eng, REGEXP *old_re,
7338 bool *is_bare_re, const U32 orig_rx_flags, const U32 pm_flags)
7341 REGEXP *Rx; /* Capital 'R' means points to a REGEXP */
7349 SV** new_patternp = patternp;
7351 /* these are all flags - maybe they should be turned
7352 * into a single int with different bit masks */
7353 I32 sawlookahead = 0;
7358 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
7360 bool runtime_code = 0;
7362 RExC_state_t RExC_state;
7363 RExC_state_t * const pRExC_state = &RExC_state;
7364 #ifdef TRIE_STUDY_OPT
7366 RExC_state_t copyRExC_state;
7368 GET_RE_DEBUG_FLAGS_DECL;
7370 PERL_ARGS_ASSERT_RE_OP_COMPILE;
7372 DEBUG_r(if (!PL_colorset) reginitcolors());
7374 /* Initialize these here instead of as-needed, as is quick and avoids
7375 * having to test them each time otherwise */
7376 if (! PL_InBitmap) {
7378 char * dump_len_string;
7381 /* This is calculated here, because the Perl program that generates the
7382 * static global ones doesn't currently have access to
7383 * NUM_ANYOF_CODE_POINTS */
7384 PL_InBitmap = _new_invlist(2);
7385 PL_InBitmap = _add_range_to_invlist(PL_InBitmap, 0,
7386 NUM_ANYOF_CODE_POINTS - 1);
7388 dump_len_string = PerlEnv_getenv("PERL_DUMP_RE_MAX_LEN");
7389 if ( ! dump_len_string
7390 || ! grok_atoUV(dump_len_string, (UV *)&PL_dump_re_max_len, NULL))
7392 PL_dump_re_max_len = 60; /* A reasonable default */
7397 pRExC_state->warn_text = NULL;
7398 pRExC_state->unlexed_names = NULL;
7399 pRExC_state->code_blocks = NULL;
7402 *is_bare_re = FALSE;
7404 if (expr && (expr->op_type == OP_LIST ||
7405 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
7406 /* allocate code_blocks if needed */
7410 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o))
7411 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
7412 ncode++; /* count of DO blocks */
7415 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_ ncode);
7419 /* compile-time pattern with just OP_CONSTs and DO blocks */
7424 /* find how many CONSTs there are */
7427 if (expr->op_type == OP_CONST)
7430 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
7431 if (o->op_type == OP_CONST)
7435 /* fake up an SV array */
7437 assert(!new_patternp);
7438 Newx(new_patternp, n, SV*);
7439 SAVEFREEPV(new_patternp);
7443 if (expr->op_type == OP_CONST)
7444 new_patternp[n] = cSVOPx_sv(expr);
7446 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
7447 if (o->op_type == OP_CONST)
7448 new_patternp[n++] = cSVOPo_sv;
7453 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7454 "Assembling pattern from %d elements%s\n", pat_count,
7455 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
7457 /* set expr to the first arg op */
7459 if (pRExC_state->code_blocks && pRExC_state->code_blocks->count
7460 && expr->op_type != OP_CONST)
7462 expr = cLISTOPx(expr)->op_first;
7463 assert( expr->op_type == OP_PUSHMARK
7464 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
7465 || expr->op_type == OP_PADRANGE);
7466 expr = OpSIBLING(expr);
7469 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
7470 expr, &recompile, NULL);
7472 /* handle bare (possibly after overloading) regex: foo =~ $re */
7477 if (SvTYPE(re) == SVt_REGEXP) {
7481 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7482 "Precompiled pattern%s\n",
7483 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
7489 exp = SvPV_nomg(pat, plen);
7491 if (!eng->op_comp) {
7492 if ((SvUTF8(pat) && IN_BYTES)
7493 || SvGMAGICAL(pat) || SvAMAGIC(pat))
7495 /* make a temporary copy; either to convert to bytes,
7496 * or to avoid repeating get-magic / overloaded stringify */
7497 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
7498 (IN_BYTES ? 0 : SvUTF8(pat)));
7500 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
7503 /* ignore the utf8ness if the pattern is 0 length */
7504 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
7505 RExC_uni_semantics = 0;
7506 RExC_contains_locale = 0;
7507 RExC_strict = cBOOL(pm_flags & RXf_PMf_STRICT);
7508 RExC_in_script_run = 0;
7509 RExC_study_started = 0;
7510 pRExC_state->runtime_code_qr = NULL;
7511 RExC_frame_head= NULL;
7512 RExC_frame_last= NULL;
7513 RExC_frame_count= 0;
7514 RExC_latest_warn_offset = 0;
7515 RExC_use_BRANCHJ = 0;
7516 RExC_total_parens = 0;
7517 RExC_open_parens = NULL;
7518 RExC_close_parens = NULL;
7519 RExC_paren_names = NULL;
7521 RExC_seen_d_op = FALSE;
7523 RExC_paren_name_list = NULL;
7527 RExC_mysv1= sv_newmortal();
7528 RExC_mysv2= sv_newmortal();
7532 SV *dsv= sv_newmortal();
7533 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, PL_dump_re_max_len);
7534 Perl_re_printf( aTHX_ "%sCompiling REx%s %s\n",
7535 PL_colors[4], PL_colors[5], s);
7538 /* we jump here if we have to recompile, e.g., from upgrading the pattern
7541 if ((pm_flags & PMf_USE_RE_EVAL)
7542 /* this second condition covers the non-regex literal case,
7543 * i.e. $foo =~ '(?{})'. */
7544 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
7546 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
7549 /* return old regex if pattern hasn't changed */
7550 /* XXX: note in the below we have to check the flags as well as the
7553 * Things get a touch tricky as we have to compare the utf8 flag
7554 * independently from the compile flags. */
7558 && !!RX_UTF8(old_re) == !!RExC_utf8
7559 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
7560 && RX_PRECOMP(old_re)
7561 && RX_PRELEN(old_re) == plen
7562 && memEQ(RX_PRECOMP(old_re), exp, plen)
7563 && !runtime_code /* with runtime code, always recompile */ )
7568 /* Allocate the pattern's SV */
7569 RExC_rx_sv = Rx = (REGEXP*) newSV_type(SVt_REGEXP);
7570 RExC_rx = ReANY(Rx);
7571 if ( RExC_rx == NULL )
7572 FAIL("Regexp out of space");
7574 rx_flags = orig_rx_flags;
7576 if ( (UTF || RExC_uni_semantics)
7577 && initial_charset == REGEX_DEPENDS_CHARSET)
7580 /* Set to use unicode semantics if the pattern is in utf8 and has the
7581 * 'depends' charset specified, as it means unicode when utf8 */
7582 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
7583 RExC_uni_semantics = 1;
7586 RExC_pm_flags = pm_flags;
7589 assert(TAINTING_get || !TAINT_get);
7591 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
7593 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
7594 /* whoops, we have a non-utf8 pattern, whilst run-time code
7595 * got compiled as utf8. Try again with a utf8 pattern */
7596 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7597 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7601 assert(!pRExC_state->runtime_code_qr);
7607 RExC_in_lookbehind = 0;
7608 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
7610 RExC_recode_x_to_native = 0;
7612 RExC_in_multi_char_class = 0;
7614 RExC_start = RExC_copy_start_in_constructed = RExC_copy_start_in_input = RExC_precomp = exp;
7615 RExC_precomp_end = RExC_end = exp + plen;
7617 RExC_whilem_seen = 0;
7619 RExC_recurse = NULL;
7620 RExC_study_chunk_recursed = NULL;
7621 RExC_study_chunk_recursed_bytes= 0;
7622 RExC_recurse_count = 0;
7623 pRExC_state->code_index = 0;
7625 /* Initialize the string in the compiled pattern. This is so that there is
7626 * something to output if necessary */
7627 set_regex_pv(pRExC_state, Rx);
7630 Perl_re_printf( aTHX_
7631 "Starting parse and generation\n");
7633 RExC_lastparse=NULL;
7636 /* Allocate space and zero-initialize. Note, the two step process
7637 of zeroing when in debug mode, thus anything assigned has to
7638 happen after that */
7641 /* On the first pass of the parse, we guess how big this will be. Then
7642 * we grow in one operation to that amount and then give it back. As
7643 * we go along, we re-allocate what we need.
7645 * XXX Currently the guess is essentially that the pattern will be an
7646 * EXACT node with one byte input, one byte output. This is crude, and
7647 * better heuristics are welcome.
7649 * On any subsequent passes, we guess what we actually computed in the
7650 * latest earlier pass. Such a pass probably didn't complete so is
7651 * missing stuff. We could improve those guesses by knowing where the
7652 * parse stopped, and use the length so far plus apply the above
7653 * assumption to what's left. */
7654 RExC_size = STR_SZ(RExC_end - RExC_start);
7657 Newxc(RExC_rxi, sizeof(regexp_internal) + RExC_size, char, regexp_internal);
7658 if ( RExC_rxi == NULL )
7659 FAIL("Regexp out of space");
7661 Zero(RExC_rxi, sizeof(regexp_internal) + RExC_size, char);
7662 RXi_SET( RExC_rx, RExC_rxi );
7664 /* We start from 0 (over from 0 in the case this is a reparse. The first
7665 * node parsed will give back any excess memory we have allocated so far).
7669 /* non-zero initialization begins here */
7670 RExC_rx->engine= eng;
7671 RExC_rx->extflags = rx_flags;
7672 RXp_COMPFLAGS(RExC_rx) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
7674 if (pm_flags & PMf_IS_QR) {
7675 RExC_rxi->code_blocks = pRExC_state->code_blocks;
7676 if (RExC_rxi->code_blocks) {
7677 RExC_rxi->code_blocks->refcnt++;
7681 RExC_rx->intflags = 0;
7683 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
7686 /* This NUL is guaranteed because the pattern comes from an SV*, and the sv
7687 * code makes sure the final byte is an uncounted NUL. But should this
7688 * ever not be the case, lots of things could read beyond the end of the
7689 * buffer: loops like
7690 * while(isFOO(*RExC_parse)) RExC_parse++;
7691 * strchr(RExC_parse, "foo");
7692 * etc. So it is worth noting. */
7693 assert(*RExC_end == '\0');
7697 RExC_parens_buf_size = 0;
7698 RExC_emit_start = RExC_rxi->program;
7699 pRExC_state->code_index = 0;
7701 *((char*) RExC_emit_start) = (char) REG_MAGIC;
7705 if (reg(pRExC_state, 0, &flags, 1)) {
7707 /* Success!, But we may need to redo the parse knowing how many parens
7708 * there actually are */
7709 if (IN_PARENS_PASS) {
7710 flags |= RESTART_PARSE;
7713 /* We have that number in RExC_npar */
7714 RExC_total_parens = RExC_npar;
7716 else if (! MUST_RESTART(flags)) {
7718 Perl_croak(aTHX_ "panic: reg returned failure to re_op_compile, flags=%#" UVxf, (UV) flags);
7721 /* Here, we either have success, or we have to redo the parse for some reason */
7722 if (MUST_RESTART(flags)) {
7724 /* It's possible to write a regexp in ascii that represents Unicode
7725 codepoints outside of the byte range, such as via \x{100}. If we
7726 detect such a sequence we have to convert the entire pattern to utf8
7727 and then recompile, as our sizing calculation will have been based
7728 on 1 byte == 1 character, but we will need to use utf8 to encode
7729 at least some part of the pattern, and therefore must convert the whole
7732 if (flags & NEED_UTF8) {
7734 /* We have stored the offset of the final warning output so far.
7735 * That must be adjusted. Any variant characters between the start
7736 * of the pattern and this warning count for 2 bytes in the final,
7737 * so just add them again */
7738 if (UNLIKELY(RExC_latest_warn_offset > 0)) {
7739 RExC_latest_warn_offset +=
7740 variant_under_utf8_count((U8 *) exp, (U8 *) exp
7741 + RExC_latest_warn_offset);
7743 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7744 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7745 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "Need to redo parse after upgrade\n"));
7748 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "Need to redo parse\n"));
7751 if (ALL_PARENS_COUNTED) {
7752 /* Make enough room for all the known parens, and zero it */
7753 Renew(RExC_open_parens, RExC_total_parens, regnode_offset);
7754 Zero(RExC_open_parens, RExC_total_parens, regnode_offset);
7755 RExC_open_parens[0] = 1; /* +1 for REG_MAGIC */
7757 Renew(RExC_close_parens, RExC_total_parens, regnode_offset);
7758 Zero(RExC_close_parens, RExC_total_parens, regnode_offset);
7760 else { /* Parse did not complete. Reinitialize the parentheses
7762 RExC_total_parens = 0;
7763 if (RExC_open_parens) {
7764 Safefree(RExC_open_parens);
7765 RExC_open_parens = NULL;
7767 if (RExC_close_parens) {
7768 Safefree(RExC_close_parens);
7769 RExC_close_parens = NULL;
7773 /* Clean up what we did in this parse */
7774 SvREFCNT_dec_NN(RExC_rx_sv);
7779 /* Here, we have successfully parsed and generated the pattern's program
7780 * for the regex engine. We are ready to finish things up and look for
7783 /* Update the string to compile, with correct modifiers, etc */
7784 set_regex_pv(pRExC_state, Rx);
7786 RExC_rx->nparens = RExC_total_parens - 1;
7788 /* Uses the upper 4 bits of the FLAGS field, so keep within that size */
7789 if (RExC_whilem_seen > 15)
7790 RExC_whilem_seen = 15;
7793 Perl_re_printf( aTHX_
7794 "Required size %" IVdf " nodes\n", (IV)RExC_size);
7796 RExC_lastparse=NULL;
7799 #ifdef RE_TRACK_PATTERN_OFFSETS
7800 DEBUG_OFFSETS_r(Perl_re_printf( aTHX_
7801 "%s %" UVuf " bytes for offset annotations.\n",
7802 RExC_offsets ? "Got" : "Couldn't get",
7803 (UV)((RExC_offsets[0] * 2 + 1))));
7804 DEBUG_OFFSETS_r(if (RExC_offsets) {
7805 const STRLEN len = RExC_offsets[0];
7807 GET_RE_DEBUG_FLAGS_DECL;
7808 Perl_re_printf( aTHX_
7809 "Offsets: [%" UVuf "]\n\t", (UV)RExC_offsets[0]);
7810 for (i = 1; i <= len; i++) {
7811 if (RExC_offsets[i*2-1] || RExC_offsets[i*2])
7812 Perl_re_printf( aTHX_ "%" UVuf ":%" UVuf "[%" UVuf "] ",
7813 (UV)i, (UV)RExC_offsets[i*2-1], (UV)RExC_offsets[i*2]);
7815 Perl_re_printf( aTHX_ "\n");
7819 SetProgLen(RExC_rxi,RExC_size);
7823 Perl_re_printf( aTHX_ "Starting post parse optimization\n");
7826 /* XXXX To minimize changes to RE engine we always allocate
7827 3-units-long substrs field. */
7828 Newx(RExC_rx->substrs, 1, struct reg_substr_data);
7829 if (RExC_recurse_count) {
7830 Newx(RExC_recurse, RExC_recurse_count, regnode *);
7831 SAVEFREEPV(RExC_recurse);
7834 if (RExC_seen & REG_RECURSE_SEEN) {
7835 /* Note, RExC_total_parens is 1 + the number of parens in a pattern.
7836 * So its 1 if there are no parens. */
7837 RExC_study_chunk_recursed_bytes= (RExC_total_parens >> 3) +
7838 ((RExC_total_parens & 0x07) != 0);
7839 Newx(RExC_study_chunk_recursed,
7840 RExC_study_chunk_recursed_bytes * RExC_total_parens, U8);
7841 SAVEFREEPV(RExC_study_chunk_recursed);
7845 RExC_rx->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
7847 RExC_study_chunk_recursed_count= 0;
7849 Zero(RExC_rx->substrs, 1, struct reg_substr_data);
7850 if (RExC_study_chunk_recursed) {
7851 Zero(RExC_study_chunk_recursed,
7852 RExC_study_chunk_recursed_bytes * RExC_total_parens, U8);
7856 #ifdef TRIE_STUDY_OPT
7858 StructCopy(&zero_scan_data, &data, scan_data_t);
7859 copyRExC_state = RExC_state;
7862 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "Restudying\n"));
7864 RExC_state = copyRExC_state;
7865 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
7866 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
7868 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
7869 StructCopy(&zero_scan_data, &data, scan_data_t);
7872 StructCopy(&zero_scan_data, &data, scan_data_t);
7875 /* Dig out information for optimizations. */
7876 RExC_rx->extflags = RExC_flags; /* was pm_op */
7877 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
7880 SvUTF8_on(Rx); /* Unicode in it? */
7881 RExC_rxi->regstclass = NULL;
7882 if (RExC_naughty >= TOO_NAUGHTY) /* Probably an expensive pattern. */
7883 RExC_rx->intflags |= PREGf_NAUGHTY;
7884 scan = RExC_rxi->program + 1; /* First BRANCH. */
7886 /* testing for BRANCH here tells us whether there is "must appear"
7887 data in the pattern. If there is then we can use it for optimisations */
7888 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
7891 STRLEN longest_length[2];
7892 regnode_ssc ch_class; /* pointed to by data */
7894 SSize_t last_close = 0; /* pointed to by data */
7895 regnode *first= scan;
7896 regnode *first_next= regnext(first);
7900 * Skip introductions and multiplicators >= 1
7901 * so that we can extract the 'meat' of the pattern that must
7902 * match in the large if() sequence following.
7903 * NOTE that EXACT is NOT covered here, as it is normally
7904 * picked up by the optimiser separately.
7906 * This is unfortunate as the optimiser isnt handling lookahead
7907 * properly currently.
7910 while ((OP(first) == OPEN && (sawopen = 1)) ||
7911 /* An OR of *one* alternative - should not happen now. */
7912 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
7913 /* for now we can't handle lookbehind IFMATCH*/
7914 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
7915 (OP(first) == PLUS) ||
7916 (OP(first) == MINMOD) ||
7917 /* An {n,m} with n>0 */
7918 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
7919 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
7922 * the only op that could be a regnode is PLUS, all the rest
7923 * will be regnode_1 or regnode_2.
7925 * (yves doesn't think this is true)
7927 if (OP(first) == PLUS)
7930 if (OP(first) == MINMOD)
7932 first += regarglen[OP(first)];
7934 first = NEXTOPER(first);
7935 first_next= regnext(first);
7938 /* Starting-point info. */
7940 DEBUG_PEEP("first:", first, 0, 0);
7941 /* Ignore EXACT as we deal with it later. */
7942 if (PL_regkind[OP(first)] == EXACT) {
7943 if ( OP(first) == EXACT
7944 || OP(first) == EXACT_ONLY8
7945 || OP(first) == EXACTL)
7947 NOOP; /* Empty, get anchored substr later. */
7950 RExC_rxi->regstclass = first;
7953 else if (PL_regkind[OP(first)] == TRIE &&
7954 ((reg_trie_data *)RExC_rxi->data->data[ ARG(first) ])->minlen>0)
7956 /* this can happen only on restudy */
7957 RExC_rxi->regstclass = construct_ahocorasick_from_trie(pRExC_state, (regnode *)first, 0);
7960 else if (REGNODE_SIMPLE(OP(first)))
7961 RExC_rxi->regstclass = first;
7962 else if (PL_regkind[OP(first)] == BOUND ||
7963 PL_regkind[OP(first)] == NBOUND)
7964 RExC_rxi->regstclass = first;
7965 else if (PL_regkind[OP(first)] == BOL) {
7966 RExC_rx->intflags |= (OP(first) == MBOL
7969 first = NEXTOPER(first);
7972 else if (OP(first) == GPOS) {
7973 RExC_rx->intflags |= PREGf_ANCH_GPOS;
7974 first = NEXTOPER(first);
7977 else if ((!sawopen || !RExC_sawback) &&
7979 (OP(first) == STAR &&
7980 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
7981 !(RExC_rx->intflags & PREGf_ANCH) && !pRExC_state->code_blocks)
7983 /* turn .* into ^.* with an implied $*=1 */
7985 (OP(NEXTOPER(first)) == REG_ANY)
7988 RExC_rx->intflags |= (type | PREGf_IMPLICIT);
7989 first = NEXTOPER(first);
7992 if (sawplus && !sawminmod && !sawlookahead
7993 && (!sawopen || !RExC_sawback)
7994 && !pRExC_state->code_blocks) /* May examine pos and $& */
7995 /* x+ must match at the 1st pos of run of x's */
7996 RExC_rx->intflags |= PREGf_SKIP;
7998 /* Scan is after the zeroth branch, first is atomic matcher. */
7999 #ifdef TRIE_STUDY_OPT
8002 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
8003 (IV)(first - scan + 1))
8007 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
8008 (IV)(first - scan + 1))
8014 * If there's something expensive in the r.e., find the
8015 * longest literal string that must appear and make it the
8016 * regmust. Resolve ties in favor of later strings, since
8017 * the regstart check works with the beginning of the r.e.
8018 * and avoiding duplication strengthens checking. Not a
8019 * strong reason, but sufficient in the absence of others.
8020 * [Now we resolve ties in favor of the earlier string if
8021 * it happens that c_offset_min has been invalidated, since the
8022 * earlier string may buy us something the later one won't.]
8025 data.substrs[0].str = newSVpvs("");
8026 data.substrs[1].str = newSVpvs("");
8027 data.last_found = newSVpvs("");
8028 data.cur_is_floating = 0; /* initially any found substring is fixed */
8029 ENTER_with_name("study_chunk");
8030 SAVEFREESV(data.substrs[0].str);
8031 SAVEFREESV(data.substrs[1].str);
8032 SAVEFREESV(data.last_found);
8034 if (!RExC_rxi->regstclass) {
8035 ssc_init(pRExC_state, &ch_class);
8036 data.start_class = &ch_class;
8037 stclass_flag = SCF_DO_STCLASS_AND;
8038 } else /* XXXX Check for BOUND? */
8040 data.last_closep = &last_close;
8044 * MAIN ENTRY FOR study_chunk() FOR m/PATTERN/
8045 * (NO top level branches)
8047 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
8048 scan + RExC_size, /* Up to end */
8050 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
8051 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
8055 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
8058 if ( RExC_total_parens == 1 && !data.cur_is_floating
8059 && data.last_start_min == 0 && data.last_end > 0
8060 && !RExC_seen_zerolen
8061 && !(RExC_seen & REG_VERBARG_SEEN)
8062 && !(RExC_seen & REG_GPOS_SEEN)
8064 RExC_rx->extflags |= RXf_CHECK_ALL;
8066 scan_commit(pRExC_state, &data,&minlen, 0);
8069 /* XXX this is done in reverse order because that's the way the
8070 * code was before it was parameterised. Don't know whether it
8071 * actually needs doing in reverse order. DAPM */
8072 for (i = 1; i >= 0; i--) {
8073 longest_length[i] = CHR_SVLEN(data.substrs[i].str);
8076 && SvCUR(data.substrs[0].str) /* ok to leave SvCUR */
8077 && data.substrs[0].min_offset
8078 == data.substrs[1].min_offset
8079 && SvCUR(data.substrs[0].str)
8080 == SvCUR(data.substrs[1].str)
8082 && S_setup_longest (aTHX_ pRExC_state,
8083 &(RExC_rx->substrs->data[i]),
8087 RExC_rx->substrs->data[i].min_offset =
8088 data.substrs[i].min_offset - data.substrs[i].lookbehind;
8090 RExC_rx->substrs->data[i].max_offset = data.substrs[i].max_offset;
8091 /* Don't offset infinity */
8092 if (data.substrs[i].max_offset < SSize_t_MAX)
8093 RExC_rx->substrs->data[i].max_offset -= data.substrs[i].lookbehind;
8094 SvREFCNT_inc_simple_void_NN(data.substrs[i].str);
8097 RExC_rx->substrs->data[i].substr = NULL;
8098 RExC_rx->substrs->data[i].utf8_substr = NULL;
8099 longest_length[i] = 0;
8103 LEAVE_with_name("study_chunk");
8105 if (RExC_rxi->regstclass
8106 && (OP(RExC_rxi->regstclass) == REG_ANY || OP(RExC_rxi->regstclass) == SANY))
8107 RExC_rxi->regstclass = NULL;
8109 if ((!(RExC_rx->substrs->data[0].substr || RExC_rx->substrs->data[0].utf8_substr)
8110 || RExC_rx->substrs->data[0].min_offset)
8112 && ! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
8113 && is_ssc_worth_it(pRExC_state, data.start_class))
8115 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
8117 ssc_finalize(pRExC_state, data.start_class);
8119 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
8120 StructCopy(data.start_class,
8121 (regnode_ssc*)RExC_rxi->data->data[n],
8123 RExC_rxi->regstclass = (regnode*)RExC_rxi->data->data[n];
8124 RExC_rx->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
8125 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
8126 regprop(RExC_rx, sv, (regnode*)data.start_class, NULL, pRExC_state);
8127 Perl_re_printf( aTHX_
8128 "synthetic stclass \"%s\".\n",
8129 SvPVX_const(sv));});
8130 data.start_class = NULL;
8133 /* A temporary algorithm prefers floated substr to fixed one of
8134 * same length to dig more info. */
8135 i = (longest_length[0] <= longest_length[1]);
8136 RExC_rx->substrs->check_ix = i;
8137 RExC_rx->check_end_shift = RExC_rx->substrs->data[i].end_shift;
8138 RExC_rx->check_substr = RExC_rx->substrs->data[i].substr;
8139 RExC_rx->check_utf8 = RExC_rx->substrs->data[i].utf8_substr;
8140 RExC_rx->check_offset_min = RExC_rx->substrs->data[i].min_offset;
8141 RExC_rx->check_offset_max = RExC_rx->substrs->data[i].max_offset;
8142 if (!i && (RExC_rx->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS)))
8143 RExC_rx->intflags |= PREGf_NOSCAN;
8145 if ((RExC_rx->check_substr || RExC_rx->check_utf8) ) {
8146 RExC_rx->extflags |= RXf_USE_INTUIT;
8147 if (SvTAIL(RExC_rx->check_substr ? RExC_rx->check_substr : RExC_rx->check_utf8))
8148 RExC_rx->extflags |= RXf_INTUIT_TAIL;
8151 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
8152 if ( (STRLEN)minlen < longest_length[1] )
8153 minlen= longest_length[1];
8154 if ( (STRLEN)minlen < longest_length[0] )
8155 minlen= longest_length[0];
8159 /* Several toplevels. Best we can is to set minlen. */
8161 regnode_ssc ch_class;
8162 SSize_t last_close = 0;
8164 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "\nMulti Top Level\n"));
8166 scan = RExC_rxi->program + 1;
8167 ssc_init(pRExC_state, &ch_class);
8168 data.start_class = &ch_class;
8169 data.last_closep = &last_close;
8173 * MAIN ENTRY FOR study_chunk() FOR m/P1|P2|.../
8174 * (patterns WITH top level branches)
8176 minlen = study_chunk(pRExC_state,
8177 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
8178 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
8179 ? SCF_TRIE_DOING_RESTUDY
8183 CHECK_RESTUDY_GOTO_butfirst(NOOP);
8185 RExC_rx->check_substr = NULL;
8186 RExC_rx->check_utf8 = NULL;
8187 RExC_rx->substrs->data[0].substr = NULL;
8188 RExC_rx->substrs->data[0].utf8_substr = NULL;
8189 RExC_rx->substrs->data[1].substr = NULL;
8190 RExC_rx->substrs->data[1].utf8_substr = NULL;
8192 if (! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
8193 && is_ssc_worth_it(pRExC_state, data.start_class))
8195 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
8197 ssc_finalize(pRExC_state, data.start_class);
8199 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
8200 StructCopy(data.start_class,
8201 (regnode_ssc*)RExC_rxi->data->data[n],
8203 RExC_rxi->regstclass = (regnode*)RExC_rxi->data->data[n];
8204 RExC_rx->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
8205 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
8206 regprop(RExC_rx, sv, (regnode*)data.start_class, NULL, pRExC_state);
8207 Perl_re_printf( aTHX_
8208 "synthetic stclass \"%s\".\n",
8209 SvPVX_const(sv));});
8210 data.start_class = NULL;
8214 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
8215 RExC_rx->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
8216 RExC_rx->maxlen = REG_INFTY;
8219 RExC_rx->maxlen = RExC_maxlen;
8222 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
8223 the "real" pattern. */
8225 Perl_re_printf( aTHX_ "minlen: %" IVdf " RExC_rx->minlen:%" IVdf " maxlen:%" IVdf "\n",
8226 (IV)minlen, (IV)RExC_rx->minlen, (IV)RExC_maxlen);
8228 RExC_rx->minlenret = minlen;
8229 if (RExC_rx->minlen < minlen)
8230 RExC_rx->minlen = minlen;
8232 if (RExC_seen & REG_RECURSE_SEEN ) {
8233 RExC_rx->intflags |= PREGf_RECURSE_SEEN;
8234 Newx(RExC_rx->recurse_locinput, RExC_rx->nparens + 1, char *);
8236 if (RExC_seen & REG_GPOS_SEEN)
8237 RExC_rx->intflags |= PREGf_GPOS_SEEN;
8238 if (RExC_seen & REG_LOOKBEHIND_SEEN)
8239 RExC_rx->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
8241 if (pRExC_state->code_blocks)
8242 RExC_rx->extflags |= RXf_EVAL_SEEN;
8243 if (RExC_seen & REG_VERBARG_SEEN)
8245 RExC_rx->intflags |= PREGf_VERBARG_SEEN;
8246 RExC_rx->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
8248 if (RExC_seen & REG_CUTGROUP_SEEN)
8249 RExC_rx->intflags |= PREGf_CUTGROUP_SEEN;
8250 if (pm_flags & PMf_USE_RE_EVAL)
8251 RExC_rx->intflags |= PREGf_USE_RE_EVAL;
8252 if (RExC_paren_names)
8253 RXp_PAREN_NAMES(RExC_rx) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
8255 RXp_PAREN_NAMES(RExC_rx) = NULL;
8257 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
8258 * so it can be used in pp.c */
8259 if (RExC_rx->intflags & PREGf_ANCH)
8260 RExC_rx->extflags |= RXf_IS_ANCHORED;
8264 /* this is used to identify "special" patterns that might result
8265 * in Perl NOT calling the regex engine and instead doing the match "itself",
8266 * particularly special cases in split//. By having the regex compiler
8267 * do this pattern matching at a regop level (instead of by inspecting the pattern)
8268 * we avoid weird issues with equivalent patterns resulting in different behavior,
8269 * AND we allow non Perl engines to get the same optimizations by the setting the
8270 * flags appropriately - Yves */
8271 regnode *first = RExC_rxi->program + 1;
8273 regnode *next = regnext(first);
8276 if (PL_regkind[fop] == NOTHING && nop == END)
8277 RExC_rx->extflags |= RXf_NULL;
8278 else if ((fop == MBOL || (fop == SBOL && !first->flags)) && nop == END)
8279 /* when fop is SBOL first->flags will be true only when it was
8280 * produced by parsing /\A/, and not when parsing /^/. This is
8281 * very important for the split code as there we want to
8282 * treat /^/ as /^/m, but we do not want to treat /\A/ as /^/m.
8283 * See rt #122761 for more details. -- Yves */
8284 RExC_rx->extflags |= RXf_START_ONLY;
8285 else if (fop == PLUS
8286 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
8288 RExC_rx->extflags |= RXf_WHITE;
8289 else if ( RExC_rx->extflags & RXf_SPLIT
8290 && (fop == EXACT || fop == EXACT_ONLY8 || fop == EXACTL)
8291 && STR_LEN(first) == 1
8292 && *(STRING(first)) == ' '
8294 RExC_rx->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
8298 if (RExC_contains_locale) {
8299 RXp_EXTFLAGS(RExC_rx) |= RXf_TAINTED;
8303 if (RExC_paren_names) {
8304 RExC_rxi->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
8305 RExC_rxi->data->data[RExC_rxi->name_list_idx]
8306 = (void*)SvREFCNT_inc(RExC_paren_name_list);
8309 RExC_rxi->name_list_idx = 0;
8311 while ( RExC_recurse_count > 0 ) {
8312 const regnode *scan = RExC_recurse[ --RExC_recurse_count ];
8314 * This data structure is set up in study_chunk() and is used
8315 * to calculate the distance between a GOSUB regopcode and
8316 * the OPEN/CURLYM (CURLYM's are special and can act like OPEN's)
8319 * If for some reason someone writes code that optimises
8320 * away a GOSUB opcode then the assert should be changed to
8321 * an if(scan) to guard the ARG2L_SET() - Yves
8324 assert(scan && OP(scan) == GOSUB);
8325 ARG2L_SET( scan, RExC_open_parens[ARG(scan)] - REGNODE_OFFSET(scan));
8328 Newxz(RExC_rx->offs, RExC_total_parens, regexp_paren_pair);
8329 /* assume we don't need to swap parens around before we match */
8331 Perl_re_printf( aTHX_ "study_chunk_recursed_count: %lu\n",
8332 (unsigned long)RExC_study_chunk_recursed_count);
8336 Perl_re_printf( aTHX_ "Final program:\n");
8340 if (RExC_open_parens) {
8341 Safefree(RExC_open_parens);
8342 RExC_open_parens = NULL;
8344 if (RExC_close_parens) {
8345 Safefree(RExC_close_parens);
8346 RExC_close_parens = NULL;
8350 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
8351 * by setting the regexp SV to readonly-only instead. If the
8352 * pattern's been recompiled, the USEDness should remain. */
8353 if (old_re && SvREADONLY(old_re))
8361 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
8364 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
8366 PERL_UNUSED_ARG(value);
8368 if (flags & RXapif_FETCH) {
8369 return reg_named_buff_fetch(rx, key, flags);
8370 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
8371 Perl_croak_no_modify();
8373 } else if (flags & RXapif_EXISTS) {
8374 return reg_named_buff_exists(rx, key, flags)
8377 } else if (flags & RXapif_REGNAMES) {
8378 return reg_named_buff_all(rx, flags);
8379 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
8380 return reg_named_buff_scalar(rx, flags);
8382 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
8388 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
8391 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
8392 PERL_UNUSED_ARG(lastkey);
8394 if (flags & RXapif_FIRSTKEY)
8395 return reg_named_buff_firstkey(rx, flags);
8396 else if (flags & RXapif_NEXTKEY)
8397 return reg_named_buff_nextkey(rx, flags);
8399 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
8406 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
8410 struct regexp *const rx = ReANY(r);
8412 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
8414 if (rx && RXp_PAREN_NAMES(rx)) {
8415 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
8418 SV* sv_dat=HeVAL(he_str);
8419 I32 *nums=(I32*)SvPVX(sv_dat);
8420 AV * const retarray = (flags & RXapif_ALL) ? newAV() : NULL;
8421 for ( i=0; i<SvIVX(sv_dat); i++ ) {
8422 if ((I32)(rx->nparens) >= nums[i]
8423 && rx->offs[nums[i]].start != -1
8424 && rx->offs[nums[i]].end != -1)
8427 CALLREG_NUMBUF_FETCH(r, nums[i], ret);
8432 ret = newSVsv(&PL_sv_undef);
8435 av_push(retarray, ret);
8438 return newRV_noinc(MUTABLE_SV(retarray));
8445 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
8448 struct regexp *const rx = ReANY(r);
8450 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
8452 if (rx && RXp_PAREN_NAMES(rx)) {
8453 if (flags & RXapif_ALL) {
8454 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
8456 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
8458 SvREFCNT_dec_NN(sv);
8470 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
8472 struct regexp *const rx = ReANY(r);
8474 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
8476 if ( rx && RXp_PAREN_NAMES(rx) ) {
8477 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
8479 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
8486 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
8488 struct regexp *const rx = ReANY(r);
8489 GET_RE_DEBUG_FLAGS_DECL;
8491 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
8493 if (rx && RXp_PAREN_NAMES(rx)) {
8494 HV *hv = RXp_PAREN_NAMES(rx);
8496 while ( (temphe = hv_iternext_flags(hv, 0)) ) {
8499 SV* sv_dat = HeVAL(temphe);
8500 I32 *nums = (I32*)SvPVX(sv_dat);
8501 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8502 if ((I32)(rx->lastparen) >= nums[i] &&
8503 rx->offs[nums[i]].start != -1 &&
8504 rx->offs[nums[i]].end != -1)
8510 if (parno || flags & RXapif_ALL) {
8511 return newSVhek(HeKEY_hek(temphe));
8519 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
8524 struct regexp *const rx = ReANY(r);
8526 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
8528 if (rx && RXp_PAREN_NAMES(rx)) {
8529 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
8530 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
8531 } else if (flags & RXapif_ONE) {
8532 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
8533 av = MUTABLE_AV(SvRV(ret));
8534 length = av_tindex(av);
8535 SvREFCNT_dec_NN(ret);
8536 return newSViv(length + 1);
8538 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
8543 return &PL_sv_undef;
8547 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
8549 struct regexp *const rx = ReANY(r);
8552 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
8554 if (rx && RXp_PAREN_NAMES(rx)) {
8555 HV *hv= RXp_PAREN_NAMES(rx);
8557 (void)hv_iterinit(hv);
8558 while ( (temphe = hv_iternext_flags(hv, 0)) ) {
8561 SV* sv_dat = HeVAL(temphe);
8562 I32 *nums = (I32*)SvPVX(sv_dat);
8563 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8564 if ((I32)(rx->lastparen) >= nums[i] &&
8565 rx->offs[nums[i]].start != -1 &&
8566 rx->offs[nums[i]].end != -1)
8572 if (parno || flags & RXapif_ALL) {
8573 av_push(av, newSVhek(HeKEY_hek(temphe)));
8578 return newRV_noinc(MUTABLE_SV(av));
8582 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
8585 struct regexp *const rx = ReANY(r);
8591 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
8593 if ( n == RX_BUFF_IDX_CARET_PREMATCH
8594 || n == RX_BUFF_IDX_CARET_FULLMATCH
8595 || n == RX_BUFF_IDX_CARET_POSTMATCH
8598 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8600 /* on something like
8603 * the KEEPCOPY is set on the PMOP rather than the regex */
8604 if (PL_curpm && r == PM_GETRE(PL_curpm))
8605 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8614 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
8615 /* no need to distinguish between them any more */
8616 n = RX_BUFF_IDX_FULLMATCH;
8618 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
8619 && rx->offs[0].start != -1)
8621 /* $`, ${^PREMATCH} */
8622 i = rx->offs[0].start;
8626 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
8627 && rx->offs[0].end != -1)
8629 /* $', ${^POSTMATCH} */
8630 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
8631 i = rx->sublen + rx->suboffset - rx->offs[0].end;
8634 if ( 0 <= n && n <= (I32)rx->nparens &&
8635 (s1 = rx->offs[n].start) != -1 &&
8636 (t1 = rx->offs[n].end) != -1)
8638 /* $&, ${^MATCH}, $1 ... */
8640 s = rx->subbeg + s1 - rx->suboffset;
8645 assert(s >= rx->subbeg);
8646 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
8648 #ifdef NO_TAINT_SUPPORT
8649 sv_setpvn(sv, s, i);
8651 const int oldtainted = TAINT_get;
8653 sv_setpvn(sv, s, i);
8654 TAINT_set(oldtainted);
8656 if (RXp_MATCH_UTF8(rx))
8661 if (RXp_MATCH_TAINTED(rx)) {
8662 if (SvTYPE(sv) >= SVt_PVMG) {
8663 MAGIC* const mg = SvMAGIC(sv);
8666 SvMAGIC_set(sv, mg->mg_moremagic);
8668 if ((mgt = SvMAGIC(sv))) {
8669 mg->mg_moremagic = mgt;
8670 SvMAGIC_set(sv, mg);
8687 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
8688 SV const * const value)
8690 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
8692 PERL_UNUSED_ARG(rx);
8693 PERL_UNUSED_ARG(paren);
8694 PERL_UNUSED_ARG(value);
8697 Perl_croak_no_modify();
8701 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
8704 struct regexp *const rx = ReANY(r);
8708 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
8710 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
8711 || paren == RX_BUFF_IDX_CARET_FULLMATCH
8712 || paren == RX_BUFF_IDX_CARET_POSTMATCH
8715 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8717 /* on something like
8720 * the KEEPCOPY is set on the PMOP rather than the regex */
8721 if (PL_curpm && r == PM_GETRE(PL_curpm))
8722 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8728 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
8730 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
8731 case RX_BUFF_IDX_PREMATCH: /* $` */
8732 if (rx->offs[0].start != -1) {
8733 i = rx->offs[0].start;
8742 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
8743 case RX_BUFF_IDX_POSTMATCH: /* $' */
8744 if (rx->offs[0].end != -1) {
8745 i = rx->sublen - rx->offs[0].end;
8747 s1 = rx->offs[0].end;
8754 default: /* $& / ${^MATCH}, $1, $2, ... */
8755 if (paren <= (I32)rx->nparens &&
8756 (s1 = rx->offs[paren].start) != -1 &&
8757 (t1 = rx->offs[paren].end) != -1)
8763 if (ckWARN(WARN_UNINITIALIZED))
8764 report_uninit((const SV *)sv);
8769 if (i > 0 && RXp_MATCH_UTF8(rx)) {
8770 const char * const s = rx->subbeg - rx->suboffset + s1;
8775 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
8782 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
8784 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
8785 PERL_UNUSED_ARG(rx);
8789 return newSVpvs("Regexp");
8792 /* Scans the name of a named buffer from the pattern.
8793 * If flags is REG_RSN_RETURN_NULL returns null.
8794 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
8795 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
8796 * to the parsed name as looked up in the RExC_paren_names hash.
8797 * If there is an error throws a vFAIL().. type exception.
8800 #define REG_RSN_RETURN_NULL 0
8801 #define REG_RSN_RETURN_NAME 1
8802 #define REG_RSN_RETURN_DATA 2
8805 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
8807 char *name_start = RExC_parse;
8810 PERL_ARGS_ASSERT_REG_SCAN_NAME;
8812 assert (RExC_parse <= RExC_end);
8813 if (RExC_parse == RExC_end) NOOP;
8814 else if (isIDFIRST_lazy_if_safe(RExC_parse, RExC_end, UTF)) {
8815 /* Note that the code here assumes well-formed UTF-8. Skip IDFIRST by
8816 * using do...while */
8819 RExC_parse += UTF8SKIP(RExC_parse);
8820 } while ( RExC_parse < RExC_end
8821 && isWORDCHAR_utf8_safe((U8*)RExC_parse, (U8*) RExC_end));
8825 } while (RExC_parse < RExC_end && isWORDCHAR(*RExC_parse));
8827 RExC_parse++; /* so the <- from the vFAIL is after the offending
8829 vFAIL("Group name must start with a non-digit word character");
8831 sv_name = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
8832 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
8833 if ( flags == REG_RSN_RETURN_NAME)
8835 else if (flags==REG_RSN_RETURN_DATA) {
8838 if ( ! sv_name ) /* should not happen*/
8839 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
8840 if (RExC_paren_names)
8841 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
8843 sv_dat = HeVAL(he_str);
8844 if ( ! sv_dat ) { /* Didn't find group */
8846 /* It might be a forward reference; we can't fail until we
8847 * know, by completing the parse to get all the groups, and
8849 if (ALL_PARENS_COUNTED) {
8850 vFAIL("Reference to nonexistent named group");
8853 REQUIRE_PARENS_PASS;
8859 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
8860 (unsigned long) flags);
8863 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
8864 if (RExC_lastparse!=RExC_parse) { \
8865 Perl_re_printf( aTHX_ "%s", \
8866 Perl_pv_pretty(aTHX_ RExC_mysv1, RExC_parse, \
8867 RExC_end - RExC_parse, 16, \
8869 PERL_PV_ESCAPE_UNI_DETECT | \
8870 PERL_PV_PRETTY_ELLIPSES | \
8871 PERL_PV_PRETTY_LTGT | \
8872 PERL_PV_ESCAPE_RE | \
8873 PERL_PV_PRETTY_EXACTSIZE \
8877 Perl_re_printf( aTHX_ "%16s",""); \
8879 if (RExC_lastnum!=RExC_emit) \
8880 Perl_re_printf( aTHX_ "|%4d", RExC_emit); \
8882 Perl_re_printf( aTHX_ "|%4s",""); \
8883 Perl_re_printf( aTHX_ "|%*s%-4s", \
8884 (int)((depth*2)), "", \
8887 RExC_lastnum=RExC_emit; \
8888 RExC_lastparse=RExC_parse; \
8893 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
8894 DEBUG_PARSE_MSG((funcname)); \
8895 Perl_re_printf( aTHX_ "%4s","\n"); \
8897 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({\
8898 DEBUG_PARSE_MSG((funcname)); \
8899 Perl_re_printf( aTHX_ fmt "\n",args); \
8902 /* This section of code defines the inversion list object and its methods. The
8903 * interfaces are highly subject to change, so as much as possible is static to
8904 * this file. An inversion list is here implemented as a malloc'd C UV array
8905 * as an SVt_INVLIST scalar.
8907 * An inversion list for Unicode is an array of code points, sorted by ordinal
8908 * number. Each element gives the code point that begins a range that extends
8909 * up-to but not including the code point given by the next element. The final
8910 * element gives the first code point of a range that extends to the platform's
8911 * infinity. The even-numbered elements (invlist[0], invlist[2], invlist[4],
8912 * ...) give ranges whose code points are all in the inversion list. We say
8913 * that those ranges are in the set. The odd-numbered elements give ranges
8914 * whose code points are not in the inversion list, and hence not in the set.
8915 * Thus, element [0] is the first code point in the list. Element [1]
8916 * is the first code point beyond that not in the list; and element [2] is the
8917 * first code point beyond that that is in the list. In other words, the first
8918 * range is invlist[0]..(invlist[1]-1), and all code points in that range are
8919 * in the inversion list. The second range is invlist[1]..(invlist[2]-1), and
8920 * all code points in that range are not in the inversion list. The third
8921 * range invlist[2]..(invlist[3]-1) gives code points that are in the inversion
8922 * list, and so forth. Thus every element whose index is divisible by two
8923 * gives the beginning of a range that is in the list, and every element whose
8924 * index is not divisible by two gives the beginning of a range not in the
8925 * list. If the final element's index is divisible by two, the inversion list
8926 * extends to the platform's infinity; otherwise the highest code point in the
8927 * inversion list is the contents of that element minus 1.
8929 * A range that contains just a single code point N will look like
8931 * invlist[i+1] == N+1
8933 * If N is UV_MAX (the highest representable code point on the machine), N+1 is
8934 * impossible to represent, so element [i+1] is omitted. The single element
8936 * invlist[0] == UV_MAX
8937 * contains just UV_MAX, but is interpreted as matching to infinity.
8939 * Taking the complement (inverting) an inversion list is quite simple, if the
8940 * first element is 0, remove it; otherwise add a 0 element at the beginning.
8941 * This implementation reserves an element at the beginning of each inversion
8942 * list to always contain 0; there is an additional flag in the header which
8943 * indicates if the list begins at the 0, or is offset to begin at the next
8944 * element. This means that the inversion list can be inverted without any
8945 * copying; just flip the flag.
8947 * More about inversion lists can be found in "Unicode Demystified"
8948 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
8950 * The inversion list data structure is currently implemented as an SV pointing
8951 * to an array of UVs that the SV thinks are bytes. This allows us to have an
8952 * array of UV whose memory management is automatically handled by the existing
8953 * facilities for SV's.
8955 * Some of the methods should always be private to the implementation, and some
8956 * should eventually be made public */
8958 /* The header definitions are in F<invlist_inline.h> */
8960 #ifndef PERL_IN_XSUB_RE
8962 PERL_STATIC_INLINE UV*
8963 S__invlist_array_init(SV* const invlist, const bool will_have_0)
8965 /* Returns a pointer to the first element in the inversion list's array.
8966 * This is called upon initialization of an inversion list. Where the
8967 * array begins depends on whether the list has the code point U+0000 in it
8968 * or not. The other parameter tells it whether the code that follows this
8969 * call is about to put a 0 in the inversion list or not. The first
8970 * element is either the element reserved for 0, if TRUE, or the element
8971 * after it, if FALSE */
8973 bool* offset = get_invlist_offset_addr(invlist);
8974 UV* zero_addr = (UV *) SvPVX(invlist);
8976 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
8979 assert(! _invlist_len(invlist));
8983 /* 1^1 = 0; 1^0 = 1 */
8984 *offset = 1 ^ will_have_0;
8985 return zero_addr + *offset;
8988 PERL_STATIC_INLINE void
8989 S_invlist_set_len(pTHX_ SV* const invlist, const UV len, const bool offset)
8991 /* Sets the current number of elements stored in the inversion list.
8992 * Updates SvCUR correspondingly */
8993 PERL_UNUSED_CONTEXT;
8994 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
8996 assert(is_invlist(invlist));
9001 : TO_INTERNAL_SIZE(len + offset));
9002 assert(SvLEN(invlist) == 0 || SvCUR(invlist) <= SvLEN(invlist));
9006 S_invlist_replace_list_destroys_src(pTHX_ SV * dest, SV * src)
9008 /* Replaces the inversion list in 'dest' with the one from 'src'. It
9009 * steals the list from 'src', so 'src' is made to have a NULL list. This
9010 * is similar to what SvSetMagicSV() would do, if it were implemented on
9011 * inversion lists, though this routine avoids a copy */
9013 const UV src_len = _invlist_len(src);
9014 const bool src_offset = *get_invlist_offset_addr(src);
9015 const STRLEN src_byte_len = SvLEN(src);
9016 char * array = SvPVX(src);
9018 const int oldtainted = TAINT_get;
9020 PERL_ARGS_ASSERT_INVLIST_REPLACE_LIST_DESTROYS_SRC;
9022 assert(is_invlist(src));
9023 assert(is_invlist(dest));
9024 assert(! invlist_is_iterating(src));
9025 assert(SvCUR(src) == 0 || SvCUR(src) < SvLEN(src));
9027 /* Make sure it ends in the right place with a NUL, as our inversion list
9028 * manipulations aren't careful to keep this true, but sv_usepvn_flags()
9030 array[src_byte_len - 1] = '\0';
9032 TAINT_NOT; /* Otherwise it breaks */
9033 sv_usepvn_flags(dest,
9037 /* This flag is documented to cause a copy to be avoided */
9038 SV_HAS_TRAILING_NUL);
9039 TAINT_set(oldtainted);
9044 /* Finish up copying over the other fields in an inversion list */
9045 *get_invlist_offset_addr(dest) = src_offset;
9046 invlist_set_len(dest, src_len, src_offset);
9047 *get_invlist_previous_index_addr(dest) = 0;
9048 invlist_iterfinish(dest);
9051 PERL_STATIC_INLINE IV*
9052 S_get_invlist_previous_index_addr(SV* invlist)
9054 /* Return the address of the IV that is reserved to hold the cached index
9056 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
9058 assert(is_invlist(invlist));
9060 return &(((XINVLIST*) SvANY(invlist))->prev_index);
9063 PERL_STATIC_INLINE IV
9064 S_invlist_previous_index(SV* const invlist)
9066 /* Returns cached index of previous search */
9068 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
9070 return *get_invlist_previous_index_addr(invlist);
9073 PERL_STATIC_INLINE void
9074 S_invlist_set_previous_index(SV* const invlist, const IV index)
9076 /* Caches <index> for later retrieval */
9078 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
9080 assert(index == 0 || index < (int) _invlist_len(invlist));
9082 *get_invlist_previous_index_addr(invlist) = index;
9085 PERL_STATIC_INLINE void
9086 S_invlist_trim(SV* invlist)
9088 /* Free the not currently-being-used space in an inversion list */
9090 /* But don't free up the space needed for the 0 UV that is always at the
9091 * beginning of the list, nor the trailing NUL */
9092 const UV min_size = TO_INTERNAL_SIZE(1) + 1;
9094 PERL_ARGS_ASSERT_INVLIST_TRIM;
9096 assert(is_invlist(invlist));
9098 SvPV_renew(invlist, MAX(min_size, SvCUR(invlist) + 1));
9101 PERL_STATIC_INLINE void
9102 S_invlist_clear(pTHX_ SV* invlist) /* Empty the inversion list */
9104 PERL_ARGS_ASSERT_INVLIST_CLEAR;
9106 assert(is_invlist(invlist));
9108 invlist_set_len(invlist, 0, 0);
9109 invlist_trim(invlist);
9112 #endif /* ifndef PERL_IN_XSUB_RE */
9114 PERL_STATIC_INLINE bool
9115 S_invlist_is_iterating(SV* const invlist)
9117 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
9119 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
9122 #ifndef PERL_IN_XSUB_RE
9124 PERL_STATIC_INLINE UV
9125 S_invlist_max(SV* const invlist)
9127 /* Returns the maximum number of elements storable in the inversion list's
9128 * array, without having to realloc() */
9130 PERL_ARGS_ASSERT_INVLIST_MAX;
9132 assert(is_invlist(invlist));
9134 /* Assumes worst case, in which the 0 element is not counted in the
9135 * inversion list, so subtracts 1 for that */
9136 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
9137 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
9138 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
9142 S_initialize_invlist_guts(pTHX_ SV* invlist, const Size_t initial_size)
9144 PERL_ARGS_ASSERT_INITIALIZE_INVLIST_GUTS;
9146 /* First 1 is in case the zero element isn't in the list; second 1 is for
9148 SvGROW(invlist, TO_INTERNAL_SIZE(initial_size + 1) + 1);
9149 invlist_set_len(invlist, 0, 0);
9151 /* Force iterinit() to be used to get iteration to work */
9152 invlist_iterfinish(invlist);
9154 *get_invlist_previous_index_addr(invlist) = 0;
9158 Perl__new_invlist(pTHX_ IV initial_size)
9161 /* Return a pointer to a newly constructed inversion list, with enough
9162 * space to store 'initial_size' elements. If that number is negative, a
9163 * system default is used instead */
9167 if (initial_size < 0) {
9171 new_list = newSV_type(SVt_INVLIST);
9172 initialize_invlist_guts(new_list, initial_size);
9178 Perl__new_invlist_C_array(pTHX_ const UV* const list)
9180 /* Return a pointer to a newly constructed inversion list, initialized to
9181 * point to <list>, which has to be in the exact correct inversion list
9182 * form, including internal fields. Thus this is a dangerous routine that
9183 * should not be used in the wrong hands. The passed in 'list' contains
9184 * several header fields at the beginning that are not part of the
9185 * inversion list body proper */
9187 const STRLEN length = (STRLEN) list[0];
9188 const UV version_id = list[1];
9189 const bool offset = cBOOL(list[2]);
9190 #define HEADER_LENGTH 3
9191 /* If any of the above changes in any way, you must change HEADER_LENGTH
9192 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
9193 * perl -E 'say int(rand 2**31-1)'
9195 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
9196 data structure type, so that one being
9197 passed in can be validated to be an
9198 inversion list of the correct vintage.
9201 SV* invlist = newSV_type(SVt_INVLIST);
9203 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
9205 if (version_id != INVLIST_VERSION_ID) {
9206 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
9209 /* The generated array passed in includes header elements that aren't part
9210 * of the list proper, so start it just after them */
9211 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
9213 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
9214 shouldn't touch it */
9216 *(get_invlist_offset_addr(invlist)) = offset;
9218 /* The 'length' passed to us is the physical number of elements in the
9219 * inversion list. But if there is an offset the logical number is one
9221 invlist_set_len(invlist, length - offset, offset);
9223 invlist_set_previous_index(invlist, 0);
9225 /* Initialize the iteration pointer. */
9226 invlist_iterfinish(invlist);
9228 SvREADONLY_on(invlist);
9234 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
9236 /* Grow the maximum size of an inversion list */
9238 PERL_ARGS_ASSERT_INVLIST_EXTEND;
9240 assert(is_invlist(invlist));
9242 /* Add one to account for the zero element at the beginning which may not
9243 * be counted by the calling parameters */
9244 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max + 1));
9248 S__append_range_to_invlist(pTHX_ SV* const invlist,
9249 const UV start, const UV end)
9251 /* Subject to change or removal. Append the range from 'start' to 'end' at
9252 * the end of the inversion list. The range must be above any existing
9256 UV max = invlist_max(invlist);
9257 UV len = _invlist_len(invlist);
9260 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
9262 if (len == 0) { /* Empty lists must be initialized */
9263 offset = start != 0;
9264 array = _invlist_array_init(invlist, ! offset);
9267 /* Here, the existing list is non-empty. The current max entry in the
9268 * list is generally the first value not in the set, except when the
9269 * set extends to the end of permissible values, in which case it is
9270 * the first entry in that final set, and so this call is an attempt to
9271 * append out-of-order */
9273 UV final_element = len - 1;
9274 array = invlist_array(invlist);
9275 if ( array[final_element] > start
9276 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
9278 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",
9279 array[final_element], start,
9280 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
9283 /* Here, it is a legal append. If the new range begins 1 above the end
9284 * of the range below it, it is extending the range below it, so the
9285 * new first value not in the set is one greater than the newly
9286 * extended range. */
9287 offset = *get_invlist_offset_addr(invlist);
9288 if (array[final_element] == start) {
9289 if (end != UV_MAX) {
9290 array[final_element] = end + 1;
9293 /* But if the end is the maximum representable on the machine,
9294 * assume that infinity was actually what was meant. Just let
9295 * the range that this would extend to have no end */
9296 invlist_set_len(invlist, len - 1, offset);
9302 /* Here the new range doesn't extend any existing set. Add it */
9304 len += 2; /* Includes an element each for the start and end of range */
9306 /* If wll overflow the existing space, extend, which may cause the array to
9309 invlist_extend(invlist, len);
9311 /* Have to set len here to avoid assert failure in invlist_array() */
9312 invlist_set_len(invlist, len, offset);
9314 array = invlist_array(invlist);
9317 invlist_set_len(invlist, len, offset);
9320 /* The next item on the list starts the range, the one after that is
9321 * one past the new range. */
9322 array[len - 2] = start;
9323 if (end != UV_MAX) {
9324 array[len - 1] = end + 1;
9327 /* But if the end is the maximum representable on the machine, just let
9328 * the range have no end */
9329 invlist_set_len(invlist, len - 1, offset);
9334 Perl__invlist_search(SV* const invlist, const UV cp)
9336 /* Searches the inversion list for the entry that contains the input code
9337 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
9338 * return value is the index into the list's array of the range that
9339 * contains <cp>, that is, 'i' such that
9340 * array[i] <= cp < array[i+1]
9345 IV high = _invlist_len(invlist);
9346 const IV highest_element = high - 1;
9349 PERL_ARGS_ASSERT__INVLIST_SEARCH;
9351 /* If list is empty, return failure. */
9356 /* (We can't get the array unless we know the list is non-empty) */
9357 array = invlist_array(invlist);
9359 mid = invlist_previous_index(invlist);
9361 if (mid > highest_element) {
9362 mid = highest_element;
9365 /* <mid> contains the cache of the result of the previous call to this
9366 * function (0 the first time). See if this call is for the same result,
9367 * or if it is for mid-1. This is under the theory that calls to this
9368 * function will often be for related code points that are near each other.
9369 * And benchmarks show that caching gives better results. We also test
9370 * here if the code point is within the bounds of the list. These tests
9371 * replace others that would have had to be made anyway to make sure that
9372 * the array bounds were not exceeded, and these give us extra information
9373 * at the same time */
9374 if (cp >= array[mid]) {
9375 if (cp >= array[highest_element]) {
9376 return highest_element;
9379 /* Here, array[mid] <= cp < array[highest_element]. This means that
9380 * the final element is not the answer, so can exclude it; it also
9381 * means that <mid> is not the final element, so can refer to 'mid + 1'
9383 if (cp < array[mid + 1]) {
9389 else { /* cp < aray[mid] */
9390 if (cp < array[0]) { /* Fail if outside the array */
9394 if (cp >= array[mid - 1]) {
9399 /* Binary search. What we are looking for is <i> such that
9400 * array[i] <= cp < array[i+1]
9401 * The loop below converges on the i+1. Note that there may not be an
9402 * (i+1)th element in the array, and things work nonetheless */
9403 while (low < high) {
9404 mid = (low + high) / 2;
9405 assert(mid <= highest_element);
9406 if (array[mid] <= cp) { /* cp >= array[mid] */
9409 /* We could do this extra test to exit the loop early.
9410 if (cp < array[low]) {
9415 else { /* cp < array[mid] */
9422 invlist_set_previous_index(invlist, high);
9427 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9428 const bool complement_b, SV** output)
9430 /* Take the union of two inversion lists and point '*output' to it. On
9431 * input, '*output' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9432 * even 'a' or 'b'). If to an inversion list, the contents of the original
9433 * list will be replaced by the union. The first list, 'a', may be
9434 * NULL, in which case a copy of the second list is placed in '*output'.
9435 * If 'complement_b' is TRUE, the union is taken of the complement
9436 * (inversion) of 'b' instead of b itself.
9438 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9439 * Richard Gillam, published by Addison-Wesley, and explained at some
9440 * length there. The preface says to incorporate its examples into your
9441 * code at your own risk.
9443 * The algorithm is like a merge sort. */
9445 const UV* array_a; /* a's array */
9447 UV len_a; /* length of a's array */
9450 SV* u; /* the resulting union */
9454 UV i_a = 0; /* current index into a's array */
9458 /* running count, as explained in the algorithm source book; items are
9459 * stopped accumulating and are output when the count changes to/from 0.
9460 * The count is incremented when we start a range that's in an input's set,
9461 * and decremented when we start a range that's not in a set. So this
9462 * variable can be 0, 1, or 2. When it is 0 neither input is in their set,
9463 * and hence nothing goes into the union; 1, just one of the inputs is in
9464 * its set (and its current range gets added to the union); and 2 when both
9465 * inputs are in their sets. */
9468 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
9470 assert(*output == NULL || is_invlist(*output));
9472 len_b = _invlist_len(b);
9475 /* Here, 'b' is empty, hence it's complement is all possible code
9476 * points. So if the union includes the complement of 'b', it includes
9477 * everything, and we need not even look at 'a'. It's easiest to
9478 * create a new inversion list that matches everything. */
9480 SV* everything = _add_range_to_invlist(NULL, 0, UV_MAX);
9482 if (*output == NULL) { /* If the output didn't exist, just point it
9484 *output = everything;
9486 else { /* Otherwise, replace its contents with the new list */
9487 invlist_replace_list_destroys_src(*output, everything);
9488 SvREFCNT_dec_NN(everything);
9494 /* Here, we don't want the complement of 'b', and since 'b' is empty,
9495 * the union will come entirely from 'a'. If 'a' is NULL or empty, the
9496 * output will be empty */
9498 if (a == NULL || _invlist_len(a) == 0) {
9499 if (*output == NULL) {
9500 *output = _new_invlist(0);
9503 invlist_clear(*output);
9508 /* Here, 'a' is not empty, but 'b' is, so 'a' entirely determines the
9509 * union. We can just return a copy of 'a' if '*output' doesn't point
9510 * to an existing list */
9511 if (*output == NULL) {
9512 *output = invlist_clone(a, NULL);
9516 /* If the output is to overwrite 'a', we have a no-op, as it's
9522 /* Here, '*output' is to be overwritten by 'a' */
9523 u = invlist_clone(a, NULL);
9524 invlist_replace_list_destroys_src(*output, u);
9530 /* Here 'b' is not empty. See about 'a' */
9532 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
9534 /* Here, 'a' is empty (and b is not). That means the union will come
9535 * entirely from 'b'. If '*output' is NULL, we can directly return a
9536 * clone of 'b'. Otherwise, we replace the contents of '*output' with
9539 SV ** dest = (*output == NULL) ? output : &u;
9540 *dest = invlist_clone(b, NULL);
9542 _invlist_invert(*dest);
9546 invlist_replace_list_destroys_src(*output, u);
9553 /* Here both lists exist and are non-empty */
9554 array_a = invlist_array(a);
9555 array_b = invlist_array(b);
9557 /* If are to take the union of 'a' with the complement of b, set it
9558 * up so are looking at b's complement. */
9561 /* To complement, we invert: if the first element is 0, remove it. To
9562 * do this, we just pretend the array starts one later */
9563 if (array_b[0] == 0) {
9569 /* But if the first element is not zero, we pretend the list starts
9570 * at the 0 that is always stored immediately before the array. */
9576 /* Size the union for the worst case: that the sets are completely
9578 u = _new_invlist(len_a + len_b);
9580 /* Will contain U+0000 if either component does */
9581 array_u = _invlist_array_init(u, ( len_a > 0 && array_a[0] == 0)
9582 || (len_b > 0 && array_b[0] == 0));
9584 /* Go through each input list item by item, stopping when have exhausted
9586 while (i_a < len_a && i_b < len_b) {
9587 UV cp; /* The element to potentially add to the union's array */
9588 bool cp_in_set; /* is it in the the input list's set or not */
9590 /* We need to take one or the other of the two inputs for the union.
9591 * Since we are merging two sorted lists, we take the smaller of the
9592 * next items. In case of a tie, we take first the one that is in its
9593 * set. If we first took the one not in its set, it would decrement
9594 * the count, possibly to 0 which would cause it to be output as ending
9595 * the range, and the next time through we would take the same number,
9596 * and output it again as beginning the next range. By doing it the
9597 * opposite way, there is no possibility that the count will be
9598 * momentarily decremented to 0, and thus the two adjoining ranges will
9599 * be seamlessly merged. (In a tie and both are in the set or both not
9600 * in the set, it doesn't matter which we take first.) */
9601 if ( array_a[i_a] < array_b[i_b]
9602 || ( array_a[i_a] == array_b[i_b]
9603 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9605 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9606 cp = array_a[i_a++];
9609 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9610 cp = array_b[i_b++];
9613 /* Here, have chosen which of the two inputs to look at. Only output
9614 * if the running count changes to/from 0, which marks the
9615 * beginning/end of a range that's in the set */
9618 array_u[i_u++] = cp;
9625 array_u[i_u++] = cp;
9631 /* The loop above increments the index into exactly one of the input lists
9632 * each iteration, and ends when either index gets to its list end. That
9633 * means the other index is lower than its end, and so something is
9634 * remaining in that one. We decrement 'count', as explained below, if
9635 * that list is in its set. (i_a and i_b each currently index the element
9636 * beyond the one we care about.) */
9637 if ( (i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9638 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9643 /* Above we decremented 'count' if the list that had unexamined elements in
9644 * it was in its set. This has made it so that 'count' being non-zero
9645 * means there isn't anything left to output; and 'count' equal to 0 means
9646 * that what is left to output is precisely that which is left in the
9647 * non-exhausted input list.
9649 * To see why, note first that the exhausted input obviously has nothing
9650 * left to add to the union. If it was in its set at its end, that means
9651 * the set extends from here to the platform's infinity, and hence so does
9652 * the union and the non-exhausted set is irrelevant. The exhausted set
9653 * also contributed 1 to 'count'. If 'count' was 2, it got decremented to
9654 * 1, but if it was 1, the non-exhausted set wasn't in its set, and so
9655 * 'count' remains at 1. This is consistent with the decremented 'count'
9656 * != 0 meaning there's nothing left to add to the union.
9658 * But if the exhausted input wasn't in its set, it contributed 0 to
9659 * 'count', and the rest of the union will be whatever the other input is.
9660 * If 'count' was 0, neither list was in its set, and 'count' remains 0;
9661 * otherwise it gets decremented to 0. This is consistent with 'count'
9662 * == 0 meaning the remainder of the union is whatever is left in the
9663 * non-exhausted list. */
9668 IV copy_count = len_a - i_a;
9669 if (copy_count > 0) { /* The non-exhausted input is 'a' */
9670 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
9672 else { /* The non-exhausted input is b */
9673 copy_count = len_b - i_b;
9674 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
9676 len_u = i_u + copy_count;
9679 /* Set the result to the final length, which can change the pointer to
9680 * array_u, so re-find it. (Note that it is unlikely that this will
9681 * change, as we are shrinking the space, not enlarging it) */
9682 if (len_u != _invlist_len(u)) {
9683 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
9685 array_u = invlist_array(u);
9688 if (*output == NULL) { /* Simply return the new inversion list */
9692 /* Otherwise, overwrite the inversion list that was in '*output'. We
9693 * could instead free '*output', and then set it to 'u', but experience
9694 * has shown [perl #127392] that if the input is a mortal, we can get a
9695 * huge build-up of these during regex compilation before they get
9697 invlist_replace_list_destroys_src(*output, u);
9705 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9706 const bool complement_b, SV** i)
9708 /* Take the intersection of two inversion lists and point '*i' to it. On
9709 * input, '*i' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9710 * even 'a' or 'b'). If to an inversion list, the contents of the original
9711 * list will be replaced by the intersection. The first list, 'a', may be
9712 * NULL, in which case '*i' will be an empty list. If 'complement_b' is
9713 * TRUE, the result will be the intersection of 'a' and the complement (or
9714 * inversion) of 'b' instead of 'b' directly.
9716 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9717 * Richard Gillam, published by Addison-Wesley, and explained at some
9718 * length there. The preface says to incorporate its examples into your
9719 * code at your own risk. In fact, it had bugs
9721 * The algorithm is like a merge sort, and is essentially the same as the
9725 const UV* array_a; /* a's array */
9727 UV len_a; /* length of a's array */
9730 SV* r; /* the resulting intersection */
9734 UV i_a = 0; /* current index into a's array */
9738 /* running count of how many of the two inputs are postitioned at ranges
9739 * that are in their sets. As explained in the algorithm source book,
9740 * items are stopped accumulating and are output when the count changes
9741 * to/from 2. The count is incremented when we start a range that's in an
9742 * input's set, and decremented when we start a range that's not in a set.
9743 * Only when it is 2 are we in the intersection. */
9746 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
9748 assert(*i == NULL || is_invlist(*i));
9750 /* Special case if either one is empty */
9751 len_a = (a == NULL) ? 0 : _invlist_len(a);
9752 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
9753 if (len_a != 0 && complement_b) {
9755 /* Here, 'a' is not empty, therefore from the enclosing 'if', 'b'
9756 * must be empty. Here, also we are using 'b's complement, which
9757 * hence must be every possible code point. Thus the intersection
9760 if (*i == a) { /* No-op */
9765 *i = invlist_clone(a, NULL);
9769 r = invlist_clone(a, NULL);
9770 invlist_replace_list_destroys_src(*i, r);
9775 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
9776 * intersection must be empty */
9778 *i = _new_invlist(0);
9786 /* Here both lists exist and are non-empty */
9787 array_a = invlist_array(a);
9788 array_b = invlist_array(b);
9790 /* If are to take the intersection of 'a' with the complement of b, set it
9791 * up so are looking at b's complement. */
9794 /* To complement, we invert: if the first element is 0, remove it. To
9795 * do this, we just pretend the array starts one later */
9796 if (array_b[0] == 0) {
9802 /* But if the first element is not zero, we pretend the list starts
9803 * at the 0 that is always stored immediately before the array. */
9809 /* Size the intersection for the worst case: that the intersection ends up
9810 * fragmenting everything to be completely disjoint */
9811 r= _new_invlist(len_a + len_b);
9813 /* Will contain U+0000 iff both components do */
9814 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
9815 && len_b > 0 && array_b[0] == 0);
9817 /* Go through each list item by item, stopping when have exhausted one of
9819 while (i_a < len_a && i_b < len_b) {
9820 UV cp; /* The element to potentially add to the intersection's
9822 bool cp_in_set; /* Is it in the input list's set or not */
9824 /* We need to take one or the other of the two inputs for the
9825 * intersection. Since we are merging two sorted lists, we take the
9826 * smaller of the next items. In case of a tie, we take first the one
9827 * that is not in its set (a difference from the union algorithm). If
9828 * we first took the one in its set, it would increment the count,
9829 * possibly to 2 which would cause it to be output as starting a range
9830 * in the intersection, and the next time through we would take that
9831 * same number, and output it again as ending the set. By doing the
9832 * opposite of this, there is no possibility that the count will be
9833 * momentarily incremented to 2. (In a tie and both are in the set or
9834 * both not in the set, it doesn't matter which we take first.) */
9835 if ( array_a[i_a] < array_b[i_b]
9836 || ( array_a[i_a] == array_b[i_b]
9837 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9839 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9840 cp = array_a[i_a++];
9843 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9847 /* Here, have chosen which of the two inputs to look at. Only output
9848 * if the running count changes to/from 2, which marks the
9849 * beginning/end of a range that's in the intersection */
9853 array_r[i_r++] = cp;
9858 array_r[i_r++] = cp;
9865 /* The loop above increments the index into exactly one of the input lists
9866 * each iteration, and ends when either index gets to its list end. That
9867 * means the other index is lower than its end, and so something is
9868 * remaining in that one. We increment 'count', as explained below, if the
9869 * exhausted list was in its set. (i_a and i_b each currently index the
9870 * element beyond the one we care about.) */
9871 if ( (i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9872 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9877 /* Above we incremented 'count' if the exhausted list was in its set. This
9878 * has made it so that 'count' being below 2 means there is nothing left to
9879 * output; otheriwse what's left to add to the intersection is precisely
9880 * that which is left in the non-exhausted input list.
9882 * To see why, note first that the exhausted input obviously has nothing
9883 * left to affect the intersection. If it was in its set at its end, that
9884 * means the set extends from here to the platform's infinity, and hence
9885 * anything in the non-exhausted's list will be in the intersection, and
9886 * anything not in it won't be. Hence, the rest of the intersection is
9887 * precisely what's in the non-exhausted list The exhausted set also
9888 * contributed 1 to 'count', meaning 'count' was at least 1. Incrementing
9889 * it means 'count' is now at least 2. This is consistent with the
9890 * incremented 'count' being >= 2 means to add the non-exhausted list to
9893 * But if the exhausted input wasn't in its set, it contributed 0 to
9894 * 'count', and the intersection can't include anything further; the
9895 * non-exhausted set is irrelevant. 'count' was at most 1, and doesn't get
9896 * incremented. This is consistent with 'count' being < 2 meaning nothing
9897 * further to add to the intersection. */
9898 if (count < 2) { /* Nothing left to put in the intersection. */
9901 else { /* copy the non-exhausted list, unchanged. */
9902 IV copy_count = len_a - i_a;
9903 if (copy_count > 0) { /* a is the one with stuff left */
9904 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
9906 else { /* b is the one with stuff left */
9907 copy_count = len_b - i_b;
9908 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
9910 len_r = i_r + copy_count;
9913 /* Set the result to the final length, which can change the pointer to
9914 * array_r, so re-find it. (Note that it is unlikely that this will
9915 * change, as we are shrinking the space, not enlarging it) */
9916 if (len_r != _invlist_len(r)) {
9917 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
9919 array_r = invlist_array(r);
9922 if (*i == NULL) { /* Simply return the calculated intersection */
9925 else { /* Otherwise, replace the existing inversion list in '*i'. We could
9926 instead free '*i', and then set it to 'r', but experience has
9927 shown [perl #127392] that if the input is a mortal, we can get a
9928 huge build-up of these during regex compilation before they get
9931 invlist_replace_list_destroys_src(*i, r);
9943 Perl__add_range_to_invlist(pTHX_ SV* invlist, UV start, UV end)
9945 /* Add the range from 'start' to 'end' inclusive to the inversion list's
9946 * set. A pointer to the inversion list is returned. This may actually be
9947 * a new list, in which case the passed in one has been destroyed. The
9948 * passed-in inversion list can be NULL, in which case a new one is created
9949 * with just the one range in it. The new list is not necessarily
9950 * NUL-terminated. Space is not freed if the inversion list shrinks as a
9951 * result of this function. The gain would not be large, and in many
9952 * cases, this is called multiple times on a single inversion list, so
9953 * anything freed may almost immediately be needed again.
9955 * This used to mostly call the 'union' routine, but that is much more
9956 * heavyweight than really needed for a single range addition */
9958 UV* array; /* The array implementing the inversion list */
9959 UV len; /* How many elements in 'array' */
9960 SSize_t i_s; /* index into the invlist array where 'start'
9962 SSize_t i_e = 0; /* And the index where 'end' should go */
9963 UV cur_highest; /* The highest code point in the inversion list
9964 upon entry to this function */
9966 /* This range becomes the whole inversion list if none already existed */
9967 if (invlist == NULL) {
9968 invlist = _new_invlist(2);
9969 _append_range_to_invlist(invlist, start, end);
9973 /* Likewise, if the inversion list is currently empty */
9974 len = _invlist_len(invlist);
9976 _append_range_to_invlist(invlist, start, end);
9980 /* Starting here, we have to know the internals of the list */
9981 array = invlist_array(invlist);
9983 /* If the new range ends higher than the current highest ... */
9984 cur_highest = invlist_highest(invlist);
9985 if (end > cur_highest) {
9987 /* If the whole range is higher, we can just append it */
9988 if (start > cur_highest) {
9989 _append_range_to_invlist(invlist, start, end);
9993 /* Otherwise, add the portion that is higher ... */
9994 _append_range_to_invlist(invlist, cur_highest + 1, end);
9996 /* ... and continue on below to handle the rest. As a result of the
9997 * above append, we know that the index of the end of the range is the
9998 * final even numbered one of the array. Recall that the final element
9999 * always starts a range that extends to infinity. If that range is in
10000 * the set (meaning the set goes from here to infinity), it will be an
10001 * even index, but if it isn't in the set, it's odd, and the final
10002 * range in the set is one less, which is even. */
10003 if (end == UV_MAX) {
10011 /* We have dealt with appending, now see about prepending. If the new
10012 * range starts lower than the current lowest ... */
10013 if (start < array[0]) {
10015 /* Adding something which has 0 in it is somewhat tricky, and uncommon.
10016 * Let the union code handle it, rather than having to know the
10017 * trickiness in two code places. */
10018 if (UNLIKELY(start == 0)) {
10021 range_invlist = _new_invlist(2);
10022 _append_range_to_invlist(range_invlist, start, end);
10024 _invlist_union(invlist, range_invlist, &invlist);
10026 SvREFCNT_dec_NN(range_invlist);
10031 /* If the whole new range comes before the first entry, and doesn't
10032 * extend it, we have to insert it as an additional range */
10033 if (end < array[0] - 1) {
10035 goto splice_in_new_range;
10038 /* Here the new range adjoins the existing first range, extending it
10042 /* And continue on below to handle the rest. We know that the index of
10043 * the beginning of the range is the first one of the array */
10046 else { /* Not prepending any part of the new range to the existing list.
10047 * Find where in the list it should go. This finds i_s, such that:
10048 * invlist[i_s] <= start < array[i_s+1]
10050 i_s = _invlist_search(invlist, start);
10053 /* At this point, any extending before the beginning of the inversion list
10054 * and/or after the end has been done. This has made it so that, in the
10055 * code below, each endpoint of the new range is either in a range that is
10056 * in the set, or is in a gap between two ranges that are. This means we
10057 * don't have to worry about exceeding the array bounds.
10059 * Find where in the list the new range ends (but we can skip this if we
10060 * have already determined what it is, or if it will be the same as i_s,
10061 * which we already have computed) */
10063 i_e = (start == end)
10065 : _invlist_search(invlist, end);
10068 /* Here generally invlist[i_e] <= end < array[i_e+1]. But if invlist[i_e]
10069 * is a range that goes to infinity there is no element at invlist[i_e+1],
10070 * so only the first relation holds. */
10072 if ( ! ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
10074 /* Here, the ranges on either side of the beginning of the new range
10075 * are in the set, and this range starts in the gap between them.
10077 * The new range extends the range above it downwards if the new range
10078 * ends at or above that range's start */
10079 const bool extends_the_range_above = ( end == UV_MAX
10080 || end + 1 >= array[i_s+1]);
10082 /* The new range extends the range below it upwards if it begins just
10083 * after where that range ends */
10084 if (start == array[i_s]) {
10086 /* If the new range fills the entire gap between the other ranges,
10087 * they will get merged together. Other ranges may also get
10088 * merged, depending on how many of them the new range spans. In
10089 * the general case, we do the merge later, just once, after we
10090 * figure out how many to merge. But in the case where the new
10091 * range exactly spans just this one gap (possibly extending into
10092 * the one above), we do the merge here, and an early exit. This
10093 * is done here to avoid having to special case later. */
10094 if (i_e - i_s <= 1) {
10096 /* If i_e - i_s == 1, it means that the new range terminates
10097 * within the range above, and hence 'extends_the_range_above'
10098 * must be true. (If the range above it extends to infinity,
10099 * 'i_s+2' will be above the array's limit, but 'len-i_s-2'
10100 * will be 0, so no harm done.) */
10101 if (extends_the_range_above) {
10102 Move(array + i_s + 2, array + i_s, len - i_s - 2, UV);
10103 invlist_set_len(invlist,
10105 *(get_invlist_offset_addr(invlist)));
10109 /* Here, i_e must == i_s. We keep them in sync, as they apply
10110 * to the same range, and below we are about to decrement i_s
10115 /* Here, the new range is adjacent to the one below. (It may also
10116 * span beyond the range above, but that will get resolved later.)
10117 * Extend the range below to include this one. */
10118 array[i_s] = (end == UV_MAX) ? UV_MAX : end + 1;
10120 start = array[i_s];
10122 else if (extends_the_range_above) {
10124 /* Here the new range only extends the range above it, but not the
10125 * one below. It merges with the one above. Again, we keep i_e
10126 * and i_s in sync if they point to the same range */
10131 array[i_s] = start;
10135 /* Here, we've dealt with the new range start extending any adjoining
10138 * If the new range extends to infinity, it is now the final one,
10139 * regardless of what was there before */
10140 if (UNLIKELY(end == UV_MAX)) {
10141 invlist_set_len(invlist, i_s + 1, *(get_invlist_offset_addr(invlist)));
10145 /* If i_e started as == i_s, it has also been dealt with,
10146 * and been updated to the new i_s, which will fail the following if */
10147 if (! ELEMENT_RANGE_MATCHES_INVLIST(i_e)) {
10149 /* Here, the ranges on either side of the end of the new range are in
10150 * the set, and this range ends in the gap between them.
10152 * If this range is adjacent to (hence extends) the range above it, it
10153 * becomes part of that range; likewise if it extends the range below,
10154 * it becomes part of that range */
10155 if (end + 1 == array[i_e+1]) {
10157 array[i_e] = start;
10159 else if (start <= array[i_e]) {
10160 array[i_e] = end + 1;
10167 /* If the range fits entirely in an existing range (as possibly already
10168 * extended above), it doesn't add anything new */
10169 if (ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
10173 /* Here, no part of the range is in the list. Must add it. It will
10174 * occupy 2 more slots */
10175 splice_in_new_range:
10177 invlist_extend(invlist, len + 2);
10178 array = invlist_array(invlist);
10179 /* Move the rest of the array down two slots. Don't include any
10181 Move(array + i_e + 1, array + i_e + 3, len - i_e - 1, UV);
10183 /* Do the actual splice */
10184 array[i_e+1] = start;
10185 array[i_e+2] = end + 1;
10186 invlist_set_len(invlist, len + 2, *(get_invlist_offset_addr(invlist)));
10190 /* Here the new range crossed the boundaries of a pre-existing range. The
10191 * code above has adjusted things so that both ends are in ranges that are
10192 * in the set. This means everything in between must also be in the set.
10193 * Just squash things together */
10194 Move(array + i_e + 1, array + i_s + 1, len - i_e - 1, UV);
10195 invlist_set_len(invlist,
10197 *(get_invlist_offset_addr(invlist)));
10203 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
10204 UV** other_elements_ptr)
10206 /* Create and return an inversion list whose contents are to be populated
10207 * by the caller. The caller gives the number of elements (in 'size') and
10208 * the very first element ('element0'). This function will set
10209 * '*other_elements_ptr' to an array of UVs, where the remaining elements
10210 * are to be placed.
10212 * Obviously there is some trust involved that the caller will properly
10213 * fill in the other elements of the array.
10215 * (The first element needs to be passed in, as the underlying code does
10216 * things differently depending on whether it is zero or non-zero) */
10218 SV* invlist = _new_invlist(size);
10221 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
10223 invlist = add_cp_to_invlist(invlist, element0);
10224 offset = *get_invlist_offset_addr(invlist);
10226 invlist_set_len(invlist, size, offset);
10227 *other_elements_ptr = invlist_array(invlist) + 1;
10233 PERL_STATIC_INLINE SV*
10234 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
10235 return _add_range_to_invlist(invlist, cp, cp);
10238 #ifndef PERL_IN_XSUB_RE
10240 Perl__invlist_invert(pTHX_ SV* const invlist)
10242 /* Complement the input inversion list. This adds a 0 if the list didn't
10243 * have a zero; removes it otherwise. As described above, the data
10244 * structure is set up so that this is very efficient */
10246 PERL_ARGS_ASSERT__INVLIST_INVERT;
10248 assert(! invlist_is_iterating(invlist));
10250 /* The inverse of matching nothing is matching everything */
10251 if (_invlist_len(invlist) == 0) {
10252 _append_range_to_invlist(invlist, 0, UV_MAX);
10256 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
10260 Perl_invlist_clone(pTHX_ SV* const invlist, SV* new_invlist)
10262 /* Return a new inversion list that is a copy of the input one, which is
10263 * unchanged. The new list will not be mortal even if the old one was. */
10265 const STRLEN nominal_length = _invlist_len(invlist);
10266 const STRLEN physical_length = SvCUR(invlist);
10267 const bool offset = *(get_invlist_offset_addr(invlist));
10269 PERL_ARGS_ASSERT_INVLIST_CLONE;
10271 if (new_invlist == NULL) {
10272 new_invlist = _new_invlist(nominal_length);
10275 sv_upgrade(new_invlist, SVt_INVLIST);
10276 initialize_invlist_guts(new_invlist, nominal_length);
10279 *(get_invlist_offset_addr(new_invlist)) = offset;
10280 invlist_set_len(new_invlist, nominal_length, offset);
10281 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
10283 return new_invlist;
10288 PERL_STATIC_INLINE STRLEN*
10289 S_get_invlist_iter_addr(SV* invlist)
10291 /* Return the address of the UV that contains the current iteration
10294 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
10296 assert(is_invlist(invlist));
10298 return &(((XINVLIST*) SvANY(invlist))->iterator);
10301 PERL_STATIC_INLINE void
10302 S_invlist_iterinit(SV* invlist) /* Initialize iterator for invlist */
10304 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
10306 *get_invlist_iter_addr(invlist) = 0;
10309 PERL_STATIC_INLINE void
10310 S_invlist_iterfinish(SV* invlist)
10312 /* Terminate iterator for invlist. This is to catch development errors.
10313 * Any iteration that is interrupted before completed should call this
10314 * function. Functions that add code points anywhere else but to the end
10315 * of an inversion list assert that they are not in the middle of an
10316 * iteration. If they were, the addition would make the iteration
10317 * problematical: if the iteration hadn't reached the place where things
10318 * were being added, it would be ok */
10320 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
10322 *get_invlist_iter_addr(invlist) = (STRLEN) UV_MAX;
10326 S_invlist_iternext(SV* invlist, UV* start, UV* end)
10328 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
10329 * This call sets in <*start> and <*end>, the next range in <invlist>.
10330 * Returns <TRUE> if successful and the next call will return the next
10331 * range; <FALSE> if was already at the end of the list. If the latter,
10332 * <*start> and <*end> are unchanged, and the next call to this function
10333 * will start over at the beginning of the list */
10335 STRLEN* pos = get_invlist_iter_addr(invlist);
10336 UV len = _invlist_len(invlist);
10339 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
10342 *pos = (STRLEN) UV_MAX; /* Force iterinit() to be required next time */
10346 array = invlist_array(invlist);
10348 *start = array[(*pos)++];
10354 *end = array[(*pos)++] - 1;
10360 PERL_STATIC_INLINE UV
10361 S_invlist_highest(SV* const invlist)
10363 /* Returns the highest code point that matches an inversion list. This API
10364 * has an ambiguity, as it returns 0 under either the highest is actually
10365 * 0, or if the list is empty. If this distinction matters to you, check
10366 * for emptiness before calling this function */
10368 UV len = _invlist_len(invlist);
10371 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
10377 array = invlist_array(invlist);
10379 /* The last element in the array in the inversion list always starts a
10380 * range that goes to infinity. That range may be for code points that are
10381 * matched in the inversion list, or it may be for ones that aren't
10382 * matched. In the latter case, the highest code point in the set is one
10383 * less than the beginning of this range; otherwise it is the final element
10384 * of this range: infinity */
10385 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
10387 : array[len - 1] - 1;
10391 S_invlist_contents(pTHX_ SV* const invlist, const bool traditional_style)
10393 /* Get the contents of an inversion list into a string SV so that they can
10394 * be printed out. If 'traditional_style' is TRUE, it uses the format
10395 * traditionally done for debug tracing; otherwise it uses a format
10396 * suitable for just copying to the output, with blanks between ranges and
10397 * a dash between range components */
10401 const char intra_range_delimiter = (traditional_style ? '\t' : '-');
10402 const char inter_range_delimiter = (traditional_style ? '\n' : ' ');
10404 if (traditional_style) {
10405 output = newSVpvs("\n");
10408 output = newSVpvs("");
10411 PERL_ARGS_ASSERT_INVLIST_CONTENTS;
10413 assert(! invlist_is_iterating(invlist));
10415 invlist_iterinit(invlist);
10416 while (invlist_iternext(invlist, &start, &end)) {
10417 if (end == UV_MAX) {
10418 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%cINFTY%c",
10419 start, intra_range_delimiter,
10420 inter_range_delimiter);
10422 else if (end != start) {
10423 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c%04" UVXf "%c",
10425 intra_range_delimiter,
10426 end, inter_range_delimiter);
10429 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c",
10430 start, inter_range_delimiter);
10434 if (SvCUR(output) && ! traditional_style) {/* Get rid of trailing blank */
10435 SvCUR_set(output, SvCUR(output) - 1);
10441 #ifndef PERL_IN_XSUB_RE
10443 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
10444 const char * const indent, SV* const invlist)
10446 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
10447 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
10448 * the string 'indent'. The output looks like this:
10449 [0] 0x000A .. 0x000D
10451 [4] 0x2028 .. 0x2029
10452 [6] 0x3104 .. INFTY
10453 * This means that the first range of code points matched by the list are
10454 * 0xA through 0xD; the second range contains only the single code point
10455 * 0x85, etc. An inversion list is an array of UVs. Two array elements
10456 * are used to define each range (except if the final range extends to
10457 * infinity, only a single element is needed). The array index of the
10458 * first element for the corresponding range is given in brackets. */
10463 PERL_ARGS_ASSERT__INVLIST_DUMP;
10465 if (invlist_is_iterating(invlist)) {
10466 Perl_dump_indent(aTHX_ level, file,
10467 "%sCan't dump inversion list because is in middle of iterating\n",
10472 invlist_iterinit(invlist);
10473 while (invlist_iternext(invlist, &start, &end)) {
10474 if (end == UV_MAX) {
10475 Perl_dump_indent(aTHX_ level, file,
10476 "%s[%" UVuf "] 0x%04" UVXf " .. INFTY\n",
10477 indent, (UV)count, start);
10479 else if (end != start) {
10480 Perl_dump_indent(aTHX_ level, file,
10481 "%s[%" UVuf "] 0x%04" UVXf " .. 0x%04" UVXf "\n",
10482 indent, (UV)count, start, end);
10485 Perl_dump_indent(aTHX_ level, file, "%s[%" UVuf "] 0x%04" UVXf "\n",
10486 indent, (UV)count, start);
10494 #if defined(PERL_ARGS_ASSERT__INVLISTEQ) && !defined(PERL_IN_XSUB_RE)
10496 Perl__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
10498 /* Return a boolean as to if the two passed in inversion lists are
10499 * identical. The final argument, if TRUE, says to take the complement of
10500 * the second inversion list before doing the comparison */
10502 const UV len_a = _invlist_len(a);
10503 UV len_b = _invlist_len(b);
10505 const UV* array_a = NULL;
10506 const UV* array_b = NULL;
10508 PERL_ARGS_ASSERT__INVLISTEQ;
10510 /* This code avoids accessing the arrays unless it knows the length is
10515 return ! complement_b;
10519 array_a = invlist_array(a);
10523 array_b = invlist_array(b);
10526 /* If are to compare 'a' with the complement of b, set it
10527 * up so are looking at b's complement. */
10528 if (complement_b) {
10530 /* The complement of nothing is everything, so <a> would have to have
10531 * just one element, starting at zero (ending at infinity) */
10533 return (len_a == 1 && array_a[0] == 0);
10535 if (array_b[0] == 0) {
10537 /* Otherwise, to complement, we invert. Here, the first element is
10538 * 0, just remove it. To do this, we just pretend the array starts
10546 /* But if the first element is not zero, we pretend the list starts
10547 * at the 0 that is always stored immediately before the array. */
10553 return len_a == len_b
10554 && memEQ(array_a, array_b, len_a * sizeof(array_a[0]));
10560 * As best we can, determine the characters that can match the start of
10561 * the given EXACTF-ish node. This is for use in creating ssc nodes, so there
10562 * can be false positive matches
10564 * Returns the invlist as a new SV*; it is the caller's responsibility to
10565 * call SvREFCNT_dec() when done with it.
10568 S__make_exactf_invlist(pTHX_ RExC_state_t *pRExC_state, regnode *node)
10571 const U8 * s = (U8*)STRING(node);
10572 SSize_t bytelen = STR_LEN(node);
10574 /* Start out big enough for 2 separate code points */
10575 SV* invlist = _new_invlist(4);
10577 PERL_ARGS_ASSERT__MAKE_EXACTF_INVLIST;
10582 /* We punt and assume can match anything if the node begins
10583 * with a multi-character fold. Things are complicated. For
10584 * example, /ffi/i could match any of:
10585 * "\N{LATIN SMALL LIGATURE FFI}"
10586 * "\N{LATIN SMALL LIGATURE FF}I"
10587 * "F\N{LATIN SMALL LIGATURE FI}"
10588 * plus several other things; and making sure we have all the
10589 * possibilities is hard. */
10590 if (is_MULTI_CHAR_FOLD_latin1_safe(s, s + bytelen)) {
10591 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10594 /* Any Latin1 range character can potentially match any
10595 * other depending on the locale, and in Turkic locales, U+130 and
10597 if (OP(node) == EXACTFL) {
10598 _invlist_union(invlist, PL_Latin1, &invlist);
10599 invlist = add_cp_to_invlist(invlist,
10600 LATIN_SMALL_LETTER_DOTLESS_I);
10601 invlist = add_cp_to_invlist(invlist,
10602 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
10605 /* But otherwise, it matches at least itself. We can
10606 * quickly tell if it has a distinct fold, and if so,
10607 * it matches that as well */
10608 invlist = add_cp_to_invlist(invlist, uc);
10609 if (IS_IN_SOME_FOLD_L1(uc))
10610 invlist = add_cp_to_invlist(invlist, PL_fold_latin1[uc]);
10613 /* Some characters match above-Latin1 ones under /i. This
10614 * is true of EXACTFL ones when the locale is UTF-8 */
10615 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(uc)
10616 && (! isASCII(uc) || (OP(node) != EXACTFAA
10617 && OP(node) != EXACTFAA_NO_TRIE)))
10619 add_above_Latin1_folds(pRExC_state, (U8) uc, &invlist);
10623 else { /* Pattern is UTF-8 */
10624 U8 folded[UTF8_MAX_FOLD_CHAR_EXPAND * UTF8_MAXBYTES_CASE + 1] = { '\0' };
10625 const U8* e = s + bytelen;
10628 fc = uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
10630 /* The only code points that aren't folded in a UTF EXACTFish
10631 * node are are the problematic ones in EXACTFL nodes */
10632 if (OP(node) == EXACTFL && is_PROBLEMATIC_LOCALE_FOLDEDS_START_cp(uc)) {
10633 /* We need to check for the possibility that this EXACTFL
10634 * node begins with a multi-char fold. Therefore we fold
10635 * the first few characters of it so that we can make that
10641 for (i = 0; i < UTF8_MAX_FOLD_CHAR_EXPAND && s < e; i++) {
10643 *(d++) = (U8) toFOLD(*s);
10644 if (fc < 0) { /* Save the first fold */
10651 UV fold = toFOLD_utf8_safe(s, e, d, &len);
10652 if (fc < 0) { /* Save the first fold */
10660 /* And set up so the code below that looks in this folded
10661 * buffer instead of the node's string */
10666 /* When we reach here 's' points to the fold of the first
10667 * character(s) of the node; and 'e' points to far enough along
10668 * the folded string to be just past any possible multi-char
10671 * Unlike the non-UTF-8 case, the macro for determining if a
10672 * string is a multi-char fold requires all the characters to
10673 * already be folded. This is because of all the complications
10674 * if not. Note that they are folded anyway, except in EXACTFL
10675 * nodes. Like the non-UTF case above, we punt if the node
10676 * begins with a multi-char fold */
10678 if (is_MULTI_CHAR_FOLD_utf8_safe(s, e)) {
10679 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10681 else { /* Single char fold */
10683 unsigned int first_fold;
10684 const unsigned int * remaining_folds;
10685 Size_t folds_count;
10687 /* It matches itself */
10688 invlist = add_cp_to_invlist(invlist, fc);
10690 /* ... plus all the things that fold to it, which are found in
10691 * PL_utf8_foldclosures */
10692 folds_count = _inverse_folds(fc, &first_fold,
10694 for (k = 0; k < folds_count; k++) {
10695 UV c = (k == 0) ? first_fold : remaining_folds[k-1];
10697 /* /aa doesn't allow folds between ASCII and non- */
10698 if ( (OP(node) == EXACTFAA || OP(node) == EXACTFAA_NO_TRIE)
10699 && isASCII(c) != isASCII(fc))
10704 invlist = add_cp_to_invlist(invlist, c);
10707 if (OP(node) == EXACTFL) {
10709 /* If either [iI] are present in an EXACTFL node the above code
10710 * should have added its normal case pair, but under a Turkish
10711 * locale they could match instead the case pairs from it. Add
10712 * those as potential matches as well */
10713 if (isALPHA_FOLD_EQ(fc, 'I')) {
10714 invlist = add_cp_to_invlist(invlist,
10715 LATIN_SMALL_LETTER_DOTLESS_I);
10716 invlist = add_cp_to_invlist(invlist,
10717 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
10719 else if (fc == LATIN_SMALL_LETTER_DOTLESS_I) {
10720 invlist = add_cp_to_invlist(invlist, 'I');
10722 else if (fc == LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE) {
10723 invlist = add_cp_to_invlist(invlist, 'i');
10732 #undef HEADER_LENGTH
10733 #undef TO_INTERNAL_SIZE
10734 #undef FROM_INTERNAL_SIZE
10735 #undef INVLIST_VERSION_ID
10737 /* End of inversion list object */
10740 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
10742 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
10743 * constructs, and updates RExC_flags with them. On input, RExC_parse
10744 * should point to the first flag; it is updated on output to point to the
10745 * final ')' or ':'. There needs to be at least one flag, or this will
10748 /* for (?g), (?gc), and (?o) warnings; warning
10749 about (?c) will warn about (?g) -- japhy */
10751 #define WASTED_O 0x01
10752 #define WASTED_G 0x02
10753 #define WASTED_C 0x04
10754 #define WASTED_GC (WASTED_G|WASTED_C)
10755 I32 wastedflags = 0x00;
10756 U32 posflags = 0, negflags = 0;
10757 U32 *flagsp = &posflags;
10758 char has_charset_modifier = '\0';
10760 bool has_use_defaults = FALSE;
10761 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
10762 int x_mod_count = 0;
10764 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
10766 /* '^' as an initial flag sets certain defaults */
10767 if (UCHARAT(RExC_parse) == '^') {
10769 has_use_defaults = TRUE;
10770 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
10771 cs = (RExC_uni_semantics)
10772 ? REGEX_UNICODE_CHARSET
10773 : REGEX_DEPENDS_CHARSET;
10774 set_regex_charset(&RExC_flags, cs);
10777 cs = get_regex_charset(RExC_flags);
10778 if ( cs == REGEX_DEPENDS_CHARSET
10779 && RExC_uni_semantics)
10781 cs = REGEX_UNICODE_CHARSET;
10785 while (RExC_parse < RExC_end) {
10786 /* && strchr("iogcmsx", *RExC_parse) */
10787 /* (?g), (?gc) and (?o) are useless here
10788 and must be globally applied -- japhy */
10789 switch (*RExC_parse) {
10791 /* Code for the imsxn flags */
10792 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp, x_mod_count);
10794 case LOCALE_PAT_MOD:
10795 if (has_charset_modifier) {
10796 goto excess_modifier;
10798 else if (flagsp == &negflags) {
10801 cs = REGEX_LOCALE_CHARSET;
10802 has_charset_modifier = LOCALE_PAT_MOD;
10804 case UNICODE_PAT_MOD:
10805 if (has_charset_modifier) {
10806 goto excess_modifier;
10808 else if (flagsp == &negflags) {
10811 cs = REGEX_UNICODE_CHARSET;
10812 has_charset_modifier = UNICODE_PAT_MOD;
10814 case ASCII_RESTRICT_PAT_MOD:
10815 if (flagsp == &negflags) {
10818 if (has_charset_modifier) {
10819 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
10820 goto excess_modifier;
10822 /* Doubled modifier implies more restricted */
10823 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
10826 cs = REGEX_ASCII_RESTRICTED_CHARSET;
10828 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
10830 case DEPENDS_PAT_MOD:
10831 if (has_use_defaults) {
10832 goto fail_modifiers;
10834 else if (flagsp == &negflags) {
10837 else if (has_charset_modifier) {
10838 goto excess_modifier;
10841 /* The dual charset means unicode semantics if the
10842 * pattern (or target, not known until runtime) are
10843 * utf8, or something in the pattern indicates unicode
10845 cs = (RExC_uni_semantics)
10846 ? REGEX_UNICODE_CHARSET
10847 : REGEX_DEPENDS_CHARSET;
10848 has_charset_modifier = DEPENDS_PAT_MOD;
10852 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
10853 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
10855 else if (has_charset_modifier == *(RExC_parse - 1)) {
10856 vFAIL2("Regexp modifier \"%c\" may not appear twice",
10857 *(RExC_parse - 1));
10860 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
10862 NOT_REACHED; /*NOTREACHED*/
10865 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
10866 *(RExC_parse - 1));
10867 NOT_REACHED; /*NOTREACHED*/
10868 case ONCE_PAT_MOD: /* 'o' */
10869 case GLOBAL_PAT_MOD: /* 'g' */
10870 if (ckWARN(WARN_REGEXP)) {
10871 const I32 wflagbit = *RExC_parse == 'o'
10874 if (! (wastedflags & wflagbit) ) {
10875 wastedflags |= wflagbit;
10876 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10879 "Useless (%s%c) - %suse /%c modifier",
10880 flagsp == &negflags ? "?-" : "?",
10882 flagsp == &negflags ? "don't " : "",
10889 case CONTINUE_PAT_MOD: /* 'c' */
10890 if (ckWARN(WARN_REGEXP)) {
10891 if (! (wastedflags & WASTED_C) ) {
10892 wastedflags |= WASTED_GC;
10893 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10896 "Useless (%sc) - %suse /gc modifier",
10897 flagsp == &negflags ? "?-" : "?",
10898 flagsp == &negflags ? "don't " : ""
10903 case KEEPCOPY_PAT_MOD: /* 'p' */
10904 if (flagsp == &negflags) {
10905 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
10907 *flagsp |= RXf_PMf_KEEPCOPY;
10911 /* A flag is a default iff it is following a minus, so
10912 * if there is a minus, it means will be trying to
10913 * re-specify a default which is an error */
10914 if (has_use_defaults || flagsp == &negflags) {
10915 goto fail_modifiers;
10917 flagsp = &negflags;
10918 wastedflags = 0; /* reset so (?g-c) warns twice */
10924 if ((posflags & (RXf_PMf_EXTENDED|RXf_PMf_EXTENDED_MORE)) == RXf_PMf_EXTENDED) {
10925 negflags |= RXf_PMf_EXTENDED_MORE;
10927 RExC_flags |= posflags;
10929 if (negflags & RXf_PMf_EXTENDED) {
10930 negflags |= RXf_PMf_EXTENDED_MORE;
10932 RExC_flags &= ~negflags;
10933 set_regex_charset(&RExC_flags, cs);
10938 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
10939 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10940 vFAIL2utf8f("Sequence (%" UTF8f "...) not recognized",
10941 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
10942 NOT_REACHED; /*NOTREACHED*/
10945 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10948 vFAIL("Sequence (?... not terminated");
10952 - reg - regular expression, i.e. main body or parenthesized thing
10954 * Caller must absorb opening parenthesis.
10956 * Combining parenthesis handling with the base level of regular expression
10957 * is a trifle forced, but the need to tie the tails of the branches to what
10958 * follows makes it hard to avoid.
10960 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
10962 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
10964 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
10967 PERL_STATIC_INLINE regnode_offset
10968 S_handle_named_backref(pTHX_ RExC_state_t *pRExC_state,
10970 char * parse_start,
10974 regnode_offset ret;
10975 char* name_start = RExC_parse;
10977 SV *sv_dat = reg_scan_name(pRExC_state, REG_RSN_RETURN_DATA);
10978 GET_RE_DEBUG_FLAGS_DECL;
10980 PERL_ARGS_ASSERT_HANDLE_NAMED_BACKREF;
10982 if (RExC_parse == name_start || *RExC_parse != ch) {
10983 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
10984 vFAIL2("Sequence %.3s... not terminated", parse_start);
10988 num = add_data( pRExC_state, STR_WITH_LEN("S"));
10989 RExC_rxi->data->data[num]=(void*)sv_dat;
10990 SvREFCNT_inc_simple_void_NN(sv_dat);
10993 ret = reganode(pRExC_state,
10996 : (ASCII_FOLD_RESTRICTED)
10998 : (AT_LEAST_UNI_SEMANTICS)
11004 *flagp |= HASWIDTH;
11006 Set_Node_Offset(REGNODE_p(ret), parse_start+1);
11007 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
11009 nextchar(pRExC_state);
11013 /* On success, returns the offset at which any next node should be placed into
11014 * the regex engine program being compiled.
11016 * Returns 0 otherwise, with *flagp set to indicate why:
11017 * TRYAGAIN at the end of (?) that only sets flags.
11018 * RESTART_PARSE if the parse needs to be restarted, or'd with
11019 * NEED_UTF8 if the pattern needs to be upgraded to UTF-8.
11020 * Otherwise would only return 0 if regbranch() returns 0, which cannot
11022 STATIC regnode_offset
11023 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp, U32 depth)
11024 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
11025 * 2 is like 1, but indicates that nextchar() has been called to advance
11026 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
11027 * this flag alerts us to the need to check for that */
11029 regnode_offset ret = 0; /* Will be the head of the group. */
11031 regnode_offset lastbr;
11032 regnode_offset ender = 0;
11035 U32 oregflags = RExC_flags;
11036 bool have_branch = 0;
11038 I32 freeze_paren = 0;
11039 I32 after_freeze = 0;
11040 I32 num; /* numeric backreferences */
11041 SV * max_open; /* Max number of unclosed parens */
11043 char * parse_start = RExC_parse; /* MJD */
11044 char * const oregcomp_parse = RExC_parse;
11046 GET_RE_DEBUG_FLAGS_DECL;
11048 PERL_ARGS_ASSERT_REG;
11049 DEBUG_PARSE("reg ");
11052 max_open = get_sv(RE_COMPILE_RECURSION_LIMIT, GV_ADD);
11054 if (!SvIOK(max_open)) {
11055 sv_setiv(max_open, RE_COMPILE_RECURSION_INIT);
11057 if (depth > 4 * (UV) SvIV(max_open)) { /* We increase depth by 4 for each
11059 vFAIL("Too many nested open parens");
11062 *flagp = 0; /* Tentatively. */
11064 /* Having this true makes it feasible to have a lot fewer tests for the
11065 * parse pointer being in scope. For example, we can write
11066 * while(isFOO(*RExC_parse)) RExC_parse++;
11068 * while(RExC_parse < RExC_end && isFOO(*RExC_parse)) RExC_parse++;
11070 assert(*RExC_end == '\0');
11072 /* Make an OPEN node, if parenthesized. */
11075 /* Under /x, space and comments can be gobbled up between the '(' and
11076 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
11077 * intervening space, as the sequence is a token, and a token should be
11079 bool has_intervening_patws = (paren == 2)
11080 && *(RExC_parse - 1) != '(';
11082 if (RExC_parse >= RExC_end) {
11083 vFAIL("Unmatched (");
11086 if (paren == 'r') { /* Atomic script run */
11090 else if ( *RExC_parse == '*') { /* (*VERB:ARG), (*construct:...) */
11091 char *start_verb = RExC_parse + 1;
11093 char *start_arg = NULL;
11094 unsigned char op = 0;
11095 int arg_required = 0;
11096 int internal_argval = -1; /* if >-1 we are not allowed an argument*/
11097 bool has_upper = FALSE;
11099 if (has_intervening_patws) {
11100 RExC_parse++; /* past the '*' */
11102 /* For strict backwards compatibility, don't change the message
11103 * now that we also have lowercase operands */
11104 if (isUPPER(*RExC_parse)) {
11105 vFAIL("In '(*VERB...)', the '(' and '*' must be adjacent");
11108 vFAIL("In '(*...)', the '(' and '*' must be adjacent");
11111 while (RExC_parse < RExC_end && *RExC_parse != ')' ) {
11112 if ( *RExC_parse == ':' ) {
11113 start_arg = RExC_parse + 1;
11117 if (isUPPER(*RExC_parse)) {
11123 RExC_parse += UTF8SKIP(RExC_parse);
11126 verb_len = RExC_parse - start_verb;
11128 if (RExC_parse >= RExC_end) {
11129 goto unterminated_verb_pattern;
11132 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11133 while ( RExC_parse < RExC_end && *RExC_parse != ')' ) {
11134 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11136 if ( RExC_parse >= RExC_end || *RExC_parse != ')' ) {
11137 unterminated_verb_pattern:
11139 vFAIL("Unterminated verb pattern argument");
11142 vFAIL("Unterminated '(*...' argument");
11146 if ( RExC_parse >= RExC_end || *RExC_parse != ')' ) {
11148 vFAIL("Unterminated verb pattern");
11151 vFAIL("Unterminated '(*...' construct");
11156 /* Here, we know that RExC_parse < RExC_end */
11158 switch ( *start_verb ) {
11159 case 'A': /* (*ACCEPT) */
11160 if ( memEQs(start_verb, verb_len,"ACCEPT") ) {
11162 internal_argval = RExC_nestroot;
11165 case 'C': /* (*COMMIT) */
11166 if ( memEQs(start_verb, verb_len,"COMMIT") )
11169 case 'F': /* (*FAIL) */
11170 if ( verb_len==1 || memEQs(start_verb, verb_len,"FAIL") ) {
11174 case ':': /* (*:NAME) */
11175 case 'M': /* (*MARK:NAME) */
11176 if ( verb_len==0 || memEQs(start_verb, verb_len,"MARK") ) {
11181 case 'P': /* (*PRUNE) */
11182 if ( memEQs(start_verb, verb_len,"PRUNE") )
11185 case 'S': /* (*SKIP) */
11186 if ( memEQs(start_verb, verb_len,"SKIP") )
11189 case 'T': /* (*THEN) */
11190 /* [19:06] <TimToady> :: is then */
11191 if ( memEQs(start_verb, verb_len,"THEN") ) {
11193 RExC_seen |= REG_CUTGROUP_SEEN;
11197 if ( memEQs(start_verb, verb_len, "asr")
11198 || memEQs(start_verb, verb_len, "atomic_script_run"))
11200 paren = 'r'; /* Mnemonic: recursed run */
11203 else if (memEQs(start_verb, verb_len, "atomic")) {
11204 paren = 't'; /* AtOMIC */
11205 goto alpha_assertions;
11209 if ( memEQs(start_verb, verb_len, "plb")
11210 || memEQs(start_verb, verb_len, "positive_lookbehind"))
11213 goto lookbehind_alpha_assertions;
11215 else if ( memEQs(start_verb, verb_len, "pla")
11216 || memEQs(start_verb, verb_len, "positive_lookahead"))
11219 goto alpha_assertions;
11223 if ( memEQs(start_verb, verb_len, "nlb")
11224 || memEQs(start_verb, verb_len, "negative_lookbehind"))
11227 goto lookbehind_alpha_assertions;
11229 else if ( memEQs(start_verb, verb_len, "nla")
11230 || memEQs(start_verb, verb_len, "negative_lookahead"))
11233 goto alpha_assertions;
11237 if ( memEQs(start_verb, verb_len, "sr")
11238 || memEQs(start_verb, verb_len, "script_run"))
11240 regnode_offset atomic;
11246 /* This indicates Unicode rules. */
11247 REQUIRE_UNI_RULES(flagp, 0);
11253 RExC_parse = start_arg;
11255 if (RExC_in_script_run) {
11257 /* Nested script runs are treated as no-ops, because
11258 * if the nested one fails, the outer one must as
11259 * well. It could fail sooner, and avoid (??{} with
11260 * side effects, but that is explicitly documented as
11261 * undefined behavior. */
11265 if (paren == 's') {
11270 /* But, the atomic part of a nested atomic script run
11271 * isn't a no-op, but can be treated just like a '(?>'
11277 /* By doing this here, we avoid extra warnings for nested
11279 ckWARNexperimental(RExC_parse,
11280 WARN_EXPERIMENTAL__SCRIPT_RUN,
11281 "The script_run feature is experimental");
11283 if (paren == 's') {
11284 /* Here, we're starting a new regular script run */
11285 ret = reg_node(pRExC_state, SROPEN);
11286 RExC_in_script_run = 1;
11291 /* Here, we are starting an atomic script run. This is
11292 * handled by recursing to deal with the atomic portion
11293 * separately, enclosed in SROPEN ... SRCLOSE nodes */
11295 ret = reg_node(pRExC_state, SROPEN);
11297 RExC_in_script_run = 1;
11299 atomic = reg(pRExC_state, 'r', &flags, depth);
11300 if (flags & (RESTART_PARSE|NEED_UTF8)) {
11301 *flagp = flags & (RESTART_PARSE|NEED_UTF8);
11305 REGTAIL(pRExC_state, ret, atomic);
11307 REGTAIL(pRExC_state, atomic,
11308 reg_node(pRExC_state, SRCLOSE));
11310 RExC_in_script_run = 0;
11316 lookbehind_alpha_assertions:
11317 RExC_seen |= REG_LOOKBEHIND_SEEN;
11318 RExC_in_lookbehind++;
11322 ckWARNexperimental(RExC_parse,
11323 WARN_EXPERIMENTAL__ALPHA_ASSERTIONS,
11324 "The alpha_assertions feature is experimental");
11326 RExC_seen_zerolen++;
11332 /* An empty negative lookahead assertion simply is failure */
11333 if (paren == 'A' && RExC_parse == start_arg) {
11334 ret=reganode(pRExC_state, OPFAIL, 0);
11335 nextchar(pRExC_state);
11339 RExC_parse = start_arg;
11344 "'(*%" UTF8f "' requires a terminating ':'",
11345 UTF8fARG(UTF, verb_len, start_verb));
11346 NOT_REACHED; /*NOTREACHED*/
11348 } /* End of switch */
11351 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
11353 if (has_upper || verb_len == 0) {
11355 "Unknown verb pattern '%" UTF8f "'",
11356 UTF8fARG(UTF, verb_len, start_verb));
11360 "Unknown '(*...)' construct '%" UTF8f "'",
11361 UTF8fARG(UTF, verb_len, start_verb));
11364 if ( RExC_parse == start_arg ) {
11367 if ( arg_required && !start_arg ) {
11368 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
11369 verb_len, start_verb);
11371 if (internal_argval == -1) {
11372 ret = reganode(pRExC_state, op, 0);
11374 ret = reg2Lanode(pRExC_state, op, 0, internal_argval);
11376 RExC_seen |= REG_VERBARG_SEEN;
11378 SV *sv = newSVpvn( start_arg,
11379 RExC_parse - start_arg);
11380 ARG(REGNODE_p(ret)) = add_data( pRExC_state,
11381 STR_WITH_LEN("S"));
11382 RExC_rxi->data->data[ARG(REGNODE_p(ret))]=(void*)sv;
11383 FLAGS(REGNODE_p(ret)) = 1;
11385 FLAGS(REGNODE_p(ret)) = 0;
11387 if ( internal_argval != -1 )
11388 ARG2L_SET(REGNODE_p(ret), internal_argval);
11389 nextchar(pRExC_state);
11392 else if (*RExC_parse == '?') { /* (?...) */
11393 bool is_logical = 0;
11394 const char * const seqstart = RExC_parse;
11395 const char * endptr;
11396 if (has_intervening_patws) {
11398 vFAIL("In '(?...)', the '(' and '?' must be adjacent");
11401 RExC_parse++; /* past the '?' */
11402 paren = *RExC_parse; /* might be a trailing NUL, if not
11404 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11405 if (RExC_parse > RExC_end) {
11408 ret = 0; /* For look-ahead/behind. */
11411 case 'P': /* (?P...) variants for those used to PCRE/Python */
11412 paren = *RExC_parse;
11413 if ( paren == '<') { /* (?P<...>) named capture */
11415 if (RExC_parse >= RExC_end) {
11416 vFAIL("Sequence (?P<... not terminated");
11418 goto named_capture;
11420 else if (paren == '>') { /* (?P>name) named recursion */
11422 if (RExC_parse >= RExC_end) {
11423 vFAIL("Sequence (?P>... not terminated");
11425 goto named_recursion;
11427 else if (paren == '=') { /* (?P=...) named backref */
11429 return handle_named_backref(pRExC_state, flagp,
11432 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
11433 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11434 vFAIL3("Sequence (%.*s...) not recognized",
11435 RExC_parse-seqstart, seqstart);
11436 NOT_REACHED; /*NOTREACHED*/
11437 case '<': /* (?<...) */
11438 if (*RExC_parse == '!')
11440 else if (*RExC_parse != '=')
11447 case '\'': /* (?'...') */
11448 name_start = RExC_parse;
11449 svname = reg_scan_name(pRExC_state, REG_RSN_RETURN_NAME);
11450 if ( RExC_parse == name_start
11451 || RExC_parse >= RExC_end
11452 || *RExC_parse != paren)
11454 vFAIL2("Sequence (?%c... not terminated",
11455 paren=='>' ? '<' : paren);
11460 if (!svname) /* shouldn't happen */
11462 "panic: reg_scan_name returned NULL");
11463 if (!RExC_paren_names) {
11464 RExC_paren_names= newHV();
11465 sv_2mortal(MUTABLE_SV(RExC_paren_names));
11467 RExC_paren_name_list= newAV();
11468 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
11471 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
11473 sv_dat = HeVAL(he_str);
11475 /* croak baby croak */
11477 "panic: paren_name hash element allocation failed");
11478 } else if ( SvPOK(sv_dat) ) {
11479 /* (?|...) can mean we have dupes so scan to check
11480 its already been stored. Maybe a flag indicating
11481 we are inside such a construct would be useful,
11482 but the arrays are likely to be quite small, so
11483 for now we punt -- dmq */
11484 IV count = SvIV(sv_dat);
11485 I32 *pv = (I32*)SvPVX(sv_dat);
11487 for ( i = 0 ; i < count ; i++ ) {
11488 if ( pv[i] == RExC_npar ) {
11494 pv = (I32*)SvGROW(sv_dat,
11495 SvCUR(sv_dat) + sizeof(I32)+1);
11496 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
11497 pv[count] = RExC_npar;
11498 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
11501 (void)SvUPGRADE(sv_dat, SVt_PVNV);
11502 sv_setpvn(sv_dat, (char *)&(RExC_npar),
11505 SvIV_set(sv_dat, 1);
11508 /* Yes this does cause a memory leak in debugging Perls
11510 if (!av_store(RExC_paren_name_list,
11511 RExC_npar, SvREFCNT_inc_NN(svname)))
11512 SvREFCNT_dec_NN(svname);
11515 /*sv_dump(sv_dat);*/
11517 nextchar(pRExC_state);
11519 goto capturing_parens;
11522 RExC_seen |= REG_LOOKBEHIND_SEEN;
11523 RExC_in_lookbehind++;
11525 if (RExC_parse >= RExC_end) {
11526 vFAIL("Sequence (?... not terminated");
11530 case '=': /* (?=...) */
11531 RExC_seen_zerolen++;
11533 case '!': /* (?!...) */
11534 RExC_seen_zerolen++;
11535 /* check if we're really just a "FAIL" assertion */
11536 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
11537 FALSE /* Don't force to /x */ );
11538 if (*RExC_parse == ')') {
11539 ret=reganode(pRExC_state, OPFAIL, 0);
11540 nextchar(pRExC_state);
11544 case '|': /* (?|...) */
11545 /* branch reset, behave like a (?:...) except that
11546 buffers in alternations share the same numbers */
11548 after_freeze = freeze_paren = RExC_npar;
11550 /* XXX This construct currently requires an extra pass.
11551 * Investigation would be required to see if that could be
11553 REQUIRE_PARENS_PASS;
11555 case ':': /* (?:...) */
11556 case '>': /* (?>...) */
11558 case '$': /* (?$...) */
11559 case '@': /* (?@...) */
11560 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
11562 case '0' : /* (?0) */
11563 case 'R' : /* (?R) */
11564 if (RExC_parse == RExC_end || *RExC_parse != ')')
11565 FAIL("Sequence (?R) not terminated");
11567 RExC_seen |= REG_RECURSE_SEEN;
11569 /* XXX These constructs currently require an extra pass.
11570 * It probably could be changed */
11571 REQUIRE_PARENS_PASS;
11573 *flagp |= POSTPONED;
11574 goto gen_recurse_regop;
11576 /* named and numeric backreferences */
11577 case '&': /* (?&NAME) */
11578 parse_start = RExC_parse - 1;
11581 SV *sv_dat = reg_scan_name(pRExC_state,
11582 REG_RSN_RETURN_DATA);
11583 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
11585 if (RExC_parse >= RExC_end || *RExC_parse != ')')
11586 vFAIL("Sequence (?&... not terminated");
11587 goto gen_recurse_regop;
11590 if (! inRANGE(RExC_parse[0], '1', '9')) {
11592 vFAIL("Illegal pattern");
11594 goto parse_recursion;
11596 case '-': /* (?-1) */
11597 if (! inRANGE(RExC_parse[0], '1', '9')) {
11598 RExC_parse--; /* rewind to let it be handled later */
11602 case '1': case '2': case '3': case '4': /* (?1) */
11603 case '5': case '6': case '7': case '8': case '9':
11604 RExC_parse = (char *) seqstart + 1; /* Point to the digit */
11607 bool is_neg = FALSE;
11609 parse_start = RExC_parse - 1; /* MJD */
11610 if (*RExC_parse == '-') {
11615 if (grok_atoUV(RExC_parse, &unum, &endptr)
11619 RExC_parse = (char*)endptr;
11623 /* Some limit for num? */
11627 if (*RExC_parse!=')')
11628 vFAIL("Expecting close bracket");
11631 if ( paren == '-' ) {
11633 Diagram of capture buffer numbering.
11634 Top line is the normal capture buffer numbers
11635 Bottom line is the negative indexing as from
11639 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
11643 num = RExC_npar + num;
11646 /* It might be a forward reference; we can't fail until
11647 * we know, by completing the parse to get all the
11648 * groups, and then reparsing */
11649 if (ALL_PARENS_COUNTED) {
11651 vFAIL("Reference to nonexistent group");
11654 REQUIRE_PARENS_PASS;
11657 } else if ( paren == '+' ) {
11658 num = RExC_npar + num - 1;
11660 /* We keep track how many GOSUB items we have produced.
11661 To start off the ARG2L() of the GOSUB holds its "id",
11662 which is used later in conjunction with RExC_recurse
11663 to calculate the offset we need to jump for the GOSUB,
11664 which it will store in the final representation.
11665 We have to defer the actual calculation until much later
11666 as the regop may move.
11669 ret = reg2Lanode(pRExC_state, GOSUB, num, RExC_recurse_count);
11670 if (num >= RExC_npar) {
11672 /* It might be a forward reference; we can't fail until we
11673 * know, by completing the parse to get all the groups, and
11674 * then reparsing */
11675 if (ALL_PARENS_COUNTED) {
11676 if (num >= RExC_total_parens) {
11678 vFAIL("Reference to nonexistent group");
11682 REQUIRE_PARENS_PASS;
11685 RExC_recurse_count++;
11686 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11687 "%*s%*s Recurse #%" UVuf " to %" IVdf "\n",
11688 22, "| |", (int)(depth * 2 + 1), "",
11689 (UV)ARG(REGNODE_p(ret)),
11690 (IV)ARG2L(REGNODE_p(ret))));
11691 RExC_seen |= REG_RECURSE_SEEN;
11693 Set_Node_Length(REGNODE_p(ret),
11694 1 + regarglen[OP(REGNODE_p(ret))]); /* MJD */
11695 Set_Node_Offset(REGNODE_p(ret), parse_start); /* MJD */
11697 *flagp |= POSTPONED;
11698 assert(*RExC_parse == ')');
11699 nextchar(pRExC_state);
11704 case '?': /* (??...) */
11706 if (*RExC_parse != '{') {
11707 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
11708 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11710 "Sequence (%" UTF8f "...) not recognized",
11711 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
11712 NOT_REACHED; /*NOTREACHED*/
11714 *flagp |= POSTPONED;
11718 case '{': /* (?{...}) */
11721 struct reg_code_block *cb;
11724 RExC_seen_zerolen++;
11726 if ( !pRExC_state->code_blocks
11727 || pRExC_state->code_index
11728 >= pRExC_state->code_blocks->count
11729 || pRExC_state->code_blocks->cb[pRExC_state->code_index].start
11730 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
11733 if (RExC_pm_flags & PMf_USE_RE_EVAL)
11734 FAIL("panic: Sequence (?{...}): no code block found\n");
11735 FAIL("Eval-group not allowed at runtime, use re 'eval'");
11737 /* this is a pre-compiled code block (?{...}) */
11738 cb = &pRExC_state->code_blocks->cb[pRExC_state->code_index];
11739 RExC_parse = RExC_start + cb->end;
11741 if (cb->src_regex) {
11742 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
11743 RExC_rxi->data->data[n] =
11744 (void*)SvREFCNT_inc((SV*)cb->src_regex);
11745 RExC_rxi->data->data[n+1] = (void*)o;
11748 n = add_data(pRExC_state,
11749 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
11750 RExC_rxi->data->data[n] = (void*)o;
11752 pRExC_state->code_index++;
11753 nextchar(pRExC_state);
11756 regnode_offset eval;
11757 ret = reg_node(pRExC_state, LOGICAL);
11759 eval = reg2Lanode(pRExC_state, EVAL,
11762 /* for later propagation into (??{})
11764 RExC_flags & RXf_PMf_COMPILETIME
11766 FLAGS(REGNODE_p(ret)) = 2;
11767 REGTAIL(pRExC_state, ret, eval);
11768 /* deal with the length of this later - MJD */
11771 ret = reg2Lanode(pRExC_state, EVAL, n, 0);
11772 Set_Node_Length(REGNODE_p(ret), RExC_parse - parse_start + 1);
11773 Set_Node_Offset(REGNODE_p(ret), parse_start);
11776 case '(': /* (?(?{...})...) and (?(?=...)...) */
11779 const int DEFINE_len = sizeof("DEFINE") - 1;
11780 if ( RExC_parse < RExC_end - 1
11781 && ( ( RExC_parse[0] == '?' /* (?(?...)) */
11782 && ( RExC_parse[1] == '='
11783 || RExC_parse[1] == '!'
11784 || RExC_parse[1] == '<'
11785 || RExC_parse[1] == '{'))
11786 || ( RExC_parse[0] == '*' /* (?(*...)) */
11787 && ( memBEGINs(RExC_parse + 1,
11788 (Size_t) (RExC_end - (RExC_parse + 1)),
11790 || memBEGINs(RExC_parse + 1,
11791 (Size_t) (RExC_end - (RExC_parse + 1)),
11793 || memBEGINs(RExC_parse + 1,
11794 (Size_t) (RExC_end - (RExC_parse + 1)),
11796 || memBEGINs(RExC_parse + 1,
11797 (Size_t) (RExC_end - (RExC_parse + 1)),
11799 || memBEGINs(RExC_parse + 1,
11800 (Size_t) (RExC_end - (RExC_parse + 1)),
11801 "positive_lookahead:")
11802 || memBEGINs(RExC_parse + 1,
11803 (Size_t) (RExC_end - (RExC_parse + 1)),
11804 "positive_lookbehind:")
11805 || memBEGINs(RExC_parse + 1,
11806 (Size_t) (RExC_end - (RExC_parse + 1)),
11807 "negative_lookahead:")
11808 || memBEGINs(RExC_parse + 1,
11809 (Size_t) (RExC_end - (RExC_parse + 1)),
11810 "negative_lookbehind:"))))
11811 ) { /* Lookahead or eval. */
11813 regnode_offset tail;
11815 ret = reg_node(pRExC_state, LOGICAL);
11816 FLAGS(REGNODE_p(ret)) = 1;
11818 tail = reg(pRExC_state, 1, &flag, depth+1);
11819 RETURN_FAIL_ON_RESTART(flag, flagp);
11820 REGTAIL(pRExC_state, ret, tail);
11823 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
11824 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
11826 char ch = RExC_parse[0] == '<' ? '>' : '\'';
11827 char *name_start= RExC_parse++;
11829 SV *sv_dat=reg_scan_name(pRExC_state, REG_RSN_RETURN_DATA);
11830 if ( RExC_parse == name_start
11831 || RExC_parse >= RExC_end
11832 || *RExC_parse != ch)
11834 vFAIL2("Sequence (?(%c... not terminated",
11835 (ch == '>' ? '<' : ch));
11839 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11840 RExC_rxi->data->data[num]=(void*)sv_dat;
11841 SvREFCNT_inc_simple_void_NN(sv_dat);
11843 ret = reganode(pRExC_state, NGROUPP, num);
11844 goto insert_if_check_paren;
11846 else if (memBEGINs(RExC_parse,
11847 (STRLEN) (RExC_end - RExC_parse),
11850 ret = reganode(pRExC_state, DEFINEP, 0);
11851 RExC_parse += DEFINE_len;
11853 goto insert_if_check_paren;
11855 else if (RExC_parse[0] == 'R') {
11857 /* parno == 0 => /(?(R)YES|NO)/ "in any form of recursion OR eval"
11858 * parno == 1 => /(?(R0)YES|NO)/ "in GOSUB (?0) / (?R)"
11859 * parno == 2 => /(?(R1)YES|NO)/ "in GOSUB (?1) (parno-1)"
11862 if (RExC_parse[0] == '0') {
11866 else if (inRANGE(RExC_parse[0], '1', '9')) {
11869 if (grok_atoUV(RExC_parse, &uv, &endptr)
11872 parno = (I32)uv + 1;
11873 RExC_parse = (char*)endptr;
11875 /* else "Switch condition not recognized" below */
11876 } else if (RExC_parse[0] == '&') {
11879 sv_dat = reg_scan_name(pRExC_state,
11880 REG_RSN_RETURN_DATA);
11882 parno = 1 + *((I32 *)SvPVX(sv_dat));
11884 ret = reganode(pRExC_state, INSUBP, parno);
11885 goto insert_if_check_paren;
11887 else if (inRANGE(RExC_parse[0], '1', '9')) {
11892 if (grok_atoUV(RExC_parse, &uv, &endptr)
11896 RExC_parse = (char*)endptr;
11899 vFAIL("panic: grok_atoUV returned FALSE");
11901 ret = reganode(pRExC_state, GROUPP, parno);
11903 insert_if_check_paren:
11904 if (UCHARAT(RExC_parse) != ')') {
11906 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
11908 vFAIL("Switch condition not recognized");
11910 nextchar(pRExC_state);
11912 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
11913 br = regbranch(pRExC_state, &flags, 1, depth+1);
11915 RETURN_FAIL_ON_RESTART(flags,flagp);
11916 FAIL2("panic: regbranch returned failure, flags=%#" UVxf,
11919 REGTAIL(pRExC_state, br, reganode(pRExC_state,
11921 c = UCHARAT(RExC_parse);
11922 nextchar(pRExC_state);
11923 if (flags&HASWIDTH)
11924 *flagp |= HASWIDTH;
11927 vFAIL("(?(DEFINE)....) does not allow branches");
11929 /* Fake one for optimizer. */
11930 lastbr = reganode(pRExC_state, IFTHEN, 0);
11932 if (!regbranch(pRExC_state, &flags, 1, depth+1)) {
11933 RETURN_FAIL_ON_RESTART(flags, flagp);
11934 FAIL2("panic: regbranch returned failure, flags=%#" UVxf,
11937 REGTAIL(pRExC_state, ret, lastbr);
11938 if (flags&HASWIDTH)
11939 *flagp |= HASWIDTH;
11940 c = UCHARAT(RExC_parse);
11941 nextchar(pRExC_state);
11946 if (RExC_parse >= RExC_end)
11947 vFAIL("Switch (?(condition)... not terminated");
11949 vFAIL("Switch (?(condition)... contains too many branches");
11951 ender = reg_node(pRExC_state, TAIL);
11952 REGTAIL(pRExC_state, br, ender);
11954 REGTAIL(pRExC_state, lastbr, ender);
11955 REGTAIL(pRExC_state, REGNODE_OFFSET(
11957 NEXTOPER(REGNODE_p(lastbr)))),
11961 REGTAIL(pRExC_state, ret, ender);
11962 #if 0 /* Removing this doesn't cause failures in the test suite -- khw */
11963 RExC_size++; /* XXX WHY do we need this?!!
11964 For large programs it seems to be required
11965 but I can't figure out why. -- dmq*/
11970 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
11972 vFAIL("Unknown switch condition (?(...))");
11974 case '[': /* (?[ ... ]) */
11975 return handle_regex_sets(pRExC_state, NULL, flagp, depth+1,
11977 case 0: /* A NUL */
11978 RExC_parse--; /* for vFAIL to print correctly */
11979 vFAIL("Sequence (? incomplete");
11983 if (RExC_strict) { /* [perl #132851] */
11984 ckWARNreg(RExC_parse, "Empty (?) without any modifiers");
11987 default: /* e.g., (?i) */
11988 RExC_parse = (char *) seqstart + 1;
11990 parse_lparen_question_flags(pRExC_state);
11991 if (UCHARAT(RExC_parse) != ':') {
11992 if (RExC_parse < RExC_end)
11993 nextchar(pRExC_state);
11998 nextchar(pRExC_state);
12004 if (*RExC_parse == '{') {
12005 ckWARNregdep(RExC_parse + 1,
12006 "Unescaped left brace in regex is "
12007 "deprecated here (and will be fatal "
12008 "in Perl 5.32), passed through");
12010 /* Not bothering to indent here, as the above 'else' is temporary
12012 if (!(RExC_flags & RXf_PMf_NOCAPTURE)) { /* (...) */
12016 if (! ALL_PARENS_COUNTED) {
12017 /* If we are in our first pass through (and maybe only pass),
12018 * we need to allocate memory for the capturing parentheses
12022 if (!RExC_parens_buf_size) {
12023 /* first guess at number of parens we might encounter */
12024 RExC_parens_buf_size = 10;
12026 /* setup RExC_open_parens, which holds the address of each
12027 * OPEN tag, and to make things simpler for the 0 index the
12028 * start of the program - this is used later for offsets */
12029 Newxz(RExC_open_parens, RExC_parens_buf_size,
12031 RExC_open_parens[0] = 1; /* +1 for REG_MAGIC */
12033 /* setup RExC_close_parens, which holds the address of each
12034 * CLOSE tag, and to make things simpler for the 0 index
12035 * the end of the program - this is used later for offsets
12037 Newxz(RExC_close_parens, RExC_parens_buf_size,
12039 /* we dont know where end op starts yet, so we dont need to
12040 * set RExC_close_parens[0] like we do RExC_open_parens[0]
12043 else if (RExC_npar > RExC_parens_buf_size) {
12044 I32 old_size = RExC_parens_buf_size;
12046 RExC_parens_buf_size *= 2;
12048 Renew(RExC_open_parens, RExC_parens_buf_size,
12050 Zero(RExC_open_parens + old_size,
12051 RExC_parens_buf_size - old_size, regnode_offset);
12053 Renew(RExC_close_parens, RExC_parens_buf_size,
12055 Zero(RExC_close_parens + old_size,
12056 RExC_parens_buf_size - old_size, regnode_offset);
12060 ret = reganode(pRExC_state, OPEN, parno);
12061 if (!RExC_nestroot)
12062 RExC_nestroot = parno;
12063 if (RExC_open_parens && !RExC_open_parens[parno])
12065 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12066 "%*s%*s Setting open paren #%" IVdf " to %d\n",
12067 22, "| |", (int)(depth * 2 + 1), "",
12069 RExC_open_parens[parno]= ret;
12072 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
12073 Set_Node_Offset(REGNODE_p(ret), RExC_parse); /* MJD */
12076 /* with RXf_PMf_NOCAPTURE treat (...) as (?:...) */
12086 /* Pick up the branches, linking them together. */
12087 parse_start = RExC_parse; /* MJD */
12088 br = regbranch(pRExC_state, &flags, 1, depth+1);
12090 /* branch_len = (paren != 0); */
12093 RETURN_FAIL_ON_RESTART(flags, flagp);
12094 FAIL2("panic: regbranch returned failure, flags=%#" UVxf, (UV) flags);
12096 if (*RExC_parse == '|') {
12097 if (RExC_use_BRANCHJ) {
12098 reginsert(pRExC_state, BRANCHJ, br, depth+1);
12101 reginsert(pRExC_state, BRANCH, br, depth+1);
12102 Set_Node_Length(REGNODE_p(br), paren != 0);
12103 Set_Node_Offset_To_R(br, parse_start-RExC_start);
12107 else if (paren == ':') {
12108 *flagp |= flags&SIMPLE;
12110 if (is_open) { /* Starts with OPEN. */
12111 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
12113 else if (paren != '?') /* Not Conditional */
12115 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
12117 while (*RExC_parse == '|') {
12118 if (RExC_use_BRANCHJ) {
12119 ender = reganode(pRExC_state, LONGJMP, 0);
12121 /* Append to the previous. */
12122 REGTAIL(pRExC_state,
12123 REGNODE_OFFSET(NEXTOPER(NEXTOPER(REGNODE_p(lastbr)))),
12126 nextchar(pRExC_state);
12127 if (freeze_paren) {
12128 if (RExC_npar > after_freeze)
12129 after_freeze = RExC_npar;
12130 RExC_npar = freeze_paren;
12132 br = regbranch(pRExC_state, &flags, 0, depth+1);
12135 RETURN_FAIL_ON_RESTART(flags, flagp);
12136 FAIL2("panic: regbranch returned failure, flags=%#" UVxf, (UV) flags);
12138 if (! REGTAIL(pRExC_state, lastbr, br)) { /* BRANCH -> BRANCH. */
12139 REQUIRE_BRANCHJ(flagp, 0);
12142 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
12145 if (have_branch || paren != ':') {
12148 /* Make a closing node, and hook it on the end. */
12151 ender = reg_node(pRExC_state, TAIL);
12154 ender = reganode(pRExC_state, CLOSE, parno);
12155 if ( RExC_close_parens ) {
12156 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12157 "%*s%*s Setting close paren #%" IVdf " to %d\n",
12158 22, "| |", (int)(depth * 2 + 1), "",
12159 (IV)parno, ender));
12160 RExC_close_parens[parno]= ender;
12161 if (RExC_nestroot == parno)
12164 Set_Node_Offset(REGNODE_p(ender), RExC_parse+1); /* MJD */
12165 Set_Node_Length(REGNODE_p(ender), 1); /* MJD */
12168 ender = reg_node(pRExC_state, SRCLOSE);
12169 RExC_in_script_run = 0;
12179 *flagp &= ~HASWIDTH;
12181 case 't': /* aTomic */
12183 ender = reg_node(pRExC_state, SUCCEED);
12186 ender = reg_node(pRExC_state, END);
12187 assert(!RExC_end_op); /* there can only be one! */
12188 RExC_end_op = REGNODE_p(ender);
12189 if (RExC_close_parens) {
12190 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12191 "%*s%*s Setting close paren #0 (END) to %d\n",
12192 22, "| |", (int)(depth * 2 + 1), "",
12195 RExC_close_parens[0]= ender;
12200 DEBUG_PARSE_MSG("lsbr");
12201 regprop(RExC_rx, RExC_mysv1, REGNODE_p(lastbr), NULL, pRExC_state);
12202 regprop(RExC_rx, RExC_mysv2, REGNODE_p(ender), NULL, pRExC_state);
12203 Perl_re_printf( aTHX_ "~ tying lastbr %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
12204 SvPV_nolen_const(RExC_mysv1),
12206 SvPV_nolen_const(RExC_mysv2),
12208 (IV)(ender - lastbr)
12211 if (! REGTAIL(pRExC_state, lastbr, ender)) {
12212 REQUIRE_BRANCHJ(flagp, 0);
12216 char is_nothing= 1;
12218 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
12220 /* Hook the tails of the branches to the closing node. */
12221 for (br = REGNODE_p(ret); br; br = regnext(br)) {
12222 const U8 op = PL_regkind[OP(br)];
12223 if (op == BRANCH) {
12224 if (! REGTAIL_STUDY(pRExC_state,
12225 REGNODE_OFFSET(NEXTOPER(br)),
12228 REQUIRE_BRANCHJ(flagp, 0);
12230 if ( OP(NEXTOPER(br)) != NOTHING
12231 || regnext(NEXTOPER(br)) != REGNODE_p(ender))
12234 else if (op == BRANCHJ) {
12235 REGTAIL_STUDY(pRExC_state,
12236 REGNODE_OFFSET(NEXTOPER(NEXTOPER(br))),
12238 /* for now we always disable this optimisation * /
12239 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
12240 || regnext(NEXTOPER(NEXTOPER(br))) != REGNODE_p(ender))
12246 regnode * ret_as_regnode = REGNODE_p(ret);
12247 br= PL_regkind[OP(ret_as_regnode)] != BRANCH
12248 ? regnext(ret_as_regnode)
12251 DEBUG_PARSE_MSG("NADA");
12252 regprop(RExC_rx, RExC_mysv1, ret_as_regnode,
12253 NULL, pRExC_state);
12254 regprop(RExC_rx, RExC_mysv2, REGNODE_p(ender),
12255 NULL, pRExC_state);
12256 Perl_re_printf( aTHX_ "~ converting ret %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
12257 SvPV_nolen_const(RExC_mysv1),
12258 (IV)REG_NODE_NUM(ret_as_regnode),
12259 SvPV_nolen_const(RExC_mysv2),
12265 if (OP(REGNODE_p(ender)) == TAIL) {
12267 RExC_emit= REGNODE_OFFSET(br) + 1;
12270 for ( opt= br + 1; opt < REGNODE_p(ender) ; opt++ )
12271 OP(opt)= OPTIMIZED;
12272 NEXT_OFF(br)= REGNODE_p(ender) - br;
12280 /* Even/odd or x=don't care: 010101x10x */
12281 static const char parens[] = "=!aA<,>Bbt";
12282 /* flag below is set to 0 up through 'A'; 1 for larger */
12284 if (paren && (p = strchr(parens, paren))) {
12285 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
12286 int flag = (p - parens) > 3;
12288 if (paren == '>' || paren == 't') {
12289 node = SUSPEND, flag = 0;
12292 reginsert(pRExC_state, node, ret, depth+1);
12293 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
12294 Set_Node_Offset(REGNODE_p(ret), parse_start + 1);
12295 FLAGS(REGNODE_p(ret)) = flag;
12296 if (! REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL)))
12298 REQUIRE_BRANCHJ(flagp, 0);
12303 /* Check for proper termination. */
12305 /* restore original flags, but keep (?p) and, if we've encountered
12306 * something in the parse that changes /d rules into /u, keep the /u */
12307 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
12308 if (DEPENDS_SEMANTICS && RExC_uni_semantics) {
12309 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
12311 if (RExC_parse >= RExC_end || UCHARAT(RExC_parse) != ')') {
12312 RExC_parse = oregcomp_parse;
12313 vFAIL("Unmatched (");
12315 nextchar(pRExC_state);
12317 else if (!paren && RExC_parse < RExC_end) {
12318 if (*RExC_parse == ')') {
12320 vFAIL("Unmatched )");
12323 FAIL("Junk on end of regexp"); /* "Can't happen". */
12324 NOT_REACHED; /* NOTREACHED */
12327 if (RExC_in_lookbehind) {
12328 RExC_in_lookbehind--;
12330 if (after_freeze > RExC_npar)
12331 RExC_npar = after_freeze;
12336 - regbranch - one alternative of an | operator
12338 * Implements the concatenation operator.
12340 * On success, returns the offset at which any next node should be placed into
12341 * the regex engine program being compiled.
12343 * Returns 0 otherwise, setting flagp to RESTART_PARSE if the parse needs
12344 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to
12347 STATIC regnode_offset
12348 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
12350 regnode_offset ret;
12351 regnode_offset chain = 0;
12352 regnode_offset latest;
12353 I32 flags = 0, c = 0;
12354 GET_RE_DEBUG_FLAGS_DECL;
12356 PERL_ARGS_ASSERT_REGBRANCH;
12358 DEBUG_PARSE("brnc");
12363 if (RExC_use_BRANCHJ)
12364 ret = reganode(pRExC_state, BRANCHJ, 0);
12366 ret = reg_node(pRExC_state, BRANCH);
12367 Set_Node_Length(REGNODE_p(ret), 1);
12371 *flagp = WORST; /* Tentatively. */
12373 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
12374 FALSE /* Don't force to /x */ );
12375 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
12376 flags &= ~TRYAGAIN;
12377 latest = regpiece(pRExC_state, &flags, depth+1);
12379 if (flags & TRYAGAIN)
12381 RETURN_FAIL_ON_RESTART(flags, flagp);
12382 FAIL2("panic: regpiece returned failure, flags=%#" UVxf, (UV) flags);
12386 *flagp |= flags&(HASWIDTH|POSTPONED);
12387 if (chain == 0) /* First piece. */
12388 *flagp |= flags&SPSTART;
12390 /* FIXME adding one for every branch after the first is probably
12391 * excessive now we have TRIE support. (hv) */
12393 if (! REGTAIL(pRExC_state, chain, latest)) {
12394 /* XXX We could just redo this branch, but figuring out what
12395 * bookkeeping needs to be reset is a pain, and it's likely
12396 * that other branches that goto END will also be too large */
12397 REQUIRE_BRANCHJ(flagp, 0);
12403 if (chain == 0) { /* Loop ran zero times. */
12404 chain = reg_node(pRExC_state, NOTHING);
12409 *flagp |= flags&SIMPLE;
12416 - regpiece - something followed by possible quantifier * + ? {n,m}
12418 * Note that the branching code sequences used for ? and the general cases
12419 * of * and + are somewhat optimized: they use the same NOTHING node as
12420 * both the endmarker for their branch list and the body of the last branch.
12421 * It might seem that this node could be dispensed with entirely, but the
12422 * endmarker role is not redundant.
12424 * On success, returns the offset at which any next node should be placed into
12425 * the regex engine program being compiled.
12427 * Returns 0 otherwise, with *flagp set to indicate why:
12428 * TRYAGAIN if regatom() returns 0 with TRYAGAIN.
12429 * RESTART_PARSE if the parse needs to be restarted, or'd with
12430 * NEED_UTF8 if the pattern needs to be upgraded to UTF-8.
12432 STATIC regnode_offset
12433 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
12435 regnode_offset ret;
12439 const char * const origparse = RExC_parse;
12441 I32 max = REG_INFTY;
12442 #ifdef RE_TRACK_PATTERN_OFFSETS
12445 const char *maxpos = NULL;
12448 /* Save the original in case we change the emitted regop to a FAIL. */
12449 const regnode_offset orig_emit = RExC_emit;
12451 GET_RE_DEBUG_FLAGS_DECL;
12453 PERL_ARGS_ASSERT_REGPIECE;
12455 DEBUG_PARSE("piec");
12457 ret = regatom(pRExC_state, &flags, depth+1);
12459 RETURN_FAIL_ON_RESTART_OR_FLAGS(flags, flagp, TRYAGAIN);
12460 FAIL2("panic: regatom returned failure, flags=%#" UVxf, (UV) flags);
12465 if (op == '{' && regcurly(RExC_parse)) {
12467 #ifdef RE_TRACK_PATTERN_OFFSETS
12468 parse_start = RExC_parse; /* MJD */
12470 next = RExC_parse + 1;
12471 while (isDIGIT(*next) || *next == ',') {
12472 if (*next == ',') {
12480 if (*next == '}') { /* got one */
12481 const char* endptr;
12485 if (isDIGIT(*RExC_parse)) {
12487 if (!grok_atoUV(RExC_parse, &uv, &endptr))
12488 vFAIL("Invalid quantifier in {,}");
12489 if (uv >= REG_INFTY)
12490 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
12495 if (*maxpos == ',')
12498 maxpos = RExC_parse;
12499 if (isDIGIT(*maxpos)) {
12501 if (!grok_atoUV(maxpos, &uv, &endptr))
12502 vFAIL("Invalid quantifier in {,}");
12503 if (uv >= REG_INFTY)
12504 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
12507 max = REG_INFTY; /* meaning "infinity" */
12510 nextchar(pRExC_state);
12511 if (max < min) { /* If can't match, warn and optimize to fail
12513 reginsert(pRExC_state, OPFAIL, orig_emit, depth+1);
12514 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
12515 NEXT_OFF(REGNODE_p(orig_emit)) =
12516 regarglen[OPFAIL] + NODE_STEP_REGNODE;
12519 else if (min == max && *RExC_parse == '?')
12521 ckWARN2reg(RExC_parse + 1,
12522 "Useless use of greediness modifier '%c'",
12527 if ((flags&SIMPLE)) {
12528 if (min == 0 && max == REG_INFTY) {
12529 reginsert(pRExC_state, STAR, ret, depth+1);
12531 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
12534 if (min == 1 && max == REG_INFTY) {
12535 reginsert(pRExC_state, PLUS, ret, depth+1);
12537 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
12540 MARK_NAUGHTY_EXP(2, 2);
12541 reginsert(pRExC_state, CURLY, ret, depth+1);
12542 Set_Node_Offset(REGNODE_p(ret), parse_start+1); /* MJD */
12543 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
12546 const regnode_offset w = reg_node(pRExC_state, WHILEM);
12548 FLAGS(REGNODE_p(w)) = 0;
12549 REGTAIL(pRExC_state, ret, w);
12550 if (RExC_use_BRANCHJ) {
12551 reginsert(pRExC_state, LONGJMP, ret, depth+1);
12552 reginsert(pRExC_state, NOTHING, ret, depth+1);
12553 NEXT_OFF(REGNODE_p(ret)) = 3; /* Go over LONGJMP. */
12555 reginsert(pRExC_state, CURLYX, ret, depth+1);
12557 Set_Node_Offset(REGNODE_p(ret), parse_start+1);
12558 Set_Node_Length(REGNODE_p(ret),
12559 op == '{' ? (RExC_parse - parse_start) : 1);
12561 if (RExC_use_BRANCHJ)
12562 NEXT_OFF(REGNODE_p(ret)) = 3; /* Go over NOTHING to
12564 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
12565 RExC_whilem_seen++;
12566 MARK_NAUGHTY_EXP(1, 4); /* compound interest */
12568 FLAGS(REGNODE_p(ret)) = 0;
12573 *flagp |= HASWIDTH;
12574 ARG1_SET(REGNODE_p(ret), (U16)min);
12575 ARG2_SET(REGNODE_p(ret), (U16)max);
12576 if (max == REG_INFTY)
12577 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
12583 if (!ISMULT1(op)) {
12588 #if 0 /* Now runtime fix should be reliable. */
12590 /* if this is reinstated, don't forget to put this back into perldiag:
12592 =item Regexp *+ operand could be empty at {#} in regex m/%s/
12594 (F) The part of the regexp subject to either the * or + quantifier
12595 could match an empty string. The {#} shows in the regular
12596 expression about where the problem was discovered.
12600 if (!(flags&HASWIDTH) && op != '?')
12601 vFAIL("Regexp *+ operand could be empty");
12604 #ifdef RE_TRACK_PATTERN_OFFSETS
12605 parse_start = RExC_parse;
12607 nextchar(pRExC_state);
12609 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
12615 else if (op == '+') {
12619 else if (op == '?') {
12624 if (!(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
12625 ckWARN2reg(RExC_parse,
12626 "%" UTF8f " matches null string many times",
12627 UTF8fARG(UTF, (RExC_parse >= origparse
12628 ? RExC_parse - origparse
12633 if (*RExC_parse == '?') {
12634 nextchar(pRExC_state);
12635 reginsert(pRExC_state, MINMOD, ret, depth+1);
12636 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
12638 else if (*RExC_parse == '+') {
12639 regnode_offset ender;
12640 nextchar(pRExC_state);
12641 ender = reg_node(pRExC_state, SUCCEED);
12642 REGTAIL(pRExC_state, ret, ender);
12643 reginsert(pRExC_state, SUSPEND, ret, depth+1);
12644 ender = reg_node(pRExC_state, TAIL);
12645 REGTAIL(pRExC_state, ret, ender);
12648 if (ISMULT2(RExC_parse)) {
12650 vFAIL("Nested quantifiers");
12657 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state,
12658 regnode_offset * node_p,
12666 /* This routine teases apart the various meanings of \N and returns
12667 * accordingly. The input parameters constrain which meaning(s) is/are valid
12668 * in the current context.
12670 * Exactly one of <node_p> and <code_point_p> must be non-NULL.
12672 * If <code_point_p> is not NULL, the context is expecting the result to be a
12673 * single code point. If this \N instance turns out to a single code point,
12674 * the function returns TRUE and sets *code_point_p to that code point.
12676 * If <node_p> is not NULL, the context is expecting the result to be one of
12677 * the things representable by a regnode. If this \N instance turns out to be
12678 * one such, the function generates the regnode, returns TRUE and sets *node_p
12679 * to point to the offset of that regnode into the regex engine program being
12682 * If this instance of \N isn't legal in any context, this function will
12683 * generate a fatal error and not return.
12685 * On input, RExC_parse should point to the first char following the \N at the
12686 * time of the call. On successful return, RExC_parse will have been updated
12687 * to point to just after the sequence identified by this routine. Also
12688 * *flagp has been updated as needed.
12690 * When there is some problem with the current context and this \N instance,
12691 * the function returns FALSE, without advancing RExC_parse, nor setting
12692 * *node_p, nor *code_point_p, nor *flagp.
12694 * If <cp_count> is not NULL, the caller wants to know the length (in code
12695 * points) that this \N sequence matches. This is set, and the input is
12696 * parsed for errors, even if the function returns FALSE, as detailed below.
12698 * There are 6 possibilities here, as detailed in the next 6 paragraphs.
12700 * Probably the most common case is for the \N to specify a single code point.
12701 * *cp_count will be set to 1, and *code_point_p will be set to that code
12704 * Another possibility is for the input to be an empty \N{}. This is no
12705 * longer accepted, and will generate a fatal error.
12707 * Another possibility is for a custom charnames handler to be in effect which
12708 * translates the input name to an empty string. *cp_count will be set to 0.
12709 * *node_p will be set to a generated NOTHING node.
12711 * Still another possibility is for the \N to mean [^\n]. *cp_count will be
12712 * set to 0. *node_p will be set to a generated REG_ANY node.
12714 * The fifth possibility is that \N resolves to a sequence of more than one
12715 * code points. *cp_count will be set to the number of code points in the
12716 * sequence. *node_p will be set to a generated node returned by this
12717 * function calling S_reg().
12719 * The final possibility is that it is premature to be calling this function;
12720 * the parse needs to be restarted. This can happen when this changes from
12721 * /d to /u rules, or when the pattern needs to be upgraded to UTF-8. The
12722 * latter occurs only when the fifth possibility would otherwise be in
12723 * effect, and is because one of those code points requires the pattern to be
12724 * recompiled as UTF-8. The function returns FALSE, and sets the
12725 * RESTART_PARSE and NEED_UTF8 flags in *flagp, as appropriate. When this
12726 * happens, the caller needs to desist from continuing parsing, and return
12727 * this information to its caller. This is not set for when there is only one
12728 * code point, as this can be called as part of an ANYOF node, and they can
12729 * store above-Latin1 code points without the pattern having to be in UTF-8.
12731 * For non-single-quoted regexes, the tokenizer has resolved character and
12732 * sequence names inside \N{...} into their Unicode values, normalizing the
12733 * result into what we should see here: '\N{U+c1.c2...}', where c1... are the
12734 * hex-represented code points in the sequence. This is done there because
12735 * the names can vary based on what charnames pragma is in scope at the time,
12736 * so we need a way to take a snapshot of what they resolve to at the time of
12737 * the original parse. [perl #56444].
12739 * That parsing is skipped for single-quoted regexes, so here we may get
12740 * '\N{NAME}', which is parsed now. If the single-quoted regex is something
12741 * like '\N{U+41}', that code point is Unicode, and has to be translated into
12742 * the native character set for non-ASCII platforms. The other possibilities
12743 * are already native, so no translation is done. */
12745 char * endbrace; /* points to '}' following the name */
12746 char* p = RExC_parse; /* Temporary */
12748 SV * substitute_parse = NULL;
12753 GET_RE_DEBUG_FLAGS_DECL;
12755 PERL_ARGS_ASSERT_GROK_BSLASH_N;
12757 GET_RE_DEBUG_FLAGS;
12759 assert(cBOOL(node_p) ^ cBOOL(code_point_p)); /* Exactly one should be set */
12760 assert(! (node_p && cp_count)); /* At most 1 should be set */
12762 if (cp_count) { /* Initialize return for the most common case */
12766 /* The [^\n] meaning of \N ignores spaces and comments under the /x
12767 * modifier. The other meanings do not, so use a temporary until we find
12768 * out which we are being called with */
12769 skip_to_be_ignored_text(pRExC_state, &p,
12770 FALSE /* Don't force to /x */ );
12772 /* Disambiguate between \N meaning a named character versus \N meaning
12773 * [^\n]. The latter is assumed when the {...} following the \N is a legal
12774 * quantifier, or if there is no '{' at all */
12775 if (*p != '{' || regcurly(p)) {
12785 *node_p = reg_node(pRExC_state, REG_ANY);
12786 *flagp |= HASWIDTH|SIMPLE;
12788 Set_Node_Length(REGNODE_p(*(node_p)), 1); /* MJD */
12792 /* The test above made sure that the next real character is a '{', but
12793 * under the /x modifier, it could be separated by space (or a comment and
12794 * \n) and this is not allowed (for consistency with \x{...} and the
12795 * tokenizer handling of \N{NAME}). */
12796 if (*RExC_parse != '{') {
12797 vFAIL("Missing braces on \\N{}");
12800 RExC_parse++; /* Skip past the '{' */
12802 endbrace = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
12803 if (! endbrace) { /* no trailing brace */
12804 vFAIL2("Missing right brace on \\%c{}", 'N');
12807 /* Here, we have decided it should be a named character or sequence. These
12808 * imply Unicode semantics */
12809 REQUIRE_UNI_RULES(flagp, FALSE);
12811 /* \N{_} is what toke.c returns to us to indicate a name that evaluates to
12812 * nothing at all (not allowed under strict) */
12813 if (endbrace - RExC_parse == 1 && *RExC_parse == '_') {
12814 RExC_parse = endbrace;
12816 RExC_parse++; /* Position after the "}" */
12817 vFAIL("Zero length \\N{}");
12823 nextchar(pRExC_state);
12828 *node_p = reg_node(pRExC_state, NOTHING);
12832 if (endbrace - RExC_parse < 2 || ! strBEGINs(RExC_parse, "U+")) {
12834 /* Here, the name isn't of the form U+.... This can happen if the
12835 * pattern is single-quoted, so didn't get evaluated in toke.c. Now
12836 * is the time to find out what the name means */
12838 const STRLEN name_len = endbrace - RExC_parse;
12839 SV * value_sv; /* What does this name evaluate to */
12841 const U8 * value; /* string of name's value */
12842 STRLEN value_len; /* and its length */
12844 /* RExC_unlexed_names is a hash of names that weren't evaluated by
12845 * toke.c, and their values. Make sure is initialized */
12846 if (! RExC_unlexed_names) {
12847 RExC_unlexed_names = newHV();
12850 /* If we have already seen this name in this pattern, use that. This
12851 * allows us to only call the charnames handler once per name per
12852 * pattern. A broken or malicious handler could return something
12853 * different each time, which could cause the results to vary depending
12854 * on if something gets added or subtracted from the pattern that
12855 * causes the number of passes to change, for example */
12856 if ((value_svp = hv_fetch(RExC_unlexed_names, RExC_parse,
12859 value_sv = *value_svp;
12861 else { /* Otherwise we have to go out and get the name */
12862 const char * error_msg = NULL;
12863 value_sv = get_and_check_backslash_N_name(RExC_parse, endbrace,
12867 RExC_parse = endbrace;
12871 /* If no error message, should have gotten a valid return */
12874 /* Save the name's meaning for later use */
12875 if (! hv_store(RExC_unlexed_names, RExC_parse, name_len,
12878 Perl_croak(aTHX_ "panic: hv_store() unexpectedly failed");
12882 /* Here, we have the value the name evaluates to in 'value_sv' */
12883 value = (U8 *) SvPV(value_sv, value_len);
12885 /* See if the result is one code point vs 0 or multiple */
12886 if (value_len > 0 && value_len <= (UV) ((SvUTF8(value_sv))
12890 /* Here, exactly one code point. If that isn't what is wanted,
12892 if (! code_point_p) {
12897 /* Convert from string to numeric code point */
12898 *code_point_p = (SvUTF8(value_sv))
12899 ? valid_utf8_to_uvchr(value, NULL)
12902 /* Have parsed this entire single code point \N{...}. *cp_count
12903 * has already been set to 1, so don't do it again. */
12904 RExC_parse = endbrace;
12905 nextchar(pRExC_state);
12907 } /* End of is a single code point */
12909 /* Count the code points, if caller desires. The API says to do this
12910 * even if we will later return FALSE */
12914 *cp_count = (SvUTF8(value_sv))
12915 ? utf8_length(value, value + value_len)
12919 /* Fail if caller doesn't want to handle a multi-code-point sequence.
12920 * But don't back the pointer up if the caller wants to know how many
12921 * code points there are (they need to handle it themselves in this
12930 /* Convert this to a sub-pattern of the form "(?: ... )", and then call
12931 * reg recursively to parse it. That way, it retains its atomicness,
12932 * while not having to worry about any special handling that some code
12933 * points may have. */
12935 substitute_parse = newSVpvs("?:");
12936 sv_catsv(substitute_parse, value_sv);
12937 sv_catpv(substitute_parse, ")");
12940 /* The value should already be native, so no need to convert on EBCDIC
12942 assert(! RExC_recode_x_to_native);
12946 else { /* \N{U+...} */
12947 Size_t count = 0; /* code point count kept internally */
12949 /* We can get to here when the input is \N{U+...} or when toke.c has
12950 * converted a name to the \N{U+...} form. This include changing a
12951 * name that evaluates to multiple code points to \N{U+c1.c2.c3 ...} */
12953 RExC_parse += 2; /* Skip past the 'U+' */
12955 /* Code points are separated by dots. The '}' terminates the whole
12958 do { /* Loop until the ending brace */
12960 char * start_digit; /* The first of the current code point */
12961 if (! isXDIGIT(*RExC_parse)) {
12963 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12966 start_digit = RExC_parse;
12969 /* Loop through the hex digits of the current code point */
12971 /* Adding this digit will shift the result 4 bits. If that
12972 * result would be above the legal max, it's overflow */
12973 if (cp > MAX_LEGAL_CP >> 4) {
12975 /* Find the end of the code point */
12978 } while (isXDIGIT(*RExC_parse) || *RExC_parse == '_');
12980 /* Be sure to synchronize this message with the similar one
12982 vFAIL4("Use of code point 0x%.*s is not allowed; the"
12983 " permissible max is 0x%" UVxf,
12984 (int) (RExC_parse - start_digit), start_digit,
12988 /* Accumulate this (valid) digit into the running total */
12989 cp = (cp << 4) + READ_XDIGIT(RExC_parse);
12991 /* READ_XDIGIT advanced the input pointer. Ignore a single
12992 * underscore separator */
12993 if (*RExC_parse == '_' && isXDIGIT(RExC_parse[1])) {
12996 } while (isXDIGIT(*RExC_parse));
12998 /* Here, have accumulated the next code point */
12999 if (RExC_parse >= endbrace) { /* If done ... */
13004 /* Here, is a single code point; fail if doesn't want that */
13005 if (! code_point_p) {
13010 /* A single code point is easy to handle; just return it */
13011 *code_point_p = UNI_TO_NATIVE(cp);
13012 RExC_parse = endbrace;
13013 nextchar(pRExC_state);
13017 /* Here, the only legal thing would be a multiple character
13018 * sequence (of the form "\N{U+c1.c2. ... }". So the next
13019 * character must be a dot (and the one after that can't be the
13020 * endbrace, or we'd have something like \N{U+100.} ) */
13021 if (*RExC_parse != '.' || RExC_parse + 1 >= endbrace) {
13022 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
13023 ? UTF8SKIP(RExC_parse)
13025 if (RExC_parse >= endbrace) { /* Guard against malformed utf8 */
13026 RExC_parse = endbrace;
13028 vFAIL("Invalid hexadecimal number in \\N{U+...}");
13031 /* Here, looks like its really a multiple character sequence. Fail
13032 * if that's not what the caller wants. But continue with counting
13033 * and error checking if they still want a count */
13034 if (! node_p && ! cp_count) {
13038 /* What is done here is to convert this to a sub-pattern of the
13039 * form \x{char1}\x{char2}... and then call reg recursively to
13040 * parse it (enclosing in "(?: ... )" ). That way, it retains its
13041 * atomicness, while not having to worry about special handling
13042 * that some code points may have. We don't create a subpattern,
13043 * but go through the motions of code point counting and error
13044 * checking, if the caller doesn't want a node returned. */
13046 if (node_p && count == 1) {
13047 substitute_parse = newSVpvs("?:");
13053 /* Convert to notation the rest of the code understands */
13054 sv_catpvs(substitute_parse, "\\x{");
13055 sv_catpvn(substitute_parse, start_digit,
13056 RExC_parse - start_digit);
13057 sv_catpvs(substitute_parse, "}");
13060 /* Move to after the dot (or ending brace the final time through.)
13065 } while (RExC_parse < endbrace);
13067 if (! node_p) { /* Doesn't want the node */
13074 sv_catpvs(substitute_parse, ")");
13077 /* The values are Unicode, and therefore have to be converted to native
13078 * on a non-Unicode (meaning non-ASCII) platform. */
13079 RExC_recode_x_to_native = 1;
13084 /* Here, we have the string the name evaluates to, ready to be parsed,
13085 * stored in 'substitute_parse' as a series of valid "\x{...}\x{...}"
13086 * constructs. This can be called from within a substitute parse already.
13087 * The error reporting mechanism doesn't work for 2 levels of this, but the
13088 * code above has validated this new construct, so there should be no
13089 * errors generated by the below. And this isn' an exact copy, so the
13090 * mechanism to seamlessly deal with this won't work, so turn off warnings
13092 save_start = RExC_start;
13093 orig_end = RExC_end;
13095 RExC_parse = RExC_start = SvPVX(substitute_parse);
13096 RExC_end = RExC_parse + SvCUR(substitute_parse);
13097 TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE;
13099 *node_p = reg(pRExC_state, 1, &flags, depth+1);
13101 /* Restore the saved values */
13103 RExC_start = save_start;
13104 RExC_parse = endbrace;
13105 RExC_end = orig_end;
13107 RExC_recode_x_to_native = 0;
13110 SvREFCNT_dec_NN(substitute_parse);
13113 RETURN_FAIL_ON_RESTART(flags, flagp);
13114 FAIL2("panic: reg returned failure to grok_bslash_N, flags=%#" UVxf,
13117 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
13119 nextchar(pRExC_state);
13125 PERL_STATIC_INLINE U8
13126 S_compute_EXACTish(RExC_state_t *pRExC_state)
13130 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
13138 op = get_regex_charset(RExC_flags);
13139 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
13140 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
13141 been, so there is no hole */
13144 return op + EXACTF;
13148 S_new_regcurly(const char *s, const char *e)
13150 /* This is a temporary function designed to match the most lenient form of
13151 * a {m,n} quantifier we ever envision, with either number omitted, and
13152 * spaces anywhere between/before/after them.
13154 * If this function fails, then the string it matches is very unlikely to
13155 * ever be considered a valid quantifier, so we can allow the '{' that
13156 * begins it to be considered as a literal */
13158 bool has_min = FALSE;
13159 bool has_max = FALSE;
13161 PERL_ARGS_ASSERT_NEW_REGCURLY;
13163 if (s >= e || *s++ != '{')
13166 while (s < e && isSPACE(*s)) {
13169 while (s < e && isDIGIT(*s)) {
13173 while (s < e && isSPACE(*s)) {
13179 while (s < e && isSPACE(*s)) {
13182 while (s < e && isDIGIT(*s)) {
13186 while (s < e && isSPACE(*s)) {
13191 return s < e && *s == '}' && (has_min || has_max);
13194 /* Parse backref decimal value, unless it's too big to sensibly be a backref,
13195 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
13198 S_backref_value(char *p, char *e)
13200 const char* endptr = e;
13202 if (grok_atoUV(p, &val, &endptr) && val <= I32_MAX)
13209 - regatom - the lowest level
13211 Try to identify anything special at the start of the current parse position.
13212 If there is, then handle it as required. This may involve generating a
13213 single regop, such as for an assertion; or it may involve recursing, such as
13214 to handle a () structure.
13216 If the string doesn't start with something special then we gobble up
13217 as much literal text as we can. If we encounter a quantifier, we have to
13218 back off the final literal character, as that quantifier applies to just it
13219 and not to the whole string of literals.
13221 Once we have been able to handle whatever type of thing started the
13222 sequence, we return the offset into the regex engine program being compiled
13223 at which any next regnode should be placed.
13225 Returns 0, setting *flagp to TRYAGAIN if reg() returns 0 with TRYAGAIN.
13226 Returns 0, setting *flagp to RESTART_PARSE if the parse needs to be
13227 restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
13228 Otherwise does not return 0.
13230 Note: we have to be careful with escapes, as they can be both literal
13231 and special, and in the case of \10 and friends, context determines which.
13233 A summary of the code structure is:
13235 switch (first_byte) {
13236 cases for each special:
13237 handle this special;
13240 switch (2nd byte) {
13241 cases for each unambiguous special:
13242 handle this special;
13244 cases for each ambigous special/literal:
13246 if (special) handle here
13248 default: // unambiguously literal:
13251 default: // is a literal char
13254 create EXACTish node for literal;
13255 while (more input and node isn't full) {
13256 switch (input_byte) {
13257 cases for each special;
13258 make sure parse pointer is set so that the next call to
13259 regatom will see this special first
13260 goto loopdone; // EXACTish node terminated by prev. char
13262 append char to EXACTISH node;
13264 get next input byte;
13268 return the generated node;
13270 Specifically there are two separate switches for handling
13271 escape sequences, with the one for handling literal escapes requiring
13272 a dummy entry for all of the special escapes that are actually handled
13277 STATIC regnode_offset
13278 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
13281 regnode_offset ret = 0;
13288 GET_RE_DEBUG_FLAGS_DECL;
13290 *flagp = WORST; /* Tentatively. */
13292 DEBUG_PARSE("atom");
13294 PERL_ARGS_ASSERT_REGATOM;
13297 parse_start = RExC_parse;
13298 assert(RExC_parse < RExC_end);
13299 switch ((U8)*RExC_parse) {
13301 RExC_seen_zerolen++;
13302 nextchar(pRExC_state);
13303 if (RExC_flags & RXf_PMf_MULTILINE)
13304 ret = reg_node(pRExC_state, MBOL);
13306 ret = reg_node(pRExC_state, SBOL);
13307 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13310 nextchar(pRExC_state);
13312 RExC_seen_zerolen++;
13313 if (RExC_flags & RXf_PMf_MULTILINE)
13314 ret = reg_node(pRExC_state, MEOL);
13316 ret = reg_node(pRExC_state, SEOL);
13317 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13320 nextchar(pRExC_state);
13321 if (RExC_flags & RXf_PMf_SINGLELINE)
13322 ret = reg_node(pRExC_state, SANY);
13324 ret = reg_node(pRExC_state, REG_ANY);
13325 *flagp |= HASWIDTH|SIMPLE;
13327 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13331 char * const oregcomp_parse = ++RExC_parse;
13332 ret = regclass(pRExC_state, flagp, depth+1,
13333 FALSE, /* means parse the whole char class */
13334 TRUE, /* allow multi-char folds */
13335 FALSE, /* don't silence non-portable warnings. */
13336 (bool) RExC_strict,
13337 TRUE, /* Allow an optimized regnode result */
13340 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13341 FAIL2("panic: regclass returned failure to regatom, flags=%#" UVxf,
13344 if (*RExC_parse != ']') {
13345 RExC_parse = oregcomp_parse;
13346 vFAIL("Unmatched [");
13348 nextchar(pRExC_state);
13349 Set_Node_Length(REGNODE_p(ret), RExC_parse - oregcomp_parse + 1); /* MJD */
13353 nextchar(pRExC_state);
13354 ret = reg(pRExC_state, 2, &flags, depth+1);
13356 if (flags & TRYAGAIN) {
13357 if (RExC_parse >= RExC_end) {
13358 /* Make parent create an empty node if needed. */
13359 *flagp |= TRYAGAIN;
13364 RETURN_FAIL_ON_RESTART(flags, flagp);
13365 FAIL2("panic: reg returned failure to regatom, flags=%#" UVxf,
13368 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
13372 if (flags & TRYAGAIN) {
13373 *flagp |= TRYAGAIN;
13376 vFAIL("Internal urp");
13377 /* Supposed to be caught earlier. */
13383 vFAIL("Quantifier follows nothing");
13388 This switch handles escape sequences that resolve to some kind
13389 of special regop and not to literal text. Escape sequences that
13390 resolve to literal text are handled below in the switch marked
13393 Every entry in this switch *must* have a corresponding entry
13394 in the literal escape switch. However, the opposite is not
13395 required, as the default for this switch is to jump to the
13396 literal text handling code.
13399 switch ((U8)*RExC_parse) {
13400 /* Special Escapes */
13402 RExC_seen_zerolen++;
13403 ret = reg_node(pRExC_state, SBOL);
13404 /* SBOL is shared with /^/ so we set the flags so we can tell
13405 * /\A/ from /^/ in split. */
13406 FLAGS(REGNODE_p(ret)) = 1;
13408 goto finish_meta_pat;
13410 ret = reg_node(pRExC_state, GPOS);
13411 RExC_seen |= REG_GPOS_SEEN;
13413 goto finish_meta_pat;
13415 RExC_seen_zerolen++;
13416 ret = reg_node(pRExC_state, KEEPS);
13418 /* XXX:dmq : disabling in-place substitution seems to
13419 * be necessary here to avoid cases of memory corruption, as
13420 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
13422 RExC_seen |= REG_LOOKBEHIND_SEEN;
13423 goto finish_meta_pat;
13425 ret = reg_node(pRExC_state, SEOL);
13427 RExC_seen_zerolen++; /* Do not optimize RE away */
13428 goto finish_meta_pat;
13430 ret = reg_node(pRExC_state, EOS);
13432 RExC_seen_zerolen++; /* Do not optimize RE away */
13433 goto finish_meta_pat;
13435 vFAIL("\\C no longer supported");
13437 ret = reg_node(pRExC_state, CLUMP);
13438 *flagp |= HASWIDTH;
13439 goto finish_meta_pat;
13445 arg = ANYOF_WORDCHAR;
13454 regex_charset charset = get_regex_charset(RExC_flags);
13456 RExC_seen_zerolen++;
13457 RExC_seen |= REG_LOOKBEHIND_SEEN;
13458 op = BOUND + charset;
13460 if (RExC_parse >= RExC_end || *(RExC_parse + 1) != '{') {
13461 flags = TRADITIONAL_BOUND;
13462 if (op > BOUNDA) { /* /aa is same as /a */
13468 char name = *RExC_parse;
13469 char * endbrace = NULL;
13471 endbrace = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
13474 vFAIL2("Missing right brace on \\%c{}", name);
13476 /* XXX Need to decide whether to take spaces or not. Should be
13477 * consistent with \p{}, but that currently is SPACE, which
13478 * means vertical too, which seems wrong
13479 * while (isBLANK(*RExC_parse)) {
13482 if (endbrace == RExC_parse) {
13483 RExC_parse++; /* After the '}' */
13484 vFAIL2("Empty \\%c{}", name);
13486 length = endbrace - RExC_parse;
13487 /*while (isBLANK(*(RExC_parse + length - 1))) {
13490 switch (*RExC_parse) {
13493 && (memNEs(RExC_parse + 1, length - 1, "cb")))
13495 goto bad_bound_type;
13500 if (length != 2 || *(RExC_parse + 1) != 'b') {
13501 goto bad_bound_type;
13506 if (length != 2 || *(RExC_parse + 1) != 'b') {
13507 goto bad_bound_type;
13512 if (length != 2 || *(RExC_parse + 1) != 'b') {
13513 goto bad_bound_type;
13519 RExC_parse = endbrace;
13521 "'%" UTF8f "' is an unknown bound type",
13522 UTF8fARG(UTF, length, endbrace - length));
13523 NOT_REACHED; /*NOTREACHED*/
13525 RExC_parse = endbrace;
13526 REQUIRE_UNI_RULES(flagp, 0);
13531 else if (op >= BOUNDA) { /* /aa is same as /a */
13535 /* Don't have to worry about UTF-8, in this message because
13536 * to get here the contents of the \b must be ASCII */
13537 ckWARN4reg(RExC_parse + 1, /* Include the '}' in msg */
13538 "Using /u for '%.*s' instead of /%s",
13540 endbrace - length + 1,
13541 (charset == REGEX_ASCII_RESTRICTED_CHARSET)
13542 ? ASCII_RESTRICT_PAT_MODS
13543 : ASCII_MORE_RESTRICT_PAT_MODS);
13548 RExC_seen_d_op = TRUE;
13550 else if (op == BOUNDL) {
13551 RExC_contains_locale = 1;
13555 op += NBOUND - BOUND;
13558 ret = reg_node(pRExC_state, op);
13559 FLAGS(REGNODE_p(ret)) = flags;
13563 goto finish_meta_pat;
13571 if (! DEPENDS_SEMANTICS) {
13575 /* \d doesn't have any matches in the upper Latin1 range, hence /d
13576 * is equivalent to /u. Changing to /u saves some branches at
13579 goto join_posix_op_known;
13582 ret = reg_node(pRExC_state, LNBREAK);
13583 *flagp |= HASWIDTH|SIMPLE;
13584 goto finish_meta_pat;
13592 goto join_posix_op_known;
13598 arg = ANYOF_VERTWS;
13600 goto join_posix_op_known;
13610 op = POSIXD + get_regex_charset(RExC_flags);
13611 if (op > POSIXA) { /* /aa is same as /a */
13614 else if (op == POSIXL) {
13615 RExC_contains_locale = 1;
13617 else if (op == POSIXD) {
13618 RExC_seen_d_op = TRUE;
13621 join_posix_op_known:
13624 op += NPOSIXD - POSIXD;
13627 ret = reg_node(pRExC_state, op);
13628 FLAGS(REGNODE_p(ret)) = namedclass_to_classnum(arg);
13630 *flagp |= HASWIDTH|SIMPLE;
13634 if ( UCHARAT(RExC_parse + 1) == '{'
13635 && UNLIKELY(! new_regcurly(RExC_parse + 1, RExC_end)))
13638 vFAIL("Unescaped left brace in regex is illegal here");
13640 nextchar(pRExC_state);
13641 Set_Node_Length(REGNODE_p(ret), 2); /* MJD */
13647 ret = regclass(pRExC_state, flagp, depth+1,
13648 TRUE, /* means just parse this element */
13649 FALSE, /* don't allow multi-char folds */
13650 FALSE, /* don't silence non-portable warnings. It
13651 would be a bug if these returned
13653 (bool) RExC_strict,
13654 TRUE, /* Allow an optimized regnode result */
13656 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13657 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
13658 * multi-char folds are allowed. */
13660 FAIL2("panic: regclass returned failure to regatom, flags=%#" UVxf,
13665 Set_Node_Offset(REGNODE_p(ret), parse_start);
13666 Set_Node_Cur_Length(REGNODE_p(ret), parse_start - 2);
13667 nextchar(pRExC_state);
13670 /* Handle \N, \N{} and \N{NAMED SEQUENCE} (the latter meaning the
13671 * \N{...} evaluates to a sequence of more than one code points).
13672 * The function call below returns a regnode, which is our result.
13673 * The parameters cause it to fail if the \N{} evaluates to a
13674 * single code point; we handle those like any other literal. The
13675 * reason that the multicharacter case is handled here and not as
13676 * part of the EXACtish code is because of quantifiers. In
13677 * /\N{BLAH}+/, the '+' applies to the whole thing, and doing it
13678 * this way makes that Just Happen. dmq.
13679 * join_exact() will join this up with adjacent EXACTish nodes
13680 * later on, if appropriate. */
13682 if (grok_bslash_N(pRExC_state,
13683 &ret, /* Want a regnode returned */
13684 NULL, /* Fail if evaluates to a single code
13686 NULL, /* Don't need a count of how many code
13695 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13697 /* Here, evaluates to a single code point. Go get that */
13698 RExC_parse = parse_start;
13701 case 'k': /* Handle \k<NAME> and \k'NAME' */
13705 if ( RExC_parse >= RExC_end - 1
13706 || (( ch = RExC_parse[1]) != '<'
13711 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
13712 vFAIL2("Sequence %.2s... not terminated", parse_start);
13715 ret = handle_named_backref(pRExC_state,
13727 case '1': case '2': case '3': case '4':
13728 case '5': case '6': case '7': case '8': case '9':
13733 if (*RExC_parse == 'g') {
13737 if (*RExC_parse == '{') {
13741 if (*RExC_parse == '-') {
13745 if (hasbrace && !isDIGIT(*RExC_parse)) {
13746 if (isrel) RExC_parse--;
13748 goto parse_named_seq;
13751 if (RExC_parse >= RExC_end) {
13752 goto unterminated_g;
13754 num = S_backref_value(RExC_parse, RExC_end);
13756 vFAIL("Reference to invalid group 0");
13757 else if (num == I32_MAX) {
13758 if (isDIGIT(*RExC_parse))
13759 vFAIL("Reference to nonexistent group");
13762 vFAIL("Unterminated \\g... pattern");
13766 num = RExC_npar - num;
13768 vFAIL("Reference to nonexistent or unclosed group");
13772 num = S_backref_value(RExC_parse, RExC_end);
13773 /* bare \NNN might be backref or octal - if it is larger
13774 * than or equal RExC_npar then it is assumed to be an
13775 * octal escape. Note RExC_npar is +1 from the actual
13776 * number of parens. */
13777 /* Note we do NOT check if num == I32_MAX here, as that is
13778 * handled by the RExC_npar check */
13781 /* any numeric escape < 10 is always a backref */
13783 /* any numeric escape < RExC_npar is a backref */
13784 && num >= RExC_npar
13785 /* cannot be an octal escape if it starts with 8 */
13786 && *RExC_parse != '8'
13787 /* cannot be an octal escape it it starts with 9 */
13788 && *RExC_parse != '9'
13790 /* Probably not meant to be a backref, instead likely
13791 * to be an octal character escape, e.g. \35 or \777.
13792 * The above logic should make it obvious why using
13793 * octal escapes in patterns is problematic. - Yves */
13794 RExC_parse = parse_start;
13799 /* At this point RExC_parse points at a numeric escape like
13800 * \12 or \88 or something similar, which we should NOT treat
13801 * as an octal escape. It may or may not be a valid backref
13802 * escape. For instance \88888888 is unlikely to be a valid
13804 while (isDIGIT(*RExC_parse))
13807 if (*RExC_parse != '}')
13808 vFAIL("Unterminated \\g{...} pattern");
13811 if (num >= (I32)RExC_npar) {
13813 /* It might be a forward reference; we can't fail until we
13814 * know, by completing the parse to get all the groups, and
13815 * then reparsing */
13816 if (ALL_PARENS_COUNTED) {
13817 if (num >= RExC_total_parens) {
13818 vFAIL("Reference to nonexistent group");
13822 REQUIRE_PARENS_PASS;
13826 ret = reganode(pRExC_state,
13829 : (ASCII_FOLD_RESTRICTED)
13831 : (AT_LEAST_UNI_SEMANTICS)
13837 if (OP(REGNODE_p(ret)) == REFF) {
13838 RExC_seen_d_op = TRUE;
13840 *flagp |= HASWIDTH;
13842 /* override incorrect value set in reganode MJD */
13843 Set_Node_Offset(REGNODE_p(ret), parse_start);
13844 Set_Node_Cur_Length(REGNODE_p(ret), parse_start-1);
13845 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13846 FALSE /* Don't force to /x */ );
13850 if (RExC_parse >= RExC_end)
13851 FAIL("Trailing \\");
13854 /* Do not generate "unrecognized" warnings here, we fall
13855 back into the quick-grab loop below */
13856 RExC_parse = parse_start;
13858 } /* end of switch on a \foo sequence */
13863 /* '#' comments should have been spaced over before this function was
13865 assert((RExC_flags & RXf_PMf_EXTENDED) == 0);
13867 if (RExC_flags & RXf_PMf_EXTENDED) {
13868 RExC_parse = reg_skipcomment( pRExC_state, RExC_parse );
13869 if (RExC_parse < RExC_end)
13879 /* Here, we have determined that the next thing is probably a
13880 * literal character. RExC_parse points to the first byte of its
13881 * definition. (It still may be an escape sequence that evaluates
13882 * to a single character) */
13889 /* This allows us to fill a node with just enough spare so that if the final
13890 * character folds, its expansion is guaranteed to fit */
13891 #define MAX_NODE_STRING_SIZE (255-UTF8_MAXBYTES_CASE)
13894 U8 upper_parse = MAX_NODE_STRING_SIZE;
13896 /* We start out as an EXACT node, even if under /i, until we find a
13897 * character which is in a fold. The algorithm now segregates into
13898 * separate nodes, characters that fold from those that don't under
13899 * /i. (This hopefully will create nodes that are fixed strings
13900 * even under /i, giving the optimizer something to grab on to.)
13901 * So, if a node has something in it and the next character is in
13902 * the opposite category, that node is closed up, and the function
13903 * returns. Then regatom is called again, and a new node is
13904 * created for the new category. */
13905 U8 node_type = EXACT;
13907 /* Assume the node will be fully used; the excess is given back at
13908 * the end. We can't make any other length assumptions, as a byte
13909 * input sequence could shrink down. */
13910 Ptrdiff_t initial_size = STR_SZ(256);
13912 bool next_is_quantifier;
13913 char * oldp = NULL;
13915 /* We can convert EXACTF nodes to EXACTFU if they contain only
13916 * characters that match identically regardless of the target
13917 * string's UTF8ness. The reason to do this is that EXACTF is not
13918 * trie-able, EXACTFU is, and EXACTFU requires fewer operations at
13921 * Similarly, we can convert EXACTFL nodes to EXACTFLU8 if they
13922 * contain only above-Latin1 characters (hence must be in UTF8),
13923 * which don't participate in folds with Latin1-range characters,
13924 * as the latter's folds aren't known until runtime. */
13925 bool maybe_exactfu = FOLD && (DEPENDS_SEMANTICS || LOC);
13927 /* Single-character EXACTish nodes are almost always SIMPLE. This
13928 * allows us to override this as encountered */
13929 U8 maybe_SIMPLE = SIMPLE;
13931 /* Does this node contain something that can't match unless the
13932 * target string is (also) in UTF-8 */
13933 bool requires_utf8_target = FALSE;
13935 /* The sequence 'ss' is problematic in non-UTF-8 patterns. */
13936 bool has_ss = FALSE;
13938 /* So is the MICRO SIGN */
13939 bool has_micro_sign = FALSE;
13941 /* Allocate an EXACT node. The node_type may change below to
13942 * another EXACTish node, but since the size of the node doesn't
13943 * change, it works */
13944 ret = regnode_guts(pRExC_state, node_type, initial_size, "exact");
13945 FILL_NODE(ret, node_type);
13948 s = STRING(REGNODE_p(ret));
13954 /* This breaks under rare circumstances. If folding, we do not
13955 * want to split a node at a character that is a non-final in a
13956 * multi-char fold, as an input string could just happen to want to
13957 * match across the node boundary. The code at the end of the loop
13958 * looks for this, and backs off until it finds not such a
13959 * character, but it is possible (though extremely, extremely
13960 * unlikely) for all characters in the node to be non-final fold
13961 * ones, in which case we just leave the node fully filled, and
13962 * hope that it doesn't match the string in just the wrong place */
13964 assert( ! UTF /* Is at the beginning of a character */
13965 || UTF8_IS_INVARIANT(UCHARAT(RExC_parse))
13966 || UTF8_IS_START(UCHARAT(RExC_parse)));
13968 /* Here, we have a literal character. Find the maximal string of
13969 * them in the input that we can fit into a single EXACTish node.
13970 * We quit at the first non-literal or when the node gets full, or
13971 * under /i the categorization of folding/non-folding character
13973 for (p = RExC_parse; len < upper_parse && p < RExC_end; ) {
13975 /* In most cases each iteration adds one byte to the output.
13976 * The exceptions override this */
13977 Size_t added_len = 1;
13981 /* White space has already been ignored */
13982 assert( (RExC_flags & RXf_PMf_EXTENDED) == 0
13983 || ! is_PATWS_safe((p), RExC_end, UTF));
13995 /* Literal Escapes Switch
13997 This switch is meant to handle escape sequences that
13998 resolve to a literal character.
14000 Every escape sequence that represents something
14001 else, like an assertion or a char class, is handled
14002 in the switch marked 'Special Escapes' above in this
14003 routine, but also has an entry here as anything that
14004 isn't explicitly mentioned here will be treated as
14005 an unescaped equivalent literal.
14008 switch ((U8)*++p) {
14010 /* These are all the special escapes. */
14011 case 'A': /* Start assertion */
14012 case 'b': case 'B': /* Word-boundary assertion*/
14013 case 'C': /* Single char !DANGEROUS! */
14014 case 'd': case 'D': /* digit class */
14015 case 'g': case 'G': /* generic-backref, pos assertion */
14016 case 'h': case 'H': /* HORIZWS */
14017 case 'k': case 'K': /* named backref, keep marker */
14018 case 'p': case 'P': /* Unicode property */
14019 case 'R': /* LNBREAK */
14020 case 's': case 'S': /* space class */
14021 case 'v': case 'V': /* VERTWS */
14022 case 'w': case 'W': /* word class */
14023 case 'X': /* eXtended Unicode "combining
14024 character sequence" */
14025 case 'z': case 'Z': /* End of line/string assertion */
14029 /* Anything after here is an escape that resolves to a
14030 literal. (Except digits, which may or may not)
14036 case 'N': /* Handle a single-code point named character. */
14037 RExC_parse = p + 1;
14038 if (! grok_bslash_N(pRExC_state,
14039 NULL, /* Fail if evaluates to
14040 anything other than a
14041 single code point */
14042 &ender, /* The returned single code
14044 NULL, /* Don't need a count of
14045 how many code points */
14050 if (*flagp & NEED_UTF8)
14051 FAIL("panic: grok_bslash_N set NEED_UTF8");
14052 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
14054 /* Here, it wasn't a single code point. Go close
14055 * up this EXACTish node. The switch() prior to
14056 * this switch handles the other cases */
14057 RExC_parse = p = oldp;
14061 RExC_parse = parse_start;
14063 /* The \N{} means the pattern, if previously /d,
14064 * becomes /u. That means it can't be an EXACTF node,
14065 * but an EXACTFU */
14066 if (node_type == EXACTF) {
14067 node_type = EXACTFU;
14069 /* If the node already contains something that
14070 * differs between EXACTF and EXACTFU, reparse it
14072 if (! maybe_exactfu) {
14093 ender = ESC_NATIVE;
14103 const char* error_msg;
14105 bool valid = grok_bslash_o(&p,
14109 TO_OUTPUT_WARNINGS(p),
14110 (bool) RExC_strict,
14111 TRUE, /* Output warnings
14116 RExC_parse = p; /* going to die anyway; point
14117 to exact spot of failure */
14120 UPDATE_WARNINGS_LOC(p - 1);
14126 UV result = UV_MAX; /* initialize to erroneous
14128 const char* error_msg;
14130 bool valid = grok_bslash_x(&p,
14134 TO_OUTPUT_WARNINGS(p),
14135 (bool) RExC_strict,
14136 TRUE, /* Silence warnings
14141 RExC_parse = p; /* going to die anyway; point
14142 to exact spot of failure */
14145 UPDATE_WARNINGS_LOC(p - 1);
14148 if (ender < 0x100) {
14150 if (RExC_recode_x_to_native) {
14151 ender = LATIN1_TO_NATIVE(ender);
14159 ender = grok_bslash_c(*p, TO_OUTPUT_WARNINGS(p));
14160 UPDATE_WARNINGS_LOC(p);
14163 case '8': case '9': /* must be a backreference */
14165 /* we have an escape like \8 which cannot be an octal escape
14166 * so we exit the loop, and let the outer loop handle this
14167 * escape which may or may not be a legitimate backref. */
14169 case '1': case '2': case '3':case '4':
14170 case '5': case '6': case '7':
14171 /* When we parse backslash escapes there is ambiguity
14172 * between backreferences and octal escapes. Any escape
14173 * from \1 - \9 is a backreference, any multi-digit
14174 * escape which does not start with 0 and which when
14175 * evaluated as decimal could refer to an already
14176 * parsed capture buffer is a back reference. Anything
14179 * Note this implies that \118 could be interpreted as
14180 * 118 OR as "\11" . "8" depending on whether there
14181 * were 118 capture buffers defined already in the
14184 /* NOTE, RExC_npar is 1 more than the actual number of
14185 * parens we have seen so far, hence the "<" as opposed
14187 if ( !isDIGIT(p[1]) || S_backref_value(p, RExC_end) < RExC_npar)
14188 { /* Not to be treated as an octal constant, go
14196 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
14198 ender = grok_oct(p, &numlen, &flags, NULL);
14200 if ( isDIGIT(*p) /* like \08, \178 */
14201 && ckWARN(WARN_REGEXP)
14204 reg_warn_non_literal_string(
14206 form_short_octal_warning(p, numlen));
14212 FAIL("Trailing \\");
14215 if (isALPHANUMERIC(*p)) {
14216 /* An alpha followed by '{' is going to fail next
14217 * iteration, so don't output this warning in that
14219 if (! isALPHA(*p) || *(p + 1) != '{') {
14220 ckWARN2reg(p + 1, "Unrecognized escape \\%.1s"
14221 " passed through", p);
14224 goto normal_default;
14225 } /* End of switch on '\' */
14228 /* Trying to gain new uses for '{' without breaking too
14229 * much existing code is hard. The solution currently
14231 * 1) If there is no ambiguity that a '{' should always
14232 * be taken literally, at the start of a construct, we
14234 * 2) If the literal '{' conflicts with our desired use
14235 * of it as a metacharacter, we die. The deprecation
14236 * cycles for this have come and gone.
14237 * 3) If there is ambiguity, we raise a simple warning.
14238 * This could happen, for example, if the user
14239 * intended it to introduce a quantifier, but slightly
14240 * misspelled the quantifier. Without this warning,
14241 * the quantifier would silently be taken as a literal
14242 * string of characters instead of a meta construct */
14243 if (len || (p > RExC_start && isALPHA_A(*(p - 1)))) {
14245 || ( p > parse_start + 1
14246 && isALPHA_A(*(p - 1))
14247 && *(p - 2) == '\\')
14248 || new_regcurly(p, RExC_end))
14250 RExC_parse = p + 1;
14251 vFAIL("Unescaped left brace in regex is "
14254 ckWARNreg(p + 1, "Unescaped left brace in regex is"
14255 " passed through");
14257 goto normal_default;
14260 if (p > RExC_parse && RExC_strict) {
14261 ckWARN2reg(p + 1, "Unescaped literal '%c'", *p);
14264 default: /* A literal character */
14266 if (! UTF8_IS_INVARIANT(*p) && UTF) {
14268 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
14269 &numlen, UTF8_ALLOW_DEFAULT);
14275 } /* End of switch on the literal */
14277 /* Here, have looked at the literal character, and <ender>
14278 * contains its ordinal; <p> points to the character after it.
14282 REQUIRE_UTF8(flagp);
14285 /* We need to check if the next non-ignored thing is a
14286 * quantifier. Move <p> to after anything that should be
14287 * ignored, which, as a side effect, positions <p> for the next
14288 * loop iteration */
14289 skip_to_be_ignored_text(pRExC_state, &p,
14290 FALSE /* Don't force to /x */ );
14292 /* If the next thing is a quantifier, it applies to this
14293 * character only, which means that this character has to be in
14294 * its own node and can't just be appended to the string in an
14295 * existing node, so if there are already other characters in
14296 * the node, close the node with just them, and set up to do
14297 * this character again next time through, when it will be the
14298 * only thing in its new node */
14300 next_is_quantifier = LIKELY(p < RExC_end)
14301 && UNLIKELY(ISMULT2(p));
14303 if (next_is_quantifier && LIKELY(len)) {
14308 /* Ready to add 'ender' to the node */
14310 if (! FOLD) { /* The simple case, just append the literal */
14313 if (UVCHR_IS_INVARIANT(ender) || ! UTF) {
14314 *(s++) = (char) ender;
14317 U8 * new_s = uvchr_to_utf8((U8*)s, ender);
14318 added_len = (char *) new_s - s;
14319 s = (char *) new_s;
14322 requires_utf8_target = TRUE;
14326 else if (LOC && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)) {
14328 /* Here are folding under /l, and the code point is
14329 * problematic. If this is the first character in the
14330 * node, change the node type to folding. Otherwise, if
14331 * this is the first problematic character, close up the
14332 * existing node, so can start a new node with this one */
14334 node_type = EXACTFL;
14335 RExC_contains_locale = 1;
14337 else if (node_type == EXACT) {
14342 /* This problematic code point means we can't simplify
14344 maybe_exactfu = FALSE;
14346 /* Here, we are adding a problematic fold character.
14347 * "Problematic" in this context means that its fold isn't
14348 * known until runtime. (The non-problematic code points
14349 * are the above-Latin1 ones that fold to also all
14350 * above-Latin1. Their folds don't vary no matter what the
14351 * locale is.) But here we have characters whose fold
14352 * depends on the locale. We just add in the unfolded
14353 * character, and wait until runtime to fold it */
14354 goto not_fold_common;
14356 else /* regular fold; see if actually is in a fold */
14357 if ( (ender < 256 && ! IS_IN_SOME_FOLD_L1(ender))
14359 && ! _invlist_contains_cp(PL_in_some_fold, ender)))
14361 /* Here, folding, but the character isn't in a fold.
14363 * Start a new node if previous characters in the node were
14365 if (len && node_type != EXACT) {
14370 /* Here, continuing a node with non-folded characters. Add
14372 goto not_fold_common;
14374 else { /* Here, does participate in some fold */
14376 /* If this is the first character in the node, change its
14377 * type to folding. Otherwise, if this is the first
14378 * folding character in the node, close up the existing
14379 * node, so can start a new node with this one. */
14381 node_type = compute_EXACTish(pRExC_state);
14383 else if (node_type == EXACT) {
14388 if (UTF) { /* Use the folded value */
14389 if (UVCHR_IS_INVARIANT(ender)) {
14390 *(s)++ = (U8) toFOLD(ender);
14393 ender = _to_uni_fold_flags(
14397 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
14398 ? FOLD_FLAGS_NOMIX_ASCII
14403 && LIKELY(ender != GREEK_SMALL_LETTER_MU))
14405 /* U+B5 folds to the MU, so its possible for a
14406 * non-UTF-8 target to match it */
14407 requires_utf8_target = TRUE;
14413 /* Here is non-UTF8. First, see if the character's
14414 * fold differs between /d and /u. */
14415 if (PL_fold[ender] != PL_fold_latin1[ender]) {
14416 maybe_exactfu = FALSE;
14419 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
14420 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
14421 || UNICODE_DOT_DOT_VERSION > 0)
14423 /* On non-ancient Unicode versions, this includes the
14424 * multi-char fold SHARP S to 'ss' */
14426 if ( UNLIKELY(ender == LATIN_SMALL_LETTER_SHARP_S)
14427 || ( isALPHA_FOLD_EQ(ender, 's')
14429 && isALPHA_FOLD_EQ(*(s-1), 's')))
14431 /* Here, we have one of the following:
14432 * a) a SHARP S. This folds to 'ss' only under
14433 * /u rules. If we are in that situation,
14434 * fold the SHARP S to 'ss'. See the comments
14435 * for join_exact() as to why we fold this
14436 * non-UTF at compile time, and no others.
14437 * b) 'ss'. When under /u, there's nothing
14438 * special needed to be done here. The
14439 * previous iteration handled the first 's',
14440 * and this iteration will handle the second.
14441 * If, on the otherhand it's not /u, we have
14442 * to exclude the possibility of moving to /u,
14443 * so that we won't generate an unwanted
14444 * match, unless, at runtime, the target
14445 * string is in UTF-8.
14449 maybe_exactfu = FALSE; /* Can't generate an
14450 EXACTFU node (unless we
14451 already are in one) */
14452 if (UNLIKELY(ender == LATIN_SMALL_LETTER_SHARP_S)) {
14454 if (node_type == EXACTFU) {
14457 /* Let the code below add in the extra 's' */
14465 else if (UNLIKELY(ender == MICRO_SIGN)) {
14466 has_micro_sign = TRUE;
14469 *(s++) = (DEPENDS_SEMANTICS)
14470 ? (char) toFOLD(ender)
14472 /* Under /u, the fold of any character in
14473 * the 0-255 range happens to be its
14474 * lowercase equivalent, except for LATIN
14475 * SMALL LETTER SHARP S, which was handled
14476 * above, and the MICRO SIGN, whose fold
14477 * requires UTF-8 to represent. */
14478 : (char) toLOWER_L1(ender);
14480 } /* End of adding current character to the node */
14484 if (next_is_quantifier) {
14486 /* Here, the next input is a quantifier, and to get here,
14487 * the current character is the only one in the node. */
14491 } /* End of loop through literal characters */
14493 /* Here we have either exhausted the input or ran out of room in
14494 * the node. (If we encountered a character that can't be in the
14495 * node, transfer is made directly to <loopdone>, and so we
14496 * wouldn't have fallen off the end of the loop.) In the latter
14497 * case, we artificially have to split the node into two, because
14498 * we just don't have enough space to hold everything. This
14499 * creates a problem if the final character participates in a
14500 * multi-character fold in the non-final position, as a match that
14501 * should have occurred won't, due to the way nodes are matched,
14502 * and our artificial boundary. So back off until we find a non-
14503 * problematic character -- one that isn't at the beginning or
14504 * middle of such a fold. (Either it doesn't participate in any
14505 * folds, or appears only in the final position of all the folds it
14506 * does participate in.) A better solution with far fewer false
14507 * positives, and that would fill the nodes more completely, would
14508 * be to actually have available all the multi-character folds to
14509 * test against, and to back-off only far enough to be sure that
14510 * this node isn't ending with a partial one. <upper_parse> is set
14511 * further below (if we need to reparse the node) to include just
14512 * up through that final non-problematic character that this code
14513 * identifies, so when it is set to less than the full node, we can
14514 * skip the rest of this */
14515 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
14516 PERL_UINT_FAST8_T backup_count = 0;
14518 const STRLEN full_len = len;
14520 assert(len >= MAX_NODE_STRING_SIZE);
14522 /* Here, <s> points to just beyond where we have output the
14523 * final character of the node. Look backwards through the
14524 * string until find a non- problematic character */
14528 /* This has no multi-char folds to non-UTF characters */
14529 if (ASCII_FOLD_RESTRICTED) {
14533 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) {
14540 /* Point to the first byte of the final character */
14541 s = (char *) utf8_hop_back((U8 *) s, -1, (U8 *) s0);
14543 while (s >= s0) { /* Search backwards until find
14544 a non-problematic char */
14545 if (UTF8_IS_INVARIANT(*s)) {
14547 /* There are no ascii characters that participate
14548 * in multi-char folds under /aa. In EBCDIC, the
14549 * non-ascii invariants are all control characters,
14550 * so don't ever participate in any folds. */
14551 if (ASCII_FOLD_RESTRICTED
14552 || ! IS_NON_FINAL_FOLD(*s))
14557 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
14558 if (! IS_NON_FINAL_FOLD(EIGHT_BIT_UTF8_TO_NATIVE(
14564 else if (! _invlist_contains_cp(
14566 valid_utf8_to_uvchr((U8 *) s, NULL)))
14571 /* Here, the current character is problematic in that
14572 * it does occur in the non-final position of some
14573 * fold, so try the character before it, but have to
14574 * special case the very first byte in the string, so
14575 * we don't read outside the string */
14576 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
14578 } /* End of loop backwards through the string */
14580 /* If there were only problematic characters in the string,
14581 * <s> will point to before s0, in which case the length
14582 * should be 0, otherwise include the length of the
14583 * non-problematic character just found */
14584 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
14587 /* Here, have found the final character, if any, that is
14588 * non-problematic as far as ending the node without splitting
14589 * it across a potential multi-char fold. <len> contains the
14590 * number of bytes in the node up-to and including that
14591 * character, or is 0 if there is no such character, meaning
14592 * the whole node contains only problematic characters. In
14593 * this case, give up and just take the node as-is. We can't
14600 /* Here, the node does contain some characters that aren't
14601 * problematic. If we didn't have to backup any, then the
14602 * final character in the node is non-problematic, and we
14603 * can take the node as-is */
14604 if (backup_count == 0) {
14607 else if (backup_count == 1) {
14609 /* If the final character is problematic, but the
14610 * penultimate is not, back-off that last character to
14611 * later start a new node with it */
14616 /* Here, the final non-problematic character is earlier
14617 * in the input than the penultimate character. What we do
14618 * is reparse from the beginning, going up only as far as
14619 * this final ok one, thus guaranteeing that the node ends
14620 * in an acceptable character. The reason we reparse is
14621 * that we know how far in the character is, but we don't
14622 * know how to correlate its position with the input parse.
14623 * An alternate implementation would be to build that
14624 * correlation as we go along during the original parse,
14625 * but that would entail extra work for every node, whereas
14626 * this code gets executed only when the string is too
14627 * large for the node, and the final two characters are
14628 * problematic, an infrequent occurrence. Yet another
14629 * possible strategy would be to save the tail of the
14630 * string, and the next time regatom is called, initialize
14631 * with that. The problem with this is that unless you
14632 * back off one more character, you won't be guaranteed
14633 * regatom will get called again, unless regbranch,
14634 * regpiece ... are also changed. If you do back off that
14635 * extra character, so that there is input guaranteed to
14636 * force calling regatom, you can't handle the case where
14637 * just the first character in the node is acceptable. I
14638 * (khw) decided to try this method which doesn't have that
14639 * pitfall; if performance issues are found, we can do a
14640 * combination of the current approach plus that one */
14646 } /* End of verifying node ends with an appropriate char */
14648 loopdone: /* Jumped to when encounters something that shouldn't be
14651 /* Free up any over-allocated space; cast is to silence bogus
14652 * warning in MS VC */
14653 change_engine_size(pRExC_state,
14654 - (Ptrdiff_t) (initial_size - STR_SZ(len)));
14656 /* I (khw) don't know if you can get here with zero length, but the
14657 * old code handled this situation by creating a zero-length EXACT
14658 * node. Might as well be NOTHING instead */
14660 OP(REGNODE_p(ret)) = NOTHING;
14664 /* If the node type is EXACT here, check to see if it
14665 * should be EXACTL, or EXACT_ONLY8. */
14666 if (node_type == EXACT) {
14668 node_type = EXACTL;
14670 else if (requires_utf8_target) {
14671 node_type = EXACT_ONLY8;
14674 if ( UNLIKELY(has_micro_sign || has_ss)
14675 && (node_type == EXACTFU || ( node_type == EXACTF
14676 && maybe_exactfu)))
14677 { /* These two conditions are problematic in non-UTF-8
14680 node_type = EXACTFUP;
14682 else if (node_type == EXACTFL) {
14684 /* 'maybe_exactfu' is deliberately set above to
14685 * indicate this node type, where all code points in it
14687 if (maybe_exactfu) {
14688 node_type = EXACTFLU8;
14691 else if (node_type == EXACTF) { /* Means is /di */
14693 /* If 'maybe_exactfu' is clear, then we need to stay
14694 * /di. If it is set, it means there are no code
14695 * points that match differently depending on UTF8ness
14696 * of the target string, so it can become an EXACTFU
14698 if (! maybe_exactfu) {
14699 RExC_seen_d_op = TRUE;
14701 else if ( isALPHA_FOLD_EQ(* STRING(REGNODE_p(ret)), 's')
14702 || isALPHA_FOLD_EQ(ender, 's'))
14704 /* But, if the node begins or ends in an 's' we
14705 * have to defer changing it into an EXACTFU, as
14706 * the node could later get joined with another one
14707 * that ends or begins with 's' creating an 'ss'
14708 * sequence which would then wrongly match the
14709 * sharp s without the target being UTF-8. We
14710 * create a special node that we resolve later when
14711 * we join nodes together */
14713 node_type = EXACTFU_S_EDGE;
14716 node_type = EXACTFU;
14720 if (requires_utf8_target && node_type == EXACTFU) {
14721 node_type = EXACTFU_ONLY8;
14725 OP(REGNODE_p(ret)) = node_type;
14726 STR_LEN(REGNODE_p(ret)) = len;
14727 RExC_emit += STR_SZ(len);
14729 /* If the node isn't a single character, it can't be SIMPLE */
14730 if (len > (Size_t) ((UTF) ? UVCHR_SKIP(ender) : 1)) {
14734 *flagp |= HASWIDTH | maybe_SIMPLE;
14737 Set_Node_Length(REGNODE_p(ret), p - parse_start - 1);
14741 /* len is STRLEN which is unsigned, need to copy to signed */
14744 vFAIL("Internal disaster");
14747 } /* End of label 'defchar:' */
14749 } /* End of giant switch on input character */
14751 /* Position parse to next real character */
14752 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14753 FALSE /* Don't force to /x */ );
14754 if ( *RExC_parse == '{'
14755 && OP(REGNODE_p(ret)) != SBOL && ! regcurly(RExC_parse))
14757 if (RExC_strict || new_regcurly(RExC_parse, RExC_end)) {
14759 vFAIL("Unescaped left brace in regex is illegal here");
14761 ckWARNreg(RExC_parse + 1, "Unescaped left brace in regex is"
14762 " passed through");
14770 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
14772 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
14773 * sets up the bitmap and any flags, removing those code points from the
14774 * inversion list, setting it to NULL should it become completely empty */
14778 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
14779 assert(PL_regkind[OP(node)] == ANYOF);
14781 /* There is no bitmap for this node type */
14782 if (OP(node) == ANYOFH) {
14786 ANYOF_BITMAP_ZERO(node);
14787 if (*invlist_ptr) {
14789 /* This gets set if we actually need to modify things */
14790 bool change_invlist = FALSE;
14794 /* Start looking through *invlist_ptr */
14795 invlist_iterinit(*invlist_ptr);
14796 while (invlist_iternext(*invlist_ptr, &start, &end)) {
14800 if (end == UV_MAX && start <= NUM_ANYOF_CODE_POINTS) {
14801 ANYOF_FLAGS(node) |= ANYOF_MATCHES_ALL_ABOVE_BITMAP;
14804 /* Quit if are above what we should change */
14805 if (start >= NUM_ANYOF_CODE_POINTS) {
14809 change_invlist = TRUE;
14811 /* Set all the bits in the range, up to the max that we are doing */
14812 high = (end < NUM_ANYOF_CODE_POINTS - 1)
14814 : NUM_ANYOF_CODE_POINTS - 1;
14815 for (i = start; i <= (int) high; i++) {
14816 if (! ANYOF_BITMAP_TEST(node, i)) {
14817 ANYOF_BITMAP_SET(node, i);
14821 invlist_iterfinish(*invlist_ptr);
14823 /* Done with loop; remove any code points that are in the bitmap from
14824 * *invlist_ptr; similarly for code points above the bitmap if we have
14825 * a flag to match all of them anyways */
14826 if (change_invlist) {
14827 _invlist_subtract(*invlist_ptr, PL_InBitmap, invlist_ptr);
14829 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
14830 _invlist_intersection(*invlist_ptr, PL_InBitmap, invlist_ptr);
14833 /* If have completely emptied it, remove it completely */
14834 if (_invlist_len(*invlist_ptr) == 0) {
14835 SvREFCNT_dec_NN(*invlist_ptr);
14836 *invlist_ptr = NULL;
14841 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
14842 Character classes ([:foo:]) can also be negated ([:^foo:]).
14843 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
14844 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
14845 but trigger failures because they are currently unimplemented. */
14847 #define POSIXCC_DONE(c) ((c) == ':')
14848 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
14849 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
14850 #define MAYBE_POSIXCC(c) (POSIXCC(c) || (c) == '^' || (c) == ';')
14852 #define WARNING_PREFIX "Assuming NOT a POSIX class since "
14853 #define NO_BLANKS_POSIX_WARNING "no blanks are allowed in one"
14854 #define SEMI_COLON_POSIX_WARNING "a semi-colon was found instead of a colon"
14856 #define NOT_MEANT_TO_BE_A_POSIX_CLASS (OOB_NAMEDCLASS - 1)
14858 /* 'posix_warnings' and 'warn_text' are names of variables in the following
14860 #define ADD_POSIX_WARNING(p, text) STMT_START { \
14861 if (posix_warnings) { \
14862 if (! RExC_warn_text ) RExC_warn_text = \
14863 (AV *) sv_2mortal((SV *) newAV()); \
14864 av_push(RExC_warn_text, Perl_newSVpvf(aTHX_ \
14868 REPORT_LOCATION_ARGS(p))); \
14871 #define CLEAR_POSIX_WARNINGS() \
14873 if (posix_warnings && RExC_warn_text) \
14874 av_clear(RExC_warn_text); \
14877 #define CLEAR_POSIX_WARNINGS_AND_RETURN(ret) \
14879 CLEAR_POSIX_WARNINGS(); \
14884 S_handle_possible_posix(pTHX_ RExC_state_t *pRExC_state,
14886 const char * const s, /* Where the putative posix class begins.
14887 Normally, this is one past the '['. This
14888 parameter exists so it can be somewhere
14889 besides RExC_parse. */
14890 char ** updated_parse_ptr, /* Where to set the updated parse pointer, or
14892 AV ** posix_warnings, /* Where to place any generated warnings, or
14894 const bool check_only /* Don't die if error */
14897 /* This parses what the caller thinks may be one of the three POSIX
14899 * 1) a character class, like [:blank:]
14900 * 2) a collating symbol, like [. .]
14901 * 3) an equivalence class, like [= =]
14902 * In the latter two cases, it croaks if it finds a syntactically legal
14903 * one, as these are not handled by Perl.
14905 * The main purpose is to look for a POSIX character class. It returns:
14906 * a) the class number
14907 * if it is a completely syntactically and semantically legal class.
14908 * 'updated_parse_ptr', if not NULL, is set to point to just after the
14909 * closing ']' of the class
14910 * b) OOB_NAMEDCLASS
14911 * if it appears that one of the three POSIX constructs was meant, but
14912 * its specification was somehow defective. 'updated_parse_ptr', if
14913 * not NULL, is set to point to the character just after the end
14914 * character of the class. See below for handling of warnings.
14915 * c) NOT_MEANT_TO_BE_A_POSIX_CLASS
14916 * if it doesn't appear that a POSIX construct was intended.
14917 * 'updated_parse_ptr' is not changed. No warnings nor errors are
14920 * In b) there may be errors or warnings generated. If 'check_only' is
14921 * TRUE, then any errors are discarded. Warnings are returned to the
14922 * caller via an AV* created into '*posix_warnings' if it is not NULL. If
14923 * instead it is NULL, warnings are suppressed.
14925 * The reason for this function, and its complexity is that a bracketed
14926 * character class can contain just about anything. But it's easy to
14927 * mistype the very specific posix class syntax but yielding a valid
14928 * regular bracketed class, so it silently gets compiled into something
14929 * quite unintended.
14931 * The solution adopted here maintains backward compatibility except that
14932 * it adds a warning if it looks like a posix class was intended but
14933 * improperly specified. The warning is not raised unless what is input
14934 * very closely resembles one of the 14 legal posix classes. To do this,
14935 * it uses fuzzy parsing. It calculates how many single-character edits it
14936 * would take to transform what was input into a legal posix class. Only
14937 * if that number is quite small does it think that the intention was a
14938 * posix class. Obviously these are heuristics, and there will be cases
14939 * where it errs on one side or another, and they can be tweaked as
14940 * experience informs.
14942 * The syntax for a legal posix class is:
14944 * qr/(?xa: \[ : \^? [[:lower:]]{4,6} : \] )/
14946 * What this routine considers syntactically to be an intended posix class
14947 * is this (the comments indicate some restrictions that the pattern
14950 * qr/(?x: \[? # The left bracket, possibly
14952 * \h* # possibly followed by blanks
14953 * (?: \^ \h* )? # possibly a misplaced caret
14954 * [:;]? # The opening class character,
14955 * # possibly omitted. A typo
14956 * # semi-colon can also be used.
14958 * \^? # possibly a correctly placed
14959 * # caret, but not if there was also
14960 * # a misplaced one
14962 * .{3,15} # The class name. If there are
14963 * # deviations from the legal syntax,
14964 * # its edit distance must be close
14965 * # to a real class name in order
14966 * # for it to be considered to be
14967 * # an intended posix class.
14969 * [[:punct:]]? # The closing class character,
14970 * # possibly omitted. If not a colon
14971 * # nor semi colon, the class name
14972 * # must be even closer to a valid
14975 * \]? # The right bracket, possibly
14979 * In the above, \h must be ASCII-only.
14981 * These are heuristics, and can be tweaked as field experience dictates.
14982 * There will be cases when someone didn't intend to specify a posix class
14983 * that this warns as being so. The goal is to minimize these, while
14984 * maximizing the catching of things intended to be a posix class that
14985 * aren't parsed as such.
14989 const char * const e = RExC_end;
14990 unsigned complement = 0; /* If to complement the class */
14991 bool found_problem = FALSE; /* Assume OK until proven otherwise */
14992 bool has_opening_bracket = FALSE;
14993 bool has_opening_colon = FALSE;
14994 int class_number = OOB_NAMEDCLASS; /* Out-of-bounds until find
14996 const char * possible_end = NULL; /* used for a 2nd parse pass */
14997 const char* name_start; /* ptr to class name first char */
14999 /* If the number of single-character typos the input name is away from a
15000 * legal name is no more than this number, it is considered to have meant
15001 * the legal name */
15002 int max_distance = 2;
15004 /* to store the name. The size determines the maximum length before we
15005 * decide that no posix class was intended. Should be at least
15006 * sizeof("alphanumeric") */
15008 STATIC_ASSERT_DECL(C_ARRAY_LENGTH(input_text) >= sizeof "alphanumeric");
15010 PERL_ARGS_ASSERT_HANDLE_POSSIBLE_POSIX;
15012 CLEAR_POSIX_WARNINGS();
15015 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
15018 if (*(p - 1) != '[') {
15019 ADD_POSIX_WARNING(p, "it doesn't start with a '['");
15020 found_problem = TRUE;
15023 has_opening_bracket = TRUE;
15026 /* They could be confused and think you can put spaces between the
15029 found_problem = TRUE;
15033 } while (p < e && isBLANK(*p));
15035 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15038 /* For [. .] and [= =]. These are quite different internally from [: :],
15039 * so they are handled separately. */
15040 if (POSIXCC_NOTYET(*p) && p < e - 3) /* 1 for the close, and 1 for the ']'
15041 and 1 for at least one char in it
15044 const char open_char = *p;
15045 const char * temp_ptr = p + 1;
15047 /* These two constructs are not handled by perl, and if we find a
15048 * syntactically valid one, we croak. khw, who wrote this code, finds
15049 * this explanation of them very unclear:
15050 * http://pubs.opengroup.org/onlinepubs/009696899/basedefs/xbd_chap09.html
15051 * And searching the rest of the internet wasn't very helpful either.
15052 * It looks like just about any byte can be in these constructs,
15053 * depending on the locale. But unless the pattern is being compiled
15054 * under /l, which is very rare, Perl runs under the C or POSIX locale.
15055 * In that case, it looks like [= =] isn't allowed at all, and that
15056 * [. .] could be any single code point, but for longer strings the
15057 * constituent characters would have to be the ASCII alphabetics plus
15058 * the minus-hyphen. Any sensible locale definition would limit itself
15059 * to these. And any portable one definitely should. Trying to parse
15060 * the general case is a nightmare (see [perl #127604]). So, this code
15061 * looks only for interiors of these constructs that match:
15063 * Using \w relaxes the apparent rules a little, without adding much
15064 * danger of mistaking something else for one of these constructs.
15066 * [. .] in some implementations described on the internet is usable to
15067 * escape a character that otherwise is special in bracketed character
15068 * classes. For example [.].] means a literal right bracket instead of
15069 * the ending of the class
15071 * [= =] can legitimately contain a [. .] construct, but we don't
15072 * handle this case, as that [. .] construct will later get parsed
15073 * itself and croak then. And [= =] is checked for even when not under
15074 * /l, as Perl has long done so.
15076 * The code below relies on there being a trailing NUL, so it doesn't
15077 * have to keep checking if the parse ptr < e.
15079 if (temp_ptr[1] == open_char) {
15082 else while ( temp_ptr < e
15083 && (isWORDCHAR(*temp_ptr) || *temp_ptr == '-'))
15088 if (*temp_ptr == open_char) {
15090 if (*temp_ptr == ']') {
15092 if (! found_problem && ! check_only) {
15093 RExC_parse = (char *) temp_ptr;
15094 vFAIL3("POSIX syntax [%c %c] is reserved for future "
15095 "extensions", open_char, open_char);
15098 /* Here, the syntax wasn't completely valid, or else the call
15099 * is to check-only */
15100 if (updated_parse_ptr) {
15101 *updated_parse_ptr = (char *) temp_ptr;
15104 CLEAR_POSIX_WARNINGS_AND_RETURN(OOB_NAMEDCLASS);
15108 /* If we find something that started out to look like one of these
15109 * constructs, but isn't, we continue below so that it can be checked
15110 * for being a class name with a typo of '.' or '=' instead of a colon.
15114 /* Here, we think there is a possibility that a [: :] class was meant, and
15115 * we have the first real character. It could be they think the '^' comes
15118 found_problem = TRUE;
15119 ADD_POSIX_WARNING(p + 1, "the '^' must come after the colon");
15124 found_problem = TRUE;
15128 } while (p < e && isBLANK(*p));
15130 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15134 /* But the first character should be a colon, which they could have easily
15135 * mistyped on a qwerty keyboard as a semi-colon (and which may be hard to
15136 * distinguish from a colon, so treat that as a colon). */
15139 has_opening_colon = TRUE;
15141 else if (*p == ';') {
15142 found_problem = TRUE;
15144 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
15145 has_opening_colon = TRUE;
15148 found_problem = TRUE;
15149 ADD_POSIX_WARNING(p, "there must be a starting ':'");
15151 /* Consider an initial punctuation (not one of the recognized ones) to
15152 * be a left terminator */
15153 if (*p != '^' && *p != ']' && isPUNCT(*p)) {
15158 /* They may think that you can put spaces between the components */
15160 found_problem = TRUE;
15164 } while (p < e && isBLANK(*p));
15166 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15171 /* We consider something like [^:^alnum:]] to not have been intended to
15172 * be a posix class, but XXX maybe we should */
15174 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15181 /* Again, they may think that you can put spaces between the components */
15183 found_problem = TRUE;
15187 } while (p < e && isBLANK(*p));
15189 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15194 /* XXX This ']' may be a typo, and something else was meant. But
15195 * treating it as such creates enough complications, that that
15196 * possibility isn't currently considered here. So we assume that the
15197 * ']' is what is intended, and if we've already found an initial '[',
15198 * this leaves this construct looking like [:] or [:^], which almost
15199 * certainly weren't intended to be posix classes */
15200 if (has_opening_bracket) {
15201 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15204 /* But this function can be called when we parse the colon for
15205 * something like qr/[alpha:]]/, so we back up to look for the
15210 found_problem = TRUE;
15211 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
15213 else if (*p != ':') {
15215 /* XXX We are currently very restrictive here, so this code doesn't
15216 * consider the possibility that, say, /[alpha.]]/ was intended to
15217 * be a posix class. */
15218 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15221 /* Here we have something like 'foo:]'. There was no initial colon,
15222 * and we back up over 'foo. XXX Unlike the going forward case, we
15223 * don't handle typos of non-word chars in the middle */
15224 has_opening_colon = FALSE;
15227 while (p > RExC_start && isWORDCHAR(*p)) {
15232 /* Here, we have positioned ourselves to where we think the first
15233 * character in the potential class is */
15236 /* Now the interior really starts. There are certain key characters that
15237 * can end the interior, or these could just be typos. To catch both
15238 * cases, we may have to do two passes. In the first pass, we keep on
15239 * going unless we come to a sequence that matches
15240 * qr/ [[:punct:]] [[:blank:]]* \] /xa
15241 * This means it takes a sequence to end the pass, so two typos in a row if
15242 * that wasn't what was intended. If the class is perfectly formed, just
15243 * this one pass is needed. We also stop if there are too many characters
15244 * being accumulated, but this number is deliberately set higher than any
15245 * real class. It is set high enough so that someone who thinks that
15246 * 'alphanumeric' is a correct name would get warned that it wasn't.
15247 * While doing the pass, we keep track of where the key characters were in
15248 * it. If we don't find an end to the class, and one of the key characters
15249 * was found, we redo the pass, but stop when we get to that character.
15250 * Thus the key character was considered a typo in the first pass, but a
15251 * terminator in the second. If two key characters are found, we stop at
15252 * the second one in the first pass. Again this can miss two typos, but
15253 * catches a single one
15255 * In the first pass, 'possible_end' starts as NULL, and then gets set to
15256 * point to the first key character. For the second pass, it starts as -1.
15262 bool has_blank = FALSE;
15263 bool has_upper = FALSE;
15264 bool has_terminating_colon = FALSE;
15265 bool has_terminating_bracket = FALSE;
15266 bool has_semi_colon = FALSE;
15267 unsigned int name_len = 0;
15268 int punct_count = 0;
15272 /* Squeeze out blanks when looking up the class name below */
15273 if (isBLANK(*p) ) {
15275 found_problem = TRUE;
15280 /* The name will end with a punctuation */
15282 const char * peek = p + 1;
15284 /* Treat any non-']' punctuation followed by a ']' (possibly
15285 * with intervening blanks) as trying to terminate the class.
15286 * ']]' is very likely to mean a class was intended (but
15287 * missing the colon), but the warning message that gets
15288 * generated shows the error position better if we exit the
15289 * loop at the bottom (eventually), so skip it here. */
15291 if (peek < e && isBLANK(*peek)) {
15293 found_problem = TRUE;
15296 } while (peek < e && isBLANK(*peek));
15299 if (peek < e && *peek == ']') {
15300 has_terminating_bracket = TRUE;
15302 has_terminating_colon = TRUE;
15304 else if (*p == ';') {
15305 has_semi_colon = TRUE;
15306 has_terminating_colon = TRUE;
15309 found_problem = TRUE;
15316 /* Here we have punctuation we thought didn't end the class.
15317 * Keep track of the position of the key characters that are
15318 * more likely to have been class-enders */
15319 if (*p == ']' || *p == '[' || *p == ':' || *p == ';') {
15321 /* Allow just one such possible class-ender not actually
15322 * ending the class. */
15323 if (possible_end) {
15329 /* If we have too many punctuation characters, no use in
15331 if (++punct_count > max_distance) {
15335 /* Treat the punctuation as a typo. */
15336 input_text[name_len++] = *p;
15339 else if (isUPPER(*p)) { /* Use lowercase for lookup */
15340 input_text[name_len++] = toLOWER(*p);
15342 found_problem = TRUE;
15344 } else if (! UTF || UTF8_IS_INVARIANT(*p)) {
15345 input_text[name_len++] = *p;
15349 input_text[name_len++] = utf8_to_uvchr_buf((U8 *) p, e, NULL);
15353 /* The declaration of 'input_text' is how long we allow a potential
15354 * class name to be, before saying they didn't mean a class name at
15356 if (name_len >= C_ARRAY_LENGTH(input_text)) {
15361 /* We get to here when the possible class name hasn't been properly
15362 * terminated before:
15363 * 1) we ran off the end of the pattern; or
15364 * 2) found two characters, each of which might have been intended to
15365 * be the name's terminator
15366 * 3) found so many punctuation characters in the purported name,
15367 * that the edit distance to a valid one is exceeded
15368 * 4) we decided it was more characters than anyone could have
15369 * intended to be one. */
15371 found_problem = TRUE;
15373 /* In the final two cases, we know that looking up what we've
15374 * accumulated won't lead to a match, even a fuzzy one. */
15375 if ( name_len >= C_ARRAY_LENGTH(input_text)
15376 || punct_count > max_distance)
15378 /* If there was an intermediate key character that could have been
15379 * an intended end, redo the parse, but stop there */
15380 if (possible_end && possible_end != (char *) -1) {
15381 possible_end = (char *) -1; /* Special signal value to say
15382 we've done a first pass */
15387 /* Otherwise, it can't have meant to have been a class */
15388 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15391 /* If we ran off the end, and the final character was a punctuation
15392 * one, back up one, to look at that final one just below. Later, we
15393 * will restore the parse pointer if appropriate */
15394 if (name_len && p == e && isPUNCT(*(p-1))) {
15399 if (p < e && isPUNCT(*p)) {
15401 has_terminating_bracket = TRUE;
15403 /* If this is a 2nd ']', and the first one is just below this
15404 * one, consider that to be the real terminator. This gives a
15405 * uniform and better positioning for the warning message */
15407 && possible_end != (char *) -1
15408 && *possible_end == ']'
15409 && name_len && input_text[name_len - 1] == ']')
15414 /* And this is actually equivalent to having done the 2nd
15415 * pass now, so set it to not try again */
15416 possible_end = (char *) -1;
15421 has_terminating_colon = TRUE;
15423 else if (*p == ';') {
15424 has_semi_colon = TRUE;
15425 has_terminating_colon = TRUE;
15433 /* Here, we have a class name to look up. We can short circuit the
15434 * stuff below for short names that can't possibly be meant to be a
15435 * class name. (We can do this on the first pass, as any second pass
15436 * will yield an even shorter name) */
15437 if (name_len < 3) {
15438 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15441 /* Find which class it is. Initially switch on the length of the name.
15443 switch (name_len) {
15445 if (memEQs(name_start, 4, "word")) {
15446 /* this is not POSIX, this is the Perl \w */
15447 class_number = ANYOF_WORDCHAR;
15451 /* Names all of length 5: alnum alpha ascii blank cntrl digit
15452 * graph lower print punct space upper
15453 * Offset 4 gives the best switch position. */
15454 switch (name_start[4]) {
15456 if (memBEGINs(name_start, 5, "alph")) /* alpha */
15457 class_number = ANYOF_ALPHA;
15460 if (memBEGINs(name_start, 5, "spac")) /* space */
15461 class_number = ANYOF_SPACE;
15464 if (memBEGINs(name_start, 5, "grap")) /* graph */
15465 class_number = ANYOF_GRAPH;
15468 if (memBEGINs(name_start, 5, "asci")) /* ascii */
15469 class_number = ANYOF_ASCII;
15472 if (memBEGINs(name_start, 5, "blan")) /* blank */
15473 class_number = ANYOF_BLANK;
15476 if (memBEGINs(name_start, 5, "cntr")) /* cntrl */
15477 class_number = ANYOF_CNTRL;
15480 if (memBEGINs(name_start, 5, "alnu")) /* alnum */
15481 class_number = ANYOF_ALPHANUMERIC;
15484 if (memBEGINs(name_start, 5, "lowe")) /* lower */
15485 class_number = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
15486 else if (memBEGINs(name_start, 5, "uppe")) /* upper */
15487 class_number = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
15490 if (memBEGINs(name_start, 5, "digi")) /* digit */
15491 class_number = ANYOF_DIGIT;
15492 else if (memBEGINs(name_start, 5, "prin")) /* print */
15493 class_number = ANYOF_PRINT;
15494 else if (memBEGINs(name_start, 5, "punc")) /* punct */
15495 class_number = ANYOF_PUNCT;
15500 if (memEQs(name_start, 6, "xdigit"))
15501 class_number = ANYOF_XDIGIT;
15505 /* If the name exactly matches a posix class name the class number will
15506 * here be set to it, and the input almost certainly was meant to be a
15507 * posix class, so we can skip further checking. If instead the syntax
15508 * is exactly correct, but the name isn't one of the legal ones, we
15509 * will return that as an error below. But if neither of these apply,
15510 * it could be that no posix class was intended at all, or that one
15511 * was, but there was a typo. We tease these apart by doing fuzzy
15512 * matching on the name */
15513 if (class_number == OOB_NAMEDCLASS && found_problem) {
15514 const UV posix_names[][6] = {
15515 { 'a', 'l', 'n', 'u', 'm' },
15516 { 'a', 'l', 'p', 'h', 'a' },
15517 { 'a', 's', 'c', 'i', 'i' },
15518 { 'b', 'l', 'a', 'n', 'k' },
15519 { 'c', 'n', 't', 'r', 'l' },
15520 { 'd', 'i', 'g', 'i', 't' },
15521 { 'g', 'r', 'a', 'p', 'h' },
15522 { 'l', 'o', 'w', 'e', 'r' },
15523 { 'p', 'r', 'i', 'n', 't' },
15524 { 'p', 'u', 'n', 'c', 't' },
15525 { 's', 'p', 'a', 'c', 'e' },
15526 { 'u', 'p', 'p', 'e', 'r' },
15527 { 'w', 'o', 'r', 'd' },
15528 { 'x', 'd', 'i', 'g', 'i', 't' }
15530 /* The names of the above all have added NULs to make them the same
15531 * size, so we need to also have the real lengths */
15532 const UV posix_name_lengths[] = {
15533 sizeof("alnum") - 1,
15534 sizeof("alpha") - 1,
15535 sizeof("ascii") - 1,
15536 sizeof("blank") - 1,
15537 sizeof("cntrl") - 1,
15538 sizeof("digit") - 1,
15539 sizeof("graph") - 1,
15540 sizeof("lower") - 1,
15541 sizeof("print") - 1,
15542 sizeof("punct") - 1,
15543 sizeof("space") - 1,
15544 sizeof("upper") - 1,
15545 sizeof("word") - 1,
15546 sizeof("xdigit")- 1
15549 int temp_max = max_distance; /* Use a temporary, so if we
15550 reparse, we haven't changed the
15553 /* Use a smaller max edit distance if we are missing one of the
15555 if ( has_opening_bracket + has_opening_colon < 2
15556 || has_terminating_bracket + has_terminating_colon < 2)
15561 /* See if the input name is close to a legal one */
15562 for (i = 0; i < C_ARRAY_LENGTH(posix_names); i++) {
15564 /* Short circuit call if the lengths are too far apart to be
15566 if (abs( (int) (name_len - posix_name_lengths[i]))
15572 if (edit_distance(input_text,
15575 posix_name_lengths[i],
15579 { /* If it is close, it probably was intended to be a class */
15580 goto probably_meant_to_be;
15584 /* Here the input name is not close enough to a valid class name
15585 * for us to consider it to be intended to be a posix class. If
15586 * we haven't already done so, and the parse found a character that
15587 * could have been terminators for the name, but which we absorbed
15588 * as typos during the first pass, repeat the parse, signalling it
15589 * to stop at that character */
15590 if (possible_end && possible_end != (char *) -1) {
15591 possible_end = (char *) -1;
15596 /* Here neither pass found a close-enough class name */
15597 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15600 probably_meant_to_be:
15602 /* Here we think that a posix specification was intended. Update any
15604 if (updated_parse_ptr) {
15605 *updated_parse_ptr = (char *) p;
15608 /* If a posix class name was intended but incorrectly specified, we
15609 * output or return the warnings */
15610 if (found_problem) {
15612 /* We set flags for these issues in the parse loop above instead of
15613 * adding them to the list of warnings, because we can parse it
15614 * twice, and we only want one warning instance */
15616 ADD_POSIX_WARNING(p, "the name must be all lowercase letters");
15619 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15621 if (has_semi_colon) {
15622 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
15624 else if (! has_terminating_colon) {
15625 ADD_POSIX_WARNING(p, "there is no terminating ':'");
15627 if (! has_terminating_bracket) {
15628 ADD_POSIX_WARNING(p, "there is no terminating ']'");
15631 if ( posix_warnings
15633 && av_top_index(RExC_warn_text) > -1)
15635 *posix_warnings = RExC_warn_text;
15638 else if (class_number != OOB_NAMEDCLASS) {
15639 /* If it is a known class, return the class. The class number
15640 * #defines are structured so each complement is +1 to the normal
15642 CLEAR_POSIX_WARNINGS_AND_RETURN(class_number + complement);
15644 else if (! check_only) {
15646 /* Here, it is an unrecognized class. This is an error (unless the
15647 * call is to check only, which we've already handled above) */
15648 const char * const complement_string = (complement)
15651 RExC_parse = (char *) p;
15652 vFAIL3utf8f("POSIX class [:%s%" UTF8f ":] unknown",
15654 UTF8fARG(UTF, RExC_parse - name_start - 2, name_start));
15658 return OOB_NAMEDCLASS;
15660 #undef ADD_POSIX_WARNING
15662 STATIC unsigned int
15663 S_regex_set_precedence(const U8 my_operator) {
15665 /* Returns the precedence in the (?[...]) construct of the input operator,
15666 * specified by its character representation. The precedence follows
15667 * general Perl rules, but it extends this so that ')' and ']' have (low)
15668 * precedence even though they aren't really operators */
15670 switch (my_operator) {
15686 NOT_REACHED; /* NOTREACHED */
15687 return 0; /* Silence compiler warning */
15690 STATIC regnode_offset
15691 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
15692 I32 *flagp, U32 depth,
15693 char * const oregcomp_parse)
15695 /* Handle the (?[...]) construct to do set operations */
15697 U8 curchar; /* Current character being parsed */
15698 UV start, end; /* End points of code point ranges */
15699 SV* final = NULL; /* The end result inversion list */
15700 SV* result_string; /* 'final' stringified */
15701 AV* stack; /* stack of operators and operands not yet
15703 AV* fence_stack = NULL; /* A stack containing the positions in
15704 'stack' of where the undealt-with left
15705 parens would be if they were actually
15707 /* The 'volatile' is a workaround for an optimiser bug
15708 * in Solaris Studio 12.3. See RT #127455 */
15709 volatile IV fence = 0; /* Position of where most recent undealt-
15710 with left paren in stack is; -1 if none.
15712 STRLEN len; /* Temporary */
15713 regnode_offset node; /* Temporary, and final regnode returned by
15715 const bool save_fold = FOLD; /* Temporary */
15716 char *save_end, *save_parse; /* Temporaries */
15717 const bool in_locale = LOC; /* we turn off /l during processing */
15719 GET_RE_DEBUG_FLAGS_DECL;
15721 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
15723 DEBUG_PARSE("xcls");
15726 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
15729 /* The use of this operator implies /u. This is required so that the
15730 * compile time values are valid in all runtime cases */
15731 REQUIRE_UNI_RULES(flagp, 0);
15733 ckWARNexperimental(RExC_parse,
15734 WARN_EXPERIMENTAL__REGEX_SETS,
15735 "The regex_sets feature is experimental");
15737 /* Everything in this construct is a metacharacter. Operands begin with
15738 * either a '\' (for an escape sequence), or a '[' for a bracketed
15739 * character class. Any other character should be an operator, or
15740 * parenthesis for grouping. Both types of operands are handled by calling
15741 * regclass() to parse them. It is called with a parameter to indicate to
15742 * return the computed inversion list. The parsing here is implemented via
15743 * a stack. Each entry on the stack is a single character representing one
15744 * of the operators; or else a pointer to an operand inversion list. */
15746 #define IS_OPERATOR(a) SvIOK(a)
15747 #define IS_OPERAND(a) (! IS_OPERATOR(a))
15749 /* The stack is kept in Łukasiewicz order. (That's pronounced similar
15750 * to luke-a-shave-itch (or -itz), but people who didn't want to bother
15751 * with pronouncing it called it Reverse Polish instead, but now that YOU
15752 * know how to pronounce it you can use the correct term, thus giving due
15753 * credit to the person who invented it, and impressing your geek friends.
15754 * Wikipedia says that the pronounciation of "Ł" has been changing so that
15755 * it is now more like an English initial W (as in wonk) than an L.)
15757 * This means that, for example, 'a | b & c' is stored on the stack as
15765 * where the numbers in brackets give the stack [array] element number.
15766 * In this implementation, parentheses are not stored on the stack.
15767 * Instead a '(' creates a "fence" so that the part of the stack below the
15768 * fence is invisible except to the corresponding ')' (this allows us to
15769 * replace testing for parens, by using instead subtraction of the fence
15770 * position). As new operands are processed they are pushed onto the stack
15771 * (except as noted in the next paragraph). New operators of higher
15772 * precedence than the current final one are inserted on the stack before
15773 * the lhs operand (so that when the rhs is pushed next, everything will be
15774 * in the correct positions shown above. When an operator of equal or
15775 * lower precedence is encountered in parsing, all the stacked operations
15776 * of equal or higher precedence are evaluated, leaving the result as the
15777 * top entry on the stack. This makes higher precedence operations
15778 * evaluate before lower precedence ones, and causes operations of equal
15779 * precedence to left associate.
15781 * The only unary operator '!' is immediately pushed onto the stack when
15782 * encountered. When an operand is encountered, if the top of the stack is
15783 * a '!", the complement is immediately performed, and the '!' popped. The
15784 * resulting value is treated as a new operand, and the logic in the
15785 * previous paragraph is executed. Thus in the expression
15787 * the stack looks like
15793 * as 'b' gets parsed, the latter gets evaluated to '!b', and the stack
15800 * A ')' is treated as an operator with lower precedence than all the
15801 * aforementioned ones, which causes all operations on the stack above the
15802 * corresponding '(' to be evaluated down to a single resultant operand.
15803 * Then the fence for the '(' is removed, and the operand goes through the
15804 * algorithm above, without the fence.
15806 * A separate stack is kept of the fence positions, so that the position of
15807 * the latest so-far unbalanced '(' is at the top of it.
15809 * The ']' ending the construct is treated as the lowest operator of all,
15810 * so that everything gets evaluated down to a single operand, which is the
15813 sv_2mortal((SV *)(stack = newAV()));
15814 sv_2mortal((SV *)(fence_stack = newAV()));
15816 while (RExC_parse < RExC_end) {
15817 I32 top_index; /* Index of top-most element in 'stack' */
15818 SV** top_ptr; /* Pointer to top 'stack' element */
15819 SV* current = NULL; /* To contain the current inversion list
15821 SV* only_to_avoid_leaks;
15823 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
15824 TRUE /* Force /x */ );
15825 if (RExC_parse >= RExC_end) { /* Fail */
15829 curchar = UCHARAT(RExC_parse);
15833 #ifdef ENABLE_REGEX_SETS_DEBUGGING
15834 /* Enable with -Accflags=-DENABLE_REGEX_SETS_DEBUGGING */
15835 DEBUG_U(dump_regex_sets_structures(pRExC_state,
15836 stack, fence, fence_stack));
15839 top_index = av_tindex_skip_len_mg(stack);
15842 SV** stacked_ptr; /* Ptr to something already on 'stack' */
15843 char stacked_operator; /* The topmost operator on the 'stack'. */
15844 SV* lhs; /* Operand to the left of the operator */
15845 SV* rhs; /* Operand to the right of the operator */
15846 SV* fence_ptr; /* Pointer to top element of the fence
15851 if ( RExC_parse < RExC_end - 2
15852 && UCHARAT(RExC_parse + 1) == '?'
15853 && UCHARAT(RExC_parse + 2) == '^')
15855 /* If is a '(?', could be an embedded '(?^flags:(?[...])'.
15856 * This happens when we have some thing like
15858 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
15860 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
15862 * Here we would be handling the interpolated
15863 * '$thai_or_lao'. We handle this by a recursive call to
15864 * ourselves which returns the inversion list the
15865 * interpolated expression evaluates to. We use the flags
15866 * from the interpolated pattern. */
15867 U32 save_flags = RExC_flags;
15868 const char * save_parse;
15870 RExC_parse += 2; /* Skip past the '(?' */
15871 save_parse = RExC_parse;
15873 /* Parse the flags for the '(?'. We already know the first
15874 * flag to parse is a '^' */
15875 parse_lparen_question_flags(pRExC_state);
15877 if ( RExC_parse >= RExC_end - 4
15878 || UCHARAT(RExC_parse) != ':'
15879 || UCHARAT(++RExC_parse) != '('
15880 || UCHARAT(++RExC_parse) != '?'
15881 || UCHARAT(++RExC_parse) != '[')
15884 /* In combination with the above, this moves the
15885 * pointer to the point just after the first erroneous
15887 if (RExC_parse >= RExC_end - 4) {
15888 RExC_parse = RExC_end;
15890 else if (RExC_parse != save_parse) {
15891 RExC_parse += (UTF)
15892 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
15895 vFAIL("Expecting '(?flags:(?[...'");
15898 /* Recurse, with the meat of the embedded expression */
15900 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
15901 depth+1, oregcomp_parse);
15903 /* Here, 'current' contains the embedded expression's
15904 * inversion list, and RExC_parse points to the trailing
15905 * ']'; the next character should be the ')' */
15907 if (UCHARAT(RExC_parse) != ')')
15908 vFAIL("Expecting close paren for nested extended charclass");
15910 /* Then the ')' matching the original '(' handled by this
15911 * case: statement */
15913 if (UCHARAT(RExC_parse) != ')')
15914 vFAIL("Expecting close paren for wrapper for nested extended charclass");
15916 RExC_flags = save_flags;
15917 goto handle_operand;
15920 /* A regular '('. Look behind for illegal syntax */
15921 if (top_index - fence >= 0) {
15922 /* If the top entry on the stack is an operator, it had
15923 * better be a '!', otherwise the entry below the top
15924 * operand should be an operator */
15925 if ( ! (top_ptr = av_fetch(stack, top_index, FALSE))
15926 || (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) != '!')
15927 || ( IS_OPERAND(*top_ptr)
15928 && ( top_index - fence < 1
15929 || ! (stacked_ptr = av_fetch(stack,
15932 || ! IS_OPERATOR(*stacked_ptr))))
15935 vFAIL("Unexpected '(' with no preceding operator");
15939 /* Stack the position of this undealt-with left paren */
15940 av_push(fence_stack, newSViv(fence));
15941 fence = top_index + 1;
15945 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
15946 * multi-char folds are allowed. */
15947 if (!regclass(pRExC_state, flagp, depth+1,
15948 TRUE, /* means parse just the next thing */
15949 FALSE, /* don't allow multi-char folds */
15950 FALSE, /* don't silence non-portable warnings. */
15952 FALSE, /* Require return to be an ANYOF */
15955 goto regclass_failed;
15958 /* regclass() will return with parsing just the \ sequence,
15959 * leaving the parse pointer at the next thing to parse */
15961 goto handle_operand;
15963 case '[': /* Is a bracketed character class */
15965 /* See if this is a [:posix:] class. */
15966 bool is_posix_class = (OOB_NAMEDCLASS
15967 < handle_possible_posix(pRExC_state,
15971 TRUE /* checking only */));
15972 /* If it is a posix class, leave the parse pointer at the '['
15973 * to fool regclass() into thinking it is part of a
15974 * '[[:posix:]]'. */
15975 if (! is_posix_class) {
15979 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
15980 * multi-char folds are allowed. */
15981 if (!regclass(pRExC_state, flagp, depth+1,
15982 is_posix_class, /* parse the whole char
15983 class only if not a
15985 FALSE, /* don't allow multi-char folds */
15986 TRUE, /* silence non-portable warnings. */
15988 FALSE, /* Require return to be an ANYOF */
15991 goto regclass_failed;
15998 /* function call leaves parse pointing to the ']', except if we
16000 if (is_posix_class) {
16004 goto handle_operand;
16008 if (top_index >= 1) {
16009 goto join_operators;
16012 /* Only a single operand on the stack: are done */
16016 if (av_tindex_skip_len_mg(fence_stack) < 0) {
16017 if (UCHARAT(RExC_parse - 1) == ']') {
16021 vFAIL("Unexpected ')'");
16024 /* If nothing after the fence, is missing an operand */
16025 if (top_index - fence < 0) {
16029 /* If at least two things on the stack, treat this as an
16031 if (top_index - fence >= 1) {
16032 goto join_operators;
16035 /* Here only a single thing on the fenced stack, and there is a
16036 * fence. Get rid of it */
16037 fence_ptr = av_pop(fence_stack);
16039 fence = SvIV(fence_ptr);
16040 SvREFCNT_dec_NN(fence_ptr);
16047 /* Having gotten rid of the fence, we pop the operand at the
16048 * stack top and process it as a newly encountered operand */
16049 current = av_pop(stack);
16050 if (IS_OPERAND(current)) {
16051 goto handle_operand;
16063 /* These binary operators should have a left operand already
16065 if ( top_index - fence < 0
16066 || top_index - fence == 1
16067 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
16068 || ! IS_OPERAND(*top_ptr))
16070 goto unexpected_binary;
16073 /* If only the one operand is on the part of the stack visible
16074 * to us, we just place this operator in the proper position */
16075 if (top_index - fence < 2) {
16077 /* Place the operator before the operand */
16079 SV* lhs = av_pop(stack);
16080 av_push(stack, newSVuv(curchar));
16081 av_push(stack, lhs);
16085 /* But if there is something else on the stack, we need to
16086 * process it before this new operator if and only if the
16087 * stacked operation has equal or higher precedence than the
16092 /* The operator on the stack is supposed to be below both its
16094 if ( ! (stacked_ptr = av_fetch(stack, top_index - 2, FALSE))
16095 || IS_OPERAND(*stacked_ptr))
16097 /* But if not, it's legal and indicates we are completely
16098 * done if and only if we're currently processing a ']',
16099 * which should be the final thing in the expression */
16100 if (curchar == ']') {
16106 vFAIL2("Unexpected binary operator '%c' with no "
16107 "preceding operand", curchar);
16109 stacked_operator = (char) SvUV(*stacked_ptr);
16111 if (regex_set_precedence(curchar)
16112 > regex_set_precedence(stacked_operator))
16114 /* Here, the new operator has higher precedence than the
16115 * stacked one. This means we need to add the new one to
16116 * the stack to await its rhs operand (and maybe more
16117 * stuff). We put it before the lhs operand, leaving
16118 * untouched the stacked operator and everything below it
16120 lhs = av_pop(stack);
16121 assert(IS_OPERAND(lhs));
16123 av_push(stack, newSVuv(curchar));
16124 av_push(stack, lhs);
16128 /* Here, the new operator has equal or lower precedence than
16129 * what's already there. This means the operation already
16130 * there should be performed now, before the new one. */
16132 rhs = av_pop(stack);
16133 if (! IS_OPERAND(rhs)) {
16135 /* This can happen when a ! is not followed by an operand,
16136 * like in /(?[\t &!])/ */
16140 lhs = av_pop(stack);
16142 if (! IS_OPERAND(lhs)) {
16144 /* This can happen when there is an empty (), like in
16145 * /(?[[0]+()+])/ */
16149 switch (stacked_operator) {
16151 _invlist_intersection(lhs, rhs, &rhs);
16156 _invlist_union(lhs, rhs, &rhs);
16160 _invlist_subtract(lhs, rhs, &rhs);
16163 case '^': /* The union minus the intersection */
16168 _invlist_union(lhs, rhs, &u);
16169 _invlist_intersection(lhs, rhs, &i);
16170 _invlist_subtract(u, i, &rhs);
16171 SvREFCNT_dec_NN(i);
16172 SvREFCNT_dec_NN(u);
16178 /* Here, the higher precedence operation has been done, and the
16179 * result is in 'rhs'. We overwrite the stacked operator with
16180 * the result. Then we redo this code to either push the new
16181 * operator onto the stack or perform any higher precedence
16182 * stacked operation */
16183 only_to_avoid_leaks = av_pop(stack);
16184 SvREFCNT_dec(only_to_avoid_leaks);
16185 av_push(stack, rhs);
16188 case '!': /* Highest priority, right associative */
16190 /* If what's already at the top of the stack is another '!",
16191 * they just cancel each other out */
16192 if ( (top_ptr = av_fetch(stack, top_index, FALSE))
16193 && (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) == '!'))
16195 only_to_avoid_leaks = av_pop(stack);
16196 SvREFCNT_dec(only_to_avoid_leaks);
16198 else { /* Otherwise, since it's right associative, just push
16200 av_push(stack, newSVuv(curchar));
16205 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16206 if (RExC_parse >= RExC_end) {
16209 vFAIL("Unexpected character");
16213 /* Here 'current' is the operand. If something is already on the
16214 * stack, we have to check if it is a !. But first, the code above
16215 * may have altered the stack in the time since we earlier set
16218 top_index = av_tindex_skip_len_mg(stack);
16219 if (top_index - fence >= 0) {
16220 /* If the top entry on the stack is an operator, it had better
16221 * be a '!', otherwise the entry below the top operand should
16222 * be an operator */
16223 top_ptr = av_fetch(stack, top_index, FALSE);
16225 if (IS_OPERATOR(*top_ptr)) {
16227 /* The only permissible operator at the top of the stack is
16228 * '!', which is applied immediately to this operand. */
16229 curchar = (char) SvUV(*top_ptr);
16230 if (curchar != '!') {
16231 SvREFCNT_dec(current);
16232 vFAIL2("Unexpected binary operator '%c' with no "
16233 "preceding operand", curchar);
16236 _invlist_invert(current);
16238 only_to_avoid_leaks = av_pop(stack);
16239 SvREFCNT_dec(only_to_avoid_leaks);
16241 /* And we redo with the inverted operand. This allows
16242 * handling multiple ! in a row */
16243 goto handle_operand;
16245 /* Single operand is ok only for the non-binary ')'
16247 else if ((top_index - fence == 0 && curchar != ')')
16248 || (top_index - fence > 0
16249 && (! (stacked_ptr = av_fetch(stack,
16252 || IS_OPERAND(*stacked_ptr))))
16254 SvREFCNT_dec(current);
16255 vFAIL("Operand with no preceding operator");
16259 /* Here there was nothing on the stack or the top element was
16260 * another operand. Just add this new one */
16261 av_push(stack, current);
16263 } /* End of switch on next parse token */
16265 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16266 } /* End of loop parsing through the construct */
16268 vFAIL("Syntax error in (?[...])");
16272 if (RExC_parse >= RExC_end || RExC_parse[1] != ')') {
16273 if (RExC_parse < RExC_end) {
16277 vFAIL("Unexpected ']' with no following ')' in (?[...");
16280 if (av_tindex_skip_len_mg(fence_stack) >= 0) {
16281 vFAIL("Unmatched (");
16284 if (av_tindex_skip_len_mg(stack) < 0 /* Was empty */
16285 || ((final = av_pop(stack)) == NULL)
16286 || ! IS_OPERAND(final)
16287 || ! is_invlist(final)
16288 || av_tindex_skip_len_mg(stack) >= 0) /* More left on stack */
16291 SvREFCNT_dec(final);
16292 vFAIL("Incomplete expression within '(?[ ])'");
16295 /* Here, 'final' is the resultant inversion list from evaluating the
16296 * expression. Return it if so requested */
16297 if (return_invlist) {
16298 *return_invlist = final;
16302 /* Otherwise generate a resultant node, based on 'final'. regclass() is
16303 * expecting a string of ranges and individual code points */
16304 invlist_iterinit(final);
16305 result_string = newSVpvs("");
16306 while (invlist_iternext(final, &start, &end)) {
16307 if (start == end) {
16308 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}", start);
16311 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}-\\x{%" UVXf "}",
16316 /* About to generate an ANYOF (or similar) node from the inversion list we
16317 * have calculated */
16318 save_parse = RExC_parse;
16319 RExC_parse = SvPV(result_string, len);
16320 save_end = RExC_end;
16321 RExC_end = RExC_parse + len;
16322 TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE;
16324 /* We turn off folding around the call, as the class we have constructed
16325 * already has all folding taken into consideration, and we don't want
16326 * regclass() to add to that */
16327 RExC_flags &= ~RXf_PMf_FOLD;
16328 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if multi-char
16329 * folds are allowed. */
16330 node = regclass(pRExC_state, flagp, depth+1,
16331 FALSE, /* means parse the whole char class */
16332 FALSE, /* don't allow multi-char folds */
16333 TRUE, /* silence non-portable warnings. The above may very
16334 well have generated non-portable code points, but
16335 they're valid on this machine */
16336 FALSE, /* similarly, no need for strict */
16337 FALSE, /* Require return to be an ANYOF */
16342 RExC_parse = save_parse + 1;
16343 RExC_end = save_end;
16344 SvREFCNT_dec_NN(final);
16345 SvREFCNT_dec_NN(result_string);
16348 RExC_flags |= RXf_PMf_FOLD;
16352 goto regclass_failed;
16354 /* Fix up the node type if we are in locale. (We have pretended we are
16355 * under /u for the purposes of regclass(), as this construct will only
16356 * work under UTF-8 locales. But now we change the opcode to be ANYOFL (so
16357 * as to cause any warnings about bad locales to be output in regexec.c),
16358 * and add the flag that indicates to check if not in a UTF-8 locale. The
16359 * reason we above forbid optimization into something other than an ANYOF
16360 * node is simply to minimize the number of code changes in regexec.c.
16361 * Otherwise we would have to create new EXACTish node types and deal with
16362 * them. This decision could be revisited should this construct become
16365 * (One might think we could look at the resulting ANYOF node and suppress
16366 * the flag if everything is above 255, as those would be UTF-8 only,
16367 * but this isn't true, as the components that led to that result could
16368 * have been locale-affected, and just happen to cancel each other out
16369 * under UTF-8 locales.) */
16371 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
16373 assert(OP(REGNODE_p(node)) == ANYOF);
16375 OP(REGNODE_p(node)) = ANYOFL;
16376 ANYOF_FLAGS(REGNODE_p(node))
16377 |= ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
16380 nextchar(pRExC_state);
16381 Set_Node_Length(REGNODE_p(node), RExC_parse - oregcomp_parse + 1); /* MJD */
16385 FAIL2("panic: regclass returned failure to handle_sets, " "flags=%#" UVxf,
16389 #ifdef ENABLE_REGEX_SETS_DEBUGGING
16392 S_dump_regex_sets_structures(pTHX_ RExC_state_t *pRExC_state,
16393 AV * stack, const IV fence, AV * fence_stack)
16394 { /* Dumps the stacks in handle_regex_sets() */
16396 const SSize_t stack_top = av_tindex_skip_len_mg(stack);
16397 const SSize_t fence_stack_top = av_tindex_skip_len_mg(fence_stack);
16400 PERL_ARGS_ASSERT_DUMP_REGEX_SETS_STRUCTURES;
16402 PerlIO_printf(Perl_debug_log, "\nParse position is:%s\n", RExC_parse);
16404 if (stack_top < 0) {
16405 PerlIO_printf(Perl_debug_log, "Nothing on stack\n");
16408 PerlIO_printf(Perl_debug_log, "Stack: (fence=%d)\n", (int) fence);
16409 for (i = stack_top; i >= 0; i--) {
16410 SV ** element_ptr = av_fetch(stack, i, FALSE);
16411 if (! element_ptr) {
16414 if (IS_OPERATOR(*element_ptr)) {
16415 PerlIO_printf(Perl_debug_log, "[%d]: %c\n",
16416 (int) i, (int) SvIV(*element_ptr));
16419 PerlIO_printf(Perl_debug_log, "[%d] ", (int) i);
16420 sv_dump(*element_ptr);
16425 if (fence_stack_top < 0) {
16426 PerlIO_printf(Perl_debug_log, "Nothing on fence_stack\n");
16429 PerlIO_printf(Perl_debug_log, "Fence_stack: \n");
16430 for (i = fence_stack_top; i >= 0; i--) {
16431 SV ** element_ptr = av_fetch(fence_stack, i, FALSE);
16432 if (! element_ptr) {
16435 PerlIO_printf(Perl_debug_log, "[%d]: %d\n",
16436 (int) i, (int) SvIV(*element_ptr));
16447 S_add_above_Latin1_folds(pTHX_ RExC_state_t *pRExC_state, const U8 cp, SV** invlist)
16449 /* This adds the Latin1/above-Latin1 folding rules.
16451 * This should be called only for a Latin1-range code points, cp, which is
16452 * known to be involved in a simple fold with other code points above
16453 * Latin1. It would give false results if /aa has been specified.
16454 * Multi-char folds are outside the scope of this, and must be handled
16457 PERL_ARGS_ASSERT_ADD_ABOVE_LATIN1_FOLDS;
16459 assert(HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(cp));
16461 /* The rules that are valid for all Unicode versions are hard-coded in */
16466 add_cp_to_invlist(*invlist, KELVIN_SIGN);
16470 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_LONG_S);
16473 *invlist = add_cp_to_invlist(*invlist, GREEK_CAPITAL_LETTER_MU);
16474 *invlist = add_cp_to_invlist(*invlist, GREEK_SMALL_LETTER_MU);
16476 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
16477 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
16478 *invlist = add_cp_to_invlist(*invlist, ANGSTROM_SIGN);
16480 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
16481 *invlist = add_cp_to_invlist(*invlist,
16482 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
16485 default: /* Other code points are checked against the data for the
16486 current Unicode version */
16488 Size_t folds_count;
16489 unsigned int first_fold;
16490 const unsigned int * remaining_folds;
16494 folded_cp = toFOLD(cp);
16497 U8 dummy_fold[UTF8_MAXBYTES_CASE+1];
16499 folded_cp = _to_fold_latin1(cp, dummy_fold, &dummy_len, 0);
16502 if (folded_cp > 255) {
16503 *invlist = add_cp_to_invlist(*invlist, folded_cp);
16506 folds_count = _inverse_folds(folded_cp, &first_fold,
16508 if (folds_count == 0) {
16510 /* Use deprecated warning to increase the chances of this being
16512 ckWARN2reg_d(RExC_parse,
16513 "Perl folding rules are not up-to-date for 0x%02X;"
16514 " please use the perlbug utility to report;", cp);
16519 if (first_fold > 255) {
16520 *invlist = add_cp_to_invlist(*invlist, first_fold);
16522 for (i = 0; i < folds_count - 1; i++) {
16523 if (remaining_folds[i] > 255) {
16524 *invlist = add_cp_to_invlist(*invlist,
16525 remaining_folds[i]);
16535 S_output_posix_warnings(pTHX_ RExC_state_t *pRExC_state, AV* posix_warnings)
16537 /* Output the elements of the array given by '*posix_warnings' as REGEXP
16541 const bool first_is_fatal = ckDEAD(packWARN(WARN_REGEXP));
16543 PERL_ARGS_ASSERT_OUTPUT_POSIX_WARNINGS;
16545 if (! TO_OUTPUT_WARNINGS(RExC_parse)) {
16549 while ((msg = av_shift(posix_warnings)) != &PL_sv_undef) {
16550 if (first_is_fatal) { /* Avoid leaking this */
16551 av_undef(posix_warnings); /* This isn't necessary if the
16552 array is mortal, but is a
16554 (void) sv_2mortal(msg);
16557 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s", SvPVX(msg));
16558 SvREFCNT_dec_NN(msg);
16561 UPDATE_WARNINGS_LOC(RExC_parse);
16565 S_add_multi_match(pTHX_ AV* multi_char_matches, SV* multi_string, const STRLEN cp_count)
16567 /* This adds the string scalar <multi_string> to the array
16568 * <multi_char_matches>. <multi_string> is known to have exactly
16569 * <cp_count> code points in it. This is used when constructing a
16570 * bracketed character class and we find something that needs to match more
16571 * than a single character.
16573 * <multi_char_matches> is actually an array of arrays. Each top-level
16574 * element is an array that contains all the strings known so far that are
16575 * the same length. And that length (in number of code points) is the same
16576 * as the index of the top-level array. Hence, the [2] element is an
16577 * array, each element thereof is a string containing TWO code points;
16578 * while element [3] is for strings of THREE characters, and so on. Since
16579 * this is for multi-char strings there can never be a [0] nor [1] element.
16581 * When we rewrite the character class below, we will do so such that the
16582 * longest strings are written first, so that it prefers the longest
16583 * matching strings first. This is done even if it turns out that any
16584 * quantifier is non-greedy, out of this programmer's (khw) laziness. Tom
16585 * Christiansen has agreed that this is ok. This makes the test for the
16586 * ligature 'ffi' come before the test for 'ff', for example */
16589 AV** this_array_ptr;
16591 PERL_ARGS_ASSERT_ADD_MULTI_MATCH;
16593 if (! multi_char_matches) {
16594 multi_char_matches = newAV();
16597 if (av_exists(multi_char_matches, cp_count)) {
16598 this_array_ptr = (AV**) av_fetch(multi_char_matches, cp_count, FALSE);
16599 this_array = *this_array_ptr;
16602 this_array = newAV();
16603 av_store(multi_char_matches, cp_count,
16606 av_push(this_array, multi_string);
16608 return multi_char_matches;
16611 /* The names of properties whose definitions are not known at compile time are
16612 * stored in this SV, after a constant heading. So if the length has been
16613 * changed since initialization, then there is a run-time definition. */
16614 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
16615 (SvCUR(listsv) != initial_listsv_len)
16617 /* There is a restricted set of white space characters that are legal when
16618 * ignoring white space in a bracketed character class. This generates the
16619 * code to skip them.
16621 * There is a line below that uses the same white space criteria but is outside
16622 * this macro. Both here and there must use the same definition */
16623 #define SKIP_BRACKETED_WHITE_SPACE(do_skip, p) \
16626 while (isBLANK_A(UCHARAT(p))) \
16633 STATIC regnode_offset
16634 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
16635 const bool stop_at_1, /* Just parse the next thing, don't
16636 look for a full character class */
16637 bool allow_mutiple_chars,
16638 const bool silence_non_portable, /* Don't output warnings
16642 bool optimizable, /* ? Allow a non-ANYOF return
16644 SV** ret_invlist /* Return an inversion list, not a node */
16647 /* parse a bracketed class specification. Most of these will produce an
16648 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
16649 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
16650 * under /i with multi-character folds: it will be rewritten following the
16651 * paradigm of this example, where the <multi-fold>s are characters which
16652 * fold to multiple character sequences:
16653 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
16654 * gets effectively rewritten as:
16655 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
16656 * reg() gets called (recursively) on the rewritten version, and this
16657 * function will return what it constructs. (Actually the <multi-fold>s
16658 * aren't physically removed from the [abcdefghi], it's just that they are
16659 * ignored in the recursion by means of a flag:
16660 * <RExC_in_multi_char_class>.)
16662 * ANYOF nodes contain a bit map for the first NUM_ANYOF_CODE_POINTS
16663 * characters, with the corresponding bit set if that character is in the
16664 * list. For characters above this, an inversion list is used. There
16665 * are extra bits for \w, etc. in locale ANYOFs, as what these match is not
16666 * determinable at compile time
16668 * On success, returns the offset at which any next node should be placed
16669 * into the regex engine program being compiled.
16671 * Returns 0 otherwise, setting flagp to RESTART_PARSE if the parse needs
16672 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to
16677 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
16679 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
16680 regnode_offset ret = -1; /* Initialized to an illegal value */
16682 int namedclass = OOB_NAMEDCLASS;
16683 char *rangebegin = NULL;
16684 SV *listsv = NULL; /* List of \p{user-defined} whose definitions
16685 aren't available at the time this was called */
16686 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
16687 than just initialized. */
16688 SV* properties = NULL; /* Code points that match \p{} \P{} */
16689 SV* posixes = NULL; /* Code points that match classes like [:word:],
16690 extended beyond the Latin1 range. These have to
16691 be kept separate from other code points for much
16692 of this function because their handling is
16693 different under /i, and for most classes under
16695 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
16696 separate for a while from the non-complemented
16697 versions because of complications with /d
16699 SV* simple_posixes = NULL; /* But under some conditions, the classes can be
16700 treated more simply than the general case,
16701 leading to less compilation and execution
16703 UV element_count = 0; /* Number of distinct elements in the class.
16704 Optimizations may be possible if this is tiny */
16705 AV * multi_char_matches = NULL; /* Code points that fold to more than one
16706 character; used under /i */
16708 char * stop_ptr = RExC_end; /* where to stop parsing */
16710 /* ignore unescaped whitespace? */
16711 const bool skip_white = cBOOL( ret_invlist
16712 || (RExC_flags & RXf_PMf_EXTENDED_MORE));
16714 /* inversion list of code points this node matches only when the target
16715 * string is in UTF-8. These are all non-ASCII, < 256. (Because is under
16717 SV* upper_latin1_only_utf8_matches = NULL;
16719 /* Inversion list of code points this node matches regardless of things
16720 * like locale, folding, utf8ness of the target string */
16721 SV* cp_list = NULL;
16723 /* Like cp_list, but code points on this list need to be checked for things
16724 * that fold to/from them under /i */
16725 SV* cp_foldable_list = NULL;
16727 /* Like cp_list, but code points on this list are valid only when the
16728 * runtime locale is UTF-8 */
16729 SV* only_utf8_locale_list = NULL;
16731 /* In a range, if one of the endpoints is non-character-set portable,
16732 * meaning that it hard-codes a code point that may mean a different
16733 * charactger in ASCII vs. EBCDIC, as opposed to, say, a literal 'A' or a
16734 * mnemonic '\t' which each mean the same character no matter which
16735 * character set the platform is on. */
16736 unsigned int non_portable_endpoint = 0;
16738 /* Is the range unicode? which means on a platform that isn't 1-1 native
16739 * to Unicode (i.e. non-ASCII), each code point in it should be considered
16740 * to be a Unicode value. */
16741 bool unicode_range = FALSE;
16742 bool invert = FALSE; /* Is this class to be complemented */
16744 bool warn_super = ALWAYS_WARN_SUPER;
16746 const char * orig_parse = RExC_parse;
16748 /* This variable is used to mark where the end in the input is of something
16749 * that looks like a POSIX construct but isn't. During the parse, when
16750 * something looks like it could be such a construct is encountered, it is
16751 * checked for being one, but not if we've already checked this area of the
16752 * input. Only after this position is reached do we check again */
16753 char *not_posix_region_end = RExC_parse - 1;
16755 AV* posix_warnings = NULL;
16756 const bool do_posix_warnings = ckWARN(WARN_REGEXP);
16757 U8 op = END; /* The returned node-type, initialized to an impossible
16759 U8 anyof_flags = 0; /* flag bits if the node is an ANYOF-type */
16760 U32 posixl = 0; /* bit field of posix classes matched under /l */
16763 /* Flags as to what things aren't knowable until runtime. (Note that these are
16764 * mutually exclusive.) */
16765 #define HAS_USER_DEFINED_PROPERTY 0x01 /* /u any user-defined properties that
16766 haven't been defined as of yet */
16767 #define HAS_D_RUNTIME_DEPENDENCY 0x02 /* /d if the target being matched is
16769 #define HAS_L_RUNTIME_DEPENDENCY 0x04 /* /l what the posix classes match and
16770 what gets folded */
16771 U32 has_runtime_dependency = 0; /* OR of the above flags */
16773 GET_RE_DEBUG_FLAGS_DECL;
16775 PERL_ARGS_ASSERT_REGCLASS;
16777 PERL_UNUSED_ARG(depth);
16781 /* If wants an inversion list returned, we can't optimize to something
16784 optimizable = FALSE;
16787 DEBUG_PARSE("clas");
16789 #if UNICODE_MAJOR_VERSION < 3 /* no multifolds in early Unicode */ \
16790 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0 \
16791 && UNICODE_DOT_DOT_VERSION == 0)
16792 allow_mutiple_chars = FALSE;
16795 /* We include the /i status at the beginning of this so that we can
16796 * know it at runtime */
16797 listsv = sv_2mortal(Perl_newSVpvf(aTHX_ "#%d\n", cBOOL(FOLD)));
16798 initial_listsv_len = SvCUR(listsv);
16799 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
16801 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16803 assert(RExC_parse <= RExC_end);
16805 if (UCHARAT(RExC_parse) == '^') { /* Complement the class */
16808 allow_mutiple_chars = FALSE;
16810 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16813 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
16814 if (! ret_invlist && MAYBE_POSIXCC(UCHARAT(RExC_parse))) {
16815 int maybe_class = handle_possible_posix(pRExC_state,
16817 ¬_posix_region_end,
16819 TRUE /* checking only */);
16820 if (maybe_class >= OOB_NAMEDCLASS && do_posix_warnings) {
16821 ckWARN4reg(not_posix_region_end,
16822 "POSIX syntax [%c %c] belongs inside character classes%s",
16823 *RExC_parse, *RExC_parse,
16824 (maybe_class == OOB_NAMEDCLASS)
16825 ? ((POSIXCC_NOTYET(*RExC_parse))
16826 ? " (but this one isn't implemented)"
16827 : " (but this one isn't fully valid)")
16833 /* If the caller wants us to just parse a single element, accomplish this
16834 * by faking the loop ending condition */
16835 if (stop_at_1 && RExC_end > RExC_parse) {
16836 stop_ptr = RExC_parse + 1;
16839 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
16840 if (UCHARAT(RExC_parse) == ']')
16841 goto charclassloop;
16845 if ( posix_warnings
16846 && av_tindex_skip_len_mg(posix_warnings) >= 0
16847 && RExC_parse > not_posix_region_end)
16849 /* Warnings about posix class issues are considered tentative until
16850 * we are far enough along in the parse that we can no longer
16851 * change our mind, at which point we output them. This is done
16852 * each time through the loop so that a later class won't zap them
16853 * before they have been dealt with. */
16854 output_posix_warnings(pRExC_state, posix_warnings);
16857 if (RExC_parse >= stop_ptr) {
16861 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16863 if (UCHARAT(RExC_parse) == ']') {
16869 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
16870 save_value = value;
16871 save_prevvalue = prevvalue;
16874 rangebegin = RExC_parse;
16876 non_portable_endpoint = 0;
16878 if (UTF && ! UTF8_IS_INVARIANT(* RExC_parse)) {
16879 value = utf8n_to_uvchr((U8*)RExC_parse,
16880 RExC_end - RExC_parse,
16881 &numlen, UTF8_ALLOW_DEFAULT);
16882 RExC_parse += numlen;
16885 value = UCHARAT(RExC_parse++);
16887 if (value == '[') {
16888 char * posix_class_end;
16889 namedclass = handle_possible_posix(pRExC_state,
16892 do_posix_warnings ? &posix_warnings : NULL,
16893 FALSE /* die if error */);
16894 if (namedclass > OOB_NAMEDCLASS) {
16896 /* If there was an earlier attempt to parse this particular
16897 * posix class, and it failed, it was a false alarm, as this
16898 * successful one proves */
16899 if ( posix_warnings
16900 && av_tindex_skip_len_mg(posix_warnings) >= 0
16901 && not_posix_region_end >= RExC_parse
16902 && not_posix_region_end <= posix_class_end)
16904 av_undef(posix_warnings);
16907 RExC_parse = posix_class_end;
16909 else if (namedclass == OOB_NAMEDCLASS) {
16910 not_posix_region_end = posix_class_end;
16913 namedclass = OOB_NAMEDCLASS;
16916 else if ( RExC_parse - 1 > not_posix_region_end
16917 && MAYBE_POSIXCC(value))
16919 (void) handle_possible_posix(
16921 RExC_parse - 1, /* -1 because parse has already been
16923 ¬_posix_region_end,
16924 do_posix_warnings ? &posix_warnings : NULL,
16925 TRUE /* checking only */);
16927 else if ( strict && ! skip_white
16928 && ( _generic_isCC(value, _CC_VERTSPACE)
16929 || is_VERTWS_cp_high(value)))
16931 vFAIL("Literal vertical space in [] is illegal except under /x");
16933 else if (value == '\\') {
16934 /* Is a backslash; get the code point of the char after it */
16936 if (RExC_parse >= RExC_end) {
16937 vFAIL("Unmatched [");
16940 if (UTF && ! UTF8_IS_INVARIANT(UCHARAT(RExC_parse))) {
16941 value = utf8n_to_uvchr((U8*)RExC_parse,
16942 RExC_end - RExC_parse,
16943 &numlen, UTF8_ALLOW_DEFAULT);
16944 RExC_parse += numlen;
16947 value = UCHARAT(RExC_parse++);
16949 /* Some compilers cannot handle switching on 64-bit integer
16950 * values, therefore value cannot be an UV. Yes, this will
16951 * be a problem later if we want switch on Unicode.
16952 * A similar issue a little bit later when switching on
16953 * namedclass. --jhi */
16955 /* If the \ is escaping white space when white space is being
16956 * skipped, it means that that white space is wanted literally, and
16957 * is already in 'value'. Otherwise, need to translate the escape
16958 * into what it signifies. */
16959 if (! skip_white || ! isBLANK_A(value)) switch ((I32)value) {
16961 case 'w': namedclass = ANYOF_WORDCHAR; break;
16962 case 'W': namedclass = ANYOF_NWORDCHAR; break;
16963 case 's': namedclass = ANYOF_SPACE; break;
16964 case 'S': namedclass = ANYOF_NSPACE; break;
16965 case 'd': namedclass = ANYOF_DIGIT; break;
16966 case 'D': namedclass = ANYOF_NDIGIT; break;
16967 case 'v': namedclass = ANYOF_VERTWS; break;
16968 case 'V': namedclass = ANYOF_NVERTWS; break;
16969 case 'h': namedclass = ANYOF_HORIZWS; break;
16970 case 'H': namedclass = ANYOF_NHORIZWS; break;
16971 case 'N': /* Handle \N{NAME} in class */
16973 const char * const backslash_N_beg = RExC_parse - 2;
16976 if (! grok_bslash_N(pRExC_state,
16977 NULL, /* No regnode */
16978 &value, /* Yes single value */
16979 &cp_count, /* Multiple code pt count */
16985 if (*flagp & NEED_UTF8)
16986 FAIL("panic: grok_bslash_N set NEED_UTF8");
16988 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
16990 if (cp_count < 0) {
16991 vFAIL("\\N in a character class must be a named character: \\N{...}");
16993 else if (cp_count == 0) {
16994 ckWARNreg(RExC_parse,
16995 "Ignoring zero length \\N{} in character class");
16997 else { /* cp_count > 1 */
16998 assert(cp_count > 1);
16999 if (! RExC_in_multi_char_class) {
17000 if ( ! allow_mutiple_chars
17003 || *RExC_parse == '-')
17007 vFAIL("\\N{} in inverted character class or as a range end-point is restricted to one character");
17009 ckWARNreg(RExC_parse, "Using just the first character returned by \\N{} in character class");
17010 break; /* <value> contains the first code
17011 point. Drop out of the switch to
17015 SV * multi_char_N = newSVpvn(backslash_N_beg,
17016 RExC_parse - backslash_N_beg);
17018 = add_multi_match(multi_char_matches,
17023 } /* End of cp_count != 1 */
17025 /* This element should not be processed further in this
17028 value = save_value;
17029 prevvalue = save_prevvalue;
17030 continue; /* Back to top of loop to get next char */
17033 /* Here, is a single code point, and <value> contains it */
17034 unicode_range = TRUE; /* \N{} are Unicode */
17042 /* \p means they want Unicode semantics */
17043 REQUIRE_UNI_RULES(flagp, 0);
17045 if (RExC_parse >= RExC_end)
17046 vFAIL2("Empty \\%c", (U8)value);
17047 if (*RExC_parse == '{') {
17048 const U8 c = (U8)value;
17049 e = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
17052 vFAIL2("Missing right brace on \\%c{}", c);
17057 /* White space is allowed adjacent to the braces and after
17058 * any '^', even when not under /x */
17059 while (isSPACE(*RExC_parse)) {
17063 if (UCHARAT(RExC_parse) == '^') {
17065 /* toggle. (The rhs xor gets the single bit that
17066 * differs between P and p; the other xor inverts just
17068 value ^= 'P' ^ 'p';
17071 while (isSPACE(*RExC_parse)) {
17076 if (e == RExC_parse)
17077 vFAIL2("Empty \\%c{}", c);
17079 n = e - RExC_parse;
17080 while (isSPACE(*(RExC_parse + n - 1)))
17083 } /* The \p isn't immediately followed by a '{' */
17084 else if (! isALPHA(*RExC_parse)) {
17085 RExC_parse += (UTF)
17086 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
17088 vFAIL2("Character following \\%c must be '{' or a "
17089 "single-character Unicode property name",
17097 char* name = RExC_parse;
17099 /* Any message returned about expanding the definition */
17100 SV* msg = newSVpvs_flags("", SVs_TEMP);
17102 /* If set TRUE, the property is user-defined as opposed to
17103 * official Unicode */
17104 bool user_defined = FALSE;
17106 SV * prop_definition = parse_uniprop_string(
17107 name, n, UTF, FOLD,
17108 FALSE, /* This is compile-time */
17110 /* We can't defer this defn when
17111 * the full result is required in
17113 ! cBOOL(ret_invlist),
17119 if (SvCUR(msg)) { /* Assumes any error causes a msg */
17120 assert(prop_definition == NULL);
17121 RExC_parse = e + 1;
17122 if (SvUTF8(msg)) { /* msg being UTF-8 makes the whole
17123 thing so, or else the display is
17127 /* diag_listed_as: Can't find Unicode property definition "%s" in regex; marked by <-- HERE in m/%s/ */
17128 vFAIL2utf8f("%" UTF8f, UTF8fARG(SvUTF8(msg),
17129 SvCUR(msg), SvPVX(msg)));
17132 if (! is_invlist(prop_definition)) {
17134 /* Here, the definition isn't known, so we have gotten
17135 * returned a string that will be evaluated if and when
17136 * encountered at runtime. We add it to the list of
17137 * such properties, along with whether it should be
17138 * complemented or not */
17139 if (value == 'P') {
17140 sv_catpvs(listsv, "!");
17143 sv_catpvs(listsv, "+");
17145 sv_catsv(listsv, prop_definition);
17147 has_runtime_dependency |= HAS_USER_DEFINED_PROPERTY;
17149 /* We don't know yet what this matches, so have to flag
17151 anyof_flags |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
17154 assert (prop_definition && is_invlist(prop_definition));
17156 /* Here we do have the complete property definition
17158 * Temporary workaround for [perl #133136]. For this
17159 * precise input that is in the .t that is failing,
17160 * load utf8.pm, which is what the test wants, so that
17161 * that .t passes */
17162 if ( memEQs(RExC_start, e + 1 - RExC_start,
17164 && ! hv_common(GvHVn(PL_incgv),
17166 "utf8.pm", sizeof("utf8.pm") - 1,
17167 0, HV_FETCH_ISEXISTS, NULL, 0))
17169 require_pv("utf8.pm");
17172 if (! user_defined &&
17173 /* We warn on matching an above-Unicode code point
17174 * if the match would return true, except don't
17175 * warn for \p{All}, which has exactly one element
17177 (_invlist_contains_cp(prop_definition, 0x110000)
17178 && (! (_invlist_len(prop_definition) == 1
17179 && *invlist_array(prop_definition) == 0))))
17184 /* Invert if asking for the complement */
17185 if (value == 'P') {
17186 _invlist_union_complement_2nd(properties,
17191 _invlist_union(properties, prop_definition, &properties);
17196 RExC_parse = e + 1;
17197 namedclass = ANYOF_UNIPROP; /* no official name, but it's
17201 case 'n': value = '\n'; break;
17202 case 'r': value = '\r'; break;
17203 case 't': value = '\t'; break;
17204 case 'f': value = '\f'; break;
17205 case 'b': value = '\b'; break;
17206 case 'e': value = ESC_NATIVE; break;
17207 case 'a': value = '\a'; break;
17209 RExC_parse--; /* function expects to be pointed at the 'o' */
17211 const char* error_msg;
17212 bool valid = grok_bslash_o(&RExC_parse,
17216 TO_OUTPUT_WARNINGS(RExC_parse),
17218 silence_non_portable,
17223 UPDATE_WARNINGS_LOC(RExC_parse - 1);
17225 non_portable_endpoint++;
17228 RExC_parse--; /* function expects to be pointed at the 'x' */
17230 const char* error_msg;
17231 bool valid = grok_bslash_x(&RExC_parse,
17235 TO_OUTPUT_WARNINGS(RExC_parse),
17237 silence_non_portable,
17242 UPDATE_WARNINGS_LOC(RExC_parse - 1);
17244 non_portable_endpoint++;
17247 value = grok_bslash_c(*RExC_parse, TO_OUTPUT_WARNINGS(RExC_parse));
17248 UPDATE_WARNINGS_LOC(RExC_parse);
17250 non_portable_endpoint++;
17252 case '0': case '1': case '2': case '3': case '4':
17253 case '5': case '6': case '7':
17255 /* Take 1-3 octal digits */
17256 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
17257 numlen = (strict) ? 4 : 3;
17258 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
17259 RExC_parse += numlen;
17262 RExC_parse += (UTF)
17263 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
17265 vFAIL("Need exactly 3 octal digits");
17267 else if ( numlen < 3 /* like \08, \178 */
17268 && RExC_parse < RExC_end
17269 && isDIGIT(*RExC_parse)
17270 && ckWARN(WARN_REGEXP))
17272 reg_warn_non_literal_string(
17274 form_short_octal_warning(RExC_parse, numlen));
17277 non_portable_endpoint++;
17281 /* Allow \_ to not give an error */
17282 if (isWORDCHAR(value) && value != '_') {
17284 vFAIL2("Unrecognized escape \\%c in character class",
17288 ckWARN2reg(RExC_parse,
17289 "Unrecognized escape \\%c in character class passed through",
17294 } /* End of switch on char following backslash */
17295 } /* end of handling backslash escape sequences */
17297 /* Here, we have the current token in 'value' */
17299 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
17302 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
17303 * literal, as is the character that began the false range, i.e.
17304 * the 'a' in the examples */
17306 const int w = (RExC_parse >= rangebegin)
17307 ? RExC_parse - rangebegin
17311 "False [] range \"%" UTF8f "\"",
17312 UTF8fARG(UTF, w, rangebegin));
17315 ckWARN2reg(RExC_parse,
17316 "False [] range \"%" UTF8f "\"",
17317 UTF8fARG(UTF, w, rangebegin));
17318 cp_list = add_cp_to_invlist(cp_list, '-');
17319 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
17323 range = 0; /* this was not a true range */
17324 element_count += 2; /* So counts for three values */
17327 classnum = namedclass_to_classnum(namedclass);
17329 if (LOC && namedclass < ANYOF_POSIXL_MAX
17330 #ifndef HAS_ISASCII
17331 && classnum != _CC_ASCII
17334 SV* scratch_list = NULL;
17336 /* What the Posix classes (like \w, [:space:]) match isn't
17337 * generally knowable under locale until actual match time. A
17338 * special node is used for these which has extra space for a
17339 * bitmap, with a bit reserved for each named class that is to
17340 * be matched against. (This isn't needed for \p{} and
17341 * pseudo-classes, as they are not affected by locale, and
17342 * hence are dealt with separately.) However, if a named class
17343 * and its complement are both present, then it matches
17344 * everything, and there is no runtime dependency. Odd numbers
17345 * are the complements of the next lower number, so xor works.
17346 * (Note that something like [\w\D] should match everything,
17347 * because \d should be a proper subset of \w. But rather than
17348 * trust that the locale is well behaved, we leave this to
17349 * runtime to sort out) */
17350 if (POSIXL_TEST(posixl, namedclass ^ 1)) {
17351 cp_list = _add_range_to_invlist(cp_list, 0, UV_MAX);
17352 POSIXL_ZERO(posixl);
17353 has_runtime_dependency &= ~HAS_L_RUNTIME_DEPENDENCY;
17354 anyof_flags &= ~ANYOF_MATCHES_POSIXL;
17355 continue; /* We could ignore the rest of the class, but
17356 best to parse it for any errors */
17358 else { /* Here, isn't the complement of any already parsed
17360 POSIXL_SET(posixl, namedclass);
17361 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
17362 anyof_flags |= ANYOF_MATCHES_POSIXL;
17364 /* The above-Latin1 characters are not subject to locale
17365 * rules. Just add them to the unconditionally-matched
17368 /* Get the list of the above-Latin1 code points this
17370 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
17371 PL_XPosix_ptrs[classnum],
17373 /* Odd numbers are complements,
17374 * like NDIGIT, NASCII, ... */
17375 namedclass % 2 != 0,
17377 /* Checking if 'cp_list' is NULL first saves an extra
17378 * clone. Its reference count will be decremented at the
17379 * next union, etc, or if this is the only instance, at the
17380 * end of the routine */
17382 cp_list = scratch_list;
17385 _invlist_union(cp_list, scratch_list, &cp_list);
17386 SvREFCNT_dec_NN(scratch_list);
17388 continue; /* Go get next character */
17393 /* Here, is not /l, or is a POSIX class for which /l doesn't
17394 * matter (or is a Unicode property, which is skipped here). */
17395 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
17396 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
17398 /* Here, should be \h, \H, \v, or \V. None of /d, /i
17399 * nor /l make a difference in what these match,
17400 * therefore we just add what they match to cp_list. */
17401 if (classnum != _CC_VERTSPACE) {
17402 assert( namedclass == ANYOF_HORIZWS
17403 || namedclass == ANYOF_NHORIZWS);
17405 /* It turns out that \h is just a synonym for
17407 classnum = _CC_BLANK;
17410 _invlist_union_maybe_complement_2nd(
17412 PL_XPosix_ptrs[classnum],
17413 namedclass % 2 != 0, /* Complement if odd
17414 (NHORIZWS, NVERTWS)
17419 else if ( AT_LEAST_UNI_SEMANTICS
17420 || classnum == _CC_ASCII
17421 || (DEPENDS_SEMANTICS && ( classnum == _CC_DIGIT
17422 || classnum == _CC_XDIGIT)))
17424 /* We usually have to worry about /d affecting what POSIX
17425 * classes match, with special code needed because we won't
17426 * know until runtime what all matches. But there is no
17427 * extra work needed under /u and /a; and [:ascii:] is
17428 * unaffected by /d; and :digit: and :xdigit: don't have
17429 * runtime differences under /d. So we can special case
17430 * these, and avoid some extra work below, and at runtime.
17432 _invlist_union_maybe_complement_2nd(
17434 ((AT_LEAST_ASCII_RESTRICTED)
17435 ? PL_Posix_ptrs[classnum]
17436 : PL_XPosix_ptrs[classnum]),
17437 namedclass % 2 != 0,
17440 else { /* Garden variety class. If is NUPPER, NALPHA, ...
17441 complement and use nposixes */
17442 SV** posixes_ptr = namedclass % 2 == 0
17445 _invlist_union_maybe_complement_2nd(
17447 PL_XPosix_ptrs[classnum],
17448 namedclass % 2 != 0,
17452 } /* end of namedclass \blah */
17454 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
17456 /* If 'range' is set, 'value' is the ending of a range--check its
17457 * validity. (If value isn't a single code point in the case of a
17458 * range, we should have figured that out above in the code that
17459 * catches false ranges). Later, we will handle each individual code
17460 * point in the range. If 'range' isn't set, this could be the
17461 * beginning of a range, so check for that by looking ahead to see if
17462 * the next real character to be processed is the range indicator--the
17467 /* For unicode ranges, we have to test that the Unicode as opposed
17468 * to the native values are not decreasing. (Above 255, there is
17469 * no difference between native and Unicode) */
17470 if (unicode_range && prevvalue < 255 && value < 255) {
17471 if (NATIVE_TO_LATIN1(prevvalue) > NATIVE_TO_LATIN1(value)) {
17472 goto backwards_range;
17477 if (prevvalue > value) /* b-a */ {
17482 w = RExC_parse - rangebegin;
17484 "Invalid [] range \"%" UTF8f "\"",
17485 UTF8fARG(UTF, w, rangebegin));
17486 NOT_REACHED; /* NOTREACHED */
17490 prevvalue = value; /* save the beginning of the potential range */
17491 if (! stop_at_1 /* Can't be a range if parsing just one thing */
17492 && *RExC_parse == '-')
17494 char* next_char_ptr = RExC_parse + 1;
17496 /* Get the next real char after the '-' */
17497 SKIP_BRACKETED_WHITE_SPACE(skip_white, next_char_ptr);
17499 /* If the '-' is at the end of the class (just before the ']',
17500 * it is a literal minus; otherwise it is a range */
17501 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
17502 RExC_parse = next_char_ptr;
17504 /* a bad range like \w-, [:word:]- ? */
17505 if (namedclass > OOB_NAMEDCLASS) {
17506 if (strict || ckWARN(WARN_REGEXP)) {
17507 const int w = RExC_parse >= rangebegin
17508 ? RExC_parse - rangebegin
17511 vFAIL4("False [] range \"%*.*s\"",
17516 "False [] range \"%*.*s\"",
17520 cp_list = add_cp_to_invlist(cp_list, '-');
17523 range = 1; /* yeah, it's a range! */
17524 continue; /* but do it the next time */
17529 if (namedclass > OOB_NAMEDCLASS) {
17533 /* Here, we have a single value this time through the loop, and
17534 * <prevvalue> is the beginning of the range, if any; or <value> if
17537 /* non-Latin1 code point implies unicode semantics. */
17539 REQUIRE_UNI_RULES(flagp, 0);
17542 /* Ready to process either the single value, or the completed range.
17543 * For single-valued non-inverted ranges, we consider the possibility
17544 * of multi-char folds. (We made a conscious decision to not do this
17545 * for the other cases because it can often lead to non-intuitive
17546 * results. For example, you have the peculiar case that:
17547 * "s s" =~ /^[^\xDF]+$/i => Y
17548 * "ss" =~ /^[^\xDF]+$/i => N
17550 * See [perl #89750] */
17551 if (FOLD && allow_mutiple_chars && value == prevvalue) {
17552 if ( value == LATIN_SMALL_LETTER_SHARP_S
17553 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
17556 /* Here <value> is indeed a multi-char fold. Get what it is */
17558 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
17561 UV folded = _to_uni_fold_flags(
17565 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
17566 ? FOLD_FLAGS_NOMIX_ASCII
17570 /* Here, <folded> should be the first character of the
17571 * multi-char fold of <value>, with <foldbuf> containing the
17572 * whole thing. But, if this fold is not allowed (because of
17573 * the flags), <fold> will be the same as <value>, and should
17574 * be processed like any other character, so skip the special
17576 if (folded != value) {
17578 /* Skip if we are recursed, currently parsing the class
17579 * again. Otherwise add this character to the list of
17580 * multi-char folds. */
17581 if (! RExC_in_multi_char_class) {
17582 STRLEN cp_count = utf8_length(foldbuf,
17583 foldbuf + foldlen);
17584 SV* multi_fold = sv_2mortal(newSVpvs(""));
17586 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%" UVXf "}", value);
17589 = add_multi_match(multi_char_matches,
17595 /* This element should not be processed further in this
17598 value = save_value;
17599 prevvalue = save_prevvalue;
17605 if (strict && ckWARN(WARN_REGEXP)) {
17608 /* If the range starts above 255, everything is portable and
17609 * likely to be so for any forseeable character set, so don't
17611 if (unicode_range && non_portable_endpoint && prevvalue < 256) {
17612 vWARN(RExC_parse, "Both or neither range ends should be Unicode");
17614 else if (prevvalue != value) {
17616 /* Under strict, ranges that stop and/or end in an ASCII
17617 * printable should have each end point be a portable value
17618 * for it (preferably like 'A', but we don't warn if it is
17619 * a (portable) Unicode name or code point), and the range
17620 * must be be all digits or all letters of the same case.
17621 * Otherwise, the range is non-portable and unclear as to
17622 * what it contains */
17623 if ( (isPRINT_A(prevvalue) || isPRINT_A(value))
17624 && ( non_portable_endpoint
17625 || ! ( (isDIGIT_A(prevvalue) && isDIGIT_A(value))
17626 || (isLOWER_A(prevvalue) && isLOWER_A(value))
17627 || (isUPPER_A(prevvalue) && isUPPER_A(value))
17629 vWARN(RExC_parse, "Ranges of ASCII printables should"
17630 " be some subset of \"0-9\","
17631 " \"A-Z\", or \"a-z\"");
17633 else if (prevvalue >= FIRST_NON_ASCII_DECIMAL_DIGIT) {
17634 SSize_t index_start;
17635 SSize_t index_final;
17637 /* But the nature of Unicode and languages mean we
17638 * can't do the same checks for above-ASCII ranges,
17639 * except in the case of digit ones. These should
17640 * contain only digits from the same group of 10. The
17641 * ASCII case is handled just above. Hence here, the
17642 * range could be a range of digits. First some
17643 * unlikely special cases. Grandfather in that a range
17644 * ending in 19DA (NEW TAI LUE THAM DIGIT ONE) is bad
17645 * if its starting value is one of the 10 digits prior
17646 * to it. This is because it is an alternate way of
17647 * writing 19D1, and some people may expect it to be in
17648 * that group. But it is bad, because it won't give
17649 * the expected results. In Unicode 5.2 it was
17650 * considered to be in that group (of 11, hence), but
17651 * this was fixed in the next version */
17653 if (UNLIKELY(value == 0x19DA && prevvalue >= 0x19D0)) {
17654 goto warn_bad_digit_range;
17656 else if (UNLIKELY( prevvalue >= 0x1D7CE
17657 && value <= 0x1D7FF))
17659 /* This is the only other case currently in Unicode
17660 * where the algorithm below fails. The code
17661 * points just above are the end points of a single
17662 * range containing only decimal digits. It is 5
17663 * different series of 0-9. All other ranges of
17664 * digits currently in Unicode are just a single
17665 * series. (And mktables will notify us if a later
17666 * Unicode version breaks this.)
17668 * If the range being checked is at most 9 long,
17669 * and the digit values represented are in
17670 * numerical order, they are from the same series.
17672 if ( value - prevvalue > 9
17673 || ((( value - 0x1D7CE) % 10)
17674 <= (prevvalue - 0x1D7CE) % 10))
17676 goto warn_bad_digit_range;
17681 /* For all other ranges of digits in Unicode, the
17682 * algorithm is just to check if both end points
17683 * are in the same series, which is the same range.
17685 index_start = _invlist_search(
17686 PL_XPosix_ptrs[_CC_DIGIT],
17689 /* Warn if the range starts and ends with a digit,
17690 * and they are not in the same group of 10. */
17691 if ( index_start >= 0
17692 && ELEMENT_RANGE_MATCHES_INVLIST(index_start)
17694 _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
17695 value)) != index_start
17696 && index_final >= 0
17697 && ELEMENT_RANGE_MATCHES_INVLIST(index_final))
17699 warn_bad_digit_range:
17700 vWARN(RExC_parse, "Ranges of digits should be"
17701 " from the same group of"
17708 if ((! range || prevvalue == value) && non_portable_endpoint) {
17709 if (isPRINT_A(value)) {
17712 if (isBACKSLASHED_PUNCT(value)) {
17713 literal[d++] = '\\';
17715 literal[d++] = (char) value;
17716 literal[d++] = '\0';
17719 "\"%.*s\" is more clearly written simply as \"%s\"",
17720 (int) (RExC_parse - rangebegin),
17725 else if isMNEMONIC_CNTRL(value) {
17727 "\"%.*s\" is more clearly written simply as \"%s\"",
17728 (int) (RExC_parse - rangebegin),
17730 cntrl_to_mnemonic((U8) value)
17736 /* Deal with this element of the class */
17739 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
17742 /* On non-ASCII platforms, for ranges that span all of 0..255, and ones
17743 * that don't require special handling, we can just add the range like
17744 * we do for ASCII platforms */
17745 if ((UNLIKELY(prevvalue == 0) && value >= 255)
17746 || ! (prevvalue < 256
17748 || (! non_portable_endpoint
17749 && ((isLOWER_A(prevvalue) && isLOWER_A(value))
17750 || (isUPPER_A(prevvalue)
17751 && isUPPER_A(value)))))))
17753 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
17757 /* Here, requires special handling. This can be because it is a
17758 * range whose code points are considered to be Unicode, and so
17759 * must be individually translated into native, or because its a
17760 * subrange of 'A-Z' or 'a-z' which each aren't contiguous in
17761 * EBCDIC, but we have defined them to include only the "expected"
17762 * upper or lower case ASCII alphabetics. Subranges above 255 are
17763 * the same in native and Unicode, so can be added as a range */
17764 U8 start = NATIVE_TO_LATIN1(prevvalue);
17766 U8 end = (value < 256) ? NATIVE_TO_LATIN1(value) : 255;
17767 for (j = start; j <= end; j++) {
17768 cp_foldable_list = add_cp_to_invlist(cp_foldable_list, LATIN1_TO_NATIVE(j));
17771 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
17777 range = 0; /* this range (if it was one) is done now */
17778 } /* End of loop through all the text within the brackets */
17780 if ( posix_warnings && av_tindex_skip_len_mg(posix_warnings) >= 0) {
17781 output_posix_warnings(pRExC_state, posix_warnings);
17784 /* If anything in the class expands to more than one character, we have to
17785 * deal with them by building up a substitute parse string, and recursively
17786 * calling reg() on it, instead of proceeding */
17787 if (multi_char_matches) {
17788 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
17791 char *save_end = RExC_end;
17792 char *save_parse = RExC_parse;
17793 char *save_start = RExC_start;
17794 Size_t constructed_prefix_len = 0; /* This gives the length of the
17795 constructed portion of the
17796 substitute parse. */
17797 bool first_time = TRUE; /* First multi-char occurrence doesn't get
17802 /* Only one level of recursion allowed */
17803 assert(RExC_copy_start_in_constructed == RExC_precomp);
17805 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
17806 because too confusing */
17808 sv_catpvs(substitute_parse, "(?:");
17812 /* Look at the longest folds first */
17813 for (cp_count = av_tindex_skip_len_mg(multi_char_matches);
17818 if (av_exists(multi_char_matches, cp_count)) {
17819 AV** this_array_ptr;
17822 this_array_ptr = (AV**) av_fetch(multi_char_matches,
17824 while ((this_sequence = av_pop(*this_array_ptr)) !=
17827 if (! first_time) {
17828 sv_catpvs(substitute_parse, "|");
17830 first_time = FALSE;
17832 sv_catpv(substitute_parse, SvPVX(this_sequence));
17837 /* If the character class contains anything else besides these
17838 * multi-character folds, have to include it in recursive parsing */
17839 if (element_count) {
17840 sv_catpvs(substitute_parse, "|[");
17841 constructed_prefix_len = SvCUR(substitute_parse);
17842 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
17844 /* Put in a closing ']' only if not going off the end, as otherwise
17845 * we are adding something that really isn't there */
17846 if (RExC_parse < RExC_end) {
17847 sv_catpvs(substitute_parse, "]");
17851 sv_catpvs(substitute_parse, ")");
17854 /* This is a way to get the parse to skip forward a whole named
17855 * sequence instead of matching the 2nd character when it fails the
17857 sv_catpvs(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
17861 /* Set up the data structure so that any errors will be properly
17862 * reported. See the comments at the definition of
17863 * REPORT_LOCATION_ARGS for details */
17864 RExC_copy_start_in_input = (char *) orig_parse;
17865 RExC_start = RExC_parse = SvPV(substitute_parse, len);
17866 RExC_copy_start_in_constructed = RExC_start + constructed_prefix_len;
17867 RExC_end = RExC_parse + len;
17868 RExC_in_multi_char_class = 1;
17870 ret = reg(pRExC_state, 1, ®_flags, depth+1);
17872 *flagp |= reg_flags & (HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_PARSE|NEED_UTF8);
17874 /* And restore so can parse the rest of the pattern */
17875 RExC_parse = save_parse;
17876 RExC_start = RExC_copy_start_in_constructed = RExC_copy_start_in_input = save_start;
17877 RExC_end = save_end;
17878 RExC_in_multi_char_class = 0;
17879 SvREFCNT_dec_NN(multi_char_matches);
17883 /* If folding, we calculate all characters that could fold to or from the
17884 * ones already on the list */
17885 if (cp_foldable_list) {
17887 UV start, end; /* End points of code point ranges */
17889 SV* fold_intersection = NULL;
17892 /* Our calculated list will be for Unicode rules. For locale
17893 * matching, we have to keep a separate list that is consulted at
17894 * runtime only when the locale indicates Unicode rules (and we
17895 * don't include potential matches in the ASCII/Latin1 range, as
17896 * any code point could fold to any other, based on the run-time
17897 * locale). For non-locale, we just use the general list */
17899 use_list = &only_utf8_locale_list;
17902 use_list = &cp_list;
17905 /* Only the characters in this class that participate in folds need
17906 * be checked. Get the intersection of this class and all the
17907 * possible characters that are foldable. This can quickly narrow
17908 * down a large class */
17909 _invlist_intersection(PL_in_some_fold, cp_foldable_list,
17910 &fold_intersection);
17912 /* Now look at the foldable characters in this class individually */
17913 invlist_iterinit(fold_intersection);
17914 while (invlist_iternext(fold_intersection, &start, &end)) {
17918 /* Look at every character in the range */
17919 for (j = start; j <= end; j++) {
17920 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
17923 Size_t folds_count;
17924 unsigned int first_fold;
17925 const unsigned int * remaining_folds;
17929 /* Under /l, we don't know what code points below 256
17930 * fold to, except we do know the MICRO SIGN folds to
17931 * an above-255 character if the locale is UTF-8, so we
17932 * add it to the special list (in *use_list) Otherwise
17933 * we know now what things can match, though some folds
17934 * are valid under /d only if the target is UTF-8.
17935 * Those go in a separate list */
17936 if ( IS_IN_SOME_FOLD_L1(j)
17937 && ! (LOC && j != MICRO_SIGN))
17940 /* ASCII is always matched; non-ASCII is matched
17941 * only under Unicode rules (which could happen
17942 * under /l if the locale is a UTF-8 one */
17943 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
17944 *use_list = add_cp_to_invlist(*use_list,
17945 PL_fold_latin1[j]);
17947 else if (j != PL_fold_latin1[j]) {
17948 upper_latin1_only_utf8_matches
17949 = add_cp_to_invlist(
17950 upper_latin1_only_utf8_matches,
17951 PL_fold_latin1[j]);
17955 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(j)
17956 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
17958 add_above_Latin1_folds(pRExC_state,
17965 /* Here is an above Latin1 character. We don't have the
17966 * rules hard-coded for it. First, get its fold. This is
17967 * the simple fold, as the multi-character folds have been
17968 * handled earlier and separated out */
17969 folded = _to_uni_fold_flags(j, foldbuf, &foldlen,
17970 (ASCII_FOLD_RESTRICTED)
17971 ? FOLD_FLAGS_NOMIX_ASCII
17974 /* Single character fold of above Latin1. Add everything
17975 * in its fold closure to the list that this node should
17977 folds_count = _inverse_folds(folded, &first_fold,
17979 for (k = 0; k <= folds_count; k++) {
17980 UV c = (k == 0) /* First time through use itself */
17982 : (k == 1) /* 2nd time use, the first fold */
17985 /* Then the remaining ones */
17986 : remaining_folds[k-2];
17988 /* /aa doesn't allow folds between ASCII and non- */
17989 if (( ASCII_FOLD_RESTRICTED
17990 && (isASCII(c) != isASCII(j))))
17995 /* Folds under /l which cross the 255/256 boundary are
17996 * added to a separate list. (These are valid only
17997 * when the locale is UTF-8.) */
17998 if (c < 256 && LOC) {
17999 *use_list = add_cp_to_invlist(*use_list, c);
18003 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
18005 cp_list = add_cp_to_invlist(cp_list, c);
18008 /* Similarly folds involving non-ascii Latin1
18009 * characters under /d are added to their list */
18010 upper_latin1_only_utf8_matches
18011 = add_cp_to_invlist(
18012 upper_latin1_only_utf8_matches,
18018 SvREFCNT_dec_NN(fold_intersection);
18021 /* Now that we have finished adding all the folds, there is no reason
18022 * to keep the foldable list separate */
18023 _invlist_union(cp_list, cp_foldable_list, &cp_list);
18024 SvREFCNT_dec_NN(cp_foldable_list);
18027 /* And combine the result (if any) with any inversion lists from posix
18028 * classes. The lists are kept separate up to now because we don't want to
18029 * fold the classes */
18030 if (simple_posixes) { /* These are the classes known to be unaffected by
18033 _invlist_union(cp_list, simple_posixes, &cp_list);
18034 SvREFCNT_dec_NN(simple_posixes);
18037 cp_list = simple_posixes;
18040 if (posixes || nposixes) {
18041 if (! DEPENDS_SEMANTICS) {
18043 /* For everything but /d, we can just add the current 'posixes' and
18044 * 'nposixes' to the main list */
18047 _invlist_union(cp_list, posixes, &cp_list);
18048 SvREFCNT_dec_NN(posixes);
18056 _invlist_union(cp_list, nposixes, &cp_list);
18057 SvREFCNT_dec_NN(nposixes);
18060 cp_list = nposixes;
18065 /* Under /d, things like \w match upper Latin1 characters only if
18066 * the target string is in UTF-8. But things like \W match all the
18067 * upper Latin1 characters if the target string is not in UTF-8.
18069 * Handle the case with something like \W separately */
18071 SV* only_non_utf8_list = invlist_clone(PL_UpperLatin1, NULL);
18073 /* A complemented posix class matches all upper Latin1
18074 * characters if not in UTF-8. And it matches just certain
18075 * ones when in UTF-8. That means those certain ones are
18076 * matched regardless, so can just be added to the
18077 * unconditional list */
18079 _invlist_union(cp_list, nposixes, &cp_list);
18080 SvREFCNT_dec_NN(nposixes);
18084 cp_list = nposixes;
18087 /* Likewise for 'posixes' */
18088 _invlist_union(posixes, cp_list, &cp_list);
18090 /* Likewise for anything else in the range that matched only
18092 if (upper_latin1_only_utf8_matches) {
18093 _invlist_union(cp_list,
18094 upper_latin1_only_utf8_matches,
18096 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
18097 upper_latin1_only_utf8_matches = NULL;
18100 /* If we don't match all the upper Latin1 characters regardless
18101 * of UTF-8ness, we have to set a flag to match the rest when
18103 _invlist_subtract(only_non_utf8_list, cp_list,
18104 &only_non_utf8_list);
18105 if (_invlist_len(only_non_utf8_list) != 0) {
18106 anyof_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
18108 SvREFCNT_dec_NN(only_non_utf8_list);
18111 /* Here there were no complemented posix classes. That means
18112 * the upper Latin1 characters in 'posixes' match only when the
18113 * target string is in UTF-8. So we have to add them to the
18114 * list of those types of code points, while adding the
18115 * remainder to the unconditional list.
18117 * First calculate what they are */
18118 SV* nonascii_but_latin1_properties = NULL;
18119 _invlist_intersection(posixes, PL_UpperLatin1,
18120 &nonascii_but_latin1_properties);
18122 /* And add them to the final list of such characters. */
18123 _invlist_union(upper_latin1_only_utf8_matches,
18124 nonascii_but_latin1_properties,
18125 &upper_latin1_only_utf8_matches);
18127 /* Remove them from what now becomes the unconditional list */
18128 _invlist_subtract(posixes, nonascii_but_latin1_properties,
18131 /* And add those unconditional ones to the final list */
18133 _invlist_union(cp_list, posixes, &cp_list);
18134 SvREFCNT_dec_NN(posixes);
18141 SvREFCNT_dec(nonascii_but_latin1_properties);
18143 /* Get rid of any characters from the conditional list that we
18144 * now know are matched unconditionally, which may make that
18146 _invlist_subtract(upper_latin1_only_utf8_matches,
18148 &upper_latin1_only_utf8_matches);
18149 if (_invlist_len(upper_latin1_only_utf8_matches) == 0) {
18150 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
18151 upper_latin1_only_utf8_matches = NULL;
18157 /* And combine the result (if any) with any inversion list from properties.
18158 * The lists are kept separate up to now so that we can distinguish the two
18159 * in regards to matching above-Unicode. A run-time warning is generated
18160 * if a Unicode property is matched against a non-Unicode code point. But,
18161 * we allow user-defined properties to match anything, without any warning,
18162 * and we also suppress the warning if there is a portion of the character
18163 * class that isn't a Unicode property, and which matches above Unicode, \W
18164 * or [\x{110000}] for example.
18165 * (Note that in this case, unlike the Posix one above, there is no
18166 * <upper_latin1_only_utf8_matches>, because having a Unicode property
18167 * forces Unicode semantics */
18171 /* If it matters to the final outcome, see if a non-property
18172 * component of the class matches above Unicode. If so, the
18173 * warning gets suppressed. This is true even if just a single
18174 * such code point is specified, as, though not strictly correct if
18175 * another such code point is matched against, the fact that they
18176 * are using above-Unicode code points indicates they should know
18177 * the issues involved */
18179 warn_super = ! (invert
18180 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
18183 _invlist_union(properties, cp_list, &cp_list);
18184 SvREFCNT_dec_NN(properties);
18187 cp_list = properties;
18192 |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
18194 /* Because an ANYOF node is the only one that warns, this node
18195 * can't be optimized into something else */
18196 optimizable = FALSE;
18200 /* Here, we have calculated what code points should be in the character
18203 * Now we can see about various optimizations. Fold calculation (which we
18204 * did above) needs to take place before inversion. Otherwise /[^k]/i
18205 * would invert to include K, which under /i would match k, which it
18206 * shouldn't. Therefore we can't invert folded locale now, as it won't be
18207 * folded until runtime */
18209 /* If we didn't do folding, it's because some information isn't available
18210 * until runtime; set the run-time fold flag for these We know to set the
18211 * flag if we have a non-NULL list for UTF-8 locales, or the class matches
18212 * at least one 0-255 range code point */
18215 /* Some things on the list might be unconditionally included because of
18216 * other components. Remove them, and clean up the list if it goes to
18218 if (only_utf8_locale_list && cp_list) {
18219 _invlist_subtract(only_utf8_locale_list, cp_list,
18220 &only_utf8_locale_list);
18222 if (_invlist_len(only_utf8_locale_list) == 0) {
18223 SvREFCNT_dec_NN(only_utf8_locale_list);
18224 only_utf8_locale_list = NULL;
18227 if ( only_utf8_locale_list
18228 || (cp_list && ( _invlist_contains_cp(cp_list, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE)
18229 || _invlist_contains_cp(cp_list, LATIN_SMALL_LETTER_DOTLESS_I))))
18231 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
18234 | ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
18236 else if (cp_list) { /* Look to see if a 0-255 code point is in list */
18238 invlist_iterinit(cp_list);
18239 if (invlist_iternext(cp_list, &start, &end) && start < 256) {
18240 anyof_flags |= ANYOFL_FOLD;
18241 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
18243 invlist_iterfinish(cp_list);
18246 else if ( DEPENDS_SEMANTICS
18247 && ( upper_latin1_only_utf8_matches
18248 || (anyof_flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)))
18250 RExC_seen_d_op = TRUE;
18251 has_runtime_dependency |= HAS_D_RUNTIME_DEPENDENCY;
18254 /* Optimize inverted patterns (e.g. [^a-z]) when everything is known at
18258 && ! has_runtime_dependency)
18260 _invlist_invert(cp_list);
18262 /* Clear the invert flag since have just done it here */
18267 *ret_invlist = cp_list;
18272 /* All possible optimizations below still have these characteristics.
18273 * (Multi-char folds aren't SIMPLE, but they don't get this far in this
18275 *flagp |= HASWIDTH|SIMPLE;
18277 if (anyof_flags & ANYOF_LOCALE_FLAGS) {
18278 RExC_contains_locale = 1;
18281 /* Some character classes are equivalent to other nodes. Such nodes take
18282 * up less room, and some nodes require fewer operations to execute, than
18283 * ANYOF nodes. EXACTish nodes may be joinable with adjacent nodes to
18284 * improve efficiency. */
18287 PERL_UINT_FAST8_T i;
18288 Size_t partial_cp_count = 0;
18289 UV start[MAX_FOLD_FROMS+1] = { 0 }; /* +1 for the folded-to char */
18290 UV end[MAX_FOLD_FROMS+1] = { 0 };
18292 if (cp_list) { /* Count the code points in enough ranges that we would
18293 see all the ones possible in any fold in this version
18296 invlist_iterinit(cp_list);
18297 for (i = 0; i <= MAX_FOLD_FROMS; i++) {
18298 if (! invlist_iternext(cp_list, &start[i], &end[i])) {
18301 partial_cp_count += end[i] - start[i] + 1;
18304 invlist_iterfinish(cp_list);
18307 /* If we know at compile time that this matches every possible code
18308 * point, any run-time dependencies don't matter */
18309 if (start[0] == 0 && end[0] == UV_MAX) {
18311 ret = reganode(pRExC_state, OPFAIL, 0);
18314 ret = reg_node(pRExC_state, SANY);
18320 /* Similarly, for /l posix classes, if both a class and its
18321 * complement match, any run-time dependencies don't matter */
18323 for (namedclass = 0; namedclass < ANYOF_POSIXL_MAX;
18326 if ( POSIXL_TEST(posixl, namedclass) /* class */
18327 && POSIXL_TEST(posixl, namedclass + 1)) /* its complement */
18330 ret = reganode(pRExC_state, OPFAIL, 0);
18333 ret = reg_node(pRExC_state, SANY);
18339 /* For well-behaved locales, some classes are subsets of others,
18340 * so complementing the subset and including the non-complemented
18341 * superset should match everything, like [\D[:alnum:]], and
18342 * [[:^alpha:][:alnum:]], but some implementations of locales are
18343 * buggy, and khw thinks its a bad idea to have optimization change
18344 * behavior, even if it avoids an OS bug in a given case */
18346 #define isSINGLE_BIT_SET(n) isPOWER_OF_2(n)
18348 /* If is a single posix /l class, can optimize to just that op.
18349 * Such a node will not match anything in the Latin1 range, as that
18350 * is not determinable until runtime, but will match whatever the
18351 * class does outside that range. (Note that some classes won't
18352 * match anything outside the range, like [:ascii:]) */
18353 if ( isSINGLE_BIT_SET(posixl)
18354 && (partial_cp_count == 0 || start[0] > 255))
18357 SV * class_above_latin1 = NULL;
18358 bool already_inverted;
18359 bool are_equivalent;
18361 /* Compute which bit is set, which is the same thing as, e.g.,
18362 * ANYOF_CNTRL. From
18363 * https://graphics.stanford.edu/~seander/bithacks.html#IntegerLogDeBruijn
18365 static const int MultiplyDeBruijnBitPosition2[32] =
18367 0, 1, 28, 2, 29, 14, 24, 3, 30, 22, 20, 15, 25, 17, 4, 8,
18368 31, 27, 13, 23, 21, 19, 16, 7, 26, 12, 18, 6, 11, 5, 10, 9
18371 namedclass = MultiplyDeBruijnBitPosition2[(posixl
18372 * 0x077CB531U) >> 27];
18373 classnum = namedclass_to_classnum(namedclass);
18375 /* The named classes are such that the inverted number is one
18376 * larger than the non-inverted one */
18377 already_inverted = namedclass
18378 - classnum_to_namedclass(classnum);
18380 /* Create an inversion list of the official property, inverted
18381 * if the constructed node list is inverted, and restricted to
18382 * only the above latin1 code points, which are the only ones
18383 * known at compile time */
18384 _invlist_intersection_maybe_complement_2nd(
18386 PL_XPosix_ptrs[classnum],
18388 &class_above_latin1);
18389 are_equivalent = _invlistEQ(class_above_latin1, cp_list,
18391 SvREFCNT_dec_NN(class_above_latin1);
18393 if (are_equivalent) {
18395 /* Resolve the run-time inversion flag with this possibly
18396 * inverted class */
18397 invert = invert ^ already_inverted;
18399 ret = reg_node(pRExC_state,
18400 POSIXL + invert * (NPOSIXL - POSIXL));
18401 FLAGS(REGNODE_p(ret)) = classnum;
18407 /* khw can't think of any other possible transformation involving
18409 if (has_runtime_dependency & HAS_USER_DEFINED_PROPERTY) {
18413 if (! has_runtime_dependency) {
18415 /* If the list is empty, nothing matches. This happens, for
18416 * example, when a Unicode property that doesn't match anything is
18417 * the only element in the character class (perluniprops.pod notes
18418 * such properties). */
18419 if (partial_cp_count == 0) {
18421 ret = reg_node(pRExC_state, SANY);
18424 ret = reganode(pRExC_state, OPFAIL, 0);
18430 /* If matches everything but \n */
18431 if ( start[0] == 0 && end[0] == '\n' - 1
18432 && start[1] == '\n' + 1 && end[1] == UV_MAX)
18435 ret = reg_node(pRExC_state, REG_ANY);
18441 /* Next see if can optimize classes that contain just a few code points
18442 * into an EXACTish node. The reason to do this is to let the
18443 * optimizer join this node with adjacent EXACTish ones.
18445 * An EXACTFish node can be generated even if not under /i, and vice
18446 * versa. But care must be taken. An EXACTFish node has to be such
18447 * that it only matches precisely the code points in the class, but we
18448 * want to generate the least restrictive one that does that, to
18449 * increase the odds of being able to join with an adjacent node. For
18450 * example, if the class contains [kK], we have to make it an EXACTFAA
18451 * node to prevent the KELVIN SIGN from matching. Whether we are under
18452 * /i or not is irrelevant in this case. Less obvious is the pattern
18453 * qr/[\x{02BC}]n/i. U+02BC is MODIFIER LETTER APOSTROPHE. That is
18454 * supposed to match the single character U+0149 LATIN SMALL LETTER N
18455 * PRECEDED BY APOSTROPHE. And so even though there is no simple fold
18456 * that includes \X{02BC}, there is a multi-char fold that does, and so
18457 * the node generated for it must be an EXACTFish one. On the other
18458 * hand qr/:/i should generate a plain EXACT node since the colon
18459 * participates in no fold whatsoever, and having it EXACT tells the
18460 * optimizer the target string cannot match unless it has a colon in
18463 * We don't typically generate an EXACTish node if doing so would
18464 * require changing the pattern to UTF-8, as that affects /d and
18465 * otherwise is slower. However, under /i, not changing to UTF-8 can
18466 * miss some potential multi-character folds. We calculate the
18467 * EXACTish node, and then decide if something would be missed if we
18472 /* Only try if there are no more code points in the class than
18473 * in the max possible fold */
18474 && partial_cp_count > 0 && partial_cp_count <= MAX_FOLD_FROMS + 1
18476 && (start[0] < 256 || UTF || FOLD))
18478 if (partial_cp_count == 1 && ! upper_latin1_only_utf8_matches)
18480 /* We can always make a single code point class into an
18481 * EXACTish node. */
18485 /* Here is /l: Use EXACTL, except /li indicates EXACTFL,
18486 * as that means there is a fold not known until runtime so
18487 * shows as only a single code point here. */
18488 op = (FOLD) ? EXACTFL : EXACTL;
18490 else if (! FOLD) { /* Not /l and not /i */
18491 op = (start[0] < 256) ? EXACT : EXACT_ONLY8;
18493 else if (start[0] < 256) { /* /i, not /l, and the code point is
18496 /* Under /i, it gets a little tricky. A code point that
18497 * doesn't participate in a fold should be an EXACT node.
18498 * We know this one isn't the result of a simple fold, or
18499 * there'd be more than one code point in the list, but it
18500 * could be part of a multi- character fold. In that case
18501 * we better not create an EXACT node, as we would wrongly
18502 * be telling the optimizer that this code point must be in
18503 * the target string, and that is wrong. This is because
18504 * if the sequence around this code point forms a
18505 * multi-char fold, what needs to be in the string could be
18506 * the code point that folds to the sequence.
18508 * This handles the case of below-255 code points, as we
18509 * have an easy look up for those. The next clause handles
18510 * the above-256 one */
18511 op = IS_IN_SOME_FOLD_L1(start[0])
18515 else { /* /i, larger code point. Since we are under /i, and
18516 have just this code point, we know that it can't
18517 fold to something else, so PL_InMultiCharFold
18519 op = _invlist_contains_cp(PL_InMultiCharFold,
18527 else if ( ! (has_runtime_dependency & ~HAS_D_RUNTIME_DEPENDENCY)
18528 && _invlist_contains_cp(PL_in_some_fold, start[0]))
18530 /* Here, the only runtime dependency, if any, is from /d, and
18531 * the class matches more than one code point, and the lowest
18532 * code point participates in some fold. It might be that the
18533 * other code points are /i equivalent to this one, and hence
18534 * they would representable by an EXACTFish node. Above, we
18535 * eliminated classes that contain too many code points to be
18536 * EXACTFish, with the test for MAX_FOLD_FROMS
18538 * First, special case the ASCII fold pairs, like 'B' and 'b'.
18539 * We do this because we have EXACTFAA at our disposal for the
18541 if (partial_cp_count == 2 && isASCII(start[0])) {
18543 /* The only ASCII characters that participate in folds are
18545 assert(isALPHA(start[0]));
18546 if ( end[0] == start[0] /* First range is a single
18547 character, so 2nd exists */
18548 && isALPHA_FOLD_EQ(start[0], start[1]))
18551 /* Here, is part of an ASCII fold pair */
18553 if ( ASCII_FOLD_RESTRICTED
18554 || HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(start[0]))
18556 /* If the second clause just above was true, it
18557 * means we can't be under /i, or else the list
18558 * would have included more than this fold pair.
18559 * Therefore we have to exclude the possibility of
18560 * whatever else it is that folds to these, by
18561 * using EXACTFAA */
18564 else if (HAS_NONLATIN1_FOLD_CLOSURE(start[0])) {
18566 /* Here, there's no simple fold that start[0] is part
18567 * of, but there is a multi-character one. If we
18568 * are not under /i, we want to exclude that
18569 * possibility; if under /i, we want to include it
18571 op = (FOLD) ? EXACTFU : EXACTFAA;
18575 /* Here, the only possible fold start[0] particpates in
18576 * is with start[1]. /i or not isn't relevant */
18580 value = toFOLD(start[0]);
18583 else if ( ! upper_latin1_only_utf8_matches
18584 || ( _invlist_len(upper_latin1_only_utf8_matches)
18587 invlist_highest(upper_latin1_only_utf8_matches)]
18590 /* Here, the smallest character is non-ascii or there are
18591 * more than 2 code points matched by this node. Also, we
18592 * either don't have /d UTF-8 dependent matches, or if we
18593 * do, they look like they could be a single character that
18594 * is the fold of the lowest one in the always-match list.
18595 * This test quickly excludes most of the false positives
18596 * when there are /d UTF-8 depdendent matches. These are
18597 * like LATIN CAPITAL LETTER A WITH GRAVE matching LATIN
18598 * SMALL LETTER A WITH GRAVE iff the target string is
18599 * UTF-8. (We don't have to worry above about exceeding
18600 * the array bounds of PL_fold_latin1[] because any code
18601 * point in 'upper_latin1_only_utf8_matches' is below 256.)
18603 * EXACTFAA would apply only to pairs (hence exactly 2 code
18604 * points) in the ASCII range, so we can't use it here to
18605 * artificially restrict the fold domain, so we check if
18606 * the class does or does not match some EXACTFish node.
18607 * Further, if we aren't under /i, and and the folded-to
18608 * character is part of a multi-character fold, we can't do
18609 * this optimization, as the sequence around it could be
18610 * that multi-character fold, and we don't here know the
18611 * context, so we have to assume it is that multi-char
18612 * fold, to prevent potential bugs.
18614 * To do the general case, we first find the fold of the
18615 * lowest code point (which may be higher than the lowest
18616 * one), then find everything that folds to it. (The data
18617 * structure we have only maps from the folded code points,
18618 * so we have to do the earlier step.) */
18621 U8 foldbuf[UTF8_MAXBYTES_CASE];
18622 UV folded = _to_uni_fold_flags(start[0],
18623 foldbuf, &foldlen, 0);
18624 unsigned int first_fold;
18625 const unsigned int * remaining_folds;
18626 Size_t folds_to_this_cp_count = _inverse_folds(
18630 Size_t folds_count = folds_to_this_cp_count + 1;
18631 SV * fold_list = _new_invlist(folds_count);
18634 /* If there are UTF-8 dependent matches, create a temporary
18635 * list of what this node matches, including them. */
18636 SV * all_cp_list = NULL;
18637 SV ** use_this_list = &cp_list;
18639 if (upper_latin1_only_utf8_matches) {
18640 all_cp_list = _new_invlist(0);
18641 use_this_list = &all_cp_list;
18642 _invlist_union(cp_list,
18643 upper_latin1_only_utf8_matches,
18647 /* Having gotten everything that participates in the fold
18648 * containing the lowest code point, we turn that into an
18649 * inversion list, making sure everything is included. */
18650 fold_list = add_cp_to_invlist(fold_list, start[0]);
18651 fold_list = add_cp_to_invlist(fold_list, folded);
18652 if (folds_to_this_cp_count > 0) {
18653 fold_list = add_cp_to_invlist(fold_list, first_fold);
18654 for (i = 0; i + 1 < folds_to_this_cp_count; i++) {
18655 fold_list = add_cp_to_invlist(fold_list,
18656 remaining_folds[i]);
18660 /* If the fold list is identical to what's in this ANYOF
18661 * node, the node can be represented by an EXACTFish one
18663 if (_invlistEQ(*use_this_list, fold_list,
18664 0 /* Don't complement */ )
18667 /* But, we have to be careful, as mentioned above.
18668 * Just the right sequence of characters could match
18669 * this if it is part of a multi-character fold. That
18670 * IS what we want if we are under /i. But it ISN'T
18671 * what we want if not under /i, as it could match when
18672 * it shouldn't. So, when we aren't under /i and this
18673 * character participates in a multi-char fold, we
18674 * don't optimize into an EXACTFish node. So, for each
18675 * case below we have to check if we are folding
18676 * and if not, if it is not part of a multi-char fold.
18678 if (start[0] > 255) { /* Highish code point */
18679 if (FOLD || ! _invlist_contains_cp(
18680 PL_InMultiCharFold, folded))
18684 : (ASCII_FOLD_RESTRICTED)
18689 } /* Below, the lowest code point < 256 */
18692 && DEPENDS_SEMANTICS)
18693 { /* An EXACTF node containing a single character
18694 's', can be an EXACTFU if it doesn't get
18695 joined with an adjacent 's' */
18696 op = EXACTFU_S_EDGE;
18700 || ! HAS_NONLATIN1_FOLD_CLOSURE(start[0]))
18702 if (upper_latin1_only_utf8_matches) {
18705 /* We can't use the fold, as that only matches
18709 else if ( UNLIKELY(start[0] == MICRO_SIGN)
18711 { /* EXACTFUP is a special node for this
18713 op = (ASCII_FOLD_RESTRICTED)
18716 value = MICRO_SIGN;
18718 else if ( ASCII_FOLD_RESTRICTED
18719 && ! isASCII(start[0]))
18720 { /* For ASCII under /iaa, we can use EXACTFU
18732 SvREFCNT_dec_NN(fold_list);
18733 SvREFCNT_dec(all_cp_list);
18739 /* Here, we have calculated what EXACTish node we would use.
18740 * But we don't use it if it would require converting the
18741 * pattern to UTF-8, unless not using it could cause us to miss
18742 * some folds (hence be buggy) */
18744 if (! UTF && value > 255) {
18745 SV * in_multis = NULL;
18749 /* If there is no code point that is part of a multi-char
18750 * fold, then there aren't any matches, so we don't do this
18751 * optimization. Otherwise, it could match depending on
18752 * the context around us, so we do upgrade */
18753 _invlist_intersection(PL_InMultiCharFold, cp_list, &in_multis);
18754 if (UNLIKELY(_invlist_len(in_multis) != 0)) {
18755 REQUIRE_UTF8(flagp);
18763 U8 len = (UTF) ? UVCHR_SKIP(value) : 1;
18765 ret = regnode_guts(pRExC_state, op, len, "exact");
18766 FILL_NODE(ret, op);
18767 RExC_emit += 1 + STR_SZ(len);
18768 STR_LEN(REGNODE_p(ret)) = len;
18770 *STRING(REGNODE_p(ret)) = (U8) value;
18773 uvchr_to_utf8((U8 *) STRING(REGNODE_p(ret)), value);
18780 if (! has_runtime_dependency) {
18782 /* See if this can be turned into an ANYOFM node. Think about the
18783 * bit patterns in two different bytes. In some positions, the
18784 * bits in each will be 1; and in other positions both will be 0;
18785 * and in some positions the bit will be 1 in one byte, and 0 in
18786 * the other. Let 'n' be the number of positions where the bits
18787 * differ. We create a mask which has exactly 'n' 0 bits, each in
18788 * a position where the two bytes differ. Now take the set of all
18789 * bytes that when ANDed with the mask yield the same result. That
18790 * set has 2**n elements, and is representable by just two 8 bit
18791 * numbers: the result and the mask. Importantly, matching the set
18792 * can be vectorized by creating a word full of the result bytes,
18793 * and a word full of the mask bytes, yielding a significant speed
18794 * up. Here, see if this node matches such a set. As a concrete
18795 * example consider [01], and the byte representing '0' which is
18796 * 0x30 on ASCII machines. It has the bits 0011 0000. Take the
18797 * mask 1111 1110. If we AND 0x31 and 0x30 with that mask we get
18798 * 0x30. Any other bytes ANDed yield something else. So [01],
18799 * which is a common usage, is optimizable into ANYOFM, and can
18800 * benefit from the speed up. We can only do this on UTF-8
18801 * invariant bytes, because they have the same bit patterns under
18803 PERL_UINT_FAST8_T inverted = 0;
18805 const PERL_UINT_FAST8_T max_permissible = 0xFF;
18807 const PERL_UINT_FAST8_T max_permissible = 0x7F;
18809 /* If doesn't fit the criteria for ANYOFM, invert and try again.
18810 * If that works we will instead later generate an NANYOFM, and
18811 * invert back when through */
18812 if (invlist_highest(cp_list) > max_permissible) {
18813 _invlist_invert(cp_list);
18817 if (invlist_highest(cp_list) <= max_permissible) {
18818 UV this_start, this_end;
18819 UV lowest_cp = UV_MAX; /* inited to suppress compiler warn */
18820 U8 bits_differing = 0;
18821 Size_t full_cp_count = 0;
18822 bool first_time = TRUE;
18824 /* Go through the bytes and find the bit positions that differ
18826 invlist_iterinit(cp_list);
18827 while (invlist_iternext(cp_list, &this_start, &this_end)) {
18828 unsigned int i = this_start;
18831 if (! UVCHR_IS_INVARIANT(i)) {
18835 first_time = FALSE;
18836 lowest_cp = this_start;
18838 /* We have set up the code point to compare with.
18839 * Don't compare it with itself */
18843 /* Find the bit positions that differ from the lowest code
18844 * point in the node. Keep track of all such positions by
18846 for (; i <= this_end; i++) {
18847 if (! UVCHR_IS_INVARIANT(i)) {
18851 bits_differing |= i ^ lowest_cp;
18854 full_cp_count += this_end - this_start + 1;
18856 invlist_iterfinish(cp_list);
18858 /* At the end of the loop, we count how many bits differ from
18859 * the bits in lowest code point, call the count 'd'. If the
18860 * set we found contains 2**d elements, it is the closure of
18861 * all code points that differ only in those bit positions. To
18862 * convince yourself of that, first note that the number in the
18863 * closure must be a power of 2, which we test for. The only
18864 * way we could have that count and it be some differing set,
18865 * is if we got some code points that don't differ from the
18866 * lowest code point in any position, but do differ from each
18867 * other in some other position. That means one code point has
18868 * a 1 in that position, and another has a 0. But that would
18869 * mean that one of them differs from the lowest code point in
18870 * that position, which possibility we've already excluded. */
18871 if ( (inverted || full_cp_count > 1)
18872 && full_cp_count == 1U << PL_bitcount[bits_differing])
18876 op = ANYOFM + inverted;;
18878 /* We need to make the bits that differ be 0's */
18879 ANYOFM_mask = ~ bits_differing; /* This goes into FLAGS */
18881 /* The argument is the lowest code point */
18882 ret = reganode(pRExC_state, op, lowest_cp);
18883 FLAGS(REGNODE_p(ret)) = ANYOFM_mask;
18889 _invlist_invert(cp_list);
18897 if (! (anyof_flags & ANYOF_LOCALE_FLAGS)) {
18898 PERL_UINT_FAST8_T type;
18899 SV * intersection = NULL;
18900 SV* d_invlist = NULL;
18902 /* See if this matches any of the POSIX classes. The POSIXA and
18903 * POSIXD ones are about the same speed as ANYOF ops, but take less
18904 * room; the ones that have above-Latin1 code point matches are
18905 * somewhat faster than ANYOF. */
18907 for (type = POSIXA; type >= POSIXD; type--) {
18910 if (type == POSIXL) { /* But not /l posix classes */
18914 for (posix_class = 0;
18915 posix_class <= _HIGHEST_REGCOMP_DOT_H_SYNC;
18918 SV** our_code_points = &cp_list;
18919 SV** official_code_points;
18922 if (type == POSIXA) {
18923 official_code_points = &PL_Posix_ptrs[posix_class];
18926 official_code_points = &PL_XPosix_ptrs[posix_class];
18929 /* Skip non-existent classes of this type. e.g. \v only
18930 * has an entry in PL_XPosix_ptrs */
18931 if (! *official_code_points) {
18935 /* Try both the regular class, and its inversion */
18936 for (try_inverted = 0; try_inverted < 2; try_inverted++) {
18937 bool this_inverted = invert ^ try_inverted;
18939 if (type != POSIXD) {
18941 /* This class that isn't /d can't match if we have
18942 * /d dependencies */
18943 if (has_runtime_dependency
18944 & HAS_D_RUNTIME_DEPENDENCY)
18949 else /* is /d */ if (! this_inverted) {
18951 /* /d classes don't match anything non-ASCII below
18952 * 256 unconditionally (which cp_list contains) */
18953 _invlist_intersection(cp_list, PL_UpperLatin1,
18955 if (_invlist_len(intersection) != 0) {
18959 SvREFCNT_dec(d_invlist);
18960 d_invlist = invlist_clone(cp_list, NULL);
18962 /* But under UTF-8 it turns into using /u rules.
18963 * Add the things it matches under these conditions
18964 * so that we check below that these are identical
18965 * to what the tested class should match */
18966 if (upper_latin1_only_utf8_matches) {
18969 upper_latin1_only_utf8_matches,
18972 our_code_points = &d_invlist;
18974 else { /* POSIXD, inverted. If this doesn't have this
18975 flag set, it isn't /d. */
18976 if (! (anyof_flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
18980 our_code_points = &cp_list;
18983 /* Here, have weeded out some things. We want to see
18984 * if the list of characters this node contains
18985 * ('*our_code_points') precisely matches those of the
18986 * class we are currently checking against
18987 * ('*official_code_points'). */
18988 if (_invlistEQ(*our_code_points,
18989 *official_code_points,
18992 /* Here, they precisely match. Optimize this ANYOF
18993 * node into its equivalent POSIX one of the
18994 * correct type, possibly inverted */
18995 ret = reg_node(pRExC_state, (try_inverted)
18999 FLAGS(REGNODE_p(ret)) = posix_class;
19000 SvREFCNT_dec(d_invlist);
19001 SvREFCNT_dec(intersection);
19007 SvREFCNT_dec(d_invlist);
19008 SvREFCNT_dec(intersection);
19011 /* If didn't find an optimization and there is no need for a
19012 * bitmap, optimize to indicate that */
19013 if ( start[0] >= NUM_ANYOF_CODE_POINTS
19015 && ! upper_latin1_only_utf8_matches
19016 && anyof_flags == 0)
19018 UV highest_cp = invlist_highest(cp_list);
19020 /* If the lowest and highest code point in the class have the same
19021 * UTF-8 first byte, then all do, and we can store that byte for
19022 * regexec.c to use so that it can more quickly scan the target
19023 * string for potential matches for this class. We co-opt the the
19024 * flags field for this. Zero means, they don't have the same
19025 * first byte. We do accept here very large code points (for
19026 * future use), but don't bother with this optimization for them,
19027 * as it would cause other complications */
19028 if (highest_cp > IV_MAX) {
19032 U8 low_utf8[UTF8_MAXBYTES+1];
19033 U8 high_utf8[UTF8_MAXBYTES+1];
19035 (void) uvchr_to_utf8(low_utf8, start[0]);
19036 (void) uvchr_to_utf8(high_utf8, invlist_highest(cp_list));
19038 anyof_flags = (low_utf8[0] == high_utf8[0])
19045 } /* End of seeing if can optimize it into a different node */
19047 is_anyof: /* It's going to be an ANYOF node. */
19048 if (op != ANYOFH) {
19049 op = (has_runtime_dependency & HAS_D_RUNTIME_DEPENDENCY)
19058 ret = regnode_guts(pRExC_state, op, regarglen[op], "anyof");
19059 FILL_NODE(ret, op); /* We set the argument later */
19060 RExC_emit += 1 + regarglen[op];
19061 ANYOF_FLAGS(REGNODE_p(ret)) = anyof_flags;
19063 /* Here, <cp_list> contains all the code points we can determine at
19064 * compile time that match under all conditions. Go through it, and
19065 * for things that belong in the bitmap, put them there, and delete from
19066 * <cp_list>. While we are at it, see if everything above 255 is in the
19067 * list, and if so, set a flag to speed up execution */
19069 populate_ANYOF_from_invlist(REGNODE_p(ret), &cp_list);
19072 ANYOF_POSIXL_SET_TO_BITMAP(REGNODE_p(ret), posixl);
19076 ANYOF_FLAGS(REGNODE_p(ret)) |= ANYOF_INVERT;
19079 /* Here, the bitmap has been populated with all the Latin1 code points that
19080 * always match. Can now add to the overall list those that match only
19081 * when the target string is UTF-8 (<upper_latin1_only_utf8_matches>).
19083 if (upper_latin1_only_utf8_matches) {
19085 _invlist_union(cp_list,
19086 upper_latin1_only_utf8_matches,
19088 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
19091 cp_list = upper_latin1_only_utf8_matches;
19093 ANYOF_FLAGS(REGNODE_p(ret)) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
19096 set_ANYOF_arg(pRExC_state, REGNODE_p(ret), cp_list,
19097 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
19099 only_utf8_locale_list);
19104 /* Here, the node is getting optimized into something that's not an ANYOF
19105 * one. Finish up. */
19107 Set_Node_Offset_Length(REGNODE_p(ret), orig_parse - RExC_start,
19108 RExC_parse - orig_parse);;
19109 SvREFCNT_dec(cp_list);;
19113 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
19116 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
19117 regnode* const node,
19119 SV* const runtime_defns,
19120 SV* const only_utf8_locale_list)
19122 /* Sets the arg field of an ANYOF-type node 'node', using information about
19123 * the node passed-in. If there is nothing outside the node's bitmap, the
19124 * arg is set to ANYOF_ONLY_HAS_BITMAP. Otherwise, it sets the argument to
19125 * the count returned by add_data(), having allocated and stored an array,
19128 * av[0] stores the inversion list defining this class as far as known at
19129 * this time, or PL_sv_undef if nothing definite is now known.
19130 * av[1] stores the inversion list of code points that match only if the
19131 * current locale is UTF-8, or if none, PL_sv_undef if there is an
19132 * av[2], or no entry otherwise.
19133 * av[2] stores the list of user-defined properties whose subroutine
19134 * definitions aren't known at this time, or no entry if none. */
19138 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
19140 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
19141 assert(! (ANYOF_FLAGS(node)
19142 & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP));
19143 ARG_SET(node, ANYOF_ONLY_HAS_BITMAP);
19146 AV * const av = newAV();
19150 av_store(av, INVLIST_INDEX, cp_list);
19153 if (only_utf8_locale_list) {
19154 av_store(av, ONLY_LOCALE_MATCHES_INDEX, only_utf8_locale_list);
19157 if (runtime_defns) {
19158 av_store(av, DEFERRED_USER_DEFINED_INDEX, SvREFCNT_inc(runtime_defns));
19161 rv = newRV_noinc(MUTABLE_SV(av));
19162 n = add_data(pRExC_state, STR_WITH_LEN("s"));
19163 RExC_rxi->data->data[n] = (void*)rv;
19168 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
19170 Perl__get_regclass_nonbitmap_data(pTHX_ const regexp *prog,
19171 const regnode* node,
19174 SV** only_utf8_locale_ptr,
19175 SV** output_invlist)
19178 /* For internal core use only.
19179 * Returns the inversion list for the input 'node' in the regex 'prog'.
19180 * If <doinit> is 'true', will attempt to create the inversion list if not
19182 * If <listsvp> is non-null, will return the printable contents of the
19183 * property definition. This can be used to get debugging information
19184 * even before the inversion list exists, by calling this function with
19185 * 'doinit' set to false, in which case the components that will be used
19186 * to eventually create the inversion list are returned (in a printable
19188 * If <only_utf8_locale_ptr> is not NULL, it is where this routine is to
19189 * store an inversion list of code points that should match only if the
19190 * execution-time locale is a UTF-8 one.
19191 * If <output_invlist> is not NULL, it is where this routine is to store an
19192 * inversion list of the code points that would be instead returned in
19193 * <listsvp> if this were NULL. Thus, what gets output in <listsvp>
19194 * when this parameter is used, is just the non-code point data that
19195 * will go into creating the inversion list. This currently should be just
19196 * user-defined properties whose definitions were not known at compile
19197 * time. Using this parameter allows for easier manipulation of the
19198 * inversion list's data by the caller. It is illegal to call this
19199 * function with this parameter set, but not <listsvp>
19201 * Tied intimately to how S_set_ANYOF_arg sets up the data structure. Note
19202 * that, in spite of this function's name, the inversion list it returns
19203 * may include the bitmap data as well */
19205 SV *si = NULL; /* Input initialization string */
19206 SV* invlist = NULL;
19208 RXi_GET_DECL(prog, progi);
19209 const struct reg_data * const data = prog ? progi->data : NULL;
19211 PERL_ARGS_ASSERT__GET_REGCLASS_NONBITMAP_DATA;
19212 assert(! output_invlist || listsvp);
19214 if (data && data->count) {
19215 const U32 n = ARG(node);
19217 if (data->what[n] == 's') {
19218 SV * const rv = MUTABLE_SV(data->data[n]);
19219 AV * const av = MUTABLE_AV(SvRV(rv));
19220 SV **const ary = AvARRAY(av);
19222 invlist = ary[INVLIST_INDEX];
19224 if (av_tindex_skip_len_mg(av) >= ONLY_LOCALE_MATCHES_INDEX) {
19225 *only_utf8_locale_ptr = ary[ONLY_LOCALE_MATCHES_INDEX];
19228 if (av_tindex_skip_len_mg(av) >= DEFERRED_USER_DEFINED_INDEX) {
19229 si = ary[DEFERRED_USER_DEFINED_INDEX];
19232 if (doinit && (si || invlist)) {
19235 SV * msg = newSVpvs_flags("", SVs_TEMP);
19237 SV * prop_definition = handle_user_defined_property(
19238 "", 0, FALSE, /* There is no \p{}, \P{} */
19239 SvPVX_const(si)[1] - '0', /* /i or not has been
19240 stored here for just
19242 TRUE, /* run time */
19243 FALSE, /* This call must find the defn */
19244 si, /* The property definition */
19247 0 /* base level call */
19251 assert(prop_definition == NULL);
19253 Perl_croak(aTHX_ "%" UTF8f,
19254 UTF8fARG(SvUTF8(msg), SvCUR(msg), SvPVX(msg)));
19258 _invlist_union(invlist, prop_definition, &invlist);
19259 SvREFCNT_dec_NN(prop_definition);
19262 invlist = prop_definition;
19265 STATIC_ASSERT_STMT(ONLY_LOCALE_MATCHES_INDEX == 1 + INVLIST_INDEX);
19266 STATIC_ASSERT_STMT(DEFERRED_USER_DEFINED_INDEX == 1 + ONLY_LOCALE_MATCHES_INDEX);
19268 av_store(av, INVLIST_INDEX, invlist);
19269 av_fill(av, (ary[ONLY_LOCALE_MATCHES_INDEX])
19270 ? ONLY_LOCALE_MATCHES_INDEX:
19278 /* If requested, return a printable version of what this ANYOF node matches
19281 SV* matches_string = NULL;
19283 /* This function can be called at compile-time, before everything gets
19284 * resolved, in which case we return the currently best available
19285 * information, which is the string that will eventually be used to do
19286 * that resolving, 'si' */
19288 /* Here, we only have 'si' (and possibly some passed-in data in
19289 * 'invlist', which is handled below) If the caller only wants
19290 * 'si', use that. */
19291 if (! output_invlist) {
19292 matches_string = newSVsv(si);
19295 /* But if the caller wants an inversion list of the node, we
19296 * need to parse 'si' and place as much as possible in the
19297 * desired output inversion list, making 'matches_string' only
19298 * contain the currently unresolvable things */
19299 const char *si_string = SvPVX(si);
19300 STRLEN remaining = SvCUR(si);
19304 /* Ignore everything before the first new-line */
19305 while (*si_string != '\n' && remaining > 0) {
19309 assert(remaining > 0);
19314 while (remaining > 0) {
19316 /* The data consists of just strings defining user-defined
19317 * property names, but in prior incarnations, and perhaps
19318 * somehow from pluggable regex engines, it could still
19319 * hold hex code point definitions. Each component of a
19320 * range would be separated by a tab, and each range by a
19321 * new-line. If these are found, instead add them to the
19322 * inversion list */
19323 I32 grok_flags = PERL_SCAN_SILENT_ILLDIGIT
19324 |PERL_SCAN_SILENT_NON_PORTABLE;
19325 STRLEN len = remaining;
19326 UV cp = grok_hex(si_string, &len, &grok_flags, NULL);
19328 /* If the hex decode routine found something, it should go
19329 * up to the next \n */
19330 if ( *(si_string + len) == '\n') {
19331 if (count) { /* 2nd code point on line */
19332 *output_invlist = _add_range_to_invlist(*output_invlist, prev_cp, cp);
19335 *output_invlist = add_cp_to_invlist(*output_invlist, cp);
19338 goto prepare_for_next_iteration;
19341 /* If the hex decode was instead for the lower range limit,
19342 * save it, and go parse the upper range limit */
19343 if (*(si_string + len) == '\t') {
19344 assert(count == 0);
19348 prepare_for_next_iteration:
19349 si_string += len + 1;
19350 remaining -= len + 1;
19354 /* Here, didn't find a legal hex number. Just add it from
19355 * here to the next \n */
19358 while (*(si_string + len) != '\n' && remaining > 0) {
19362 if (*(si_string + len) == '\n') {
19366 if (matches_string) {
19367 sv_catpvn(matches_string, si_string, len - 1);
19370 matches_string = newSVpvn(si_string, len - 1);
19373 sv_catpvs(matches_string, " ");
19374 } /* end of loop through the text */
19376 assert(matches_string);
19377 if (SvCUR(matches_string)) { /* Get rid of trailing blank */
19378 SvCUR_set(matches_string, SvCUR(matches_string) - 1);
19380 } /* end of has an 'si' */
19383 /* Add the stuff that's already known */
19386 /* Again, if the caller doesn't want the output inversion list, put
19387 * everything in 'matches-string' */
19388 if (! output_invlist) {
19389 if ( ! matches_string) {
19390 matches_string = newSVpvs("\n");
19392 sv_catsv(matches_string, invlist_contents(invlist,
19393 TRUE /* traditional style */
19396 else if (! *output_invlist) {
19397 *output_invlist = invlist_clone(invlist, NULL);
19400 _invlist_union(*output_invlist, invlist, output_invlist);
19404 *listsvp = matches_string;
19409 #endif /* !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION) */
19411 /* reg_skipcomment()
19413 Absorbs an /x style # comment from the input stream,
19414 returning a pointer to the first character beyond the comment, or if the
19415 comment terminates the pattern without anything following it, this returns
19416 one past the final character of the pattern (in other words, RExC_end) and
19417 sets the REG_RUN_ON_COMMENT_SEEN flag.
19419 Note it's the callers responsibility to ensure that we are
19420 actually in /x mode
19424 PERL_STATIC_INLINE char*
19425 S_reg_skipcomment(RExC_state_t *pRExC_state, char* p)
19427 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
19431 while (p < RExC_end) {
19432 if (*(++p) == '\n') {
19437 /* we ran off the end of the pattern without ending the comment, so we have
19438 * to add an \n when wrapping */
19439 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
19444 S_skip_to_be_ignored_text(pTHX_ RExC_state_t *pRExC_state,
19446 const bool force_to_xmod
19449 /* If the text at the current parse position '*p' is a '(?#...)' comment,
19450 * or if we are under /x or 'force_to_xmod' is TRUE, and the text at '*p'
19451 * is /x whitespace, advance '*p' so that on exit it points to the first
19452 * byte past all such white space and comments */
19454 const bool use_xmod = force_to_xmod || (RExC_flags & RXf_PMf_EXTENDED);
19456 PERL_ARGS_ASSERT_SKIP_TO_BE_IGNORED_TEXT;
19458 assert( ! UTF || UTF8_IS_INVARIANT(**p) || UTF8_IS_START(**p));
19461 if (RExC_end - (*p) >= 3
19463 && *(*p + 1) == '?'
19464 && *(*p + 2) == '#')
19466 while (*(*p) != ')') {
19467 if ((*p) == RExC_end)
19468 FAIL("Sequence (?#... not terminated");
19476 const char * save_p = *p;
19477 while ((*p) < RExC_end) {
19479 if ((len = is_PATWS_safe((*p), RExC_end, UTF))) {
19482 else if (*(*p) == '#') {
19483 (*p) = reg_skipcomment(pRExC_state, (*p));
19489 if (*p != save_p) {
19502 Advances the parse position by one byte, unless that byte is the beginning
19503 of a '(?#...)' style comment, or is /x whitespace and /x is in effect. In
19504 those two cases, the parse position is advanced beyond all such comments and
19507 This is the UTF, (?#...), and /x friendly way of saying RExC_parse++.
19511 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
19513 PERL_ARGS_ASSERT_NEXTCHAR;
19515 if (RExC_parse < RExC_end) {
19517 || UTF8_IS_INVARIANT(*RExC_parse)
19518 || UTF8_IS_START(*RExC_parse));
19520 RExC_parse += (UTF)
19521 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
19524 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
19525 FALSE /* Don't force /x */ );
19530 S_change_engine_size(pTHX_ RExC_state_t *pRExC_state, const Ptrdiff_t size)
19532 /* 'size' is the delta to add or subtract from the current memory allocated
19533 * to the regex engine being constructed */
19535 PERL_ARGS_ASSERT_CHANGE_ENGINE_SIZE;
19540 sizeof(regexp_internal) + (RExC_size + 1) * sizeof(regnode),
19541 /* +1 for REG_MAGIC */
19544 if ( RExC_rxi == NULL )
19545 FAIL("Regexp out of space");
19546 RXi_SET(RExC_rx, RExC_rxi);
19548 RExC_emit_start = RExC_rxi->program;
19550 Zero(REGNODE_p(RExC_emit), size, regnode);
19553 #ifdef RE_TRACK_PATTERN_OFFSETS
19554 Renew(RExC_offsets, 2*RExC_size+1, U32);
19556 Zero(RExC_offsets + 2*(RExC_size - size) + 1, 2 * size, U32);
19558 RExC_offsets[0] = RExC_size;
19562 STATIC regnode_offset
19563 S_regnode_guts(pTHX_ RExC_state_t *pRExC_state, const U8 op, const STRLEN extra_size, const char* const name)
19565 /* Allocate a regnode for 'op', with 'extra_size' extra space. It aligns
19566 * and increments RExC_size and RExC_emit
19568 * It returns the regnode's offset into the regex engine program */
19570 const regnode_offset ret = RExC_emit;
19572 GET_RE_DEBUG_FLAGS_DECL;
19574 PERL_ARGS_ASSERT_REGNODE_GUTS;
19576 SIZE_ALIGN(RExC_size);
19577 change_engine_size(pRExC_state, (Ptrdiff_t) 1 + extra_size);
19578 NODE_ALIGN_FILL(REGNODE_p(ret));
19579 #ifndef RE_TRACK_PATTERN_OFFSETS
19580 PERL_UNUSED_ARG(name);
19581 PERL_UNUSED_ARG(op);
19583 assert(extra_size >= regarglen[op] || PL_regkind[op] == ANYOF);
19585 if (RExC_offsets) { /* MJD */
19587 ("%s:%d: (op %s) %s %" UVuf " (len %" UVuf ") (max %" UVuf ").\n",
19590 (UV)(RExC_emit) > RExC_offsets[0]
19591 ? "Overwriting end of array!\n" : "OK",
19593 (UV)(RExC_parse - RExC_start),
19594 (UV)RExC_offsets[0]));
19595 Set_Node_Offset(REGNODE_p(RExC_emit), RExC_parse + (op == END));
19602 - reg_node - emit a node
19604 STATIC regnode_offset /* Location. */
19605 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
19607 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reg_node");
19608 regnode_offset ptr = ret;
19610 PERL_ARGS_ASSERT_REG_NODE;
19612 assert(regarglen[op] == 0);
19614 FILL_ADVANCE_NODE(ptr, op);
19620 - reganode - emit a node with an argument
19622 STATIC regnode_offset /* Location. */
19623 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
19625 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reganode");
19626 regnode_offset ptr = ret;
19628 PERL_ARGS_ASSERT_REGANODE;
19630 /* ANYOF are special cased to allow non-length 1 args */
19631 assert(regarglen[op] == 1);
19633 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
19638 STATIC regnode_offset
19639 S_reg2Lanode(pTHX_ RExC_state_t *pRExC_state, const U8 op, const U32 arg1, const I32 arg2)
19641 /* emit a node with U32 and I32 arguments */
19643 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reg2Lanode");
19644 regnode_offset ptr = ret;
19646 PERL_ARGS_ASSERT_REG2LANODE;
19648 assert(regarglen[op] == 2);
19650 FILL_ADVANCE_NODE_2L_ARG(ptr, op, arg1, arg2);
19656 - reginsert - insert an operator in front of already-emitted operand
19658 * That means that on exit 'operand' is the offset of the newly inserted
19659 * operator, and the original operand has been relocated.
19661 * IMPORTANT NOTE - it is the *callers* responsibility to correctly
19662 * set up NEXT_OFF() of the inserted node if needed. Something like this:
19664 * reginsert(pRExC, OPFAIL, orig_emit, depth+1);
19665 * NEXT_OFF(orig_emit) = regarglen[OPFAIL] + NODE_STEP_REGNODE;
19667 * ALSO NOTE - FLAGS(newly-inserted-operator) will be set to 0 as well.
19670 S_reginsert(pTHX_ RExC_state_t *pRExC_state, const U8 op,
19671 const regnode_offset operand, const U32 depth)
19676 const int offset = regarglen[(U8)op];
19677 const int size = NODE_STEP_REGNODE + offset;
19678 GET_RE_DEBUG_FLAGS_DECL;
19680 PERL_ARGS_ASSERT_REGINSERT;
19681 PERL_UNUSED_CONTEXT;
19682 PERL_UNUSED_ARG(depth);
19683 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
19684 DEBUG_PARSE_FMT("inst"," - %s", PL_reg_name[op]);
19685 assert(!RExC_study_started); /* I believe we should never use reginsert once we have started
19686 studying. If this is wrong then we need to adjust RExC_recurse
19687 below like we do with RExC_open_parens/RExC_close_parens. */
19688 change_engine_size(pRExC_state, (Ptrdiff_t) size);
19689 src = REGNODE_p(RExC_emit);
19691 dst = REGNODE_p(RExC_emit);
19693 /* If we are in a "count the parentheses" pass, the numbers are unreliable,
19694 * and [perl #133871] shows this can lead to problems, so skip this
19695 * realignment of parens until a later pass when they are reliable */
19696 if (! IN_PARENS_PASS && RExC_open_parens) {
19698 /*DEBUG_PARSE_FMT("inst"," - %" IVdf, (IV)RExC_npar);*/
19699 /* remember that RExC_npar is rex->nparens + 1,
19700 * iow it is 1 more than the number of parens seen in
19701 * the pattern so far. */
19702 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
19703 /* note, RExC_open_parens[0] is the start of the
19704 * regex, it can't move. RExC_close_parens[0] is the end
19705 * of the regex, it *can* move. */
19706 if ( paren && RExC_open_parens[paren] >= operand ) {
19707 /*DEBUG_PARSE_FMT("open"," - %d", size);*/
19708 RExC_open_parens[paren] += size;
19710 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
19712 if ( RExC_close_parens[paren] >= operand ) {
19713 /*DEBUG_PARSE_FMT("close"," - %d", size);*/
19714 RExC_close_parens[paren] += size;
19716 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
19721 RExC_end_op += size;
19723 while (src > REGNODE_p(operand)) {
19724 StructCopy(--src, --dst, regnode);
19725 #ifdef RE_TRACK_PATTERN_OFFSETS
19726 if (RExC_offsets) { /* MJD 20010112 */
19728 ("%s(%d): (op %s) %s copy %" UVuf " -> %" UVuf " (max %" UVuf ").\n",
19732 (UV)(REGNODE_OFFSET(dst)) > RExC_offsets[0]
19733 ? "Overwriting end of array!\n" : "OK",
19734 (UV)REGNODE_OFFSET(src),
19735 (UV)REGNODE_OFFSET(dst),
19736 (UV)RExC_offsets[0]));
19737 Set_Node_Offset_To_R(REGNODE_OFFSET(dst), Node_Offset(src));
19738 Set_Node_Length_To_R(REGNODE_OFFSET(dst), Node_Length(src));
19743 place = REGNODE_p(operand); /* Op node, where operand used to be. */
19744 #ifdef RE_TRACK_PATTERN_OFFSETS
19745 if (RExC_offsets) { /* MJD */
19747 ("%s(%d): (op %s) %s %" UVuf " <- %" UVuf " (max %" UVuf ").\n",
19751 (UV)REGNODE_OFFSET(place) > RExC_offsets[0]
19752 ? "Overwriting end of array!\n" : "OK",
19753 (UV)REGNODE_OFFSET(place),
19754 (UV)(RExC_parse - RExC_start),
19755 (UV)RExC_offsets[0]));
19756 Set_Node_Offset(place, RExC_parse);
19757 Set_Node_Length(place, 1);
19760 src = NEXTOPER(place);
19762 FILL_NODE(operand, op);
19764 /* Zero out any arguments in the new node */
19765 Zero(src, offset, regnode);
19769 - regtail - set the next-pointer at the end of a node chain of p to val. If
19770 that value won't fit in the space available, instead returns FALSE.
19771 (Except asserts if we can't fit in the largest space the regex
19772 engine is designed for.)
19773 - SEE ALSO: regtail_study
19776 S_regtail(pTHX_ RExC_state_t * pRExC_state,
19777 const regnode_offset p,
19778 const regnode_offset val,
19781 regnode_offset scan;
19782 GET_RE_DEBUG_FLAGS_DECL;
19784 PERL_ARGS_ASSERT_REGTAIL;
19786 PERL_UNUSED_ARG(depth);
19789 /* Find last node. */
19790 scan = (regnode_offset) p;
19792 regnode * const temp = regnext(REGNODE_p(scan));
19794 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
19795 regprop(RExC_rx, RExC_mysv, REGNODE_p(scan), NULL, pRExC_state);
19796 Perl_re_printf( aTHX_ "~ %s (%d) %s %s\n",
19797 SvPV_nolen_const(RExC_mysv), scan,
19798 (temp == NULL ? "->" : ""),
19799 (temp == NULL ? PL_reg_name[OP(REGNODE_p(val))] : "")
19804 scan = REGNODE_OFFSET(temp);
19807 if (reg_off_by_arg[OP(REGNODE_p(scan))]) {
19808 assert((UV) (val - scan) <= U32_MAX);
19809 ARG_SET(REGNODE_p(scan), val - scan);
19812 if (val - scan > U16_MAX) {
19813 /* Since not all callers check the return value, populate this with
19814 * something that won't loop and will likely lead to a crash if
19815 * execution continues */
19816 NEXT_OFF(REGNODE_p(scan)) = U16_MAX;
19819 NEXT_OFF(REGNODE_p(scan)) = val - scan;
19827 - regtail_study - set the next-pointer at the end of a node chain of p to val.
19828 - Look for optimizable sequences at the same time.
19829 - currently only looks for EXACT chains.
19831 This is experimental code. The idea is to use this routine to perform
19832 in place optimizations on branches and groups as they are constructed,
19833 with the long term intention of removing optimization from study_chunk so
19834 that it is purely analytical.
19836 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
19837 to control which is which.
19839 This used to return a value that was ignored. It was a problem that it is
19840 #ifdef'd to be another function that didn't return a value. khw has changed it
19841 so both currently return a pass/fail return.
19844 /* TODO: All four parms should be const */
19847 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode_offset p,
19848 const regnode_offset val, U32 depth)
19850 regnode_offset scan;
19852 #ifdef EXPERIMENTAL_INPLACESCAN
19855 GET_RE_DEBUG_FLAGS_DECL;
19857 PERL_ARGS_ASSERT_REGTAIL_STUDY;
19860 /* Find last node. */
19864 regnode * const temp = regnext(REGNODE_p(scan));
19865 #ifdef EXPERIMENTAL_INPLACESCAN
19866 if (PL_regkind[OP(REGNODE_p(scan))] == EXACT) {
19867 bool unfolded_multi_char; /* Unexamined in this routine */
19868 if (join_exact(pRExC_state, scan, &min,
19869 &unfolded_multi_char, 1, REGNODE_p(val), depth+1))
19870 return TRUE; /* Was return EXACT */
19874 switch (OP(REGNODE_p(scan))) {
19879 case EXACTFU_S_EDGE:
19880 case EXACTFAA_NO_TRIE:
19883 case EXACTFU_ONLY8:
19887 if( exact == PSEUDO )
19888 exact= OP(REGNODE_p(scan));
19889 else if ( exact != OP(REGNODE_p(scan)) )
19898 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
19899 regprop(RExC_rx, RExC_mysv, REGNODE_p(scan), NULL, pRExC_state);
19900 Perl_re_printf( aTHX_ "~ %s (%d) -> %s\n",
19901 SvPV_nolen_const(RExC_mysv),
19903 PL_reg_name[exact]);
19907 scan = REGNODE_OFFSET(temp);
19910 DEBUG_PARSE_MSG("");
19911 regprop(RExC_rx, RExC_mysv, REGNODE_p(val), NULL, pRExC_state);
19912 Perl_re_printf( aTHX_
19913 "~ attach to %s (%" IVdf ") offset to %" IVdf "\n",
19914 SvPV_nolen_const(RExC_mysv),
19919 if (reg_off_by_arg[OP(REGNODE_p(scan))]) {
19920 assert((UV) (val - scan) <= U32_MAX);
19921 ARG_SET(REGNODE_p(scan), val - scan);
19924 if (val - scan > U16_MAX) {
19925 NEXT_OFF(REGNODE_p(scan)) = U16_MAX;
19928 NEXT_OFF(REGNODE_p(scan)) = val - scan;
19931 return TRUE; /* Was 'return exact' */
19936 S_get_ANYOFM_contents(pTHX_ const regnode * n) {
19938 /* Returns an inversion list of all the code points matched by the
19939 * ANYOFM/NANYOFM node 'n' */
19941 SV * cp_list = _new_invlist(-1);
19942 const U8 lowest = (U8) ARG(n);
19945 U8 needed = 1U << PL_bitcount[ (U8) ~ FLAGS(n)];
19947 PERL_ARGS_ASSERT_GET_ANYOFM_CONTENTS;
19949 /* Starting with the lowest code point, any code point that ANDed with the
19950 * mask yields the lowest code point is in the set */
19951 for (i = lowest; i <= 0xFF; i++) {
19952 if ((i & FLAGS(n)) == ARG(n)) {
19953 cp_list = add_cp_to_invlist(cp_list, i);
19956 /* We know how many code points (a power of two) that are in the
19957 * set. No use looking once we've got that number */
19958 if (count >= needed) break;
19962 if (OP(n) == NANYOFM) {
19963 _invlist_invert(cp_list);
19969 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
19974 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
19979 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
19981 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
19982 if (flags & (1<<bit)) {
19983 if (!set++ && lead)
19984 Perl_re_printf( aTHX_ "%s", lead);
19985 Perl_re_printf( aTHX_ "%s ", PL_reg_intflags_name[bit]);
19990 Perl_re_printf( aTHX_ "\n");
19992 Perl_re_printf( aTHX_ "%s[none-set]\n", lead);
19997 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
20003 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
20005 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
20006 if (flags & (1<<bit)) {
20007 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
20010 if (!set++ && lead)
20011 Perl_re_printf( aTHX_ "%s", lead);
20012 Perl_re_printf( aTHX_ "%s ", PL_reg_extflags_name[bit]);
20015 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
20016 if (!set++ && lead) {
20017 Perl_re_printf( aTHX_ "%s", lead);
20020 case REGEX_UNICODE_CHARSET:
20021 Perl_re_printf( aTHX_ "UNICODE");
20023 case REGEX_LOCALE_CHARSET:
20024 Perl_re_printf( aTHX_ "LOCALE");
20026 case REGEX_ASCII_RESTRICTED_CHARSET:
20027 Perl_re_printf( aTHX_ "ASCII-RESTRICTED");
20029 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
20030 Perl_re_printf( aTHX_ "ASCII-MORE_RESTRICTED");
20033 Perl_re_printf( aTHX_ "UNKNOWN CHARACTER SET");
20039 Perl_re_printf( aTHX_ "\n");
20041 Perl_re_printf( aTHX_ "%s[none-set]\n", lead);
20047 Perl_regdump(pTHX_ const regexp *r)
20051 SV * const sv = sv_newmortal();
20052 SV *dsv= sv_newmortal();
20053 RXi_GET_DECL(r, ri);
20054 GET_RE_DEBUG_FLAGS_DECL;
20056 PERL_ARGS_ASSERT_REGDUMP;
20058 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
20060 /* Header fields of interest. */
20061 for (i = 0; i < 2; i++) {
20062 if (r->substrs->data[i].substr) {
20063 RE_PV_QUOTED_DECL(s, 0, dsv,
20064 SvPVX_const(r->substrs->data[i].substr),
20065 RE_SV_DUMPLEN(r->substrs->data[i].substr),
20066 PL_dump_re_max_len);
20067 Perl_re_printf( aTHX_
20068 "%s %s%s at %" IVdf "..%" UVuf " ",
20069 i ? "floating" : "anchored",
20071 RE_SV_TAIL(r->substrs->data[i].substr),
20072 (IV)r->substrs->data[i].min_offset,
20073 (UV)r->substrs->data[i].max_offset);
20075 else if (r->substrs->data[i].utf8_substr) {
20076 RE_PV_QUOTED_DECL(s, 1, dsv,
20077 SvPVX_const(r->substrs->data[i].utf8_substr),
20078 RE_SV_DUMPLEN(r->substrs->data[i].utf8_substr),
20080 Perl_re_printf( aTHX_
20081 "%s utf8 %s%s at %" IVdf "..%" UVuf " ",
20082 i ? "floating" : "anchored",
20084 RE_SV_TAIL(r->substrs->data[i].utf8_substr),
20085 (IV)r->substrs->data[i].min_offset,
20086 (UV)r->substrs->data[i].max_offset);
20090 if (r->check_substr || r->check_utf8)
20091 Perl_re_printf( aTHX_
20093 ( r->check_substr == r->substrs->data[1].substr
20094 && r->check_utf8 == r->substrs->data[1].utf8_substr
20095 ? "(checking floating" : "(checking anchored"));
20096 if (r->intflags & PREGf_NOSCAN)
20097 Perl_re_printf( aTHX_ " noscan");
20098 if (r->extflags & RXf_CHECK_ALL)
20099 Perl_re_printf( aTHX_ " isall");
20100 if (r->check_substr || r->check_utf8)
20101 Perl_re_printf( aTHX_ ") ");
20103 if (ri->regstclass) {
20104 regprop(r, sv, ri->regstclass, NULL, NULL);
20105 Perl_re_printf( aTHX_ "stclass %s ", SvPVX_const(sv));
20107 if (r->intflags & PREGf_ANCH) {
20108 Perl_re_printf( aTHX_ "anchored");
20109 if (r->intflags & PREGf_ANCH_MBOL)
20110 Perl_re_printf( aTHX_ "(MBOL)");
20111 if (r->intflags & PREGf_ANCH_SBOL)
20112 Perl_re_printf( aTHX_ "(SBOL)");
20113 if (r->intflags & PREGf_ANCH_GPOS)
20114 Perl_re_printf( aTHX_ "(GPOS)");
20115 Perl_re_printf( aTHX_ " ");
20117 if (r->intflags & PREGf_GPOS_SEEN)
20118 Perl_re_printf( aTHX_ "GPOS:%" UVuf " ", (UV)r->gofs);
20119 if (r->intflags & PREGf_SKIP)
20120 Perl_re_printf( aTHX_ "plus ");
20121 if (r->intflags & PREGf_IMPLICIT)
20122 Perl_re_printf( aTHX_ "implicit ");
20123 Perl_re_printf( aTHX_ "minlen %" IVdf " ", (IV)r->minlen);
20124 if (r->extflags & RXf_EVAL_SEEN)
20125 Perl_re_printf( aTHX_ "with eval ");
20126 Perl_re_printf( aTHX_ "\n");
20128 regdump_extflags("r->extflags: ", r->extflags);
20129 regdump_intflags("r->intflags: ", r->intflags);
20132 PERL_ARGS_ASSERT_REGDUMP;
20133 PERL_UNUSED_CONTEXT;
20134 PERL_UNUSED_ARG(r);
20135 #endif /* DEBUGGING */
20138 /* Should be synchronized with ANYOF_ #defines in regcomp.h */
20141 # if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 \
20142 || _CC_LOWER != 3 || _CC_UPPER != 4 || _CC_PUNCT != 5 \
20143 || _CC_PRINT != 6 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 \
20144 || _CC_CASED != 9 || _CC_SPACE != 10 || _CC_BLANK != 11 \
20145 || _CC_XDIGIT != 12 || _CC_CNTRL != 13 || _CC_ASCII != 14 \
20146 || _CC_VERTSPACE != 15
20147 # error Need to adjust order of anyofs[]
20149 static const char * const anyofs[] = {
20186 - regprop - printable representation of opcode, with run time support
20190 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo, const RExC_state_t *pRExC_state)
20195 RXi_GET_DECL(prog, progi);
20196 GET_RE_DEBUG_FLAGS_DECL;
20198 PERL_ARGS_ASSERT_REGPROP;
20202 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
20203 /* It would be nice to FAIL() here, but this may be called from
20204 regexec.c, and it would be hard to supply pRExC_state. */
20205 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
20206 (int)OP(o), (int)REGNODE_MAX);
20207 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
20209 k = PL_regkind[OP(o)];
20212 sv_catpvs(sv, " ");
20213 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
20214 * is a crude hack but it may be the best for now since
20215 * we have no flag "this EXACTish node was UTF-8"
20217 pv_pretty(sv, STRING(o), STR_LEN(o), PL_dump_re_max_len,
20218 PL_colors[0], PL_colors[1],
20219 PERL_PV_ESCAPE_UNI_DETECT |
20220 PERL_PV_ESCAPE_NONASCII |
20221 PERL_PV_PRETTY_ELLIPSES |
20222 PERL_PV_PRETTY_LTGT |
20223 PERL_PV_PRETTY_NOCLEAR
20225 } else if (k == TRIE) {
20226 /* print the details of the trie in dumpuntil instead, as
20227 * progi->data isn't available here */
20228 const char op = OP(o);
20229 const U32 n = ARG(o);
20230 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
20231 (reg_ac_data *)progi->data->data[n] :
20233 const reg_trie_data * const trie
20234 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
20236 Perl_sv_catpvf(aTHX_ sv, "-%s", PL_reg_name[o->flags]);
20237 DEBUG_TRIE_COMPILE_r({
20239 sv_catpvs(sv, "(JUMP)");
20240 Perl_sv_catpvf(aTHX_ sv,
20241 "<S:%" UVuf "/%" IVdf " W:%" UVuf " L:%" UVuf "/%" UVuf " C:%" UVuf "/%" UVuf ">",
20242 (UV)trie->startstate,
20243 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
20244 (UV)trie->wordcount,
20247 (UV)TRIE_CHARCOUNT(trie),
20248 (UV)trie->uniquecharcount
20251 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
20252 sv_catpvs(sv, "[");
20253 (void) put_charclass_bitmap_innards(sv,
20254 ((IS_ANYOF_TRIE(op))
20256 : TRIE_BITMAP(trie)),
20262 sv_catpvs(sv, "]");
20264 } else if (k == CURLY) {
20265 U32 lo = ARG1(o), hi = ARG2(o);
20266 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
20267 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
20268 Perl_sv_catpvf(aTHX_ sv, "{%u,", (unsigned) lo);
20269 if (hi == REG_INFTY)
20270 sv_catpvs(sv, "INFTY");
20272 Perl_sv_catpvf(aTHX_ sv, "%u", (unsigned) hi);
20273 sv_catpvs(sv, "}");
20275 else if (k == WHILEM && o->flags) /* Ordinal/of */
20276 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
20277 else if (k == REF || k == OPEN || k == CLOSE
20278 || k == GROUPP || OP(o)==ACCEPT)
20280 AV *name_list= NULL;
20281 U32 parno= OP(o) == ACCEPT ? (U32)ARG2L(o) : ARG(o);
20282 Perl_sv_catpvf(aTHX_ sv, "%" UVuf, (UV)parno); /* Parenth number */
20283 if ( RXp_PAREN_NAMES(prog) ) {
20284 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
20285 } else if ( pRExC_state ) {
20286 name_list= RExC_paren_name_list;
20289 if ( k != REF || (OP(o) < NREF)) {
20290 SV **name= av_fetch(name_list, parno, 0 );
20292 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
20295 SV *sv_dat= MUTABLE_SV(progi->data->data[ parno ]);
20296 I32 *nums=(I32*)SvPVX(sv_dat);
20297 SV **name= av_fetch(name_list, nums[0], 0 );
20300 for ( n=0; n<SvIVX(sv_dat); n++ ) {
20301 Perl_sv_catpvf(aTHX_ sv, "%s%" IVdf,
20302 (n ? "," : ""), (IV)nums[n]);
20304 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
20308 if ( k == REF && reginfo) {
20309 U32 n = ARG(o); /* which paren pair */
20310 I32 ln = prog->offs[n].start;
20311 if (prog->lastparen < n || ln == -1 || prog->offs[n].end == -1)
20312 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
20313 else if (ln == prog->offs[n].end)
20314 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
20316 const char *s = reginfo->strbeg + ln;
20317 Perl_sv_catpvf(aTHX_ sv, ": ");
20318 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
20319 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
20322 } else if (k == GOSUB) {
20323 AV *name_list= NULL;
20324 if ( RXp_PAREN_NAMES(prog) ) {
20325 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
20326 } else if ( pRExC_state ) {
20327 name_list= RExC_paren_name_list;
20330 /* Paren and offset */
20331 Perl_sv_catpvf(aTHX_ sv, "%d[%+d:%d]", (int)ARG(o),(int)ARG2L(o),
20332 (int)((o + (int)ARG2L(o)) - progi->program) );
20334 SV **name= av_fetch(name_list, ARG(o), 0 );
20336 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
20339 else if (k == LOGICAL)
20340 /* 2: embedded, otherwise 1 */
20341 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
20342 else if (k == ANYOF) {
20343 const U8 flags = (OP(o) == ANYOFH) ? 0 : ANYOF_FLAGS(o);
20344 bool do_sep = FALSE; /* Do we need to separate various components of
20346 /* Set if there is still an unresolved user-defined property */
20347 SV *unresolved = NULL;
20349 /* Things that are ignored except when the runtime locale is UTF-8 */
20350 SV *only_utf8_locale_invlist = NULL;
20352 /* Code points that don't fit in the bitmap */
20353 SV *nonbitmap_invlist = NULL;
20355 /* And things that aren't in the bitmap, but are small enough to be */
20356 SV* bitmap_range_not_in_bitmap = NULL;
20358 const bool inverted = flags & ANYOF_INVERT;
20360 if (OP(o) == ANYOFL || OP(o) == ANYOFPOSIXL) {
20361 if (ANYOFL_UTF8_LOCALE_REQD(flags)) {
20362 sv_catpvs(sv, "{utf8-locale-reqd}");
20364 if (flags & ANYOFL_FOLD) {
20365 sv_catpvs(sv, "{i}");
20369 /* If there is stuff outside the bitmap, get it */
20370 if (ARG(o) != ANYOF_ONLY_HAS_BITMAP) {
20371 (void) _get_regclass_nonbitmap_data(prog, o, FALSE,
20373 &only_utf8_locale_invlist,
20374 &nonbitmap_invlist);
20375 /* The non-bitmap data may contain stuff that could fit in the
20376 * bitmap. This could come from a user-defined property being
20377 * finally resolved when this call was done; or much more likely
20378 * because there are matches that require UTF-8 to be valid, and so
20379 * aren't in the bitmap. This is teased apart later */
20380 _invlist_intersection(nonbitmap_invlist,
20382 &bitmap_range_not_in_bitmap);
20383 /* Leave just the things that don't fit into the bitmap */
20384 _invlist_subtract(nonbitmap_invlist,
20386 &nonbitmap_invlist);
20389 /* Obey this flag to add all above-the-bitmap code points */
20390 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
20391 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
20392 NUM_ANYOF_CODE_POINTS,
20396 /* Ready to start outputting. First, the initial left bracket */
20397 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
20399 if (OP(o) != ANYOFH) {
20400 /* Then all the things that could fit in the bitmap */
20401 do_sep = put_charclass_bitmap_innards(sv,
20403 bitmap_range_not_in_bitmap,
20404 only_utf8_locale_invlist,
20407 /* Can't try inverting for a
20408 * better display if there
20409 * are things that haven't
20411 unresolved != NULL);
20412 SvREFCNT_dec(bitmap_range_not_in_bitmap);
20414 /* If there are user-defined properties which haven't been defined
20415 * yet, output them. If the result is not to be inverted, it is
20416 * clearest to output them in a separate [] from the bitmap range
20417 * stuff. If the result is to be complemented, we have to show
20418 * everything in one [], as the inversion applies to the whole
20419 * thing. Use {braces} to separate them from anything in the
20420 * bitmap and anything above the bitmap. */
20423 if (! do_sep) { /* If didn't output anything in the bitmap
20425 sv_catpvs(sv, "^");
20427 sv_catpvs(sv, "{");
20430 Perl_sv_catpvf(aTHX_ sv,"%s][%s", PL_colors[1],
20433 sv_catsv(sv, unresolved);
20435 sv_catpvs(sv, "}");
20437 do_sep = ! inverted;
20441 /* And, finally, add the above-the-bitmap stuff */
20442 if (nonbitmap_invlist && _invlist_len(nonbitmap_invlist)) {
20445 /* See if truncation size is overridden */
20446 const STRLEN dump_len = (PL_dump_re_max_len > 256)
20447 ? PL_dump_re_max_len
20450 /* This is output in a separate [] */
20452 Perl_sv_catpvf(aTHX_ sv,"%s][%s", PL_colors[1], PL_colors[0]);
20455 /* And, for easy of understanding, it is shown in the
20456 * uncomplemented form if possible. The one exception being if
20457 * there are unresolved items, where the inversion has to be
20458 * delayed until runtime */
20459 if (inverted && ! unresolved) {
20460 _invlist_invert(nonbitmap_invlist);
20461 _invlist_subtract(nonbitmap_invlist, PL_InBitmap, &nonbitmap_invlist);
20464 contents = invlist_contents(nonbitmap_invlist,
20465 FALSE /* output suitable for catsv */
20468 /* If the output is shorter than the permissible maximum, just do it. */
20469 if (SvCUR(contents) <= dump_len) {
20470 sv_catsv(sv, contents);
20473 const char * contents_string = SvPVX(contents);
20474 STRLEN i = dump_len;
20476 /* Otherwise, start at the permissible max and work back to the
20477 * first break possibility */
20478 while (i > 0 && contents_string[i] != ' ') {
20481 if (i == 0) { /* Fail-safe. Use the max if we couldn't
20482 find a legal break */
20486 sv_catpvn(sv, contents_string, i);
20487 sv_catpvs(sv, "...");
20490 SvREFCNT_dec_NN(contents);
20491 SvREFCNT_dec_NN(nonbitmap_invlist);
20494 /* And finally the matching, closing ']' */
20495 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
20497 if (OP(o) == ANYOFH && FLAGS(o) != 0) {
20498 Perl_sv_catpvf(aTHX_ sv, " (First UTF-8 byte=\\x%02x)", FLAGS(o));
20502 SvREFCNT_dec(unresolved);
20504 else if (k == ANYOFM) {
20505 SV * cp_list = get_ANYOFM_contents(o);
20507 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
20508 if (OP(o) == NANYOFM) {
20509 _invlist_invert(cp_list);
20512 put_charclass_bitmap_innards(sv, NULL, cp_list, NULL, NULL, TRUE);
20513 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
20515 SvREFCNT_dec(cp_list);
20517 else if (k == POSIXD || k == NPOSIXD) {
20518 U8 index = FLAGS(o) * 2;
20519 if (index < C_ARRAY_LENGTH(anyofs)) {
20520 if (*anyofs[index] != '[') {
20521 sv_catpvs(sv, "[");
20523 sv_catpv(sv, anyofs[index]);
20524 if (*anyofs[index] != '[') {
20525 sv_catpvs(sv, "]");
20529 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
20532 else if (k == BOUND || k == NBOUND) {
20533 /* Must be synced with order of 'bound_type' in regcomp.h */
20534 const char * const bounds[] = {
20535 "", /* Traditional */
20541 assert(FLAGS(o) < C_ARRAY_LENGTH(bounds));
20542 sv_catpv(sv, bounds[FLAGS(o)]);
20544 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH)) {
20545 Perl_sv_catpvf(aTHX_ sv, "[%d", -(o->flags));
20547 Perl_sv_catpvf(aTHX_ sv, "..-%d", o->flags - o->next_off);
20549 Perl_sv_catpvf(aTHX_ sv, "]");
20551 else if (OP(o) == SBOL)
20552 Perl_sv_catpvf(aTHX_ sv, " /%s/", o->flags ? "\\A" : "^");
20554 /* add on the verb argument if there is one */
20555 if ( ( k == VERB || OP(o) == ACCEPT || OP(o) == OPFAIL ) && o->flags) {
20557 Perl_sv_catpvf(aTHX_ sv, ":%" SVf,
20558 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
20560 sv_catpvs(sv, ":NULL");
20563 PERL_UNUSED_CONTEXT;
20564 PERL_UNUSED_ARG(sv);
20565 PERL_UNUSED_ARG(o);
20566 PERL_UNUSED_ARG(prog);
20567 PERL_UNUSED_ARG(reginfo);
20568 PERL_UNUSED_ARG(pRExC_state);
20569 #endif /* DEBUGGING */
20575 Perl_re_intuit_string(pTHX_ REGEXP * const r)
20576 { /* Assume that RE_INTUIT is set */
20577 struct regexp *const prog = ReANY(r);
20578 GET_RE_DEBUG_FLAGS_DECL;
20580 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
20581 PERL_UNUSED_CONTEXT;
20585 const char * const s = SvPV_nolen_const(RX_UTF8(r)
20586 ? prog->check_utf8 : prog->check_substr);
20588 if (!PL_colorset) reginitcolors();
20589 Perl_re_printf( aTHX_
20590 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
20592 RX_UTF8(r) ? "utf8 " : "",
20593 PL_colors[5], PL_colors[0],
20596 (strlen(s) > PL_dump_re_max_len ? "..." : ""));
20599 /* use UTF8 check substring if regexp pattern itself is in UTF8 */
20600 return RX_UTF8(r) ? prog->check_utf8 : prog->check_substr;
20606 handles refcounting and freeing the perl core regexp structure. When
20607 it is necessary to actually free the structure the first thing it
20608 does is call the 'free' method of the regexp_engine associated to
20609 the regexp, allowing the handling of the void *pprivate; member
20610 first. (This routine is not overridable by extensions, which is why
20611 the extensions free is called first.)
20613 See regdupe and regdupe_internal if you change anything here.
20615 #ifndef PERL_IN_XSUB_RE
20617 Perl_pregfree(pTHX_ REGEXP *r)
20623 Perl_pregfree2(pTHX_ REGEXP *rx)
20625 struct regexp *const r = ReANY(rx);
20626 GET_RE_DEBUG_FLAGS_DECL;
20628 PERL_ARGS_ASSERT_PREGFREE2;
20633 if (r->mother_re) {
20634 ReREFCNT_dec(r->mother_re);
20636 CALLREGFREE_PVT(rx); /* free the private data */
20637 SvREFCNT_dec(RXp_PAREN_NAMES(r));
20641 for (i = 0; i < 2; i++) {
20642 SvREFCNT_dec(r->substrs->data[i].substr);
20643 SvREFCNT_dec(r->substrs->data[i].utf8_substr);
20645 Safefree(r->substrs);
20647 RX_MATCH_COPY_FREE(rx);
20648 #ifdef PERL_ANY_COW
20649 SvREFCNT_dec(r->saved_copy);
20652 SvREFCNT_dec(r->qr_anoncv);
20653 if (r->recurse_locinput)
20654 Safefree(r->recurse_locinput);
20660 Copy ssv to dsv, both of which should of type SVt_REGEXP or SVt_PVLV,
20661 except that dsv will be created if NULL.
20663 This function is used in two main ways. First to implement
20664 $r = qr/....; $s = $$r;
20666 Secondly, it is used as a hacky workaround to the structural issue of
20668 being stored in the regexp structure which is in turn stored in
20669 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
20670 could be PL_curpm in multiple contexts, and could require multiple
20671 result sets being associated with the pattern simultaneously, such
20672 as when doing a recursive match with (??{$qr})
20674 The solution is to make a lightweight copy of the regexp structure
20675 when a qr// is returned from the code executed by (??{$qr}) this
20676 lightweight copy doesn't actually own any of its data except for
20677 the starp/end and the actual regexp structure itself.
20683 Perl_reg_temp_copy(pTHX_ REGEXP *dsv, REGEXP *ssv)
20685 struct regexp *drx;
20686 struct regexp *const srx = ReANY(ssv);
20687 const bool islv = dsv && SvTYPE(dsv) == SVt_PVLV;
20689 PERL_ARGS_ASSERT_REG_TEMP_COPY;
20692 dsv = (REGEXP*) newSV_type(SVt_REGEXP);
20694 assert(SvTYPE(dsv) == SVt_REGEXP || (SvTYPE(dsv) == SVt_PVLV));
20696 /* our only valid caller, sv_setsv_flags(), should have done
20697 * a SV_CHECK_THINKFIRST_COW_DROP() by now */
20698 assert(!SvOOK(dsv));
20699 assert(!SvIsCOW(dsv));
20700 assert(!SvROK(dsv));
20702 if (SvPVX_const(dsv)) {
20704 Safefree(SvPVX(dsv));
20709 SvOK_off((SV *)dsv);
20712 /* For PVLVs, the head (sv_any) points to an XPVLV, while
20713 * the LV's xpvlenu_rx will point to a regexp body, which
20714 * we allocate here */
20715 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
20716 assert(!SvPVX(dsv));
20717 ((XPV*)SvANY(dsv))->xpv_len_u.xpvlenu_rx = temp->sv_any;
20718 temp->sv_any = NULL;
20719 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
20720 SvREFCNT_dec_NN(temp);
20721 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
20722 ing below will not set it. */
20723 SvCUR_set(dsv, SvCUR(ssv));
20726 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
20727 sv_force_normal(sv) is called. */
20731 SvFLAGS(dsv) |= SvFLAGS(ssv) & (SVf_POK|SVp_POK|SVf_UTF8);
20732 SvPV_set(dsv, RX_WRAPPED(ssv));
20733 /* We share the same string buffer as the original regexp, on which we
20734 hold a reference count, incremented when mother_re is set below.
20735 The string pointer is copied here, being part of the regexp struct.
20737 memcpy(&(drx->xpv_cur), &(srx->xpv_cur),
20738 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
20742 const I32 npar = srx->nparens+1;
20743 Newx(drx->offs, npar, regexp_paren_pair);
20744 Copy(srx->offs, drx->offs, npar, regexp_paren_pair);
20746 if (srx->substrs) {
20748 Newx(drx->substrs, 1, struct reg_substr_data);
20749 StructCopy(srx->substrs, drx->substrs, struct reg_substr_data);
20751 for (i = 0; i < 2; i++) {
20752 SvREFCNT_inc_void(drx->substrs->data[i].substr);
20753 SvREFCNT_inc_void(drx->substrs->data[i].utf8_substr);
20756 /* check_substr and check_utf8, if non-NULL, point to either their
20757 anchored or float namesakes, and don't hold a second reference. */
20759 RX_MATCH_COPIED_off(dsv);
20760 #ifdef PERL_ANY_COW
20761 drx->saved_copy = NULL;
20763 drx->mother_re = ReREFCNT_inc(srx->mother_re ? srx->mother_re : ssv);
20764 SvREFCNT_inc_void(drx->qr_anoncv);
20765 if (srx->recurse_locinput)
20766 Newx(drx->recurse_locinput, srx->nparens + 1, char *);
20773 /* regfree_internal()
20775 Free the private data in a regexp. This is overloadable by
20776 extensions. Perl takes care of the regexp structure in pregfree(),
20777 this covers the *pprivate pointer which technically perl doesn't
20778 know about, however of course we have to handle the
20779 regexp_internal structure when no extension is in use.
20781 Note this is called before freeing anything in the regexp
20786 Perl_regfree_internal(pTHX_ REGEXP * const rx)
20788 struct regexp *const r = ReANY(rx);
20789 RXi_GET_DECL(r, ri);
20790 GET_RE_DEBUG_FLAGS_DECL;
20792 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
20802 SV *dsv= sv_newmortal();
20803 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
20804 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), PL_dump_re_max_len);
20805 Perl_re_printf( aTHX_ "%sFreeing REx:%s %s\n",
20806 PL_colors[4], PL_colors[5], s);
20810 #ifdef RE_TRACK_PATTERN_OFFSETS
20812 Safefree(ri->u.offsets); /* 20010421 MJD */
20814 if (ri->code_blocks)
20815 S_free_codeblocks(aTHX_ ri->code_blocks);
20818 int n = ri->data->count;
20821 /* If you add a ->what type here, update the comment in regcomp.h */
20822 switch (ri->data->what[n]) {
20828 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
20831 Safefree(ri->data->data[n]);
20837 { /* Aho Corasick add-on structure for a trie node.
20838 Used in stclass optimization only */
20840 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
20841 #ifdef USE_ITHREADS
20845 refcount = --aho->refcount;
20848 PerlMemShared_free(aho->states);
20849 PerlMemShared_free(aho->fail);
20850 /* do this last!!!! */
20851 PerlMemShared_free(ri->data->data[n]);
20852 /* we should only ever get called once, so
20853 * assert as much, and also guard the free
20854 * which /might/ happen twice. At the least
20855 * it will make code anlyzers happy and it
20856 * doesn't cost much. - Yves */
20857 assert(ri->regstclass);
20858 if (ri->regstclass) {
20859 PerlMemShared_free(ri->regstclass);
20860 ri->regstclass = 0;
20867 /* trie structure. */
20869 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
20870 #ifdef USE_ITHREADS
20874 refcount = --trie->refcount;
20877 PerlMemShared_free(trie->charmap);
20878 PerlMemShared_free(trie->states);
20879 PerlMemShared_free(trie->trans);
20881 PerlMemShared_free(trie->bitmap);
20883 PerlMemShared_free(trie->jump);
20884 PerlMemShared_free(trie->wordinfo);
20885 /* do this last!!!! */
20886 PerlMemShared_free(ri->data->data[n]);
20891 Perl_croak(aTHX_ "panic: regfree data code '%c'",
20892 ri->data->what[n]);
20895 Safefree(ri->data->what);
20896 Safefree(ri->data);
20902 #define av_dup_inc(s, t) MUTABLE_AV(sv_dup_inc((const SV *)s, t))
20903 #define hv_dup_inc(s, t) MUTABLE_HV(sv_dup_inc((const SV *)s, t))
20904 #define SAVEPVN(p, n) ((p) ? savepvn(p, n) : NULL)
20907 re_dup_guts - duplicate a regexp.
20909 This routine is expected to clone a given regexp structure. It is only
20910 compiled under USE_ITHREADS.
20912 After all of the core data stored in struct regexp is duplicated
20913 the regexp_engine.dupe method is used to copy any private data
20914 stored in the *pprivate pointer. This allows extensions to handle
20915 any duplication it needs to do.
20917 See pregfree() and regfree_internal() if you change anything here.
20919 #if defined(USE_ITHREADS)
20920 #ifndef PERL_IN_XSUB_RE
20922 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
20926 const struct regexp *r = ReANY(sstr);
20927 struct regexp *ret = ReANY(dstr);
20929 PERL_ARGS_ASSERT_RE_DUP_GUTS;
20931 npar = r->nparens+1;
20932 Newx(ret->offs, npar, regexp_paren_pair);
20933 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
20935 if (ret->substrs) {
20936 /* Do it this way to avoid reading from *r after the StructCopy().
20937 That way, if any of the sv_dup_inc()s dislodge *r from the L1
20938 cache, it doesn't matter. */
20940 const bool anchored = r->check_substr
20941 ? r->check_substr == r->substrs->data[0].substr
20942 : r->check_utf8 == r->substrs->data[0].utf8_substr;
20943 Newx(ret->substrs, 1, struct reg_substr_data);
20944 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
20946 for (i = 0; i < 2; i++) {
20947 ret->substrs->data[i].substr =
20948 sv_dup_inc(ret->substrs->data[i].substr, param);
20949 ret->substrs->data[i].utf8_substr =
20950 sv_dup_inc(ret->substrs->data[i].utf8_substr, param);
20953 /* check_substr and check_utf8, if non-NULL, point to either their
20954 anchored or float namesakes, and don't hold a second reference. */
20956 if (ret->check_substr) {
20958 assert(r->check_utf8 == r->substrs->data[0].utf8_substr);
20960 ret->check_substr = ret->substrs->data[0].substr;
20961 ret->check_utf8 = ret->substrs->data[0].utf8_substr;
20963 assert(r->check_substr == r->substrs->data[1].substr);
20964 assert(r->check_utf8 == r->substrs->data[1].utf8_substr);
20966 ret->check_substr = ret->substrs->data[1].substr;
20967 ret->check_utf8 = ret->substrs->data[1].utf8_substr;
20969 } else if (ret->check_utf8) {
20971 ret->check_utf8 = ret->substrs->data[0].utf8_substr;
20973 ret->check_utf8 = ret->substrs->data[1].utf8_substr;
20978 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
20979 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
20980 if (r->recurse_locinput)
20981 Newx(ret->recurse_locinput, r->nparens + 1, char *);
20984 RXi_SET(ret, CALLREGDUPE_PVT(dstr, param));
20986 if (RX_MATCH_COPIED(dstr))
20987 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
20989 ret->subbeg = NULL;
20990 #ifdef PERL_ANY_COW
20991 ret->saved_copy = NULL;
20994 /* Whether mother_re be set or no, we need to copy the string. We
20995 cannot refrain from copying it when the storage points directly to
20996 our mother regexp, because that's
20997 1: a buffer in a different thread
20998 2: something we no longer hold a reference on
20999 so we need to copy it locally. */
21000 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED_const(sstr), SvCUR(sstr)+1);
21001 /* set malloced length to a non-zero value so it will be freed
21002 * (otherwise in combination with SVf_FAKE it looks like an alien
21003 * buffer). It doesn't have to be the actual malloced size, since it
21004 * should never be grown */
21005 SvLEN_set(dstr, SvCUR(sstr)+1);
21006 ret->mother_re = NULL;
21008 #endif /* PERL_IN_XSUB_RE */
21013 This is the internal complement to regdupe() which is used to copy
21014 the structure pointed to by the *pprivate pointer in the regexp.
21015 This is the core version of the extension overridable cloning hook.
21016 The regexp structure being duplicated will be copied by perl prior
21017 to this and will be provided as the regexp *r argument, however
21018 with the /old/ structures pprivate pointer value. Thus this routine
21019 may override any copying normally done by perl.
21021 It returns a pointer to the new regexp_internal structure.
21025 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
21028 struct regexp *const r = ReANY(rx);
21029 regexp_internal *reti;
21031 RXi_GET_DECL(r, ri);
21033 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
21037 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
21038 char, regexp_internal);
21039 Copy(ri->program, reti->program, len+1, regnode);
21042 if (ri->code_blocks) {
21044 Newx(reti->code_blocks, 1, struct reg_code_blocks);
21045 Newx(reti->code_blocks->cb, ri->code_blocks->count,
21046 struct reg_code_block);
21047 Copy(ri->code_blocks->cb, reti->code_blocks->cb,
21048 ri->code_blocks->count, struct reg_code_block);
21049 for (n = 0; n < ri->code_blocks->count; n++)
21050 reti->code_blocks->cb[n].src_regex = (REGEXP*)
21051 sv_dup_inc((SV*)(ri->code_blocks->cb[n].src_regex), param);
21052 reti->code_blocks->count = ri->code_blocks->count;
21053 reti->code_blocks->refcnt = 1;
21056 reti->code_blocks = NULL;
21058 reti->regstclass = NULL;
21061 struct reg_data *d;
21062 const int count = ri->data->count;
21065 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
21066 char, struct reg_data);
21067 Newx(d->what, count, U8);
21070 for (i = 0; i < count; i++) {
21071 d->what[i] = ri->data->what[i];
21072 switch (d->what[i]) {
21073 /* see also regcomp.h and regfree_internal() */
21074 case 'a': /* actually an AV, but the dup function is identical.
21075 values seem to be "plain sv's" generally. */
21076 case 'r': /* a compiled regex (but still just another SV) */
21077 case 's': /* an RV (currently only used for an RV to an AV by the ANYOF code)
21078 this use case should go away, the code could have used
21079 'a' instead - see S_set_ANYOF_arg() for array contents. */
21080 case 'S': /* actually an SV, but the dup function is identical. */
21081 case 'u': /* actually an HV, but the dup function is identical.
21082 values are "plain sv's" */
21083 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
21086 /* Synthetic Start Class - "Fake" charclass we generate to optimize
21087 * patterns which could start with several different things. Pre-TRIE
21088 * this was more important than it is now, however this still helps
21089 * in some places, for instance /x?a+/ might produce a SSC equivalent
21090 * to [xa]. This is used by Perl_re_intuit_start() and S_find_byclass()
21093 /* This is cheating. */
21094 Newx(d->data[i], 1, regnode_ssc);
21095 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
21096 reti->regstclass = (regnode*)d->data[i];
21099 /* AHO-CORASICK fail table */
21100 /* Trie stclasses are readonly and can thus be shared
21101 * without duplication. We free the stclass in pregfree
21102 * when the corresponding reg_ac_data struct is freed.
21104 reti->regstclass= ri->regstclass;
21107 /* TRIE transition table */
21109 ((reg_trie_data*)ri->data->data[i])->refcount++;
21112 case 'l': /* (?{...}) or (??{ ... }) code (cb->block) */
21113 case 'L': /* same when RExC_pm_flags & PMf_HAS_CV and code
21114 is not from another regexp */
21115 d->data[i] = ri->data->data[i];
21118 Perl_croak(aTHX_ "panic: re_dup_guts unknown data code '%c'",
21119 ri->data->what[i]);
21128 reti->name_list_idx = ri->name_list_idx;
21130 #ifdef RE_TRACK_PATTERN_OFFSETS
21131 if (ri->u.offsets) {
21132 Newx(reti->u.offsets, 2*len+1, U32);
21133 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
21136 SetProgLen(reti, len);
21139 return (void*)reti;
21142 #endif /* USE_ITHREADS */
21144 #ifndef PERL_IN_XSUB_RE
21147 - regnext - dig the "next" pointer out of a node
21150 Perl_regnext(pTHX_ regnode *p)
21157 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
21158 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
21159 (int)OP(p), (int)REGNODE_MAX);
21162 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
21172 S_re_croak2(pTHX_ bool utf8, const char* pat1, const char* pat2,...)
21175 STRLEN l1 = strlen(pat1);
21176 STRLEN l2 = strlen(pat2);
21179 const char *message;
21181 PERL_ARGS_ASSERT_RE_CROAK2;
21187 Copy(pat1, buf, l1 , char);
21188 Copy(pat2, buf + l1, l2 , char);
21189 buf[l1 + l2] = '\n';
21190 buf[l1 + l2 + 1] = '\0';
21191 va_start(args, pat2);
21192 msv = vmess(buf, &args);
21194 message = SvPV_const(msv, l1);
21197 Copy(message, buf, l1 , char);
21198 /* l1-1 to avoid \n */
21199 Perl_croak(aTHX_ "%" UTF8f, UTF8fARG(utf8, l1-1, buf));
21202 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
21204 #ifndef PERL_IN_XSUB_RE
21206 Perl_save_re_context(pTHX)
21211 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
21214 const REGEXP * const rx = PM_GETRE(PL_curpm);
21216 nparens = RX_NPARENS(rx);
21219 /* RT #124109. This is a complete hack; in the SWASHNEW case we know
21220 * that PL_curpm will be null, but that utf8.pm and the modules it
21221 * loads will only use $1..$3.
21222 * The t/porting/re_context.t test file checks this assumption.
21227 for (i = 1; i <= nparens; i++) {
21228 char digits[TYPE_CHARS(long)];
21229 const STRLEN len = my_snprintf(digits, sizeof(digits),
21231 GV *const *const gvp
21232 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
21235 GV * const gv = *gvp;
21236 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
21246 S_put_code_point(pTHX_ SV *sv, UV c)
21248 PERL_ARGS_ASSERT_PUT_CODE_POINT;
21251 Perl_sv_catpvf(aTHX_ sv, "\\x{%04" UVXf "}", c);
21253 else if (isPRINT(c)) {
21254 const char string = (char) c;
21256 /* We use {phrase} as metanotation in the class, so also escape literal
21258 if (isBACKSLASHED_PUNCT(c) || c == '{' || c == '}')
21259 sv_catpvs(sv, "\\");
21260 sv_catpvn(sv, &string, 1);
21262 else if (isMNEMONIC_CNTRL(c)) {
21263 Perl_sv_catpvf(aTHX_ sv, "%s", cntrl_to_mnemonic((U8) c));
21266 Perl_sv_catpvf(aTHX_ sv, "\\x%02X", (U8) c);
21270 #define MAX_PRINT_A MAX_PRINT_A_FOR_USE_ONLY_BY_REGCOMP_DOT_C
21273 S_put_range(pTHX_ SV *sv, UV start, const UV end, const bool allow_literals)
21275 /* Appends to 'sv' a displayable version of the range of code points from
21276 * 'start' to 'end'. Mnemonics (like '\r') are used for the few controls
21277 * that have them, when they occur at the beginning or end of the range.
21278 * It uses hex to output the remaining code points, unless 'allow_literals'
21279 * is true, in which case the printable ASCII ones are output as-is (though
21280 * some of these will be escaped by put_code_point()).
21282 * NOTE: This is designed only for printing ranges of code points that fit
21283 * inside an ANYOF bitmap. Higher code points are simply suppressed
21286 const unsigned int min_range_count = 3;
21288 assert(start <= end);
21290 PERL_ARGS_ASSERT_PUT_RANGE;
21292 while (start <= end) {
21294 const char * format;
21296 if (end - start < min_range_count) {
21298 /* Output chars individually when they occur in short ranges */
21299 for (; start <= end; start++) {
21300 put_code_point(sv, start);
21305 /* If permitted by the input options, and there is a possibility that
21306 * this range contains a printable literal, look to see if there is
21308 if (allow_literals && start <= MAX_PRINT_A) {
21310 /* If the character at the beginning of the range isn't an ASCII
21311 * printable, effectively split the range into two parts:
21312 * 1) the portion before the first such printable,
21314 * and output them separately. */
21315 if (! isPRINT_A(start)) {
21316 UV temp_end = start + 1;
21318 /* There is no point looking beyond the final possible
21319 * printable, in MAX_PRINT_A */
21320 UV max = MIN(end, MAX_PRINT_A);
21322 while (temp_end <= max && ! isPRINT_A(temp_end)) {
21326 /* Here, temp_end points to one beyond the first printable if
21327 * found, or to one beyond 'max' if not. If none found, make
21328 * sure that we use the entire range */
21329 if (temp_end > MAX_PRINT_A) {
21330 temp_end = end + 1;
21333 /* Output the first part of the split range: the part that
21334 * doesn't have printables, with the parameter set to not look
21335 * for literals (otherwise we would infinitely recurse) */
21336 put_range(sv, start, temp_end - 1, FALSE);
21338 /* The 2nd part of the range (if any) starts here. */
21341 /* We do a continue, instead of dropping down, because even if
21342 * the 2nd part is non-empty, it could be so short that we want
21343 * to output it as individual characters, as tested for at the
21344 * top of this loop. */
21348 /* Here, 'start' is a printable ASCII. If it is an alphanumeric,
21349 * output a sub-range of just the digits or letters, then process
21350 * the remaining portion as usual. */
21351 if (isALPHANUMERIC_A(start)) {
21352 UV mask = (isDIGIT_A(start))
21357 UV temp_end = start + 1;
21359 /* Find the end of the sub-range that includes just the
21360 * characters in the same class as the first character in it */
21361 while (temp_end <= end && _generic_isCC_A(temp_end, mask)) {
21366 /* For short ranges, don't duplicate the code above to output
21367 * them; just call recursively */
21368 if (temp_end - start < min_range_count) {
21369 put_range(sv, start, temp_end, FALSE);
21371 else { /* Output as a range */
21372 put_code_point(sv, start);
21373 sv_catpvs(sv, "-");
21374 put_code_point(sv, temp_end);
21376 start = temp_end + 1;
21380 /* We output any other printables as individual characters */
21381 if (isPUNCT_A(start) || isSPACE_A(start)) {
21382 while (start <= end && (isPUNCT_A(start)
21383 || isSPACE_A(start)))
21385 put_code_point(sv, start);
21390 } /* End of looking for literals */
21392 /* Here is not to output as a literal. Some control characters have
21393 * mnemonic names. Split off any of those at the beginning and end of
21394 * the range to print mnemonically. It isn't possible for many of
21395 * these to be in a row, so this won't overwhelm with output */
21397 && (isMNEMONIC_CNTRL(start) || isMNEMONIC_CNTRL(end)))
21399 while (isMNEMONIC_CNTRL(start) && start <= end) {
21400 put_code_point(sv, start);
21404 /* If this didn't take care of the whole range ... */
21405 if (start <= end) {
21407 /* Look backwards from the end to find the final non-mnemonic
21410 while (isMNEMONIC_CNTRL(temp_end)) {
21414 /* And separately output the interior range that doesn't start
21415 * or end with mnemonics */
21416 put_range(sv, start, temp_end, FALSE);
21418 /* Then output the mnemonic trailing controls */
21419 start = temp_end + 1;
21420 while (start <= end) {
21421 put_code_point(sv, start);
21428 /* As a final resort, output the range or subrange as hex. */
21430 this_end = (end < NUM_ANYOF_CODE_POINTS)
21432 : NUM_ANYOF_CODE_POINTS - 1;
21433 #if NUM_ANYOF_CODE_POINTS > 256
21434 format = (this_end < 256)
21435 ? "\\x%02" UVXf "-\\x%02" UVXf
21436 : "\\x{%04" UVXf "}-\\x{%04" UVXf "}";
21438 format = "\\x%02" UVXf "-\\x%02" UVXf;
21440 GCC_DIAG_IGNORE_STMT(-Wformat-nonliteral);
21441 Perl_sv_catpvf(aTHX_ sv, format, start, this_end);
21442 GCC_DIAG_RESTORE_STMT;
21448 S_put_charclass_bitmap_innards_invlist(pTHX_ SV *sv, SV* invlist)
21450 /* Concatenate onto the PV in 'sv' a displayable form of the inversion list
21454 bool allow_literals = TRUE;
21456 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_INVLIST;
21458 /* Generally, it is more readable if printable characters are output as
21459 * literals, but if a range (nearly) spans all of them, it's best to output
21460 * it as a single range. This code will use a single range if all but 2
21461 * ASCII printables are in it */
21462 invlist_iterinit(invlist);
21463 while (invlist_iternext(invlist, &start, &end)) {
21465 /* If the range starts beyond the final printable, it doesn't have any
21467 if (start > MAX_PRINT_A) {
21471 /* In both ASCII and EBCDIC, a SPACE is the lowest printable. To span
21472 * all but two, the range must start and end no later than 2 from
21474 if (start < ' ' + 2 && end > MAX_PRINT_A - 2) {
21475 if (end > MAX_PRINT_A) {
21481 if (end - start >= MAX_PRINT_A - ' ' - 2) {
21482 allow_literals = FALSE;
21487 invlist_iterfinish(invlist);
21489 /* Here we have figured things out. Output each range */
21490 invlist_iterinit(invlist);
21491 while (invlist_iternext(invlist, &start, &end)) {
21492 if (start >= NUM_ANYOF_CODE_POINTS) {
21495 put_range(sv, start, end, allow_literals);
21497 invlist_iterfinish(invlist);
21503 S_put_charclass_bitmap_innards_common(pTHX_
21504 SV* invlist, /* The bitmap */
21505 SV* posixes, /* Under /l, things like [:word:], \S */
21506 SV* only_utf8, /* Under /d, matches iff the target is UTF-8 */
21507 SV* not_utf8, /* /d, matches iff the target isn't UTF-8 */
21508 SV* only_utf8_locale, /* Under /l, matches if the locale is UTF-8 */
21509 const bool invert /* Is the result to be inverted? */
21512 /* Create and return an SV containing a displayable version of the bitmap
21513 * and associated information determined by the input parameters. If the
21514 * output would have been only the inversion indicator '^', NULL is instead
21520 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_COMMON;
21523 output = newSVpvs("^");
21526 output = newSVpvs("");
21529 /* First, the code points in the bitmap that are unconditionally there */
21530 put_charclass_bitmap_innards_invlist(output, invlist);
21532 /* Traditionally, these have been placed after the main code points */
21534 sv_catsv(output, posixes);
21537 if (only_utf8 && _invlist_len(only_utf8)) {
21538 Perl_sv_catpvf(aTHX_ output, "%s{utf8}%s", PL_colors[1], PL_colors[0]);
21539 put_charclass_bitmap_innards_invlist(output, only_utf8);
21542 if (not_utf8 && _invlist_len(not_utf8)) {
21543 Perl_sv_catpvf(aTHX_ output, "%s{not utf8}%s", PL_colors[1], PL_colors[0]);
21544 put_charclass_bitmap_innards_invlist(output, not_utf8);
21547 if (only_utf8_locale && _invlist_len(only_utf8_locale)) {
21548 Perl_sv_catpvf(aTHX_ output, "%s{utf8 locale}%s", PL_colors[1], PL_colors[0]);
21549 put_charclass_bitmap_innards_invlist(output, only_utf8_locale);
21551 /* This is the only list in this routine that can legally contain code
21552 * points outside the bitmap range. The call just above to
21553 * 'put_charclass_bitmap_innards_invlist' will simply suppress them, so
21554 * output them here. There's about a half-dozen possible, and none in
21555 * contiguous ranges longer than 2 */
21556 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
21558 SV* above_bitmap = NULL;
21560 _invlist_subtract(only_utf8_locale, PL_InBitmap, &above_bitmap);
21562 invlist_iterinit(above_bitmap);
21563 while (invlist_iternext(above_bitmap, &start, &end)) {
21566 for (i = start; i <= end; i++) {
21567 put_code_point(output, i);
21570 invlist_iterfinish(above_bitmap);
21571 SvREFCNT_dec_NN(above_bitmap);
21575 if (invert && SvCUR(output) == 1) {
21583 S_put_charclass_bitmap_innards(pTHX_ SV *sv,
21585 SV *nonbitmap_invlist,
21586 SV *only_utf8_locale_invlist,
21587 const regnode * const node,
21588 const bool force_as_is_display)
21590 /* Appends to 'sv' a displayable version of the innards of the bracketed
21591 * character class defined by the other arguments:
21592 * 'bitmap' points to the bitmap, or NULL if to ignore that.
21593 * 'nonbitmap_invlist' is an inversion list of the code points that are in
21594 * the bitmap range, but for some reason aren't in the bitmap; NULL if
21595 * none. The reasons for this could be that they require some
21596 * condition such as the target string being or not being in UTF-8
21597 * (under /d), or because they came from a user-defined property that
21598 * was not resolved at the time of the regex compilation (under /u)
21599 * 'only_utf8_locale_invlist' is an inversion list of the code points that
21600 * are valid only if the runtime locale is a UTF-8 one; NULL if none
21601 * 'node' is the regex pattern ANYOF node. It is needed only when the
21602 * above two parameters are not null, and is passed so that this
21603 * routine can tease apart the various reasons for them.
21604 * 'force_as_is_display' is TRUE if this routine should definitely NOT try
21605 * to invert things to see if that leads to a cleaner display. If
21606 * FALSE, this routine is free to use its judgment about doing this.
21608 * It returns TRUE if there was actually something output. (It may be that
21609 * the bitmap, etc is empty.)
21611 * When called for outputting the bitmap of a non-ANYOF node, just pass the
21612 * bitmap, with the succeeding parameters set to NULL, and the final one to
21616 /* In general, it tries to display the 'cleanest' representation of the
21617 * innards, choosing whether to display them inverted or not, regardless of
21618 * whether the class itself is to be inverted. However, there are some
21619 * cases where it can't try inverting, as what actually matches isn't known
21620 * until runtime, and hence the inversion isn't either. */
21623 bool inverting_allowed = ! force_as_is_display;
21626 STRLEN orig_sv_cur = SvCUR(sv);
21628 SV* invlist; /* Inversion list we accumulate of code points that
21629 are unconditionally matched */
21630 SV* only_utf8 = NULL; /* Under /d, list of matches iff the target is
21632 SV* not_utf8 = NULL; /* /d, list of matches iff the target isn't UTF-8
21634 SV* posixes = NULL; /* Under /l, string of things like [:word:], \D */
21635 SV* only_utf8_locale = NULL; /* Under /l, list of matches if the locale
21638 SV* as_is_display; /* The output string when we take the inputs
21640 SV* inverted_display; /* The output string when we invert the inputs */
21642 U8 flags = (node) ? ANYOF_FLAGS(node) : 0;
21644 bool invert = cBOOL(flags & ANYOF_INVERT); /* Is the input to be inverted
21646 /* We are biased in favor of displaying things without them being inverted,
21647 * as that is generally easier to understand */
21648 const int bias = 5;
21650 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS;
21652 /* Start off with whatever code points are passed in. (We clone, so we
21653 * don't change the caller's list) */
21654 if (nonbitmap_invlist) {
21655 assert(invlist_highest(nonbitmap_invlist) < NUM_ANYOF_CODE_POINTS);
21656 invlist = invlist_clone(nonbitmap_invlist, NULL);
21658 else { /* Worst case size is every other code point is matched */
21659 invlist = _new_invlist(NUM_ANYOF_CODE_POINTS / 2);
21663 if (OP(node) == ANYOFD) {
21665 /* This flag indicates that the code points below 0x100 in the
21666 * nonbitmap list are precisely the ones that match only when the
21667 * target is UTF-8 (they should all be non-ASCII). */
21668 if (flags & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)
21670 _invlist_intersection(invlist, PL_UpperLatin1, &only_utf8);
21671 _invlist_subtract(invlist, only_utf8, &invlist);
21674 /* And this flag for matching all non-ASCII 0xFF and below */
21675 if (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)
21677 not_utf8 = invlist_clone(PL_UpperLatin1, NULL);
21680 else if (OP(node) == ANYOFL || OP(node) == ANYOFPOSIXL) {
21682 /* If either of these flags are set, what matches isn't
21683 * determinable except during execution, so don't know enough here
21685 if (flags & (ANYOFL_FOLD|ANYOF_MATCHES_POSIXL)) {
21686 inverting_allowed = FALSE;
21689 /* What the posix classes match also varies at runtime, so these
21690 * will be output symbolically. */
21691 if (ANYOF_POSIXL_TEST_ANY_SET(node)) {
21694 posixes = newSVpvs("");
21695 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
21696 if (ANYOF_POSIXL_TEST(node, i)) {
21697 sv_catpv(posixes, anyofs[i]);
21704 /* Accumulate the bit map into the unconditional match list */
21706 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
21707 if (BITMAP_TEST(bitmap, i)) {
21710 i < NUM_ANYOF_CODE_POINTS && BITMAP_TEST(bitmap, i);
21713 invlist = _add_range_to_invlist(invlist, start, i-1);
21718 /* Make sure that the conditional match lists don't have anything in them
21719 * that match unconditionally; otherwise the output is quite confusing.
21720 * This could happen if the code that populates these misses some
21723 _invlist_subtract(only_utf8, invlist, &only_utf8);
21726 _invlist_subtract(not_utf8, invlist, ¬_utf8);
21729 if (only_utf8_locale_invlist) {
21731 /* Since this list is passed in, we have to make a copy before
21733 only_utf8_locale = invlist_clone(only_utf8_locale_invlist, NULL);
21735 _invlist_subtract(only_utf8_locale, invlist, &only_utf8_locale);
21737 /* And, it can get really weird for us to try outputting an inverted
21738 * form of this list when it has things above the bitmap, so don't even
21740 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
21741 inverting_allowed = FALSE;
21745 /* Calculate what the output would be if we take the input as-is */
21746 as_is_display = put_charclass_bitmap_innards_common(invlist,
21753 /* If have to take the output as-is, just do that */
21754 if (! inverting_allowed) {
21755 if (as_is_display) {
21756 sv_catsv(sv, as_is_display);
21757 SvREFCNT_dec_NN(as_is_display);
21760 else { /* But otherwise, create the output again on the inverted input, and
21761 use whichever version is shorter */
21763 int inverted_bias, as_is_bias;
21765 /* We will apply our bias to whichever of the the results doesn't have
21775 inverted_bias = bias;
21778 /* Now invert each of the lists that contribute to the output,
21779 * excluding from the result things outside the possible range */
21781 /* For the unconditional inversion list, we have to add in all the
21782 * conditional code points, so that when inverted, they will be gone
21784 _invlist_union(only_utf8, invlist, &invlist);
21785 _invlist_union(not_utf8, invlist, &invlist);
21786 _invlist_union(only_utf8_locale, invlist, &invlist);
21787 _invlist_invert(invlist);
21788 _invlist_intersection(invlist, PL_InBitmap, &invlist);
21791 _invlist_invert(only_utf8);
21792 _invlist_intersection(only_utf8, PL_UpperLatin1, &only_utf8);
21794 else if (not_utf8) {
21796 /* If a code point matches iff the target string is not in UTF-8,
21797 * then complementing the result has it not match iff not in UTF-8,
21798 * which is the same thing as matching iff it is UTF-8. */
21799 only_utf8 = not_utf8;
21803 if (only_utf8_locale) {
21804 _invlist_invert(only_utf8_locale);
21805 _invlist_intersection(only_utf8_locale,
21807 &only_utf8_locale);
21810 inverted_display = put_charclass_bitmap_innards_common(
21815 only_utf8_locale, invert);
21817 /* Use the shortest representation, taking into account our bias
21818 * against showing it inverted */
21819 if ( inverted_display
21820 && ( ! as_is_display
21821 || ( SvCUR(inverted_display) + inverted_bias
21822 < SvCUR(as_is_display) + as_is_bias)))
21824 sv_catsv(sv, inverted_display);
21826 else if (as_is_display) {
21827 sv_catsv(sv, as_is_display);
21830 SvREFCNT_dec(as_is_display);
21831 SvREFCNT_dec(inverted_display);
21834 SvREFCNT_dec_NN(invlist);
21835 SvREFCNT_dec(only_utf8);
21836 SvREFCNT_dec(not_utf8);
21837 SvREFCNT_dec(posixes);
21838 SvREFCNT_dec(only_utf8_locale);
21840 return SvCUR(sv) > orig_sv_cur;
21843 #define CLEAR_OPTSTART \
21844 if (optstart) STMT_START { \
21845 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ \
21846 " (%" IVdf " nodes)\n", (IV)(node - optstart))); \
21850 #define DUMPUNTIL(b,e) \
21852 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
21854 STATIC const regnode *
21855 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
21856 const regnode *last, const regnode *plast,
21857 SV* sv, I32 indent, U32 depth)
21859 U8 op = PSEUDO; /* Arbitrary non-END op. */
21860 const regnode *next;
21861 const regnode *optstart= NULL;
21863 RXi_GET_DECL(r, ri);
21864 GET_RE_DEBUG_FLAGS_DECL;
21866 PERL_ARGS_ASSERT_DUMPUNTIL;
21868 #ifdef DEBUG_DUMPUNTIL
21869 Perl_re_printf( aTHX_ "--- %d : %d - %d - %d\n", indent, node-start,
21870 last ? last-start : 0, plast ? plast-start : 0);
21873 if (plast && plast < last)
21876 while (PL_regkind[op] != END && (!last || node < last)) {
21878 /* While that wasn't END last time... */
21881 if (op == CLOSE || op == SRCLOSE || op == WHILEM)
21883 next = regnext((regnode *)node);
21886 if (OP(node) == OPTIMIZED) {
21887 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
21894 regprop(r, sv, node, NULL, NULL);
21895 Perl_re_printf( aTHX_ "%4" IVdf ":%*s%s", (IV)(node - start),
21896 (int)(2*indent + 1), "", SvPVX_const(sv));
21898 if (OP(node) != OPTIMIZED) {
21899 if (next == NULL) /* Next ptr. */
21900 Perl_re_printf( aTHX_ " (0)");
21901 else if (PL_regkind[(U8)op] == BRANCH
21902 && PL_regkind[OP(next)] != BRANCH )
21903 Perl_re_printf( aTHX_ " (FAIL)");
21905 Perl_re_printf( aTHX_ " (%" IVdf ")", (IV)(next - start));
21906 Perl_re_printf( aTHX_ "\n");
21910 if (PL_regkind[(U8)op] == BRANCHJ) {
21913 const regnode *nnode = (OP(next) == LONGJMP
21914 ? regnext((regnode *)next)
21916 if (last && nnode > last)
21918 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
21921 else if (PL_regkind[(U8)op] == BRANCH) {
21923 DUMPUNTIL(NEXTOPER(node), next);
21925 else if ( PL_regkind[(U8)op] == TRIE ) {
21926 const regnode *this_trie = node;
21927 const char op = OP(node);
21928 const U32 n = ARG(node);
21929 const reg_ac_data * const ac = op>=AHOCORASICK ?
21930 (reg_ac_data *)ri->data->data[n] :
21932 const reg_trie_data * const trie =
21933 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
21935 AV *const trie_words
21936 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
21938 const regnode *nextbranch= NULL;
21941 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
21942 SV ** const elem_ptr = av_fetch(trie_words, word_idx, 0);
21944 Perl_re_indentf( aTHX_ "%s ",
21947 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
21948 SvCUR(*elem_ptr), PL_dump_re_max_len,
21949 PL_colors[0], PL_colors[1],
21951 ? PERL_PV_ESCAPE_UNI
21953 | PERL_PV_PRETTY_ELLIPSES
21954 | PERL_PV_PRETTY_LTGT
21959 U16 dist= trie->jump[word_idx+1];
21960 Perl_re_printf( aTHX_ "(%" UVuf ")\n",
21961 (UV)((dist ? this_trie + dist : next) - start));
21964 nextbranch= this_trie + trie->jump[0];
21965 DUMPUNTIL(this_trie + dist, nextbranch);
21967 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
21968 nextbranch= regnext((regnode *)nextbranch);
21970 Perl_re_printf( aTHX_ "\n");
21973 if (last && next > last)
21978 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
21979 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
21980 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
21982 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
21984 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
21986 else if ( op == PLUS || op == STAR) {
21987 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
21989 else if (PL_regkind[(U8)op] == EXACT) {
21990 /* Literal string, where present. */
21991 node += NODE_SZ_STR(node) - 1;
21992 node = NEXTOPER(node);
21995 node = NEXTOPER(node);
21996 node += regarglen[(U8)op];
21998 if (op == CURLYX || op == OPEN || op == SROPEN)
22002 #ifdef DEBUG_DUMPUNTIL
22003 Perl_re_printf( aTHX_ "--- %d\n", (int)indent);
22008 #endif /* DEBUGGING */
22010 #ifndef PERL_IN_XSUB_RE
22012 #include "uni_keywords.h"
22015 Perl_init_uniprops(pTHX)
22019 PL_user_def_props = newHV();
22021 #ifdef USE_ITHREADS
22023 HvSHAREKEYS_off(PL_user_def_props);
22024 PL_user_def_props_aTHX = aTHX;
22028 /* Set up the inversion list global variables */
22030 PL_XPosix_ptrs[_CC_ASCII] = _new_invlist_C_array(uni_prop_ptrs[UNI_ASCII]);
22031 PL_XPosix_ptrs[_CC_ALPHANUMERIC] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXALNUM]);
22032 PL_XPosix_ptrs[_CC_ALPHA] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXALPHA]);
22033 PL_XPosix_ptrs[_CC_BLANK] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXBLANK]);
22034 PL_XPosix_ptrs[_CC_CASED] = _new_invlist_C_array(uni_prop_ptrs[UNI_CASED]);
22035 PL_XPosix_ptrs[_CC_CNTRL] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXCNTRL]);
22036 PL_XPosix_ptrs[_CC_DIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXDIGIT]);
22037 PL_XPosix_ptrs[_CC_GRAPH] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXGRAPH]);
22038 PL_XPosix_ptrs[_CC_LOWER] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXLOWER]);
22039 PL_XPosix_ptrs[_CC_PRINT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXPRINT]);
22040 PL_XPosix_ptrs[_CC_PUNCT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXPUNCT]);
22041 PL_XPosix_ptrs[_CC_SPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXSPACE]);
22042 PL_XPosix_ptrs[_CC_UPPER] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXUPPER]);
22043 PL_XPosix_ptrs[_CC_VERTSPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_VERTSPACE]);
22044 PL_XPosix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXWORD]);
22045 PL_XPosix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXXDIGIT]);
22047 PL_Posix_ptrs[_CC_ASCII] = _new_invlist_C_array(uni_prop_ptrs[UNI_ASCII]);
22048 PL_Posix_ptrs[_CC_ALPHANUMERIC] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXALNUM]);
22049 PL_Posix_ptrs[_CC_ALPHA] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXALPHA]);
22050 PL_Posix_ptrs[_CC_BLANK] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXBLANK]);
22051 PL_Posix_ptrs[_CC_CASED] = PL_Posix_ptrs[_CC_ALPHA];
22052 PL_Posix_ptrs[_CC_CNTRL] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXCNTRL]);
22053 PL_Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXDIGIT]);
22054 PL_Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXGRAPH]);
22055 PL_Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXLOWER]);
22056 PL_Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXPRINT]);
22057 PL_Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXPUNCT]);
22058 PL_Posix_ptrs[_CC_SPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXSPACE]);
22059 PL_Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXUPPER]);
22060 PL_Posix_ptrs[_CC_VERTSPACE] = NULL;
22061 PL_Posix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXWORD]);
22062 PL_Posix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXXDIGIT]);
22064 PL_GCB_invlist = _new_invlist_C_array(_Perl_GCB_invlist);
22065 PL_SB_invlist = _new_invlist_C_array(_Perl_SB_invlist);
22066 PL_WB_invlist = _new_invlist_C_array(_Perl_WB_invlist);
22067 PL_LB_invlist = _new_invlist_C_array(_Perl_LB_invlist);
22068 PL_SCX_invlist = _new_invlist_C_array(_Perl_SCX_invlist);
22070 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
22071 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
22072 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
22074 PL_Assigned_invlist = _new_invlist_C_array(uni_prop_ptrs[UNI_ASSIGNED]);
22076 PL_utf8_perl_idstart = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_IDSTART]);
22077 PL_utf8_perl_idcont = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_IDCONT]);
22079 PL_utf8_charname_begin = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_CHARNAME_BEGIN]);
22080 PL_utf8_charname_continue = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_CHARNAME_CONTINUE]);
22082 PL_in_some_fold = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_ANY_FOLDS]);
22083 PL_HasMultiCharFold = _new_invlist_C_array(uni_prop_ptrs[
22084 UNI__PERL_FOLDS_TO_MULTI_CHAR]);
22085 PL_InMultiCharFold = _new_invlist_C_array(uni_prop_ptrs[
22086 UNI__PERL_IS_IN_MULTI_CHAR_FOLD]);
22087 PL_NonFinalFold = _new_invlist_C_array(uni_prop_ptrs[
22088 UNI__PERL_NON_FINAL_FOLDS]);
22090 PL_utf8_toupper = _new_invlist_C_array(Uppercase_Mapping_invlist);
22091 PL_utf8_tolower = _new_invlist_C_array(Lowercase_Mapping_invlist);
22092 PL_utf8_totitle = _new_invlist_C_array(Titlecase_Mapping_invlist);
22093 PL_utf8_tofold = _new_invlist_C_array(Case_Folding_invlist);
22094 PL_utf8_tosimplefold = _new_invlist_C_array(Simple_Case_Folding_invlist);
22095 PL_utf8_foldclosures = _new_invlist_C_array(_Perl_IVCF_invlist);
22096 PL_utf8_mark = _new_invlist_C_array(uni_prop_ptrs[UNI_M]);
22097 PL_CCC_non0_non230 = _new_invlist_C_array(_Perl_CCC_non0_non230_invlist);
22098 PL_Private_Use = _new_invlist_C_array(uni_prop_ptrs[UNI_CO]);
22101 /* The below are used only by deprecated functions. They could be removed */
22102 PL_utf8_xidcont = _new_invlist_C_array(uni_prop_ptrs[UNI_XIDC]);
22103 PL_utf8_idcont = _new_invlist_C_array(uni_prop_ptrs[UNI_IDC]);
22104 PL_utf8_xidstart = _new_invlist_C_array(uni_prop_ptrs[UNI_XIDS]);
22110 This code was mainly added for backcompat to give a warning for non-portable
22111 code points in user-defined properties. But experiments showed that the
22112 warning in earlier perls were only omitted on overflow, which should be an
22113 error, so there really isnt a backcompat issue, and actually adding the
22114 warning when none was present before might cause breakage, for little gain. So
22115 khw left this code in, but not enabled. Tests were never added.
22118 Ei |const char *|get_extended_utf8_msg|const UV cp
22120 PERL_STATIC_INLINE const char *
22121 S_get_extended_utf8_msg(pTHX_ const UV cp)
22123 U8 dummy[UTF8_MAXBYTES + 1];
22127 uvchr_to_utf8_flags_msgs(dummy, cp, UNICODE_WARN_PERL_EXTENDED,
22130 msg = hv_fetchs(msgs, "text", 0);
22133 (void) sv_2mortal((SV *) msgs);
22135 return SvPVX(*msg);
22141 Perl_handle_user_defined_property(pTHX_
22143 /* Parses the contents of a user-defined property definition; returning the
22144 * expanded definition if possible. If so, the return is an inversion
22147 * If there are subroutines that are part of the expansion and which aren't
22148 * known at the time of the call to this function, this returns what
22149 * parse_uniprop_string() returned for the first one encountered.
22151 * If an error was found, NULL is returned, and 'msg' gets a suitable
22152 * message appended to it. (Appending allows the back trace of how we got
22153 * to the faulty definition to be displayed through nested calls of
22154 * user-defined subs.)
22156 * The caller IS responsible for freeing any returned SV.
22158 * The syntax of the contents is pretty much described in perlunicode.pod,
22159 * but we also allow comments on each line */
22161 const char * name, /* Name of property */
22162 const STRLEN name_len, /* The name's length in bytes */
22163 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
22164 const bool to_fold, /* ? Is this under /i */
22165 const bool runtime, /* ? Are we in compile- or run-time */
22166 const bool deferrable, /* Is it ok for this property's full definition
22167 to be deferred until later? */
22168 SV* contents, /* The property's definition */
22169 bool *user_defined_ptr, /* This will be set TRUE as we wouldn't be
22170 getting called unless this is thought to be
22171 a user-defined property */
22172 SV * msg, /* Any error or warning msg(s) are appended to
22174 const STRLEN level) /* Recursion level of this call */
22177 const char * string = SvPV_const(contents, len);
22178 const char * const e = string + len;
22179 const bool is_contents_utf8 = cBOOL(SvUTF8(contents));
22180 const STRLEN msgs_length_on_entry = SvCUR(msg);
22182 const char * s0 = string; /* Points to first byte in the current line
22183 being parsed in 'string' */
22184 const char overflow_msg[] = "Code point too large in \"";
22185 SV* running_definition = NULL;
22187 PERL_ARGS_ASSERT_HANDLE_USER_DEFINED_PROPERTY;
22189 *user_defined_ptr = TRUE;
22191 /* Look at each line */
22193 const char * s; /* Current byte */
22194 char op = '+'; /* Default operation is 'union' */
22195 IV min = 0; /* range begin code point */
22196 IV max = -1; /* and range end */
22197 SV* this_definition;
22199 /* Skip comment lines */
22201 s0 = strchr(s0, '\n');
22209 /* For backcompat, allow an empty first line */
22215 /* First character in the line may optionally be the operation */
22224 /* If the line is one or two hex digits separated by blank space, its
22225 * a range; otherwise it is either another user-defined property or an
22230 if (! isXDIGIT(*s)) {
22231 goto check_if_property;
22234 do { /* Each new hex digit will add 4 bits. */
22235 if (min > ( (IV) MAX_LEGAL_CP >> 4)) {
22236 s = strchr(s, '\n');
22240 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
22241 sv_catpv(msg, overflow_msg);
22242 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
22243 UTF8fARG(is_contents_utf8, s - s0, s0));
22244 sv_catpvs(msg, "\"");
22245 goto return_failure;
22248 /* Accumulate this digit into the value */
22249 min = (min << 4) + READ_XDIGIT(s);
22250 } while (isXDIGIT(*s));
22252 while (isBLANK(*s)) { s++; }
22254 /* We allow comments at the end of the line */
22256 s = strchr(s, '\n');
22262 else if (s < e && *s != '\n') {
22263 if (! isXDIGIT(*s)) {
22264 goto check_if_property;
22267 /* Look for the high point of the range */
22270 if (max > ( (IV) MAX_LEGAL_CP >> 4)) {
22271 s = strchr(s, '\n');
22275 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
22276 sv_catpv(msg, overflow_msg);
22277 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
22278 UTF8fARG(is_contents_utf8, s - s0, s0));
22279 sv_catpvs(msg, "\"");
22280 goto return_failure;
22283 max = (max << 4) + READ_XDIGIT(s);
22284 } while (isXDIGIT(*s));
22286 while (isBLANK(*s)) { s++; }
22289 s = strchr(s, '\n');
22294 else if (s < e && *s != '\n') {
22295 goto check_if_property;
22299 if (max == -1) { /* The line only had one entry */
22302 else if (max < min) {
22303 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
22304 sv_catpvs(msg, "Illegal range in \"");
22305 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
22306 UTF8fARG(is_contents_utf8, s - s0, s0));
22307 sv_catpvs(msg, "\"");
22308 goto return_failure;
22311 #if 0 /* See explanation at definition above of get_extended_utf8_msg() */
22313 if ( UNICODE_IS_PERL_EXTENDED(min)
22314 || UNICODE_IS_PERL_EXTENDED(max))
22316 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
22318 /* If both code points are non-portable, warn only on the lower
22320 sv_catpv(msg, get_extended_utf8_msg(
22321 (UNICODE_IS_PERL_EXTENDED(min))
22323 sv_catpvs(msg, " in \"");
22324 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
22325 UTF8fARG(is_contents_utf8, s - s0, s0));
22326 sv_catpvs(msg, "\"");
22331 /* Here, this line contains a legal range */
22332 this_definition = sv_2mortal(_new_invlist(2));
22333 this_definition = _add_range_to_invlist(this_definition, min, max);
22338 /* Here it isn't a legal range line. See if it is a legal property
22339 * line. First find the end of the meat of the line */
22340 s = strpbrk(s, "#\n");
22345 /* Ignore trailing blanks in keeping with the requirements of
22346 * parse_uniprop_string() */
22348 while (s > s0 && isBLANK_A(*s)) {
22353 this_definition = parse_uniprop_string(s0, s - s0,
22354 is_utf8, to_fold, runtime,
22356 user_defined_ptr, msg,
22358 ? level /* Don't increase level
22359 if input is empty */
22362 if (this_definition == NULL) {
22363 goto return_failure; /* 'msg' should have had the reason
22364 appended to it by the above call */
22367 if (! is_invlist(this_definition)) { /* Unknown at this time */
22368 return newSVsv(this_definition);
22372 s = strchr(s, '\n');
22382 _invlist_union(running_definition, this_definition,
22383 &running_definition);
22386 _invlist_subtract(running_definition, this_definition,
22387 &running_definition);
22390 _invlist_intersection(running_definition, this_definition,
22391 &running_definition);
22394 _invlist_union_complement_2nd(running_definition,
22395 this_definition, &running_definition);
22398 Perl_croak(aTHX_ "panic: %s: %d: Unexpected operation %d",
22399 __FILE__, __LINE__, op);
22403 /* Position past the '\n' */
22405 } /* End of loop through the lines of 'contents' */
22407 /* Here, we processed all the lines in 'contents' without error. If we
22408 * didn't add any warnings, simply return success */
22409 if (msgs_length_on_entry == SvCUR(msg)) {
22411 /* If the expansion was empty, the answer isn't nothing: its an empty
22412 * inversion list */
22413 if (running_definition == NULL) {
22414 running_definition = _new_invlist(1);
22417 return running_definition;
22420 /* Otherwise, add some explanatory text, but we will return success */
22424 running_definition = NULL;
22428 if (name_len > 0) {
22429 sv_catpvs(msg, " in expansion of ");
22430 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8, name_len, name));
22433 return running_definition;
22436 /* As explained below, certain operations need to take place in the first
22437 * thread created. These macros switch contexts */
22438 #ifdef USE_ITHREADS
22439 # define DECLARATION_FOR_GLOBAL_CONTEXT \
22440 PerlInterpreter * save_aTHX = aTHX;
22441 # define SWITCH_TO_GLOBAL_CONTEXT \
22442 PERL_SET_CONTEXT((aTHX = PL_user_def_props_aTHX))
22443 # define RESTORE_CONTEXT PERL_SET_CONTEXT((aTHX = save_aTHX));
22444 # define CUR_CONTEXT aTHX
22445 # define ORIGINAL_CONTEXT save_aTHX
22447 # define DECLARATION_FOR_GLOBAL_CONTEXT
22448 # define SWITCH_TO_GLOBAL_CONTEXT NOOP
22449 # define RESTORE_CONTEXT NOOP
22450 # define CUR_CONTEXT NULL
22451 # define ORIGINAL_CONTEXT NULL
22455 S_delete_recursion_entry(pTHX_ void *key)
22457 /* Deletes the entry used to detect recursion when expanding user-defined
22458 * properties. This is a function so it can be set up to be called even if
22459 * the program unexpectedly quits */
22462 SV ** current_entry;
22463 const STRLEN key_len = strlen((const char *) key);
22464 DECLARATION_FOR_GLOBAL_CONTEXT;
22466 SWITCH_TO_GLOBAL_CONTEXT;
22468 /* If the entry is one of these types, it is a permanent entry, and not the
22469 * one used to detect recursions. This function should delete only the
22470 * recursion entry */
22471 current_entry = hv_fetch(PL_user_def_props, (const char *) key, key_len, 0);
22473 && ! is_invlist(*current_entry)
22474 && ! SvPOK(*current_entry))
22476 (void) hv_delete(PL_user_def_props, (const char *) key, key_len,
22484 S_get_fq_name(pTHX_
22485 const char * const name, /* The first non-blank in the \p{}, \P{} */
22486 const Size_t name_len, /* Its length in bytes, not including any trailing space */
22487 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
22488 const bool has_colon_colon
22491 /* Returns a mortal SV containing the fully qualified version of the input
22496 fq_name = newSVpvs_flags("", SVs_TEMP);
22498 /* Use the current package if it wasn't included in our input */
22499 if (! has_colon_colon) {
22500 const HV * pkg = (IN_PERL_COMPILETIME)
22502 : CopSTASH(PL_curcop);
22503 const char* pkgname = HvNAME(pkg);
22505 Perl_sv_catpvf(aTHX_ fq_name, "%" UTF8f,
22506 UTF8fARG(is_utf8, strlen(pkgname), pkgname));
22507 sv_catpvs(fq_name, "::");
22510 Perl_sv_catpvf(aTHX_ fq_name, "%" UTF8f,
22511 UTF8fARG(is_utf8, name_len, name));
22516 Perl_parse_uniprop_string(pTHX_
22518 /* Parse the interior of a \p{}, \P{}. Returns its definition if knowable
22519 * now. If so, the return is an inversion list.
22521 * If the property is user-defined, it is a subroutine, which in turn
22522 * may call other subroutines. This function will call the whole nest of
22523 * them to get the definition they return; if some aren't known at the time
22524 * of the call to this function, the fully qualified name of the highest
22525 * level sub is returned. It is an error to call this function at runtime
22526 * without every sub defined.
22528 * If an error was found, NULL is returned, and 'msg' gets a suitable
22529 * message appended to it. (Appending allows the back trace of how we got
22530 * to the faulty definition to be displayed through nested calls of
22531 * user-defined subs.)
22533 * The caller should NOT try to free any returned inversion list.
22535 * Other parameters will be set on return as described below */
22537 const char * const name, /* The first non-blank in the \p{}, \P{} */
22538 const Size_t name_len, /* Its length in bytes, not including any
22540 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
22541 const bool to_fold, /* ? Is this under /i */
22542 const bool runtime, /* TRUE if this is being called at run time */
22543 const bool deferrable, /* TRUE if it's ok for the definition to not be
22544 known at this call */
22545 bool *user_defined_ptr, /* Upon return from this function it will be
22546 set to TRUE if any component is a
22547 user-defined property */
22548 SV * msg, /* Any error or warning msg(s) are appended to
22550 const STRLEN level) /* Recursion level of this call */
22553 char* lookup_name; /* normalized name for lookup in our tables */
22554 unsigned lookup_len; /* Its length */
22555 bool stricter = FALSE; /* Some properties have stricter name
22556 normalization rules, which we decide upon
22557 based on parsing */
22559 /* nv= or numeric_value=, or possibly one of the cjk numeric properties
22560 * (though it requires extra effort to download them from Unicode and
22561 * compile perl to know about them) */
22562 bool is_nv_type = FALSE;
22564 unsigned int i, j = 0;
22565 int equals_pos = -1; /* Where the '=' is found, or negative if none */
22566 int slash_pos = -1; /* Where the '/' is found, or negative if none */
22567 int table_index = 0; /* The entry number for this property in the table
22568 of all Unicode property names */
22569 bool starts_with_In_or_Is = FALSE; /* ? Does the name start with 'In' or
22571 Size_t lookup_offset = 0; /* Used to ignore the first few characters of
22572 the normalized name in certain situations */
22573 Size_t non_pkg_begin = 0; /* Offset of first byte in 'name' that isn't
22574 part of a package name */
22575 bool could_be_user_defined = TRUE; /* ? Could this be a user-defined
22576 property rather than a Unicode
22578 SV * prop_definition = NULL; /* The returned definition of 'name' or NULL
22579 if an error. If it is an inversion list,
22580 it is the definition. Otherwise it is a
22581 string containing the fully qualified sub
22583 SV * fq_name = NULL; /* For user-defined properties, the fully
22585 bool invert_return = FALSE; /* ? Do we need to complement the result before
22588 PERL_ARGS_ASSERT_PARSE_UNIPROP_STRING;
22590 /* The input will be normalized into 'lookup_name' */
22591 Newx(lookup_name, name_len, char);
22592 SAVEFREEPV(lookup_name);
22594 /* Parse the input. */
22595 for (i = 0; i < name_len; i++) {
22596 char cur = name[i];
22598 /* Most of the characters in the input will be of this ilk, being parts
22600 if (isIDCONT_A(cur)) {
22602 /* Case differences are ignored. Our lookup routine assumes
22603 * everything is lowercase, so normalize to that */
22604 if (isUPPER_A(cur)) {
22605 lookup_name[j++] = toLOWER_A(cur);
22609 if (cur == '_') { /* Don't include these in the normalized name */
22613 lookup_name[j++] = cur;
22615 /* The first character in a user-defined name must be of this type.
22617 if (i - non_pkg_begin == 0 && ! isIDFIRST_A(cur)) {
22618 could_be_user_defined = FALSE;
22624 /* Here, the character is not something typically in a name, But these
22625 * two types of characters (and the '_' above) can be freely ignored in
22626 * most situations. Later it may turn out we shouldn't have ignored
22627 * them, and we have to reparse, but we don't have enough information
22628 * yet to make that decision */
22629 if (cur == '-' || isSPACE_A(cur)) {
22630 could_be_user_defined = FALSE;
22634 /* An equals sign or single colon mark the end of the first part of
22635 * the property name */
22637 || (cur == ':' && (i >= name_len - 1 || name[i+1] != ':')))
22639 lookup_name[j++] = '='; /* Treat the colon as an '=' */
22640 equals_pos = j; /* Note where it occurred in the input */
22641 could_be_user_defined = FALSE;
22645 /* Otherwise, this character is part of the name. */
22646 lookup_name[j++] = cur;
22648 /* Here it isn't a single colon, so if it is a colon, it must be a
22652 /* A double colon should be a package qualifier. We note its
22653 * position and continue. Note that one could have
22654 * pkg1::pkg2::...::foo
22655 * so that the position at the end of the loop will be just after
22656 * the final qualifier */
22659 non_pkg_begin = i + 1;
22660 lookup_name[j++] = ':';
22662 else { /* Only word chars (and '::') can be in a user-defined name */
22663 could_be_user_defined = FALSE;
22665 } /* End of parsing through the lhs of the property name (or all of it if
22668 #define STRLENs(s) (sizeof("" s "") - 1)
22670 /* If there is a single package name 'utf8::', it is ambiguous. It could
22671 * be for a user-defined property, or it could be a Unicode property, as
22672 * all of them are considered to be for that package. For the purposes of
22673 * parsing the rest of the property, strip it off */
22674 if (non_pkg_begin == STRLENs("utf8::") && memBEGINPs(name, name_len, "utf8::")) {
22675 lookup_name += STRLENs("utf8::");
22676 j -= STRLENs("utf8::");
22677 equals_pos -= STRLENs("utf8::");
22680 /* Here, we are either done with the whole property name, if it was simple;
22681 * or are positioned just after the '=' if it is compound. */
22683 if (equals_pos >= 0) {
22684 assert(! stricter); /* We shouldn't have set this yet */
22686 /* Space immediately after the '=' is ignored */
22688 for (; i < name_len; i++) {
22689 if (! isSPACE_A(name[i])) {
22694 /* Most punctuation after the equals indicates a subpattern, like
22696 if ( isPUNCT_A(name[i])
22702 /* Find the property. The table includes the equals sign, so we
22704 table_index = match_uniprop((U8 *) lookup_name, j);
22706 const char * const * prop_values
22707 = UNI_prop_value_ptrs[table_index];
22709 Size_t subpattern_len;
22710 REGEXP * subpattern_re;
22711 char open = name[i++];
22713 const char * pos_in_brackets;
22716 /* A backslash means the real delimitter is the next character.
22718 if (open == '\\') {
22723 /* This data structure is constructed so that the matching
22724 * closing bracket is 3 past its matching opening. The second
22725 * set of closing is so that if the opening is something like
22726 * ']', the closing will be that as well. Something similar is
22727 * done in toke.c */
22728 pos_in_brackets = strchr("([<)]>)]>", open);
22729 close = (pos_in_brackets) ? pos_in_brackets[3] : open;
22732 || name[name_len-1] != close
22733 || (escaped && name[name_len-2] != '\\'))
22735 sv_catpvs(msg, "Unicode property wildcard not terminated");
22736 goto append_name_to_msg;
22739 Perl_ck_warner_d(aTHX_
22740 packWARN(WARN_EXPERIMENTAL__UNIPROP_WILDCARDS),
22741 "The Unicode property wildcards feature is experimental");
22743 /* Now create and compile the wildcard subpattern. Use /iaa
22744 * because nothing outside of ASCII will match, and it the
22745 * property values should all match /i. Note that when the
22746 * pattern fails to compile, our added text to the user's
22747 * pattern will be displayed to the user, which is not so
22749 subpattern_len = name_len - i - 1 - escaped;
22750 subpattern = Perl_newSVpvf(aTHX_ "(?iaa:%.*s)",
22751 (unsigned) subpattern_len,
22753 subpattern = sv_2mortal(subpattern);
22754 subpattern_re = re_compile(subpattern, 0);
22755 assert(subpattern_re); /* Should have died if didn't compile
22758 /* For each legal property value, see if the supplied pattern
22760 while (*prop_values) {
22761 const char * const entry = *prop_values;
22762 const Size_t len = strlen(entry);
22763 SV* entry_sv = newSVpvn_flags(entry, len, SVs_TEMP);
22765 if (pregexec(subpattern_re,
22767 (char *) entry + len,
22771 { /* Here, matched. Add to the returned list */
22772 Size_t total_len = j + len;
22773 SV * sub_invlist = NULL;
22774 char * this_string;
22776 /* We know this is a legal \p{property=value}. Call
22777 * the function to return the list of code points that
22779 Newxz(this_string, total_len + 1, char);
22780 Copy(lookup_name, this_string, j, char);
22781 my_strlcat(this_string, entry, total_len + 1);
22782 SAVEFREEPV(this_string);
22783 sub_invlist = parse_uniprop_string(this_string,
22792 _invlist_union(prop_definition, sub_invlist,
22796 prop_values++; /* Next iteration, look at next propvalue */
22797 } /* End of looking through property values; (the data
22798 structure is terminated by a NULL ptr) */
22800 SvREFCNT_dec_NN(subpattern_re);
22802 if (prop_definition) {
22803 return prop_definition;
22806 sv_catpvs(msg, "No Unicode property value wildcard matches:");
22807 goto append_name_to_msg;
22810 /* Here's how khw thinks we should proceed to handle the properties
22811 * not yet done: Bidi Mirroring Glyph
22812 Bidi Paired Bracket
22813 Case Folding (both full and simple)
22814 Decomposition Mapping
22815 Equivalent Unified Ideograph
22818 Lowercase Mapping (both full and simple)
22820 Titlecase Mapping (both full and simple)
22821 Uppercase Mapping (both full and simple)
22822 * Move the part that looks at the property values into a perl
22823 * script, like utf8_heavy.pl is done. This makes things somewhat
22824 * easier, but most importantly, it avoids always adding all these
22825 * strings to the memory usage when the feature is little-used.
22827 * The property values would all be concatenated into a single
22828 * string per property with each value on a separate line, and the
22829 * code point it's for on alternating lines. Then we match the
22830 * user's input pattern m//mg, without having to worry about their
22831 * uses of '^' and '$'. Only the values that aren't the default
22832 * would be in the strings. Code points would be in UTF-8. The
22833 * search pattern that we would construct would look like
22834 * (?: \n (code-point_re) \n (?aam: user-re ) \n )
22835 * And so $1 would contain the code point that matched the user-re.
22836 * For properties where the default is the code point itself, such
22837 * as any of the case changing mappings, the string would otherwise
22838 * consist of all Unicode code points in UTF-8 strung together.
22839 * This would be impractical. So instead, examine their compiled
22840 * pattern, looking at the ssc. If none, reject the pattern as an
22841 * error. Otherwise run the pattern against every code point in
22842 * the ssc. The ssc is kind of like tr18's 3.9 Possible Match Sets
22843 * And it might be good to create an API to return the ssc.
22845 * For the name properties, a new function could be created in
22846 * charnames which essentially does the same thing as above,
22847 * sharing Name.pl with the other charname functions. Don't know
22848 * about loose name matching, or algorithmically determined names.
22849 * Decomposition.pl similarly.
22851 * It might be that a new pattern modifier would have to be
22852 * created, like /t for resTricTed, which changed the behavior of
22853 * some constructs in their subpattern, like \A. */
22854 } /* End of is a wildcard subppattern */
22857 /* Certain properties whose values are numeric need special handling.
22858 * They may optionally be prefixed by 'is'. Ignore that prefix for the
22859 * purposes of checking if this is one of those properties */
22860 if (memBEGINPs(lookup_name, name_len, "is")) {
22864 /* Then check if it is one of these specially-handled properties. The
22865 * possibilities are hard-coded because easier this way, and the list
22866 * is unlikely to change.
22868 * All numeric value type properties are of this ilk, and are also
22869 * special in a different way later on. So find those first. There
22870 * are several numeric value type properties in the Unihan DB (which is
22871 * unlikely to be compiled with perl, but we handle it here in case it
22872 * does get compiled). They all end with 'numeric'. The interiors
22873 * aren't checked for the precise property. This would stop working if
22874 * a cjk property were to be created that ended with 'numeric' and
22875 * wasn't a numeric type */
22876 is_nv_type = memEQs(lookup_name + lookup_offset,
22877 j - 1 - lookup_offset, "numericvalue")
22878 || memEQs(lookup_name + lookup_offset,
22879 j - 1 - lookup_offset, "nv")
22880 || ( memENDPs(lookup_name + lookup_offset,
22881 j - 1 - lookup_offset, "numeric")
22882 && ( memBEGINPs(lookup_name + lookup_offset,
22883 j - 1 - lookup_offset, "cjk")
22884 || memBEGINPs(lookup_name + lookup_offset,
22885 j - 1 - lookup_offset, "k")));
22887 || memEQs(lookup_name + lookup_offset,
22888 j - 1 - lookup_offset, "canonicalcombiningclass")
22889 || memEQs(lookup_name + lookup_offset,
22890 j - 1 - lookup_offset, "ccc")
22891 || memEQs(lookup_name + lookup_offset,
22892 j - 1 - lookup_offset, "age")
22893 || memEQs(lookup_name + lookup_offset,
22894 j - 1 - lookup_offset, "in")
22895 || memEQs(lookup_name + lookup_offset,
22896 j - 1 - lookup_offset, "presentin"))
22900 /* Since the stuff after the '=' is a number, we can't throw away
22901 * '-' willy-nilly, as those could be a minus sign. Other stricter
22902 * rules also apply. However, these properties all can have the
22903 * rhs not be a number, in which case they contain at least one
22904 * alphabetic. In those cases, the stricter rules don't apply.
22905 * But the numeric type properties can have the alphas [Ee] to
22906 * signify an exponent, and it is still a number with stricter
22907 * rules. So look for an alpha that signifies not-strict */
22909 for (k = i; k < name_len; k++) {
22910 if ( isALPHA_A(name[k])
22911 && (! is_nv_type || ! isALPHA_FOLD_EQ(name[k], 'E')))
22921 /* A number may have a leading '+' or '-'. The latter is retained
22923 if (name[i] == '+') {
22926 else if (name[i] == '-') {
22927 lookup_name[j++] = '-';
22931 /* Skip leading zeros including single underscores separating the
22932 * zeros, or between the final leading zero and the first other
22934 for (; i < name_len - 1; i++) {
22935 if ( name[i] != '0'
22936 && (name[i] != '_' || ! isDIGIT_A(name[i+1])))
22943 else { /* No '=' */
22945 /* Only a few properties without an '=' should be parsed with stricter
22946 * rules. The list is unlikely to change. */
22947 if ( memBEGINPs(lookup_name, j, "perl")
22948 && memNEs(lookup_name + 4, j - 4, "space")
22949 && memNEs(lookup_name + 4, j - 4, "word"))
22953 /* We set the inputs back to 0 and the code below will reparse,
22959 /* Here, we have either finished the property, or are positioned to parse
22960 * the remainder, and we know if stricter rules apply. Finish out, if not
22962 for (; i < name_len; i++) {
22963 char cur = name[i];
22965 /* In all instances, case differences are ignored, and we normalize to
22967 if (isUPPER_A(cur)) {
22968 lookup_name[j++] = toLOWER(cur);
22972 /* An underscore is skipped, but not under strict rules unless it
22973 * separates two digits */
22976 && ( i == 0 || (int) i == equals_pos || i == name_len- 1
22977 || ! isDIGIT_A(name[i-1]) || ! isDIGIT_A(name[i+1])))
22979 lookup_name[j++] = '_';
22984 /* Hyphens are skipped except under strict */
22985 if (cur == '-' && ! stricter) {
22989 /* XXX Bug in documentation. It says white space skipped adjacent to
22990 * non-word char. Maybe we should, but shouldn't skip it next to a dot
22992 if (isSPACE_A(cur) && ! stricter) {
22996 lookup_name[j++] = cur;
22998 /* Unless this is a non-trailing slash, we are done with it */
22999 if (i >= name_len - 1 || cur != '/') {
23005 /* A slash in the 'numeric value' property indicates that what follows
23006 * is a denominator. It can have a leading '+' and '0's that should be
23007 * skipped. But we have never allowed a negative denominator, so treat
23008 * a minus like every other character. (No need to rule out a second
23009 * '/', as that won't match anything anyway */
23012 if (i < name_len && name[i] == '+') {
23016 /* Skip leading zeros including underscores separating digits */
23017 for (; i < name_len - 1; i++) {
23018 if ( name[i] != '0'
23019 && (name[i] != '_' || ! isDIGIT_A(name[i+1])))
23025 /* Store the first real character in the denominator */
23026 lookup_name[j++] = name[i];
23030 /* Here are completely done parsing the input 'name', and 'lookup_name'
23031 * contains a copy, normalized.
23033 * This special case is grandfathered in: 'L_' and 'GC=L_' are accepted and
23034 * different from without the underscores. */
23035 if ( ( UNLIKELY(memEQs(lookup_name, j, "l"))
23036 || UNLIKELY(memEQs(lookup_name, j, "gc=l")))
23037 && UNLIKELY(name[name_len-1] == '_'))
23039 lookup_name[j++] = '&';
23042 /* If the original input began with 'In' or 'Is', it could be a subroutine
23043 * call to a user-defined property instead of a Unicode property name. */
23044 if ( non_pkg_begin + name_len > 2
23045 && name[non_pkg_begin+0] == 'I'
23046 && (name[non_pkg_begin+1] == 'n' || name[non_pkg_begin+1] == 's'))
23048 starts_with_In_or_Is = TRUE;
23051 could_be_user_defined = FALSE;
23054 if (could_be_user_defined) {
23057 /* If the user defined property returns the empty string, it could
23058 * easily be because the pattern is being compiled before the data it
23059 * actually needs to compile is available. This could be argued to be
23060 * a bug in the perl code, but this is a change of behavior for Perl,
23061 * so we handle it. This means that intentionally returning nothing
23062 * will not be resolved until runtime */
23063 bool empty_return = FALSE;
23065 /* Here, the name could be for a user defined property, which are
23066 * implemented as subs. */
23067 user_sub = get_cvn_flags(name, name_len, 0);
23069 const char insecure[] = "Insecure user-defined property";
23071 /* Here, there is a sub by the correct name. Normally we call it
23072 * to get the property definition */
23074 SV * user_sub_sv = MUTABLE_SV(user_sub);
23075 SV * error; /* Any error returned by calling 'user_sub' */
23076 SV * key; /* The key into the hash of user defined sub names
23079 SV ** saved_user_prop_ptr; /* Hash entry for this property */
23081 /* How many times to retry when another thread is in the middle of
23082 * expanding the same definition we want */
23083 PERL_INT_FAST8_T retry_countdown = 10;
23085 DECLARATION_FOR_GLOBAL_CONTEXT;
23087 /* If we get here, we know this property is user-defined */
23088 *user_defined_ptr = TRUE;
23090 /* We refuse to call a potentially tainted subroutine; returning an
23093 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23094 sv_catpvn(msg, insecure, sizeof(insecure) - 1);
23095 goto append_name_to_msg;
23098 /* In principal, we only call each subroutine property definition
23099 * once during the life of the program. This guarantees that the
23100 * property definition never changes. The results of the single
23101 * sub call are stored in a hash, which is used instead for future
23102 * references to this property. The property definition is thus
23103 * immutable. But, to allow the user to have a /i-dependent
23104 * definition, we call the sub once for non-/i, and once for /i,
23105 * should the need arise, passing the /i status as a parameter.
23107 * We start by constructing the hash key name, consisting of the
23108 * fully qualified subroutine name, preceded by the /i status, so
23109 * that there is a key for /i and a different key for non-/i */
23110 key = newSVpvn(((to_fold) ? "1" : "0"), 1);
23111 fq_name = S_get_fq_name(aTHX_ name, name_len, is_utf8,
23112 non_pkg_begin != 0);
23113 sv_catsv(key, fq_name);
23116 /* We only call the sub once throughout the life of the program
23117 * (with the /i, non-/i exception noted above). That means the
23118 * hash must be global and accessible to all threads. It is
23119 * created at program start-up, before any threads are created, so
23120 * is accessible to all children. But this creates some
23123 * 1) The keys can't be shared, or else problems arise; sharing is
23124 * turned off at hash creation time
23125 * 2) All SVs in it are there for the remainder of the life of the
23126 * program, and must be created in the same interpreter context
23127 * as the hash, or else they will be freed from the wrong pool
23128 * at global destruction time. This is handled by switching to
23129 * the hash's context to create each SV going into it, and then
23130 * immediately switching back
23131 * 3) All accesses to the hash must be controlled by a mutex, to
23132 * prevent two threads from getting an unstable state should
23133 * they simultaneously be accessing it. The code below is
23134 * crafted so that the mutex is locked whenever there is an
23135 * access and unlocked only when the next stable state is
23138 * The hash stores either the definition of the property if it was
23139 * valid, or, if invalid, the error message that was raised. We
23140 * use the type of SV to distinguish.
23142 * There's also the need to guard against the definition expansion
23143 * from infinitely recursing. This is handled by storing the aTHX
23144 * of the expanding thread during the expansion. Again the SV type
23145 * is used to distinguish this from the other two cases. If we
23146 * come to here and the hash entry for this property is our aTHX,
23147 * it means we have recursed, and the code assumes that we would
23148 * infinitely recurse, so instead stops and raises an error.
23149 * (Any recursion has always been treated as infinite recursion in
23152 * If instead, the entry is for a different aTHX, it means that
23153 * that thread has gotten here first, and hasn't finished expanding
23154 * the definition yet. We just have to wait until it is done. We
23155 * sleep and retry a few times, returning an error if the other
23156 * thread doesn't complete. */
23159 USER_PROP_MUTEX_LOCK;
23161 /* If we have an entry for this key, the subroutine has already
23162 * been called once with this /i status. */
23163 saved_user_prop_ptr = hv_fetch(PL_user_def_props,
23164 SvPVX(key), SvCUR(key), 0);
23165 if (saved_user_prop_ptr) {
23167 /* If the saved result is an inversion list, it is the valid
23168 * definition of this property */
23169 if (is_invlist(*saved_user_prop_ptr)) {
23170 prop_definition = *saved_user_prop_ptr;
23172 /* The SV in the hash won't be removed until global
23173 * destruction, so it is stable and we can unlock */
23174 USER_PROP_MUTEX_UNLOCK;
23176 /* The caller shouldn't try to free this SV */
23177 return prop_definition;
23180 /* Otherwise, if it is a string, it is the error message
23181 * that was returned when we first tried to evaluate this
23182 * property. Fail, and append the message */
23183 if (SvPOK(*saved_user_prop_ptr)) {
23184 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23185 sv_catsv(msg, *saved_user_prop_ptr);
23187 /* The SV in the hash won't be removed until global
23188 * destruction, so it is stable and we can unlock */
23189 USER_PROP_MUTEX_UNLOCK;
23194 assert(SvIOK(*saved_user_prop_ptr));
23196 /* Here, we have an unstable entry in the hash. Either another
23197 * thread is in the middle of expanding the property's
23198 * definition, or we are ourselves recursing. We use the aTHX
23199 * in it to distinguish */
23200 if (SvIV(*saved_user_prop_ptr) != PTR2IV(CUR_CONTEXT)) {
23202 /* Here, it's another thread doing the expanding. We've
23203 * looked as much as we are going to at the contents of the
23204 * hash entry. It's safe to unlock. */
23205 USER_PROP_MUTEX_UNLOCK;
23207 /* Retry a few times */
23208 if (retry_countdown-- > 0) {
23213 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23214 sv_catpvs(msg, "Timeout waiting for another thread to "
23216 goto append_name_to_msg;
23219 /* Here, we are recursing; don't dig any deeper */
23220 USER_PROP_MUTEX_UNLOCK;
23222 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23224 "Infinite recursion in user-defined property");
23225 goto append_name_to_msg;
23228 /* Here, this thread has exclusive control, and there is no entry
23229 * for this property in the hash. So we have the go ahead to
23230 * expand the definition ourselves. */
23232 PUSHSTACKi(PERLSI_MAGIC);
23235 /* Create a temporary placeholder in the hash to detect recursion
23237 SWITCH_TO_GLOBAL_CONTEXT;
23238 placeholder= newSVuv(PTR2IV(ORIGINAL_CONTEXT));
23239 (void) hv_store_ent(PL_user_def_props, key, placeholder, 0);
23242 /* Now that we have a placeholder, we can let other threads
23244 USER_PROP_MUTEX_UNLOCK;
23246 /* Make sure the placeholder always gets destroyed */
23247 SAVEDESTRUCTOR_X(S_delete_recursion_entry, SvPVX(key));
23252 /* Call the user's function, with the /i status as a parameter.
23253 * Note that we have gone to a lot of trouble to keep this call
23254 * from being within the locked mutex region. */
23255 XPUSHs(boolSV(to_fold));
23258 /* The following block was taken from swash_init(). Presumably
23259 * they apply to here as well, though we no longer use a swash --
23263 /* We might get here via a subroutine signature which uses a utf8
23264 * parameter name, at which point PL_subname will have been set
23265 * but not yet used. */
23266 save_item(PL_subname);
23268 (void) call_sv(user_sub_sv, G_EVAL|G_SCALAR);
23273 if (TAINT_get || SvTRUE(error)) {
23274 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23275 if (SvTRUE(error)) {
23276 sv_catpvs(msg, "Error \"");
23277 sv_catsv(msg, error);
23278 sv_catpvs(msg, "\"");
23281 if (SvTRUE(error)) sv_catpvs(msg, "; ");
23282 sv_catpvn(msg, insecure, sizeof(insecure) - 1);
23285 if (name_len > 0) {
23286 sv_catpvs(msg, " in expansion of ");
23287 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8,
23293 prop_definition = NULL;
23295 else { /* G_SCALAR guarantees a single return value */
23296 SV * contents = POPs;
23298 /* The contents is supposed to be the expansion of the property
23299 * definition. If the definition is deferrable, and we got an
23300 * empty string back, set a flag to later defer it (after clean
23303 && (! SvPOK(contents) || SvCUR(contents) == 0))
23305 empty_return = TRUE;
23307 else { /* Otherwise, call a function to check for valid syntax,
23310 prop_definition = handle_user_defined_property(
23312 is_utf8, to_fold, runtime,
23314 contents, user_defined_ptr,
23320 /* Here, we have the results of the expansion. Delete the
23321 * placeholder, and if the definition is now known, replace it with
23322 * that definition. We need exclusive access to the hash, and we
23323 * can't let anyone else in, between when we delete the placeholder
23324 * and add the permanent entry */
23325 USER_PROP_MUTEX_LOCK;
23327 S_delete_recursion_entry(aTHX_ SvPVX(key));
23329 if ( ! empty_return
23330 && (! prop_definition || is_invlist(prop_definition)))
23332 /* If we got success we use the inversion list defining the
23333 * property; otherwise use the error message */
23334 SWITCH_TO_GLOBAL_CONTEXT;
23335 (void) hv_store_ent(PL_user_def_props,
23338 ? newSVsv(prop_definition)
23344 /* All done, and the hash now has a permanent entry for this
23345 * property. Give up exclusive control */
23346 USER_PROP_MUTEX_UNLOCK;
23352 if (empty_return) {
23353 goto definition_deferred;
23356 if (prop_definition) {
23358 /* If the definition is for something not known at this time,
23359 * we toss it, and go return the main property name, as that's
23360 * the one the user will be aware of */
23361 if (! is_invlist(prop_definition)) {
23362 SvREFCNT_dec_NN(prop_definition);
23363 goto definition_deferred;
23366 sv_2mortal(prop_definition);
23370 return prop_definition;
23372 } /* End of calling the subroutine for the user-defined property */
23373 } /* End of it could be a user-defined property */
23375 /* Here it wasn't a user-defined property that is known at this time. See
23376 * if it is a Unicode property */
23378 lookup_len = j; /* This is a more mnemonic name than 'j' */
23380 /* Get the index into our pointer table of the inversion list corresponding
23381 * to the property */
23382 table_index = match_uniprop((U8 *) lookup_name, lookup_len);
23384 /* If it didn't find the property ... */
23385 if (table_index == 0) {
23387 /* Try again stripping off any initial 'In' or 'Is' */
23388 if (starts_with_In_or_Is) {
23394 table_index = match_uniprop((U8 *) lookup_name, lookup_len);
23397 if (table_index == 0) {
23400 /* Here, we didn't find it. If not a numeric type property, and
23401 * can't be a user-defined one, it isn't a legal property */
23402 if (! is_nv_type) {
23403 if (! could_be_user_defined) {
23407 /* Here, the property name is legal as a user-defined one. At
23408 * compile time, it might just be that the subroutine for that
23409 * property hasn't been encountered yet, but at runtime, it's
23410 * an error to try to use an undefined one */
23411 if (! deferrable) {
23412 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23413 sv_catpvs(msg, "Unknown user-defined property name");
23414 goto append_name_to_msg;
23417 goto definition_deferred;
23418 } /* End of isn't a numeric type property */
23420 /* The numeric type properties need more work to decide. What we
23421 * do is make sure we have the number in canonical form and look
23424 if (slash_pos < 0) { /* No slash */
23426 /* When it isn't a rational, take the input, convert it to a
23427 * NV, then create a canonical string representation of that
23432 /* Get the value */
23433 if (my_atof3(lookup_name + equals_pos, &value,
23434 lookup_len - equals_pos)
23435 != lookup_name + lookup_len)
23440 /* If the value is an integer, the canonical value is integral
23442 if (Perl_ceil(value) == value) {
23443 canonical = Perl_form(aTHX_ "%.*s%.0" NVff,
23444 equals_pos, lookup_name, value);
23446 else { /* Otherwise, it is %e with a known precision */
23449 canonical = Perl_form(aTHX_ "%.*s%.*" NVef,
23450 equals_pos, lookup_name,
23451 PL_E_FORMAT_PRECISION, value);
23453 /* The exponent generated is expecting two digits, whereas
23454 * %e on some systems will generate three. Remove leading
23455 * zeros in excess of 2 from the exponent. We start
23456 * looking for them after the '=' */
23457 exp_ptr = strchr(canonical + equals_pos, 'e');
23459 char * cur_ptr = exp_ptr + 2; /* past the 'e[+-]' */
23460 SSize_t excess_exponent_len = strlen(cur_ptr) - 2;
23462 assert(*(cur_ptr - 1) == '-' || *(cur_ptr - 1) == '+');
23464 if (excess_exponent_len > 0) {
23465 SSize_t leading_zeros = strspn(cur_ptr, "0");
23466 SSize_t excess_leading_zeros
23467 = MIN(leading_zeros, excess_exponent_len);
23468 if (excess_leading_zeros > 0) {
23469 Move(cur_ptr + excess_leading_zeros,
23471 strlen(cur_ptr) - excess_leading_zeros
23472 + 1, /* Copy the NUL as well */
23479 else { /* Has a slash. Create a rational in canonical form */
23480 UV numerator, denominator, gcd, trial;
23481 const char * end_ptr;
23482 const char * sign = "";
23484 /* We can't just find the numerator, denominator, and do the
23485 * division, then use the method above, because that is
23486 * inexact. And the input could be a rational that is within
23487 * epsilon (given our precision) of a valid rational, and would
23488 * then incorrectly compare valid.
23490 * We're only interested in the part after the '=' */
23491 const char * this_lookup_name = lookup_name + equals_pos;
23492 lookup_len -= equals_pos;
23493 slash_pos -= equals_pos;
23495 /* Handle any leading minus */
23496 if (this_lookup_name[0] == '-') {
23498 this_lookup_name++;
23503 /* Convert the numerator to numeric */
23504 end_ptr = this_lookup_name + slash_pos;
23505 if (! grok_atoUV(this_lookup_name, &numerator, &end_ptr)) {
23509 /* It better have included all characters before the slash */
23510 if (*end_ptr != '/') {
23514 /* Set to look at just the denominator */
23515 this_lookup_name += slash_pos;
23516 lookup_len -= slash_pos;
23517 end_ptr = this_lookup_name + lookup_len;
23519 /* Convert the denominator to numeric */
23520 if (! grok_atoUV(this_lookup_name, &denominator, &end_ptr)) {
23524 /* It better be the rest of the characters, and don't divide by
23526 if ( end_ptr != this_lookup_name + lookup_len
23527 || denominator == 0)
23532 /* Get the greatest common denominator using
23533 http://en.wikipedia.org/wiki/Euclidean_algorithm */
23535 trial = denominator;
23536 while (trial != 0) {
23538 trial = gcd % trial;
23542 /* If already in lowest possible terms, we have already tried
23543 * looking this up */
23548 /* Reduce the rational, which should put it in canonical form
23551 denominator /= gcd;
23553 canonical = Perl_form(aTHX_ "%.*s%s%" UVuf "/%" UVuf,
23554 equals_pos, lookup_name, sign, numerator, denominator);
23557 /* Here, we have the number in canonical form. Try that */
23558 table_index = match_uniprop((U8 *) canonical, strlen(canonical));
23559 if (table_index == 0) {
23562 } /* End of still didn't find the property in our table */
23563 } /* End of didn't find the property in our table */
23565 /* Here, we have a non-zero return, which is an index into a table of ptrs.
23566 * A negative return signifies that the real index is the absolute value,
23567 * but the result needs to be inverted */
23568 if (table_index < 0) {
23569 invert_return = TRUE;
23570 table_index = -table_index;
23573 /* Out-of band indices indicate a deprecated property. The proper index is
23574 * modulo it with the table size. And dividing by the table size yields
23575 * an offset into a table constructed by regen/mk_invlists.pl to contain
23576 * the corresponding warning message */
23577 if (table_index > MAX_UNI_KEYWORD_INDEX) {
23578 Size_t warning_offset = table_index / MAX_UNI_KEYWORD_INDEX;
23579 table_index %= MAX_UNI_KEYWORD_INDEX;
23580 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED),
23581 "Use of '%.*s' in \\p{} or \\P{} is deprecated because: %s",
23582 (int) name_len, name, deprecated_property_msgs[warning_offset]);
23585 /* In a few properties, a different property is used under /i. These are
23586 * unlikely to change, so are hard-coded here. */
23588 if ( table_index == UNI_XPOSIXUPPER
23589 || table_index == UNI_XPOSIXLOWER
23590 || table_index == UNI_TITLE)
23592 table_index = UNI_CASED;
23594 else if ( table_index == UNI_UPPERCASELETTER
23595 || table_index == UNI_LOWERCASELETTER
23596 # ifdef UNI_TITLECASELETTER /* Missing from early Unicodes */
23597 || table_index == UNI_TITLECASELETTER
23600 table_index = UNI_CASEDLETTER;
23602 else if ( table_index == UNI_POSIXUPPER
23603 || table_index == UNI_POSIXLOWER)
23605 table_index = UNI_POSIXALPHA;
23609 /* Create and return the inversion list */
23610 prop_definition =_new_invlist_C_array(uni_prop_ptrs[table_index]);
23611 sv_2mortal(prop_definition);
23614 /* See if there is a private use override to add to this definition */
23616 COPHH * hinthash = (IN_PERL_COMPILETIME)
23617 ? CopHINTHASH_get(&PL_compiling)
23618 : CopHINTHASH_get(PL_curcop);
23619 SV * pu_overrides = cophh_fetch_pv(hinthash, "private_use", 0, 0);
23621 if (UNLIKELY(pu_overrides && SvPOK(pu_overrides))) {
23623 /* See if there is an element in the hints hash for this table */
23624 SV * pu_lookup = Perl_newSVpvf(aTHX_ "%d=", table_index);
23625 const char * pos = strstr(SvPVX(pu_overrides), SvPVX(pu_lookup));
23629 SV * pu_definition;
23631 SV * expanded_prop_definition =
23632 sv_2mortal(invlist_clone(prop_definition, NULL));
23634 /* If so, it's definition is the string from here to the next
23635 * \a character. And its format is the same as a user-defined
23637 pos += SvCUR(pu_lookup);
23638 pu_definition = newSVpvn(pos, strchr(pos, '\a') - pos);
23639 pu_invlist = handle_user_defined_property(lookup_name,
23642 0, /* Not folded */
23650 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23651 sv_catpvs(msg, "Insecure private-use override");
23652 goto append_name_to_msg;
23655 /* For now, as a safety measure, make sure that it doesn't
23656 * override non-private use code points */
23657 _invlist_intersection(pu_invlist, PL_Private_Use, &pu_invlist);
23659 /* Add it to the list to be returned */
23660 _invlist_union(prop_definition, pu_invlist,
23661 &expanded_prop_definition);
23662 prop_definition = expanded_prop_definition;
23663 Perl_ck_warner_d(aTHX_ packWARN(WARN_EXPERIMENTAL__PRIVATE_USE), "The private_use feature is experimental");
23668 if (invert_return) {
23669 _invlist_invert(prop_definition);
23671 return prop_definition;
23675 if (non_pkg_begin != 0) {
23676 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23677 sv_catpvs(msg, "Illegal user-defined property name");
23680 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23681 sv_catpvs(msg, "Can't find Unicode property definition");
23685 append_name_to_msg:
23687 const char * prefix = (runtime && level == 0) ? " \\p{" : " \"";
23688 const char * suffix = (runtime && level == 0) ? "}" : "\"";
23690 sv_catpv(msg, prefix);
23691 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8, name_len, name));
23692 sv_catpv(msg, suffix);
23697 definition_deferred:
23699 /* Here it could yet to be defined, so defer evaluation of this
23700 * until its needed at runtime. We need the fully qualified property name
23701 * to avoid ambiguity, and a trailing newline */
23703 fq_name = S_get_fq_name(aTHX_ name, name_len, is_utf8,
23704 non_pkg_begin != 0 /* If has "::" */
23707 sv_catpvs(fq_name, "\n");
23709 *user_defined_ptr = TRUE;
23716 * ex: set ts=8 sts=4 sw=4 et: